d/freebsd-dev
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
58883a1fe5
freebsd-dev/sys/ufs/ffs/ffs_softdep.c

5966 lines
177 KiB
C
Raw Normal View History

Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
Update to new copyright.
2000-06-22 00:29:53 +00:00
* Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
* The soft updates code is derived from the appendix of a University
* of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
* "Soft Updates: A Solution to the Metadata Update Problem in File
* Systems", CSE-TR-254-95, August 1995).
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*
Update to new copyright.
2000-06-22 00:29:53 +00:00
* Further information about soft updates can be obtained from:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*
Update to new copyright.
2000-06-22 00:29:53 +00:00
* Marshall Kirk McKusick http://www.mckusick.com/softdep/
* 1614 Oxford Street mckusick@mckusick.com
* Berkeley, CA 94709-1608 +1-510-843-9542
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* USA
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
Update to new copyright.
2000-06-22 00:29:53 +00:00
* from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
*/
Quiet a warning on the Alpha.
2002-03-15 04:06:10 +00:00
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
/*
* For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
#ifndef DIAGNOSTIC
#define DIAGNOSTIC
#endif
#ifndef DEBUG
#define DEBUG
#endif
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#include <sys/param.h>
#include <sys/kernel.h>
Keep the inlines for <sys/buf.h> happy..
1999-06-27 13:26:23 +00:00
#include <sys/systm.h>
Separate the struct bio related stuff out of <sys/buf.h> into <sys/bio.h>. <sys/bio.h> is now a prerequisite for <sys/buf.h> but it shall not be made a nested include according to bdes teachings on the subject of nested includes. Diskdrivers and similar stuff below specfs::strategy() should no longer need to include <sys/buf.> unless they need caching of data. Still a few bogus uses of struct buf to track down. Repocopy by: peter
2000-05-05 09:59:14 +00:00
#include <sys/bio.h>
Keep the inlines for <sys/buf.h> happy..
1999-06-27 13:26:23 +00:00
#include <sys/buf.h>
Update for the KDB debugger framework: o Make debugging code conditional upon KDB. o Use kdb_backtrace() instead of backtrace(). o Remove inclusion of opt_ddb.h.
2004-07-10 20:45:47 +00:00
#include <sys/kdb.h>
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#include <sys/malloc.h>
#include <sys/mount.h>
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
#include <sys/proc.h>
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
#include <sys/stat.h>
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#include <sys/syslog.h>
#include <sys/vnode.h>
Decommision miscfs/specfs/specdev.h. Most of it goes into <sys/conf.h>, a few lines into <sys/vnode.h>. Add a few fields to struct specinfo, paving the way for the fun part.
1999-08-08 18:43:05 +00:00
#include <sys/conf.h>
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#include <ufs/ufs/dir.h>
Introduce extended attribute support for FFS, allowing arbitrary (name, value) pairs to be associated with inodes. This support is used for ACLs, MAC labels, and Capabilities in the TrustedBSD security extensions, which are currently under development. In this implementation, attributes are backed to data vnodes in the style of the quota support in FFS. Support for FFS extended attributes may be enabled using the FFS_EXTATTR kernel option (disabled by default). Userland utilities and man pages will be committed in the next batch. VFS interfaces and man pages have been in the repo since 4.0-RELEASE and are unchanged. o ufs/ufs/extattr.h: UFS-specific extattr defines o ufs/ufs/ufs_extattr.c: bulk of support routines o ufs/{ufs,ffs,mfs}/*.[ch]: hooks and extattr.h includes o contrib/softupdates/ffs_softdep.c: extattr.h includes o conf/options, conf/files, i386/conf/LINT: added FFS_EXTATTR o coda/coda_vfsops.c: XXX required extattr.h due to ufsmount.h (This should not be the case, and will be fixed in a future commit) Currently attributes are not supported in MFS. This will be fixed. Reviewed by: adrian, bp, freebsd-fs, other unthanked souls Obtained from: TrustedBSD Project
2000-04-15 03:34:27 +00:00
#include <ufs/ufs/extattr.h>
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#include <ufs/ufs/quota.h>
#include <ufs/ufs/inode.h>
#include <ufs/ufs/ufsmount.h>
#include <ufs/ffs/fs.h>
#include <ufs/ffs/softdep.h>
#include <ufs/ffs/ffs_extern.h>
#include <ufs/ufs/ufs_extern.h>
/*
Merge with Kirk's version as of Feb 20 His version 9.23 == our version 1.5 of ffs_softdep.c His version 9.5 == our version 1.4 of softdep.c
1998-05-19 22:54:53 +00:00
* These definitions need to be adapted to the system to which
* this file is being ported.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
/*
* malloc types defined for the softdep system.
*/
Staticize some malloc M_ instances.
2000-12-08 20:09:00 +00:00
static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
Back out M_* changes, per decision of the TRB. Approved by: trb
2003-02-19 05:47:46 +00:00
#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
#define D_PAGEDEP 0
#define D_INODEDEP 1
#define D_NEWBLK 2
#define D_BMSAFEMAP 3
#define D_ALLOCDIRECT 4
#define D_INDIRDEP 5
#define D_ALLOCINDIR 6
#define D_FREEFRAG 7
#define D_FREEBLKS 8
#define D_FREEFILE 9
#define D_DIRADD 10
#define D_MKDIR 11
#define D_DIRREM 12
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
#define D_NEWDIRBLK 13
#define D_LAST D_NEWDIRBLK
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
/*
* translate from workitem type to memory type
* MUST match the defines above, such that memtype[D_XXX] == M_XXX
*/
static struct malloc_type *memtype[] = {
M_PAGEDEP,
M_INODEDEP,
M_NEWBLK,
M_BMSAFEMAP,
M_ALLOCDIRECT,
M_INDIRDEP,
M_ALLOCINDIR,
M_FREEFRAG,
M_FREEBLKS,
M_FREEFILE,
M_DIRADD,
M_MKDIR,
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
M_DIRREM,
M_NEWDIRBLK
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
};
#define DtoM(type) (memtype[type])
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Names of malloc types.
*/
Merge with Kirk's version as of Feb 20 His version 9.23 == our version 1.5 of ffs_softdep.c His version 9.5 == our version 1.4 of softdep.c
1998-05-19 22:54:53 +00:00
#define TYPENAME(type) \
((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
/*
* End system adaptaion definitions.
*/
/*
* Internal function prototypes.
*/
Remove __P.
2002-03-19 22:40:48 +00:00
static void softdep_error(char *, int);
static void drain_output(struct vnode *, int);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
static struct buf *getdirtybuf(struct buf **, struct mtx *, int);
Remove __P.
2002-03-19 22:40:48 +00:00
static void clear_remove(struct thread *);
static void clear_inodedeps(struct thread *);
static int flush_pagedep_deps(struct vnode *, struct mount *,
struct diraddhd *);
static int flush_inodedep_deps(struct fs *, ino_t);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
static int flush_deplist(struct allocdirectlst *, int, int *);
Remove __P.
2002-03-19 22:40:48 +00:00
static int handle_written_filepage(struct pagedep *, struct buf *);
static void diradd_inode_written(struct diradd *, struct inodedep *);
static int handle_written_inodeblock(struct inodedep *, struct buf *);
static void handle_allocdirect_partdone(struct allocdirect *);
static void handle_allocindir_partdone(struct allocindir *);
static void initiate_write_filepage(struct pagedep *, struct buf *);
static void handle_written_mkdir(struct mkdir *, int);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
Remove __P.
2002-03-19 22:40:48 +00:00
static void handle_workitem_freefile(struct freefile *);
static void handle_workitem_remove(struct dirrem *, struct vnode *);
static struct dirrem *newdirrem(struct buf *, struct inode *,
struct inode *, int, struct dirrem **);
static void free_diradd(struct diradd *);
static void free_allocindir(struct allocindir *, struct inodedep *);
static void free_newdirblk(struct newdirblk *);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
ufs2_daddr_t *);
Remove __P.
2002-03-19 22:40:48 +00:00
static void deallocate_dependencies(struct buf *, struct inodedep *);
static void free_allocdirect(struct allocdirectlst *,
struct allocdirect *, int);
static int check_inode_unwritten(struct inodedep *);
static int free_inodedep(struct inodedep *);
static void handle_workitem_freeblocks(struct freeblks *, int);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
Remove __P.
2002-03-19 22:40:48 +00:00
static void setup_allocindir_phase2(struct buf *, struct inode *,
struct allocindir *);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
ufs2_daddr_t);
Remove __P.
2002-03-19 22:40:48 +00:00
static void handle_workitem_freefrag(struct freefrag *);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
Remove __P.
2002-03-19 22:40:48 +00:00
static void allocdirect_merge(struct allocdirectlst *,
struct allocdirect *, struct allocdirect *);
static struct bmsafemap *bmsafemap_lookup(struct buf *);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
Remove __P.
2002-03-19 22:40:48 +00:00
static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
static void pause_timer(void *);
static int request_cleanup(int, int);
static int process_worklist_item(struct mount *, int);
static void add_to_worklist(struct worklist *);
Merge with Kirk's version as of Feb 20 His version 9.23 == our version 1.5 of ffs_softdep.c His version 9.5 == our version 1.4 of softdep.c
1998-05-19 22:54:53 +00:00
/*
* Exported softdep operations.
*/
When I was a kid my work table was one cluttered mess an cleaning it up were a rather overwhelming task. I soon learned that if you don't know where you're going to store something, at least try to pile it next to something slightly related in the hope that a pattern emerges. Apply the same principle to the ffs/snapshot/softupdates code which have leaked into specfs: Add yet a buf-quasi-method and call it from the only two places I can see it can make a difference and implement the magic in ffs_softdep.c where it belongs. It's not pretty, but at least it's one less layer violated.
2004-03-11 18:50:33 +00:00
static int softdep_disk_prewrite(struct vnode *vp, struct buf *bp);
Remove __P.
2002-03-19 22:40:48 +00:00
static void softdep_disk_io_initiation(struct buf *);
static void softdep_disk_write_complete(struct buf *);
static void softdep_deallocate_dependencies(struct buf *);
static void softdep_move_dependencies(struct buf *, struct buf *);
static int softdep_count_dependencies(struct buf *bp, int);
Make static non-exported functions from soft updates.
2000-01-09 22:40:09 +00:00
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Locking primitives.
*
* For a uniprocessor, all we need to do is protect against disk
* interrupts. For a multiprocessor, this lock would have to be
* a mutex. A single mutex is used throughout this file, though
* finer grain locking could be used if contention warranted it.
*
* For a multiprocessor, the sleep call would accept a lock and
* release it after the sleep processing was complete. In a uniprocessor
* implementation there is no such interlock, so we simple mark
* the places where it needs to be done with the `interlocked' form
* of the lock calls. Since the uniprocessor sleep already interlocks
* the spl, there is nothing that really needs to be done.
*/
#ifndef /* NOT */ DEBUG
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
static struct lockit {
int lkt_spl;
} lk = { 0 };
#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio()
#define FREE_LOCK(lk) splx((lk)->lkt_spl)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#else /* DEBUG */
- Fix some minor whitespace nits. - Move the SPECIAL_FLAG #define up next to the NOHOLDER #define and fix a little nit that caused it to be defined as -(sizeof (struct thread) + 1) instead of -2.
2001-09-27 21:04:13 +00:00
#define NOHOLDER ((struct thread *)-1)
#define SPECIAL_FLAG ((struct thread *)-2)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
static struct lockit {
int lkt_spl;
- Fix some minor whitespace nits. - Move the SPECIAL_FLAG #define up next to the NOHOLDER #define and fix a little nit that caused it to be defined as -(sizeof (struct thread) + 1) instead of -2.
2001-09-27 21:04:13 +00:00
struct thread *lkt_held;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
} lk = { 0, NOHOLDER };
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
Remove __P.
2002-03-19 22:40:48 +00:00
static void acquire_lock(struct lockit *);
static void free_lock(struct lockit *);
void softdep_panic(char *);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#define ACQUIRE_LOCK(lk) acquire_lock(lk)
#define FREE_LOCK(lk) free_lock(lk)
static void
acquire_lock(lk)
struct lockit *lk;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *holder;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk->lkt_held != NOHOLDER) {
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
holder = lk->lkt_held;
FREE_LOCK(lk);
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (holder == curthread)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_lock: locking against myself");
else
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
panic("softdep_lock: lock held by %p", holder);
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
lk->lkt_spl = splbio();
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk->lkt_held = curthread;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
static void
free_lock(lk)
struct lockit *lk;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk->lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_unlock: lock not held");
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk->lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
splx(lk->lkt_spl);
}
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
/*
* Function to release soft updates lock and panic.
*/
void
softdep_panic(msg)
char *msg;
{
if (lk.lkt_held != NOHOLDER)
FREE_LOCK(&lk);
panic(msg);
}
#endif /* DEBUG */
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
static int interlocked_sleep(struct lockit *, int, void *, struct mtx *, int,
Remove __P.
2002-03-19 22:40:48 +00:00
const char *, int);
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
/*
* When going to sleep, we must save our SPL so that it does
* not get lost if some other process uses the lock while we
* are sleeping. We restore it after we have slept. This routine
* wraps the interlocking with functions that sleep. The list
* below enumerates the available set of operations.
*/
#define UNKNOWN 0
#define SLEEP 1
#define LOCKBUF 2
static int
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
interlocked_sleep(lk, op, ident, mtx, flags, wmesg, timo)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct lockit *lk;
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
int op;
void *ident;
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
struct mtx *mtx;
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
int flags;
const char *wmesg;
int timo;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *holder;
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
int s, retval;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
s = lk->lkt_spl;
# ifdef DEBUG
if (lk->lkt_held == NOHOLDER)
panic("interlocked_sleep: lock not held");
lk->lkt_held = NOHOLDER;
# endif /* DEBUG */
switch (op) {
case SLEEP:
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
retval = msleep(ident, mtx, flags, wmesg, timo);
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
break;
case LOCKBUF:
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
retval = BUF_LOCK((struct buf *)ident, flags, mtx);
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
break;
default:
panic("interlocked_sleep: unknown operation");
}
# ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk->lkt_held != NOHOLDER) {
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
holder = lk->lkt_held;
FREE_LOCK(lk);
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (holder == curthread)
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
panic("interlocked_sleep: locking against self");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
else
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
panic("interlocked_sleep: lock held by %p", holder);
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
}
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk->lkt_held = curthread;
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
# endif /* DEBUG */
lk->lkt_spl = s;
return (retval);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Place holder for real semaphores.
*/
struct sema {
int value;
- Fix some minor whitespace nits. - Move the SPECIAL_FLAG #define up next to the NOHOLDER #define and fix a little nit that caused it to be defined as -(sizeof (struct thread) + 1) instead of -2.
2001-09-27 21:04:13 +00:00
struct thread *holder;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
char *name;
int prio;
int timo;
};
Remove __P.
2002-03-19 22:40:48 +00:00
static void sema_init(struct sema *, char *, int, int);
static int sema_get(struct sema *, struct lockit *);
static void sema_release(struct sema *);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
static void
sema_init(semap, name, prio, timo)
struct sema *semap;
char *name;
int prio, timo;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
semap->holder = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
semap->value = 0;
semap->name = name;
semap->prio = prio;
semap->timo = timo;
}
static int
sema_get(semap, interlock)
struct sema *semap;
struct lockit *interlock;
{
if (semap->value++ > 0) {
if (interlock != NULL) {
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
NULL, semap->prio, semap->name,
semap->timo);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(interlock);
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
} else {
More low-hanging fruit: kill caddr_t in calls to wakeup(9) / [mt]sleep(9).
2003-03-02 16:54:40 +00:00
tsleep(semap, semap->prio, semap->name,
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
semap->timo);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
return (0);
}
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
semap->holder = curthread;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (interlock != NULL)
FREE_LOCK(interlock);
return (1);
}
static void
sema_release(semap)
struct sema *semap;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (semap->value <= 0 || semap->holder != curthread) {
if (lk.lkt_held != NOHOLDER)
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("sema_release: not held");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (--semap->value > 0) {
semap->value = 0;
wakeup(semap);
}
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
semap->holder = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Worklist queue management.
* These routines require that the lock be held.
*/
#ifndef /* NOT */ DEBUG
#define WORKLIST_INSERT(head, item) do { \
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
(item)->wk_state |= ONWORKLIST; \
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_INSERT_HEAD(head, item, wk_list); \
} while (0)
#define WORKLIST_REMOVE(item) do { \
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
(item)->wk_state &= ~ONWORKLIST; \
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_REMOVE(item, wk_list); \
} while (0)
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
#define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#else /* DEBUG */
Remove __P.
2002-03-19 22:40:48 +00:00
static void worklist_insert(struct workhead *, struct worklist *);
static void worklist_remove(struct worklist *);
static void workitem_free(struct worklist *, int);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#define WORKLIST_INSERT(head, item) worklist_insert(head, item)
#define WORKLIST_REMOVE(item) worklist_remove(item)
#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
static void
worklist_insert(head, item)
struct workhead *head;
struct worklist *item;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("worklist_insert: lock not held");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (item->wk_state & ONWORKLIST) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("worklist_insert: already on list");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
item->wk_state |= ONWORKLIST;
LIST_INSERT_HEAD(head, item, wk_list);
}
static void
worklist_remove(item)
struct worklist *item;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("worklist_remove: lock not held");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((item->wk_state & ONWORKLIST) == 0) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("worklist_remove: not on list");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
item->wk_state &= ~ONWORKLIST;
LIST_REMOVE(item, wk_list);
}
static void
workitem_free(item, type)
struct worklist *item;
int type;
{
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (item->wk_state & ONWORKLIST) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held != NOHOLDER)
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("workitem_free: still on list");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
if (item->wk_type != type) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held != NOHOLDER)
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("workitem_free: type mismatch");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
FREE(item, DtoM(type));
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
#endif /* DEBUG */
/*
* Workitem queue management
*/
static struct workhead softdep_workitem_pending;
When removing the last item from a non-empty worklist, the worklist tail pointer must be updated. Reported by: Kris Kennaway <kris@obsecurity.org> Sponsored by: DARPA & NAI Labs.
2003-02-24 07:28:41 +00:00
static struct worklist *worklist_tail;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
static int num_on_worklist; /* number of worklist items to be processed */
static int softdep_worklist_busy; /* 1 => trying to do unmount */
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
static int softdep_worklist_req; /* serialized waiters */
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
static int max_softdeps; /* maximum number of structs before slowdown */
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
static int maxindirdeps = 50; /* max number of indirdeps before slowdown */
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
static int tickdelay = 2; /* number of ticks to pause during slowdown */
static int proc_waiting; /* tracks whether we have a timeout posted */
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
static int *stat_countp; /* statistic to count in proc_waiting timeout */
static struct callout_handle handle; /* handle on posted proc_waiting timeout */
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
static struct thread *filesys_syncer; /* proc of filesystem syncer process */
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
static int req_clear_inodedeps; /* syncer process flush some inodedeps */
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
#define FLUSH_INODES 1
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
static int req_clear_remove; /* syncer process flush some freeblks */
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
#define FLUSH_REMOVE 2
#define FLUSH_REMOVE_WAIT 3
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
/*
* runtime statistics
*/
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
static int stat_worklist_push; /* number of worklist cleanups */
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
static int stat_blk_limit_push; /* number of times block limit neared */
static int stat_ino_limit_push; /* number of times inode limit neared */
static int stat_blk_limit_hit; /* number of times block slowdown imposed */
static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
#ifdef DEBUG
#include <vm/vm.h>
#include <sys/sysctl.h>
SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
#endif /* DEBUG */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Add an item to the end of the work queue.
* This routine requires that the lock be held.
* This is the only routine that adds items to the list.
* The following routine is the only one that removes items
* and does so in order from first to last.
*/
static void
add_to_worklist(wk)
struct worklist *wk;
{
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (wk->wk_state & ONWORKLIST) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held != NOHOLDER)
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("add_to_worklist: already on list");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
wk->wk_state |= ONWORKLIST;
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
if (LIST_FIRST(&softdep_workitem_pending) == NULL)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
else
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
worklist_tail = wk;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
num_on_worklist += 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Process that runs once per second to handle items in the background queue.
*
* Note that we ensure that everything is done in the order in which they
* appear in the queue. The code below depends on this property to ensure
* that blocks of a file are freed before the inode itself is freed. This
* ordering ensures that no new <vfsid, inum, lbn> triples will be generated
* until all the old ones have been purged from the dependency lists.
*/
ARGH! I have too many source trees :-( Fix prototype errors in last commit.
2000-06-16 13:00:33 +00:00
int
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_process_worklist(matchmnt)
struct mount *matchmnt;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td = curthread;
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
int cnt, matchcnt, loopcount;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
long starttime;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
/*
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
* Record the process identifier of our caller so that we can give
* this process preferential treatment in request_cleanup below.
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
*/
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
filesys_syncer = td;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
matchcnt = 0;
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* There is no danger of having multiple processes run this
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
* code, but we have to single-thread it when softdep_flushfiles()
* is in operation to get an accurate count of the number of items
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* related to its mount point that are in the list.
*/
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
if (matchmnt == NULL) {
if (softdep_worklist_busy < 0)
return(-1);
softdep_worklist_busy += 1;
}
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* If requested, try removing inode or removal dependencies.
*/
if (req_clear_inodedeps) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
clear_inodedeps(td);
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
req_clear_inodedeps -= 1;
wakeup_one(&proc_waiting);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
if (req_clear_remove) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
clear_remove(td);
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
req_clear_remove -= 1;
wakeup_one(&proc_waiting);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
loopcount = 1;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
starttime = time_second;
while (num_on_worklist > 0) {
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if ((cnt = process_worklist_item(matchmnt, 0)) == -1)
break;
else
matchcnt += cnt;
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
/*
* If a umount operation wants to run the worklist
* accurately, abort.
*/
if (softdep_worklist_req && matchmnt == NULL) {
matchcnt = -1;
break;
}
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* If requested, try removing inode or removal dependencies.
*/
if (req_clear_inodedeps) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
clear_inodedeps(td);
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
req_clear_inodedeps -= 1;
wakeup_one(&proc_waiting);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
if (req_clear_remove) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
clear_remove(td);
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
req_clear_remove -= 1;
wakeup_one(&proc_waiting);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
/*
* We do not generally want to stop for buffer space, but if
* we are really being a buffer hog, we will stop and wait.
*/
if (loopcount++ % 128 == 0)
bwillwrite();
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
/*
* Never allow processing to run for more than one
* second. Otherwise the other syncer tasks may get
* excessively backlogged.
*/
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
if (starttime != time_second && matchmnt == NULL) {
matchcnt = -1;
break;
}
}
if (matchmnt == NULL) {
softdep_worklist_busy -= 1;
if (softdep_worklist_req && softdep_worklist_busy == 0)
wakeup(&softdep_worklist_req);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
return (matchcnt);
}
/*
* Process one item on the worklist.
*/
static int
process_worklist_item(matchmnt, flags)
struct mount *matchmnt;
int flags;
{
When removing the last item from a non-empty worklist, the worklist tail pointer must be updated. Reported by: Kris Kennaway <kris@obsecurity.org> Sponsored by: DARPA & NAI Labs.
2003-02-24 07:28:41 +00:00
struct worklist *wk, *wkend;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
struct mount *mp;
struct vnode *vp;
int matchcnt = 0;
We must be careful to avoid recursive copy-on-write faults when trying to clean up during disk-full senarios. Sponsored by: DARPA & NAI Labs.
2002-10-23 21:47:02 +00:00
/*
* If we are being called because of a process doing a
* copy-on-write, then it is not safe to write as we may
* recurse into the copy-on-write routine.
*/
Move the P_COWINPROGRESS flag from being a per-process p_flag to being a per-thread td_pflag which doesn't require any locks to read or write as it is only read or written by curthread on itself. Glanced at by: mckusick
2003-10-23 21:14:08 +00:00
if (curthread->td_pflags & TDP_COWINPROGRESS)
We must be careful to avoid recursive copy-on-write faults when trying to clean up during disk-full senarios. Sponsored by: DARPA & NAI Labs.
2002-10-23 21:47:02 +00:00
return (-1);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
ACQUIRE_LOCK(&lk);
/*
* Normally we just process each item on the worklist in order.
* However, if we are in a situation where we cannot lock any
* inodes, we have to skip over any dirrem requests whose
* vnodes are resident and locked.
*/
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
vp = NULL;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if (wk->wk_state & INPROGRESS)
continue;
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
break;
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
wk->wk_state |= INPROGRESS;
FREE_LOCK(&lk);
VFS_VGET(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum,
LK_NOWAIT | LK_EXCLUSIVE, &vp);
ACQUIRE_LOCK(&lk);
wk->wk_state &= ~INPROGRESS;
if (vp != NULL)
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
break;
}
Free lock before returning from process_worklist_item. Obtained from: Constantine Sapuntzakis <csapuntz@stanford.edu>
2001-03-01 21:43:46 +00:00
if (wk == 0) {
FREE_LOCK(&lk);
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
return (-1);
Free lock before returning from process_worklist_item. Obtained from: Constantine Sapuntzakis <csapuntz@stanford.edu>
2001-03-01 21:43:46 +00:00
}
When removing the last item from a non-empty worklist, the worklist tail pointer must be updated. Reported by: Kris Kennaway <kris@obsecurity.org> Sponsored by: DARPA & NAI Labs.
2003-02-24 07:28:41 +00:00
/*
* Remove the item to be processed. If we are removing the last
* item on the list, we need to recalculate the tail pointer.
* As this happens rarely and usually when the list is short,
* we just run down the list to find it rather than tracking it
* in the above loop.
*/
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
WORKLIST_REMOVE(wk);
When removing the last item from a non-empty worklist, the worklist tail pointer must be updated. Reported by: Kris Kennaway <kris@obsecurity.org> Sponsored by: DARPA & NAI Labs.
2003-02-24 07:28:41 +00:00
if (wk == worklist_tail) {
LIST_FOREACH(wkend, &softdep_workitem_pending, wk_list)
if (LIST_NEXT(wkend, wk_list) == NULL)
break;
worklist_tail = wkend;
}
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
num_on_worklist -= 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
switch (wk->wk_type) {
case D_DIRREM:
/* removal of a directory entry */
mp = WK_DIRREM(wk)->dm_mnt;
if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
panic("%s: dirrem on suspended filesystem",
"process_worklist_item");
if (mp == matchmnt)
matchcnt += 1;
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
handle_workitem_remove(WK_DIRREM(wk), vp);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
break;
case D_FREEBLKS:
/* releasing blocks and/or fragments from a file */
mp = WK_FREEBLKS(wk)->fb_mnt;
if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
panic("%s: freeblks on suspended filesystem",
"process_worklist_item");
if (mp == matchmnt)
matchcnt += 1;
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
break;
case D_FREEFRAG:
/* releasing a fragment when replaced as a file grows */
mp = WK_FREEFRAG(wk)->ff_mnt;
if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
panic("%s: freefrag on suspended filesystem",
"process_worklist_item");
if (mp == matchmnt)
matchcnt += 1;
handle_workitem_freefrag(WK_FREEFRAG(wk));
break;
case D_FREEFILE:
/* releasing an inode when its link count drops to 0 */
mp = WK_FREEFILE(wk)->fx_mnt;
if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
panic("%s: freefile on suspended filesystem",
"process_worklist_item");
if (mp == matchmnt)
matchcnt += 1;
handle_workitem_freefile(WK_FREEFILE(wk));
break;
default:
panic("%s_process_worklist: Unknown type %s",
"softdep", TYPENAME(wk->wk_type));
/* NOTREACHED */
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (matchcnt);
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
/*
* Move dependencies from one buffer to another.
*/
static void
softdep_move_dependencies(oldbp, newbp)
struct buf *oldbp;
struct buf *newbp;
{
struct worklist *wk, *wktail;
if (LIST_FIRST(&newbp->b_dep) != NULL)
panic("softdep_move_dependencies: need merge code");
wktail = 0;
ACQUIRE_LOCK(&lk);
Non-operational change, fix compiler warning. Reviewed by: mckusick
2000-01-14 04:39:28 +00:00
while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
LIST_REMOVE(wk, wk_list);
if (wktail == 0)
LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
else
LIST_INSERT_AFTER(wktail, wk, wk_list);
wktail = wk;
}
FREE_LOCK(&lk);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Purge the work list of all items associated with a particular mount point.
*/
int
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
softdep_flushworklist(oldmnt, countp, td)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct mount *oldmnt;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
int *countp;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct vnode *devvp;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
int count, error = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
* Await our turn to clear out the queue, then serialize access.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
while (softdep_worklist_busy) {
softdep_worklist_req += 1;
tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
softdep_worklist_req -= 1;
}
softdep_worklist_busy = -1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Alternately flush the block device associated with the mount
* point and process any dependencies that the flushing
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
* creates. We continue until no more worklist dependencies
* are found.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
*countp = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
devvp = VFSTOUFS(oldmnt)->um_devvp;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
while ((count = softdep_process_worklist(oldmnt)) > 0) {
*countp += count;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
error = VOP_FSYNC(devvp, td->td_ucred, MNT_WAIT, td);
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
VOP_UNLOCK(devvp, 0, td);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (error)
break;
}
softdep_worklist_busy = 0;
Fix a race between the syncer and umount. When you umount a softupdates filesystem softdep_process_worklist() is called in a loop until it indicates that no dependancies remain, but the determination of that fact depends on there only being one softdep_process_worklist() instance running. It was possible for the syncer to also be running softdep_process_worklist() and the pre-existing checks in the code to prevent this were not sufficient to prevent the race. This patch solves the problem. Approved-by: mckusick
2001-01-30 06:31:59 +00:00
if (softdep_worklist_req)
wakeup(&softdep_worklist_req);
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
return (error);
}
/*
* Flush all vnodes and worklist items associated with a specified mount point.
*/
int
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
softdep_flushfiles(oldmnt, flags, td)
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
struct mount *oldmnt;
int flags;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
{
int error, count, loopcnt;
Quiet a warning on the Alpha.
2002-03-15 04:06:10 +00:00
error = 0;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
/*
* Alternately flush the vnodes associated with the mount
* point and process any dependencies that the flushing
* creates. In theory, this loop can happen at most twice,
* but we give it a few extra just to be sure.
*/
for (loopcnt = 10; loopcnt > 0; loopcnt--) {
/*
* Do another flush in case any vnodes were brought in
* as part of the cleanup operations.
*/
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
break;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
count == 0)
break;
}
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
/*
* If we are unmounting then it is an error to fail. If we
* are simply trying to downgrade to read-only, then filesystem
* activity can keep us busy forever, so we just fail with EBUSY.
*/
if (loopcnt == 0) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
panic("softdep_flushfiles: looping");
error = EBUSY;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (error);
}
/*
* Structure hashing.
*
* There are three types of structures that can be looked up:
* 1) pagedep structures identified by mount point, inode number,
* and logical block.
* 2) inodedep structures identified by mount point and inode number.
* 3) newblk structures identified by mount point and
* physical block number.
*
* The "pagedep" and "inodedep" dependency structures are hashed
* separately from the file blocks and inodes to which they correspond.
* This separation helps when the in-memory copy of an inode or
* file block must be replaced. It also obviates the need to access
* an inode or file page when simply updating (or de-allocating)
* dependency structures. Lookup of newblk structures is needed to
* find newly allocated blocks when trying to associate them with
* their allocdirect or allocindir structure.
*
* The lookup routines optionally create and hash a new instance when
* an existing entry is not found.
*/
#define DEPALLOC 0x0001 /* allocate structure if lookup fails */
This patch corrects two problems with the rate limiting code that was introduced in revision 1.80. The problem manifested itself with a `locking against myself' panic and could also result in soft updates inconsistences associated with inodedeps. The two problems are: 1) One of the background operations could manipulate the bitmap while holding it locked with intent to create. This held lock results in a `locking against myself' panic, when the background processing that we have been coopted to do tries to lock the bitmap which we are already holding locked. To understand how to fix this problem, first, observe that we can do the background cleanups in inodedep_lookup only when allocating inodedeps (DEPALLOC is set in the call to inodedep_lookup). Second observe that calls to inodedep_lookup with DEPALLOC set can only happen from the following calls into the softdep code: softdep_setup_inomapdep softdep_setup_allocdirect softdep_setup_remove softdep_setup_freeblocks softdep_setup_directory_change softdep_setup_directory_add softdep_change_linkcnt Only the first two of these can come from ffs_alloc.c while holding a bitmap locked. Thus, inodedep_lookup must not go off to do request_cleanups when being called from these functions. This change adds a flag, NODELAY, that can be passed to inodedep_lookup to let it know that it should not do background processing in those cases. 2) The return value from request_cleanup when helping out with the cleanup was 0 instead of 1. This meant that despite the fact that we may have slept while doing the cleanups, the code did not recheck for the appearance of an inodedep (e.g., goto top in inodedep_lookup). This lead to the softdep inconsistency in which we ended up with two inodedep's for the same inode. Reviewed by: Peter Wemm <peter@yahoo-inc.com>, Matt Dillon <dillon@earth.backplane.com>
2001-02-20 11:14:38 +00:00
#define NODELAY 0x0002 /* cannot do background work */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Structures and routines associated with pagedep caching.
*/
Back out the previous change to the queue(3) interface. It was not discussed and should probably not happen. Requested by: msmith and others
2000-05-26 02:09:24 +00:00
LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
u_long pagedep_hash; /* size of hash table - 1 */
#define PAGEDEP_HASH(mp, inum, lbn) \
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
pagedep_hash])
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
static struct sema pagedep_in_progress;
/*
Occationally deleted files would hang around for hours or days without being reclaimed. This bug was introduced in revision 1.95 dealing with filenames placed in newly allocated directory blocks, thus is not present in 4.X systems. The bug is triggered when a new entry is made in a directory after the data block containing the original new entry has been written, but before the inode that references the data block has been written. Submitted by: Bill Fenner <fenner@research.att.com>
2002-02-07 00:54:32 +00:00
* Look up a pagedep. Return 1 if found, 0 if not found or found
* when asked to allocate but not associated with any buffer.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* If not found, allocate if DEPALLOC flag is passed.
* Found or allocated entry is returned in pagedeppp.
* This routine must be called with splbio interrupts blocked.
*/
static int
pagedep_lookup(ip, lbn, flags, pagedeppp)
struct inode *ip;
ufs_lbn_t lbn;
int flags;
struct pagedep **pagedeppp;
{
struct pagedep *pagedep;
struct pagedep_hashhead *pagedephd;
struct mount *mp;
int i;
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("pagedep_lookup: lock not held");
#endif
mp = ITOV(ip)->v_mount;
pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
top:
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(pagedep, pagedephd, pd_hash)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (ip->i_number == pagedep->pd_ino &&
lbn == pagedep->pd_lbn &&
mp == pagedep->pd_mnt)
break;
if (pagedep) {
*pagedeppp = pagedep;
Occationally deleted files would hang around for hours or days without being reclaimed. This bug was introduced in revision 1.95 dealing with filenames placed in newly allocated directory blocks, thus is not present in 4.X systems. The bug is triggered when a new entry is made in a directory after the data block containing the original new entry has been written, but before the inode that references the data block has been written. Submitted by: Bill Fenner <fenner@research.att.com>
2002-02-07 00:54:32 +00:00
if ((flags & DEPALLOC) != 0 &&
(pagedep->pd_state & ONWORKLIST) == 0)
return (0);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (1);
}
if ((flags & DEPALLOC) == 0) {
*pagedeppp = NULL;
return (0);
}
if (sema_get(&pagedep_in_progress, &lk) == 0) {
ACQUIRE_LOCK(&lk);
goto top;
}
MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
pagedep->pd_list.wk_type = D_PAGEDEP;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
pagedep->pd_mnt = mp;
pagedep->pd_ino = ip->i_number;
pagedep->pd_lbn = lbn;
LIST_INIT(&pagedep->pd_dirremhd);
LIST_INIT(&pagedep->pd_pendinghd);
for (i = 0; i < DAHASHSZ; i++)
LIST_INIT(&pagedep->pd_diraddhd[i]);
ACQUIRE_LOCK(&lk);
LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
sema_release(&pagedep_in_progress);
*pagedeppp = pagedep;
return (0);
}
/*
* Structures and routines associated with inodedep caching.
*/
Back out the previous change to the queue(3) interface. It was not discussed and should probably not happen. Requested by: msmith and others
2000-05-26 02:09:24 +00:00
LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
static u_long inodedep_hash; /* size of hash table - 1 */
static long num_inodedep; /* number of inodedep allocated */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#define INODEDEP_HASH(fs, inum) \
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
(&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
static struct sema inodedep_in_progress;
/*
Correct typos, mostly s/ a / an / where appropriate. Some whitespace cleanup, especially in troff files.
2003-01-01 18:49:04 +00:00
* Look up an inodedep. Return 1 if found, 0 if not found.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* If not found, allocate if DEPALLOC flag is passed.
* Found or allocated entry is returned in inodedeppp.
* This routine must be called with splbio interrupts blocked.
*/
static int
inodedep_lookup(fs, inum, flags, inodedeppp)
struct fs *fs;
ino_t inum;
int flags;
struct inodedep **inodedeppp;
{
struct inodedep *inodedep;
struct inodedep_hashhead *inodedephd;
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
int firsttry;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("inodedep_lookup: lock not held");
#endif
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
firsttry = 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedephd = INODEDEP_HASH(fs, inum);
top:
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(inodedep, inodedephd, id_hash)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (inum == inodedep->id_ino && fs == inodedep->id_fs)
break;
if (inodedep) {
*inodedeppp = inodedep;
return (1);
}
if ((flags & DEPALLOC) == 0) {
*inodedeppp = NULL;
return (0);
}
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
/*
* If we are over our limit, try to improve the situation.
*/
This patch corrects two problems with the rate limiting code that was introduced in revision 1.80. The problem manifested itself with a `locking against myself' panic and could also result in soft updates inconsistences associated with inodedeps. The two problems are: 1) One of the background operations could manipulate the bitmap while holding it locked with intent to create. This held lock results in a `locking against myself' panic, when the background processing that we have been coopted to do tries to lock the bitmap which we are already holding locked. To understand how to fix this problem, first, observe that we can do the background cleanups in inodedep_lookup only when allocating inodedeps (DEPALLOC is set in the call to inodedep_lookup). Second observe that calls to inodedep_lookup with DEPALLOC set can only happen from the following calls into the softdep code: softdep_setup_inomapdep softdep_setup_allocdirect softdep_setup_remove softdep_setup_freeblocks softdep_setup_directory_change softdep_setup_directory_add softdep_change_linkcnt Only the first two of these can come from ffs_alloc.c while holding a bitmap locked. Thus, inodedep_lookup must not go off to do request_cleanups when being called from these functions. This change adds a flag, NODELAY, that can be passed to inodedep_lookup to let it know that it should not do background processing in those cases. 2) The return value from request_cleanup when helping out with the cleanup was 0 instead of 1. This meant that despite the fact that we may have slept while doing the cleanups, the code did not recheck for the appearance of an inodedep (e.g., goto top in inodedep_lookup). This lead to the softdep inconsistency in which we ended up with two inodedep's for the same inode. Reviewed by: Peter Wemm <peter@yahoo-inc.com>, Matt Dillon <dillon@earth.backplane.com>
2001-02-20 11:14:38 +00:00
if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 &&
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
request_cleanup(FLUSH_INODES, 1)) {
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
firsttry = 0;
goto top;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (sema_get(&inodedep_in_progress, &lk) == 0) {
ACQUIRE_LOCK(&lk);
goto top;
}
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
num_inodedep += 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_INODEDEP, M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
inodedep->id_list.wk_type = D_INODEDEP;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_fs = fs;
inodedep->id_ino = inum;
inodedep->id_state = ALLCOMPLETE;
inodedep->id_nlinkdelta = 0;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
inodedep->id_savedino1 = NULL;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_savedsize = -1;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
inodedep->id_savedextsize = -1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_buf = NULL;
LIST_INIT(&inodedep->id_pendinghd);
LIST_INIT(&inodedep->id_inowait);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
LIST_INIT(&inodedep->id_bufwait);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
TAILQ_INIT(&inodedep->id_inoupdt);
TAILQ_INIT(&inodedep->id_newinoupdt);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
TAILQ_INIT(&inodedep->id_extupdt);
TAILQ_INIT(&inodedep->id_newextupdt);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ACQUIRE_LOCK(&lk);
LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
sema_release(&inodedep_in_progress);
*inodedeppp = inodedep;
return (0);
}
/*
* Structures and routines associated with newblk caching.
*/
Back out the previous change to the queue(3) interface. It was not discussed and should probably not happen. Requested by: msmith and others
2000-05-26 02:09:24 +00:00
LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
u_long newblk_hash; /* size of hash table - 1 */
#define NEWBLK_HASH(fs, inum) \
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
static struct sema newblk_in_progress;
/*
* Look up a newblk. Return 1 if found, 0 if not found.
* If not found, allocate if DEPALLOC flag is passed.
* Found or allocated entry is returned in newblkpp.
*/
static int
newblk_lookup(fs, newblkno, flags, newblkpp)
struct fs *fs;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t newblkno;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int flags;
struct newblk **newblkpp;
{
struct newblk *newblk;
struct newblk_hashhead *newblkhd;
newblkhd = NEWBLK_HASH(fs, newblkno);
top:
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(newblk, newblkhd, nb_hash)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
break;
if (newblk) {
*newblkpp = newblk;
return (1);
}
if ((flags & DEPALLOC) == 0) {
*newblkpp = NULL;
return (0);
}
if (sema_get(&newblk_in_progress, 0) == 0)
goto top;
MALLOC(newblk, struct newblk *, sizeof(struct newblk),
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_NEWBLK, M_SOFTDEP_FLAGS);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
newblk->nb_state = 0;
newblk->nb_fs = fs;
newblk->nb_newblkno = newblkno;
LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
sema_release(&newblk_in_progress);
*newblkpp = newblk;
return (0);
}
/*
* Executed during filesystem system initialization before
More s/file system/filesystem/g
2002-05-16 21:28:32 +00:00
* mounting any filesystems.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
void
softdep_initialize()
{
LIST_INIT(&mkdirlisthd);
LIST_INIT(&softdep_workitem_pending);
The target for the maximum number of dependencies has been cut in half because of reports that under heavy load the kernel could exhaust its memory pool. The limit is now (desiredvnodes * 4) rather than (desiredvnodes * 8), so it will still scale with larger systems, just not as quickly. Sponsored by: DARPA & NAI Labs.
2002-11-20 05:16:11 +00:00
max_softdeps = desiredvnodes * 4;
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
&pagedep_hash);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
Initialise the bioops vector hack at runtime rather than at link time. This avoids the use of common variables. Reviewed by: mckusick
2002-01-08 19:32:18 +00:00
Use indirect function pointer hooks instead of #ifdef SOFTUPDATES direct calls for the two places where the kernel calls into soft updates code. Set up the hooks in softdep_initialize() and NULL them out in softdep_uninitialize(). This change allows soft updates to function correctly when ufs is loaded as a module. Reviewed by: mckusick
2002-07-01 17:59:40 +00:00
/* hooks through which the main kernel code calls us */
softdep_process_worklist_hook = softdep_process_worklist;
softdep_fsync_hook = softdep_fsync;
Initialise the bioops vector hack at runtime rather than at link time. This avoids the use of common variables. Reviewed by: mckusick
2002-01-08 19:32:18 +00:00
/* initialise bioops hack */
When I was a kid my work table was one cluttered mess an cleaning it up were a rather overwhelming task. I soon learned that if you don't know where you're going to store something, at least try to pile it next to something slightly related in the hope that a pattern emerges. Apply the same principle to the ffs/snapshot/softupdates code which have leaked into specfs: Add yet a buf-quasi-method and call it from the only two places I can see it can make a difference and implement the magic in ffs_softdep.c where it belongs. It's not pretty, but at least it's one less layer violated.
2004-03-11 18:50:33 +00:00
bioops.io_prewrite = softdep_disk_prewrite;
Initialise the bioops vector hack at runtime rather than at link time. This avoids the use of common variables. Reviewed by: mckusick
2002-01-08 19:32:18 +00:00
bioops.io_start = softdep_disk_io_initiation;
bioops.io_complete = softdep_disk_write_complete;
bioops.io_deallocate = softdep_deallocate_dependencies;
bioops.io_movedeps = softdep_move_dependencies;
bioops.io_countdeps = softdep_count_dependencies;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Add the ffs bits necessary to support unloading of the ufs kernel module. This adds an ffs_uninit() function that calls ufs_uninit() and also calls a new softdep_uninitialize() function. Add a stub for softdep_uninitialize() to cover the non-SOFTUPDATES case. Reviewed by: mckusick
2002-07-01 11:00:47 +00:00
/*
* Executed after all filesystems have been unmounted during
* filesystem module unload.
*/
void
softdep_uninitialize()
{
Use indirect function pointer hooks instead of #ifdef SOFTUPDATES direct calls for the two places where the kernel calls into soft updates code. Set up the hooks in softdep_initialize() and NULL them out in softdep_uninitialize(). This change allows soft updates to function correctly when ufs is loaded as a module. Reviewed by: mckusick
2002-07-01 17:59:40 +00:00
softdep_process_worklist_hook = NULL;
softdep_fsync_hook = NULL;
Add the ffs bits necessary to support unloading of the ufs kernel module. This adds an ffs_uninit() function that calls ufs_uninit() and also calls a new softdep_uninitialize() function. Add a stub for softdep_uninitialize() to cover the non-SOFTUPDATES case. Reviewed by: mckusick
2002-07-01 11:00:47 +00:00
hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Called at mount time to notify the dependency code that a
* filesystem wishes to use it.
*/
int
softdep_mount(devvp, mp, fs, cred)
struct vnode *devvp;
struct mount *mp;
struct fs *fs;
struct ucred *cred;
{
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
struct csum_total cstotal;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct cg *cgp;
struct buf *bp;
int error, cyl;
Bring up-to-date with Whistle's current version Includes some debugging code.
1998-05-19 23:07:25 +00:00
mp->mnt_flag &= ~MNT_ASYNC;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
mp->mnt_flag |= MNT_SOFTDEP;
/*
* When doing soft updates, the counters in the
* superblock may have gotten out of sync, so we have
* to scan the cylinder groups and recalculate them.
*/
if (fs->fs_clean != 0)
return (0);
bzero(&cstotal, sizeof cstotal);
for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
fs->fs_cgsize, cred, &bp)) != 0) {
brelse(bp);
return (error);
}
cgp = (struct cg *)bp->b_data;
cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
fs->fs_cs(fs, cyl) = cgp->cg_cs;
brelse(bp);
}
#ifdef DEBUG
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif
bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
return (0);
}
/*
* Protecting the freemaps (or bitmaps).
*
More s/file system/filesystem/g
2002-05-16 21:28:32 +00:00
* To eliminate the need to execute fsck before mounting a filesystem
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* after a power failure, one must (conservatively) guarantee that the
* on-disk copy of the bitmaps never indicate that a live inode or block is
* free. So, when a block or inode is allocated, the bitmap should be
* updated (on disk) before any new pointers. When a block or inode is
* freed, the bitmap should not be updated until all pointers have been
* reset. The latter dependency is handled by the delayed de-allocation
* approach described below for block and inode de-allocation. The former
* dependency is handled by calling the following procedure when a block or
* inode is allocated. When an inode is allocated an "inodedep" is created
* with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
* Each "inodedep" is also inserted into the hash indexing structure so
* that any additional link additions can be made dependent on the inode
* allocation.
*
More s/file system/filesystem/g
2002-05-16 21:28:32 +00:00
* The ufs filesystem maintains a number of free block counts (e.g., per
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* cylinder group, per cylinder and per <cylinder, rotational position> pair)
* in addition to the bitmaps. These counts are used to improve efficiency
* during allocation and therefore must be consistent with the bitmaps.
* There is no convenient way to guarantee post-crash consistency of these
* counts with simple update ordering, for two main reasons: (1) The counts
* and bitmaps for a single cylinder group block are not in the same disk
* sector. If a disk write is interrupted (e.g., by power failure), one may
* be written and the other not. (2) Some of the counts are located in the
* superblock rather than the cylinder group block. So, we focus our soft
* updates implementation on protecting the bitmaps. When mounting a
* filesystem, we recompute the auxiliary counts from the bitmaps.
*/
/*
* Called just after updating the cylinder group block to allocate an inode.
*/
void
softdep_setup_inomapdep(bp, ip, newinum)
struct buf *bp; /* buffer for cylgroup block with inode map */
struct inode *ip; /* inode related to allocation */
ino_t newinum; /* new inode number being allocated */
{
struct inodedep *inodedep;
struct bmsafemap *bmsafemap;
/*
* Create a dependency for the newly allocated inode.
* Panic if it already exists as something is seriously wrong.
* Otherwise add it to the dependency list for the buffer holding
* the cylinder group map from which it was allocated.
*/
ACQUIRE_LOCK(&lk);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_setup_inomapdep: found inode");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_buf = bp;
inodedep->id_state &= ~DEPCOMPLETE;
bmsafemap = bmsafemap_lookup(bp);
LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
FREE_LOCK(&lk);
}
/*
* Called just after updating the cylinder group block to
* allocate block or fragment.
*/
void
softdep_setup_blkmapdep(bp, fs, newblkno)
struct buf *bp; /* buffer for cylgroup block with block map */
struct fs *fs; /* filesystem doing allocation */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t newblkno; /* number of newly allocated block */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct newblk *newblk;
struct bmsafemap *bmsafemap;
/*
* Create a dependency for the newly allocated block.
* Add it to the dependency list for the buffer holding
* the cylinder group map from which it was allocated.
*/
if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
panic("softdep_setup_blkmapdep: found block");
ACQUIRE_LOCK(&lk);
newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
FREE_LOCK(&lk);
}
/*
* Find the bmsafemap associated with a cylinder group buffer.
* If none exists, create one. The buffer must be locked when
* this routine is called and this routine must be called with
* splbio interrupts blocked.
*/
static struct bmsafemap *
bmsafemap_lookup(bp)
struct buf *bp;
{
struct bmsafemap *bmsafemap;
struct worklist *wk;
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("bmsafemap_lookup: lock not held");
#endif
Mechanical change to use <sys/queue.h> macro API instead of fondling implementation details. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 13:13:25 +00:00
LIST_FOREACH(wk, &bp->b_dep, wk_list)
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
if (wk->wk_type == D_BMSAFEMAP)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (WK_BMSAFEMAP(wk));
FREE_LOCK(&lk);
MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_BMSAFEMAP, M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bmsafemap->sm_list.wk_state = 0;
bmsafemap->sm_buf = bp;
LIST_INIT(&bmsafemap->sm_allocdirecthd);
LIST_INIT(&bmsafemap->sm_allocindirhd);
LIST_INIT(&bmsafemap->sm_inodedephd);
LIST_INIT(&bmsafemap->sm_newblkhd);
ACQUIRE_LOCK(&lk);
WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
return (bmsafemap);
}
/*
* Direct block allocation dependencies.
*
* When a new block is allocated, the corresponding disk locations must be
* initialized (with zeros or new data) before the on-disk inode points to
* them. Also, the freemap from which the block was allocated must be
* updated (on disk) before the inode's pointer. These two dependencies are
* independent of each other and are needed for all file blocks and indirect
* blocks that are pointed to directly by the inode. Just before the
* "in-core" version of the inode is updated with a newly allocated block
* number, a procedure (below) is called to setup allocation dependency
* structures. These structures are removed when the corresponding
* dependencies are satisfied or when the block allocation becomes obsolete
* (i.e., the file is deleted, the block is de-allocated, or the block is a
* fragment that gets upgraded). All of these cases are handled in
* procedures described later.
*
* When a file extension causes a fragment to be upgraded, either to a larger
* fragment or to a full block, the on-disk location may change (if the
* previous fragment could not simply be extended). In this case, the old
* fragment must be de-allocated, but not until after the inode's pointer has
* been updated. In most cases, this is handled by later procedures, which
* will construct a "freefrag" structure to be added to the workitem queue
* when the inode update is complete (or obsolete). The main exception to
* this is when an allocation occurs while a pending allocation dependency
* (for the same block pointer) remains. This case is handled in the main
* allocation dependency setup procedure by immediately freeing the
* unreferenced fragments.
*/
void
softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
struct inode *ip; /* inode to which block is being added */
ufs_lbn_t lbn; /* block pointer within inode */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t newblkno; /* disk block number being added */
ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
long newsize; /* size of new block */
long oldsize; /* size of new block */
struct buf *bp; /* bp for allocated block */
{
struct allocdirect *adp, *oldadp;
struct allocdirectlst *adphead;
struct bmsafemap *bmsafemap;
struct inodedep *inodedep;
struct pagedep *pagedep;
struct newblk *newblk;
MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
adp->ad_list.wk_type = D_ALLOCDIRECT;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
adp->ad_lbn = lbn;
adp->ad_newblkno = newblkno;
adp->ad_oldblkno = oldblkno;
adp->ad_newsize = newsize;
adp->ad_oldsize = oldsize;
adp->ad_state = ATTACHED;
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
LIST_INIT(&adp->ad_newdirblk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (newblkno == oldblkno)
adp->ad_freefrag = NULL;
else
adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
panic("softdep_setup_allocdirect: lost block");
ACQUIRE_LOCK(&lk);
This patch corrects two problems with the rate limiting code that was introduced in revision 1.80. The problem manifested itself with a `locking against myself' panic and could also result in soft updates inconsistences associated with inodedeps. The two problems are: 1) One of the background operations could manipulate the bitmap while holding it locked with intent to create. This held lock results in a `locking against myself' panic, when the background processing that we have been coopted to do tries to lock the bitmap which we are already holding locked. To understand how to fix this problem, first, observe that we can do the background cleanups in inodedep_lookup only when allocating inodedeps (DEPALLOC is set in the call to inodedep_lookup). Second observe that calls to inodedep_lookup with DEPALLOC set can only happen from the following calls into the softdep code: softdep_setup_inomapdep softdep_setup_allocdirect softdep_setup_remove softdep_setup_freeblocks softdep_setup_directory_change softdep_setup_directory_add softdep_change_linkcnt Only the first two of these can come from ffs_alloc.c while holding a bitmap locked. Thus, inodedep_lookup must not go off to do request_cleanups when being called from these functions. This change adds a flag, NODELAY, that can be passed to inodedep_lookup to let it know that it should not do background processing in those cases. 2) The return value from request_cleanup when helping out with the cleanup was 0 instead of 1. This meant that despite the fact that we may have slept while doing the cleanups, the code did not recheck for the appearance of an inodedep (e.g., goto top in inodedep_lookup). This lead to the softdep inconsistency in which we ended up with two inodedep's for the same inode. Reviewed by: Peter Wemm <peter@yahoo-inc.com>, Matt Dillon <dillon@earth.backplane.com>
2001-02-20 11:14:38 +00:00
inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
adp->ad_inodedep = inodedep;
if (newblk->nb_state == DEPCOMPLETE) {
adp->ad_state |= DEPCOMPLETE;
adp->ad_buf = NULL;
} else {
bmsafemap = newblk->nb_bmsafemap;
adp->ad_buf = bmsafemap->sm_buf;
LIST_REMOVE(newblk, nb_deps);
LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
}
LIST_REMOVE(newblk, nb_hash);
FREE(newblk, M_NEWBLK);
WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
if (lbn >= NDADDR) {
/* allocating an indirect block */
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (oldblkno != 0) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_setup_allocdirect: non-zero indir");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
} else {
/*
* Allocating a direct block.
*
* If we are allocating a directory block, then we must
* allocate an associated pagedep to track additions and
* deletions.
*/
if ((ip->i_mode & IFMT) == IFDIR &&
pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
}
/*
* The list of allocdirects must be kept in sorted and ascending
* order so that the rollback routines can quickly determine the
* first uncommitted block (the size of the file stored on disk
* ends at the end of the lowest committed fragment, or if there
* are no fragments, at the end of the highest committed block).
* Since files generally grow, the typical case is that the new
* block is to be added at the end of the list. We speed this
* special case by checking against the last allocdirect in the
* list before laboriously traversing the list looking for the
* insertion point.
*/
adphead = &inodedep->id_newinoupdt;
oldadp = TAILQ_LAST(adphead, allocdirectlst);
if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
/* insert at end of list */
TAILQ_INSERT_TAIL(adphead, adp, ad_next);
if (oldadp != NULL && oldadp->ad_lbn == lbn)
allocdirect_merge(adphead, adp, oldadp);
FREE_LOCK(&lk);
return;
}
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
TAILQ_FOREACH(oldadp, adphead, ad_next) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (oldadp->ad_lbn >= lbn)
break;
}
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (oldadp == NULL) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_setup_allocdirect: lost entry");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/* insert in middle of list */
TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
if (oldadp->ad_lbn == lbn)
allocdirect_merge(adphead, adp, oldadp);
FREE_LOCK(&lk);
}
/*
* Replace an old allocdirect dependency with a newer one.
* This routine must be called with splbio interrupts blocked.
*/
static void
allocdirect_merge(adphead, newadp, oldadp)
struct allocdirectlst *adphead; /* head of list holding allocdirects */
struct allocdirect *newadp; /* allocdirect being added */
struct allocdirect *oldadp; /* existing allocdirect being checked */
{
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
struct worklist *wk;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct freefrag *freefrag;
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
struct newdirblk *newdirblk;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("allocdirect_merge: lock not held");
#endif
if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
newadp->ad_oldsize != oldadp->ad_newsize ||
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
newadp->ad_lbn >= NDADDR) {
FREE_LOCK(&lk);
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s %jd != new %jd || old size %ld != new %ld",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"allocdirect_merge: old blkno",
(intmax_t)newadp->ad_oldblkno,
(intmax_t)oldadp->ad_newblkno,
newadp->ad_oldsize, oldadp->ad_newsize);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
newadp->ad_oldblkno = oldadp->ad_oldblkno;
newadp->ad_oldsize = oldadp->ad_oldsize;
/*
* If the old dependency had a fragment to free or had never
* previously had a block allocated, then the new dependency
* can immediately post its freefrag and adopt the old freefrag.
* This action is done by swapping the freefrag dependencies.
* The new dependency gains the old one's freefrag, and the
* old one gets the new one and then immediately puts it on
* the worklist when it is freed by free_allocdirect. It is
* not possible to do this swap when the old dependency had a
* non-zero size but no previous fragment to free. This condition
* arises when the new block is an extension of the old block.
* Here, the first part of the fragment allocated to the new
* dependency is part of the block currently claimed on disk by
* the old dependency, so cannot legitimately be freed until the
* conditions for the new dependency are fulfilled.
*/
if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
freefrag = newadp->ad_freefrag;
newadp->ad_freefrag = oldadp->ad_freefrag;
oldadp->ad_freefrag = freefrag;
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
/*
* If we are tracking a new directory-block allocation,
* move it from the old allocdirect to the new allocdirect.
*/
if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
newdirblk = WK_NEWDIRBLK(wk);
WORKLIST_REMOVE(&newdirblk->db_list);
if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL)
panic("allocdirect_merge: extra newdirblk");
WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
free_allocdirect(adphead, oldadp, 0);
}
/*
* Allocate a new freefrag structure if needed.
*/
static struct freefrag *
newfreefrag(ip, blkno, size)
struct inode *ip;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t blkno;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
long size;
{
struct freefrag *freefrag;
struct fs *fs;
if (blkno == 0)
return (NULL);
fs = ip->i_fs;
if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
panic("newfreefrag: frag size");
MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_FREEFRAG, M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
freefrag->ff_list.wk_type = D_FREEFRAG;
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
freefrag->ff_state = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
freefrag->ff_inum = ip->i_number;
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
freefrag->ff_mnt = ITOV(ip)->v_mount;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
freefrag->ff_blkno = blkno;
freefrag->ff_fragsize = size;
return (freefrag);
}
/*
* This workitem de-allocates fragments that were replaced during
* file block allocation.
*/
static void
handle_workitem_freefrag(freefrag)
struct freefrag *freefrag;
{
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt);
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ffs_blkfree(ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
freefrag->ff_fragsize, freefrag->ff_inum);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE(freefrag, M_FREEFRAG);
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
/*
* Set up a dependency structure for an external attributes data block.
* This routine follows much of the structure of softdep_setup_allocdirect.
* See the description of softdep_setup_allocdirect above for details.
*/
void
softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
struct inode *ip;
ufs_lbn_t lbn;
ufs2_daddr_t newblkno;
ufs2_daddr_t oldblkno;
long newsize;
long oldsize;
struct buf *bp;
{
struct allocdirect *adp, *oldadp;
struct allocdirectlst *adphead;
struct bmsafemap *bmsafemap;
struct inodedep *inodedep;
struct newblk *newblk;
MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
adp->ad_list.wk_type = D_ALLOCDIRECT;
adp->ad_lbn = lbn;
adp->ad_newblkno = newblkno;
adp->ad_oldblkno = oldblkno;
adp->ad_newsize = newsize;
adp->ad_oldsize = oldsize;
adp->ad_state = ATTACHED | EXTDATA;
LIST_INIT(&adp->ad_newdirblk);
if (newblkno == oldblkno)
adp->ad_freefrag = NULL;
else
adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
panic("softdep_setup_allocext: lost block");
ACQUIRE_LOCK(&lk);
inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
adp->ad_inodedep = inodedep;
if (newblk->nb_state == DEPCOMPLETE) {
adp->ad_state |= DEPCOMPLETE;
adp->ad_buf = NULL;
} else {
bmsafemap = newblk->nb_bmsafemap;
adp->ad_buf = bmsafemap->sm_buf;
LIST_REMOVE(newblk, nb_deps);
LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
}
LIST_REMOVE(newblk, nb_hash);
FREE(newblk, M_NEWBLK);
WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
if (lbn >= NXADDR) {
FREE_LOCK(&lk);
Fix a warning: ffs_softdep.c:1630: warning: int format, different type arg (arg 2)
2002-07-20 01:09:35 +00:00
panic("softdep_setup_allocext: lbn %lld > NXADDR",
(long long)lbn);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
}
/*
* The list of allocdirects must be kept in sorted and ascending
* order so that the rollback routines can quickly determine the
* first uncommitted block (the size of the file stored on disk
* ends at the end of the lowest committed fragment, or if there
* are no fragments, at the end of the highest committed block).
* Since files generally grow, the typical case is that the new
* block is to be added at the end of the list. We speed this
* special case by checking against the last allocdirect in the
* list before laboriously traversing the list looking for the
* insertion point.
*/
adphead = &inodedep->id_newextupdt;
oldadp = TAILQ_LAST(adphead, allocdirectlst);
if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
/* insert at end of list */
TAILQ_INSERT_TAIL(adphead, adp, ad_next);
if (oldadp != NULL && oldadp->ad_lbn == lbn)
allocdirect_merge(adphead, adp, oldadp);
FREE_LOCK(&lk);
return;
}
TAILQ_FOREACH(oldadp, adphead, ad_next) {
if (oldadp->ad_lbn >= lbn)
break;
}
if (oldadp == NULL) {
FREE_LOCK(&lk);
panic("softdep_setup_allocext: lost entry");
}
/* insert in middle of list */
TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
if (oldadp->ad_lbn == lbn)
allocdirect_merge(adphead, adp, oldadp);
FREE_LOCK(&lk);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Indirect block allocation dependencies.
*
* The same dependencies that exist for a direct block also exist when
* a new block is allocated and pointed to by an entry in a block of
* indirect pointers. The undo/redo states described above are also
* used here. Because an indirect block contains many pointers that
* may have dependencies, a second copy of the entire in-memory indirect
* block is kept. The buffer cache copy is always completely up-to-date.
* The second copy, which is used only as a source for disk writes,
* contains only the safe pointers (i.e., those that have no remaining
* update dependencies). The second copy is freed when all pointers
* are safe. The cache is not allowed to replace indirect blocks with
* pending update dependencies. If a buffer containing an indirect
* block with dependencies is written, these routines will mark it
* dirty again. It can only be successfully written once all the
* dependencies are removed. The ffs_fsync routine in conjunction with
* softdep_sync_metadata work together to get all the dependencies
* removed so that a file can be successfully written to disk. Three
* procedures are used when setting up indirect block pointer
* dependencies. The division is necessary because of the organization
* of the "balloc" routine and because of the distinction between file
* pages and file metadata blocks.
*/
/*
* Allocate a new allocindir structure.
*/
static struct allocindir *
newallocindir(ip, ptrno, newblkno, oldblkno)
struct inode *ip; /* inode for file being extended */
int ptrno; /* offset of pointer in indirect block */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t newblkno; /* disk block number being added */
ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct allocindir *aip;
MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
aip->ai_list.wk_type = D_ALLOCINDIR;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
aip->ai_state = ATTACHED;
aip->ai_offset = ptrno;
aip->ai_newblkno = newblkno;
aip->ai_oldblkno = oldblkno;
aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
return (aip);
}
/*
* Called just before setting an indirect block pointer
* to a newly allocated file page.
*/
void
softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
struct inode *ip; /* inode for file being extended */
ufs_lbn_t lbn; /* allocated block number within file */
struct buf *bp; /* buffer with indirect blk referencing page */
int ptrno; /* offset of pointer in indirect block */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t newblkno; /* disk block number being added */
ufs2_daddr_t oldblkno; /* previous block number, 0 if none */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct buf *nbp; /* buffer holding allocated page */
{
struct allocindir *aip;
struct pagedep *pagedep;
aip = newallocindir(ip, ptrno, newblkno, oldblkno);
ACQUIRE_LOCK(&lk);
/*
* If we are allocating a directory page, then we must
* allocate an associated pagedep to track additions and
* deletions.
*/
if ((ip->i_mode & IFMT) == IFDIR &&
pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
FREE_LOCK(&lk);
setup_allocindir_phase2(bp, ip, aip);
}
/*
* Called just before setting an indirect block pointer to a
* newly allocated indirect block.
*/
void
softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
struct buf *nbp; /* newly allocated indirect block */
struct inode *ip; /* inode for file being extended */
struct buf *bp; /* indirect block referencing allocated block */
int ptrno; /* offset of pointer in indirect block */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t newblkno; /* disk block number being added */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct allocindir *aip;
aip = newallocindir(ip, ptrno, newblkno, 0);
ACQUIRE_LOCK(&lk);
WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
FREE_LOCK(&lk);
setup_allocindir_phase2(bp, ip, aip);
}
/*
* Called to finish the allocation of the "aip" allocated
* by one of the two routines above.
*/
static void
setup_allocindir_phase2(bp, ip, aip)
struct buf *bp; /* in-memory copy of the indirect block */
struct inode *ip; /* inode for file being extended */
struct allocindir *aip; /* allocindir allocated by the above routines */
{
struct worklist *wk;
struct indirdep *indirdep, *newindirdep;
struct bmsafemap *bmsafemap;
struct allocindir *oldaip;
struct freefrag *freefrag;
struct newblk *newblk;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t blkno;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (bp->b_lblkno >= 0)
panic("setup_allocindir_phase2: not indir blk");
for (indirdep = NULL, newindirdep = NULL; ; ) {
ACQUIRE_LOCK(&lk);
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(wk, &bp->b_dep, wk_list) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
if (wk->wk_type != D_INDIRDEP)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
indirdep = WK_INDIRDEP(wk);
break;
}
if (indirdep == NULL && newindirdep) {
indirdep = newindirdep;
WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
newindirdep = NULL;
}
FREE_LOCK(&lk);
if (indirdep) {
if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
&newblk) == 0)
panic("setup_allocindir: lost block");
ACQUIRE_LOCK(&lk);
if (newblk->nb_state == DEPCOMPLETE) {
aip->ai_state |= DEPCOMPLETE;
aip->ai_buf = NULL;
} else {
bmsafemap = newblk->nb_bmsafemap;
aip->ai_buf = bmsafemap->sm_buf;
LIST_REMOVE(newblk, nb_deps);
LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
aip, ai_deps);
}
LIST_REMOVE(newblk, nb_hash);
FREE(newblk, M_NEWBLK);
aip->ai_indirdep = indirdep;
/*
* Check to see if there is an existing dependency
* for this block. If there is, merge the old
* dependency into the new one.
*/
if (aip->ai_oldblkno == 0)
oldaip = NULL;
else
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (oldaip->ai_offset == aip->ai_offset)
break;
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
freefrag = NULL;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (oldaip != NULL) {
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (oldaip->ai_newblkno != aip->ai_oldblkno) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("setup_allocindir_phase2: blkno");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
aip->ai_oldblkno = oldaip->ai_oldblkno;
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
freefrag = aip->ai_freefrag;
aip->ai_freefrag = oldaip->ai_freefrag;
oldaip->ai_freefrag = NULL;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
free_allocindir(oldaip, NULL);
}
LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (ip->i_ump->um_fstype == UFS1)
((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
[aip->ai_offset] = aip->ai_oldblkno;
else
((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
[aip->ai_offset] = aip->ai_oldblkno;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
if (freefrag != NULL)
handle_workitem_freefrag(freefrag);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
if (newindirdep) {
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
brelse(newindirdep->ir_savebp);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
if (indirdep)
break;
MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_INDIRDEP, M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
newindirdep->ir_list.wk_type = D_INDIRDEP;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
newindirdep->ir_state = ATTACHED;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (ip->i_ump->um_fstype == UFS1)
newindirdep->ir_state |= UFS1FMT;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_INIT(&newindirdep->ir_deplisthd);
LIST_INIT(&newindirdep->ir_donehd);
Introduce the new 64-bit size disk block, daddr64_t. Change the bio and buffer structures to have daddr64_t bio_pblkno, b_blkno, and b_lblkno fields which allows access to disks larger than a Terabyte in size. This change also requires that the VOP_BMAP vnode operation accept and return daddr64_t blocks. This delta should not affect system operation in any way. It merely sets up the necessary interfaces to allow the development of disk drivers that work with these larger disk block addresses. It also allows for the development of UFS2 which will use 64-bit block addresses.
2002-03-15 18:49:47 +00:00
if (bp->b_blkno == bp->b_lblkno) {
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
NULL, NULL);
Introduce the new 64-bit size disk block, daddr64_t. Change the bio and buffer structures to have daddr64_t bio_pblkno, b_blkno, and b_lblkno fields which allows access to disks larger than a Terabyte in size. This change also requires that the VOP_BMAP vnode operation accept and return daddr64_t blocks. This delta should not affect system operation in any way. It merely sets up the necessary interfaces to allow the development of disk drivers that work with these larger disk block addresses. It also allows for the development of UFS2 which will use 64-bit block addresses.
2002-03-15 18:49:47 +00:00
bp->b_blkno = blkno;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
newindirdep->ir_savebp =
- Add a new 'flags' parameter to getblk(). - Define one flag GB_LOCK_NOWAIT that tells getblk() to pass the LK_NOWAIT flag to the initial BUF_LOCK(). This will eventually be used in cases were we want to use a buffer only if it is not currently in use. - Convert all consumers of the getblk() api to use this extra parameter. Reviwed by: arch Not objected to by: mckusick
2003-03-04 00:04:44 +00:00
getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
Convert buffer locking from using the B_BUSY and B_WANTED flags to using lockmgr locks. This commit should be functionally equivalent to the old semantics. That is, all buffer locking is done with LK_EXCLUSIVE requests. Changes to take advantage of LK_SHARED and LK_RECURSIVE will be done in future commits.
1999-06-26 02:47:16 +00:00
BUF_KERNPROC(newindirdep->ir_savebp);
Bring up-to-date with Whistle's current version Includes some debugging code.
1998-05-19 23:07:25 +00:00
bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
}
/*
* Block de-allocation dependencies.
*
* When blocks are de-allocated, the on-disk pointers must be nullified before
* the blocks are made available for use by other files. (The true
* requirement is that old pointers must be nullified before new on-disk
* pointers are set. We chose this slightly more stringent requirement to
* reduce complexity.) Our implementation handles this dependency by updating
* the inode (or indirect block) appropriately but delaying the actual block
* de-allocation (i.e., freemap and free space count manipulation) until
* after the updated versions reach stable storage. After the disk is
* updated, the blocks can be safely de-allocated whenever it is convenient.
* This implementation handles only the common case of reducing a file's
* length to zero. Other cases are handled by the conventional synchronous
* write approach.
*
* The ffs implementation with which we worked double-checks
* the state of the block pointers and file size as it reduces
* a file's length. Some of this code is replicated here in our
* soft updates implementation. The freeblks->fb_chkcnt field is
* used to transfer a part of this information to the procedure
* that eventually de-allocates the blocks.
*
* This routine should be called from the routine that shortens
* a file's length, before the inode's size or block pointers
* are modified. It will save the block pointer information for
* later release and zero the inode so that the calling routine
* can release it.
*/
void
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
softdep_setup_freeblocks(ip, length, flags)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inode *ip; /* The inode whose length is to be reduced */
off_t length; /* The new length for the file */
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
int flags; /* IO_EXT and/or IO_NORMAL */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct freeblks *freeblks;
struct inodedep *inodedep;
struct allocdirect *adp;
struct vnode *vp;
struct buf *bp;
struct fs *fs;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
ufs2_daddr_t extblocks, datablocks;
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
int i, delay, error;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
fs = ip->i_fs;
if (length != 0)
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
panic("softdep_setup_freeblocks: non-zero length");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
freeblks->fb_list.wk_type = D_FREEBLKS;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
freeblks->fb_uid = ip->i_uid;
freeblks->fb_previousinum = ip->i_number;
freeblks->fb_devvp = ip->i_devvp;
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
freeblks->fb_mnt = ITOV(ip)->v_mount;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
extblocks = 0;
if (fs->fs_magic == FS_UFS2_MAGIC)
extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
datablocks = DIP(ip, i_blocks) - extblocks;
if ((flags & IO_NORMAL) == 0) {
freeblks->fb_oldsize = 0;
freeblks->fb_chkcnt = 0;
} else {
freeblks->fb_oldsize = ip->i_size;
ip->i_size = 0;
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_size, 0);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
freeblks->fb_chkcnt = datablocks;
for (i = 0; i < NDADDR; i++) {
freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_db[i], 0);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
}
for (i = 0; i < NIADDR; i++) {
freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_ib[i], 0);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
}
/*
* If the file was removed, then the space being freed was
* accounted for then (see softdep_filereleased()). If the
* file is merely being truncated, then we account for it now.
*/
if ((ip->i_flag & IN_SPACECOUNTED) == 0)
fs->fs_pendingblocks += datablocks;
}
if ((flags & IO_EXT) == 0) {
freeblks->fb_oldextsize = 0;
} else {
freeblks->fb_oldextsize = ip->i_din2->di_extsize;
ip->i_din2->di_extsize = 0;
freeblks->fb_chkcnt += extblocks;
for (i = 0; i < NXADDR; i++) {
freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
ip->i_din2->di_extb[i] = 0;
}
}
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Push the zero'ed inode to to its disk buffer so that we are free
* to delete its dependencies below. Once the dependencies are gone
* the buffer can be safely released.
*/
if ((error = bread(ip->i_devvp,
fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
(int)fs->fs_bsize, NOCRED, &bp)) != 0) {
brelse(bp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error("softdep_setup_freeblocks", error);
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
}
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (ip->i_ump->um_fstype == UFS1)
*((struct ufs1_dinode *)bp->b_data +
ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
else
*((struct ufs2_dinode *)bp->b_data +
ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Find and eliminate any inode dependencies.
*/
ACQUIRE_LOCK(&lk);
(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((inodedep->id_state & IOSTARTED) != 0) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_setup_freeblocks: inode busy");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
When deleting a file, the ordering of events imposed by soft updates is to first write the deleted directory entry to disk, second write the zero'ed inode to disk, and finally to release the freed blocks and the inode back to the cylinder-group map. As this ordering requires two disk writes to occur which are normally spaced about 30 seconds apart (except when memory is under duress), it takes about a minute from the time that a file is deleted until its inode and data blocks show up in the cylinder-group map for reallocation. If a file has had only a brief lifetime (less than 30 seconds from creation to deletion), neither its inode nor its directory entry may have been written to disk. If its directory entry has not been written to disk, then we need not wait for that directory block to be written as the on-disk directory block does not reference the inode. Similarly, if the allocated inode has never been written to disk, we do not have to wait for it to be written back either as its on-disk representation is still zero'ed out. Thus, in the case of a short lived file, we can simply release the blocks and inode to the cylinder-group map immediately. As the inode and its blocks are released immediately, they are immediately available for other uses. If they are not released for a minute, then other inodes and blocks must be allocated for short lived files, cluttering up the vnode and buffer caches. The previous code was a bit too aggressive in trying to release the blocks and inode back to the cylinder-group map resulting in their being made available when in fact the inode on disk had not yet been zero'ed. This patch takes a more conservative approach to doing the release which avoids doing the release prematurely.
2000-11-14 09:00:25 +00:00
/*
* Add the freeblks structure to the list of operations that
* must await the zero'ed inode being written to disk. If we
* still have a bitmap dependency (delay == 0), then the inode
* has never been written to disk, so we can process the
* freeblks below once we have deleted the dependencies.
*/
delay = (inodedep->id_state & DEPCOMPLETE);
if (delay)
WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Because the file length has been truncated to zero, any
* pending block allocation dependency structures associated
* with this inode are obsolete and can simply be de-allocated.
* We must first merge the two dependency lists to get rid of
* any duplicate freefrag structures, then purge the merged list.
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
* If we still have a bitmap dependency, then the inode has never
* been written to disk, so we can free any fragments without delay.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (flags & IO_NORMAL) {
merge_inode_lists(&inodedep->id_newinoupdt,
&inodedep->id_inoupdt);
while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
free_allocdirect(&inodedep->id_inoupdt, adp, delay);
}
if (flags & IO_EXT) {
merge_inode_lists(&inodedep->id_newextupdt,
&inodedep->id_extupdt);
while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
free_allocdirect(&inodedep->id_extupdt, adp, delay);
}
We cannot proceed to free the blocks of the file until the dependencies have been cleaned up by deallocte_dependencies(). Once that is done, it is safe to post the request to free the blocks. A similar change is also needed for the freefile case.
2000-01-11 06:52:35 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bdwrite(bp);
/*
* We must wait for any I/O in progress to finish so that
* all potential buffers on the dirty list will be visible.
* Once they are all there, walk the list and get rid of
* any dependencies.
*/
vp = ITOV(ip);
Reorganize locking to avoid holding the lock during calls to bdwrite and brelse (which may sleep in some systems). Obtained from: Matthew Dillon <dillon@apollo.backplane.com>
1999-03-02 06:38:07 +00:00
ACQUIRE_LOCK(&lk);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
drain_output(vp, 1);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
restart:
Loose the v_dirty* and v_clean* alias macros. Check the count field where we just want to know the full/empty state, rather than using TAILQ_EMPTY() or TAILQ_FIRST().
2004-10-25 09:14:03 +00:00
TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
((flags & IO_NORMAL) == 0 &&
(bp->b_xflags & BX_ALTDATA) == 0))
continue;
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if ((bp = getdirtybuf(&bp, VI_MTX(vp), MNT_WAIT)) == NULL)
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
goto restart;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
deallocate_dependencies(bp, inodedep);
Add B_NOCACHE to several cases where BSD4.4 only required a B_INVAL. Change worked out by john and kirk in consort.
1998-06-11 17:44:32 +00:00
bp->b_flags |= B_INVAL | B_NOCACHE;
Reorganize locking to avoid holding the lock during calls to bdwrite and brelse (which may sleep in some systems). Obtained from: Matthew Dillon <dillon@apollo.backplane.com>
1999-03-02 06:38:07 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
brelse(bp);
Reorganize locking to avoid holding the lock during calls to bdwrite and brelse (which may sleep in some systems). Obtained from: Matthew Dillon <dillon@apollo.backplane.com>
1999-03-02 06:38:07 +00:00
ACQUIRE_LOCK(&lk);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
goto restart;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
- Convert locks to use standard macros. - Lock access to the buflists. - Document broken locking. - Use vrefcnt().
2002-09-25 02:49:48 +00:00
VI_UNLOCK(vp);
When deleting a file, the ordering of events imposed by soft updates is to first write the deleted directory entry to disk, second write the zero'ed inode to disk, and finally to release the freed blocks and the inode back to the cylinder-group map. As this ordering requires two disk writes to occur which are normally spaced about 30 seconds apart (except when memory is under duress), it takes about a minute from the time that a file is deleted until its inode and data blocks show up in the cylinder-group map for reallocation. If a file has had only a brief lifetime (less than 30 seconds from creation to deletion), neither its inode nor its directory entry may have been written to disk. If its directory entry has not been written to disk, then we need not wait for that directory block to be written as the on-disk directory block does not reference the inode. Similarly, if the allocated inode has never been written to disk, we do not have to wait for it to be written back either as its on-disk representation is still zero'ed out. Thus, in the case of a short lived file, we can simply release the blocks and inode to the cylinder-group map immediately. As the inode and its blocks are released immediately, they are immediately available for other uses. If they are not released for a minute, then other inodes and blocks must be allocated for short lived files, cluttering up the vnode and buffer caches. The previous code was a bit too aggressive in trying to release the blocks and inode back to the cylinder-group map resulting in their being made available when in fact the inode on disk had not yet been zero'ed. This patch takes a more conservative approach to doing the release which avoids doing the release prematurely.
2000-11-14 09:00:25 +00:00
if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
(void) free_inodedep(inodedep);
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
When deleting a file, the ordering of events imposed by soft updates is to first write the deleted directory entry to disk, second write the zero'ed inode to disk, and finally to release the freed blocks and the inode back to the cylinder-group map. As this ordering requires two disk writes to occur which are normally spaced about 30 seconds apart (except when memory is under duress), it takes about a minute from the time that a file is deleted until its inode and data blocks show up in the cylinder-group map for reallocation. If a file has had only a brief lifetime (less than 30 seconds from creation to deletion), neither its inode nor its directory entry may have been written to disk. If its directory entry has not been written to disk, then we need not wait for that directory block to be written as the on-disk directory block does not reference the inode. Similarly, if the allocated inode has never been written to disk, we do not have to wait for it to be written back either as its on-disk representation is still zero'ed out. Thus, in the case of a short lived file, we can simply release the blocks and inode to the cylinder-group map immediately. As the inode and its blocks are released immediately, they are immediately available for other uses. If they are not released for a minute, then other inodes and blocks must be allocated for short lived files, cluttering up the vnode and buffer caches. The previous code was a bit too aggressive in trying to release the blocks and inode back to the cylinder-group map resulting in their being made available when in fact the inode on disk had not yet been zero'ed. This patch takes a more conservative approach to doing the release which avoids doing the release prematurely.
2000-11-14 09:00:25 +00:00
* If the inode has never been written to disk (delay == 0),
* then we can process the freeblks now that we have deleted
* the dependencies.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
When deleting a file, the ordering of events imposed by soft updates is to first write the deleted directory entry to disk, second write the zero'ed inode to disk, and finally to release the freed blocks and the inode back to the cylinder-group map. As this ordering requires two disk writes to occur which are normally spaced about 30 seconds apart (except when memory is under duress), it takes about a minute from the time that a file is deleted until its inode and data blocks show up in the cylinder-group map for reallocation. If a file has had only a brief lifetime (less than 30 seconds from creation to deletion), neither its inode nor its directory entry may have been written to disk. If its directory entry has not been written to disk, then we need not wait for that directory block to be written as the on-disk directory block does not reference the inode. Similarly, if the allocated inode has never been written to disk, we do not have to wait for it to be written back either as its on-disk representation is still zero'ed out. Thus, in the case of a short lived file, we can simply release the blocks and inode to the cylinder-group map immediately. As the inode and its blocks are released immediately, they are immediately available for other uses. If they are not released for a minute, then other inodes and blocks must be allocated for short lived files, cluttering up the vnode and buffer caches. The previous code was a bit too aggressive in trying to release the blocks and inode back to the cylinder-group map resulting in their being made available when in fact the inode on disk had not yet been zero'ed. This patch takes a more conservative approach to doing the release which avoids doing the release prematurely.
2000-11-14 09:00:25 +00:00
if (!delay)
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
handle_workitem_freeblocks(freeblks, 0);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Reclaim any dependency structures from a buffer that is about to
* be reallocated to a new vnode. The buffer must be locked, thus,
* no I/O completion operations can occur while we are manipulating
* its associated dependencies. The mutex is held so that other I/O's
* associated with related dependencies do not occur.
*/
static void
deallocate_dependencies(bp, inodedep)
struct buf *bp;
struct inodedep *inodedep;
{
struct worklist *wk;
struct indirdep *indirdep;
struct allocindir *aip;
struct pagedep *pagedep;
struct dirrem *dirrem;
struct diradd *dap;
int i;
while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
switch (wk->wk_type) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INDIRDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
indirdep = WK_INDIRDEP(wk);
/*
* None of the indirect pointers will ever be visible,
* so they can simply be tossed. GOINGAWAY ensures
* that allocated pointers will be saved in the buffer
* cache until they are freed. Note that they will
* only be able to be found by their physical address
* since the inode mapping the logical address will
* be gone. The save buffer used for the safe copy
* was allocated in setup_allocindir_phase2 using
* the physical address so it could be used for this
* purpose. Hence we swap the safe copy with the real
* copy, allowing the safe copy to be freed and holding
* on to the real copy for later use in indir_trunc.
*/
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (indirdep->ir_state & GOINGAWAY) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("deallocate_dependencies: already gone");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
indirdep->ir_state |= GOINGAWAY;
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
free_allocindir(aip, inodedep);
if (bp->b_lblkno >= 0 ||
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("deallocate_dependencies: not indir");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Change to stop a silly panic. This should be understood better. Change a buffer swizzle trick to a bcopy. It would be nice if the efficient trick could be used in the future.
1998-05-19 20:50:41 +00:00
bcopy(bp->b_data, indirdep->ir_savebp->b_data,
bp->b_bcount);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
WORKLIST_REMOVE(wk);
WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_PAGEDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
pagedep = WK_PAGEDEP(wk);
/*
* None of the directory additions will ever be
* visible, so they can simply be tossed.
*/
for (i = 0; i < DAHASHSZ; i++)
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
while ((dap =
LIST_FIRST(&pagedep->pd_diraddhd[i])))
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
free_diradd(dap);
while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
free_diradd(dap);
/*
* Copy any directory remove dependencies to the list
* to be processed after the zero'ed inode is written.
* If the inode has already been written, then they
* can be dumped directly onto the work list.
*/
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_REMOVE(dirrem, dm_next);
dirrem->dm_dirinum = pagedep->pd_ino;
Better bounding on softdep_flushfiles; other minor tweeks to checks.
2000-01-17 06:35:11 +00:00
if (inodedep == NULL ||
(inodedep->id_state & ALLCOMPLETE) ==
ALLCOMPLETE)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
add_to_worklist(&dirrem->dm_list);
else
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
WORKLIST_INSERT(&inodedep->id_bufwait,
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
&dirrem->dm_list);
}
There seems to be a problem that the order of disk write operation being incorrect due to a missing check for some dependency. This change avoids the freelist corruption (but not the temporarily inconsistent state of the file system). A message is printed as a reminder of the under lying problem when a pagedep structure is not freed due to the NEWBLOCK flag being set. Submitted by: Tor.Egge@fast.no
2001-06-05 01:49:37 +00:00
if ((pagedep->pd_state & NEWBLOCK) != 0) {
Build on the change in revision 1.98 by Tor.Egge@fast.no. The symptom being treated in 1.98 was to avoid freeing a pagedep dependency if there was still a newdirblk dependency referencing it. That change is correct and no longer prints a warning message when it occurs. The other part of revision 1.98 was to panic when a newdirblk dependency was encountered during a file truncation. This fix removes that panic and replaces it with code to find and delete the newdirblk dependency so that the truncation can succeed.
2001-06-13 23:13:13 +00:00
LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
if (wk->wk_type == D_NEWDIRBLK &&
WK_NEWDIRBLK(wk)->db_pagedep ==
pagedep)
break;
if (wk != NULL) {
WORKLIST_REMOVE(wk);
free_newdirblk(WK_NEWDIRBLK(wk));
} else {
FREE_LOCK(&lk);
panic("deallocate_dependencies: "
"lost pagedep");
}
There seems to be a problem that the order of disk write operation being incorrect due to a missing check for some dependency. This change avoids the freelist corruption (but not the temporarily inconsistent state of the file system). A message is printed as a reminder of the under lying problem when a pagedep structure is not freed due to the NEWBLOCK flag being set. Submitted by: Tor.Egge@fast.no
2001-06-05 01:49:37 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
WORKLIST_REMOVE(&pagedep->pd_list);
LIST_REMOVE(pagedep, pd_hash);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(pagedep, D_PAGEDEP);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_ALLOCINDIR:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
free_allocindir(WK_ALLOCINDIR(wk), inodedep);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_ALLOCDIRECT:
case D_INODEDEP:
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("deallocate_dependencies: Unexpected type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
default:
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("deallocate_dependencies: Unknown type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
}
}
}
/*
* Free an allocdirect. Generate a new freefrag work request if appropriate.
* This routine must be called with splbio interrupts blocked.
*/
static void
free_allocdirect(adphead, adp, delay)
struct allocdirectlst *adphead;
struct allocdirect *adp;
int delay;
{
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
struct newdirblk *newdirblk;
struct worklist *wk;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("free_allocdirect: lock not held");
#endif
if ((adp->ad_state & DEPCOMPLETE) == 0)
LIST_REMOVE(adp, ad_deps);
TAILQ_REMOVE(adphead, adp, ad_next);
if ((adp->ad_state & COMPLETE) == 0)
WORKLIST_REMOVE(&adp->ad_list);
if (adp->ad_freefrag != NULL) {
if (delay)
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
&adp->ad_freefrag->ff_list);
else
add_to_worklist(&adp->ad_freefrag->ff_list);
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
newdirblk = WK_NEWDIRBLK(wk);
WORKLIST_REMOVE(&newdirblk->db_list);
if (LIST_FIRST(&adp->ad_newdirblk) != NULL)
panic("free_allocdirect: extra newdirblk");
if (delay)
WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
&newdirblk->db_list);
else
free_newdirblk(newdirblk);
}
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(adp, D_ALLOCDIRECT);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
/*
* Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
* This routine must be called with splbio interrupts blocked.
*/
static void
free_newdirblk(newdirblk)
struct newdirblk *newdirblk;
{
struct pagedep *pagedep;
struct diradd *dap;
int i;
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
panic("free_newdirblk: lock not held");
#endif
/*
Must ensure that all the entries on the pd_pendinghd list have been committed to disk before clearing them. More specifically, when free_newdirblk is called, we know that the inode claims the new directory block. However, if the associated pagedep is still linked onto the directory buffer dependency chain, then some of the entries on the pd_pendinghd list may not be committed to disk yet. In this case, we will simply note that the inode claims the block and let the pd_pendinghd list be processed when the pagedep is next written. If the pagedep is no longer on the buffer dependency chain, then all the entries on the pd_pending list are committed to disk and we can free them in free_newdirblk. This corrects a window of vulnerability introduced in the code added in version 1.95.
2001-05-19 19:24:26 +00:00
* If the pagedep is still linked onto the directory buffer
* dependency chain, then some of the entries on the
* pd_pendinghd list may not be committed to disk yet. In
* this case, we will simply clear the NEWBLOCK flag and
* let the pd_pendinghd list be processed when the pagedep
* is next written. If the pagedep is no longer on the buffer
* dependency chain, then all the entries on the pd_pending
* list are committed to disk and we can free them here.
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
*/
pagedep = newdirblk->db_pagedep;
pagedep->pd_state &= ~NEWBLOCK;
Must ensure that all the entries on the pd_pendinghd list have been committed to disk before clearing them. More specifically, when free_newdirblk is called, we know that the inode claims the new directory block. However, if the associated pagedep is still linked onto the directory buffer dependency chain, then some of the entries on the pd_pendinghd list may not be committed to disk yet. In this case, we will simply note that the inode claims the block and let the pd_pendinghd list be processed when the pagedep is next written. If the pagedep is no longer on the buffer dependency chain, then all the entries on the pd_pending list are committed to disk and we can free them in free_newdirblk. This corrects a window of vulnerability introduced in the code added in version 1.95.
2001-05-19 19:24:26 +00:00
if ((pagedep->pd_state & ONWORKLIST) == 0)
while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
free_diradd(dap);
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
/*
* If no dependencies remain, the pagedep will be freed.
*/
for (i = 0; i < DAHASHSZ; i++)
if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
break;
Must be a bit less aggressive about freeing pagedep structures. Obtained from: Robert Watson <rwatson@FreeBSD.org> and Matthew Jacob <mjacob@feral.com>
2001-05-18 22:16:28 +00:00
if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
LIST_REMOVE(pagedep, pd_hash);
WORKITEM_FREE(pagedep, D_PAGEDEP);
}
WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Prepare an inode to be freed. The actual free operation is not
* done until the zero'ed inode has been written to disk.
*/
void
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
softdep_freefile(pvp, ino, mode)
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
struct vnode *pvp;
ino_t ino;
int mode;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
struct inode *ip = VTOI(pvp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inodedep *inodedep;
struct freefile *freefile;
/*
* This sets up the inode de-allocation dependency.
*/
MALLOC(freefile, struct freefile *, sizeof(struct freefile),
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_FREEFILE, M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
freefile->fx_list.wk_type = D_FREEFILE;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
freefile->fx_list.wk_state = 0;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
freefile->fx_mode = mode;
freefile->fx_oldinum = ino;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
freefile->fx_devvp = ip->i_devvp;
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
freefile->fx_mnt = ITOV(ip)->v_mount;
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
if ((ip->i_flag & IN_SPACECOUNTED) == 0)
ip->i_fs->fs_pendinginodes += 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* If the inodedep does not exist, then the zero'ed inode has
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
* been written to disk. If the allocated inode has never been
* written to disk, then the on-disk inode is zero'ed. In either
* case we can free the file immediately.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
ACQUIRE_LOCK(&lk);
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
check_inode_unwritten(inodedep)) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
handle_workitem_freefile(freefile);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return;
}
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
FREE_LOCK(&lk);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
/*
* Check to see if an inode has never been written to disk. If
* so free the inodedep and return success, otherwise return failure.
* This routine must be called with splbio interrupts blocked.
*
* If we still have a bitmap dependency, then the inode has never
* been written to disk. Drop the dependency as it is no longer
* necessary since the inode is being deallocated. We set the
* ALLCOMPLETE flags since the bitmap now properly shows that the
* inode is not allocated. Even if the inode is actively being
* written, it has been rolled back to its zero'ed state, so we
* are ensured that a zero inode is what is on the disk. For short
* lived files, this change will usually result in removing all the
* dependencies from the inode so that it can be freed immediately.
*/
static int
check_inode_unwritten(inodedep)
struct inodedep *inodedep;
{
if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
LIST_FIRST(&inodedep->id_bufwait) != NULL ||
LIST_FIRST(&inodedep->id_inowait) != NULL ||
TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
inodedep->id_nlinkdelta != 0)
return (0);
inodedep->id_state |= ALLCOMPLETE;
LIST_REMOVE(inodedep, id_deps);
inodedep->id_buf = NULL;
When running with quotas enabled on a filesystem using soft updates, the system would panic when a user's inode quota was exceeded (see PR 18959 for details). This fixes that problem. PR: 18959 Submitted by: Jason Godsey <jason@unixguy.fidalgo.net>
2000-06-18 22:14:28 +00:00
if (inodedep->id_state & ONWORKLIST)
WORKLIST_REMOVE(&inodedep->id_list);
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (inodedep->id_savedino1 != NULL) {
FREE(inodedep->id_savedino1, M_INODEDEP);
inodedep->id_savedino1 = NULL;
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
}
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (free_inodedep(inodedep) == 0) {
FREE_LOCK(&lk);
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
panic("check_inode_unwritten: busy inode");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
return (1);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Try to free an inodedep structure. Return 1 if it could be freed.
*/
static int
free_inodedep(inodedep)
struct inodedep *inodedep;
{
if ((inodedep->id_state & ONWORKLIST) != 0 ||
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
(inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
LIST_FIRST(&inodedep->id_bufwait) != NULL ||
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_FIRST(&inodedep->id_inowait) != NULL ||
TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (0);
LIST_REMOVE(inodedep, id_hash);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(inodedep, D_INODEDEP);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
num_inodedep -= 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (1);
}
/*
* This workitem routine performs the block de-allocation.
* The workitem is added to the pending list after the updated
* inode block has been written to disk. As mentioned above,
* checks regarding the number of blocks de-allocated (compared
* to the number of blocks allocated for the file) are also
* performed in this function.
*/
static void
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
handle_workitem_freeblocks(freeblks, flags)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct freeblks *freeblks;
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
int flags;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
struct inode *ip;
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
struct vnode *vp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct fs *fs;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
int i, nblocks, level, bsize;
ufs2_daddr_t bn, blocksreleased = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int error, allerror = 0;
ufs_lbn_t baselbns[NIADDR], tmpval;
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
tmpval = 1;
baselbns[0] = NDADDR;
for (i = 1; i < NIADDR; i++) {
tmpval *= NINDIR(fs);
baselbns[i] = baselbns[i - 1] + tmpval;
}
nblocks = btodb(fs->fs_bsize);
blocksreleased = 0;
/*
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
* Release all extended attribute blocks or frags.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (freeblks->fb_oldextsize > 0) {
for (i = (NXADDR - 1); i >= 0; i--) {
if ((bn = freeblks->fb_eblks[i]) == 0)
continue;
bsize = sblksize(fs, freeblks->fb_oldextsize, i);
ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
freeblks->fb_previousinum);
blocksreleased += btodb(bsize);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
* Release all data blocks or frags.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (freeblks->fb_oldsize > 0) {
/*
* Indirect blocks first.
*/
for (level = (NIADDR - 1); level >= 0; level--) {
if ((bn = freeblks->fb_iblks[level]) == 0)
continue;
if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
level, baselbns[level], &blocksreleased)) == 0)
allerror = error;
ffs_blkfree(fs, freeblks->fb_devvp, bn, fs->fs_bsize,
freeblks->fb_previousinum);
fs->fs_pendingblocks -= nblocks;
blocksreleased += nblocks;
}
/*
* All direct blocks or frags.
*/
for (i = (NDADDR - 1); i >= 0; i--) {
if ((bn = freeblks->fb_dblks[i]) == 0)
continue;
bsize = sblksize(fs, freeblks->fb_oldsize, i);
ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
freeblks->fb_previousinum);
fs->fs_pendingblocks -= btodb(bsize);
blocksreleased += btodb(bsize);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
/*
* If we still have not finished background cleanup, then check
* to see if the block count needs to be adjusted.
*/
if (freeblks->fb_chkcnt != blocksreleased &&
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
(fs->fs_flags & FS_UNCLEAN) != 0 &&
VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum,
(flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
ip = VTOI(vp);
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \
freeblks->fb_chkcnt - blocksreleased);
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
ip->i_flag |= IN_CHANGE;
vput(vp);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#ifdef DIAGNOSTIC
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
if (freeblks->fb_chkcnt != blocksreleased &&
((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
When spamming me with a printf(9), under DIAGNOSTIC, at least be nice enough to include a newline. MFC after: 4 days Sponsored by: Bright Path Solutions
2002-09-28 19:04:49 +00:00
printf("handle_workitem_freeblocks: block count\n");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (allerror)
softdep_error("handle_workitem_freeblks", allerror);
#endif /* DIAGNOSTIC */
Add kernel support for running fsck on active filesystems.
2001-03-21 04:09:01 +00:00
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(freeblks, D_FREEBLKS);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Release blocks associated with the inode ip and stored in the indirect
* block dbn. If level is greater than SINGLE, the block is an indirect block
* and recursive calls to indirtrunc must be used to cleanse other indirect
* blocks.
*/
static int
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
indir_trunc(freeblks, dbn, level, lbn, countp)
struct freeblks *freeblks;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t dbn;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int level;
ufs_lbn_t lbn;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs2_daddr_t *countp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct buf *bp;
struct fs *fs;
struct worklist *wk;
struct indirdep *indirdep;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs1_daddr_t *bap1 = 0;
ufs2_daddr_t nb, *bap2 = 0;
ufs_lbn_t lbnadd;
int i, nblocks, ufs1fmt;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int error, allerror = 0;
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
lbnadd = 1;
for (i = level; i > 0; i--)
lbnadd *= NINDIR(fs);
/*
* Get buffer of block pointers to be freed. This routine is not
* called until the zero'ed inode has been written, so it is safe
* to free blocks as they are encountered. Because the inode has
* been zero'ed, calls to bmap on these blocks will fail. So, we
* have to use the on-disk address and the block device for the
* filesystem to look them up. If the file was deleted before its
* indirect blocks were all written to disk, the routine that set
* us up (deallocate_dependencies) will have arranged to leave
* a complete copy of the indirect block in memory for our use.
* Otherwise we have to read the blocks in from the disk.
*/
- My last commit to this file is still not safe, I believe that it may be due to the recursion in indir_trunc().
2003-10-06 03:28:03 +00:00
#ifdef notyet
- Reinstate 1.142 this was fixed by 1.144.
2003-10-06 02:39:37 +00:00
bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
GB_NOCREAT);
- My last commit to this file is still not safe, I believe that it may be due to the recursion in indir_trunc().
2003-10-06 03:28:03 +00:00
#else
Add b_bufobj to struct buf which eventually will eliminate the need for b_vp. Initialize b_bufobj for all buffers. Make incore() and gbincore() take a bufobj instead of a vnode. Make inmem() local to vfs_bio.c Change a lot of VI_[UN]LOCK(bp->b_vp) to BO_[UN]LOCK(bp->b_bufobj) also VI_MTX() to BO_MTX(), Make buf_vlist_add() take a bufobj instead of a vnode. Eliminate other uses of bp->b_vp where bp->b_bufobj will do. Various minor polishing: remove "register", turn panic into KASSERT, use new function declarations, TAILQ_FOREACH_SAFE() etc.
2004-10-22 08:47:20 +00:00
bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
- My last commit to this file is still not safe, I believe that it may be due to the recursion in indir_trunc().
2003-10-06 03:28:03 +00:00
#endif
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ACQUIRE_LOCK(&lk);
- Reinstate 1.142 this was fixed by 1.144.
2003-10-06 02:39:37 +00:00
if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
if (wk->wk_type != D_INDIRDEP ||
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
(indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
(indirdep->ir_state & GOINGAWAY) == 0) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("indir_trunc: lost indirdep");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
WORKLIST_REMOVE(wk);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(indirdep, D_INDIRDEP);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (LIST_FIRST(&bp->b_dep) != NULL) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("indir_trunc: dangling dep");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
VFSTOUFS(freeblks->fb_mnt)->um_numindirdeps -= 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
} else {
- My last commit to this file is still not safe, I believe that it may be due to the recursion in indir_trunc().
2003-10-06 03:28:03 +00:00
#ifdef notyet
- Reinstate 1.142 this was fixed by 1.144.
2003-10-06 02:39:37 +00:00
if (bp)
brelse(bp);
- My last commit to this file is still not safe, I believe that it may be due to the recursion in indir_trunc().
2003-10-06 03:28:03 +00:00
#endif
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
NOCRED, &bp);
if (error) {
brelse(bp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (error);
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Recursively free indirect blocks.
*/
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) {
ufs1fmt = 1;
bap1 = (ufs1_daddr_t *)bp->b_data;
} else {
ufs1fmt = 0;
bap2 = (ufs2_daddr_t *)bp->b_data;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
nblocks = btodb(fs->fs_bsize);
for (i = NINDIR(fs) - 1; i >= 0; i--) {
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (ufs1fmt)
nb = bap1[i];
else
nb = bap2[i];
if (nb == 0)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
if (level != 0) {
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
level - 1, lbn + (i * lbnadd), countp)) != 0)
allerror = error;
}
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
ffs_blkfree(fs, freeblks->fb_devvp, nb, fs->fs_bsize,
freeblks->fb_previousinum);
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
fs->fs_pendingblocks -= nblocks;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*countp += nblocks;
}
Add B_NOCACHE to several cases where BSD4.4 only required a B_INVAL. Change worked out by john and kirk in consort.
1998-06-11 17:44:32 +00:00
bp->b_flags |= B_INVAL | B_NOCACHE;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
brelse(bp);
return (allerror);
}
/*
* Free an allocindir.
* This routine must be called with splbio interrupts blocked.
*/
static void
free_allocindir(aip, inodedep)
struct allocindir *aip;
struct inodedep *inodedep;
{
struct freefrag *freefrag;
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("free_allocindir: lock not held");
#endif
if ((aip->ai_state & DEPCOMPLETE) == 0)
LIST_REMOVE(aip, ai_deps);
if (aip->ai_state & ONWORKLIST)
WORKLIST_REMOVE(&aip->ai_list);
LIST_REMOVE(aip, ai_next);
if ((freefrag = aip->ai_freefrag) != NULL) {
if (inodedep == NULL)
add_to_worklist(&freefrag->ff_list);
else
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
WORKLIST_INSERT(&inodedep->id_bufwait,
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
&freefrag->ff_list);
}
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(aip, D_ALLOCINDIR);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Directory entry addition dependencies.
*
* When adding a new directory entry, the inode (with its incremented link
* count) must be written to disk before the directory entry's pointer to it.
* Also, if the inode is newly allocated, the corresponding freemap must be
* updated (on disk) before the directory entry's pointer. These requirements
* are met via undo/redo on the directory entry's pointer, which consists
* simply of the inode number.
*
* As directory entries are added and deleted, the free space within a
More s/file system/filesystem/g
2002-05-16 21:28:32 +00:00
* directory block can become fragmented. The ufs filesystem will compact
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* a fragmented directory block to make space for a new entry. When this
* occurs, the offsets of previously added entries change. Any "diradd"
* dependency structures corresponding to these entries must be updated with
* the new offsets.
*/
/*
* This routine is called after the in-memory inode's link
* count has been incremented, but before the directory entry's
* pointer to the inode has been set.
*/
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
int
softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct buf *bp; /* buffer containing directory block */
struct inode *dp; /* inode for directory */
off_t diroffset; /* offset of new entry in directory */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ino_t newinum; /* inode referenced by new directory entry */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct buf *newdirbp; /* non-NULL => contents of new mkdir */
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
int isnewblk; /* entry is in a newly allocated block */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
int offset; /* offset of new entry within directory block */
ufs_lbn_t lbn; /* block in directory containing new entry */
struct fs *fs;
struct diradd *dap;
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
struct allocdirect *adp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct pagedep *pagedep;
struct inodedep *inodedep;
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
struct newdirblk *newdirblk = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct mkdir *mkdir1, *mkdir2;
/*
* Whiteouts have no dependencies.
*/
if (newinum == WINO) {
if (newdirbp != NULL)
bdwrite(newdirbp);
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
return (0);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
fs = dp->i_fs;
lbn = lblkno(fs, diroffset);
offset = blkoff(fs, diroffset);
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
dap->da_list.wk_type = D_DIRADD;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dap->da_offset = offset;
dap->da_newinum = newinum;
dap->da_state = ATTACHED;
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
M_NEWDIRBLK, M_SOFTDEP_FLAGS);
newdirblk->db_list.wk_type = D_NEWDIRBLK;
newdirblk->db_state = 0;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (newdirbp == NULL) {
dap->da_state |= DEPCOMPLETE;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
ACQUIRE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
} else {
dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
mkdir1->md_list.wk_type = D_MKDIR;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
mkdir1->md_state = MKDIR_BODY;
mkdir1->md_diradd = dap;
MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_SOFTDEP_FLAGS);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
mkdir2->md_list.wk_type = D_MKDIR;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
mkdir2->md_state = MKDIR_PARENT;
mkdir2->md_diradd = dap;
/*
Merge with Kirk's version as of Feb 20 His version 9.23 == our version 1.5 of ffs_softdep.c His version 9.5 == our version 1.4 of softdep.c
1998-05-19 22:54:53 +00:00
* Dependency on "." and ".." being written to disk.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
mkdir1->md_buf = newdirbp;
Reorganize locking to avoid holding the lock during calls to bdwrite and brelse (which may sleep in some systems). Obtained from: Matthew Dillon <dillon@apollo.backplane.com>
1999-03-02 06:38:07 +00:00
ACQUIRE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
Reorganize locking to avoid holding the lock during calls to bdwrite and brelse (which may sleep in some systems). Obtained from: Matthew Dillon <dillon@apollo.backplane.com>
1999-03-02 06:38:07 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bdwrite(newdirbp);
/*
* Dependency on link count increase for parent directory
*/
Reorganize locking to avoid holding the lock during calls to bdwrite and brelse (which may sleep in some systems). Obtained from: Matthew Dillon <dillon@apollo.backplane.com>
1999-03-02 06:38:07 +00:00
ACQUIRE_LOCK(&lk);
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
|| (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
dap->da_state &= ~MKDIR_PARENT;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(mkdir2, D_MKDIR);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
} else {
LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
}
/*
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
* Link into parent directory pagedep to await its being written.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
dap->da_pagedep = pagedep;
LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
da_pdlist);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
/*
* Link into its inodedep. Put it on the id_bufwait list if the inode
* is not yet written. If it is written, do the post-inode write
* processing to put it on the id_pendinghd list.
*/
(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
diradd_inode_written(dap, inodedep);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
else
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if (isnewblk) {
/*
* Directories growing into indirect blocks are rare
* enough and the frequency of new block allocation
* in those cases even more rare, that we choose not
* to bother tracking them. Rather we simply force the
* new directory entry to disk.
*/
if (lbn >= NDADDR) {
FREE_LOCK(&lk);
/*
* We only have a new allocation when at the
* beginning of a new block, not when we are
* expanding into an existing block.
*/
if (blkoff(fs, diroffset) == 0)
return (1);
return (0);
}
/*
* We only have a new allocation when at the beginning
* of a new fragment, not when we are expanding into an
* existing fragment. Also, there is nothing to do if we
* are already tracking this block.
*/
if (fragoff(fs, diroffset) != 0) {
FREE_LOCK(&lk);
return (0);
}
if ((pagedep->pd_state & NEWBLOCK) != 0) {
WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
FREE_LOCK(&lk);
return (0);
}
/*
* Find our associated allocdirect and have it track us.
*/
if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0)
panic("softdep_setup_directory_add: lost inodedep");
adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
if (adp == NULL || adp->ad_lbn != lbn) {
FREE_LOCK(&lk);
panic("softdep_setup_directory_add: lost entry");
}
pagedep->pd_state |= NEWBLOCK;
newdirblk->db_pagedep = pagedep;
WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
return (0);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* This procedure is called to change the offset of a directory
* entry when compacting a directory block which must be owned
* exclusively by the caller. Note that the actual entry movement
* must be done in this procedure to ensure that no I/O completions
* occur while the move is in progress.
*/
void
softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
struct inode *dp; /* inode for directory */
caddr_t base; /* address of dp->i_offset */
caddr_t oldloc; /* address of old directory location */
caddr_t newloc; /* address of new directory location */
int entrysize; /* size of directory entry */
{
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
int offset, oldoffset, newoffset;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct pagedep *pagedep;
struct diradd *dap;
ufs_lbn_t lbn;
ACQUIRE_LOCK(&lk);
lbn = lblkno(dp->i_fs, dp->i_offset);
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
offset = blkoff(dp->i_fs, dp->i_offset);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
goto done;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
oldoffset = offset + (oldloc - base);
newoffset = offset + (newloc - base);
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (dap->da_offset != oldoffset)
continue;
dap->da_offset = newoffset;
if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
break;
LIST_REMOVE(dap, da_pdlist);
LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
dap, da_pdlist);
break;
}
fix double LIST_REMOVE; other cosmetic changes to match version 9.32. Obtained from: Jeffrey Hsu <hsu@FreeBSD.ORG>
1999-02-17 20:01:20 +00:00
if (dap == NULL) {
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
fix double LIST_REMOVE; other cosmetic changes to match version 9.32. Obtained from: Jeffrey Hsu <hsu@FreeBSD.ORG>
1999-02-17 20:01:20 +00:00
if (dap->da_offset == oldoffset) {
dap->da_offset = newoffset;
break;
}
}
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
done:
bcopy(oldloc, newloc, entrysize);
FREE_LOCK(&lk);
}
/*
* Free a diradd dependency structure. This routine must be called
* with splbio interrupts blocked.
*/
static void
free_diradd(dap)
struct diradd *dap;
{
struct dirrem *dirrem;
struct pagedep *pagedep;
struct inodedep *inodedep;
struct mkdir *mkdir, *nextmd;
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held == NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("free_diradd: lock not held");
#endif
WORKLIST_REMOVE(&dap->da_list);
LIST_REMOVE(dap, da_pdlist);
if ((dap->da_state & DIRCHG) == 0) {
pagedep = dap->da_pagedep;
} else {
dirrem = dap->da_previous;
pagedep = dirrem->dm_pagedep;
Handle the case of moving a directory onto the top of a sibling's child of the same name. Submitted by: Kirk Mckusick with fixes from luoqi Chen Obtained from: Whistle test tree.
1998-08-12 20:46:47 +00:00
dirrem->dm_dirinum = pagedep->pd_ino;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
add_to_worklist(&dirrem->dm_list);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
0, &inodedep) != 0)
(void) free_inodedep(inodedep);
if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
nextmd = LIST_NEXT(mkdir, md_mkdirs);
if (mkdir->md_diradd != dap)
continue;
dap->da_state &= ~mkdir->md_state;
WORKLIST_REMOVE(&mkdir->md_list);
LIST_REMOVE(mkdir, md_mkdirs);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(mkdir, D_MKDIR);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("free_diradd: unfound ref");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(dap, D_DIRADD);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Directory entry removal dependencies.
*
* When removing a directory entry, the entry's inode pointer must be
* zero'ed on disk before the corresponding inode's link count is decremented
* (possibly freeing the inode for re-use). This dependency is handled by
* updating the directory entry but delaying the inode count reduction until
* after the directory block has been written to disk. After this point, the
* inode count can be decremented whenever it is convenient.
*/
/*
* This routine should be called immediately after removing
* a directory entry. The inode's link count should not be
* decremented by the calling procedure -- the soft updates
* code will do this task when it is safe.
*/
void
softdep_setup_remove(bp, dp, ip, isrmdir)
struct buf *bp; /* buffer containing directory block */
struct inode *dp; /* inode for the directory being modified */
struct inode *ip; /* inode for directory entry being removed */
int isrmdir; /* indicates if doing RMDIR */
{
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
struct dirrem *dirrem, *prevdirrem;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
*/
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
/*
* If the COMPLETE flag is clear, then there were no active
* entries and we want to roll back to a zeroed entry until
* the new inode is committed to disk. If the COMPLETE flag is
* set then we have deleted an entry that never made it to
* disk. If the entry we deleted resulted from a name change,
* then the old name still resides on disk. We cannot delete
* its inode (returned to us in prevdirrem) until the zeroed
* directory entry gets to disk. The new inode has never been
* referenced on the disk, so can be deleted immediately.
*/
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if ((dirrem->dm_state & COMPLETE) == 0) {
LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
dm_next);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
} else {
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
if (prevdirrem != NULL)
LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
prevdirrem, dm_next);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
FREE_LOCK(&lk);
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
handle_workitem_remove(dirrem, NULL);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
}
/*
* Allocate a new dirrem if appropriate and return it along with
* its associated pagedep. Called without a lock, returns with lock.
*/
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
static long num_dirrem; /* number of dirrem allocated */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
static struct dirrem *
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
newdirrem(bp, dp, ip, isrmdir, prevdirremp)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct buf *bp; /* buffer containing directory block */
struct inode *dp; /* inode for the directory being modified */
struct inode *ip; /* inode for directory entry being removed */
int isrmdir; /* indicates if doing RMDIR */
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
struct dirrem **prevdirremp; /* previously referenced inode, if any */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
int offset;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ufs_lbn_t lbn;
struct diradd *dap;
struct dirrem *dirrem;
struct pagedep *pagedep;
/*
* Whiteouts have no deletion dependencies.
*/
if (ip == NULL)
panic("newdirrem: whiteout");
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
/*
* If we are over our limit, try to improve the situation.
* Limiting the number of dirrem structures will also limit
* the number of freefile and freeblks structures.
*/
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
if (num_dirrem > max_softdeps / 2)
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
(void) request_cleanup(FLUSH_REMOVE, 0);
num_dirrem += 1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
dirrem->dm_list.wk_type = D_DIRREM;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dirrem->dm_state = isrmdir ? RMDIR : 0;
dirrem->dm_mnt = ITOV(ip)->v_mount;
dirrem->dm_oldinum = ip->i_number;
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
*prevdirremp = NULL;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ACQUIRE_LOCK(&lk);
lbn = lblkno(dp->i_fs, dp->i_offset);
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
offset = blkoff(dp->i_fs, dp->i_offset);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
dirrem->dm_pagedep = pagedep;
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
/*
* Check for a diradd dependency for the same directory entry.
* If present, then both dependencies become obsolete and can
* be de-allocated. Check for an entry on both the pd_dirraddhd
* list and the pd_pendinghd list.
*/
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
if (dap->da_offset == offset)
break;
if (dap == NULL) {
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
if (dap->da_offset == offset)
break;
if (dap == NULL)
return (dirrem);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
/*
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
* Must be ATTACHED at this point.
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
*/
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((dap->da_state & ATTACHED) == 0) {
FREE_LOCK(&lk);
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
panic("newdirrem: not ATTACHED");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
if (dap->da_newinum != ip->i_number) {
FREE_LOCK(&lk);
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
panic("newdirrem: inum %d should be %d",
ip->i_number, dap->da_newinum);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
/*
* If we are deleting a changed name that never made it to disk,
* then return the dirrem describing the previous inode (which
* represents the inode currently referenced from this entry on disk).
*/
if ((dap->da_state & DIRCHG) != 0) {
*prevdirremp = dap->da_previous;
dap->da_state &= ~DIRCHG;
dap->da_pagedep = pagedep;
}
/*
* We are deleting an entry that never made it to disk.
* Mark it COMPLETE so we can delete its inode immediately.
*/
Fix for "live inode" panic. Submitted by: Kirk McKusick <mckusick@McKusick.COM> Reviewed by: yeah right...
1998-06-10 20:45:46 +00:00
dirrem->dm_state |= COMPLETE;
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
free_diradd(dap);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (dirrem);
}
/*
* Directory entry change dependencies.
*
* Changing an existing directory entry requires that an add operation
* be completed first followed by a deletion. The semantics for the addition
* are identical to the description of adding a new entry above except
* that the rollback is to the old inode number rather than zero. Once
* the addition dependency is completed, the removal is done as described
* in the removal routine above.
*/
/*
* This routine should be called immediately after changing
* a directory entry. The inode's link count should not be
* decremented by the calling procedure -- the soft updates
* code will perform this task when it is safe.
*/
void
softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
struct buf *bp; /* buffer containing directory block */
struct inode *dp; /* inode for the directory being modified */
struct inode *ip; /* inode for directory entry being removed */
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ino_t newinum; /* new inode number for changed entry */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int isrmdir; /* indicates if doing RMDIR */
{
int offset;
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
struct diradd *dap = NULL;
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
struct dirrem *dirrem, *prevdirrem;
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
struct pagedep *pagedep;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inodedep *inodedep;
offset = blkoff(dp->i_fs, dp->i_offset);
/*
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
* Whiteouts do not need diradd dependencies.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
if (newinum != WINO) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
MALLOC(dap, struct diradd *, sizeof(struct diradd),
Convert more malloc+bzero to malloc+M_ZERO. Submitted by: josh@zipperup.org Submitted by: Robert Drehmel <robd@gmx.net>
2000-12-08 21:51:06 +00:00
M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
dap->da_list.wk_type = D_DIRADD;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
dap->da_offset = offset;
dap->da_newinum = newinum;
}
/*
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
* Allocate a new dirrem and ACQUIRE_LOCK.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
pagedep = dirrem->dm_pagedep;
Handle the case of moving a directory onto the top of a sibling's child of the same name. Submitted by: Kirk Mckusick with fixes from luoqi Chen Obtained from: Whistle test tree.
1998-08-12 20:46:47 +00:00
/*
* The possible values for isrmdir:
* 0 - non-directory file rename
* 1 - directory rename within same directory
* inum - directory rename to new directory of given inode number
* When renaming to a new directory, we are both deleting and
* creating a new directory entry, so the link count on the new
* directory should not change. Thus we do not need the followup
* dirrem which is usually done in handle_workitem_remove. We set
* the DIRCHG flag to tell handle_workitem_remove to skip the
* followup dirrem.
*/
if (isrmdir > 1)
dirrem->dm_state |= DIRCHG;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
* Whiteouts have no additional dependencies,
* so just put the dirrem on the correct list.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
if (newinum == WINO) {
if ((dirrem->dm_state & COMPLETE) == 0) {
LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
dm_next);
} else {
dirrem->dm_dirinum = pagedep->pd_ino;
add_to_worklist(&dirrem->dm_list);
}
FREE_LOCK(&lk);
return;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Must track multiple uncommitted renames until one ultimately gets committed to disk or is removed.
2000-01-17 06:28:18 +00:00
/*
* If the COMPLETE flag is clear, then there were no active
* entries and we want to roll back to the previous inode until
* the new inode is committed to disk. If the COMPLETE flag is
* set, then we have deleted an entry that never made it to disk.
* If the entry we deleted resulted from a name change, then the old
* inode reference still resides on disk. Any rollback that we do
* needs to be to that old inode (returned to us in prevdirrem). If
* the entry we deleted resulted from a create, then there is
* no entry on the disk, so we want to roll back to zero rather
* than the uncommitted inode. In either of the COMPLETE cases we
* want to immediately free the unwritten and unreferenced inode.
*/
if ((dirrem->dm_state & COMPLETE) == 0) {
dap->da_previous = dirrem;
} else {
if (prevdirrem != NULL) {
dap->da_previous = prevdirrem;
} else {
dap->da_state &= ~DIRCHG;
dap->da_pagedep = pagedep;
}
dirrem->dm_dirinum = pagedep->pd_ino;
add_to_worklist(&dirrem->dm_list);
}
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
/*
* Link into its inodedep. Put it on the id_bufwait list if the inode
* is not yet written. If it is written, do the post-inode write
* processing to put it on the id_pendinghd list.
*/
if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
(inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
dap->da_state |= COMPLETE;
LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
} else {
LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dap, da_pdlist);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
Fix the case when renaming to a file that you've just created and deleted, that had an inode that has not yet been written to disk, when the inode of the new file is also not yet written to disk, and your old directory entry is not yet on disk but you need to remove it and the new name exists in memory but has been deleted but the transaction to write the deleted name to disk exists and has not yet been cancelled by the request to delete the non existant name. I don't know how kirk could have missed such a glaring problem for so long. :-) Especially since the inconsitency survived on the disk for a whole 4 second on average before being fixed by other code. This was not a crashing bug but just led to filesystem inconsitencies if you crashed. Submitted by: Kirk McKusick (mckusick@mckusick.com)
1998-06-12 20:48:30 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
}
/*
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
* Called whenever the link count on an inode is changed.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* It creates an inode dependency so that the new reference(s)
* to the inode cannot be committed to disk until the updated
* inode has been written.
*/
void
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
softdep_change_linkcnt(ip)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inode *ip; /* the inode with the increased link count */
{
struct inodedep *inodedep;
ACQUIRE_LOCK(&lk);
(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (ip->i_nlink < ip->i_effnlink) {
FREE_LOCK(&lk);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
panic("softdep_change_linkcnt: bad delta");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
}
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
/*
* Called when the effective link count and the reference count
* on an inode drops to zero. At this point there are no names
* referencing the file in the filesystem and no active file
* references. The space associated with the file will be freed
* as soon as the necessary soft dependencies are cleared.
*/
void
softdep_releasefile(ip)
struct inode *ip; /* inode with the zero effective link count */
{
struct inodedep *inodedep;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
struct fs *fs;
int extblocks;
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
if (ip->i_effnlink > 0)
panic("softdep_filerelease: file still referenced");
/*
* We may be called several times as the real reference count
* drops to zero. We only want to account for the space once.
*/
if (ip->i_flag & IN_SPACECOUNTED)
return;
/*
* We have to deactivate a snapshot otherwise copyonwrites may
* add blocks and the cleanup may remove blocks after we have
* tried to account for them.
*/
if ((ip->i_flags & SF_SNAPSHOT) != 0)
ffs_snapremove(ITOV(ip));
/*
* If we are tracking an nlinkdelta, we have to also remember
* whether we accounted for the freed space yet.
*/
ACQUIRE_LOCK(&lk);
if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep)))
inodedep->id_state |= SPACECOUNTED;
FREE_LOCK(&lk);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
fs = ip->i_fs;
extblocks = 0;
if (fs->fs_magic == FS_UFS2_MAGIC)
extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
ip->i_fs->fs_pendinginodes += 1;
ip->i_flag |= IN_SPACECOUNTED;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* This workitem decrements the inode's link count.
* If the link count reaches zero, the file is removed.
*/
static void
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
handle_workitem_remove(dirrem, xp)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct dirrem *dirrem;
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
struct vnode *xp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td = curthread;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inodedep *inodedep;
struct vnode *vp;
struct inode *ip;
We cannot proceed to free the blocks of the file until the dependencies have been cleaned up by deallocte_dependencies(). Once that is done, it is safe to post the request to free the blocks. A similar change is also needed for the freefile case.
2000-01-11 06:52:35 +00:00
ino_t oldinum;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int error;
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if ((vp = xp) == NULL &&
(error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE,
&vp)) != 0) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error("handle_workitem_remove: vget", error);
return;
}
ip = VTOI(vp);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
ACQUIRE_LOCK(&lk);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
FREE_LOCK(&lk);
We cannot proceed to free the blocks of the file until the dependencies have been cleaned up by deallocte_dependencies(). Once that is done, it is safe to post the request to free the blocks. A similar change is also needed for the freefile case.
2000-01-11 06:52:35 +00:00
panic("handle_workitem_remove: lost inodedep");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Normal file deletion.
*/
if ((dirrem->dm_state & RMDIR) == 0) {
ip->i_nlink--;
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_nlink, ip->i_nlink);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
ip->i_flag |= IN_CHANGE;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (ip->i_nlink < ip->i_effnlink) {
FREE_LOCK(&lk);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
panic("handle_workitem_remove: bad file delta");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
vput(vp);
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
num_dirrem -= 1;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(dirrem, D_DIRREM);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return;
}
/*
* Directory deletion. Decrement reference count for both the
* just deleted parent directory entry and the reference for ".".
* Next truncate the directory to length zero. When the
* truncation completes, arrange to have the reference count on
* the parent decremented to account for the loss of "..".
*/
ip->i_nlink -= 2;
Avoid using casts as lvalues. Introduce DIP_SET macro which sets proper inode field based on UFS version. Use DIP ro read values and DIP_SET to modify them throughout FFS code base.
2004-07-28 06:41:27 +00:00
DIP_SET(ip, i_nlink, ip->i_nlink);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
ip->i_flag |= IN_CHANGE;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (ip->i_nlink < ip->i_effnlink) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_workitem_remove: bad dir delta");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
FREE_LOCK(&lk);
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error("handle_workitem_remove: truncate", error);
Handle the case of moving a directory onto the top of a sibling's child of the same name. Submitted by: Kirk Mckusick with fixes from luoqi Chen Obtained from: Whistle test tree.
1998-08-12 20:46:47 +00:00
/*
* Rename a directory to a new parent. Since, we are both deleting
* and creating a new directory entry, the link count on the new
* directory should not change. Thus we skip the followup dirrem.
*/
if (dirrem->dm_state & DIRCHG) {
vput(vp);
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
num_dirrem -= 1;
Handle the case of moving a directory onto the top of a sibling's child of the same name. Submitted by: Kirk Mckusick with fixes from luoqi Chen Obtained from: Whistle test tree.
1998-08-12 20:46:47 +00:00
WORKITEM_FREE(dirrem, D_DIRREM);
return;
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
/*
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
* If the inodedep does not exist, then the zero'ed inode has
* been written to disk. If the allocated inode has never been
* written to disk, then the on-disk inode is zero'ed. In either
* case we can remove the file immediately.
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
*/
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ACQUIRE_LOCK(&lk);
We cannot proceed to free the blocks of the file until the dependencies have been cleaned up by deallocte_dependencies(). Once that is done, it is safe to post the request to free the blocks. A similar change is also needed for the freefile case.
2000-01-11 06:52:35 +00:00
dirrem->dm_state = 0;
oldinum = dirrem->dm_oldinum;
dirrem->dm_oldinum = dirrem->dm_dirinum;
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
check_inode_unwritten(inodedep)) {
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
FREE_LOCK(&lk);
vput(vp);
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
handle_workitem_remove(dirrem, NULL);
We cannot proceed to free the blocks of the file until the dependencies have been cleaned up by deallocte_dependencies(). Once that is done, it is safe to post the request to free the blocks. A similar change is also needed for the freefile case.
2000-01-11 06:52:35 +00:00
return;
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
}
During fastpath processing for removal of a short-lived inode, the set of restrictions for cancelling an inode dependency (inodedep) is somewhat stronger than originally coded. Since this check appears in two places, we codify it into the function check_inode_unwritten which we then call from the two sites, one freeing blocks and the other freeing directory entries. Submitted by: Steinar Haug via Matthew Dillon
2000-01-18 01:33:05 +00:00
WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
We cannot proceed to free the blocks of the file until the dependencies have been cleaned up by deallocte_dependencies(). Once that is done, it is safe to post the request to free the blocks. A similar change is also needed for the freefile case.
2000-01-11 06:52:35 +00:00
FREE_LOCK(&lk);
vput(vp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Inode de-allocation dependencies.
*
* When an inode's link count is reduced to zero, it can be de-allocated. We
* found it convenient to postpone de-allocation until after the inode is
* written to disk with its new link count (zero). At this point, all of the
* on-disk inode's block pointers are nullified and, with careful dependency
* list ordering, all dependencies related to the inode will be satisfied and
* the corresponding dependency structures de-allocated. So, if/when the
* inode is reused, there will be no mixing of old dependencies with new
* ones. This artificial dependency is set up by the block de-allocation
* procedure above (softdep_setup_freeblocks) and completed by the
* following procedure.
*/
static void
handle_workitem_freefile(freefile)
struct freefile *freefile;
{
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
struct fs *fs;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inodedep *idp;
int error;
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#ifdef DEBUG
ACQUIRE_LOCK(&lk);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (error)
panic("handle_workitem_freefile: inodedep survived");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
fs->fs_pendinginodes -= 1;
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
if ((error = ffs_freefile(fs, freefile->fx_devvp, freefile->fx_oldinum,
freefile->fx_mode)) != 0)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error("handle_workitem_freefile", error);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(freefile, D_FREEFILE);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
When I was a kid my work table was one cluttered mess an cleaning it up were a rather overwhelming task. I soon learned that if you don't know where you're going to store something, at least try to pile it next to something slightly related in the hope that a pattern emerges. Apply the same principle to the ffs/snapshot/softupdates code which have leaked into specfs: Add yet a buf-quasi-method and call it from the only two places I can see it can make a difference and implement the magic in ffs_softdep.c where it belongs. It's not pretty, but at least it's one less layer violated.
2004-03-11 18:50:33 +00:00
static int
softdep_disk_prewrite(struct vnode *vp, struct buf *bp)
{
int error;
if (bp->b_iocmd != BIO_WRITE)
return (0);
if ((bp->b_flags & B_VALIDSUSPWRT) == 0 &&
bp->b_vp != NULL && bp->b_vp->v_mount != NULL &&
Fix a paste-o from the buf_prewrite() cleanup commit and check for the MNTK_SUSPEND flag on the correct vnode pointer in softdep_disk_prewrite(). Reviewed by: phk Tested by: kensmith
2004-04-06 19:20:24 +00:00
(bp->b_vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED) != 0)
When I was a kid my work table was one cluttered mess an cleaning it up were a rather overwhelming task. I soon learned that if you don't know where you're going to store something, at least try to pile it next to something slightly related in the hope that a pattern emerges. Apply the same principle to the ffs/snapshot/softupdates code which have leaked into specfs: Add yet a buf-quasi-method and call it from the only two places I can see it can make a difference and implement the magic in ffs_softdep.c where it belongs. It's not pretty, but at least it's one less layer violated.
2004-03-11 18:50:33 +00:00
panic("softdep_disk_prewrite: bad I/O");
bp->b_flags &= ~B_VALIDSUSPWRT;
if (LIST_FIRST(&bp->b_dep) != NULL)
buf_start(bp);
mp_fixme("This should require the vnode lock.");
if ((vp->v_vflag & VV_COPYONWRITE) &&
Degeneralize the per cdev copyonwrite callback. The only possible value is ffs_copyonwrite() and the only place it can be called from is FFS which would never want to call another filesystems copyonwrite method, should one exist, so there is no reason why anything generic should know about this.
2004-10-26 06:25:56 +00:00
vp->v_rdev->si_snapdata != NULL &&
(error = (ffs_copyonwrite)(vp, bp)) != 0 &&
When I was a kid my work table was one cluttered mess an cleaning it up were a rather overwhelming task. I soon learned that if you don't know where you're going to store something, at least try to pile it next to something slightly related in the hope that a pattern emerges. Apply the same principle to the ffs/snapshot/softupdates code which have leaked into specfs: Add yet a buf-quasi-method and call it from the only two places I can see it can make a difference and implement the magic in ffs_softdep.c where it belongs. It's not pretty, but at least it's one less layer violated.
2004-03-11 18:50:33 +00:00
error != EOPNOTSUPP) {
use bufdone() not biodone().
2004-08-08 13:23:05 +00:00
bp->b_error = error;
bp->b_ioflags |= BIO_ERROR;
bufdone(bp);
When I was a kid my work table was one cluttered mess an cleaning it up were a rather overwhelming task. I soon learned that if you don't know where you're going to store something, at least try to pile it next to something slightly related in the hope that a pattern emerges. Apply the same principle to the ffs/snapshot/softupdates code which have leaked into specfs: Add yet a buf-quasi-method and call it from the only two places I can see it can make a difference and implement the magic in ffs_softdep.c where it belongs. It's not pretty, but at least it's one less layer violated.
2004-03-11 18:50:33 +00:00
return (1);
}
return (0);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Disk writes.
*
* The dependency structures constructed above are most actively used when file
* system blocks are written to disk. No constraints are placed on when a
* block can be written, but unsatisfied update dependencies are made safe by
* modifying (or replacing) the source memory for the duration of the disk
* write. When the disk write completes, the memory block is again brought
* up-to-date.
*
* In-core inode structure reclamation.
*
* Because there are a finite number of "in-core" inode structures, they are
* reused regularly. By transferring all inode-related dependencies to the
* in-memory inode block and indexing them separately (via "inodedep"s), we
* can allow "in-core" inode structures to be reused at any time and avoid
* any increase in contention.
*
* Called just before entering the device driver to initiate a new disk I/O.
* The buffer must be locked, thus, no I/O completion operations can occur
* while we are manipulating its associated dependencies.
*/
Make static non-exported functions from soft updates.
2000-01-09 22:40:09 +00:00
static void
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_disk_io_initiation(bp)
struct buf *bp; /* structure describing disk write to occur */
{
struct worklist *wk, *nextwk;
struct indirdep *indirdep;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
struct inodedep *inodedep;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* We only care about write operations. There should never
* be dependencies for reads.
*/
Remove B_READ, B_WRITE and B_FREEBUF and replace them with a new field in struct buf: b_iocmd. The b_iocmd is enforced to have exactly one bit set. B_WRITE was bogusly defined as zero giving rise to obvious coding mistakes. Also eliminate the redundant struct buf flag B_CALL, it can just as efficiently be done by comparing b_iodone to NULL. Should you get a panic or drop into the debugger, complaining about "b_iocmd", don't continue. It is likely to write on your disk where it should have been reading. This change is a step in the direction towards a stackable BIO capability. A lot of this patch were machine generated (Thanks to style(9) compliance!) Vinum users: Greg has not had time to test this yet, be careful.
2000-03-20 10:44:49 +00:00
if (bp->b_iocmd == BIO_READ)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_disk_io_initiation: read");
/*
* Do any necessary pre-I/O processing.
*/
for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
nextwk = LIST_NEXT(wk, wk_list);
switch (wk->wk_type) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_PAGEDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
initiate_write_filepage(WK_PAGEDEP(wk), bp);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INODEDEP:
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
inodedep = WK_INODEDEP(wk);
if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
initiate_write_inodeblock_ufs1(inodedep, bp);
else
initiate_write_inodeblock_ufs2(inodedep, bp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INDIRDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
indirdep = WK_INDIRDEP(wk);
if (indirdep->ir_state & GOINGAWAY)
panic("disk_io_initiation: indirdep gone");
/*
* If there are no remaining dependencies, this
* will be writing the real pointers, so the
* dependency can be freed.
*/
if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
indirdep->ir_savebp->b_flags |=
B_INVAL | B_NOCACHE;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
brelse(indirdep->ir_savebp);
/* inline expand WORKLIST_REMOVE(wk); */
wk->wk_state &= ~ONWORKLIST;
LIST_REMOVE(wk, wk_list);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(indirdep, D_INDIRDEP);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
}
/*
* Replace up-to-date version with safe version.
*/
Cannot do MALLOC with M_WAITOK while holding ACQUIRE_LOCK Obtained from: Ethan Solomita <ethan@geocast.com>
2000-09-07 23:02:55 +00:00
MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
Implement a low-memory deadlock solution. Removed most of the hacks that were trying to deal with low-memory situations prior to now. The new code is based on the concept that I/O must be able to function in a low memory situation. All major modules related to I/O (except networking) have been adjusted to allow allocation out of the system reserve memory pool. These modules now detect a low memory situation but rather then block they instead continue to operate, then return resources to the memory pool instead of cache them or leave them wired. Code has been added to stall in a low-memory situation prior to a vnode being locked. Thus situations where a process blocks in a low-memory condition while holding a locked vnode have been reduced to near nothing. Not only will I/O continue to operate, but many prior deadlock conditions simply no longer exist. Implement a number of VFS/BIO fixes (found by Ian): in biodone(), bogus-page replacement code, the loop was not properly incrementing loop variables prior to a continue statement. We do not believe this code can be hit anyway but we aren't taking any chances. We'll turn the whole section into a panic (as it already is in brelse()) after the release is rolled. In biodone(), the foff calculation was incorrectly clamped to the iosize, causing the wrong foff to be calculated for pages in the case of an I/O error or biodone() called without initiating I/O. The problem always caused a panic before. Now it doesn't. The problem is mainly an issue with NFS. Fixed casts for ~PAGE_MASK. This code worked properly before only because the calculations use signed arithmatic. Better to properly extend PAGE_MASK first before inverting it for the 64 bit masking op. In brelse(), the bogus_page fixup code was improperly throwing away the original contents of 'm' when it did the j-loop to fix the bogus pages. The result was that it would potentially invalidate parts of the *WRONG* page(!), leading to corruption. There may still be cases where a background bitmap write is being duplicated, causing potential corruption. We have identified a potentially serious bug related to this but the fix is still TBD. So instead this patch contains a KASSERT to detect the problem and panic the machine rather then continue to corrupt the filesystem. The problem does not occur very often.. it is very hard to reproduce, and it may or may not be the cause of the corruption people have reported. Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>) Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
2000-11-18 23:06:26 +00:00
M_INDIRDEP, M_SOFTDEP_FLAGS);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ACQUIRE_LOCK(&lk);
indirdep->ir_state &= ~ATTACHED;
indirdep->ir_state |= UNDONE;
Bring up-to-date with Whistle's current version Includes some debugging code.
1998-05-19 23:07:25 +00:00
bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
bcopy(indirdep->ir_savebp->b_data, bp->b_data,
bp->b_bcount);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_MKDIR:
case D_BMSAFEMAP:
case D_ALLOCDIRECT:
case D_ALLOCINDIR:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
default:
panic("handle_disk_io_initiation: Unexpected type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
}
}
}
/*
* Called from within the procedure above to deal with unsatisfied
* allocation dependencies in a directory. The buffer must be locked,
* thus, no I/O completion operations can occur while we are
* manipulating its associated dependencies.
*/
static void
initiate_write_filepage(pagedep, bp)
struct pagedep *pagedep;
struct buf *bp;
{
struct diradd *dap;
struct direct *ep;
int i;
if (pagedep->pd_state & IOSTARTED) {
/*
* This can only happen if there is a driver that does not
* understand chaining. Here biodone will reissue the call
* to strategy for the incomplete buffers.
*/
printf("initiate_write_filepage: already started\n");
return;
}
pagedep->pd_state |= IOSTARTED;
ACQUIRE_LOCK(&lk);
for (i = 0; i < DAHASHSZ; i++) {
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ep = (struct direct *)
((char *)bp->b_data + dap->da_offset);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (ep->d_ino != dap->da_newinum) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("%s: dir inum %d != new %d",
"initiate_write_filepage",
ep->d_ino, dap->da_newinum);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (dap->da_state & DIRCHG)
ep->d_ino = dap->da_previous->dm_oldinum;
else
ep->d_ino = 0;
dap->da_state &= ~ATTACHED;
dap->da_state |= UNDONE;
}
}
FREE_LOCK(&lk);
}
/*
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
* Version of initiate_write_inodeblock that handles UFS1 dinodes.
* Note that any bug fixes made to this routine must be done in the
* version found below.
*
* Called from within the procedure above to deal with unsatisfied
* allocation dependencies in an inodeblock. The buffer must be
* locked, thus, no I/O completion operations can occur while we
* are manipulating its associated dependencies.
*/
static void
initiate_write_inodeblock_ufs1(inodedep, bp)
struct inodedep *inodedep;
struct buf *bp; /* The inode block */
{
struct allocdirect *adp, *lastadp;
struct ufs1_dinode *dp;
struct fs *fs;
ufs_lbn_t i, prevlbn = 0;
int deplist;
if (inodedep->id_state & IOSTARTED)
panic("initiate_write_inodeblock_ufs1: already started");
inodedep->id_state |= IOSTARTED;
fs = inodedep->id_fs;
dp = (struct ufs1_dinode *)bp->b_data +
ino_to_fsbo(fs, inodedep->id_ino);
/*
* If the bitmap is not yet written, then the allocated
* inode cannot be written to disk.
*/
if ((inodedep->id_state & DEPCOMPLETE) == 0) {
if (inodedep->id_savedino1 != NULL)
panic("initiate_write_inodeblock_ufs1: I/O underway");
MALLOC(inodedep->id_savedino1, struct ufs1_dinode *,
sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
*inodedep->id_savedino1 = *dp;
bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
return;
}
/*
* If no dependencies, then there is nothing to roll back.
*/
inodedep->id_savedsize = dp->di_size;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
inodedep->id_savedextsize = 0;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
return;
/*
* Set the dependencies to busy.
*/
ACQUIRE_LOCK(&lk);
for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
adp = TAILQ_NEXT(adp, ad_next)) {
#ifdef DIAGNOSTIC
if (deplist != 0 && prevlbn >= adp->ad_lbn) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: lbn order");
}
prevlbn = adp->ad_lbn;
if (adp->ad_lbn < NDADDR &&
dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
FREE_LOCK(&lk);
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s: direct pointer #%jd mismatch %d != %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"softdep_write_inodeblock",
(intmax_t)adp->ad_lbn,
dp->di_db[adp->ad_lbn],
(intmax_t)adp->ad_newblkno);
}
if (adp->ad_lbn >= NDADDR &&
dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
FREE_LOCK(&lk);
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s: indirect pointer #%jd mismatch %d != %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"softdep_write_inodeblock",
(intmax_t)adp->ad_lbn - NDADDR,
dp->di_ib[adp->ad_lbn - NDADDR],
(intmax_t)adp->ad_newblkno);
}
deplist |= 1 << adp->ad_lbn;
if ((adp->ad_state & ATTACHED) == 0) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: Unknown state 0x%x",
adp->ad_state);
}
#endif /* DIAGNOSTIC */
adp->ad_state &= ~ATTACHED;
adp->ad_state |= UNDONE;
}
/*
* The on-disk inode cannot claim to be any larger than the last
* fragment that has been written. Otherwise, the on-disk inode
* might have fragments that were not the last block in the file
* which would corrupt the filesystem.
*/
for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
if (adp->ad_lbn >= NDADDR)
break;
dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
/* keep going until hitting a rollback to a frag */
if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
continue;
dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
#ifdef DIAGNOSTIC
if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: lost dep1");
}
#endif /* DIAGNOSTIC */
dp->di_db[i] = 0;
}
for (i = 0; i < NIADDR; i++) {
#ifdef DIAGNOSTIC
if (dp->di_ib[i] != 0 &&
(deplist & ((1 << NDADDR) << i)) == 0) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: lost dep2");
}
#endif /* DIAGNOSTIC */
dp->di_ib[i] = 0;
}
FREE_LOCK(&lk);
return;
}
/*
* If we have zero'ed out the last allocated block of the file,
* roll back the size to the last currently allocated block.
* We know that this last allocated block is a full-sized as
* we already checked for fragments in the loop above.
*/
if (lastadp != NULL &&
dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
for (i = lastadp->ad_lbn; i >= 0; i--)
if (dp->di_db[i] != 0)
break;
dp->di_size = (i + 1) * fs->fs_bsize;
}
/*
* The only dependencies are for indirect blocks.
*
* The file size for indirect block additions is not guaranteed.
* Such a guarantee would be non-trivial to achieve. The conventional
* synchronous write implementation also does not make this guarantee.
* Fsck should catch and fix discrepancies. Arguably, the file size
* can be over-estimated without destroying integrity when the file
* moves into the indirect blocks (i.e., is large). If we want to
* postpone fsck, we are stuck with this argument.
*/
for (; adp; adp = TAILQ_NEXT(adp, ad_next))
dp->di_ib[adp->ad_lbn - NDADDR] = 0;
FREE_LOCK(&lk);
}
/*
* Version of initiate_write_inodeblock that handles UFS2 dinodes.
* Note that any bug fixes made to this routine must be done in the
* version found above.
*
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* Called from within the procedure above to deal with unsatisfied
* allocation dependencies in an inodeblock. The buffer must be
* locked, thus, no I/O completion operations can occur while we
* are manipulating its associated dependencies.
*/
static void
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
initiate_write_inodeblock_ufs2(inodedep, bp)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct inodedep *inodedep;
struct buf *bp; /* The inode block */
{
struct allocdirect *adp, *lastadp;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
struct ufs2_dinode *dp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct fs *fs;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
ufs_lbn_t i, prevlbn = 0;
int deplist;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (inodedep->id_state & IOSTARTED)
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
panic("initiate_write_inodeblock_ufs2: already started");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_state |= IOSTARTED;
fs = inodedep->id_fs;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
dp = (struct ufs2_dinode *)bp->b_data +
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ino_to_fsbo(fs, inodedep->id_ino);
/*
* If the bitmap is not yet written, then the allocated
* inode cannot be written to disk.
*/
if ((inodedep->id_state & DEPCOMPLETE) == 0) {
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (inodedep->id_savedino2 != NULL)
panic("initiate_write_inodeblock_ufs2: I/O underway");
MALLOC(inodedep->id_savedino2, struct ufs2_dinode *,
sizeof(struct ufs2_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
*inodedep->id_savedino2 = *dp;
bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return;
}
/*
* If no dependencies, then there is nothing to roll back.
*/
inodedep->id_savedsize = dp->di_size;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
inodedep->id_savedextsize = dp->di_extsize;
if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL &&
TAILQ_FIRST(&inodedep->id_extupdt) == NULL)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return;
/*
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
* Set the ext data dependencies to busy.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
ACQUIRE_LOCK(&lk);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
adp = TAILQ_NEXT(adp, ad_next)) {
#ifdef DIAGNOSTIC
if (deplist != 0 && prevlbn >= adp->ad_lbn) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: lbn order");
}
prevlbn = adp->ad_lbn;
if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) {
FREE_LOCK(&lk);
panic("%s: direct pointer #%jd mismatch %jd != %jd",
"softdep_write_inodeblock",
(intmax_t)adp->ad_lbn,
(intmax_t)dp->di_extb[adp->ad_lbn],
(intmax_t)adp->ad_newblkno);
}
deplist |= 1 << adp->ad_lbn;
if ((adp->ad_state & ATTACHED) == 0) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: Unknown state 0x%x",
adp->ad_state);
}
#endif /* DIAGNOSTIC */
adp->ad_state &= ~ATTACHED;
adp->ad_state |= UNDONE;
}
/*
* The on-disk inode cannot claim to be any larger than the last
* fragment that has been written. Otherwise, the on-disk inode
* might have fragments that were not the last block in the ext
* data which would corrupt the filesystem.
*/
for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
/* keep going until hitting a rollback to a frag */
if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
continue;
dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
#ifdef DIAGNOSTIC
if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) {
FREE_LOCK(&lk);
panic("softdep_write_inodeblock: lost dep1");
}
#endif /* DIAGNOSTIC */
dp->di_extb[i] = 0;
}
lastadp = NULL;
break;
}
/*
* If we have zero'ed out the last allocated block of the ext
* data, roll back the size to the last currently allocated block.
* We know that this last allocated block is a full-sized as
* we already checked for fragments in the loop above.
*/
if (lastadp != NULL &&
dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
for (i = lastadp->ad_lbn; i >= 0; i--)
if (dp->di_extb[i] != 0)
break;
dp->di_extsize = (i + 1) * fs->fs_bsize;
}
/*
* Set the file data dependencies to busy.
*/
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
adp = TAILQ_NEXT(adp, ad_next)) {
#ifdef DIAGNOSTIC
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (deplist != 0 && prevlbn >= adp->ad_lbn) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_write_inodeblock: lbn order");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
prevlbn = adp->ad_lbn;
if (adp->ad_lbn < NDADDR &&
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
FREE_LOCK(&lk);
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s: direct pointer #%jd mismatch %jd != %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"softdep_write_inodeblock",
(intmax_t)adp->ad_lbn,
(intmax_t)dp->di_db[adp->ad_lbn],
(intmax_t)adp->ad_newblkno);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (adp->ad_lbn >= NDADDR &&
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
FREE_LOCK(&lk);
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s indirect pointer #%jd mismatch %jd != %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"softdep_write_inodeblock:",
(intmax_t)adp->ad_lbn - NDADDR,
(intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
(intmax_t)adp->ad_newblkno);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
deplist |= 1 << adp->ad_lbn;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((adp->ad_state & ATTACHED) == 0) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_write_inodeblock: Unknown state 0x%x",
adp->ad_state);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif /* DIAGNOSTIC */
adp->ad_state &= ~ATTACHED;
adp->ad_state |= UNDONE;
}
/*
* The on-disk inode cannot claim to be any larger than the last
* fragment that has been written. Otherwise, the on-disk inode
* might have fragments that were not the last block in the file
* which would corrupt the filesystem.
*/
for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
if (adp->ad_lbn >= NDADDR)
break;
dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
/* keep going until hitting a rollback to a frag */
if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
continue;
dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
#ifdef DIAGNOSTIC
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
FREE_LOCK(&lk);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
panic("softdep_write_inodeblock: lost dep2");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif /* DIAGNOSTIC */
dp->di_db[i] = 0;
}
for (i = 0; i < NIADDR; i++) {
#ifdef DIAGNOSTIC
if (dp->di_ib[i] != 0 &&
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
(deplist & ((1 << NDADDR) << i)) == 0) {
FREE_LOCK(&lk);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
panic("softdep_write_inodeblock: lost dep3");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif /* DIAGNOSTIC */
dp->di_ib[i] = 0;
}
FREE_LOCK(&lk);
return;
}
/*
* If we have zero'ed out the last allocated block of the file,
* roll back the size to the last currently allocated block.
* We know that this last allocated block is a full-sized as
* we already checked for fragments in the loop above.
*/
if (lastadp != NULL &&
dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
for (i = lastadp->ad_lbn; i >= 0; i--)
if (dp->di_db[i] != 0)
break;
dp->di_size = (i + 1) * fs->fs_bsize;
}
/*
* The only dependencies are for indirect blocks.
*
* The file size for indirect block additions is not guaranteed.
* Such a guarantee would be non-trivial to achieve. The conventional
* synchronous write implementation also does not make this guarantee.
* Fsck should catch and fix discrepancies. Arguably, the file size
* can be over-estimated without destroying integrity when the file
* moves into the indirect blocks (i.e., is large). If we want to
* postpone fsck, we are stuck with this argument.
*/
for (; adp; adp = TAILQ_NEXT(adp, ad_next))
dp->di_ib[adp->ad_lbn - NDADDR] = 0;
FREE_LOCK(&lk);
}
/*
* This routine is called during the completion interrupt
* service routine for a disk write (from the procedure called
More s/file system/filesystem/g
2002-05-16 21:28:32 +00:00
* by the device driver to inform the filesystem caches of
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* a request completion). It should be called early in this
* procedure, before the block is made available to other
* processes or other routines are called.
*/
Make static non-exported functions from soft updates.
2000-01-09 22:40:09 +00:00
static void
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_disk_write_complete(bp)
struct buf *bp; /* describes the completed disk write */
{
struct worklist *wk;
struct workhead reattach;
struct newblk *newblk;
struct allocindir *aip;
struct allocdirect *adp;
struct indirdep *indirdep;
struct inodedep *inodedep;
struct bmsafemap *bmsafemap;
If an error occurs while writing a buffer, then the data will not have hit the disk and the dependencies cannot be unrolled. In this case, the system will mark the buffer as dirty again so that the write can be retried in the future. When the write succeeds or the system gives up on the buffer and marks it as invalid (B_INVAL), the dependencies will be cleared. Sponsored by: DARPA & NAI Labs.
2002-11-20 05:14:16 +00:00
/*
* If an error occurred while doing the write, then the data
* has not hit the disk and the dependencies cannot be unrolled.
*/
if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
return;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held != NOHOLDER)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_disk_write_complete: lock is held");
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = SPECIAL_FLAG;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif
LIST_INIT(&reattach);
while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
WORKLIST_REMOVE(wk);
switch (wk->wk_type) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_PAGEDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (handle_written_filepage(WK_PAGEDEP(wk), bp))
WORKLIST_INSERT(&reattach, wk);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INODEDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
WORKLIST_INSERT(&reattach, wk);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_BMSAFEMAP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bmsafemap = WK_BMSAFEMAP(wk);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
newblk->nb_state |= DEPCOMPLETE;
newblk->nb_bmsafemap = NULL;
LIST_REMOVE(newblk, nb_deps);
}
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
while ((adp =
LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
adp->ad_state |= DEPCOMPLETE;
adp->ad_buf = NULL;
LIST_REMOVE(adp, ad_deps);
handle_allocdirect_partdone(adp);
}
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
while ((aip =
LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
aip->ai_state |= DEPCOMPLETE;
aip->ai_buf = NULL;
LIST_REMOVE(aip, ai_deps);
handle_allocindir_partdone(aip);
}
while ((inodedep =
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_state |= DEPCOMPLETE;
LIST_REMOVE(inodedep, id_deps);
inodedep->id_buf = NULL;
}
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_MKDIR:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_ALLOCDIRECT:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
adp = WK_ALLOCDIRECT(wk);
adp->ad_state |= COMPLETE;
handle_allocdirect_partdone(adp);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_ALLOCINDIR:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
aip = WK_ALLOCINDIR(wk);
aip->ai_state |= COMPLETE;
handle_allocindir_partdone(aip);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INDIRDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
indirdep = WK_INDIRDEP(wk);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (indirdep->ir_state & GOINGAWAY) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("disk_write_complete: indirdep gone");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Bring up-to-date with Whistle's current version Includes some debugging code.
1998-05-19 23:07:25 +00:00
bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
FREE(indirdep->ir_saveddata, M_INDIRDEP);
indirdep->ir_saveddata = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
indirdep->ir_state &= ~UNDONE;
indirdep->ir_state |= ATTACHED;
while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
handle_allocindir_partdone(aip);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
fix double LIST_REMOVE; other cosmetic changes to match version 9.32. Obtained from: Jeffrey Hsu <hsu@FreeBSD.ORG>
1999-02-17 20:01:20 +00:00
panic("disk_write_complete: not gone");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
WORKLIST_INSERT(&reattach, wk);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
if ((bp->b_flags & B_DELWRI) == 0)
stat_indir_blk_ptrs++;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bdirty(bp);
continue;
default:
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_disk_write_complete: Unknown type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
}
}
/*
* Reattach any requests that must be redone.
*/
while ((wk = LIST_FIRST(&reattach)) != NULL) {
WORKLIST_REMOVE(wk);
WORKLIST_INSERT(&bp->b_dep, wk);
}
#ifdef DEBUG
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (lk.lkt_held != SPECIAL_FLAG)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_disk_write_complete: lock lost");
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif
}
/*
* Called from within softdep_disk_write_complete above. Note that
* this routine is always called from interrupt level with further
* splbio interrupts blocked.
*/
static void
handle_allocdirect_partdone(adp)
struct allocdirect *adp; /* the completed allocdirect */
{
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
struct allocdirectlst *listhead;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct allocdirect *listadp;
struct inodedep *inodedep;
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
long bsize, delay;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
return;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (adp->ad_buf != NULL) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_allocdirect_partdone: dangling dep");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* The on-disk inode cannot claim to be any larger than the last
* fragment that has been written. Otherwise, the on-disk inode
* might have fragments that were not the last block in the file
* which would corrupt the filesystem. Thus, we cannot free any
* allocdirects after one whose ad_oldblkno claims a fragment as
* these blocks must be rolled back to zero before writing the inode.
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
* We check the currently active set of allocdirects in id_inoupdt
* or id_extupdt as appropriate.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
inodedep = adp->ad_inodedep;
bsize = inodedep->id_fs->fs_bsize;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (adp->ad_state & EXTDATA)
listhead = &inodedep->id_extupdt;
else
listhead = &inodedep->id_inoupdt;
TAILQ_FOREACH(listadp, listhead, ad_next) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/* found our block */
if (listadp == adp)
break;
/* continue if ad_oldlbn is not a fragment */
if (listadp->ad_oldsize == 0 ||
listadp->ad_oldsize == bsize)
continue;
/* hit a fragment */
return;
}
/*
* If we have reached the end of the current list without
* finding the just finished dependency, then it must be
* on the future dependency list. Future dependencies cannot
* be freed until they are moved to the current list.
*/
if (listadp == NULL) {
#ifdef DEBUG
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (adp->ad_state & EXTDATA)
listhead = &inodedep->id_newextupdt;
else
listhead = &inodedep->id_newinoupdt;
TAILQ_FOREACH(listadp, listhead, ad_next)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/* found our block */
if (listadp == adp)
break;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (listadp == NULL) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_allocdirect_partdone: lost dep");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
#endif /* DEBUG */
return;
}
/*
* If we have found the just finished dependency, then free
* it along with anything that follows it that is complete.
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
* If the inode still has a bitmap dependency, then it has
* never been written to disk, hence the on-disk inode cannot
* reference the old fragment so we can free it without delay.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Some additional performance improvements. When freeing an inode check to see if it has been committed to disk. If it has never been written, it can be freed immediately. For short lived files this change allows the same inode to be reused repeatedly. Similarly, when upgrading a fragment to a larger size, if it has never been claimed by an inode on disk, it too can be freed immediately making it available for reuse often in the next slowly growing block of the same file.
2000-06-18 22:05:57 +00:00
delay = (inodedep->id_state & DEPCOMPLETE);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
for (; adp; adp = listadp) {
listadp = TAILQ_NEXT(adp, ad_next);
if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
return;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
free_allocdirect(listhead, adp, delay);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
}
/*
* Called from within softdep_disk_write_complete above. Note that
* this routine is always called from interrupt level with further
* splbio interrupts blocked.
*/
static void
handle_allocindir_partdone(aip)
struct allocindir *aip; /* the completed allocindir */
{
struct indirdep *indirdep;
if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
return;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (aip->ai_buf != NULL) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_allocindir_partdone: dangling dependency");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
indirdep = aip->ai_indirdep;
if (indirdep->ir_state & UNDONE) {
LIST_REMOVE(aip, ai_next);
LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
return;
}
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (indirdep->ir_state & UFS1FMT)
((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
aip->ai_newblkno;
else
((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
aip->ai_newblkno;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
LIST_REMOVE(aip, ai_next);
if (aip->ai_freefrag != NULL)
add_to_worklist(&aip->ai_freefrag->ff_list);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(aip, D_ALLOCINDIR);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Called from within softdep_disk_write_complete above to restore
* in-memory inode block contents to their most up-to-date state. Note
* that this routine is always called from interrupt level with further
* splbio interrupts blocked.
*/
static int
handle_written_inodeblock(inodedep, bp)
struct inodedep *inodedep;
struct buf *bp; /* buffer containing the inode block */
{
struct worklist *wk, *filefree;
struct allocdirect *adp, *nextadp;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
struct ufs1_dinode *dp1 = NULL;
struct ufs2_dinode *dp2 = NULL;
int hadchanges, fstype;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((inodedep->id_state & IOSTARTED) == 0) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_inodeblock: not started");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
inodedep->id_state &= ~IOSTARTED;
inodedep->id_state |= COMPLETE;
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
fstype = UFS1;
dp1 = (struct ufs1_dinode *)bp->b_data +
ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
} else {
fstype = UFS2;
dp2 = (struct ufs2_dinode *)bp->b_data +
ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* If we had to rollback the inode allocation because of
* bitmaps being incomplete, then simply restore it.
* Keep the block dirty so that it will not be reclaimed until
* all associated dependencies have been cleared and the
* corresponding updates written to disk.
*/
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (inodedep->id_savedino1 != NULL) {
if (fstype == UFS1)
*dp1 = *inodedep->id_savedino1;
else
*dp2 = *inodedep->id_savedino2;
FREE(inodedep->id_savedino1, M_INODEDEP);
inodedep->id_savedino1 = NULL;
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
if ((bp->b_flags & B_DELWRI) == 0)
stat_inode_bitmap++;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bdirty(bp);
return (1);
}
/*
* Roll forward anything that had to be rolled back before
* the inode could be updated.
*/
hadchanges = 0;
for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
nextadp = TAILQ_NEXT(adp, ad_next);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (adp->ad_state & ATTACHED) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_inodeblock: new entry");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (fstype == UFS1) {
if (adp->ad_lbn < NDADDR) {
if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
lk.lkt_held = NOHOLDER;
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s %s #%jd mismatch %d != %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"handle_written_inodeblock:",
"direct pointer",
(intmax_t)adp->ad_lbn,
dp1->di_db[adp->ad_lbn],
(intmax_t)adp->ad_oldblkno);
}
dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
} else {
if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) {
lk.lkt_held = NOHOLDER;
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s: %s #%jd allocated as %d",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"handle_written_inodeblock",
"indirect pointer",
(intmax_t)adp->ad_lbn - NDADDR,
dp1->di_ib[adp->ad_lbn - NDADDR]);
}
dp1->di_ib[adp->ad_lbn - NDADDR] =
adp->ad_newblkno;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
} else {
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (adp->ad_lbn < NDADDR) {
if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
lk.lkt_held = NOHOLDER;
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s: %s #%jd %s %jd != %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"handle_written_inodeblock",
"direct pointer",
(intmax_t)adp->ad_lbn, "mismatch",
(intmax_t)dp2->di_db[adp->ad_lbn],
(intmax_t)adp->ad_oldblkno);
}
dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
} else {
if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) {
lk.lkt_held = NOHOLDER;
Warning fixes for 64 bits platforms. This eliminates all the warnings I have had in the FFS code on sparc64. Reviewed by: mckusick
2002-06-23 18:17:27 +00:00
panic("%s: %s #%jd allocated as %jd",
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
"handle_written_inodeblock",
"indirect pointer",
(intmax_t)adp->ad_lbn - NDADDR,
(intmax_t)
dp2->di_ib[adp->ad_lbn - NDADDR]);
}
dp2->di_ib[adp->ad_lbn - NDADDR] =
adp->ad_newblkno;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
adp->ad_state &= ~UNDONE;
adp->ad_state |= ATTACHED;
hadchanges = 1;
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
nextadp = TAILQ_NEXT(adp, ad_next);
if (adp->ad_state & ATTACHED) {
lk.lkt_held = NOHOLDER;
panic("handle_written_inodeblock: new entry");
}
if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) {
lk.lkt_held = NOHOLDER;
panic("%s: direct pointers #%jd %s %jd != %jd",
"handle_written_inodeblock",
(intmax_t)adp->ad_lbn, "mismatch",
(intmax_t)dp2->di_extb[adp->ad_lbn],
(intmax_t)adp->ad_oldblkno);
}
dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
adp->ad_state &= ~UNDONE;
adp->ad_state |= ATTACHED;
hadchanges = 1;
}
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
stat_direct_blk_ptrs++;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Reset the file size to its most up-to-date value.
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_inodeblock: bad size");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
if (fstype == UFS1) {
if (dp1->di_size != inodedep->id_savedsize) {
dp1->di_size = inodedep->id_savedsize;
hadchanges = 1;
}
} else {
if (dp2->di_size != inodedep->id_savedsize) {
dp2->di_size = inodedep->id_savedsize;
hadchanges = 1;
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (dp2->di_extsize != inodedep->id_savedextsize) {
dp2->di_extsize = inodedep->id_savedextsize;
hadchanges = 1;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
inodedep->id_savedsize = -1;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
inodedep->id_savedextsize = -1;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* If there were any rollbacks in the inode block, then it must be
* marked dirty so that its will eventually get written back in
* its correct form.
*/
if (hadchanges)
bdirty(bp);
/*
* Process any allocdirects that completed during the update.
*/
if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
handle_allocdirect_partdone(adp);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
handle_allocdirect_partdone(adp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Process deallocations that were held pending until the
* inode had been written to disk. Freeing of the inode
* is delayed until after all blocks have been freed to
* avoid creation of new <vfsid, inum, lbn> triples
* before the old ones have been deleted.
*/
filefree = NULL;
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
WORKLIST_REMOVE(wk);
switch (wk->wk_type) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_FREEFILE:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* We defer adding filefree to the worklist until
* all other additions have been made to ensure
* that it will be done after all the old blocks
* have been freed.
*/
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (filefree != NULL) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_inodeblock: filefree");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
filefree = wk;
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_MKDIR:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_DIRADD:
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
diradd_inode_written(WK_DIRADD(wk), inodedep);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_FREEBLKS:
case D_FREEFRAG:
case D_DIRREM:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
add_to_worklist(wk);
continue;
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
case D_NEWDIRBLK:
free_newdirblk(WK_NEWDIRBLK(wk));
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
default:
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_inodeblock: Unknown type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
}
}
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
if (filefree != NULL) {
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (free_inodedep(inodedep) == 0) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
panic("handle_written_inodeblock: live inodedep");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
add_to_worklist(filefree);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
return (0);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* If no outstanding dependencies, free it.
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (free_inodedep(inodedep) ||
(TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
TAILQ_FIRST(&inodedep->id_extupdt) == 0))
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (0);
return (hadchanges);
}
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
/*
* Process a diradd entry after its dependent inode has been written.
* This routine must be called with splbio interrupts blocked.
*/
static void
diradd_inode_written(dap, inodedep)
struct diradd *dap;
struct inodedep *inodedep;
{
struct pagedep *pagedep;
dap->da_state |= COMPLETE;
if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
if (dap->da_state & DIRCHG)
pagedep = dap->da_previous->dm_pagedep;
else
pagedep = dap->da_pagedep;
LIST_REMOVE(dap, da_pdlist);
LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
}
WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Handle the completion of a mkdir dependency.
*/
static void
handle_written_mkdir(mkdir, type)
struct mkdir *mkdir;
int type;
{
struct diradd *dap;
struct pagedep *pagedep;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (mkdir->md_state != type) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_mkdir: bad type");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dap = mkdir->md_diradd;
dap->da_state &= ~type;
if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
dap->da_state |= DEPCOMPLETE;
if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
if (dap->da_state & DIRCHG)
pagedep = dap->da_previous->dm_pagedep;
else
pagedep = dap->da_pagedep;
LIST_REMOVE(dap, da_pdlist);
LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
}
LIST_REMOVE(mkdir, md_mkdirs);
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
WORKITEM_FREE(mkdir, D_MKDIR);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Called from within softdep_disk_write_complete above.
* A write operation was just completed. Removed inodes can
* now be freed and associated block pointers may be committed.
* Note that this routine is always called from interrupt level
* with further splbio interrupts blocked.
*/
static int
handle_written_filepage(pagedep, bp)
struct pagedep *pagedep;
struct buf *bp; /* buffer containing the written page */
{
struct dirrem *dirrem;
struct diradd *dap, *nextdap;
struct direct *ep;
int i, chgs;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((pagedep->pd_state & IOSTARTED) == 0) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_filepage: not started");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
pagedep->pd_state &= ~IOSTARTED;
/*
* Process any directory removals that have been committed.
*/
while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
LIST_REMOVE(dirrem, dm_next);
dirrem->dm_dirinum = pagedep->pd_ino;
add_to_worklist(&dirrem->dm_list);
}
/*
* Free any directory additions that have been committed.
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
* If it is a newly allocated block, we have to wait until
* the on-disk directory inode claims the new block.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if ((pagedep->pd_state & NEWBLOCK) == 0)
while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
free_diradd(dap);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Uncommitted directory entries must be restored.
*/
for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
dap = nextdap) {
nextdap = LIST_NEXT(dap, da_pdlist);
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (dap->da_state & ATTACHED) {
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
lk.lkt_held = NOHOLDER;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("handle_written_filepage: attached");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ep = (struct direct *)
((char *)bp->b_data + dap->da_offset);
ep->d_ino = dap->da_newinum;
dap->da_state &= ~UNDONE;
dap->da_state |= ATTACHED;
chgs = 1;
/*
* If the inode referenced by the directory has
* been written out, then the dependency can be
* moved to the pending list.
*/
if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
LIST_REMOVE(dap, da_pdlist);
LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
da_pdlist);
}
}
}
/*
* If there were any rollbacks in the directory, then it must be
* marked dirty so that its will eventually get written back in
* its correct form.
*/
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
if (chgs) {
if ((bp->b_flags & B_DELWRI) == 0)
stat_dir_entry++;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bdirty(bp);
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
return (1);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
Occationally deleted files would hang around for hours or days without being reclaimed. This bug was introduced in revision 1.95 dealing with filenames placed in newly allocated directory blocks, thus is not present in 4.X systems. The bug is triggered when a new entry is made in a directory after the data block containing the original new entry has been written, but before the inode that references the data block has been written. Submitted by: Bill Fenner <fenner@research.att.com>
2002-02-07 00:54:32 +00:00
* If we are not waiting for a new directory block to be
* claimed by its inode, then the pagedep will be freed.
* Otherwise it will remain to track any new entries on
* the page in case they are fsync'ed.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
Occationally deleted files would hang around for hours or days without being reclaimed. This bug was introduced in revision 1.95 dealing with filenames placed in newly allocated directory blocks, thus is not present in 4.X systems. The bug is triggered when a new entry is made in a directory after the data block containing the original new entry has been written, but before the inode that references the data block has been written. Submitted by: Bill Fenner <fenner@research.att.com>
2002-02-07 00:54:32 +00:00
if ((pagedep->pd_state & NEWBLOCK) == 0) {
Build on the change in revision 1.98 by Tor.Egge@fast.no. The symptom being treated in 1.98 was to avoid freeing a pagedep dependency if there was still a newdirblk dependency referencing it. That change is correct and no longer prints a warning message when it occurs. The other part of revision 1.98 was to panic when a newdirblk dependency was encountered during a file truncation. This fix removes that panic and replaces it with code to find and delete the newdirblk dependency so that the truncation can succeed.
2001-06-13 23:13:13 +00:00
LIST_REMOVE(pagedep, pd_hash);
WORKITEM_FREE(pagedep, D_PAGEDEP);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
return (0);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Writing back in-core inode structures.
*
More s/file system/filesystem/g
2002-05-16 21:28:32 +00:00
* The filesystem only accesses an inode's contents when it occupies an
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
* "in-core" inode structure. These "in-core" structures are separate from
* the page frames used to cache inode blocks. Only the latter are
* transferred to/from the disk. So, when the updated contents of the
* "in-core" inode structure are copied to the corresponding in-memory inode
* block, the dependencies are also transferred. The following procedure is
* called when copying a dirty "in-core" inode to a cached inode block.
*/
/*
* Called when an inode is loaded from disk. If the effective link count
* differed from the actual link count when it was last flushed, then we
* need to ensure that the correct effective link count is put back.
*/
void
softdep_load_inodeblock(ip)
struct inode *ip; /* the "in_core" copy of the inode */
{
struct inodedep *inodedep;
/*
* Check for alternate nlink count.
*/
ip->i_effnlink = ip->i_nlink;
ACQUIRE_LOCK(&lk);
if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
FREE_LOCK(&lk);
return;
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
ip->i_effnlink -= inodedep->id_nlinkdelta;
When running with soft updates, track the number of blocks and files that are committed to being freed and reflect these blocks in the counts returned by statfs (and thus also by the `df' command). This change allows programs such as those that do news expiration to know when to stop if they are trying to create a certain percentage of free space. Note that this change does not solve the much harder problem of making this to-be-freed space available to applications that want it (thus on a nearly full filesystem, you may still encounter out-of-space conditions even though the free space will show up eventually). Hopefully this harder problem will be the subject of a future enhancement.
2001-05-08 07:42:20 +00:00
if (inodedep->id_state & SPACECOUNTED)
ip->i_flag |= IN_SPACECOUNTED;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
}
/*
* This routine is called just before the "in-core" inode
* information is to be copied to the in-memory inode block.
* Recall that an inode block contains several inodes. If
* the force flag is set, then the dependencies will be
* cleared so that the update can always be made. Note that
* the buffer is locked when this routine is called, so we
* will never be in the middle of writing the inode block
* to disk.
*/
void
softdep_update_inodeblock(ip, bp, waitfor)
struct inode *ip; /* the "in_core" copy of the inode */
struct buf *bp; /* the buffer containing the inode block */
UFS_UPDATE() takes a boolean `waitfor' arg, so don't pass it the value MNT_WAIT when we mean boolean `true' or check for that value not being passed. There was no problem in practice because MNT_WAIT had the magic value of 1.
1999-01-06 18:18:06 +00:00
int waitfor; /* nonzero => update must be allowed */
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct inodedep *inodedep;
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
struct worklist *wk;
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
struct buf *ibp;
int error;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* If the effective link count is not equal to the actual link
* count, then we must track the difference in an inodedep while
* the inode is (potentially) tossed out of the cache. Otherwise,
* if there is no existing inodedep, then there are no dependencies
* to track.
*/
ACQUIRE_LOCK(&lk);
Better bounding on softdep_flushfiles; other minor tweeks to checks.
2000-01-17 06:35:11 +00:00
if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Better bounding on softdep_flushfiles; other minor tweeks to checks.
2000-01-17 06:35:11 +00:00
if (ip->i_effnlink != ip->i_nlink)
panic("softdep_update_inodeblock: bad link count");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return;
}
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_update_inodeblock: bad delta");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Changes have been initiated. Anything depending on these
* changes cannot occur until this inode has been written.
*/
inodedep->id_state &= ~COMPLETE;
if ((inodedep->id_state & ONWORKLIST) == 0)
WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
/*
* Any new dependencies associated with the incore inode must
* now be moved to the list associated with the buffer holding
* the in-memory copy of the inode. Once merged process any
* allocdirects that are completed by the merger.
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL)
handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
/*
* Now that the inode has been pushed into the buffer, the
* operations dependent on the inode being written to disk
* can be moved to the id_bufwait so that they will be
* processed when the buffer I/O completes.
*/
while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
WORKLIST_REMOVE(wk);
WORKLIST_INSERT(&inodedep->id_bufwait, wk);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Newly allocated inodes cannot be written until the bitmap
* that allocates them have been written (indicated by
* DEPCOMPLETE being set in id_state). If we are doing a
* forced sync (e.g., an fsync on a file), we force the bitmap
* to be written so that the update can be done.
*/
if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
FREE_LOCK(&lk);
return;
}
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
ibp = inodedep->id_buf;
ibp = getdirtybuf(&ibp, NULL, MNT_WAIT);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
if (ibp && (error = bwrite(ibp)) != 0)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error("softdep_update_inodeblock: bwrite", error);
if ((inodedep->id_state & DEPCOMPLETE) == 0)
panic("softdep_update_inodeblock: update failed");
}
/*
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
* Merge the a new inode dependency list (such as id_newinoupdt) into an
* old inode dependency list (such as id_inoupdt). This routine must be
* called with splbio interrupts blocked.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
static void
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
merge_inode_lists(newlisthead, oldlisthead)
struct allocdirectlst *newlisthead;
struct allocdirectlst *oldlisthead;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
struct allocdirect *listadp, *newadp;
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
newadp = TAILQ_FIRST(newlisthead);
for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (listadp->ad_lbn < newadp->ad_lbn) {
listadp = TAILQ_NEXT(listadp, ad_next);
continue;
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
TAILQ_REMOVE(newlisthead, newadp, ad_next);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
if (listadp->ad_lbn == newadp->ad_lbn) {
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
allocdirect_merge(oldlisthead, newadp,
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
listadp);
listadp = newadp;
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
newadp = TAILQ_FIRST(newlisthead);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
TAILQ_REMOVE(newlisthead, newadp, ad_next);
TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
}
/*
* If we are doing an fsync, then we must ensure that any directory
* entries for the inode have been written after the inode gets to disk.
*/
ARGH! I have too many source trees :-( Fix prototype errors in last commit.
2000-06-16 13:00:33 +00:00
int
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_fsync(vp)
struct vnode *vp; /* the "in_core" copy of the inode */
{
struct inodedep *inodedep;
struct pagedep *pagedep;
struct worklist *wk;
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
struct diradd *dap;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct mount *mnt;
struct vnode *pvp;
struct inode *ip;
struct buf *bp;
struct fs *fs;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td = curthread;
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
int error, flushparent;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ino_t parentino;
ufs_lbn_t lbn;
ip = VTOI(vp);
fs = ip->i_fs;
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
ACQUIRE_LOCK(&lk);
if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
FREE_LOCK(&lk);
return (0);
}
if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
LIST_FIRST(&inodedep->id_bufwait) != NULL ||
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
FREE_LOCK(&lk);
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
panic("softdep_fsync: pending ops");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
for (error = 0, flushparent = 0; ; ) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
break;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (wk->wk_type != D_DIRADD) {
FREE_LOCK(&lk);
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
panic("softdep_fsync: Unexpected type %s",
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
TYPENAME(wk->wk_type));
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
dap = WK_DIRADD(wk);
/*
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
* Flush our parent if this directory entry has a MKDIR_PARENT
* dependency or is contained in a newly allocated block.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
if (dap->da_state & DIRCHG)
pagedep = dap->da_previous->dm_pagedep;
else
pagedep = dap->da_pagedep;
mnt = pagedep->pd_mnt;
parentino = pagedep->pd_ino;
lbn = pagedep->pd_lbn;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_fsync: dirty");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if ((dap->da_state & MKDIR_PARENT) ||
(pagedep->pd_state & NEWBLOCK))
flushparent = 1;
else
flushparent = 0;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* If we are being fsync'ed as part of vgone'ing this vnode,
* then we will not be able to release and recover the
* vnode below, so we just have to give up on writing its
* directory entry out. It will eventually be written, just
* not now, but then the user was not asking to have it
* written, so we are not breaking any promises.
*/
- Remove a mp_fixme() and some locks that weren't necessary. I now understand how this works.
2003-10-04 11:06:43 +00:00
if (vp->v_iflag & VI_XLOCK)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
break;
/*
* We prevent deadlock by always fetching inodes from the
* root, moving down the directory tree. Thus, when fetching
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
* our parent directory, we first try to get the lock. If
* that fails, we must unlock ourselves before requesting
* the lock on our parent. See the comment in ufs_lookup
* for details on possible races.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
FREE_LOCK(&lk);
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if (VFS_VGET(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
VOP_UNLOCK(vp, 0, td);
error = VFS_VGET(mnt, parentino, LK_EXCLUSIVE, &pvp);
vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
if (error != 0)
return (error);
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
/*
* All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
* that are contained in direct blocks will be resolved by
* doing a UFS_UPDATE. Pagedeps contained in indirect blocks
* may require a complete sync'ing of the directory. So, we
* try the cheap and fast UFS_UPDATE first, and if that fails,
* then we do the slower VOP_FSYNC of the directory.
*/
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (flushparent) {
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
if ((error = UFS_UPDATE(pvp, 1)) != 0) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
vput(pvp);
return (error);
}
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
if ((pagedep->pd_state & NEWBLOCK) &&
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
(error = VOP_FSYNC(pvp, td->td_ucred, MNT_WAIT, td))) {
When a new block is allocated to a directory, an fsync of a file whose name is within that block must ensure not only that the block containing the file name has been written, but also that the on-disk directory inode references that block. When a new directory block is created, we allocate a newdirblk structure which is linked to the associated allocdirect (on its ad_newdirblk list). When the allocdirect has been satisfied, the newdirblk structure is moved to the inodedep id_bufwait list of its directory to await the inode being written. When the inode is written, the directory entries are fully committed and can be deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd lists.
2001-05-17 07:24:03 +00:00
vput(pvp);
return (error);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Flush directory page containing the inode's name.
*/
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
&bp);
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
if (error == 0)
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
error = bwrite(bp);
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
else
brelse(bp);
Merge in Kirk's changes to stop softupdates from hogging all of memory.
1998-05-19 21:45:53 +00:00
vput(pvp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (error != 0)
return (error);
Reorganize softdep_fsync so that it only does the inode-is-flushed check before the inode is unlocked while grabbing its parent directory. Once it is unlocked, other operations may slip in that could make the inode-is-flushed check fail. Allowing other writes to the inode before returning from fsync does not break the semantics of fsync since we have flushed everything that was dirty at the time of the fsync call.
2000-01-09 23:14:57 +00:00
ACQUIRE_LOCK(&lk);
if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
break;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
FREE_LOCK(&lk);
return (0);
}
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
/*
* Flush all the dirty bitmaps associated with the block device
* before flushing the rest of the dirty blocks so as to reduce
* the number of dependencies that will have to be rolled back.
*/
void
softdep_fsync_mountdev(vp)
struct vnode *vp;
{
struct buf *bp, *nbp;
struct worklist *wk;
Give vn_isdisk() a second argument where it can return a suitable errno. Suggested by: bde
2000-01-10 12:04:27 +00:00
if (!vn_isdisk(vp, NULL))
Convert various pieces of code to use vn_isdisk() rather than checking for vp->v_type == VBLK. In ccd: we don't need to call VOP_GETATTR to find the type of a vnode. Reviewed by: sos
1999-11-22 10:33:55 +00:00
panic("softdep_fsync_mountdev: vnode not a disk");
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
ACQUIRE_LOCK(&lk);
- Convert locks to use standard macros. - Lock access to the buflists. - Document broken locking. - Use vrefcnt().
2002-09-25 02:49:48 +00:00
VI_LOCK(vp);
Loose the v_dirty* and v_clean* alias macros. Check the count field where we just want to know the full/empty state, rather than using TAILQ_EMPTY() or TAILQ_FIRST().
2004-10-25 09:14:03 +00:00
TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
/*
* If it is already scheduled, skip to the next buffer.
*/
- Move BX_BKGRDWAIT and BX_BKGRDINPROG to BV_ and the b_vflags field. - Surround all accesses of the BKGRD{WAIT,INPROG} flags with the vnode interlock. - Don't use the B_LOCKED flag and QUEUE_LOCKED for background write buffers. Check for the BKGRDINPROG flag before recycling or throwing away a buffer. We do this instead because it is not safe for us to move the original buffer to a new queue from the callback on the background write buffer. - Remove the B_LOCKED flag and the locked buffer queue. They are no longer used. - The vnode interlock is used around checks for BKGRDINPROG where it may not be strictly necessary. If we hold the buf lock the a back-ground write will not be started without our knowledge, one may only be completed while we're not looking. Rather than remove the code, Document two of the places where this extra locking is done. A pass should be done to verify and minimize the locking later.
2003-08-28 06:55:18 +00:00
if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
continue;
- Move BX_BKGRDWAIT and BX_BKGRDINPROG to BV_ and the b_vflags field. - Surround all accesses of the BKGRD{WAIT,INPROG} flags with the vnode interlock. - Don't use the B_LOCKED flag and QUEUE_LOCKED for background write buffers. Check for the BKGRDINPROG flag before recycling or throwing away a buffer. We do this instead because it is not safe for us to move the original buffer to a new queue from the callback on the background write buffer. - Remove the B_LOCKED flag and the locked buffer queue. They are no longer used. - The vnode interlock is used around checks for BKGRDINPROG where it may not be strictly necessary. If we hold the buf lock the a back-ground write will not be started without our knowledge, one may only be completed while we're not looking. Rather than remove the code, Document two of the places where this extra locking is done. A pass should be done to verify and minimize the locking later.
2003-08-28 06:55:18 +00:00
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if ((bp->b_flags & B_DELWRI) == 0) {
FREE_LOCK(&lk);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
panic("softdep_fsync_mountdev: not dirty");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
/*
* We are only interested in bitmaps with outstanding
* dependencies.
*/
if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
When writing out bitmap buffers, need to skip over ones that already have a write in progress. Otherwise one can get in an infinite loop trying to get them all flushed. Submitted by: Matthew Dillon <dillon@apollo.backplane.com>
2000-01-30 20:32:59 +00:00
wk->wk_type != D_BMSAFEMAP ||
- Move BX_BKGRDWAIT and BX_BKGRDINPROG to BV_ and the b_vflags field. - Surround all accesses of the BKGRD{WAIT,INPROG} flags with the vnode interlock. - Don't use the B_LOCKED flag and QUEUE_LOCKED for background write buffers. Check for the BKGRDINPROG flag before recycling or throwing away a buffer. We do this instead because it is not safe for us to move the original buffer to a new queue from the callback on the background write buffer. - Remove the B_LOCKED flag and the locked buffer queue. They are no longer used. - The vnode interlock is used around checks for BKGRDINPROG where it may not be strictly necessary. If we hold the buf lock the a back-ground write will not be started without our knowledge, one may only be completed while we're not looking. Rather than remove the code, Document two of the places where this extra locking is done. A pass should be done to verify and minimize the locking later.
2003-08-28 06:55:18 +00:00
(bp->b_vflags & BV_BKGRDINPROG)) {
Convert buffer locking from using the B_BUSY and B_WANTED flags to using lockmgr locks. This commit should be functionally equivalent to the old semantics. That is, all buffer locking is done with LK_EXCLUSIVE requests. Changes to take advantage of LK_SHARED and LK_RECURSIVE will be done in future commits.
1999-06-26 02:47:16 +00:00
BUF_UNLOCK(bp);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
continue;
Convert buffer locking from using the B_BUSY and B_WANTED flags to using lockmgr locks. This commit should be functionally equivalent to the old semantics. That is, all buffer locking is done with LK_EXCLUSIVE requests. Changes to take advantage of LK_SHARED and LK_RECURSIVE will be done in future commits.
1999-06-26 02:47:16 +00:00
}
- Move BX_BKGRDWAIT and BX_BKGRDINPROG to BV_ and the b_vflags field. - Surround all accesses of the BKGRD{WAIT,INPROG} flags with the vnode interlock. - Don't use the B_LOCKED flag and QUEUE_LOCKED for background write buffers. Check for the BKGRDINPROG flag before recycling or throwing away a buffer. We do this instead because it is not safe for us to move the original buffer to a new queue from the callback on the background write buffer. - Remove the B_LOCKED flag and the locked buffer queue. They are no longer used. - The vnode interlock is used around checks for BKGRDINPROG where it may not be strictly necessary. If we hold the buf lock the a back-ground write will not be started without our knowledge, one may only be completed while we're not looking. Rather than remove the code, Document two of the places where this extra locking is done. A pass should be done to verify and minimize the locking later.
2003-08-28 06:55:18 +00:00
VI_UNLOCK(vp);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
bremfree(bp);
FREE_LOCK(&lk);
(void) bawrite(bp);
ACQUIRE_LOCK(&lk);
/*
* Since we may have slept during the I/O, we need
* to start from a known point.
*/
- Convert locks to use standard macros. - Lock access to the buflists. - Document broken locking. - Use vrefcnt().
2002-09-25 02:49:48 +00:00
VI_LOCK(vp);
Loose the v_dirty* and v_clean* alias macros. Check the count field where we just want to know the full/empty state, rather than using TAILQ_EMPTY() or TAILQ_FIRST().
2004-10-25 09:14:03 +00:00
nbp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
}
drain_output(vp, 1);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_UNLOCK(vp);
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
FREE_LOCK(&lk);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* This routine is called when we are trying to synchronously flush a
* file. This routine must eliminate any filesystem metadata dependencies
* so that the syncing routine can succeed by pushing the dirty blocks
* associated with the file. If any I/O errors occur, they are returned.
*/
int
softdep_sync_metadata(ap)
struct vop_fsync_args /* {
struct vnode *a_vp;
struct ucred *a_cred;
int a_waitfor;
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *a_td;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
} */ *ap;
{
struct vnode *vp = ap->a_vp;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
struct pagedep *pagedep;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct allocdirect *adp;
struct allocindir *aip;
struct buf *bp, *nbp;
struct worklist *wk;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
int i, error, waitfor;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Check whether this vnode is involved in a filesystem
* that is doing soft dependency processing.
*/
Give vn_isdisk() a second argument where it can return a suitable errno. Suggested by: bde
2000-01-10 12:04:27 +00:00
if (!vn_isdisk(vp, NULL)) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (!DOINGSOFTDEP(vp))
return (0);
} else
Blow away the v_specmountpoint define, replacing it with what it was defined as (rdev->si_mountpoint)
2000-10-09 17:31:39 +00:00
if (vp->v_rdev->si_mountpoint == NULL ||
(vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (0);
/*
* Ensure that any direct block dependencies have been cleared.
*/
ACQUIRE_LOCK(&lk);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
return (error);
}
/*
* For most files, the only metadata dependencies are the
* cylinder group maps that allocate their inode or blocks.
* The block allocation dependencies can be found by traversing
* the dependency lists for any buffers that remain on their
* dirty buffer list. The inode allocation dependency will
* be resolved when the inode is updated with MNT_WAIT.
* This work is done in two passes. The first pass grabs most
* of the buffers and begins asynchronously writing them. The
* only way to wait for these asynchronous writes is to sleep
* on the filesystem vnode which may stay busy for a long time
* if the filesystem is active. So, instead, we make a second
* pass over the dependencies blocking on each write. In the
* usual case we will be blocking against a write that we
* initiated, so when it is done the dependency will have been
* resolved. Thus the second pass is expected to end quickly.
*/
waitfor = MNT_NOWAIT;
top:
Must call drain_output() before checking the dirty block list in softdep_sync_metadata(). Otherwise we may miss dependencies that need to be flushed which will result in a later panic with the message ``vinvalbuf: dirty bufs''. Submitted by: Matthew Dillon <dillon@apollo.backplane.com> MFC after: 1 week
2002-01-11 19:59:27 +00:00
/*
* We must wait for any I/O in progress to finish so that
* all potential buffers on the dirty list will be visible.
*/
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Must call drain_output() before checking the dirty block list in softdep_sync_metadata(). Otherwise we may miss dependencies that need to be flushed which will result in a later panic with the message ``vinvalbuf: dirty bufs''. Submitted by: Matthew Dillon <dillon@apollo.backplane.com> MFC after: 1 week
2002-01-11 19:59:27 +00:00
drain_output(vp, 1);
Loose the v_dirty* and v_clean* alias macros. Check the count field where we just want to know the full/empty state, rather than using TAILQ_EMPTY() or TAILQ_FIRST().
2004-10-25 09:14:03 +00:00
bp = getdirtybuf(&TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd),
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_MTX(vp), MNT_WAIT);
if (bp == NULL) {
VI_UNLOCK(vp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
return (0);
}
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
/* While syncing snapshots, we must allow recursive lookups */
bp->b_lock.lk_flags |= LK_CANRECURSE;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
loop:
/*
* As we hold the buffer locked, none of its dependencies
* will disappear.
*/
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(wk, &bp->b_dep, wk_list) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
switch (wk->wk_type) {
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_ALLOCDIRECT:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
adp = WK_ALLOCDIRECT(wk);
if (adp->ad_state & DEPCOMPLETE)
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
nbp = adp->ad_buf;
nbp = getdirtybuf(&nbp, NULL, waitfor);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (nbp == NULL)
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
if (waitfor == MNT_NOWAIT) {
bawrite(nbp);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
} else if ((error = bwrite(nbp)) != 0) {
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
ACQUIRE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_ALLOCINDIR:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
aip = WK_ALLOCINDIR(wk);
if (aip->ai_state & DEPCOMPLETE)
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
nbp = aip->ai_buf;
nbp = getdirtybuf(&nbp, NULL, waitfor);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (nbp == NULL)
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
if (waitfor == MNT_NOWAIT) {
bawrite(nbp);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
} else if ((error = bwrite(nbp)) != 0) {
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
ACQUIRE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INDIRDEP:
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
restart:
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (aip->ai_state & DEPCOMPLETE)
continue;
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
nbp = aip->ai_buf;
nbp = getdirtybuf(&nbp, NULL, MNT_WAIT);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (nbp == NULL)
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
goto restart;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
if ((error = bwrite(nbp)) != 0) {
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
ACQUIRE_LOCK(&lk);
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
goto restart;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_INODEDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
WK_INODEDEP(wk)->id_ino)) != 0) {
FREE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
First published FreeBSD version of soft updates Feb 5.
1998-05-19 20:18:42 +00:00
case D_PAGEDEP:
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* We are trying to sync a directory that may
* have dependencies on both its own metadata
* and/or dependencies on the inodes of any
* recently allocated files. We walk its diradd
* lists pushing out the associated inode.
*/
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
pagedep = WK_PAGEDEP(wk);
for (i = 0; i < DAHASHSZ; i++) {
if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
continue;
Cosmetic changes to make it compile without errors in gcc -Wall
1999-05-22 04:43:04 +00:00
if ((error =
flush_pagedep_deps(vp, pagedep->pd_mnt,
&pagedep->pd_diraddhd[i]))) {
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
FREE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
case D_MKDIR:
/*
* This case should never happen if the vnode has
* been properly sync'ed. However, if this function
* is used at a place where the vnode has not yet
* been sync'ed, this dependency can show up. So,
* rather than panic, just flush it.
*/
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
nbp = WK_MKDIR(wk)->md_buf;
nbp = getdirtybuf(&nbp, NULL, waitfor);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (nbp == NULL)
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
FREE_LOCK(&lk);
if (waitfor == MNT_NOWAIT) {
bawrite(nbp);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
} else if ((error = bwrite(nbp)) != 0) {
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
}
ACQUIRE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
case D_BMSAFEMAP:
/*
* This case should never happen if the vnode has
* been properly sync'ed. However, if this function
* is used at a place where the vnode has not yet
* been sync'ed, this dependency can show up. So,
* rather than panic, just flush it.
*/
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
nbp = WK_BMSAFEMAP(wk)->sm_buf;
nbp = getdirtybuf(&nbp, NULL, waitfor);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (nbp == NULL)
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
FREE_LOCK(&lk);
if (waitfor == MNT_NOWAIT) {
bawrite(nbp);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
} else if ((error = bwrite(nbp)) != 0) {
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
break;
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
}
ACQUIRE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
continue;
Ensure that softdep_sync_metadata can handle bmsafemap and mkdir entries if they ever arise (which should not happen as softdep_sync_metadata is currently used).
1999-03-02 00:19:47 +00:00
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
default:
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
panic("softdep_sync_metadata: Unknown type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
}
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
/* We reach here only in error and unlocked */
if (error == 0)
panic("softdep_sync_metadata: zero error");
bp->b_lock.lk_flags &= ~LK_CANRECURSE;
bawrite(bp);
return (error);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Move the VI_BWAIT flag into no bo_flag element of bufobj and call it BO_WWAIT Add bufobj_wref(), bufobj_wdrop() and bufobj_wwait() to handle the write count on a bufobj. Bufobj_wdrop() replaces vwakeup(). Use these functions all relevant places except in ffs_softdep.c where the use if interlocked_sleep() makes this impossible. Rename b_vnbufs to b_bobufs now that we touch all the relevant files anyway.
2004-10-21 15:53:54 +00:00
nbp = getdirtybuf(&TAILQ_NEXT(bp, b_bobufs), VI_MTX(vp), MNT_WAIT);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (nbp == NULL)
VI_UNLOCK(vp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
When syncing out snapshot metadata, we must temporarily allow recursive buffer locking so as to avoid locking against ourselves if we need to write filesystem metadata.
2001-05-08 07:13:00 +00:00
bp->b_lock.lk_flags &= ~LK_CANRECURSE;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bawrite(bp);
ACQUIRE_LOCK(&lk);
if (nbp != NULL) {
bp = nbp;
goto loop;
}
/*
* The brief unlock is to allow any pent up dependency
Must call drain_output() before checking the dirty block list in softdep_sync_metadata(). Otherwise we may miss dependencies that need to be flushed which will result in a later panic with the message ``vinvalbuf: dirty bufs''. Submitted by: Matthew Dillon <dillon@apollo.backplane.com> MFC after: 1 week
2002-01-11 19:59:27 +00:00
* processing to be done. Then proceed with the second pass.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
if (waitfor == MNT_NOWAIT) {
waitfor = MNT_WAIT;
FREE_LOCK(&lk);
ACQUIRE_LOCK(&lk);
goto top;
}
/*
* If we have managed to get rid of all the dirty buffers,
* then we are done. For certain directories and block
* devices, we may need to do further work.
Must call drain_output() before checking the dirty block list in softdep_sync_metadata(). Otherwise we may miss dependencies that need to be flushed which will result in a later panic with the message ``vinvalbuf: dirty bufs''. Submitted by: Matthew Dillon <dillon@apollo.backplane.com> MFC after: 1 week
2002-01-11 19:59:27 +00:00
*
* We must wait for any I/O in progress to finish so that
* all potential buffers on the dirty list will be visible.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Must call drain_output() before checking the dirty block list in softdep_sync_metadata(). Otherwise we may miss dependencies that need to be flushed which will result in a later panic with the message ``vinvalbuf: dirty bufs''. Submitted by: Matthew Dillon <dillon@apollo.backplane.com> MFC after: 1 week
2002-01-11 19:59:27 +00:00
drain_output(vp, 1);
Loose the v_dirty* and v_clean* alias macros. Check the count field where we just want to know the full/empty state, rather than using TAILQ_EMPTY() or TAILQ_FIRST().
2004-10-25 09:14:03 +00:00
if (vp->v_bufobj.bo_dirty.bv_cnt == 0) {
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_UNLOCK(vp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
return (0);
}
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_UNLOCK(vp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
/*
* If we are trying to sync a block device, some of its buffers may
* contain metadata that cannot be written until the contents of some
* partially written files have been written to disk. The only easy
* way to accomplish this is to sync the entire filesystem (luckily
* this happens rarely).
*/
Give vn_isdisk() a second argument where it can return a suitable errno. Suggested by: bde
2000-01-10 12:04:27 +00:00
if (vn_isdisk(vp, NULL) &&
Blow away the v_specmountpoint define, replacing it with what it was defined as (rdev->si_mountpoint)
2000-10-09 17:31:39 +00:00
vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
(error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred,
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
ap->a_td)) != 0)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
return (error);
return (0);
}
/*
* Flush the dependencies associated with an inodedep.
* Called with splbio blocked.
*/
static int
flush_inodedep_deps(fs, ino)
struct fs *fs;
ino_t ino;
{
struct inodedep *inodedep;
int error, waitfor;
/*
* This work is done in two passes. The first pass grabs most
* of the buffers and begins asynchronously writing them. The
* only way to wait for these asynchronous writes is to sleep
* on the filesystem vnode which may stay busy for a long time
* if the filesystem is active. So, instead, we make a second
* pass over the dependencies blocking on each write. In the
* usual case we will be blocking against a write that we
* initiated, so when it is done the dependency will have been
* resolved. Thus the second pass is expected to end quickly.
* We give a brief window at the top of the loop to allow
* any pending I/O to complete.
*/
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
for (error = 0, waitfor = MNT_NOWAIT; ; ) {
if (error)
return (error);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
ACQUIRE_LOCK(&lk);
if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
return (0);
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
continue;
/*
* If pass2, we are done, otherwise do pass 2.
*/
if (waitfor == MNT_WAIT)
break;
waitfor = MNT_WAIT;
}
/*
* Try freeing inodedep in case all dependencies have been removed.
*/
if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
(void) free_inodedep(inodedep);
return (0);
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
/*
* Flush an inode dependency list.
* Called with splbio blocked.
*/
static int
flush_deplist(listhead, waitfor, errorp)
struct allocdirectlst *listhead;
int waitfor;
int *errorp;
{
struct allocdirect *adp;
struct buf *bp;
TAILQ_FOREACH(adp, listhead, ad_next) {
if (adp->ad_state & DEPCOMPLETE)
continue;
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
bp = adp->ad_buf;
bp = getdirtybuf(&bp, NULL, waitfor);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (bp == NULL) {
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (waitfor == MNT_NOWAIT)
continue;
return (1);
}
FREE_LOCK(&lk);
if (waitfor == MNT_NOWAIT) {
bawrite(bp);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
} else if ((*errorp = bwrite(bp)) != 0) {
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
ACQUIRE_LOCK(&lk);
return (1);
}
ACQUIRE_LOCK(&lk);
return (1);
}
return (0);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Eliminate a pagedep dependency by flushing out all its diradd dependencies.
* Called with splbio blocked.
*/
static int
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
flush_pagedep_deps(pvp, mp, diraddhdp)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct vnode *pvp;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
struct mount *mp;
struct diraddhd *diraddhdp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td = curthread;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
struct inodedep *inodedep;
struct ufsmount *ump;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct diradd *dap;
struct vnode *vp;
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
int error = 0;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
struct buf *bp;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
ino_t inum;
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
ump = VFSTOUFS(mp);
while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
/*
* Flush ourselves if this directory entry
* has a MKDIR_PARENT dependency.
*/
if (dap->da_state & MKDIR_PARENT) {
FREE_LOCK(&lk);
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
if ((error = UFS_UPDATE(pvp, 1)) != 0)
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
break;
ACQUIRE_LOCK(&lk);
/*
* If that cleared dependencies, go on to next.
*/
if (dap != LIST_FIRST(diraddhdp))
continue;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (dap->da_state & MKDIR_PARENT) {
FREE_LOCK(&lk);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
panic("flush_pagedep_deps: MKDIR_PARENT");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
}
/*
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
* A newly allocated directory must have its "." and
* ".." entries written out before its name can be
* committed in its parent. We do not want or need
* the full semantics of a synchronous VOP_FSYNC as
* that may end up here again, once for each directory
* level in the filesystem. Instead, we push the blocks
* and wait for them to clear. We have to fsync twice
* because the first call may choose to defer blocks
* that still have dependencies, but deferral will
* happen at most once.
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
*/
inum = dap->da_newinum;
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
if (dap->da_state & MKDIR_BODY) {
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
FREE_LOCK(&lk);
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if ((error = VFS_VGET(mp, inum, LK_EXCLUSIVE, &vp)))
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
break;
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
if ((error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)) ||
(error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) {
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
vput(vp);
break;
}
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
drain_output(vp, 0);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_UNLOCK(vp);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
vput(vp);
ACQUIRE_LOCK(&lk);
/*
* If that cleared dependencies, go on to next.
*/
if (dap != LIST_FIRST(diraddhdp))
continue;
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (dap->da_state & MKDIR_BODY) {
FREE_LOCK(&lk);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
panic("flush_pagedep_deps: MKDIR_BODY");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
/*
* Flush the inode on which the directory entry depends.
* Having accounted for MKDIR_PARENT and MKDIR_BODY above,
* the only remaining dependency is that the updated inode
* count must get pushed to disk. The inode has already
* been pushed into its inode buffer (via VOP_UPDATE) at
* the time of the reference count change. So we need only
* locate that buffer, ensure that there will be no rollback
* caused by a bitmap dependency, then write the inode buffer.
*/
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
FREE_LOCK(&lk);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
panic("flush_pagedep_deps: lost inode");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
/*
* If the inode still has bitmap dependencies,
* push them to disk.
*/
if ((inodedep->id_state & DEPCOMPLETE) == 0) {
- Several of the callers to getdirtybuf() were erroneously changed to pass in a list head instead of a pointer to the first element at the time of the first call. These lists are subject to change, and getdirtybuf() would refetch from the wrong list in some cases. Spottedy by: tegge Pointy hat to: me
2003-09-03 04:08:15 +00:00
bp = inodedep->id_buf;
bp = getdirtybuf(&bp, NULL, MNT_WAIT);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
FREE_LOCK(&lk);
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
if (bp && (error = bwrite(bp)) != 0)
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
break;
ACQUIRE_LOCK(&lk);
if (dap != LIST_FIRST(diraddhdp))
continue;
}
/*
* If the inode is still sitting in a buffer waiting
* to be written, push it to disk.
*/
FREE_LOCK(&lk);
if ((error = bread(ump->um_devvp,
fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
(int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
brelse(bp);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
break;
When taking a snapshot, we must check for active files that have been unlinked (e.g., with a zero link count). We have to expunge all trace of these files from the snapshot so that they are neither reclaimed prematurely by fsck nor saved unnecessarily by dump.
2002-02-02 01:42:44 +00:00
}
Properly vector all bwrite() and BUF_WRITE() calls through the same path and s/BUF_WRITE()/bwrite()/ since it now does the same as bwrite().
2004-03-11 18:02:36 +00:00
if ((error = bwrite(bp)) != 0)
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
break;
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
ACQUIRE_LOCK(&lk);
Import the next version received from kirk after some FreeBSD feedback.
1998-05-19 20:03:29 +00:00
/*
* If we have failed to get rid of all the dependencies
* then something is seriously wrong.
*/
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (dap == LIST_FIRST(diraddhdp)) {
FREE_LOCK(&lk);
Need to reorganize the flushing of directory entry (pagedep) dependencies so that they never try to lock an inode corresponding to ".." as this can lead to deadlock. We observe that any inode with an updated link count is always pushed into its buffer at the time of the link count change, so we do not need to do a VOP_UPDATE, but merely find its buffer and write it. The only time we need to get the inode itself is from the result of a mkdir whose name will never be ".." and hence locking such an inode will never request a lock above us in the filesystem tree. Thanks to Brian Fundakowski Feldman for providing the test program that tickled soft updates into hanging in "inode" sleep. Submitted by: Brian Fundakowski Feldman <green@FreeBSD.org>
2000-01-18 01:30:03 +00:00
panic("flush_pagedep_deps: flush failed");
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
if (error)
ACQUIRE_LOCK(&lk);
return (error);
}
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* A large burst of file addition or deletion activity can drive the
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
* memory load excessively high. First attempt to slow things down
* using the techniques below. If that fails, this routine requests
* the offending operations to fall back to running synchronously
* until the memory load returns to a reasonable level.
*/
int
softdep_slowdown(vp)
struct vnode *vp;
{
int max_softdeps_hard;
max_softdeps_hard = max_softdeps * 11 / 10;
if (num_dirrem < max_softdeps_hard / 2 &&
This patch fixes a problem caused by applications that rapidly and repeatedly truncate the same file. Each time the file is truncated, a buffer is grabbed to store the indirect block numbers that need to be freed. Those blocks cannot be freed until the inode claiming them is written to disk. Thus, the number of buffers being held by soft updates explodes and in extreme cases can run the kernel out of buffers. The problem can be avoided by doing an fsync on the file every debug.maxindirdep truncates (currently defaulted to 50). The fsync causes the inode to be written so that the held buffers can be freed. The check for excessive buffers is checked as part of the existing hook for excessive dependencies (softdep_slowdown) in the truncate code. Reported by: David Schultz <dschultz@uclink.Berkeley.EDU> Sponsored by: DARPA & NAI Labs. MFC after: 3 weeks
2003-01-07 18:23:50 +00:00
num_inodedep < max_softdeps_hard &&
VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps)
return (0);
if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
speedup_syncer();
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
stat_sync_limit_hit += 1;
return (1);
}
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
/*
* Called by the allocation routines when they are about to fail
* in the hope that we can free up some disk space.
*
* First check to see if the work list has anything on it. If it has,
* clean up entries until we successfully free some space. Because this
* process holds inodes locked, we cannot handle any remove requests
* that might block on a locked inode as that could lead to deadlock.
* If the worklist yields no free space, encourage the syncer daemon
* to help us. In no event will we try for longer than tickdelay seconds.
*/
int
softdep_request_cleanup(fs, vp)
struct fs *fs;
struct vnode *vp;
{
This commit adds basic support for the UFS2 filesystem. The UFS2 filesystem expands the inode to 256 bytes to make space for 64-bit block pointers. It also adds a file-creation time field, an ability to use jumbo blocks per inode to allow extent like pointer density, and space for extended attributes (up to twice the filesystem block size worth of attributes, e.g., on a 16K filesystem, there is space for 32K of attributes). UFS2 fully supports and runs existing UFS1 filesystems. New filesystems built using newfs can be built in either UFS1 or UFS2 format using the -O option. In this commit UFS1 is the default format, so if you want to build UFS2 format filesystems, you must specify -O 2. This default will be changed to UFS2 when UFS2 proves itself to be stable. In this commit the boot code for reading UFS2 filesystems is not compiled (see /sys/boot/common/ufsread.c) as there is insufficient space in the boot block. Once the size of the boot block is increased, this code can be defined. Things to note: the definition of SBSIZE has changed to SBLOCKSIZE. The header file <ufs/ufs/dinode.h> must be included before <ufs/ffs/fs.h> so as to get the definitions of ufs2_daddr_t and ufs_lbn_t. Still TODO: Verify that the first level bootstraps work for all the architectures. Convert the utility ffsinfo to understand UFS2 and test growfs. Add support for the extended attribute storage. Update soft updates to ensure integrity of extended attribute storage. Switch the current extended attribute interfaces to use the extended attribute storage. Add the extent like functionality (framework is there, but is currently never used). Sponsored by: DARPA & NAI Labs. Reviewed by: Poul-Henning Kamp <phk@freebsd.org>
2002-06-21 06:18:05 +00:00
long starttime;
ufs2_daddr_t needed;
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
starttime = time_second + tickdelay;
We must be careful to avoid recursive copy-on-write faults when trying to clean up during disk-full senarios. Sponsored by: DARPA & NAI Labs.
2002-10-23 21:47:02 +00:00
/*
* If we are being called because of a process doing a
* copy-on-write, then it is not safe to update the vnode
* as we may recurse into the copy-on-write routine.
*/
Move the P_COWINPROGRESS flag from being a per-process p_flag to being a per-thread td_pflag which doesn't require any locks to read or write as it is only read or written by curthread on itself. Glanced at by: mckusick
2003-10-23 21:14:08 +00:00
if (!(curthread->td_pflags & TDP_COWINPROGRESS) &&
We must be careful to avoid recursive copy-on-write faults when trying to clean up during disk-full senarios. Sponsored by: DARPA & NAI Labs.
2002-10-23 21:47:02 +00:00
UFS_UPDATE(vp, 1) != 0)
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
return (0);
while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
if (time_second > starttime)
return (0);
if (num_on_worklist > 0 &&
process_worklist_item(NULL, LK_NOWAIT) != -1) {
stat_worklist_push += 1;
continue;
}
request_cleanup(FLUSH_REMOVE_WAIT, 0);
}
return (1);
}
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
/*
* If memory utilization has gotten too high, deliberately slow things
* down and speed up the I/O processing.
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
*/
static int
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
request_cleanup(resource, islocked)
int resource;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
int islocked;
{
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
struct thread *td = curthread;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* We never hold up the filesystem syncer process.
*/
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
if (td == filesys_syncer)
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
return (0);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
/*
* First check to see if the work list has gotten backlogged.
* If it has, co-opt this process to help clean up two entries.
* Because this process may hold inodes locked, we cannot
* handle any remove requests that might block on a locked
* inode as that could lead to deadlock.
*/
if (num_on_worklist > max_softdeps / 10) {
This patch corrects two problems with the rate limiting code that was introduced in revision 1.80. The problem manifested itself with a `locking against myself' panic and could also result in soft updates inconsistences associated with inodedeps. The two problems are: 1) One of the background operations could manipulate the bitmap while holding it locked with intent to create. This held lock results in a `locking against myself' panic, when the background processing that we have been coopted to do tries to lock the bitmap which we are already holding locked. To understand how to fix this problem, first, observe that we can do the background cleanups in inodedep_lookup only when allocating inodedeps (DEPALLOC is set in the call to inodedep_lookup). Second observe that calls to inodedep_lookup with DEPALLOC set can only happen from the following calls into the softdep code: softdep_setup_inomapdep softdep_setup_allocdirect softdep_setup_remove softdep_setup_freeblocks softdep_setup_directory_change softdep_setup_directory_add softdep_change_linkcnt Only the first two of these can come from ffs_alloc.c while holding a bitmap locked. Thus, inodedep_lookup must not go off to do request_cleanups when being called from these functions. This change adds a flag, NODELAY, that can be passed to inodedep_lookup to let it know that it should not do background processing in those cases. 2) The return value from request_cleanup when helping out with the cleanup was 0 instead of 1. This meant that despite the fact that we may have slept while doing the cleanups, the code did not recheck for the appearance of an inodedep (e.g., goto top in inodedep_lookup). This lead to the softdep inconsistency in which we ended up with two inodedep's for the same inode. Reviewed by: Peter Wemm <peter@yahoo-inc.com>, Matt Dillon <dillon@earth.backplane.com>
2001-02-20 11:14:38 +00:00
if (islocked)
FREE_LOCK(&lk);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
process_worklist_item(NULL, LK_NOWAIT);
process_worklist_item(NULL, LK_NOWAIT);
stat_worklist_push += 2;
This patch corrects two problems with the rate limiting code that was introduced in revision 1.80. The problem manifested itself with a `locking against myself' panic and could also result in soft updates inconsistences associated with inodedeps. The two problems are: 1) One of the background operations could manipulate the bitmap while holding it locked with intent to create. This held lock results in a `locking against myself' panic, when the background processing that we have been coopted to do tries to lock the bitmap which we are already holding locked. To understand how to fix this problem, first, observe that we can do the background cleanups in inodedep_lookup only when allocating inodedeps (DEPALLOC is set in the call to inodedep_lookup). Second observe that calls to inodedep_lookup with DEPALLOC set can only happen from the following calls into the softdep code: softdep_setup_inomapdep softdep_setup_allocdirect softdep_setup_remove softdep_setup_freeblocks softdep_setup_directory_change softdep_setup_directory_add softdep_change_linkcnt Only the first two of these can come from ffs_alloc.c while holding a bitmap locked. Thus, inodedep_lookup must not go off to do request_cleanups when being called from these functions. This change adds a flag, NODELAY, that can be passed to inodedep_lookup to let it know that it should not do background processing in those cases. 2) The return value from request_cleanup when helping out with the cleanup was 0 instead of 1. This meant that despite the fact that we may have slept while doing the cleanups, the code did not recheck for the appearance of an inodedep (e.g., goto top in inodedep_lookup). This lead to the softdep inconsistency in which we ended up with two inodedep's for the same inode. Reviewed by: Peter Wemm <peter@yahoo-inc.com>, Matt Dillon <dillon@earth.backplane.com>
2001-02-20 11:14:38 +00:00
if (islocked)
ACQUIRE_LOCK(&lk);
return(1);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
}
/*
* Next, we attempt to speed up the syncer process. If that
* is successful, then we allow the process to continue.
*/
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT)
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
return(0);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* If we are resource constrained on inode dependencies, try
* flushing some dirty inodes. Otherwise, we are constrained
* by file deletions, so try accelerating flushes of directories
* with removal dependencies. We would like to do the cleanup
* here, but we probably hold an inode locked at this point and
* that might deadlock against one that we try to clean. So,
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
* the best that we can do is request the syncer daemon to do
* the cleanup for us.
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
*/
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
switch (resource) {
case FLUSH_INODES:
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
stat_ino_limit_push += 1;
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
req_clear_inodedeps += 1;
stat_countp = &stat_ino_limit_hit;
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
break;
case FLUSH_REMOVE:
This patch fixes a long standing complaint with soft updates in which small and/or nearly full filesystems would fail with `file system full' messages when trying to replace a number of existing files (for example during a system installation). When the allocation routines are about to fail with a file system full condition, they make a call to softdep_request_cleanup() which attempts to accelerate the flushing of pending deletion requests in an effort to free up space. In the face of filesystem I/O requests that exceed the available disk transfer capacity, the cleanup request could take an unbounded amount of time. Thus, the softdep_request_cleanup() routine will only try for tickdelay seconds (default 2 seconds) before giving up and returning a filesystem full error. Under typical conditions, the softdep_request_cleanup() routine is able to free up space in under fifty milliseconds.
2002-01-22 06:17:22 +00:00
case FLUSH_REMOVE_WAIT:
Add a hook to ffs_fsync to allow soft updates to get first chance at doing a sync on the block device for the filesystem. That allows it to push the bitmap blocks before the inode blocks which greatly reduces the number of inode rollbacks that need to be done.
1999-05-14 01:26:46 +00:00
stat_blk_limit_push += 1;
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
req_clear_remove += 1;
stat_countp = &stat_blk_limit_hit;
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
break;
default:
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
if (islocked)
FREE_LOCK(&lk);
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
panic("request_cleanup: unknown type");
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
/*
* Hopefully the syncer daemon will catch up and awaken us.
* We wait at most tickdelay before proceeding in any case.
*/
if (islocked == 0)
ACQUIRE_LOCK(&lk);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
proc_waiting += 1;
if (handle.callout == NULL)
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, NULL, PPAUSE,
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
"softupdate", 0);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
proc_waiting -= 1;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
if (islocked == 0)
FREE_LOCK(&lk);
return (1);
}
/*
Get rid of the global variable rushjob and replace it with a function in kern/vfs_subr.c named speedup_syncer() which handles the speedup request. Change the various clients of rushjob to use the new function.
1999-06-15 23:37:29 +00:00
* Awaken processes pausing in request_cleanup and clear proc_waiting
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
* to indicate that there is no longer a timer running.
*/
Be consistent about "static" functions: if the function is marked static in its prototype, mark it static at the definition too. Inspired by: FlexeLint warning #512
2002-09-28 17:15:38 +00:00
static void
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
pause_timer(arg)
void *arg;
{
More aggressively rate limit the growth of soft dependency structures in the face of multiple processes doing massive numbers of filesystem operations. While this patch will work in nearly all situations, there are still some perverse workloads that can overwhelm the system. Detecting and handling these perverse workloads will be the subject of another patch. Reviewed by: Paul Saab <ps@yahoo-inc.com> Obtained from: Ethan Solomita <ethan@geocast.com>
2000-11-20 06:22:39 +00:00
*stat_countp += 1;
wakeup_one(&proc_waiting);
Preventing runaway kernel soft updates memory, take three. Previously, the syncer process was the only process in the system that could process the soft updates background work list. If enough other processes were adding requests to that list, it would eventually grow without bound. Because some of the work list requests require vnodes to be locked, it was not generally safe to let random processes process the work list while they already held vnodes locked. By adding a flag to the work list queue processing function to indicate whether the calling process could safely lock vnodes, it becomes possible to co-opt other processes into helping out with the work list. Now when the worklist gets too large, other processes can safely help out by picking off those work requests that can be handled without locking a vnode, leaving only the small number of requests requiring a vnode lock for the syncer process. With this change, it appears possible to keep even the nastiest workloads under control. Submitted by: Paul Saab <ps@yahoo-inc.com>
2000-12-13 08:30:35 +00:00
if (proc_waiting > 0)
handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
else
handle.callout = NULL;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
/*
Keep tighter control of removal dependencies by limiting the number of dirrem structure rather than the collaterally created freeblks and freefile structures. Limit the rate of buffer dirtying by the syncer process during periods of intense file removal.
2000-01-09 23:35:38 +00:00
* Flush out a directory with at least one removal dependency in an effort to
* reduce the number of dirrem, freefile, and freeblks dependency structures.
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
*/
static void
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
clear_remove(td)
struct thread *td;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
{
struct pagedep_hashhead *pagedephd;
struct pagedep *pagedep;
static int next = 0;
struct mount *mp;
struct vnode *vp;
int error, cnt;
ino_t ino;
ACQUIRE_LOCK(&lk);
for (cnt = 0; cnt < pagedep_hash; cnt++) {
pagedephd = &pagedep_hashtbl[next++];
if (next >= pagedep_hash)
next = 0;
Another round of the <sys/queue.h> FOREACH transmogriffer. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 16:08:18 +00:00
LIST_FOREACH(pagedep, pagedephd, pd_hash) {
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
continue;
mp = pagedep->pd_mnt;
ino = pagedep->pd_ino;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
continue;
Missplaced FREE_LOCK causes a panic when hit while taking a snapshot. Sponsored by: DARPA & NAI Labs.
2002-10-23 05:14:06 +00:00
FREE_LOCK(&lk);
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp))) {
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
softdep_error("clear_remove: vget", error);
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
vn_finished_write(mp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
return;
}
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
softdep_error("clear_remove: fsync", error);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
drain_output(vp, 0);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_UNLOCK(vp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
vput(vp);
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
vn_finished_write(mp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
return;
}
}
FREE_LOCK(&lk);
}
/*
* Clear out a block of dirty inodes in an effort to reduce
* the number of inodedep dependency structures.
*/
static void
KSE Milestone 2 Note ALL MODULES MUST BE RECOMPILED make the kernel aware that there are smaller units of scheduling than the process. (but only allow one thread per process at this time). This is functionally equivalent to teh previousl -current except that there is a thread associated with each process. Sorry john! (your next MFC will be a doosie!) Reviewed by: peter@freebsd.org, dillon@freebsd.org X-MFC after: ha ha ha ha
2001-09-12 08:38:13 +00:00
clear_inodedeps(td)
struct thread *td;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
{
struct inodedep_hashhead *inodedephd;
struct inodedep *inodedep;
static int next = 0;
struct mount *mp;
struct vnode *vp;
struct fs *fs;
int error, cnt;
ino_t firstino, lastino, ino;
ACQUIRE_LOCK(&lk);
/*
* Pick a random inode dependency to be cleared.
* We will then gather up all the inodes in its block
* that have dependencies and flush them out.
*/
for (cnt = 0; cnt < inodedep_hash; cnt++) {
inodedephd = &inodedep_hashtbl[next++];
if (next >= inodedep_hash)
next = 0;
if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
break;
}
When cleaning up excess inode dependencies, check for being done. Reviewed by: Jan Koum <jkb@yahoo-inc.com>
2001-02-22 10:17:57 +00:00
if (inodedep == NULL)
return;
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* Ugly code to find mount point given pointer to superblock.
*/
fs = inodedep->id_fs;
struct mountlist and struct mount.mnt_list have no business being a CIRCLEQ. Change them to TAILQ_HEAD and TAILQ_ENTRY respectively. This removes ugly mp != (void*)&mountlist comparisons. Requested by: phk Submitted by: Jake Burkholder jake@checker.org PR: 14967
1999-11-20 10:00:46 +00:00
TAILQ_FOREACH(mp, &mountlist, mnt_list)
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
break;
/*
* Find the last inode in the block with dependencies.
*/
firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
break;
/*
* Asynchronously push all but the last inode with dependencies.
* Synchronously push the last inode with dependencies to ensure
* that the inode block gets written to free up the inodedeps.
*/
for (ino = firstino; ino <= lastino; ino++) {
if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
continue;
This patch corrects the first round of panics and hangs reported with the new snapshot code. Update addaliasu to correctly implement the semantics of the old checkalias function. When a device vnode first comes into existence, check to see if an anonymous vnode for the same device was created at boot time by bdevvp(). If so, adopt the bdevvp vnode rather than creating a new vnode for the device. This corrects a problem which caused the kernel to panic when taking a snapshot of the root filesystem. Change the calling convention of vn_write_suspend_wait() to be the same as vn_start_write(). Split out softdep_flushworklist() from softdep_flushfiles() so that it can be used to clear the work queue when suspending filesystem operations. Access to buffers becomes recursive so that snapshots can recursively traverse their indirect blocks using ffs_copyonwrite() when checking for the need for copy on write when flushing one of their own indirect blocks. This eliminates a deadlock between the syncer daemon and a process taking a snapshot. Ensure that softdep_process_worklist() can never block because of a snapshot being taken. This eliminates a problem with buffer starvation. Cleanup change in ffs_sync() which did not synchronously wait when MNT_WAIT was specified. The result was an unclean filesystem panic when doing forcible unmount with heavy filesystem I/O in progress. Return a zero'ed block when reading a block that was not in use at the time that a snapshot was taken. Normally, these blocks should never be read. However, the readahead code will occationally read them which can cause unexpected behavior. Clean up the debugging code that ensures that no blocks be written on a filesystem while it is suspended. Snapshots must explicitly label the blocks that they are writing during the suspension so that they do not cause a `write on suspended filesystem' panic. Reorganize ffs_copyonwrite() to eliminate a deadlock and also to prevent a race condition that would permit the same block to be copied twice. This change eliminates an unexpected soft updates inconsistency in fsck caused by the double allocation. Use bqrelse rather than brelse for buffers that will be needed soon again by the snapshot code. This improves snapshot performance.
2000-07-24 05:28:33 +00:00
if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
continue;
In the function clear_inodedeps(), a FREE_LOCK() should be called AFTER the call to vn_start_write(), not before it. Otherwise, it is possible to unlock it multiple times if the vn_start_write() fails. Submitted by: Juergen Hannken-Illjes <hannken@eis.cs.tu-bs.de>
2004-02-23 06:56:31 +00:00
FREE_LOCK(&lk);
Add a flags parameter to VFS_VGET to pass through the desired locking flags when acquiring a vnode. The immediate purpose is to allow polling lock requests (LK_NOWAIT) needed by soft updates to avoid deadlock when enlisting other processes to help with the background cleanup. For the future it will allow the use of shared locks for read access to vnodes. This change touches a lot of files as it affects most filesystems within the system. It has been well tested on FFS, loopback, and CD-ROM filesystems. only lightly on the others, so if you find a problem there, please let me (mckusick@mckusick.com) know.
2002-03-17 01:25:47 +00:00
if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
softdep_error("clear_inodedeps: vget", error);
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
vn_finished_write(mp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
return;
}
if (ino == lastino) {
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_WAIT, td)))
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
softdep_error("clear_inodedeps: fsync1", error);
} else {
Simple p_ucred -> td_ucred changes to start using the per-thread ucred reference.
2002-02-27 18:32:23 +00:00
if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
softdep_error("clear_inodedeps: fsync2", error);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_LOCK(vp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
drain_output(vp, 0);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
VI_UNLOCK(vp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
vput(vp);
Add snapshots to the fast filesystem. Most of the changes support the gating of system calls that cause modifications to the underlying filesystem. The gating can be enabled by any filesystem that needs to consistently suspend operations by adding the vop_stdgetwritemount to their set of vnops. Once gating is enabled, the function vfs_write_suspend stops all new write operations to a filesystem, allows any filesystem modifying system calls already in progress to complete, then sync's the filesystem to disk and returns. The function vfs_write_resume allows the suspended write operations to begin again. Gating is not added by default for all filesystems as for SMP systems it adds two extra locks to such critical kernel paths as the write system call. Thus, gating should only be added as needed. Details on the use and current status of snapshots in FFS can be found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness is not included here. Unless and until you create a snapshot file, these changes should have no effect on your system (famous last words).
2000-07-11 22:07:57 +00:00
vn_finished_write(mp);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
ACQUIRE_LOCK(&lk);
}
FREE_LOCK(&lk);
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
/*
* Function to determine if the buffer has outstanding dependencies
* that will cause a roll-back if the buffer is written. If wantcount
* is set, return number of dependencies, otherwise just yes or no.
*/
static int
softdep_count_dependencies(bp, wantcount)
struct buf *bp;
int wantcount;
{
struct worklist *wk;
struct inodedep *inodedep;
struct indirdep *indirdep;
struct allocindir *aip;
struct pagedep *pagedep;
struct diradd *dap;
int i, retval;
retval = 0;
ACQUIRE_LOCK(&lk);
Mechanical change to use <sys/queue.h> macro API instead of fondling implementation details. Created with: sed(1) Reviewed by: md5(1)
2001-02-04 13:13:25 +00:00
LIST_FOREACH(wk, &bp->b_dep, wk_list) {
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
switch (wk->wk_type) {
case D_INODEDEP:
inodedep = WK_INODEDEP(wk);
if ((inodedep->id_state & DEPCOMPLETE) == 0) {
/* bitmap allocation dependency */
retval += 1;
if (!wantcount)
goto out;
}
if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
/* direct block pointer dependency */
retval += 1;
if (!wantcount)
goto out;
}
Add support to UFS2 to provide storage for extended attributes. As this code is not actually used by any of the existing interfaces, it seems unlikely to break anything (famous last words). The internal kernel interface to manipulate these attributes is invoked using two new IO_ flags: IO_NORMAL and IO_EXT. These flags may be specified in the ioflags word of VOP_READ, VOP_WRITE, and VOP_TRUNCATE. Specifying IO_NORMAL means that you want to do I/O to the normal data part of the file and IO_EXT means that you want to do I/O to the extended attributes part of the file. IO_NORMAL and IO_EXT are mutually exclusive for VOP_READ and VOP_WRITE, but may be specified individually or together in the case of VOP_TRUNCATE. For example, when removing a file, VOP_TRUNCATE is called with both IO_NORMAL and IO_EXT set. For backward compatibility, if neither IO_NORMAL nor IO_EXT is set, then IO_NORMAL is assumed. Note that the BA_ and IO_ flags have been `merged' so that they may both be used in the same flags word. This merger is possible by assigning the IO_ flags to the low sixteen bits and the BA_ flags the high sixteen bits. This works because the high sixteen bits of the IO_ word is reserved for read-ahead and help with write clustering so will never be used for flags. This merge lets us get away from code of the form: if (ioflags & IO_SYNC) flags |= BA_SYNC; For the future, I have considered adding a new field to the vattr structure, va_extsize. This addition could then be exported through the stat structure to allow applications to find out the size of the extended attribute storage and also would provide a more standard interface for truncating them (via VOP_SETATTR rather than VOP_TRUNCATE). I am also contemplating adding a pathconf parameter (for concreteness, lets call it _PC_MAX_EXTSIZE) which would let an application determine the maximum size of the extended atribute storage. Sponsored by: DARPA & NAI Labs.
2002-07-19 07:29:39 +00:00
if (TAILQ_FIRST(&inodedep->id_extupdt)) {
/* direct block pointer dependency */
retval += 1;
if (!wantcount)
goto out;
}
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
continue;
case D_INDIRDEP:
indirdep = WK_INDIRDEP(wk);
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
/* indirect block pointer dependency */
retval += 1;
if (!wantcount)
goto out;
}
continue;
case D_PAGEDEP:
pagedep = WK_PAGEDEP(wk);
for (i = 0; i < DAHASHSZ; i++) {
Use <sys/queue.h> macro API.
2001-02-04 12:37:48 +00:00
LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
/* directory entry dependency */
retval += 1;
if (!wantcount)
goto out;
}
}
continue;
case D_BMSAFEMAP:
case D_ALLOCDIRECT:
case D_ALLOCINDIR:
case D_MKDIR:
/* never a dependency on these blocks */
continue;
default:
Free lock before calling panic so that subsequent attempt to write out buffers does not re-panic with `locking against myself'. This change should not affect normal operations of soft updates in any way.
2001-02-23 09:01:31 +00:00
FREE_LOCK(&lk);
Several performance improvements for soft updates have been added: 1) Fastpath deletions. When a file is being deleted, check to see if it was so recently created that its inode has not yet been written to disk. If so, the delete can proceed to immediately free the inode. 2) Background writes: No file or block allocations can be done while the bitmap is being written to disk. To avoid these stalls, the bitmap is copied to another buffer which is written thus leaving the original available for futher allocations. 3) Link count tracking. Constantly track the difference in i_effnlink and i_nlink so that inodes that have had no change other than i_effnlink need not be written. 4) Identify buffers with rollback dependencies so that the buffer flushing daemon can choose to skip over them.
2000-01-10 00:24:24 +00:00
panic("softdep_check_for_rollback: Unexpected type %s",
TYPENAME(wk->wk_type));
/* NOTREACHED */
}
}
out:
FREE_LOCK(&lk);
return retval;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Acquire exclusive access to a buffer.
* Must be called with splbio blocked.
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
* Return acquired buffer or NULL on failure. mtx, if provided, will be
* released on success but held on failure.
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
*/
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
static struct buf *
getdirtybuf(bpp, mtx, waitfor)
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
struct buf **bpp;
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
struct mtx *mtx;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
int waitfor;
{
struct buf *bp;
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
int error;
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
- Move BX_BKGRDWAIT and BX_BKGRDINPROG to BV_ and the b_vflags field. - Surround all accesses of the BKGRD{WAIT,INPROG} flags with the vnode interlock. - Don't use the B_LOCKED flag and QUEUE_LOCKED for background write buffers. Check for the BKGRDINPROG flag before recycling or throwing away a buffer. We do this instead because it is not safe for us to move the original buffer to a new queue from the callback on the background write buffer. - Remove the B_LOCKED flag and the locked buffer queue. They are no longer used. - The vnode interlock is used around checks for BKGRDINPROG where it may not be strictly necessary. If we hold the buf lock the a back-ground write will not be started without our knowledge, one may only be completed while we're not looking. Rather than remove the code, Document two of the places where this extra locking is done. A pass should be done to verify and minimize the locking later.
2003-08-28 06:55:18 +00:00
/*
* XXX This code and the code that calls it need to be reviewed to
* verify its use of the vnode interlock.
*/
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
for (;;) {
if ((bp = *bpp) == NULL)
return (0);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (bp->b_vp == NULL)
Update for the KDB debugger framework: o Make debugging code conditional upon KDB. o Use kdb_backtrace() instead of backtrace(). o Remove inclusion of opt_ddb.h.
2004-07-10 20:45:47 +00:00
kdb_backtrace();
- Add an interlock argument to BUF_LOCK and BUF_TIMELOCK. - Remove the buftimelock mutex and acquire the buf's interlock to protect these fields instead. - Hold the vnode interlock while locking bufs on the clean/dirty queues. This reduces some cases from one BUF_LOCK with a LK_NOWAIT and another BUF_LOCK with a LK_TIMEFAIL to a single lock. Reviewed by: arch, mckusick
2003-02-25 03:37:48 +00:00
if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) {
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if ((bp->b_vflags & BV_BKGRDINPROG) == 0)
Because cylinder group blocks are now written in background, it is no longer sufficient to get a lock on a buffer to know that its write has been completed. We have to first get the lock on the buffer, then check to see if it is doing a background write. If it is doing background write, we have to wait for the background write to finish, then check to see if that fullfilled our dependency, and if not to start another write. Luckily the explanation is longer than the fix.
2000-01-13 07:20:01 +00:00
break;
BUF_UNLOCK(bp);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (waitfor != MNT_WAIT)
return (NULL);
/*
* The mtx argument must be bp->b_vp's mutex in
* this case.
*/
- The VI assert in getdirtybuf() is only valid if we're not on a VCHR vnode. VCHR vnodes don't do background writes. Reported by: kan
2003-10-04 15:57:05 +00:00
#ifdef DEBUG_VFS_LOCKS
if (bp->b_vp->v_type != VCHR)
ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf");
#endif
- Move BX_BKGRDWAIT and BX_BKGRDINPROG to BV_ and the b_vflags field. - Surround all accesses of the BKGRD{WAIT,INPROG} flags with the vnode interlock. - Don't use the B_LOCKED flag and QUEUE_LOCKED for background write buffers. Check for the BKGRDINPROG flag before recycling or throwing away a buffer. We do this instead because it is not safe for us to move the original buffer to a new queue from the callback on the background write buffer. - Remove the B_LOCKED flag and the locked buffer queue. They are no longer used. - The vnode interlock is used around checks for BKGRDINPROG where it may not be strictly necessary. If we hold the buf lock the a back-ground write will not be started without our knowledge, one may only be completed while we're not looking. Rather than remove the code, Document two of the places where this extra locking is done. A pass should be done to verify and minimize the locking later.
2003-08-28 06:55:18 +00:00
bp->b_vflags |= BV_BKGRDWAIT;
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
interlocked_sleep(&lk, SLEEP, &bp->b_xflags, mtx,
PRIBIO, "getbuf", 0);
Because cylinder group blocks are now written in background, it is no longer sufficient to get a lock on a buffer to know that its write has been completed. We have to first get the lock on the buffer, then check to see if it is doing a background write. If it is doing background write, we have to wait for the background write to finish, then check to see if that fullfilled our dependency, and if not to start another write. Luckily the explanation is longer than the fix.
2000-01-13 07:20:01 +00:00
continue;
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
if (waitfor != MNT_WAIT)
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
return (NULL);
if (mtx) {
error = interlocked_sleep(&lk, LOCKBUF, bp, mtx,
LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 0, 0);
mtx_lock(mtx);
} else
error = interlocked_sleep(&lk, LOCKBUF, bp, NULL,
LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
if (error != ENOLCK) {
FREE_LOCK(&lk);
Convert buffer locking from using the B_BUSY and B_WANTED flags to using lockmgr locks. This commit should be functionally equivalent to the old semantics. That is, all buffer locking is done with LK_EXCLUSIVE requests. Changes to take advantage of LK_SHARED and LK_RECURSIVE will be done in future commits.
1999-06-26 02:47:16 +00:00
panic("getdirtybuf: inconsistent lock");
When going to sleep, we must save our SPL so that it does not get lost if some other process uses the lock while we are sleeping. We restore it after we have slept. This functionality is provided by a new routine interlocked_sleep() that wraps the interlocking with functions that sleep. This function is then used in place of the old ACQUIRE_LOCK_INTERLOCKED() and FREE_LOCK_INTERLOCKED() macros. Submitted by: Debbie Chu <dchu@juniper.net>
2002-01-12 20:57:36 +00:00
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Convert buffer locking from using the B_BUSY and B_WANTED flags to using lockmgr locks. This commit should be functionally equivalent to the old semantics. That is, all buffer locking is done with LK_EXCLUSIVE requests. Changes to take advantage of LK_SHARED and LK_RECURSIVE will be done in future commits.
1999-06-26 02:47:16 +00:00
if ((bp->b_flags & B_DELWRI) == 0) {
BUF_UNLOCK(bp);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
return (NULL);
Convert buffer locking from using the B_BUSY and B_WANTED flags to using lockmgr locks. This commit should be functionally equivalent to the old semantics. That is, all buffer locking is done with LK_EXCLUSIVE requests. Changes to take advantage of LK_SHARED and LK_RECURSIVE will be done in future commits.
1999-06-26 02:47:16 +00:00
}
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
if (mtx)
mtx_unlock(mtx);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
bremfree(bp);
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
return (bp);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
/*
* Wait for pending output on a vnode to complete.
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
* Must be called with vnode lock and interlock locked.
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
*/
static void
drain_output(vp, islocked)
struct vnode *vp;
int islocked;
{
- Don't acquire the vnode interlock in drain_output(). Instead, require the caller to acquire it. This permits drain_output() to be done atomically with other operations as well as reducing the number of lock operations. - Assert that the proper locks are held in drain_output(). - Change getdirtybuf() to accept a mutex as an argument. This mutex is used to protect the vnode's buf list and the BKGRDWAIT flag. This lock is dropped when we successfully acquire a buffer and held on return otherwise. These semantics reduce the number of cumbersome cases in calling code. - Pass the mtx from getdirtybuf() into interlocked_sleep() and allow this mutex to be used as the interlock argument to BUF_LOCK() in the LOCKBUF case of interlocked_sleep(). - Change the return value of getdirtybuf() to be the resulting locked buffer or NULL otherwise. This is for callers who pass in a list head that requires a lock. It is necessary since the lock that protects the list head must be dropped in getdirtybuf() so that we don't have a lock order reversal with the buf queues lock in bremfree(). - Adjust all callers of getdirtybuf() to match the new semantics. - Add a comment in indir_trunc() that points at unlocked access to a buf. This may also be one of the last instances of incore() in the tree.
2003-08-31 07:29:34 +00:00
ASSERT_VOP_LOCKED(vp, "drain_output");
ASSERT_VI_LOCKED(vp, "drain_output");
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
if (!islocked)
ACQUIRE_LOCK(&lk);
Move the VI_BWAIT flag into no bo_flag element of bufobj and call it BO_WWAIT Add bufobj_wref(), bufobj_wdrop() and bufobj_wwait() to handle the write count on a bufobj. Bufobj_wdrop() replaces vwakeup(). Use these functions all relevant places except in ffs_softdep.c where the use if interlocked_sleep() makes this impossible. Rename b_vnbufs to b_bobufs now that we touch all the relevant files anyway.
2004-10-21 15:53:54 +00:00
while (vp->v_bufobj.bo_numoutput) {
vp->v_bufobj.bo_flag |= BO_WWAIT;
interlocked_sleep(&lk, SLEEP,
(caddr_t)&vp->v_bufobj.bo_numoutput,
- Replace v_flag with v_iflag and v_vflag - v_vflag is protected by the vnode lock and is used when synchronization with VOP calls is needed. - v_iflag is protected by interlock and is used for dealing with vnode management issues. These flags include X/O LOCK, FREE, DOOMED, etc. - All accesses to v_iflag and v_vflag have either been locked or marked with mp_fixme's. - Many ASSERT_VOP_LOCKED calls have been added where the locking was not clear. - Many functions in vfs_subr.c were restructured to provide for stronger locking. Idea stolen from: BSD/OS
2002-08-04 10:29:36 +00:00
VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
Severe slowdowns have been reported when creating or removing many files at once on a filesystem running soft updates. The root of the problem is that soft updates limits the amount of memory that may be allocated to dependency structures so as to avoid hogging kernel memory. The original algorithm just waited for the disk I/O to catch up and reduce the number of dependencies. This new code takes a much more aggressive approach. Basically there are two resources that routinely hit the limit. Inode dependencies during periods with a high file creation rate and file and block removal dependencies during periods with a high file removal rate. I have attacked these problems from two fronts. When the inode dependency limits are reached, I pick a random inode dependency, UFS_UPDATE it together with all the other dirty inodes contained within its disk block and then write that disk block. This trick usually clears 5-50 inode dependencies in a single disk I/O. For block and file removal dependencies, I pick a random directory page that has at least one remove pending and VOP_FSYNC its directory. That releases all its removal dependencies to the work queue. To further hasten things along, I also immediately start the work queue process rather than waiting for its next one second scheduled run.
1999-05-07 02:26:47 +00:00
}
if (!islocked)
FREE_LOCK(&lk);
}
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/*
* Called whenever a buffer that is being invalidated or reallocated
* contains dependencies. This should only happen if an I/O error has
* occurred. The routine is called with the buffer locked.
*/
Make static non-exported functions from soft updates.
2000-01-09 22:40:09 +00:00
static void
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_deallocate_dependencies(bp)
struct buf *bp;
{
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
Move B_ERROR flag to b_ioflags and call it BIO_ERROR. (Much of this done by script) Move B_ORDERED flag to b_ioflags and call it BIO_ORDERED. Move b_pblkno and b_iodone_chain to struct bio while we transition, they will be obsoleted once bio structs chain/stack. Add bio_queue field for struct bio aware disksort. Address a lot of stylistic issues brought up by bde.
2000-04-02 15:24:56 +00:00
if ((bp->b_ioflags & BIO_ERROR) == 0)
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
panic("softdep_deallocate_dependencies: dangling deps");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
panic("softdep_deallocate_dependencies: unrecovered I/O error");
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
/*
* Function to handle asynchronous write errors in the filesystem.
*/
Be consistent about "static" functions: if the function is marked static in its prototype, mark it static at the definition too. Inspired by: FlexeLint warning #512
2002-09-28 17:15:38 +00:00
static void
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
softdep_error(func, error)
char *func;
int error;
{
Get rid of random debugging cruft; sync up with latest version.
1999-05-07 05:11:31 +00:00
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
/* XXX should do something better! */
Gutted softdep_deallocate_dependencies and replaced it with a panic. It turns out to not be useful to unwind the dependencies and continue in the face of a fatal error. Also changed the log() to a printf() in softdep_error() so that it will be output in the case of a impending panic. Submitted by: Kirk McKusick <mckusick@mckusick.com>
1999-01-22 09:07:32 +00:00
printf("%s: got error %d while accessing filesystem\n", func, error);
Import the earliest version of the soft update code that I have.
1998-05-19 19:47:22 +00:00
}
Reference in New Issue Copy Permalink