1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Copyright (c) 1998 Matthew Dillon,
|
1994-05-25 09:21:21 +00:00
|
|
|
* Copyright (c) 1994 John S. Dyson
|
1994-05-24 10:09:53 +00:00
|
|
|
* Copyright (c) 1990 University of Utah.
|
|
|
|
* Copyright (c) 1991, 1993
|
|
|
|
* The Regents of the University of California. All rights reserved.
|
|
|
|
*
|
|
|
|
* This code is derived from software contributed to Berkeley by
|
|
|
|
* the Systems Programming Group of the University of Utah Computer
|
|
|
|
* Science Department.
|
|
|
|
*
|
|
|
|
* 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.
|
|
|
|
* 3. All advertising materials mentioning features or use of this software
|
2000-03-27 20:41:17 +00:00
|
|
|
* must display the following acknowledgement:
|
1994-05-24 10:09:53 +00:00
|
|
|
* This product includes software developed by the University of
|
|
|
|
* California, Berkeley and its contributors.
|
|
|
|
* 4. Neither the name of the University nor the names of its contributors
|
|
|
|
* may be used to endorse or promote products derived from this software
|
|
|
|
* without specific prior written permission.
|
|
|
|
*
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``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 THE REGENTS OR CONTRIBUTORS 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.
|
|
|
|
*
|
1999-01-21 08:29:12 +00:00
|
|
|
* New Swap System
|
|
|
|
* Matthew Dillon
|
|
|
|
*
|
|
|
|
* Radix Bitmap 'blists'.
|
|
|
|
*
|
|
|
|
* - The new swapper uses the new radix bitmap code. This should scale
|
|
|
|
* to arbitrarily small or arbitrarily large swap spaces and an almost
|
|
|
|
* arbitrary degree of fragmentation.
|
|
|
|
*
|
|
|
|
* Features:
|
|
|
|
*
|
|
|
|
* - on the fly reallocation of swap during putpages. The new system
|
|
|
|
* does not try to keep previously allocated swap blocks for dirty
|
|
|
|
* pages.
|
|
|
|
*
|
|
|
|
* - on the fly deallocation of swap
|
|
|
|
*
|
|
|
|
* - No more garbage collection required. Unnecessarily allocated swap
|
|
|
|
* blocks only exist for dirty vm_page_t's now and these are already
|
|
|
|
* cycled (in a high-load system) by the pager. We also do on-the-fly
|
|
|
|
* removal of invalidated swap blocks when a page is destroyed
|
|
|
|
* or renamed.
|
|
|
|
*
|
1994-05-24 10:09:53 +00:00
|
|
|
* from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
|
|
|
|
*
|
|
|
|
* @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
|
1999-01-21 08:29:12 +00:00
|
|
|
*
|
1999-08-28 01:08:13 +00:00
|
|
|
* $FreeBSD$
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
#include <sys/param.h>
|
|
|
|
#include <sys/systm.h>
|
1999-08-23 23:55:03 +00:00
|
|
|
#include <sys/conf.h>
|
1995-04-16 13:58:42 +00:00
|
|
|
#include <sys/kernel.h>
|
1994-05-24 10:09:53 +00:00
|
|
|
#include <sys/proc.h>
|
2000-05-05 09:59:14 +00:00
|
|
|
#include <sys/bio.h>
|
1994-05-24 10:09:53 +00:00
|
|
|
#include <sys/buf.h>
|
|
|
|
#include <sys/vnode.h>
|
|
|
|
#include <sys/malloc.h>
|
1995-12-07 12:48:31 +00:00
|
|
|
#include <sys/vmmeter.h>
|
1999-02-18 19:57:33 +00:00
|
|
|
#include <sys/sysctl.h>
|
1999-01-21 08:29:12 +00:00
|
|
|
#include <sys/blist.h>
|
|
|
|
#include <sys/lock.h>
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1998-02-23 08:22:48 +00:00
|
|
|
#ifndef MAX_PAGEOUT_CLUSTER
|
1998-03-01 04:18:54 +00:00
|
|
|
#define MAX_PAGEOUT_CLUSTER 16
|
1998-02-23 08:22:48 +00:00
|
|
|
#endif
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
#define SWB_NPAGES MAX_PAGEOUT_CLUSTER
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
#include "opt_swap.h"
|
1994-05-24 10:09:53 +00:00
|
|
|
#include <vm/vm.h>
|
1995-12-07 12:48:31 +00:00
|
|
|
#include <vm/vm_object.h>
|
1994-05-24 10:09:53 +00:00
|
|
|
#include <vm/vm_page.h>
|
1995-12-07 12:48:31 +00:00
|
|
|
#include <vm/vm_pager.h>
|
1994-05-24 10:09:53 +00:00
|
|
|
#include <vm/vm_pageout.h>
|
|
|
|
#include <vm/swap_pager.h>
|
1995-12-07 12:48:31 +00:00
|
|
|
#include <vm/vm_extern.h>
|
1999-01-21 08:29:12 +00:00
|
|
|
#include <vm/vm_zone.h>
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
#define SWM_FREE 0x02 /* free, period */
|
|
|
|
#define SWM_POP 0x04 /* pop out */
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
|
|
|
|
* in the old system.
|
|
|
|
*/
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
extern int vm_swap_size; /* number of free swap blocks, in pages */
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-02-21 08:30:49 +00:00
|
|
|
int swap_pager_full; /* swap space exhaustion (task killing) */
|
|
|
|
static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
|
1999-01-21 08:29:12 +00:00
|
|
|
static int nsw_rcount; /* free read buffers */
|
1999-02-18 19:57:33 +00:00
|
|
|
static int nsw_wcount_sync; /* limit write buffers / synchronous */
|
|
|
|
static int nsw_wcount_async; /* limit write buffers / asynchronous */
|
|
|
|
static int nsw_wcount_async_max;/* assigned maximum */
|
|
|
|
static int nsw_cluster_max; /* maximum VOP I/O allowed */
|
1999-01-21 08:29:12 +00:00
|
|
|
static int sw_alloc_interlock; /* swap pager allocation interlock */
|
1995-12-14 09:55:16 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
struct blist *swapblist;
|
|
|
|
static struct swblock **swhash;
|
|
|
|
static int swhash_mask;
|
1999-02-18 19:57:33 +00:00
|
|
|
static int swap_async_max = 4; /* maximum in-progress async I/O's */
|
|
|
|
|
1999-12-28 07:30:55 +00:00
|
|
|
extern struct vnode *swapdev_vp; /* from vm_swap.c */
|
1999-11-22 15:27:09 +00:00
|
|
|
|
1999-02-18 19:57:33 +00:00
|
|
|
SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
|
|
|
|
CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
|
1995-12-14 09:55:16 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* "named" and "unnamed" anon region objects. Try to reduce the overhead
|
|
|
|
* of searching a named list by hashing it just a little.
|
|
|
|
*/
|
1995-12-14 09:55:16 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
#define NOBJLISTS 8
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
#define NOBJLIST(handle) \
|
1999-08-23 23:55:03 +00:00
|
|
|
(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
static struct pagerlst swap_pager_object_list[NOBJLISTS];
|
|
|
|
struct pagerlst swap_pager_un_object_list;
|
|
|
|
vm_zone_t swap_zone;
|
1994-05-24 10:09:53 +00:00
|
|
|
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
|
|
|
|
* calls hooked from other parts of the VM system and do not appear here.
|
|
|
|
* (see vm/swap_pager.h).
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1995-11-16 09:51:22 +00:00
|
|
|
static vm_object_t
|
1998-10-13 08:24:45 +00:00
|
|
|
swap_pager_alloc __P((void *handle, vm_ooffset_t size,
|
1995-12-11 04:58:34 +00:00
|
|
|
vm_prot_t prot, vm_ooffset_t offset));
|
1995-11-16 09:51:22 +00:00
|
|
|
static void swap_pager_dealloc __P((vm_object_t object));
|
1995-12-14 09:55:16 +00:00
|
|
|
static int swap_pager_getpages __P((vm_object_t, vm_page_t *, int, int));
|
1995-11-16 09:51:22 +00:00
|
|
|
static void swap_pager_init __P((void));
|
1999-01-21 08:29:12 +00:00
|
|
|
static void swap_pager_unswapped __P((vm_page_t));
|
2000-05-03 07:47:46 +00:00
|
|
|
static void swap_pager_strategy __P((vm_object_t, struct bio *));
|
1995-12-14 09:55:16 +00:00
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
struct pagerops swappagerops = {
|
1999-01-21 08:29:12 +00:00
|
|
|
swap_pager_init, /* early system initialization of pager */
|
|
|
|
swap_pager_alloc, /* allocate an OBJT_SWAP object */
|
|
|
|
swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
|
|
|
|
swap_pager_getpages, /* pagein */
|
|
|
|
swap_pager_putpages, /* pageout */
|
|
|
|
swap_pager_haspage, /* get backing store status for page */
|
1999-03-14 09:20:01 +00:00
|
|
|
swap_pager_unswapped, /* remove swap related to page */
|
|
|
|
swap_pager_strategy /* pager strategy call */
|
1994-05-24 10:09:53 +00:00
|
|
|
};
|
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
static struct buf *getchainbuf(struct bio *bp, struct vnode *vp, int flags);
|
2000-05-01 19:38:51 +00:00
|
|
|
static void flushchainbuf(struct buf *nbp);
|
2000-05-03 07:47:46 +00:00
|
|
|
static void waitchainbuf(struct bio *bp, int count, int done);
|
2000-05-01 19:38:51 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* dmmax is in page-sized chunks with the new swap system. It was
|
2000-10-13 16:44:34 +00:00
|
|
|
* dev-bsized chunks in the old. dmmax is always a power of 2.
|
1999-01-21 08:29:12 +00:00
|
|
|
*
|
|
|
|
* swap_*() routines are externally accessible. swp_*() routines are
|
|
|
|
* internal.
|
|
|
|
*/
|
|
|
|
|
1995-12-14 09:55:16 +00:00
|
|
|
int dmmax;
|
1999-01-21 08:29:12 +00:00
|
|
|
static int dmmax_mask;
|
1999-02-21 08:30:49 +00:00
|
|
|
int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
|
|
|
|
int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
static __inline void swp_sizecheck __P((void));
|
|
|
|
static void swp_pager_sync_iodone __P((struct buf *bp));
|
|
|
|
static void swp_pager_async_iodone __P((struct buf *bp));
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Swap bitmap functions
|
|
|
|
*/
|
|
|
|
|
|
|
|
static __inline void swp_pager_freeswapspace __P((daddr_t blk, int npages));
|
|
|
|
static __inline daddr_t swp_pager_getswapspace __P((int npages));
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Metadata functions
|
|
|
|
*/
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
static void swp_pager_meta_build __P((vm_object_t, vm_pindex_t, daddr_t));
|
|
|
|
static void swp_pager_meta_free __P((vm_object_t, vm_pindex_t, daddr_t));
|
1999-01-21 08:29:12 +00:00
|
|
|
static void swp_pager_meta_free_all __P((vm_object_t));
|
|
|
|
static daddr_t swp_pager_meta_ctl __P((vm_object_t, vm_pindex_t, int));
|
|
|
|
|
|
|
|
/*
|
|
|
|
* SWP_SIZECHECK() - update swap_pager_full indication
|
|
|
|
*
|
1999-02-21 08:30:49 +00:00
|
|
|
* update the swap_pager_almost_full indication and warn when we are
|
|
|
|
* about to run out of swap space, using lowat/hiwat hysteresis.
|
|
|
|
*
|
|
|
|
* Clear swap_pager_full ( task killing ) indication when lowat is met.
|
1999-01-21 08:29:12 +00:00
|
|
|
*
|
|
|
|
* No restrictions on call
|
|
|
|
* This routine may not block.
|
|
|
|
* This routine must be called at splvm()
|
|
|
|
*/
|
1996-03-02 02:54:24 +00:00
|
|
|
|
1998-04-15 17:47:40 +00:00
|
|
|
static __inline void
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_sizecheck()
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
if (vm_swap_size < nswap_lowat) {
|
1999-02-21 08:30:49 +00:00
|
|
|
if (swap_pager_almost_full == 0) {
|
1996-01-31 13:14:21 +00:00
|
|
|
printf("swap_pager: out of swap space\n");
|
1999-02-21 08:30:49 +00:00
|
|
|
swap_pager_almost_full = 1;
|
1999-02-06 07:22:21 +00:00
|
|
|
}
|
1999-02-21 08:30:49 +00:00
|
|
|
} else {
|
1994-05-25 09:21:21 +00:00
|
|
|
swap_pager_full = 0;
|
1999-02-21 08:30:49 +00:00
|
|
|
if (vm_swap_size > nswap_hiwat)
|
|
|
|
swap_pager_almost_full = 0;
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* SWAP_PAGER_INIT() - initialize the swap pager!
|
|
|
|
*
|
|
|
|
* Expected to be started from system init. NOTE: This code is run
|
|
|
|
* before much else so be careful what you depend on. Most of the VM
|
|
|
|
* system has yet to be initialized at this point.
|
|
|
|
*/
|
|
|
|
|
1995-11-14 20:53:20 +00:00
|
|
|
static void
|
1994-05-24 10:09:53 +00:00
|
|
|
swap_pager_init()
|
|
|
|
{
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Initialize object lists
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < NOBJLISTS; ++i)
|
|
|
|
TAILQ_INIT(&swap_pager_object_list[i]);
|
|
|
|
TAILQ_INIT(&swap_pager_un_object_list);
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Device Stripe, in PAGE_SIZE'd blocks
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
dmmax = SWB_NPAGES * 2;
|
|
|
|
dmmax_mask = ~(dmmax - 1);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
|
|
|
|
*
|
|
|
|
* Expected to be started from pageout process once, prior to entering
|
|
|
|
* its main loop.
|
|
|
|
*/
|
|
|
|
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
void
|
|
|
|
swap_pager_swap_init()
|
1994-05-24 10:09:53 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
int n;
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Number of in-transit swap bp operations. Don't
|
|
|
|
* exhaust the pbufs completely. Make sure we
|
|
|
|
* initialize workable values (0 will work for hysteresis
|
|
|
|
* but it isn't very efficient).
|
|
|
|
*
|
1999-02-18 19:57:33 +00:00
|
|
|
* The nsw_cluster_max is constrained by the bp->b_pages[]
|
1999-01-21 08:29:12 +00:00
|
|
|
* array (MAXPHYS/PAGE_SIZE) and our locally defined
|
|
|
|
* MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
|
|
|
|
* constrained by the swap device interleave stripe size.
|
1999-02-18 19:57:33 +00:00
|
|
|
*
|
|
|
|
* Currently we hardwire nsw_wcount_async to 4. This limit is
|
|
|
|
* designed to prevent other I/O from having high latencies due to
|
|
|
|
* our pageout I/O. The value 4 works well for one or two active swap
|
|
|
|
* devices but is probably a little low if you have more. Even so,
|
|
|
|
* a higher value would probably generate only a limited improvement
|
|
|
|
* with three or four active swap devices since the system does not
|
|
|
|
* typically have to pageout at extreme bandwidths. We will want
|
|
|
|
* at least 2 per swap devices, and 4 is a pretty good value if you
|
|
|
|
* have one NFS swap device due to the command/ack latency over NFS.
|
|
|
|
* So it all works out pretty well.
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
|
1999-02-21 08:34:15 +00:00
|
|
|
nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
|
1999-02-18 19:57:33 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
nsw_rcount = (nswbuf + 1) / 2;
|
1999-02-18 19:57:33 +00:00
|
|
|
nsw_wcount_sync = (nswbuf + 3) / 4;
|
|
|
|
nsw_wcount_async = 4;
|
|
|
|
nsw_wcount_async_max = nsw_wcount_async;
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Initialize our zone. Right now I'm just guessing on the number
|
|
|
|
* we need based on the number of pages in the system. Each swblock
|
|
|
|
* can hold 16 pages, so this is probably overkill.
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
n = cnt.v_page_count * 2;
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
swap_zone = zinit(
|
|
|
|
"SWAPMETA",
|
|
|
|
sizeof(struct swblock),
|
|
|
|
n,
|
|
|
|
ZONE_INTERRUPT,
|
|
|
|
1
|
|
|
|
);
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Initialize our meta-data hash table. The swapper does not need to
|
|
|
|
* be quite as efficient as the VM system, so we do not use an
|
|
|
|
* oversized hash table.
|
|
|
|
*
|
|
|
|
* n: size of hash table, must be power of 2
|
|
|
|
* swhash_mask: hash table index mask
|
|
|
|
*/
|
|
|
|
|
|
|
|
for (n = 1; n < cnt.v_page_count / 4; n <<= 1)
|
|
|
|
;
|
|
|
|
|
|
|
|
swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK);
|
|
|
|
bzero(swhash, sizeof(struct swblock *) * n);
|
|
|
|
|
|
|
|
swhash_mask = n - 1;
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
|
|
|
|
* its metadata structures.
|
|
|
|
*
|
|
|
|
* This routine is called from the mmap and fork code to create a new
|
|
|
|
* OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
|
|
|
|
* and then converting it with swp_pager_meta_build().
|
|
|
|
*
|
|
|
|
* This routine may block in vm_object_allocate() and create a named
|
|
|
|
* object lookup race, so we must interlock. We must also run at
|
|
|
|
* splvm() for the object lookup to handle races with interrupts, but
|
|
|
|
* we do not have to maintain splvm() in between the lookup and the
|
|
|
|
* add because (I believe) it is not possible to attempt to create
|
|
|
|
* a new swap object w/handle when a default object with that handle
|
|
|
|
* already exists.
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1995-11-14 20:53:20 +00:00
|
|
|
static vm_object_t
|
1998-10-13 08:24:45 +00:00
|
|
|
swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
|
1997-08-25 22:15:31 +00:00
|
|
|
vm_ooffset_t offset)
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
{
|
|
|
|
vm_object_t object;
|
|
|
|
|
|
|
|
if (handle) {
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Reference existing named region or allocate new one. There
|
|
|
|
* should not be a race here against swp_pager_meta_build()
|
|
|
|
* as called from vm_page_remove() in regards to the lookup
|
|
|
|
* of the handle.
|
|
|
|
*/
|
|
|
|
|
|
|
|
while (sw_alloc_interlock) {
|
|
|
|
sw_alloc_interlock = -1;
|
|
|
|
tsleep(&sw_alloc_interlock, PVM, "swpalc", 0);
|
|
|
|
}
|
|
|
|
sw_alloc_interlock = 1;
|
|
|
|
|
|
|
|
object = vm_pager_object_lookup(NOBJLIST(handle), handle);
|
|
|
|
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
if (object != NULL) {
|
|
|
|
vm_object_reference(object);
|
|
|
|
} else {
|
1999-01-21 08:29:12 +00:00
|
|
|
object = vm_object_allocate(OBJT_DEFAULT,
|
1998-10-13 08:24:45 +00:00
|
|
|
OFF_TO_IDX(offset + PAGE_MASK + size));
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
object->handle = handle;
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_meta_build(object, 0, SWAPBLK_NONE);
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
if (sw_alloc_interlock < 0)
|
|
|
|
wakeup(&sw_alloc_interlock);
|
|
|
|
|
|
|
|
sw_alloc_interlock = 0;
|
1994-05-24 10:09:53 +00:00
|
|
|
} else {
|
1999-01-21 08:29:12 +00:00
|
|
|
object = vm_object_allocate(OBJT_DEFAULT,
|
1998-10-13 08:24:45 +00:00
|
|
|
OFF_TO_IDX(offset + PAGE_MASK + size));
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_meta_build(object, 0, SWAPBLK_NONE);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
|
|
|
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
return (object);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_DEALLOC() - remove swap metadata from object
|
|
|
|
*
|
|
|
|
* The swap backing for the object is destroyed. The code is
|
|
|
|
* designed such that we can reinstantiate it later, but this
|
|
|
|
* routine is typically called only when the entire object is
|
|
|
|
* about to be destroyed.
|
|
|
|
*
|
|
|
|
* This routine may block, but no longer does.
|
|
|
|
*
|
|
|
|
* The object must be locked or unreferenceable.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
static void
|
|
|
|
swap_pager_dealloc(object)
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
vm_object_t object;
|
1994-05-25 09:21:21 +00:00
|
|
|
{
|
1999-09-17 05:09:24 +00:00
|
|
|
int s;
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Remove from list right away so lookups will fail if we block for
|
|
|
|
* pageout completion.
|
|
|
|
*/
|
|
|
|
|
|
|
|
if (object->handle == NULL) {
|
|
|
|
TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
|
|
|
|
} else {
|
|
|
|
TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_object_pip_wait(object, "swpdea");
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Free all remaining metadata. We only bother to free it from
|
|
|
|
* the swap meta data. We do not attempt to free swapblk's still
|
|
|
|
* associated with vm_page_t's for this object. We do not care
|
|
|
|
* if paging is still in progress on some objects.
|
|
|
|
*/
|
1999-09-17 05:09:24 +00:00
|
|
|
s = splvm();
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_pager_meta_free_all(object);
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/************************************************************************
|
|
|
|
* SWAP PAGER BITMAP ROUTINES *
|
|
|
|
************************************************************************/
|
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
|
|
|
|
*
|
|
|
|
* Allocate swap for the requested number of pages. The starting
|
|
|
|
* swap block number (a page index) is returned or SWAPBLK_NONE
|
|
|
|
* if the allocation failed.
|
|
|
|
*
|
|
|
|
* Also has the side effect of advising that somebody made a mistake
|
|
|
|
* when they configured swap and didn't configure enough.
|
|
|
|
*
|
|
|
|
* Must be called at splvm() to avoid races with bitmap frees from
|
|
|
|
* vm_page_remove() aka swap_pager_page_removed().
|
|
|
|
*
|
|
|
|
* This routine may not block
|
|
|
|
* This routine must be called at splvm().
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
static __inline daddr_t
|
|
|
|
swp_pager_getswapspace(npages)
|
|
|
|
int npages;
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
daddr_t blk;
|
1997-12-24 15:05:25 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
|
1999-02-06 07:22:21 +00:00
|
|
|
if (swap_pager_full != 2) {
|
|
|
|
printf("swap_pager_getswapspace: failed\n");
|
|
|
|
swap_pager_full = 2;
|
1999-02-21 08:30:49 +00:00
|
|
|
swap_pager_almost_full = 1;
|
1999-02-06 07:22:21 +00:00
|
|
|
}
|
1994-11-13 15:36:48 +00:00
|
|
|
} else {
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_swap_size -= npages;
|
|
|
|
swp_sizecheck();
|
1994-11-13 15:36:48 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
return(blk);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWP_PAGER_FREESWAPSPACE() - free raw swap space
|
|
|
|
*
|
|
|
|
* This routine returns the specified swap blocks back to the bitmap.
|
|
|
|
*
|
|
|
|
* Note: This routine may not block (it could in the old swap code),
|
|
|
|
* and through the use of the new blist routines it does not block.
|
|
|
|
*
|
|
|
|
* We must be called at splvm() to avoid races with bitmap frees from
|
|
|
|
* vm_page_remove() aka swap_pager_page_removed().
|
|
|
|
*
|
|
|
|
* This routine may not block
|
|
|
|
* This routine must be called at splvm().
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
static __inline void
|
|
|
|
swp_pager_freeswapspace(blk, npages)
|
|
|
|
daddr_t blk;
|
|
|
|
int npages;
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
blist_free(swapblist, blk, npages);
|
|
|
|
vm_swap_size += npages;
|
|
|
|
swp_sizecheck();
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
|
|
|
|
* range within an object.
|
|
|
|
*
|
|
|
|
* This is a globally accessible routine.
|
|
|
|
*
|
|
|
|
* This routine removes swapblk assignments from swap metadata.
|
|
|
|
*
|
|
|
|
* The external callers of this routine typically have already destroyed
|
|
|
|
* or renamed vm_page_t's associated with this range in the object so
|
|
|
|
* we should be ok.
|
1999-09-17 05:09:24 +00:00
|
|
|
*
|
|
|
|
* This routine may be called at any spl. We up our spl to splvm temporarily
|
|
|
|
* in order to perform the metadata removal.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
void
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
swap_pager_freespace(object, start, size)
|
|
|
|
vm_object_t object;
|
1995-12-11 04:58:34 +00:00
|
|
|
vm_pindex_t start;
|
|
|
|
vm_size_t size;
|
1994-05-25 09:21:21 +00:00
|
|
|
{
|
1999-09-17 05:09:24 +00:00
|
|
|
int s = splvm();
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_pager_meta_free(object, start, size);
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* SWAP_PAGER_RESERVE() - reserve swap blocks in object
|
|
|
|
*
|
|
|
|
* Assigns swap blocks to the specified range within the object. The
|
|
|
|
* swap blocks are not zerod. Any previous swap assignment is destroyed.
|
|
|
|
*
|
|
|
|
* Returns 0 on success, -1 on failure.
|
|
|
|
*/
|
|
|
|
|
|
|
|
int
|
|
|
|
swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
|
|
|
|
{
|
|
|
|
int s;
|
|
|
|
int n = 0;
|
|
|
|
daddr_t blk = SWAPBLK_NONE;
|
|
|
|
vm_pindex_t beg = start; /* save start index */
|
|
|
|
|
|
|
|
s = splvm();
|
|
|
|
while (size) {
|
|
|
|
if (n == 0) {
|
|
|
|
n = BLIST_MAX_ALLOC;
|
|
|
|
while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
|
|
|
|
n >>= 1;
|
|
|
|
if (n == 0) {
|
|
|
|
swp_pager_meta_free(object, beg, start - beg);
|
|
|
|
splx(s);
|
|
|
|
return(-1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
swp_pager_meta_build(object, start, blk);
|
|
|
|
--size;
|
|
|
|
++start;
|
|
|
|
++blk;
|
|
|
|
--n;
|
|
|
|
}
|
|
|
|
swp_pager_meta_free(object, start, n);
|
|
|
|
splx(s);
|
|
|
|
return(0);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1996-05-23 00:45:58 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
|
|
|
|
* and destroy the source.
|
|
|
|
*
|
|
|
|
* Copy any valid swapblks from the source to the destination. In
|
|
|
|
* cases where both the source and destination have a valid swapblk,
|
|
|
|
* we keep the destination's.
|
|
|
|
*
|
|
|
|
* This routine is allowed to block. It may block allocating metadata
|
|
|
|
* indirectly through swp_pager_meta_build() or if paging is still in
|
|
|
|
* progress on the source.
|
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* This routine can be called at any spl
|
|
|
|
*
|
1999-01-21 08:29:12 +00:00
|
|
|
* XXX vm_page_collapse() kinda expects us not to block because we
|
|
|
|
* supposedly do not need to allocate memory, but for the moment we
|
|
|
|
* *may* have to get a little memory from the zone allocator, but
|
|
|
|
* it is taken from the interrupt memory. We should be ok.
|
|
|
|
*
|
|
|
|
* The source object contains no vm_page_t's (which is just as well)
|
|
|
|
*
|
|
|
|
* The source object is of type OBJT_SWAP.
|
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* The source and destination objects must be locked or
|
|
|
|
* inaccessible (XXX are they ?)
|
1996-05-23 00:45:58 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1996-05-23 00:45:58 +00:00
|
|
|
void
|
1999-01-21 08:29:12 +00:00
|
|
|
swap_pager_copy(srcobject, dstobject, offset, destroysource)
|
|
|
|
vm_object_t srcobject;
|
|
|
|
vm_object_t dstobject;
|
|
|
|
vm_pindex_t offset;
|
|
|
|
int destroysource;
|
1996-05-23 00:45:58 +00:00
|
|
|
{
|
|
|
|
vm_pindex_t i;
|
1999-09-17 05:09:24 +00:00
|
|
|
int s;
|
|
|
|
|
|
|
|
s = splvm();
|
1996-05-23 00:45:58 +00:00
|
|
|
|
1995-02-02 09:09:15 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* If destroysource is set, we remove the source object from the
|
|
|
|
* swap_pager internal queue now.
|
1995-02-02 09:09:15 +00:00
|
|
|
*/
|
1998-01-17 09:17:02 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (destroysource) {
|
|
|
|
if (srcobject->handle == NULL) {
|
|
|
|
TAILQ_REMOVE(
|
|
|
|
&swap_pager_un_object_list,
|
|
|
|
srcobject,
|
|
|
|
pager_object_list
|
|
|
|
);
|
|
|
|
} else {
|
|
|
|
TAILQ_REMOVE(
|
|
|
|
NOBJLIST(srcobject->handle),
|
|
|
|
srcobject,
|
|
|
|
pager_object_list
|
|
|
|
);
|
1995-02-02 09:09:15 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* transfer source to destination.
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
for (i = 0; i < dstobject->size; ++i) {
|
|
|
|
daddr_t dstaddr;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Locate (without changing) the swapblk on the destination,
|
|
|
|
* unless it is invalid in which case free it silently, or
|
|
|
|
* if the destination is a resident page, in which case the
|
|
|
|
* source is thrown away.
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (dstaddr == SWAPBLK_NONE) {
|
|
|
|
/*
|
|
|
|
* Destination has no swapblk and is not resident,
|
|
|
|
* copy source.
|
|
|
|
*/
|
|
|
|
daddr_t srcaddr;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
srcaddr = swp_pager_meta_ctl(
|
|
|
|
srcobject,
|
|
|
|
i + offset,
|
|
|
|
SWM_POP
|
|
|
|
);
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (srcaddr != SWAPBLK_NONE)
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_meta_build(dstobject, i, srcaddr);
|
1999-01-21 08:29:12 +00:00
|
|
|
} else {
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Destination has valid swapblk or it is represented
|
|
|
|
* by a resident page. We destroy the sourceblock.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
}
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Free left over swap blocks in source.
|
|
|
|
*
|
|
|
|
* We have to revert the type to OBJT_DEFAULT so we do not accidently
|
|
|
|
* double-remove the object from the swap queues.
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
if (destroysource) {
|
|
|
|
swp_pager_meta_free_all(srcobject);
|
|
|
|
/*
|
|
|
|
* Reverting the type is not necessary, the caller is going
|
|
|
|
* to destroy srcobject directly, but I'm doing it here
|
2000-03-26 15:20:23 +00:00
|
|
|
* for consistency since we've removed the object from its
|
1999-01-21 08:29:12 +00:00
|
|
|
* queues.
|
|
|
|
*/
|
|
|
|
srcobject->type = OBJT_DEFAULT;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_HASPAGE() - determine if we have good backing store for
|
|
|
|
* the requested page.
|
|
|
|
*
|
|
|
|
* We determine whether good backing store exists for the requested
|
|
|
|
* page and return TRUE if it does, FALSE if it doesn't.
|
|
|
|
*
|
|
|
|
* If TRUE, we also try to determine how much valid, contiguous backing
|
|
|
|
* store exists before and after the requested page within a reasonable
|
|
|
|
* distance. We do not try to restrict it to the swap device stripe
|
|
|
|
* (that is handled in getpages/putpages). It probably isn't worth
|
|
|
|
* doing here.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
boolean_t
|
|
|
|
swap_pager_haspage(object, pindex, before, after)
|
|
|
|
vm_object_t object;
|
|
|
|
vm_pindex_t pindex;
|
|
|
|
int *before;
|
|
|
|
int *after;
|
1994-05-25 09:21:21 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
daddr_t blk0;
|
2000-03-27 21:33:32 +00:00
|
|
|
int s;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* do we have good backing store at the requested index ?
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
*/
|
|
|
|
|
2000-03-27 21:33:32 +00:00
|
|
|
s = splvm();
|
1999-01-21 08:29:12 +00:00
|
|
|
blk0 = swp_pager_meta_ctl(object, pindex, 0);
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (blk0 == SWAPBLK_NONE) {
|
2000-03-27 21:33:32 +00:00
|
|
|
splx(s);
|
1999-01-21 08:29:12 +00:00
|
|
|
if (before)
|
|
|
|
*before = 0;
|
|
|
|
if (after)
|
|
|
|
*after = 0;
|
|
|
|
return (FALSE);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* find backwards-looking contiguous good backing store
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
if (before != NULL) {
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 1; i < (SWB_NPAGES/2); ++i) {
|
|
|
|
daddr_t blk;
|
|
|
|
|
|
|
|
if (i > pindex)
|
|
|
|
break;
|
|
|
|
blk = swp_pager_meta_ctl(object, pindex - i, 0);
|
|
|
|
if (blk != blk0 - i)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
*before = (i - 1);
|
|
|
|
}
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* find forward-looking contiguous good backing store
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
*/
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (after != NULL) {
|
|
|
|
int i;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
for (i = 1; i < (SWB_NPAGES/2); ++i) {
|
|
|
|
daddr_t blk;
|
1995-09-04 04:44:26 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
blk = swp_pager_meta_ctl(object, pindex + i, 0);
|
|
|
|
if (blk != blk0 + i)
|
|
|
|
break;
|
1995-09-04 04:44:26 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
*after = (i - 1);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
2000-03-27 21:33:32 +00:00
|
|
|
splx(s);
|
1999-01-21 08:29:12 +00:00
|
|
|
return (TRUE);
|
1998-02-23 08:22:48 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
|
|
|
|
*
|
|
|
|
* This removes any associated swap backing store, whether valid or
|
|
|
|
* not, from the page.
|
|
|
|
*
|
|
|
|
* This routine is typically called when a page is made dirty, at
|
|
|
|
* which point any associated swap can be freed. MADV_FREE also
|
|
|
|
* calls us in a special-case situation
|
|
|
|
*
|
|
|
|
* NOTE!!! If the page is clean and the swap was valid, the caller
|
|
|
|
* should make the page dirty before calling this routine. This routine
|
|
|
|
* does NOT change the m->dirty status of the page. Also: MADV_FREE
|
|
|
|
* depends on it.
|
|
|
|
*
|
|
|
|
* This routine may not block
|
1999-09-17 05:09:24 +00:00
|
|
|
* This routine must be called at splvm()
|
1998-02-23 08:22:48 +00:00
|
|
|
*/
|
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
static void
|
1999-01-21 08:29:12 +00:00
|
|
|
swap_pager_unswapped(m)
|
|
|
|
vm_page_t m;
|
1994-05-25 09:21:21 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
|
|
|
|
1999-03-14 09:20:01 +00:00
|
|
|
/*
|
|
|
|
* SWAP_PAGER_STRATEGY() - read, write, free blocks
|
|
|
|
*
|
|
|
|
* This implements the vm_pager_strategy() interface to swap and allows
|
|
|
|
* other parts of the system to directly access swap as backing store
|
|
|
|
* through vm_objects of type OBJT_SWAP. This is intended to be a
|
|
|
|
* cacheless interface ( i.e. caching occurs at higher levels ).
|
|
|
|
* Therefore we do not maintain any resident pages. All I/O goes
|
1999-09-17 05:09:24 +00:00
|
|
|
* directly to and from the swap device.
|
1999-03-14 09:20:01 +00:00
|
|
|
*
|
|
|
|
* Note that b_blkno is scaled for PAGE_SIZE
|
|
|
|
*
|
|
|
|
* We currently attempt to run I/O synchronously or asynchronously as
|
|
|
|
* the caller requests. This isn't perfect because we loose error
|
|
|
|
* sequencing when we run multiple ops in parallel to satisfy a request.
|
|
|
|
* But this is swap, so we let it all hang out.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void
|
2000-05-03 07:47:46 +00:00
|
|
|
swap_pager_strategy(vm_object_t object, struct bio *bp)
|
1999-03-14 09:20:01 +00:00
|
|
|
{
|
|
|
|
vm_pindex_t start;
|
|
|
|
int count;
|
1999-09-17 05:09:24 +00:00
|
|
|
int s;
|
1999-03-14 09:20:01 +00:00
|
|
|
char *data;
|
|
|
|
struct buf *nbp = NULL;
|
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
/* XXX: KASSERT instead ? */
|
|
|
|
if (bp->bio_bcount & PAGE_MASK) {
|
|
|
|
bp->bio_error = EINVAL;
|
|
|
|
bp->bio_flags |= BIO_ERROR;
|
|
|
|
biodone(bp);
|
|
|
|
printf("swap_pager_strategy: bp %p blk %d size %d, not page bounded\n", bp, (int)bp->bio_pblkno, (int)bp->bio_bcount);
|
1999-03-14 09:20:01 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Clear error indication, initialize page index, count, data pointer.
|
|
|
|
*/
|
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_error = 0;
|
|
|
|
bp->bio_flags &= ~BIO_ERROR;
|
|
|
|
bp->bio_resid = bp->bio_bcount;
|
1999-03-14 09:20:01 +00:00
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
start = bp->bio_pblkno;
|
|
|
|
count = howmany(bp->bio_bcount, PAGE_SIZE);
|
|
|
|
data = bp->bio_data;
|
1999-03-14 09:20:01 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
s = splvm();
|
|
|
|
|
1999-03-14 09:20:01 +00:00
|
|
|
/*
|
2000-03-20 10:44:49 +00:00
|
|
|
* Deal with BIO_DELETE
|
1999-03-14 09:20:01 +00:00
|
|
|
*/
|
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
if (bp->bio_cmd == BIO_DELETE) {
|
1999-03-14 09:20:01 +00:00
|
|
|
/*
|
|
|
|
* FREE PAGE(s) - destroy underlying swap that is no longer
|
|
|
|
* needed.
|
|
|
|
*/
|
|
|
|
swp_pager_meta_free(object, start, count);
|
|
|
|
splx(s);
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_resid = 0;
|
|
|
|
biodone(bp);
|
1999-09-17 05:09:24 +00:00
|
|
|
return;
|
|
|
|
}
|
1999-03-14 09:20:01 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
/*
|
|
|
|
* Execute read or write
|
|
|
|
*/
|
1999-03-14 09:20:01 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
while (count > 0) {
|
|
|
|
daddr_t blk;
|
1999-03-14 09:20:01 +00:00
|
|
|
|
|
|
|
/*
|
1999-09-17 05:09:24 +00:00
|
|
|
* Obtain block. If block not found and writing, allocate a
|
|
|
|
* new block and build it into the object.
|
1999-03-14 09:20:01 +00:00
|
|
|
*/
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
blk = swp_pager_meta_ctl(object, start, 0);
|
2000-05-03 07:47:46 +00:00
|
|
|
if ((blk == SWAPBLK_NONE) && (bp->bio_cmd == BIO_WRITE)) {
|
1999-09-17 05:09:24 +00:00
|
|
|
blk = swp_pager_getswapspace(1);
|
|
|
|
if (blk == SWAPBLK_NONE) {
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_error = ENOMEM;
|
|
|
|
bp->bio_flags |= BIO_ERROR;
|
1999-03-14 09:20:01 +00:00
|
|
|
break;
|
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_meta_build(object, start, blk);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Do we have to flush our current collection? Yes if:
|
|
|
|
*
|
|
|
|
* - no swap block at this index
|
|
|
|
* - swap block is not contiguous
|
|
|
|
* - we cross a physical disk boundry in the
|
|
|
|
* stripe.
|
|
|
|
*/
|
1999-03-14 09:20:01 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (
|
|
|
|
nbp && (nbp->b_blkno + btoc(nbp->b_bcount) != blk ||
|
|
|
|
((nbp->b_blkno ^ blk) & dmmax_mask)
|
|
|
|
)
|
|
|
|
) {
|
1999-03-14 09:20:01 +00:00
|
|
|
splx(s);
|
2000-05-03 07:47:46 +00:00
|
|
|
if (bp->bio_cmd == BIO_READ) {
|
1999-09-17 05:09:24 +00:00
|
|
|
++cnt.v_swapin;
|
|
|
|
cnt.v_swappgsin += btoc(nbp->b_bcount);
|
|
|
|
} else {
|
1999-03-14 09:20:01 +00:00
|
|
|
++cnt.v_swapout;
|
|
|
|
cnt.v_swappgsout += btoc(nbp->b_bcount);
|
1999-09-17 05:09:24 +00:00
|
|
|
nbp->b_dirtyend = nbp->b_bcount;
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
flushchainbuf(nbp);
|
|
|
|
s = splvm();
|
|
|
|
nbp = NULL;
|
|
|
|
}
|
1999-03-14 09:20:01 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
/*
|
|
|
|
* Add new swapblk to nbp, instantiating nbp if necessary.
|
|
|
|
* Zero-fill reads are able to take a shortcut.
|
|
|
|
*/
|
1999-03-14 09:20:01 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (blk == SWAPBLK_NONE) {
|
1999-03-14 09:20:01 +00:00
|
|
|
/*
|
1999-09-17 05:09:24 +00:00
|
|
|
* We can only get here if we are reading. Since
|
|
|
|
* we are at splvm() we can safely modify b_resid,
|
|
|
|
* even if chain ops are in progress.
|
1999-03-14 09:20:01 +00:00
|
|
|
*/
|
1999-09-17 05:09:24 +00:00
|
|
|
bzero(data, PAGE_SIZE);
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_resid -= PAGE_SIZE;
|
1999-09-17 05:09:24 +00:00
|
|
|
} else {
|
|
|
|
if (nbp == NULL) {
|
2000-05-03 07:47:46 +00:00
|
|
|
nbp = getchainbuf(bp, swapdev_vp, B_ASYNC);
|
1999-09-17 05:09:24 +00:00
|
|
|
nbp->b_blkno = blk;
|
|
|
|
nbp->b_bcount = 0;
|
|
|
|
nbp->b_data = data;
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
nbp->b_bcount += PAGE_SIZE;
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
--count;
|
|
|
|
++start;
|
|
|
|
data += PAGE_SIZE;
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
1999-09-17 05:09:24 +00:00
|
|
|
* Flush out last buffer
|
1999-03-14 09:20:01 +00:00
|
|
|
*/
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
|
|
|
|
1999-03-14 09:20:01 +00:00
|
|
|
if (nbp) {
|
2000-03-20 10:44:49 +00:00
|
|
|
if (nbp->b_iocmd == BIO_READ) {
|
1999-03-14 09:20:01 +00:00
|
|
|
++cnt.v_swapin;
|
|
|
|
cnt.v_swappgsin += btoc(nbp->b_bcount);
|
|
|
|
} else {
|
|
|
|
++cnt.v_swapout;
|
|
|
|
cnt.v_swappgsout += btoc(nbp->b_bcount);
|
1999-09-17 05:09:24 +00:00
|
|
|
nbp->b_dirtyend = nbp->b_bcount;
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
|
|
|
flushchainbuf(nbp);
|
1999-09-17 05:09:24 +00:00
|
|
|
/* nbp = NULL; */
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Wait for completion.
|
|
|
|
*/
|
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
waitchainbuf(bp, 0, 1);
|
1999-03-14 09:20:01 +00:00
|
|
|
}
|
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWAP_PAGER_GETPAGES() - bring pages in from swap
|
|
|
|
*
|
|
|
|
* Attempt to retrieve (m, count) pages from backing store, but make
|
|
|
|
* sure we retrieve at least m[reqpage]. We try to load in as large
|
|
|
|
* a chunk surrounding m[reqpage] as is contiguous in swap and which
|
|
|
|
* belongs to the same object.
|
|
|
|
*
|
|
|
|
* The code is designed for asynchronous operation and
|
|
|
|
* immediate-notification of 'reqpage' but tends not to be
|
|
|
|
* used that way. Please do not optimize-out this algorithmic
|
|
|
|
* feature, I intend to improve on it in the future.
|
|
|
|
*
|
|
|
|
* The parent has a single vm_object_pip_add() reference prior to
|
|
|
|
* calling us and we should return with the same.
|
|
|
|
*
|
|
|
|
* The parent has BUSY'd the pages. We should return with 'm'
|
|
|
|
* left busy, but the others adjusted.
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1995-12-14 09:55:16 +00:00
|
|
|
static int
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
swap_pager_getpages(object, m, count, reqpage)
|
|
|
|
vm_object_t object;
|
1994-05-25 09:21:21 +00:00
|
|
|
vm_page_t *m;
|
|
|
|
int count, reqpage;
|
1994-05-24 10:09:53 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
struct buf *bp;
|
|
|
|
vm_page_t mreq;
|
|
|
|
int s;
|
1994-05-25 09:21:21 +00:00
|
|
|
int i;
|
1999-01-21 08:29:12 +00:00
|
|
|
int j;
|
|
|
|
daddr_t blk;
|
|
|
|
vm_offset_t kva;
|
|
|
|
vm_pindex_t lastpindex;
|
|
|
|
|
|
|
|
mreq = m[reqpage];
|
|
|
|
|
|
|
|
if (mreq->object != object) {
|
|
|
|
panic("swap_pager_getpages: object mismatch %p/%p",
|
|
|
|
object,
|
|
|
|
mreq->object
|
|
|
|
);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Calculate range to retrieve. The pages have already been assigned
|
|
|
|
* their swapblks. We require a *contiguous* range that falls entirely
|
|
|
|
* within a single device stripe. If we do not supply it, bad things
|
1999-09-17 05:09:24 +00:00
|
|
|
* happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
|
|
|
|
* loops are set up such that the case(s) are handled implicitly.
|
|
|
|
*
|
|
|
|
* The swp_*() calls must be made at splvm(). vm_page_free() does
|
|
|
|
* not need to be, but it will go a little faster if it is.
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
s = splvm();
|
1999-01-21 08:29:12 +00:00
|
|
|
blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
|
|
|
|
|
|
|
|
for (i = reqpage - 1; i >= 0; --i) {
|
|
|
|
daddr_t iblk;
|
|
|
|
|
|
|
|
iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
|
1999-09-17 05:09:24 +00:00
|
|
|
if (blk != iblk + (reqpage - i))
|
1999-01-21 08:29:12 +00:00
|
|
|
break;
|
|
|
|
if ((blk ^ iblk) & dmmax_mask)
|
|
|
|
break;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
++i;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
for (j = reqpage + 1; j < count; ++j) {
|
|
|
|
daddr_t jblk;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
|
1999-09-17 05:09:24 +00:00
|
|
|
if (blk != jblk - (j - reqpage))
|
1999-01-21 08:29:12 +00:00
|
|
|
break;
|
|
|
|
if ((blk ^ jblk) & dmmax_mask)
|
|
|
|
break;
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* free pages outside our collection range. Note: we never free
|
|
|
|
* mreq, it must remain busy throughout.
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
{
|
|
|
|
int k;
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
for (k = 0; k < i; ++k)
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_page_free(m[k]);
|
1999-09-17 05:09:24 +00:00
|
|
|
for (k = j; k < count; ++k)
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_page_free(m[k]);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-09-17 05:09:24 +00:00
|
|
|
* Return VM_PAGER_FAIL if we have nothing to do. Return mreq
|
|
|
|
* still busy, but the others unbusied.
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (blk == SWAPBLK_NONE)
|
1999-01-21 08:29:12 +00:00
|
|
|
return(VM_PAGER_FAIL);
|
|
|
|
|
1998-02-23 08:22:48 +00:00
|
|
|
/*
|
|
|
|
* Get a swap buffer header to perform the IO
|
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
bp = getpbuf(&nsw_rcount);
|
1998-02-23 08:22:48 +00:00
|
|
|
kva = (vm_offset_t) bp->b_data;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1994-08-06 09:15:42 +00:00
|
|
|
/*
|
|
|
|
* map our page(s) into kva for input
|
1999-01-21 08:29:12 +00:00
|
|
|
*
|
|
|
|
* NOTE: B_PAGING is set by pbgetvp()
|
1994-08-06 09:15:42 +00:00
|
|
|
*/
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
pmap_qenter(kva, m + i, j - i);
|
|
|
|
|
2000-03-20 10:44:49 +00:00
|
|
|
bp->b_iocmd = BIO_READ;
|
1999-01-21 08:29:12 +00:00
|
|
|
bp->b_iodone = swp_pager_async_iodone;
|
1999-05-06 20:00:34 +00:00
|
|
|
bp->b_rcred = bp->b_wcred = proc0.p_ucred;
|
1999-03-14 09:20:01 +00:00
|
|
|
bp->b_data = (caddr_t) kva;
|
1994-05-25 09:21:21 +00:00
|
|
|
crhold(bp->b_rcred);
|
|
|
|
crhold(bp->b_wcred);
|
1999-01-21 08:29:12 +00:00
|
|
|
bp->b_blkno = blk - (reqpage - i);
|
|
|
|
bp->b_bcount = PAGE_SIZE * (j - i);
|
|
|
|
bp->b_bufsize = PAGE_SIZE * (j - i);
|
|
|
|
bp->b_pager.pg_reqpage = reqpage - i;
|
|
|
|
|
|
|
|
{
|
|
|
|
int k;
|
|
|
|
|
|
|
|
for (k = i; k < j; ++k) {
|
|
|
|
bp->b_pages[k - i] = m[k];
|
|
|
|
vm_page_flag_set(m[k], PG_SWAPINPROG);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
bp->b_npages = j - i;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
pbgetvp(swapdev_vp, bp);
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1994-10-15 13:33:09 +00:00
|
|
|
cnt.v_swapin++;
|
1999-01-21 08:29:12 +00:00
|
|
|
cnt.v_swappgsin += bp->b_npages;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We still hold the lock on mreq, and our automatic completion routine
|
|
|
|
* does not remove it.
|
|
|
|
*/
|
|
|
|
|
|
|
|
vm_object_pip_add(mreq->object, bp->b_npages);
|
|
|
|
lastpindex = m[j-1]->pindex;
|
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* perform the I/O. NOTE!!! bp cannot be considered valid after
|
|
|
|
* this point because we automatically release it on completion.
|
|
|
|
* Instead, we look at the one page we are interested in which we
|
|
|
|
* still hold a lock on even through the I/O completion.
|
|
|
|
*
|
|
|
|
* The other pages in our m[] array are also released on completion,
|
|
|
|
* so we cannot assume they are valid anymore either.
|
|
|
|
*
|
1999-12-28 07:30:55 +00:00
|
|
|
* NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1999-06-27 22:08:38 +00:00
|
|
|
BUF_KERNPROC(bp);
|
2000-03-20 11:29:10 +00:00
|
|
|
BUF_STRATEGY(bp);
|
1994-05-25 09:21:21 +00:00
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* wait for the page we want to complete. PG_SWAPINPROG is always
|
|
|
|
* cleared on completion. If an I/O error occurs, SWAPBLK_NONE
|
|
|
|
* is set in the meta-data.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1998-02-23 08:22:48 +00:00
|
|
|
s = splvm();
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
while ((mreq->flags & PG_SWAPINPROG) != 0) {
|
|
|
|
vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
|
|
|
|
cnt.v_intrans++;
|
|
|
|
if (tsleep(mreq, PSWP, "swread", hz*20)) {
|
1998-07-11 07:46:16 +00:00
|
|
|
printf(
|
1999-01-21 08:29:12 +00:00
|
|
|
"swap_pager: indefinite wait buffer: device:"
|
1999-08-23 23:55:03 +00:00
|
|
|
" %s, blkno: %ld, size: %ld\n",
|
|
|
|
devtoname(bp->b_dev), (long)bp->b_blkno,
|
|
|
|
bp->b_bcount
|
1999-01-21 08:29:12 +00:00
|
|
|
);
|
1996-06-10 04:58:48 +00:00
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
splx(s);
|
|
|
|
|
1998-02-23 08:22:48 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* mreq is left bussied after completion, but all the other pages
|
|
|
|
* are freed. If we had an unrecoverable read error the page will
|
|
|
|
* not be valid.
|
1998-02-23 08:22:48 +00:00
|
|
|
*/
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (mreq->valid != VM_PAGE_BITS_ALL) {
|
|
|
|
return(VM_PAGER_ERROR);
|
1998-02-23 08:22:48 +00:00
|
|
|
} else {
|
1999-01-21 08:29:12 +00:00
|
|
|
return(VM_PAGER_OK);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* A final note: in a low swap situation, we cannot deallocate swap
|
|
|
|
* and mark a page dirty here because the caller is likely to mark
|
|
|
|
* the page clean when we return, causing the page to possibly revert
|
|
|
|
* to all-zero's later.
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* swap_pager_putpages:
|
|
|
|
*
|
|
|
|
* Assign swap (if necessary) and initiate I/O on the specified pages.
|
|
|
|
*
|
|
|
|
* We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
|
|
|
|
* are automatically converted to SWAP objects.
|
|
|
|
*
|
1999-12-28 07:30:55 +00:00
|
|
|
* In a low memory situation we may block in VOP_STRATEGY(), but the new
|
1999-01-21 08:29:12 +00:00
|
|
|
* vm_page reservation system coupled with properly written VFS devices
|
|
|
|
* should ensure that no low-memory deadlock occurs. This is an area
|
|
|
|
* which needs work.
|
|
|
|
*
|
|
|
|
* The parent has N vm_object_pip_add() references prior to
|
|
|
|
* calling us and will remove references for rtvals[] that are
|
|
|
|
* not set to VM_PAGER_PEND. We need to remove the rest on I/O
|
|
|
|
* completion.
|
|
|
|
*
|
|
|
|
* The parent has soft-busy'd the pages it passes us and will unbusy
|
|
|
|
* those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
|
|
|
|
* We need to unbusy the rest on I/O completion.
|
|
|
|
*/
|
|
|
|
|
1999-01-24 02:32:15 +00:00
|
|
|
void
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
swap_pager_putpages(object, m, count, sync, rtvals)
|
|
|
|
vm_object_t object;
|
1994-05-25 09:21:21 +00:00
|
|
|
vm_page_t *m;
|
|
|
|
int count;
|
NOTE: libkvm, w, ps, 'top', and any other utility which depends on struct
proc or any VM system structure will have to be rebuilt!!!
Much needed overhaul of the VM system. Included in this first round of
changes:
1) Improved pager interfaces: init, alloc, dealloc, getpages, putpages,
haspage, and sync operations are supported. The haspage interface now
provides information about clusterability. All pager routines now take
struct vm_object's instead of "pagers".
2) Improved data structures. In the previous paradigm, there is constant
confusion caused by pagers being both a data structure ("allocate a
pager") and a collection of routines. The idea of a pager structure has
escentially been eliminated. Objects now have types, and this type is
used to index the appropriate pager. In most cases, items in the pager
structure were duplicated in the object data structure and thus were
unnecessary. In the few cases that remained, a un_pager structure union
was created in the object to contain these items.
3) Because of the cleanup of #1 & #2, a lot of unnecessary layering can now
be removed. For instance, vm_object_enter(), vm_object_lookup(),
vm_object_remove(), and the associated object hash list were some of the
things that were removed.
4) simple_lock's removed. Discussion with several people reveals that the
SMP locking primitives used in the VM system aren't likely the mechanism
that we'll be adopting. Even if it were, the locking that was in the code
was very inadequate and would have to be mostly re-done anyway. The
locking in a uni-processor kernel was a no-op but went a long way toward
making the code difficult to read and debug.
5) Places that attempted to kludge-up the fact that we don't have kernel
thread support have been fixed to reflect the reality that we are really
dealing with processes, not threads. The VM system didn't have complete
thread support, so the comments and mis-named routines were just wrong.
We now use tsleep and wakeup directly in the lock routines, for instance.
6) Where appropriate, the pagers have been improved, especially in the
pager_alloc routines. Most of the pager_allocs have been rewritten and
are now faster and easier to maintain.
7) The pagedaemon pageout clustering algorithm has been rewritten and
now tries harder to output an even number of pages before and after
the requested page. This is sort of the reverse of the ideal pagein
algorithm and should provide better overall performance.
8) Unnecessary (incorrect) casts to caddr_t in calls to tsleep & wakeup
have been removed. Some other unnecessary casts have also been removed.
9) Some almost useless debugging code removed.
10) Terminology of shadow objects vs. backing objects straightened out.
The fact that the vm_object data structure escentially had this
backwards really confused things. The use of "shadow" and "backing
object" throughout the code is now internally consistent and correct
in the Mach terminology.
11) Several minor bug fixes, including one in the vm daemon that caused
0 RSS objects to not get purged as intended.
12) A "default pager" has now been created which cleans up the transition
of objects to the "swap" type. The previous checks throughout the code
for swp->pg_data != NULL were really ugly. This change also provides
the rudiments for future backing of "anonymous" memory by something
other than the swap pager (via the vnode pager, for example), and it
allows the decision about which of these pagers to use to be made
dynamically (although will need some additional decision code to do
this, of course).
13) (dyson) MAP_COPY has been deprecated and the corresponding "copy
object" code has been removed. MAP_COPY was undocumented and non-
standard. It was furthermore broken in several ways which caused its
behavior to degrade to MAP_PRIVATE. Binaries that use MAP_COPY will
continue to work correctly, but via the slightly different semantics
of MAP_PRIVATE.
14) (dyson) Sharing maps have been removed. It's marginal usefulness in a
threads design can be worked around in other ways. Both #12 and #13
were done to simplify the code and improve readability and maintain-
ability. (As were most all of these changes)
TODO:
1) Rewrite most of the vnode pager to use VOP_GETPAGES/PUTPAGES. Doing
this will reduce the vnode pager to a mere fraction of its current size.
2) Rewrite vm_fault and the swap/vnode pagers to use the clustering
information provided by the new haspage pager interface. This will
substantially reduce the overhead by eliminating a large number of
VOP_BMAP() calls. The VOP_BMAP() filesystem interface should be
improved to provide both a "behind" and "ahead" indication of
contiguousness.
3) Implement the extended features of pager_haspage in swap_pager_haspage().
It currently just says 0 pages ahead/behind.
4) Re-implement the swap device (swstrategy) in a more elegant way, perhaps
via a much more general mechanism that could also be used for disk
striping of regular filesystems.
5) Do something to improve the architecture of vm_object_collapse(). The
fact that it makes calls into the swap pager and knows too much about
how the swap pager operates really bothers me. It also doesn't allow
for collapsing of non-swap pager objects ("unnamed" objects backed by
other pagers).
1995-07-13 08:48:48 +00:00
|
|
|
boolean_t sync;
|
1994-05-25 09:21:21 +00:00
|
|
|
int *rtvals;
|
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
int i;
|
|
|
|
int n = 0;
|
|
|
|
|
|
|
|
if (count && m[0]->object != object) {
|
|
|
|
panic("swap_pager_getpages: object mismatch %p/%p",
|
|
|
|
object,
|
|
|
|
m[0]->object
|
|
|
|
);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Step 1
|
|
|
|
*
|
|
|
|
* Turn object into OBJT_SWAP
|
|
|
|
* check for bogus sysops
|
|
|
|
* force sync if not pageout process
|
|
|
|
*/
|
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (object->type != OBJT_SWAP)
|
|
|
|
swp_pager_meta_build(object, 0, SWAPBLK_NONE);
|
1998-02-23 08:22:48 +00:00
|
|
|
|
|
|
|
if (curproc != pageproc)
|
|
|
|
sync = TRUE;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Step 2
|
|
|
|
*
|
1999-03-14 09:20:01 +00:00
|
|
|
* Update nsw parameters from swap_async_max sysctl values.
|
1999-02-21 08:34:15 +00:00
|
|
|
* Do not let the sysop crash the machine with bogus numbers.
|
1999-02-18 19:57:33 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
if (swap_async_max != nsw_wcount_async_max) {
|
|
|
|
int n;
|
|
|
|
int s;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* limit range
|
|
|
|
*/
|
|
|
|
if ((n = swap_async_max) > nswbuf / 2)
|
|
|
|
n = nswbuf / 2;
|
|
|
|
if (n < 1)
|
|
|
|
n = 1;
|
|
|
|
swap_async_max = n;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Adjust difference ( if possible ). If the current async
|
|
|
|
* count is too low, we may not be able to make the adjustment
|
|
|
|
* at this time.
|
|
|
|
*/
|
|
|
|
s = splvm();
|
|
|
|
n -= nsw_wcount_async_max;
|
|
|
|
if (nsw_wcount_async + n >= 0) {
|
|
|
|
nsw_wcount_async += n;
|
|
|
|
nsw_wcount_async_max += n;
|
|
|
|
wakeup(&nsw_wcount_async);
|
|
|
|
}
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Step 3
|
|
|
|
*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Assign swap blocks and issue I/O. We reallocate swap on the fly.
|
|
|
|
* The page is left dirty until the pageout operation completes
|
|
|
|
* successfully.
|
|
|
|
*/
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
for (i = 0; i < count; i += n) {
|
|
|
|
int s;
|
|
|
|
int j;
|
|
|
|
struct buf *bp;
|
|
|
|
daddr_t blk;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Maximum I/O size is limited by a number of factors.
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
n = min(BLIST_MAX_ALLOC, count - i);
|
1999-02-18 19:57:33 +00:00
|
|
|
n = min(n, nsw_cluster_max);
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
s = splvm();
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Get biggest block of swap we can. If we fail, fall
|
|
|
|
* back and try to allocate a smaller block. Don't go
|
|
|
|
* overboard trying to allocate space if it would overly
|
|
|
|
* fragment swap.
|
|
|
|
*/
|
|
|
|
while (
|
|
|
|
(blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
|
|
|
|
n > 4
|
|
|
|
) {
|
|
|
|
n >>= 1;
|
|
|
|
}
|
|
|
|
if (blk == SWAPBLK_NONE) {
|
1999-09-17 05:09:24 +00:00
|
|
|
for (j = 0; j < n; ++j)
|
1999-01-21 08:29:12 +00:00
|
|
|
rtvals[i+j] = VM_PAGER_FAIL;
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
1999-01-21 08:29:12 +00:00
|
|
|
continue;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
1999-09-17 05:09:24 +00:00
|
|
|
* The I/O we are constructing cannot cross a physical
|
|
|
|
* disk boundry in the swap stripe. Note: we are still
|
|
|
|
* at splvm().
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
if ((blk ^ (blk + n)) & dmmax_mask) {
|
|
|
|
j = ((blk + dmmax) & dmmax_mask) - blk;
|
|
|
|
swp_pager_freeswapspace(blk + j, n - j);
|
|
|
|
n = j;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* All I/O parameters have been satisfied, build the I/O
|
|
|
|
* request and assign the swap space.
|
|
|
|
*
|
|
|
|
* NOTE: B_PAGING is set by pbgetvp()
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-02-18 19:57:33 +00:00
|
|
|
if (sync == TRUE) {
|
|
|
|
bp = getpbuf(&nsw_wcount_sync);
|
|
|
|
} else {
|
|
|
|
bp = getpbuf(&nsw_wcount_async);
|
2000-03-20 10:44:49 +00:00
|
|
|
bp->b_flags = B_ASYNC;
|
1999-02-18 19:57:33 +00:00
|
|
|
}
|
2000-03-22 08:40:13 +00:00
|
|
|
bp->b_iocmd = BIO_WRITE;
|
1999-01-21 08:29:12 +00:00
|
|
|
bp->b_spc = NULL; /* not used, but NULL-out anyway */
|
1998-02-03 00:50:36 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
|
1998-02-03 00:50:36 +00:00
|
|
|
|
1999-05-06 20:00:34 +00:00
|
|
|
bp->b_rcred = bp->b_wcred = proc0.p_ucred;
|
1999-01-21 08:29:12 +00:00
|
|
|
bp->b_bcount = PAGE_SIZE * n;
|
|
|
|
bp->b_bufsize = PAGE_SIZE * n;
|
|
|
|
bp->b_blkno = blk;
|
1998-02-03 00:50:36 +00:00
|
|
|
|
1999-03-14 09:20:01 +00:00
|
|
|
crhold(bp->b_rcred);
|
|
|
|
crhold(bp->b_wcred);
|
|
|
|
|
|
|
|
pbgetvp(swapdev_vp, bp);
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
for (j = 0; j < n; ++j) {
|
|
|
|
vm_page_t mreq = m[i+j];
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_pager_meta_build(
|
|
|
|
mreq->object,
|
|
|
|
mreq->pindex,
|
1999-09-17 05:09:24 +00:00
|
|
|
blk + j
|
1999-01-21 08:29:12 +00:00
|
|
|
);
|
1999-01-24 06:04:52 +00:00
|
|
|
vm_page_dirty(mreq);
|
1999-01-21 08:29:12 +00:00
|
|
|
rtvals[i+j] = VM_PAGER_OK;
|
|
|
|
|
|
|
|
vm_page_flag_set(mreq, PG_SWAPINPROG);
|
|
|
|
bp->b_pages[j] = mreq;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
bp->b_npages = n;
|
1999-03-14 09:20:01 +00:00
|
|
|
/*
|
|
|
|
* Must set dirty range for NFS to work.
|
|
|
|
*/
|
|
|
|
bp->b_dirtyoff = 0;
|
|
|
|
bp->b_dirtyend = bp->b_bcount;
|
1994-08-29 06:23:19 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
cnt.v_swapout++;
|
|
|
|
cnt.v_swappgsout += bp->b_npages;
|
|
|
|
swapdev_vp->v_numoutput++;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
splx(s);
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* asynchronous
|
|
|
|
*
|
1999-12-28 07:30:55 +00:00
|
|
|
* NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
if (sync == FALSE) {
|
|
|
|
bp->b_iodone = swp_pager_async_iodone;
|
1999-06-26 02:47:16 +00:00
|
|
|
BUF_KERNPROC(bp);
|
2000-03-20 11:29:10 +00:00
|
|
|
BUF_STRATEGY(bp);
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
for (j = 0; j < n; ++j)
|
|
|
|
rtvals[i+j] = VM_PAGER_PEND;
|
|
|
|
continue;
|
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* synchronous
|
|
|
|
*
|
1999-12-28 07:30:55 +00:00
|
|
|
* NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
bp->b_iodone = swp_pager_sync_iodone;
|
2000-03-20 11:29:10 +00:00
|
|
|
BUF_STRATEGY(bp);
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Wait for the sync I/O to complete, then update rtvals.
|
|
|
|
* We just set the rtvals[] to VM_PAGER_PEND so we can call
|
|
|
|
* our async completion routine at the end, thus avoiding a
|
|
|
|
* double-free.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-09-17 05:09:24 +00:00
|
|
|
s = splbio();
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
while ((bp->b_flags & B_DONE) == 0) {
|
|
|
|
tsleep(bp, PVM, "swwrt", 0);
|
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
for (j = 0; j < n; ++j)
|
|
|
|
rtvals[i+j] = VM_PAGER_PEND;
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* Now that we are through with the bp, we can call the
|
|
|
|
* normal async completion, which frees everything up.
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1994-10-15 13:33:09 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_pager_async_iodone(bp);
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1998-07-28 15:30:01 +00:00
|
|
|
splx(s);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* swap_pager_sync_iodone:
|
|
|
|
*
|
|
|
|
* Completion routine for synchronous reads and writes from/to swap.
|
|
|
|
* We just mark the bp is complete and wake up anyone waiting on it.
|
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* This routine may not block. This routine is called at splbio() or better.
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
static void
|
|
|
|
swp_pager_sync_iodone(bp)
|
|
|
|
struct buf *bp;
|
|
|
|
{
|
|
|
|
bp->b_flags |= B_DONE;
|
|
|
|
bp->b_flags &= ~B_ASYNC;
|
|
|
|
wakeup(bp);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* swp_pager_async_iodone:
|
|
|
|
*
|
|
|
|
* Completion routine for asynchronous reads and writes from/to swap.
|
|
|
|
* Also called manually by synchronous code to finish up a bp.
|
|
|
|
*
|
|
|
|
* For READ operations, the pages are PG_BUSY'd. For WRITE operations,
|
|
|
|
* the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
|
|
|
|
* unbusy all pages except the 'main' request page. For WRITE
|
|
|
|
* operations, we vm_page_t->busy'd unbusy all pages ( we can do this
|
|
|
|
* because we marked them all VM_PAGER_PEND on return from putpages ).
|
|
|
|
*
|
|
|
|
* This routine may not block.
|
1999-09-17 05:09:24 +00:00
|
|
|
* This routine is called at splbio() or better
|
|
|
|
*
|
|
|
|
* We up ourselves to splvm() as required for various vm_page related
|
|
|
|
* calls.
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
static void
|
|
|
|
swp_pager_async_iodone(bp)
|
|
|
|
register struct buf *bp;
|
|
|
|
{
|
|
|
|
int s;
|
|
|
|
int i;
|
|
|
|
vm_object_t object = NULL;
|
|
|
|
|
|
|
|
bp->b_flags |= B_DONE;
|
|
|
|
|
1994-05-24 10:09:53 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* report error
|
1994-05-24 10:09:53 +00:00
|
|
|
*/
|
1998-02-23 08:22:48 +00:00
|
|
|
|
2000-04-02 15:24:56 +00:00
|
|
|
if (bp->b_ioflags & BIO_ERROR) {
|
1998-07-11 07:46:16 +00:00
|
|
|
printf(
|
1999-01-21 08:29:12 +00:00
|
|
|
"swap_pager: I/O error - %s failed; blkno %ld,"
|
|
|
|
"size %ld, error %d\n",
|
2000-03-20 10:44:49 +00:00
|
|
|
((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
|
1999-01-21 08:29:12 +00:00
|
|
|
(long)bp->b_blkno,
|
|
|
|
(long)bp->b_bcount,
|
|
|
|
bp->b_error
|
|
|
|
);
|
1994-10-25 07:06:20 +00:00
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
1999-09-17 05:09:24 +00:00
|
|
|
* set object, raise to splvm().
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (bp->b_npages)
|
|
|
|
object = bp->b_pages[0]->object;
|
1999-09-17 05:09:24 +00:00
|
|
|
s = splvm();
|
1994-05-25 09:21:21 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* remove the mapping for kernel virtual
|
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
|
1994-05-25 09:21:21 +00:00
|
|
|
|
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* cleanup pages. If an error occurs writing to swap, we are in
|
|
|
|
* very serious trouble. If it happens to be a disk error, though,
|
|
|
|
* we may be able to recover by reassigning the swap later on. So
|
|
|
|
* in this case we remove the m->swapblk assignment for the page
|
|
|
|
* but do not free it in the rlist. The errornous block(s) are thus
|
|
|
|
* never reallocated as swap. Redirty the page and continue.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
for (i = 0; i < bp->b_npages; ++i) {
|
|
|
|
vm_page_t m = bp->b_pages[i];
|
|
|
|
|
|
|
|
vm_page_flag_clear(m, PG_SWAPINPROG);
|
|
|
|
|
2000-04-02 15:24:56 +00:00
|
|
|
if (bp->b_ioflags & BIO_ERROR) {
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* If an error occurs I'd love to throw the swapblk
|
|
|
|
* away without freeing it back to swapspace, so it
|
|
|
|
* can never be used again. But I can't from an
|
|
|
|
* interrupt.
|
|
|
|
*/
|
|
|
|
|
2000-03-20 10:44:49 +00:00
|
|
|
if (bp->b_iocmd == BIO_READ) {
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* When reading, reqpage needs to stay
|
|
|
|
* locked for the parent, but all other
|
|
|
|
* pages can be freed. We still want to
|
|
|
|
* wakeup the parent waiting on the page,
|
|
|
|
* though. ( also: pg_reqpage can be -1 and
|
|
|
|
* not match anything ).
|
|
|
|
*
|
|
|
|
* We have to wake specifically requested pages
|
|
|
|
* up too because we cleared PG_SWAPINPROG and
|
|
|
|
* someone may be waiting for that.
|
|
|
|
*
|
|
|
|
* NOTE: for reads, m->dirty will probably
|
2000-03-26 15:20:23 +00:00
|
|
|
* be overridden by the original caller of
|
1999-01-21 08:29:12 +00:00
|
|
|
* getpages so don't play cute tricks here.
|
|
|
|
*
|
2000-11-18 21:01:04 +00:00
|
|
|
* XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
|
|
|
|
* AS THIS MESSES WITH object->memq, and it is
|
|
|
|
* not legal to mess with object->memq from an
|
|
|
|
* interrupt.
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
m->valid = 0;
|
|
|
|
vm_page_flag_clear(m, PG_ZERO);
|
|
|
|
|
|
|
|
if (i != bp->b_pager.pg_reqpage)
|
|
|
|
vm_page_free(m);
|
|
|
|
else
|
|
|
|
vm_page_flash(m);
|
|
|
|
/*
|
|
|
|
* If i == bp->b_pager.pg_reqpage, do not wake
|
|
|
|
* the page up. The caller needs to.
|
|
|
|
*/
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* If a write error occurs, reactivate page
|
|
|
|
* so it doesn't clog the inactive list,
|
|
|
|
* then finish the I/O.
|
|
|
|
*/
|
1999-01-24 06:04:52 +00:00
|
|
|
vm_page_dirty(m);
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_page_activate(m);
|
|
|
|
vm_page_io_finish(m);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
2000-03-20 10:44:49 +00:00
|
|
|
} else if (bp->b_iocmd == BIO_READ) {
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* For read success, clear dirty bits. Nobody should
|
|
|
|
* have this page mapped but don't take any chances,
|
|
|
|
* make sure the pmap modify bits are also cleared.
|
|
|
|
*
|
|
|
|
* NOTE: for reads, m->dirty will probably be
|
2000-03-26 15:20:23 +00:00
|
|
|
* overridden by the original caller of getpages so
|
1999-01-21 08:29:12 +00:00
|
|
|
* we cannot set them in order to free the underlying
|
|
|
|
* swap in a low-swap situation. I don't think we'd
|
|
|
|
* want to do that anyway, but it was an optimization
|
|
|
|
* that existed in the old swapper for a time before
|
|
|
|
* it got ripped out due to precisely this problem.
|
|
|
|
*
|
|
|
|
* clear PG_ZERO in page.
|
|
|
|
*
|
|
|
|
* If not the requested page then deactivate it.
|
|
|
|
*
|
|
|
|
* Note that the requested page, reqpage, is left
|
|
|
|
* busied, but we still have to wake it up. The
|
|
|
|
* other pages are released (unbusied) by
|
|
|
|
* vm_page_wakeup(). We do not set reqpage's
|
|
|
|
* valid bits here, it is up to the caller.
|
|
|
|
*/
|
|
|
|
|
2000-05-21 12:50:18 +00:00
|
|
|
pmap_clear_modify(m);
|
1999-01-21 08:29:12 +00:00
|
|
|
m->valid = VM_PAGE_BITS_ALL;
|
1999-08-17 04:02:34 +00:00
|
|
|
vm_page_undirty(m);
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_page_flag_clear(m, PG_ZERO);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We have to wake specifically requested pages
|
|
|
|
* up too because we cleared PG_SWAPINPROG and
|
|
|
|
* could be waiting for it in getpages. However,
|
|
|
|
* be sure to not unbusy getpages specifically
|
|
|
|
* requested page - getpages expects it to be
|
|
|
|
* left busy.
|
|
|
|
*/
|
|
|
|
if (i != bp->b_pager.pg_reqpage) {
|
|
|
|
vm_page_deactivate(m);
|
|
|
|
vm_page_wakeup(m);
|
|
|
|
} else {
|
|
|
|
vm_page_flash(m);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* For write success, clear the modify and dirty
|
|
|
|
* status, then finish the I/O ( which decrements the
|
|
|
|
* busy count and possibly wakes waiter's up ).
|
|
|
|
*/
|
|
|
|
vm_page_protect(m, VM_PROT_READ);
|
2000-05-21 12:50:18 +00:00
|
|
|
pmap_clear_modify(m);
|
1999-08-17 05:56:00 +00:00
|
|
|
vm_page_undirty(m);
|
1999-01-21 08:29:12 +00:00
|
|
|
vm_page_io_finish(m);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* adjust pip. NOTE: the original parent may still have its own
|
|
|
|
* pip refs on the object.
|
|
|
|
*/
|
|
|
|
|
|
|
|
if (object)
|
|
|
|
vm_object_pip_wakeupn(object, bp->b_npages);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* release the physical I/O buffer
|
|
|
|
*/
|
|
|
|
|
1999-02-18 19:57:33 +00:00
|
|
|
relpbuf(
|
|
|
|
bp,
|
2000-03-20 10:44:49 +00:00
|
|
|
((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
|
1999-02-18 19:57:33 +00:00
|
|
|
((bp->b_flags & B_ASYNC) ?
|
|
|
|
&nsw_wcount_async :
|
|
|
|
&nsw_wcount_sync
|
|
|
|
)
|
|
|
|
)
|
|
|
|
);
|
1999-01-21 08:29:12 +00:00
|
|
|
splx(s);
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/************************************************************************
|
|
|
|
* SWAP META DATA *
|
|
|
|
************************************************************************
|
|
|
|
*
|
|
|
|
* These routines manipulate the swap metadata stored in the
|
1999-09-17 05:09:24 +00:00
|
|
|
* OBJT_SWAP object. All swp_*() routines must be called at
|
|
|
|
* splvm() because swap can be freed up by the low level vm_page
|
|
|
|
* code which might be called from interrupts beyond what splbio() covers.
|
1999-01-21 08:29:12 +00:00
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* Swap metadata is implemented with a global hash and not directly
|
|
|
|
* linked into the object. Instead the object simply contains
|
|
|
|
* appropriate tracking counters.
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* SWP_PAGER_HASH() - hash swap meta data
|
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* This is an inline helper function which hashes the swapblk given
|
1999-01-21 08:29:12 +00:00
|
|
|
* the object and page index. It returns a pointer to a pointer
|
|
|
|
* to the object, or a pointer to a NULL pointer if it could not
|
|
|
|
* find a swapblk.
|
1999-09-17 05:09:24 +00:00
|
|
|
*
|
|
|
|
* This routine must be called at splvm().
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
static __inline struct swblock **
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_hash(vm_object_t object, vm_pindex_t index)
|
1994-05-24 10:09:53 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
struct swblock **pswap;
|
|
|
|
struct swblock *swap;
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
index &= ~SWAP_META_MASK;
|
1999-08-23 23:55:03 +00:00
|
|
|
pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
|
1999-01-21 08:29:12 +00:00
|
|
|
|
|
|
|
while ((swap = *pswap) != NULL) {
|
|
|
|
if (swap->swb_object == object &&
|
|
|
|
swap->swb_index == index
|
|
|
|
) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
pswap = &swap->swb_hnext;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
return(pswap);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
|
|
|
|
*
|
|
|
|
* We first convert the object to a swap object if it is a default
|
|
|
|
* object.
|
|
|
|
*
|
|
|
|
* The specified swapblk is added to the object's swap metadata. If
|
|
|
|
* the swapblk is not valid, it is freed instead. Any previously
|
|
|
|
* assigned swapblk is freed.
|
1999-09-17 05:09:24 +00:00
|
|
|
*
|
|
|
|
* This routine must be called at splvm(), except when used to convert
|
|
|
|
* an OBJT_DEFAULT object into an OBJT_SWAP object.
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
|
|
|
|
1998-04-15 17:47:40 +00:00
|
|
|
static void
|
1999-01-21 08:29:12 +00:00
|
|
|
swp_pager_meta_build(
|
|
|
|
vm_object_t object,
|
1999-09-17 05:09:24 +00:00
|
|
|
vm_pindex_t index,
|
|
|
|
daddr_t swapblk
|
1999-01-21 08:29:12 +00:00
|
|
|
) {
|
|
|
|
struct swblock *swap;
|
|
|
|
struct swblock **pswap;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Convert default object to swap object if necessary
|
|
|
|
*/
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (object->type != OBJT_SWAP) {
|
|
|
|
object->type = OBJT_SWAP;
|
|
|
|
object->un_pager.swp.swp_bcount = 0;
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (object->handle != NULL) {
|
|
|
|
TAILQ_INSERT_TAIL(
|
|
|
|
NOBJLIST(object->handle),
|
|
|
|
object,
|
|
|
|
pager_object_list
|
|
|
|
);
|
|
|
|
} else {
|
|
|
|
TAILQ_INSERT_TAIL(
|
|
|
|
&swap_pager_un_object_list,
|
|
|
|
object,
|
|
|
|
pager_object_list
|
|
|
|
);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Locate hash entry. If not found create, but if we aren't adding
|
1999-09-17 05:09:24 +00:00
|
|
|
* anything just return. If we run out of space in the map we wait
|
|
|
|
* and, since the hash table may have changed, retry.
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
retry:
|
1999-01-21 08:29:12 +00:00
|
|
|
pswap = swp_pager_hash(object, index);
|
|
|
|
|
|
|
|
if ((swap = *pswap) == NULL) {
|
|
|
|
int i;
|
|
|
|
|
|
|
|
if (swapblk == SWAPBLK_NONE)
|
|
|
|
return;
|
|
|
|
|
|
|
|
swap = *pswap = zalloc(swap_zone);
|
1999-09-17 05:09:24 +00:00
|
|
|
if (swap == NULL) {
|
|
|
|
VM_WAIT;
|
|
|
|
goto retry;
|
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
swap->swb_hnext = NULL;
|
|
|
|
swap->swb_object = object;
|
|
|
|
swap->swb_index = index & ~SWAP_META_MASK;
|
|
|
|
swap->swb_count = 0;
|
|
|
|
|
|
|
|
++object->un_pager.swp.swp_bcount;
|
|
|
|
|
|
|
|
for (i = 0; i < SWAP_META_PAGES; ++i)
|
|
|
|
swap->swb_pages[i] = SWAPBLK_NONE;
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* Delete prior contents of metadata
|
|
|
|
*/
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
index &= SWAP_META_MASK;
|
|
|
|
|
|
|
|
if (swap->swb_pages[index] != SWAPBLK_NONE) {
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_freeswapspace(swap->swb_pages[index], 1);
|
1999-01-21 08:29:12 +00:00
|
|
|
--swap->swb_count;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Enter block into metadata
|
|
|
|
*/
|
|
|
|
|
|
|
|
swap->swb_pages[index] = swapblk;
|
1999-09-17 05:09:24 +00:00
|
|
|
if (swapblk != SWAPBLK_NONE)
|
|
|
|
++swap->swb_count;
|
1994-05-24 10:09:53 +00:00
|
|
|
}
|
|
|
|
|
1994-05-25 09:21:21 +00:00
|
|
|
/*
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
|
|
|
|
*
|
|
|
|
* The requested range of blocks is freed, with any associated swap
|
|
|
|
* returned to the swap bitmap.
|
|
|
|
*
|
|
|
|
* This routine will free swap metadata structures as they are cleaned
|
|
|
|
* out. This routine does *NOT* operate on swap metadata associated
|
|
|
|
* with resident pages.
|
|
|
|
*
|
|
|
|
* This routine must be called at splvm()
|
1994-05-25 09:21:21 +00:00
|
|
|
*/
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1995-11-14 20:53:20 +00:00
|
|
|
static void
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
|
1994-05-24 10:09:53 +00:00
|
|
|
{
|
1999-01-21 08:29:12 +00:00
|
|
|
if (object->type != OBJT_SWAP)
|
|
|
|
return;
|
1998-03-01 04:18:54 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
while (count > 0) {
|
|
|
|
struct swblock **pswap;
|
|
|
|
struct swblock *swap;
|
1994-05-24 10:09:53 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
pswap = swp_pager_hash(object, index);
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if ((swap = *pswap) != NULL) {
|
|
|
|
daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (v != SWAPBLK_NONE) {
|
|
|
|
swp_pager_freeswapspace(v, 1);
|
|
|
|
swap->swb_pages[index & SWAP_META_MASK] =
|
|
|
|
SWAPBLK_NONE;
|
|
|
|
if (--swap->swb_count == 0) {
|
|
|
|
*pswap = swap->swb_hnext;
|
|
|
|
zfree(swap_zone, swap);
|
|
|
|
--object->un_pager.swp.swp_bcount;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
--count;
|
|
|
|
++index;
|
|
|
|
} else {
|
1999-09-17 05:09:24 +00:00
|
|
|
int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
|
1999-01-21 08:29:12 +00:00
|
|
|
count -= n;
|
|
|
|
index += n;
|
1998-03-01 04:18:54 +00:00
|
|
|
}
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
|
|
|
|
*
|
|
|
|
* This routine locates and destroys all swap metadata associated with
|
|
|
|
* an object.
|
1999-09-17 05:09:24 +00:00
|
|
|
*
|
|
|
|
* This routine must be called at splvm()
|
1999-01-21 08:29:12 +00:00
|
|
|
*/
|
These changes embody the support of the fully coherent merged VM buffer cache,
much higher filesystem I/O performance, and much better paging performance. It
represents the culmination of over 6 months of R&D.
The majority of the merged VM/cache work is by John Dyson.
The following highlights the most significant changes. Additionally, there are
(mostly minor) changes to the various filesystem modules (nfs, msdosfs, etc) to
support the new VM/buffer scheme.
vfs_bio.c:
Significant rewrite of most of vfs_bio to support the merged VM buffer cache
scheme. The scheme is almost fully compatible with the old filesystem
interface. Significant improvement in the number of opportunities for write
clustering.
vfs_cluster.c, vfs_subr.c
Upgrade and performance enhancements in vfs layer code to support merged
VM/buffer cache. Fixup of vfs_cluster to eliminate the bogus pagemove stuff.
vm_object.c:
Yet more improvements in the collapse code. Elimination of some windows that
can cause list corruption.
vm_pageout.c:
Fixed it, it really works better now. Somehow in 2.0, some "enhancements"
broke the code. This code has been reworked from the ground-up.
vm_fault.c, vm_page.c, pmap.c, vm_object.c
Support for small-block filesystems with merged VM/buffer cache scheme.
pmap.c vm_map.c
Dynamic kernel VM size, now we dont have to pre-allocate excessive numbers of
kernel PTs.
vm_glue.c
Much simpler and more effective swapping code. No more gratuitous swapping.
proc.h
Fixed the problem that the p_lock flag was not being cleared on a fork.
swap_pager.c, vnode_pager.c
Removal of old vfs_bio cruft to support the past pseudo-coherency. Now the
code doesn't need it anymore.
machdep.c
Changes to better support the parameter values for the merged VM/buffer cache
scheme.
machdep.c, kern_exec.c, vm_glue.c
Implemented a seperate submap for temporary exec string space and another one
to contain process upages. This eliminates all map fragmentation problems
that previously existed.
ffs_inode.c, ufs_inode.c, ufs_readwrite.c
Changes for merged VM/buffer cache. Add "bypass" support for sneaking in on
busy buffers.
Submitted by: John Dyson and David Greenman
1995-01-09 16:06:02 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
static void
|
|
|
|
swp_pager_meta_free_all(vm_object_t object)
|
|
|
|
{
|
|
|
|
daddr_t index = 0;
|
1994-05-25 09:21:21 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
if (object->type != OBJT_SWAP)
|
|
|
|
return;
|
|
|
|
|
|
|
|
while (object->un_pager.swp.swp_bcount) {
|
|
|
|
struct swblock **pswap;
|
|
|
|
struct swblock *swap;
|
1995-02-02 09:09:15 +00:00
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
pswap = swp_pager_hash(object, index);
|
|
|
|
if ((swap = *pswap) != NULL) {
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < SWAP_META_PAGES; ++i) {
|
|
|
|
daddr_t v = swap->swb_pages[i];
|
|
|
|
if (v != SWAPBLK_NONE) {
|
|
|
|
--swap->swb_count;
|
1999-09-17 05:09:24 +00:00
|
|
|
swp_pager_freeswapspace(v, 1);
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
if (swap->swb_count != 0)
|
|
|
|
panic("swap_pager_meta_free_all: swb_count != 0");
|
|
|
|
*pswap = swap->swb_hnext;
|
|
|
|
zfree(swap_zone, swap);
|
|
|
|
--object->un_pager.swp.swp_bcount;
|
|
|
|
}
|
|
|
|
index += SWAP_META_PAGES;
|
|
|
|
if (index > 0x20000000)
|
|
|
|
panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
|
|
|
|
*
|
|
|
|
* This routine is capable of looking up, popping, or freeing
|
|
|
|
* swapblk assignments in the swap meta data or in the vm_page_t.
|
|
|
|
* The routine typically returns the swapblk being looked-up, or popped,
|
|
|
|
* or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
|
|
|
|
* was invalid. This routine will automatically free any invalid
|
|
|
|
* meta-data swapblks.
|
|
|
|
*
|
|
|
|
* It is not possible to store invalid swapblks in the swap meta data
|
|
|
|
* (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
|
|
|
|
*
|
|
|
|
* When acting on a busy resident page and paging is in progress, we
|
|
|
|
* have to wait until paging is complete but otherwise can act on the
|
|
|
|
* busy page.
|
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* This routine must be called at splvm().
|
1999-01-21 08:29:12 +00:00
|
|
|
*
|
1999-09-17 05:09:24 +00:00
|
|
|
* SWM_FREE remove and free swap block from metadata
|
1999-01-21 08:29:12 +00:00
|
|
|
* SWM_POP remove from meta data but do not free.. pop it out
|
|
|
|
*/
|
|
|
|
|
|
|
|
static daddr_t
|
|
|
|
swp_pager_meta_ctl(
|
|
|
|
vm_object_t object,
|
|
|
|
vm_pindex_t index,
|
|
|
|
int flags
|
|
|
|
) {
|
1999-09-17 05:09:24 +00:00
|
|
|
struct swblock **pswap;
|
|
|
|
struct swblock *swap;
|
|
|
|
daddr_t r1;
|
|
|
|
|
1999-01-21 08:29:12 +00:00
|
|
|
/*
|
|
|
|
* The meta data only exists of the object is OBJT_SWAP
|
|
|
|
* and even then might not be allocated yet.
|
|
|
|
*/
|
1995-02-02 09:09:15 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (object->type != OBJT_SWAP)
|
1999-01-21 08:29:12 +00:00
|
|
|
return(SWAPBLK_NONE);
|
1998-02-23 08:22:48 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
r1 = SWAPBLK_NONE;
|
|
|
|
pswap = swp_pager_hash(object, index);
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if ((swap = *pswap) != NULL) {
|
1999-01-21 08:29:12 +00:00
|
|
|
index &= SWAP_META_MASK;
|
1999-09-17 05:09:24 +00:00
|
|
|
r1 = swap->swb_pages[index];
|
1999-01-21 08:29:12 +00:00
|
|
|
|
1999-09-17 05:09:24 +00:00
|
|
|
if (r1 != SWAPBLK_NONE) {
|
|
|
|
if (flags & SWM_FREE) {
|
|
|
|
swp_pager_freeswapspace(r1, 1);
|
|
|
|
r1 = SWAPBLK_NONE;
|
|
|
|
}
|
|
|
|
if (flags & (SWM_FREE|SWM_POP)) {
|
|
|
|
swap->swb_pages[index] = SWAPBLK_NONE;
|
|
|
|
if (--swap->swb_count == 0) {
|
|
|
|
*pswap = swap->swb_hnext;
|
|
|
|
zfree(swap_zone, swap);
|
|
|
|
--object->un_pager.swp.swp_bcount;
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
}
|
|
|
|
}
|
1999-09-17 05:09:24 +00:00
|
|
|
return(r1);
|
1994-05-25 09:21:21 +00:00
|
|
|
}
|
1999-01-21 08:29:12 +00:00
|
|
|
|
2000-05-01 19:38:51 +00:00
|
|
|
/********************************************************
|
|
|
|
* CHAINING FUNCTIONS *
|
|
|
|
********************************************************
|
|
|
|
*
|
|
|
|
* These functions support recursion of I/O operations
|
|
|
|
* on bp's, typically by chaining one or more 'child' bp's
|
|
|
|
* to the parent. Synchronous, asynchronous, and semi-synchronous
|
|
|
|
* chaining is possible.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* vm_pager_chain_iodone:
|
|
|
|
*
|
|
|
|
* io completion routine for child bp. Currently we fudge a bit
|
|
|
|
* on dealing with b_resid. Since users of these routines may issue
|
|
|
|
* multiple children simultaneously, sequencing of the error can be lost.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void
|
|
|
|
vm_pager_chain_iodone(struct buf *nbp)
|
|
|
|
{
|
2000-05-03 07:47:46 +00:00
|
|
|
struct bio *bp;
|
|
|
|
u_int *count;
|
2000-05-01 19:38:51 +00:00
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
bp = nbp->b_caller1;
|
|
|
|
count = (u_int *)&(bp->bio_caller1);
|
|
|
|
if (bp != NULL) {
|
2000-05-01 19:38:51 +00:00
|
|
|
if (nbp->b_ioflags & BIO_ERROR) {
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_flags |= BIO_ERROR;
|
|
|
|
bp->bio_error = nbp->b_error;
|
2000-05-01 19:38:51 +00:00
|
|
|
} else if (nbp->b_resid != 0) {
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_flags |= BIO_ERROR;
|
|
|
|
bp->bio_error = EINVAL;
|
2000-05-01 19:38:51 +00:00
|
|
|
} else {
|
2000-05-03 07:47:46 +00:00
|
|
|
bp->bio_resid -= nbp->b_bcount;
|
2000-05-01 19:38:51 +00:00
|
|
|
}
|
2000-05-03 07:47:46 +00:00
|
|
|
nbp->b_caller1 = NULL;
|
|
|
|
--(*count);
|
|
|
|
if (bp->bio_flags & BIO_FLAG1) {
|
|
|
|
bp->bio_flags &= ~BIO_FLAG1;
|
2000-05-01 19:38:51 +00:00
|
|
|
wakeup(bp);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
nbp->b_flags |= B_DONE;
|
|
|
|
nbp->b_flags &= ~B_ASYNC;
|
|
|
|
relpbuf(nbp, NULL);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* getchainbuf:
|
|
|
|
*
|
|
|
|
* Obtain a physical buffer and chain it to its parent buffer. When
|
|
|
|
* I/O completes, the parent buffer will be B_SIGNAL'd. Errors are
|
|
|
|
* automatically propagated to the parent
|
|
|
|
*/
|
|
|
|
|
|
|
|
struct buf *
|
2000-05-03 07:47:46 +00:00
|
|
|
getchainbuf(struct bio *bp, struct vnode *vp, int flags)
|
2000-05-01 19:38:51 +00:00
|
|
|
{
|
|
|
|
struct buf *nbp = getpbuf(NULL);
|
2000-05-03 07:47:46 +00:00
|
|
|
u_int *count = (u_int *)&(bp->bio_caller1);
|
2000-05-01 19:38:51 +00:00
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
nbp->b_caller1 = bp;
|
|
|
|
++(*count);
|
2000-05-01 19:38:51 +00:00
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
if (*count > 4)
|
2000-05-01 19:38:51 +00:00
|
|
|
waitchainbuf(bp, 4, 0);
|
|
|
|
|
2000-05-03 07:47:46 +00:00
|
|
|
nbp->b_iocmd = bp->bio_cmd;
|
|
|
|
nbp->b_ioflags = bp->bio_flags & BIO_ORDERED;
|
2000-05-01 19:38:51 +00:00
|
|
|
nbp->b_flags = flags;
|
|
|
|
nbp->b_rcred = nbp->b_wcred = proc0.p_ucred;
|
|
|
|
nbp->b_iodone = vm_pager_chain_iodone;
|
|
|
|
|
|
|
|
crhold(nbp->b_rcred);
|
|
|
|
crhold(nbp->b_wcred);
|
|
|
|
|
|
|
|
if (vp)
|
|
|
|
pbgetvp(vp, nbp);
|
|
|
|
return(nbp);
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
flushchainbuf(struct buf *nbp)
|
|
|
|
{
|
|
|
|
if (nbp->b_bcount) {
|
|
|
|
nbp->b_bufsize = nbp->b_bcount;
|
|
|
|
if (nbp->b_iocmd == BIO_WRITE)
|
|
|
|
nbp->b_dirtyend = nbp->b_bcount;
|
|
|
|
BUF_KERNPROC(nbp);
|
|
|
|
BUF_STRATEGY(nbp);
|
|
|
|
} else {
|
|
|
|
bufdone(nbp);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
2000-05-03 07:47:46 +00:00
|
|
|
waitchainbuf(struct bio *bp, int limit, int done)
|
2000-05-01 19:38:51 +00:00
|
|
|
{
|
|
|
|
int s;
|
2000-05-03 07:47:46 +00:00
|
|
|
u_int *count = (u_int *)&(bp->bio_caller1);
|
2000-05-01 19:38:51 +00:00
|
|
|
|
|
|
|
s = splbio();
|
2000-05-03 07:47:46 +00:00
|
|
|
while (*count > limit) {
|
|
|
|
bp->bio_flags |= BIO_FLAG1;
|
2000-05-01 19:38:51 +00:00
|
|
|
tsleep(bp, PRIBIO + 4, "bpchain", 0);
|
|
|
|
}
|
|
|
|
if (done) {
|
2000-05-03 07:47:46 +00:00
|
|
|
if (bp->bio_resid != 0 && !(bp->bio_flags & BIO_ERROR)) {
|
|
|
|
bp->bio_flags |= BIO_ERROR;
|
|
|
|
bp->bio_error = EINVAL;
|
2000-05-01 19:38:51 +00:00
|
|
|
}
|
2000-05-03 07:47:46 +00:00
|
|
|
biodone(bp);
|
2000-05-01 19:38:51 +00:00
|
|
|
}
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|