freebsd-dev/sys/kern/kern_descrip.c

3895 lines
89 KiB
C
Raw Normal View History

/*-
1994-05-24 10:09:53 +00:00
* Copyright (c) 1982, 1986, 1989, 1991, 1993
* The Regents of the University of California. All rights reserved.
* (c) UNIX System Laboratories, Inc.
* All or some portions of this file are derived from material licensed
* to the University of California by American Telephone and Telegraph
* Co. or Unix System Laboratories, Inc. and are reproduced herein with
* the permission of UNIX System Laboratories, Inc.
*
* 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.
* 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.
*
* @(#)kern_descrip.c 8.6 (Berkeley) 4/19/94
*/
2003-06-11 00:56:59 +00:00
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_capsicum.h"
#include "opt_compat.h"
#include "opt_ddb.h"
#include "opt_ktrace.h"
#include "opt_procdesc.h"
1994-05-24 10:09:53 +00:00
#include <sys/param.h>
#include <sys/systm.h>
2004-12-03 21:29:25 +00:00
#include <sys/capability.h>
#include <sys/conf.h>
#include <sys/domain.h>
2004-12-03 21:29:25 +00:00
#include <sys/fcntl.h>
#include <sys/file.h>
1994-05-24 10:09:53 +00:00
#include <sys/filedesc.h>
2004-12-03 21:29:25 +00:00
#include <sys/filio.h>
#include <sys/jail.h>
1994-05-24 10:09:53 +00:00
#include <sys/kernel.h>
#include <sys/limits.h>
2004-12-03 21:29:25 +00:00
#include <sys/lock.h>
2002-10-16 15:14:31 +00:00
#include <sys/malloc.h>
#include <sys/mman.h>
#include <sys/mount.h>
#include <sys/mqueue.h>
2004-12-03 21:29:25 +00:00
#include <sys/mutex.h>
#include <sys/namei.h>
#include <sys/selinfo.h>
#include <sys/pipe.h>
#include <sys/priv.h>
2004-12-03 21:29:25 +00:00
#include <sys/proc.h>
#include <sys/procdesc.h>
#include <sys/protosw.h>
#include <sys/racct.h>
1994-05-24 10:09:53 +00:00
#include <sys/resourcevar.h>
#include <sys/signalvar.h>
2004-12-03 21:29:25 +00:00
#include <sys/socketvar.h>
#include <sys/stat.h>
#include <sys/sx.h>
#include <sys/syscallsubr.h>
#include <sys/sysctl.h>
#include <sys/sysproto.h>
Integrate the new MPSAFE TTY layer to the FreeBSD operating system. The last half year I've been working on a replacement TTY layer for the FreeBSD kernel. The new TTY layer was designed to improve the following: - Improved driver model: The old TTY layer has a driver model that is not abstract enough to make it friendly to use. A good example is the output path, where the device drivers directly access the output buffers. This means that an in-kernel PPP implementation must always convert network buffers into TTY buffers. If a PPP implementation would be built on top of the new TTY layer (still needs a hooks layer, though), it would allow the PPP implementation to directly hand the data to the TTY driver. - Improved hotplugging: With the old TTY layer, it isn't entirely safe to destroy TTY's from the system. This implementation has a two-step destructing design, where the driver first abandons the TTY. After all threads have left the TTY, the TTY layer calls a routine in the driver, which can be used to free resources (unit numbers, etc). The pts(4) driver also implements this feature, which means posix_openpt() will now return PTY's that are created on the fly. - Improved performance: One of the major improvements is the per-TTY mutex, which is expected to improve scalability when compared to the old Giant locking. Another change is the unbuffered copying to userspace, which is both used on TTY device nodes and PTY masters. Upgrading should be quite straightforward. Unlike previous versions, existing kernel configuration files do not need to be changed, except when they reference device drivers that are listed in UPDATING. Obtained from: //depot/projects/mpsafetty/... Approved by: philip (ex-mentor) Discussed: on the lists, at BSDCan, at the DevSummit Sponsored by: Snow B.V., the Netherlands dcons(4) fixed by: kan
2008-08-20 08:31:58 +00:00
#include <sys/tty.h>
2004-12-03 21:29:25 +00:00
#include <sys/unistd.h>
#include <sys/un.h>
#include <sys/unpcb.h>
#include <sys/user.h>
2004-12-03 21:29:25 +00:00
#include <sys/vnode.h>
#ifdef KTRACE
#include <sys/ktrace.h>
#endif
1994-05-24 10:09:53 +00:00
#include <net/vnet.h>
#include <netinet/in.h>
#include <netinet/in_pcb.h>
#include <security/audit/audit.h>
#include <vm/uma.h>
#include <vm/vm.h>
#include <ddb/ddb.h>
static MALLOC_DEFINE(M_FILEDESC, "filedesc", "Open file descriptor table");
static MALLOC_DEFINE(M_FILEDESC_TO_LEADER, "filedesc_to_leader",
"file desc to leader structures");
static MALLOC_DEFINE(M_SIGIO, "sigio", "sigio structures");
MALLOC_DECLARE(M_FADVISE);
static uma_zone_t file_zone;
/* Flags for do_dup() */
#define DUP_FIXED 0x1 /* Force fixed allocation */
#define DUP_FCNTL 0x2 /* fcntl()-style errors */
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
static int closefp(struct filedesc *fdp, int fd, struct file *fp,
struct thread *td, int holdleaders);
static int do_dup(struct thread *td, int flags, int old, int new,
register_t *retval);
static int fd_first_free(struct filedesc *fdp, int low, int size);
static int fd_last_used(struct filedesc *fdp, int size);
static void fdgrowtable(struct filedesc *fdp, int nfd);
static void fdunused(struct filedesc *fdp, int fd);
static void fdused(struct filedesc *fdp, int fd);
static int fill_pipe_info(struct pipe *pi, struct kinfo_file *kif);
static int fill_procdesc_info(struct procdesc *pdp,
struct kinfo_file *kif);
static int fill_pts_info(struct tty *tp, struct kinfo_file *kif);
static int fill_shm_info(struct file *fp, struct kinfo_file *kif);
static int fill_socket_info(struct socket *so, struct kinfo_file *kif);
static int fill_vnode_info(struct vnode *vp, struct kinfo_file *kif);
/*
* A process is initially started out with NDFILE descriptors stored within
* this structure, selected to be enough for typical applications based on
* the historical limit of 20 open files (and the usage of descriptors by
* shells). If these descriptors are exhausted, a larger descriptor table
* may be allocated, up to a process' resource limit; the internal arrays
* are then unused.
*/
#define NDFILE 20
#define NDSLOTSIZE sizeof(NDSLOTTYPE)
#define NDENTRIES (NDSLOTSIZE * __CHAR_BIT)
#define NDSLOT(x) ((x) / NDENTRIES)
#define NDBIT(x) ((NDSLOTTYPE)1 << ((x) % NDENTRIES))
#define NDSLOTS(x) (((x) + NDENTRIES - 1) / NDENTRIES)
2004-11-14 09:21:01 +00:00
/*
* Storage required per open file descriptor.
*/
#define OFILESIZE (sizeof(struct file *) + sizeof(char))
/*
* Storage to hold unused ofiles that need to be reclaimed.
*/
struct freetable {
struct file **ft_table;
SLIST_ENTRY(freetable) ft_next;
};
/*
* Basic allocation of descriptors:
* one of the above, plus arrays for NDFILE descriptors.
*/
struct filedesc0 {
struct filedesc fd_fd;
/*
* ofiles which need to be reclaimed on free.
*/
SLIST_HEAD(,freetable) fd_free;
/*
* These arrays are used when the number of open files is
* <= NDFILE, and are then pointed to by the pointers above.
*/
struct file *fd_dfiles[NDFILE];
char fd_dfileflags[NDFILE];
NDSLOTTYPE fd_dmap[NDSLOTS(NDFILE)];
};
1994-05-24 10:09:53 +00:00
/*
* Descriptor management.
*/
volatile int openfiles; /* actual number of open files */
struct mtx sigio_lock; /* mtx to protect pointers to sigio */
void (*mq_fdclose)(struct thread *td, int fd, struct file *fp);
1994-05-24 10:09:53 +00:00
2004-12-01 08:06:27 +00:00
/* A mutex to protect the association between a proc and filedesc. */
static struct mtx fdesc_mtx;
2004-12-01 08:06:27 +00:00
/*
* If low >= size, just return low. Otherwise find the first zero bit in the
* given bitmap, starting at low and not exceeding size - 1. Return size if
* not found.
*/
static int
fd_first_free(struct filedesc *fdp, int low, int size)
{
NDSLOTTYPE *map = fdp->fd_map;
NDSLOTTYPE mask;
int off, maxoff;
if (low >= size)
return (low);
off = NDSLOT(low);
if (low % NDENTRIES) {
mask = ~(~(NDSLOTTYPE)0 >> (NDENTRIES - (low % NDENTRIES)));
if ((mask &= ~map[off]) != 0UL)
return (off * NDENTRIES + ffsl(mask) - 1);
++off;
}
for (maxoff = NDSLOTS(size); off < maxoff; ++off)
if (map[off] != ~0UL)
return (off * NDENTRIES + ffsl(~map[off]) - 1);
return (size);
}
/*
* Find the highest non-zero bit in the given bitmap, starting at 0 and
* not exceeding size - 1. Return -1 if not found.
*/
static int
fd_last_used(struct filedesc *fdp, int size)
{
NDSLOTTYPE *map = fdp->fd_map;
NDSLOTTYPE mask;
int off, minoff;
off = NDSLOT(size);
if (size % NDENTRIES) {
mask = ~(~(NDSLOTTYPE)0 << (size % NDENTRIES));
if ((mask &= map[off]) != 0)
return (off * NDENTRIES + flsl(mask) - 1);
--off;
}
for (minoff = NDSLOT(0); off >= minoff; --off)
if (map[off] != 0)
return (off * NDENTRIES + flsl(map[off]) - 1);
return (-1);
}
static int
fdisused(struct filedesc *fdp, int fd)
{
FILEDESC_LOCK_ASSERT(fdp);
KASSERT(fd >= 0 && fd < fdp->fd_nfiles,
("file descriptor %d out of range (0, %d)", fd, fdp->fd_nfiles));
return ((fdp->fd_map[NDSLOT(fd)] & NDBIT(fd)) != 0);
}
/*
* Mark a file descriptor as used.
*/
static void
fdused(struct filedesc *fdp, int fd)
{
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK_ASSERT(fdp);
KASSERT(!fdisused(fdp, fd), ("fd=%d is already used", fd));
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
fdp->fd_map[NDSLOT(fd)] |= NDBIT(fd);
if (fd > fdp->fd_lastfile)
fdp->fd_lastfile = fd;
if (fd == fdp->fd_freefile)
fdp->fd_freefile = fd_first_free(fdp, fd, fdp->fd_nfiles);
}
/*
* Mark a file descriptor as unused.
*/
static void
fdunused(struct filedesc *fdp, int fd)
{
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK_ASSERT(fdp);
KASSERT(fdisused(fdp, fd), ("fd=%d is already unused", fd));
KASSERT(fdp->fd_ofiles[fd] == NULL, ("fd=%d is still in use", fd));
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
fdp->fd_map[NDSLOT(fd)] &= ~NDBIT(fd);
if (fd < fdp->fd_freefile)
fdp->fd_freefile = fd;
if (fd == fdp->fd_lastfile)
fdp->fd_lastfile = fd_last_used(fdp, fd);
}
1994-05-24 10:09:53 +00:00
/*
* System calls on descriptors.
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct getdtablesize_args {
int dummy;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_getdtablesize(struct thread *td, struct getdtablesize_args *uap)
1994-05-24 10:09:53 +00:00
{
struct proc *p = td->td_proc;
uint64_t lim;
1994-05-24 10:09:53 +00:00
Locking for the per-process resource limits structure. - struct plimit includes a mutex to protect a reference count. The plimit structure is treated similarly to struct ucred in that is is always copy on write, so having a reference to a structure is sufficient to read from it without needing a further lock. - The proc lock protects the p_limit pointer and must be held while reading limits from a process to keep the limit structure from changing out from under you while reading from it. - Various global limits that are ints are not protected by a lock since int writes are atomic on all the archs we support and thus a lock wouldn't buy us anything. - All accesses to individual resource limits from a process are abstracted behind a simple lim_rlimit(), lim_max(), and lim_cur() API that return either an rlimit, or the current or max individual limit of the specified resource from a process. - dosetrlimit() was renamed to kern_setrlimit() to match existing style of other similar syscall helper functions. - The alpha OSF/1 compat layer no longer calls getrlimit() and setrlimit() (it didn't used the stackgap when it should have) but uses lim_rlimit() and kern_setrlimit() instead. - The svr4 compat no longer uses the stackgap for resource limits calls, but uses lim_rlimit() and kern_setrlimit() instead. - The ibcs2 compat no longer uses the stackgap for resource limits. It also no longer uses the stackgap for accessing sysctl's for the ibcs2_sysconf() syscall but uses kernel_sysctl() instead. As a result, ibcs2_sysconf() no longer needs Giant. - The p_rlimit macro no longer exists. Submitted by: mtm (mostly, I only did a few cleanups and catchups) Tested on: i386 Compiled on: alpha, amd64
2004-02-04 21:52:57 +00:00
PROC_LOCK(p);
2004-01-11 19:39:14 +00:00
td->td_retval[0] =
Locking for the per-process resource limits structure. - struct plimit includes a mutex to protect a reference count. The plimit structure is treated similarly to struct ucred in that is is always copy on write, so having a reference to a structure is sufficient to read from it without needing a further lock. - The proc lock protects the p_limit pointer and must be held while reading limits from a process to keep the limit structure from changing out from under you while reading from it. - Various global limits that are ints are not protected by a lock since int writes are atomic on all the archs we support and thus a lock wouldn't buy us anything. - All accesses to individual resource limits from a process are abstracted behind a simple lim_rlimit(), lim_max(), and lim_cur() API that return either an rlimit, or the current or max individual limit of the specified resource from a process. - dosetrlimit() was renamed to kern_setrlimit() to match existing style of other similar syscall helper functions. - The alpha OSF/1 compat layer no longer calls getrlimit() and setrlimit() (it didn't used the stackgap when it should have) but uses lim_rlimit() and kern_setrlimit() instead. - The svr4 compat no longer uses the stackgap for resource limits calls, but uses lim_rlimit() and kern_setrlimit() instead. - The ibcs2 compat no longer uses the stackgap for resource limits. It also no longer uses the stackgap for accessing sysctl's for the ibcs2_sysconf() syscall but uses kernel_sysctl() instead. As a result, ibcs2_sysconf() no longer needs Giant. - The p_rlimit macro no longer exists. Submitted by: mtm (mostly, I only did a few cleanups and catchups) Tested on: i386 Compiled on: alpha, amd64
2004-02-04 21:52:57 +00:00
min((int)lim_cur(p, RLIMIT_NOFILE), maxfilesperproc);
lim = racct_get_limit(td->td_proc, RACCT_NOFILE);
Locking for the per-process resource limits structure. - struct plimit includes a mutex to protect a reference count. The plimit structure is treated similarly to struct ucred in that is is always copy on write, so having a reference to a structure is sufficient to read from it without needing a further lock. - The proc lock protects the p_limit pointer and must be held while reading limits from a process to keep the limit structure from changing out from under you while reading from it. - Various global limits that are ints are not protected by a lock since int writes are atomic on all the archs we support and thus a lock wouldn't buy us anything. - All accesses to individual resource limits from a process are abstracted behind a simple lim_rlimit(), lim_max(), and lim_cur() API that return either an rlimit, or the current or max individual limit of the specified resource from a process. - dosetrlimit() was renamed to kern_setrlimit() to match existing style of other similar syscall helper functions. - The alpha OSF/1 compat layer no longer calls getrlimit() and setrlimit() (it didn't used the stackgap when it should have) but uses lim_rlimit() and kern_setrlimit() instead. - The svr4 compat no longer uses the stackgap for resource limits calls, but uses lim_rlimit() and kern_setrlimit() instead. - The ibcs2 compat no longer uses the stackgap for resource limits. It also no longer uses the stackgap for accessing sysctl's for the ibcs2_sysconf() syscall but uses kernel_sysctl() instead. As a result, ibcs2_sysconf() no longer needs Giant. - The p_rlimit macro no longer exists. Submitted by: mtm (mostly, I only did a few cleanups and catchups) Tested on: i386 Compiled on: alpha, amd64
2004-02-04 21:52:57 +00:00
PROC_UNLOCK(p);
if (lim < td->td_retval[0])
td->td_retval[0] = lim;
1994-05-24 10:09:53 +00:00
return (0);
}
/*
* Duplicate a file descriptor to a particular value.
*
* Note: keep in mind that a potential race condition exists when closing
* descriptors from a shared descriptor table (via rfork).
1994-05-24 10:09:53 +00:00
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct dup2_args {
u_int from;
u_int to;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_dup2(struct thread *td, struct dup2_args *uap)
1994-05-24 10:09:53 +00:00
{
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
return (do_dup(td, DUP_FIXED, (int)uap->from, (int)uap->to,
td->td_retval));
1994-05-24 10:09:53 +00:00
}
/*
* Duplicate a file descriptor.
*/
#ifndef _SYS_SYSPROTO_H_
struct dup_args {
u_int fd;
};
#endif
/* ARGSUSED */
int
sys_dup(struct thread *td, struct dup_args *uap)
{
return (do_dup(td, 0, (int)uap->fd, 0, td->td_retval));
}
1994-05-24 10:09:53 +00:00
/*
* The file control system call.
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct fcntl_args {
int fd;
int cmd;
long arg;
1994-05-24 10:09:53 +00:00
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_fcntl(struct thread *td, struct fcntl_args *uap)
{
struct flock fl;
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
struct oflock ofl;
intptr_t arg;
int error;
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
int cmd;
error = 0;
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
cmd = uap->cmd;
switch (uap->cmd) {
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
case F_OGETLK:
case F_OSETLK:
case F_OSETLKW:
/*
* Convert old flock structure to new.
*/
error = copyin((void *)(intptr_t)uap->arg, &ofl, sizeof(ofl));
fl.l_start = ofl.l_start;
fl.l_len = ofl.l_len;
fl.l_pid = ofl.l_pid;
fl.l_type = ofl.l_type;
fl.l_whence = ofl.l_whence;
fl.l_sysid = 0;
switch (uap->cmd) {
case F_OGETLK:
cmd = F_GETLK;
break;
case F_OSETLK:
cmd = F_SETLK;
break;
case F_OSETLKW:
cmd = F_SETLKW;
break;
}
arg = (intptr_t)&fl;
break;
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
case F_GETLK:
case F_SETLK:
case F_SETLKW:
case F_SETLK_REMOTE:
error = copyin((void *)(intptr_t)uap->arg, &fl, sizeof(fl));
arg = (intptr_t)&fl;
break;
default:
arg = uap->arg;
break;
}
if (error)
return (error);
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
error = kern_fcntl(td, uap->fd, cmd, arg);
if (error)
return (error);
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
if (uap->cmd == F_OGETLK) {
ofl.l_start = fl.l_start;
ofl.l_len = fl.l_len;
ofl.l_pid = fl.l_pid;
ofl.l_type = fl.l_type;
ofl.l_whence = fl.l_whence;
error = copyout(&ofl, (void *)(intptr_t)uap->arg, sizeof(ofl));
} else if (uap->cmd == F_GETLK) {
error = copyout(&fl, (void *)(intptr_t)uap->arg, sizeof(fl));
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
}
return (error);
}
static inline int
fdunwrap(int fd, cap_rights_t rights, struct filedesc *fdp, struct file **fpp)
{
FILEDESC_LOCK_ASSERT(fdp);
*fpp = fget_locked(fdp, fd);
if (*fpp == NULL)
return (EBADF);
#ifdef CAPABILITIES
if ((*fpp)->f_type == DTYPE_CAPABILITY) {
int err = cap_funwrap(*fpp, rights, fpp);
if (err != 0) {
*fpp = NULL;
return (err);
}
}
#endif /* CAPABILITIES */
return (0);
}
int
kern_fcntl(struct thread *td, int fd, int cmd, intptr_t arg)
1994-05-24 10:09:53 +00:00
{
2003-01-01 01:05:54 +00:00
struct filedesc *fdp;
struct flock *flp;
2003-01-01 01:05:54 +00:00
struct file *fp;
struct proc *p;
2003-01-01 01:05:54 +00:00
char *pop;
1994-05-24 10:09:53 +00:00
struct vnode *vp;
int error, flg, tmp;
int vfslocked;
u_int old, new;
uint64_t bsize;
off_t foffset;
1994-05-24 10:09:53 +00:00
vfslocked = 0;
error = 0;
flg = F_POSIX;
p = td->td_proc;
fdp = p->p_fd;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
switch (cmd) {
1994-05-24 10:09:53 +00:00
case F_DUPFD:
tmp = arg;
error = do_dup(td, DUP_FCNTL, fd, tmp, td->td_retval);
break;
1994-05-24 10:09:53 +00:00
case F_DUP2FD:
tmp = arg;
error = do_dup(td, DUP_FIXED, fd, tmp, td->td_retval);
break;
1994-05-24 10:09:53 +00:00
case F_GETFD:
FILEDESC_SLOCK(fdp);
if ((fp = fget_locked(fdp, fd)) == NULL) {
FILEDESC_SUNLOCK(fdp);
error = EBADF;
break;
}
pop = &fdp->fd_ofileflags[fd];
td->td_retval[0] = (*pop & UF_EXCLOSE) ? FD_CLOEXEC : 0;
FILEDESC_SUNLOCK(fdp);
break;
1994-05-24 10:09:53 +00:00
case F_SETFD:
FILEDESC_XLOCK(fdp);
if ((fp = fget_locked(fdp, fd)) == NULL) {
FILEDESC_XUNLOCK(fdp);
error = EBADF;
break;
}
pop = &fdp->fd_ofileflags[fd];
*pop = (*pop &~ UF_EXCLOSE) |
(arg & FD_CLOEXEC ? UF_EXCLOSE : 0);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
break;
1994-05-24 10:09:53 +00:00
case F_GETFL:
FILEDESC_SLOCK(fdp);
error = fdunwrap(fd, CAP_FCNTL, fdp, &fp);
if (error != 0) {
FILEDESC_SUNLOCK(fdp);
break;
}
td->td_retval[0] = OFLAGS(fp->f_flag);
FILEDESC_SUNLOCK(fdp);
break;
1994-05-24 10:09:53 +00:00
case F_SETFL:
FILEDESC_SLOCK(fdp);
error = fdunwrap(fd, CAP_FCNTL, fdp, &fp);
if (error != 0) {
FILEDESC_SUNLOCK(fdp);
break;
}
fhold(fp);
FILEDESC_SUNLOCK(fdp);
do {
tmp = flg = fp->f_flag;
tmp &= ~FCNTLFLAGS;
tmp |= FFLAGS(arg & ~O_ACCMODE) & FCNTLFLAGS;
} while(atomic_cmpset_int(&fp->f_flag, flg, tmp) == 0);
1994-05-24 10:09:53 +00:00
tmp = fp->f_flag & FNONBLOCK;
error = fo_ioctl(fp, FIONBIO, &tmp, td->td_ucred, td);
if (error) {
fdrop(fp, td);
break;
}
1994-05-24 10:09:53 +00:00
tmp = fp->f_flag & FASYNC;
error = fo_ioctl(fp, FIOASYNC, &tmp, td->td_ucred, td);
if (error == 0) {
fdrop(fp, td);
break;
}
atomic_clear_int(&fp->f_flag, FNONBLOCK);
1994-05-24 10:09:53 +00:00
tmp = 0;
(void)fo_ioctl(fp, FIONBIO, &tmp, td->td_ucred, td);
fdrop(fp, td);
break;
1994-05-24 10:09:53 +00:00
case F_GETOWN:
FILEDESC_SLOCK(fdp);
error = fdunwrap(fd, CAP_FCNTL, fdp, &fp);
if (error != 0) {
FILEDESC_SUNLOCK(fdp);
break;
}
fhold(fp);
FILEDESC_SUNLOCK(fdp);
error = fo_ioctl(fp, FIOGETOWN, &tmp, td->td_ucred, td);
if (error == 0)
td->td_retval[0] = tmp;
fdrop(fp, td);
break;
1994-05-24 10:09:53 +00:00
case F_SETOWN:
FILEDESC_SLOCK(fdp);
error = fdunwrap(fd, CAP_FCNTL, fdp, &fp);
if (error != 0) {
FILEDESC_SUNLOCK(fdp);
break;
}
fhold(fp);
FILEDESC_SUNLOCK(fdp);
tmp = arg;
error = fo_ioctl(fp, FIOSETOWN, &tmp, td->td_ucred, td);
fdrop(fp, td);
break;
1994-05-24 10:09:53 +00:00
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
case F_SETLK_REMOTE:
error = priv_check(td, PRIV_NFS_LOCKD);
if (error)
return (error);
flg = F_REMOTE;
goto do_setlk;
1994-05-24 10:09:53 +00:00
case F_SETLKW:
flg |= F_WAIT;
/* FALLTHROUGH F_SETLK */
1994-05-24 10:09:53 +00:00
case F_SETLK:
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
do_setlk:
FILEDESC_SLOCK(fdp);
error = fdunwrap(fd, CAP_FLOCK, fdp, &fp);
if (error != 0) {
FILEDESC_SUNLOCK(fdp);
break;
}
if (fp->f_type != DTYPE_VNODE) {
FILEDESC_SUNLOCK(fdp);
error = EBADF;
break;
}
flp = (struct flock *)arg;
if (flp->l_whence == SEEK_CUR) {
foffset = foffset_get(fp);
if (foffset < 0 ||
(flp->l_start > 0 &&
foffset > OFF_MAX - flp->l_start)) {
FILEDESC_SUNLOCK(fdp);
error = EOVERFLOW;
break;
}
flp->l_start += foffset;
}
1994-05-24 10:09:53 +00:00
/*
* VOP_ADVLOCK() may block.
*/
fhold(fp);
FILEDESC_SUNLOCK(fdp);
vp = fp->f_vnode;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
switch (flp->l_type) {
1994-05-24 10:09:53 +00:00
case F_RDLCK:
if ((fp->f_flag & FREAD) == 0) {
error = EBADF;
break;
}
PROC_LOCK(p->p_leader);
p->p_leader->p_flag |= P_ADVLOCK;
PROC_UNLOCK(p->p_leader);
error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_SETLK,
flp, flg);
break;
1994-05-24 10:09:53 +00:00
case F_WRLCK:
if ((fp->f_flag & FWRITE) == 0) {
error = EBADF;
break;
}
PROC_LOCK(p->p_leader);
p->p_leader->p_flag |= P_ADVLOCK;
PROC_UNLOCK(p->p_leader);
error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_SETLK,
flp, flg);
break;
1994-05-24 10:09:53 +00:00
case F_UNLCK:
error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_UNLCK,
Add the new kernel-mode NFS Lock Manager. To use it instead of the user-mode lock manager, build a kernel with the NFSLOCKD option and add '-k' to 'rpc_lockd_flags' in rc.conf. Highlights include: * Thread-safe kernel RPC client - many threads can use the same RPC client handle safely with replies being de-multiplexed at the socket upcall (typically driven directly by the NIC interrupt) and handed off to whichever thread matches the reply. For UDP sockets, many RPC clients can share the same socket. This allows the use of a single privileged UDP port number to talk to an arbitrary number of remote hosts. * Single-threaded kernel RPC server. Adding support for multi-threaded server would be relatively straightforward and would follow approximately the Solaris KPI. A single thread should be sufficient for the NLM since it should rarely block in normal operation. * Kernel mode NLM server supporting cancel requests and granted callbacks. I've tested the NLM server reasonably extensively - it passes both my own tests and the NFS Connectathon locking tests running on Solaris, Mac OS X and Ubuntu Linux. * Userland NLM client supported. While the NLM server doesn't have support for the local NFS client's locking needs, it does have to field async replies and granted callbacks from remote NLMs that the local client has contacted. We relay these replies to the userland rpc.lockd over a local domain RPC socket. * Robust deadlock detection for the local lock manager. In particular it will detect deadlocks caused by a lock request that covers more than one blocking request. As required by the NLM protocol, all deadlock detection happens synchronously - a user is guaranteed that if a lock request isn't rejected immediately, the lock will eventually be granted. The old system allowed for a 'deferred deadlock' condition where a blocked lock request could wake up and find that some other deadlock-causing lock owner had beaten them to the lock. * Since both local and remote locks are managed by the same kernel locking code, local and remote processes can safely use file locks for mutual exclusion. Local processes have no fairness advantage compared to remote processes when contending to lock a region that has just been unlocked - the local lock manager enforces a strict first-come first-served model for both local and remote lockers. Sponsored by: Isilon Systems PR: 95247 107555 115524 116679 MFC after: 2 weeks
2008-03-26 15:23:12 +00:00
flp, flg);
break;
case F_UNLCKSYS:
/*
* Temporary api for testing remote lock
* infrastructure.
*/
if (flg != F_REMOTE) {
error = EINVAL;
break;
}
error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader,
F_UNLCKSYS, flp, flg);
break;
1994-05-24 10:09:53 +00:00
default:
error = EINVAL;
break;
1994-05-24 10:09:53 +00:00
}
VFS_UNLOCK_GIANT(vfslocked);
vfslocked = 0;
if (error != 0 || flp->l_type == F_UNLCK ||
flp->l_type == F_UNLCKSYS) {
fdrop(fp, td);
break;
}
/*
* Check for a race with close.
*
* The vnode is now advisory locked (or unlocked, but this case
* is not really important) as the caller requested.
* We had to drop the filedesc lock, so we need to recheck if
* the descriptor is still valid, because if it was closed
* in the meantime we need to remove advisory lock from the
* vnode - close on any descriptor leading to an advisory
* locked vnode, removes that lock.
* We will return 0 on purpose in that case, as the result of
* successful advisory lock might have been externally visible
* already. This is fine - effectively we pretend to the caller
* that the closing thread was a bit slower and that the
* advisory lock succeeded before the close.
*/
FILEDESC_SLOCK(fdp);
if (fget_locked(fdp, fd) != fp) {
FILEDESC_SUNLOCK(fdp);
flp->l_whence = SEEK_SET;
flp->l_start = 0;
flp->l_len = 0;
flp->l_type = F_UNLCK;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
(void) VOP_ADVLOCK(vp, (caddr_t)p->p_leader,
2012-06-14 12:37:41 +00:00
F_UNLCK, flp, F_POSIX);
VFS_UNLOCK_GIANT(vfslocked);
vfslocked = 0;
} else
FILEDESC_SUNLOCK(fdp);
fdrop(fp, td);
break;
1994-05-24 10:09:53 +00:00
case F_GETLK:
FILEDESC_SLOCK(fdp);
error = fdunwrap(fd, CAP_FLOCK, fdp, &fp);
if (error != 0) {
FILEDESC_SUNLOCK(fdp);
break;
}
if (fp->f_type != DTYPE_VNODE) {
FILEDESC_SUNLOCK(fdp);
error = EBADF;
break;
}
flp = (struct flock *)arg;
if (flp->l_type != F_RDLCK && flp->l_type != F_WRLCK &&
flp->l_type != F_UNLCK) {
FILEDESC_SUNLOCK(fdp);
error = EINVAL;
break;
}
if (flp->l_whence == SEEK_CUR) {
foffset = foffset_get(fp);
if ((flp->l_start > 0 &&
foffset > OFF_MAX - flp->l_start) ||
(flp->l_start < 0 &&
foffset < OFF_MIN - flp->l_start)) {
FILEDESC_SUNLOCK(fdp);
error = EOVERFLOW;
break;
}
flp->l_start += foffset;
}
/*
* VOP_ADVLOCK() may block.
*/
fhold(fp);
FILEDESC_SUNLOCK(fdp);
vp = fp->f_vnode;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_GETLK, flp,
F_POSIX);
VFS_UNLOCK_GIANT(vfslocked);
vfslocked = 0;
fdrop(fp, td);
break;
case F_RDAHEAD:
arg = arg ? 128 * 1024: 0;
/* FALLTHROUGH */
case F_READAHEAD:
FILEDESC_SLOCK(fdp);
if ((fp = fget_locked(fdp, fd)) == NULL) {
FILEDESC_SUNLOCK(fdp);
error = EBADF;
break;
}
if (fp->f_type != DTYPE_VNODE) {
FILEDESC_SUNLOCK(fdp);
error = EBADF;
break;
}
fhold(fp);
FILEDESC_SUNLOCK(fdp);
if (arg != 0) {
vp = fp->f_vnode;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
error = vn_lock(vp, LK_SHARED);
if (error != 0)
goto readahead_vnlock_fail;
bsize = fp->f_vnode->v_mount->mnt_stat.f_iosize;
VOP_UNLOCK(vp, 0);
fp->f_seqcount = (arg + bsize - 1) / bsize;
do {
new = old = fp->f_flag;
new |= FRDAHEAD;
} while (!atomic_cmpset_rel_int(&fp->f_flag, old, new));
readahead_vnlock_fail:
VFS_UNLOCK_GIANT(vfslocked);
vfslocked = 0;
} else {
do {
new = old = fp->f_flag;
new &= ~FRDAHEAD;
} while (!atomic_cmpset_rel_int(&fp->f_flag, old, new));
}
fdrop(fp, td);
break;
1994-05-24 10:09:53 +00:00
default:
error = EINVAL;
break;
1994-05-24 10:09:53 +00:00
}
VFS_UNLOCK_GIANT(vfslocked);
return (error);
1994-05-24 10:09:53 +00:00
}
/*
* Common code for dup, dup2, fcntl(F_DUPFD) and fcntl(F_DUP2FD).
1994-05-24 10:09:53 +00:00
*/
static int
do_dup(struct thread *td, int flags, int old, int new,
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
register_t *retval)
1994-05-24 10:09:53 +00:00
{
2003-01-01 01:05:54 +00:00
struct filedesc *fdp;
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
struct proc *p;
struct file *fp;
struct file *delfp;
int error, maxfd;
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
p = td->td_proc;
fdp = p->p_fd;
/*
* Verify we have a valid descriptor to dup from and possibly to
* dup to. Unlike dup() and dup2(), fcntl()'s F_DUPFD should
* return EINVAL when the new descriptor is out of bounds.
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
*/
if (old < 0)
return (EBADF);
if (new < 0)
return (flags & DUP_FCNTL ? EINVAL : EBADF);
Locking for the per-process resource limits structure. - struct plimit includes a mutex to protect a reference count. The plimit structure is treated similarly to struct ucred in that is is always copy on write, so having a reference to a structure is sufficient to read from it without needing a further lock. - The proc lock protects the p_limit pointer and must be held while reading limits from a process to keep the limit structure from changing out from under you while reading from it. - Various global limits that are ints are not protected by a lock since int writes are atomic on all the archs we support and thus a lock wouldn't buy us anything. - All accesses to individual resource limits from a process are abstracted behind a simple lim_rlimit(), lim_max(), and lim_cur() API that return either an rlimit, or the current or max individual limit of the specified resource from a process. - dosetrlimit() was renamed to kern_setrlimit() to match existing style of other similar syscall helper functions. - The alpha OSF/1 compat layer no longer calls getrlimit() and setrlimit() (it didn't used the stackgap when it should have) but uses lim_rlimit() and kern_setrlimit() instead. - The svr4 compat no longer uses the stackgap for resource limits calls, but uses lim_rlimit() and kern_setrlimit() instead. - The ibcs2 compat no longer uses the stackgap for resource limits. It also no longer uses the stackgap for accessing sysctl's for the ibcs2_sysconf() syscall but uses kernel_sysctl() instead. As a result, ibcs2_sysconf() no longer needs Giant. - The p_rlimit macro no longer exists. Submitted by: mtm (mostly, I only did a few cleanups and catchups) Tested on: i386 Compiled on: alpha, amd64
2004-02-04 21:52:57 +00:00
PROC_LOCK(p);
maxfd = min((int)lim_cur(p, RLIMIT_NOFILE), maxfilesperproc);
PROC_UNLOCK(p);
if (new >= maxfd)
return (flags & DUP_FCNTL ? EINVAL : EBADF);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
if (fget_locked(fdp, old) == NULL) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
return (EBADF);
}
if (flags & DUP_FIXED && old == new) {
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
*retval = new;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
return (0);
}
fp = fdp->fd_ofiles[old];
fhold(fp);
/*
* If the caller specified a file descriptor, make sure the file
* table is large enough to hold it, and grab it. Otherwise, just
* allocate a new descriptor the usual way. Since the filedesc
* lock may be temporarily dropped in the process, we have to look
* out for a race.
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
*/
if (flags & DUP_FIXED) {
if (new >= fdp->fd_nfiles) {
/*
* The resource limits are here instead of e.g.
* fdalloc(), because the file descriptor table may be
* shared between processes, so we can't really use
* racct_add()/racct_sub(). Instead of counting the
* number of actually allocated descriptors, just put
* the limit on the size of the file descriptor table.
*/
#ifdef RACCT
PROC_LOCK(p);
error = racct_set(p, RACCT_NOFILE, new + 1);
PROC_UNLOCK(p);
if (error != 0) {
FILEDESC_XUNLOCK(fdp);
fdrop(fp, td);
return (EMFILE);
}
#endif
fdgrowtable(fdp, new + 1);
}
if (fdp->fd_ofiles[new] == NULL)
fdused(fdp, new);
} else {
if ((error = fdalloc(td, new, &new)) != 0) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
fdrop(fp, td);
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
return (error);
}
}
KASSERT(fp == fdp->fd_ofiles[old], ("old fd has been modified"));
KASSERT(old != new, ("new fd is same as old"));
delfp = fdp->fd_ofiles[new];
/*
* Duplicate the source descriptor.
*/
fdp->fd_ofiles[new] = fp;
fdp->fd_ofileflags[new] = fdp->fd_ofileflags[old] &~ UF_EXCLOSE;
if (new > fdp->fd_lastfile)
fdp->fd_lastfile = new;
*retval = new;
if (delfp != NULL) {
(void) closefp(fdp, new, delfp, td, 1);
/* closefp() drops the FILEDESC lock for us. */
2004-08-16 03:09:01 +00:00
} else {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
}
1994-05-24 10:09:53 +00:00
return (0);
}
/*
* If sigio is on the list associated with a process or process group,
* disable signalling from the device, remove sigio from the list and
* free sigio.
*/
void
funsetown(struct sigio **sigiop)
{
struct sigio *sigio;
SIGIO_LOCK();
sigio = *sigiop;
if (sigio == NULL) {
SIGIO_UNLOCK();
return;
}
*(sigio->sio_myref) = NULL;
if ((sigio)->sio_pgid < 0) {
struct pgrp *pg = (sigio)->sio_pgrp;
PGRP_LOCK(pg);
SLIST_REMOVE(&sigio->sio_pgrp->pg_sigiolst, sigio,
sigio, sio_pgsigio);
PGRP_UNLOCK(pg);
} else {
struct proc *p = (sigio)->sio_proc;
PROC_LOCK(p);
SLIST_REMOVE(&sigio->sio_proc->p_sigiolst, sigio,
sigio, sio_pgsigio);
PROC_UNLOCK(p);
}
SIGIO_UNLOCK();
crfree(sigio->sio_ucred);
free(sigio, M_SIGIO);
}
/*
* Free a list of sigio structures.
* We only need to lock the SIGIO_LOCK because we have made ourselves
2005-03-08 00:58:50 +00:00
* inaccessible to callers of fsetown and therefore do not need to lock
* the proc or pgrp struct for the list manipulation.
*/
void
funsetownlst(struct sigiolst *sigiolst)
{
struct proc *p;
struct pgrp *pg;
struct sigio *sigio;
sigio = SLIST_FIRST(sigiolst);
if (sigio == NULL)
return;
p = NULL;
pg = NULL;
/*
* Every entry of the list should belong
* to a single proc or pgrp.
*/
if (sigio->sio_pgid < 0) {
pg = sigio->sio_pgrp;
PGRP_LOCK_ASSERT(pg, MA_NOTOWNED);
} else /* if (sigio->sio_pgid > 0) */ {
p = sigio->sio_proc;
PROC_LOCK_ASSERT(p, MA_NOTOWNED);
}
SIGIO_LOCK();
while ((sigio = SLIST_FIRST(sigiolst)) != NULL) {
*(sigio->sio_myref) = NULL;
if (pg != NULL) {
KASSERT(sigio->sio_pgid < 0,
("Proc sigio in pgrp sigio list"));
KASSERT(sigio->sio_pgrp == pg,
("Bogus pgrp in sigio list"));
PGRP_LOCK(pg);
SLIST_REMOVE(&pg->pg_sigiolst, sigio, sigio,
sio_pgsigio);
PGRP_UNLOCK(pg);
} else /* if (p != NULL) */ {
KASSERT(sigio->sio_pgid > 0,
("Pgrp sigio in proc sigio list"));
KASSERT(sigio->sio_proc == p,
("Bogus proc in sigio list"));
PROC_LOCK(p);
SLIST_REMOVE(&p->p_sigiolst, sigio, sigio,
sio_pgsigio);
PROC_UNLOCK(p);
}
SIGIO_UNLOCK();
crfree(sigio->sio_ucred);
free(sigio, M_SIGIO);
SIGIO_LOCK();
}
SIGIO_UNLOCK();
}
/*
* This is common code for FIOSETOWN ioctl called by fcntl(fd, F_SETOWN, arg).
*
* After permission checking, add a sigio structure to the sigio list for
* the process or process group.
*/
int
fsetown(pid_t pgid, struct sigio **sigiop)
{
struct proc *proc;
struct pgrp *pgrp;
struct sigio *sigio;
int ret;
if (pgid == 0) {
funsetown(sigiop);
return (0);
}
ret = 0;
/* Allocate and fill in the new sigio out of locks. */
sigio = malloc(sizeof(struct sigio), M_SIGIO, M_WAITOK);
sigio->sio_pgid = pgid;
sigio->sio_ucred = crhold(curthread->td_ucred);
sigio->sio_myref = sigiop;
sx_slock(&proctree_lock);
if (pgid > 0) {
proc = pfind(pgid);
if (proc == NULL) {
ret = ESRCH;
goto fail;
}
/*
* Policy - Don't allow a process to FSETOWN a process
* in another session.
*
* Remove this test to allow maximum flexibility or
* restrict FSETOWN to the current process or process
* group for maximum safety.
*/
PROC_UNLOCK(proc);
if (proc->p_session != curthread->td_proc->p_session) {
ret = EPERM;
goto fail;
}
pgrp = NULL;
} else /* if (pgid < 0) */ {
pgrp = pgfind(-pgid);
if (pgrp == NULL) {
ret = ESRCH;
goto fail;
}
PGRP_UNLOCK(pgrp);
/*
* Policy - Don't allow a process to FSETOWN a process
* in another session.
*
* Remove this test to allow maximum flexibility or
* restrict FSETOWN to the current process or process
* group for maximum safety.
*/
if (pgrp->pg_session != curthread->td_proc->p_session) {
ret = EPERM;
goto fail;
}
proc = NULL;
}
funsetown(sigiop);
if (pgid > 0) {
PROC_LOCK(proc);
2004-01-11 19:39:14 +00:00
/*
* Since funsetownlst() is called without the proctree
* locked, we need to check for P_WEXIT.
* XXX: is ESRCH correct?
*/
if ((proc->p_flag & P_WEXIT) != 0) {
PROC_UNLOCK(proc);
ret = ESRCH;
goto fail;
}
SLIST_INSERT_HEAD(&proc->p_sigiolst, sigio, sio_pgsigio);
sigio->sio_proc = proc;
PROC_UNLOCK(proc);
} else {
PGRP_LOCK(pgrp);
SLIST_INSERT_HEAD(&pgrp->pg_sigiolst, sigio, sio_pgsigio);
sigio->sio_pgrp = pgrp;
PGRP_UNLOCK(pgrp);
}
sx_sunlock(&proctree_lock);
SIGIO_LOCK();
*sigiop = sigio;
SIGIO_UNLOCK();
return (0);
fail:
sx_sunlock(&proctree_lock);
crfree(sigio->sio_ucred);
free(sigio, M_SIGIO);
return (ret);
}
/*
* This is common code for FIOGETOWN ioctl called by fcntl(fd, F_GETOWN, arg).
*/
pid_t
fgetown(sigiop)
struct sigio **sigiop;
{
pid_t pgid;
SIGIO_LOCK();
pgid = (*sigiop != NULL) ? (*sigiop)->sio_pgid : 0;
SIGIO_UNLOCK();
return (pgid);
}
/*
* Function drops the filedesc lock on return.
*/
static int
closefp(struct filedesc *fdp, int fd, struct file *fp, struct thread *td,
int holdleaders)
{
struct file *fp_object;
int error;
FILEDESC_XLOCK_ASSERT(fdp);
if (holdleaders) {
if (td->td_proc->p_fdtol != NULL) {
/*
* Ask fdfree() to sleep to ensure that all relevant
* process leaders can be traversed in closef().
*/
fdp->fd_holdleaderscount++;
} else {
holdleaders = 0;
}
}
/*
* We now hold the fp reference that used to be owned by the
* descriptor array. We have to unlock the FILEDESC *AFTER*
* knote_fdclose to prevent a race of the fd getting opened, a knote
* added, and deleteing a knote for the new fd.
*/
knote_fdclose(td, fd);
/*
* When we're closing an fd with a capability, we need to notify
* mqueue if the underlying object is of type mqueue.
*/
(void)cap_funwrap(fp, 0, &fp_object);
if (fp_object->f_type == DTYPE_MQUEUE)
mq_fdclose(td, fd, fp_object);
FILEDESC_XUNLOCK(fdp);
error = closef(fp, td);
if (holdleaders) {
FILEDESC_XLOCK(fdp);
fdp->fd_holdleaderscount--;
if (fdp->fd_holdleaderscount == 0 &&
fdp->fd_holdleaderswakeup != 0) {
fdp->fd_holdleaderswakeup = 0;
wakeup(&fdp->fd_holdleaderscount);
}
FILEDESC_XUNLOCK(fdp);
}
return (error);
}
1994-05-24 10:09:53 +00:00
/*
* Close a file descriptor.
*/
#ifndef _SYS_SYSPROTO_H_
struct close_args {
2004-01-11 19:39:14 +00:00
int fd;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_close(td, uap)
struct thread *td;
1994-05-24 10:09:53 +00:00
struct close_args *uap;
{
return (kern_close(td, uap->fd));
}
int
kern_close(td, fd)
struct thread *td;
int fd;
{
2003-01-01 01:05:54 +00:00
struct filedesc *fdp;
struct file *fp;
1994-05-24 10:09:53 +00:00
fdp = td->td_proc->p_fd;
AUDIT_SYSCLOSE(td, fd);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
if ((fp = fget_locked(fdp, fd)) == NULL) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
return (EBADF);
}
1994-05-24 10:09:53 +00:00
fdp->fd_ofiles[fd] = NULL;
fdp->fd_ofileflags[fd] = 0;
fdunused(fdp, fd);
/* closefp() drops the FILEDESC lock for us. */
return (closefp(fdp, fd, fp, td, 1));
1994-05-24 10:09:53 +00:00
}
/*
* Close open file descriptors.
*/
#ifndef _SYS_SYSPROTO_H_
struct closefrom_args {
int lowfd;
};
#endif
/* ARGSUSED */
int
sys_closefrom(struct thread *td, struct closefrom_args *uap)
{
struct filedesc *fdp;
int fd;
fdp = td->td_proc->p_fd;
AUDIT_ARG_FD(uap->lowfd);
/*
* Treat negative starting file descriptor values identical to
* closefrom(0) which closes all files.
*/
if (uap->lowfd < 0)
uap->lowfd = 0;
FILEDESC_SLOCK(fdp);
for (fd = uap->lowfd; fd < fdp->fd_nfiles; fd++) {
if (fdp->fd_ofiles[fd] != NULL) {
FILEDESC_SUNLOCK(fdp);
(void)kern_close(td, fd);
FILEDESC_SLOCK(fdp);
}
}
FILEDESC_SUNLOCK(fdp);
return (0);
}
#if defined(COMPAT_43)
1994-05-24 10:09:53 +00:00
/*
* Return status information about a file descriptor.
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct ofstat_args {
int fd;
struct ostat *sb;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
ofstat(struct thread *td, struct ofstat_args *uap)
1994-05-24 10:09:53 +00:00
{
struct ostat oub;
struct stat ub;
1994-05-24 10:09:53 +00:00
int error;
error = kern_fstat(td, uap->fd, &ub);
if (error == 0) {
cvtstat(&ub, &oub);
error = copyout(&oub, uap->sb, sizeof(oub));
}
1994-05-24 10:09:53 +00:00
return (error);
}
#endif /* COMPAT_43 */
1994-05-24 10:09:53 +00:00
/*
* Return status information about a file descriptor.
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct fstat_args {
int fd;
struct stat *sb;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_fstat(struct thread *td, struct fstat_args *uap)
1994-05-24 10:09:53 +00:00
{
struct stat ub;
int error;
error = kern_fstat(td, uap->fd, &ub);
1994-05-24 10:09:53 +00:00
if (error == 0)
error = copyout(&ub, uap->sb, sizeof(ub));
return (error);
}
int
kern_fstat(struct thread *td, int fd, struct stat *sbp)
{
struct file *fp;
int error;
AUDIT_ARG_FD(fd);
if ((error = fget(td, fd, CAP_FSTAT, &fp)) != 0)
return (error);
AUDIT_ARG_FILE(td->td_proc, fp);
error = fo_stat(fp, sbp, td->td_ucred, td);
fdrop(fp, td);
#ifdef KTRACE
if (error == 0 && KTRPOINT(td, KTR_STRUCT))
ktrstat(sbp);
#endif
1994-05-24 10:09:53 +00:00
return (error);
}
/*
* Return status information about a file descriptor.
*/
#ifndef _SYS_SYSPROTO_H_
struct nfstat_args {
int fd;
struct nstat *sb;
};
#endif
/* ARGSUSED */
int
sys_nfstat(struct thread *td, struct nfstat_args *uap)
{
struct nstat nub;
struct stat ub;
int error;
error = kern_fstat(td, uap->fd, &ub);
if (error == 0) {
cvtnstat(&ub, &nub);
error = copyout(&nub, uap->sb, sizeof(nub));
}
return (error);
}
1994-05-24 10:09:53 +00:00
/*
* Return pathconf information about a file descriptor.
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct fpathconf_args {
int fd;
int name;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_fpathconf(struct thread *td, struct fpathconf_args *uap)
1994-05-24 10:09:53 +00:00
{
struct file *fp;
struct vnode *vp;
int error;
1994-05-24 10:09:53 +00:00
if ((error = fget(td, uap->fd, CAP_FPATHCONF, &fp)) != 0)
return (error);
/* If asynchronous I/O is available, it works for all descriptors. */
if (uap->name == _PC_ASYNC_IO) {
td->td_retval[0] = async_io_version;
goto out;
}
vp = fp->f_vnode;
if (vp != NULL) {
int vfslocked;
2012-06-14 12:37:41 +00:00
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
vn_lock(vp, LK_SHARED | LK_RETRY);
error = VOP_PATHCONF(vp, uap->name, td->td_retval);
VOP_UNLOCK(vp, 0);
VFS_UNLOCK_GIANT(vfslocked);
} else if (fp->f_type == DTYPE_PIPE || fp->f_type == DTYPE_SOCKET) {
if (uap->name != _PC_PIPE_BUF) {
error = EINVAL;
} else {
td->td_retval[0] = PIPE_BUF;
2012-06-14 12:37:41 +00:00
error = 0;
}
} else {
error = EOPNOTSUPP;
1994-05-24 10:09:53 +00:00
}
out:
fdrop(fp, td);
return (error);
1994-05-24 10:09:53 +00:00
}
/*
* Grow the file table to accomodate (at least) nfd descriptors.
1994-05-24 10:09:53 +00:00
*/
static void
fdgrowtable(struct filedesc *fdp, int nfd)
{
struct filedesc0 *fdp0;
struct freetable *fo;
struct file **ntable;
struct file **otable;
char *nfileflags;
int nnfiles, onfiles;
NDSLOTTYPE *nmap;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK_ASSERT(fdp);
KASSERT(fdp->fd_nfiles > 0,
("zero-length file table"));
/* compute the size of the new table */
onfiles = fdp->fd_nfiles;
nnfiles = NDSLOTS(nfd) * NDENTRIES; /* round up */
if (nnfiles <= onfiles)
/* the table is already large enough */
return;
/* allocate a new table and (if required) new bitmaps */
ntable = malloc((nnfiles * OFILESIZE) + sizeof(struct freetable),
M_FILEDESC, M_ZERO | M_WAITOK);
nfileflags = (char *)&ntable[nnfiles];
if (NDSLOTS(nnfiles) > NDSLOTS(onfiles))
nmap = malloc(NDSLOTS(nnfiles) * NDSLOTSIZE,
M_FILEDESC, M_ZERO | M_WAITOK);
else
nmap = NULL;
bcopy(fdp->fd_ofiles, ntable, onfiles * sizeof(*ntable));
bcopy(fdp->fd_ofileflags, nfileflags, onfiles);
otable = fdp->fd_ofiles;
fdp->fd_ofileflags = nfileflags;
fdp->fd_ofiles = ntable;
/*
* We must preserve ofiles until the process exits because we can't
* be certain that no threads have references to the old table via
* _fget().
*/
if (onfiles > NDFILE) {
fo = (struct freetable *)&otable[onfiles];
fdp0 = (struct filedesc0 *)fdp;
fo->ft_table = otable;
SLIST_INSERT_HEAD(&fdp0->fd_free, fo, ft_next);
}
if (NDSLOTS(nnfiles) > NDSLOTS(onfiles)) {
bcopy(fdp->fd_map, nmap, NDSLOTS(onfiles) * sizeof(*nmap));
if (NDSLOTS(onfiles) > NDSLOTS(NDFILE))
free(fdp->fd_map, M_FILEDESC);
fdp->fd_map = nmap;
}
fdp->fd_nfiles = nnfiles;
}
1994-05-24 10:09:53 +00:00
/*
* Allocate a file descriptor for the process.
*/
int
fdalloc(struct thread *td, int minfd, int *result)
1994-05-24 10:09:53 +00:00
{
struct proc *p = td->td_proc;
struct filedesc *fdp = p->p_fd;
int fd = -1, maxfd, allocfd;
#ifdef RACCT
int error;
#endif
1994-05-24 10:09:53 +00:00
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK_ASSERT(fdp);
if (fdp->fd_freefile > minfd)
minfd = fdp->fd_freefile;
Locking for the per-process resource limits structure. - struct plimit includes a mutex to protect a reference count. The plimit structure is treated similarly to struct ucred in that is is always copy on write, so having a reference to a structure is sufficient to read from it without needing a further lock. - The proc lock protects the p_limit pointer and must be held while reading limits from a process to keep the limit structure from changing out from under you while reading from it. - Various global limits that are ints are not protected by a lock since int writes are atomic on all the archs we support and thus a lock wouldn't buy us anything. - All accesses to individual resource limits from a process are abstracted behind a simple lim_rlimit(), lim_max(), and lim_cur() API that return either an rlimit, or the current or max individual limit of the specified resource from a process. - dosetrlimit() was renamed to kern_setrlimit() to match existing style of other similar syscall helper functions. - The alpha OSF/1 compat layer no longer calls getrlimit() and setrlimit() (it didn't used the stackgap when it should have) but uses lim_rlimit() and kern_setrlimit() instead. - The svr4 compat no longer uses the stackgap for resource limits calls, but uses lim_rlimit() and kern_setrlimit() instead. - The ibcs2 compat no longer uses the stackgap for resource limits. It also no longer uses the stackgap for accessing sysctl's for the ibcs2_sysconf() syscall but uses kernel_sysctl() instead. As a result, ibcs2_sysconf() no longer needs Giant. - The p_rlimit macro no longer exists. Submitted by: mtm (mostly, I only did a few cleanups and catchups) Tested on: i386 Compiled on: alpha, amd64
2004-02-04 21:52:57 +00:00
PROC_LOCK(p);
maxfd = min((int)lim_cur(p, RLIMIT_NOFILE), maxfilesperproc);
PROC_UNLOCK(p);
1994-05-24 10:09:53 +00:00
/*
* Search the bitmap for a free descriptor starting at minfd.
* If none is found, grow the file table.
1994-05-24 10:09:53 +00:00
*/
fd = fd_first_free(fdp, minfd, fdp->fd_nfiles);
if (fd >= maxfd)
return (EMFILE);
if (fd >= fdp->fd_nfiles) {
allocfd = min(fd * 2, maxfd);
#ifdef RACCT
PROC_LOCK(p);
error = racct_set(p, RACCT_NOFILE, allocfd);
PROC_UNLOCK(p);
if (error != 0)
return (EMFILE);
#endif
/*
* fd is already equal to first free descriptor >= minfd, so
* we only need to grow the table and we are done.
*/
fdgrowtable(fdp, allocfd);
1994-05-24 10:09:53 +00:00
}
/*
* Perform some sanity checks, then mark the file descriptor as
* used and return it to the caller.
*/
KASSERT(fd >= 0 && fd < min(maxfd, fdp->fd_nfiles),
("invalid descriptor %d", fd));
KASSERT(!fdisused(fdp, fd),
("fd_first_free() returned non-free descriptor"));
KASSERT(fdp->fd_ofiles[fd] == NULL, ("file descriptor isn't free"));
KASSERT(fdp->fd_ofileflags[fd] == 0, ("file flags are set"));
fdused(fdp, fd);
*result = fd;
return (0);
1994-05-24 10:09:53 +00:00
}
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Check to see whether n user file descriptors are available to the process
* p.
1994-05-24 10:09:53 +00:00
*/
int
fdavail(struct thread *td, int n)
1994-05-24 10:09:53 +00:00
{
struct proc *p = td->td_proc;
2003-01-01 01:05:54 +00:00
struct filedesc *fdp = td->td_proc->p_fd;
int i, lim, last;
1994-05-24 10:09:53 +00:00
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_LOCK_ASSERT(fdp);
/*
* XXX: This is only called from uipc_usrreq.c:unp_externalize();
* call racct_add() from there instead of dealing with containers
* here.
*/
Locking for the per-process resource limits structure. - struct plimit includes a mutex to protect a reference count. The plimit structure is treated similarly to struct ucred in that is is always copy on write, so having a reference to a structure is sufficient to read from it without needing a further lock. - The proc lock protects the p_limit pointer and must be held while reading limits from a process to keep the limit structure from changing out from under you while reading from it. - Various global limits that are ints are not protected by a lock since int writes are atomic on all the archs we support and thus a lock wouldn't buy us anything. - All accesses to individual resource limits from a process are abstracted behind a simple lim_rlimit(), lim_max(), and lim_cur() API that return either an rlimit, or the current or max individual limit of the specified resource from a process. - dosetrlimit() was renamed to kern_setrlimit() to match existing style of other similar syscall helper functions. - The alpha OSF/1 compat layer no longer calls getrlimit() and setrlimit() (it didn't used the stackgap when it should have) but uses lim_rlimit() and kern_setrlimit() instead. - The svr4 compat no longer uses the stackgap for resource limits calls, but uses lim_rlimit() and kern_setrlimit() instead. - The ibcs2 compat no longer uses the stackgap for resource limits. It also no longer uses the stackgap for accessing sysctl's for the ibcs2_sysconf() syscall but uses kernel_sysctl() instead. As a result, ibcs2_sysconf() no longer needs Giant. - The p_rlimit macro no longer exists. Submitted by: mtm (mostly, I only did a few cleanups and catchups) Tested on: i386 Compiled on: alpha, amd64
2004-02-04 21:52:57 +00:00
PROC_LOCK(p);
lim = min((int)lim_cur(p, RLIMIT_NOFILE), maxfilesperproc);
PROC_UNLOCK(p);
1994-05-24 10:09:53 +00:00
if ((i = lim - fdp->fd_nfiles) > 0 && (n -= i) <= 0)
return (1);
last = min(fdp->fd_nfiles, lim);
for (i = fdp->fd_freefile; i < last; i++) {
if (fdp->fd_ofiles[i] == NULL && --n <= 0)
1994-05-24 10:09:53 +00:00
return (1);
}
1994-05-24 10:09:53 +00:00
return (0);
}
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Create a new open file structure and allocate a file decriptor for the
* process that refers to it. We add one reference to the file for the
* descriptor table and one reference for resultfp. This is to prevent us
* being preempted and the entry in the descriptor table closed after we
* release the FILEDESC lock.
1994-05-24 10:09:53 +00:00
*/
int
falloc(struct thread *td, struct file **resultfp, int *resultfd, int flags)
1994-05-24 10:09:53 +00:00
{
struct file *fp;
int error, fd;
error = falloc_noinstall(td, &fp);
if (error)
return (error); /* no reference held on error */
error = finstall(td, fp, &fd, flags);
if (error) {
fdrop(fp, td); /* one reference (fp only) */
return (error);
}
if (resultfp != NULL)
*resultfp = fp; /* copy out result */
else
fdrop(fp, td); /* release local reference */
if (resultfd != NULL)
*resultfd = fd;
return (0);
}
/*
* Create a new open file structure without allocating a file descriptor.
*/
int
falloc_noinstall(struct thread *td, struct file **resultfp)
{
struct file *fp;
int maxuserfiles = maxfiles - (maxfiles / 20);
static struct timeval lastfail;
static int curfail;
1994-05-24 10:09:53 +00:00
KASSERT(resultfp != NULL, ("%s: resultfp == NULL", __func__));
if ((openfiles >= maxuserfiles &&
priv_check(td, PRIV_MAXFILES) != 0) ||
openfiles >= maxfiles) {
if (ppsratecheck(&lastfail, &curfail, 1)) {
printf("kern.maxfiles limit exceeded by uid %i, "
"please see tuning(7).\n", td->td_ucred->cr_ruid);
2004-01-11 19:39:14 +00:00
}
1994-05-24 10:09:53 +00:00
return (ENFILE);
}
atomic_add_int(&openfiles, 1);
fp = uma_zalloc(file_zone, M_WAITOK | M_ZERO);
refcount_init(&fp->f_count, 1);
fp->f_cred = crhold(td->td_ucred);
fp->f_ops = &badfileops;
fp->f_data = NULL;
fp->f_vnode = NULL;
*resultfp = fp;
return (0);
}
/*
* Install a file in a file descriptor table.
*/
int
finstall(struct thread *td, struct file *fp, int *fd, int flags)
{
struct filedesc *fdp = td->td_proc->p_fd;
int error;
KASSERT(fd != NULL, ("%s: fd == NULL", __func__));
KASSERT(fp != NULL, ("%s: fp == NULL", __func__));
FILEDESC_XLOCK(fdp);
if ((error = fdalloc(td, 0, fd))) {
FILEDESC_XUNLOCK(fdp);
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
return (error);
}
fhold(fp);
fdp->fd_ofiles[*fd] = fp;
if ((flags & O_CLOEXEC) != 0)
fdp->fd_ofileflags[*fd] |= UF_EXCLOSE;
FILEDESC_XUNLOCK(fdp);
1994-05-24 10:09:53 +00:00
return (0);
}
/*
* Build a new filedesc structure from another.
* Copy the current, root, and jail root vnode references.
*/
struct filedesc *
fdinit(struct filedesc *fdp)
{
2003-01-01 01:05:54 +00:00
struct filedesc0 *newfdp;
newfdp = malloc(sizeof *newfdp, M_FILEDESC, M_WAITOK | M_ZERO);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_LOCK_INIT(&newfdp->fd_fd);
if (fdp != NULL) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
newfdp->fd_fd.fd_cdir = fdp->fd_cdir;
if (newfdp->fd_fd.fd_cdir)
VREF(newfdp->fd_fd.fd_cdir);
newfdp->fd_fd.fd_rdir = fdp->fd_rdir;
if (newfdp->fd_fd.fd_rdir)
VREF(newfdp->fd_fd.fd_rdir);
newfdp->fd_fd.fd_jdir = fdp->fd_jdir;
if (newfdp->fd_fd.fd_jdir)
VREF(newfdp->fd_fd.fd_jdir);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
}
/* Create the file descriptor table. */
newfdp->fd_fd.fd_refcnt = 1;
newfdp->fd_fd.fd_holdcnt = 1;
newfdp->fd_fd.fd_cmask = CMASK;
newfdp->fd_fd.fd_ofiles = newfdp->fd_dfiles;
newfdp->fd_fd.fd_ofileflags = newfdp->fd_dfileflags;
newfdp->fd_fd.fd_nfiles = NDFILE;
newfdp->fd_fd.fd_map = newfdp->fd_dmap;
newfdp->fd_fd.fd_lastfile = -1;
return (&newfdp->fd_fd);
}
static struct filedesc *
fdhold(struct proc *p)
{
struct filedesc *fdp;
mtx_lock(&fdesc_mtx);
fdp = p->p_fd;
if (fdp != NULL)
fdp->fd_holdcnt++;
mtx_unlock(&fdesc_mtx);
return (fdp);
}
static void
fddrop(struct filedesc *fdp)
{
struct filedesc0 *fdp0;
struct freetable *ft;
int i;
mtx_lock(&fdesc_mtx);
i = --fdp->fd_holdcnt;
mtx_unlock(&fdesc_mtx);
if (i > 0)
return;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_LOCK_DESTROY(fdp);
fdp0 = (struct filedesc0 *)fdp;
while ((ft = SLIST_FIRST(&fdp0->fd_free)) != NULL) {
SLIST_REMOVE_HEAD(&fdp0->fd_free, ft_next);
free(ft->ft_table, M_FILEDESC);
}
free(fdp, M_FILEDESC);
}
/*
* Share a filedesc structure.
*/
struct filedesc *
fdshare(struct filedesc *fdp)
{
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
fdp->fd_refcnt++;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
return (fdp);
}
/*
* Unshare a filedesc structure, if necessary by making a copy
*/
void
fdunshare(struct proc *p, struct thread *td)
{
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(p->p_fd);
if (p->p_fd->fd_refcnt > 1) {
struct filedesc *tmp;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(p->p_fd);
tmp = fdcopy(p->p_fd);
fdfree(td);
p->p_fd = tmp;
} else
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(p->p_fd);
}
1994-05-24 10:09:53 +00:00
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Copy a filedesc structure. A NULL pointer in returns a NULL reference,
* this is to ease callers, not catch errors.
1994-05-24 10:09:53 +00:00
*/
struct filedesc *
fdcopy(struct filedesc *fdp)
1994-05-24 10:09:53 +00:00
{
struct filedesc *newfdp;
int i;
1994-05-24 10:09:53 +00:00
/* Certain daemons might not have file descriptors. */
if (fdp == NULL)
return (NULL);
newfdp = fdinit(fdp);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_SLOCK(fdp);
while (fdp->fd_lastfile >= newfdp->fd_nfiles) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_SUNLOCK(fdp);
FILEDESC_XLOCK(newfdp);
fdgrowtable(newfdp, fdp->fd_lastfile + 1);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(newfdp);
FILEDESC_SLOCK(fdp);
}
/* copy all passable descriptors (i.e. not kqueue) */
newfdp->fd_freefile = -1;
for (i = 0; i <= fdp->fd_lastfile; ++i) {
if (fdisused(fdp, i) &&
(fdp->fd_ofiles[i]->f_ops->fo_flags & DFLAG_PASSABLE) &&
fdp->fd_ofiles[i]->f_ops != &badfileops) {
newfdp->fd_ofiles[i] = fdp->fd_ofiles[i];
newfdp->fd_ofileflags[i] = fdp->fd_ofileflags[i];
fhold(newfdp->fd_ofiles[i]);
newfdp->fd_lastfile = i;
} else {
if (newfdp->fd_freefile == -1)
newfdp->fd_freefile = i;
}
}
newfdp->fd_cmask = fdp->fd_cmask;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_SUNLOCK(fdp);
FILEDESC_XLOCK(newfdp);
for (i = 0; i <= newfdp->fd_lastfile; ++i)
if (newfdp->fd_ofiles[i] != NULL)
fdused(newfdp, i);
if (newfdp->fd_freefile == -1)
newfdp->fd_freefile = i;
FILEDESC_XUNLOCK(newfdp);
1994-05-24 10:09:53 +00:00
return (newfdp);
}
/*
* Release a filedesc structure.
*/
void
fdfree(struct thread *td)
1994-05-24 10:09:53 +00:00
{
2003-01-01 01:05:54 +00:00
struct filedesc *fdp;
int i, locked;
struct filedesc_to_leader *fdtol;
struct file *fp;
struct vnode *cdir, *jdir, *rdir, *vp;
struct flock lf;
1994-05-24 10:09:53 +00:00
/* Certain daemons might not have file descriptors. */
fdp = td->td_proc->p_fd;
if (fdp == NULL)
return;
#ifdef RACCT
PROC_LOCK(td->td_proc);
racct_set(td->td_proc, RACCT_NOFILE, 0);
PROC_UNLOCK(td->td_proc);
#endif
/* Check for special need to clear POSIX style locks */
fdtol = td->td_proc->p_fdtol;
if (fdtol != NULL) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
KASSERT(fdtol->fdl_refcount > 0,
("filedesc_to_refcount botch: fdl_refcount=%d",
fdtol->fdl_refcount));
if (fdtol->fdl_refcount == 1 &&
(td->td_proc->p_leader->p_flag & P_ADVLOCK) != 0) {
for (i = 0; i <= fdp->fd_lastfile; i++) {
fp = fdp->fd_ofiles[i];
if (fp == NULL || fp->f_type != DTYPE_VNODE)
continue;
fhold(fp);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
lf.l_whence = SEEK_SET;
lf.l_start = 0;
lf.l_len = 0;
lf.l_type = F_UNLCK;
vp = fp->f_vnode;
locked = VFS_LOCK_GIANT(vp->v_mount);
(void) VOP_ADVLOCK(vp,
2012-06-14 12:37:41 +00:00
(caddr_t)td->td_proc->p_leader, F_UNLCK,
&lf, F_POSIX);
VFS_UNLOCK_GIANT(locked);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
fdrop(fp, td);
}
}
retry:
if (fdtol->fdl_refcount == 1) {
if (fdp->fd_holdleaderscount > 0 &&
(td->td_proc->p_leader->p_flag & P_ADVLOCK) != 0) {
/*
* close() or do_dup() has cleared a reference
* in a shared file descriptor table.
*/
fdp->fd_holdleaderswakeup = 1;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
sx_sleep(&fdp->fd_holdleaderscount,
FILEDESC_LOCK(fdp), PLOCK, "fdlhold", 0);
goto retry;
}
if (fdtol->fdl_holdcount > 0) {
2004-01-11 19:39:14 +00:00
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Ensure that fdtol->fdl_leader remains
* valid in closef().
*/
fdtol->fdl_wakeup = 1;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
sx_sleep(fdtol, FILEDESC_LOCK(fdp), PLOCK,
"fdlhold", 0);
goto retry;
}
}
fdtol->fdl_refcount--;
if (fdtol->fdl_refcount == 0 &&
fdtol->fdl_holdcount == 0) {
fdtol->fdl_next->fdl_prev = fdtol->fdl_prev;
fdtol->fdl_prev->fdl_next = fdtol->fdl_next;
} else
fdtol = NULL;
td->td_proc->p_fdtol = NULL;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
if (fdtol != NULL)
free(fdtol, M_FILEDESC_TO_LEADER);
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
i = --fdp->fd_refcnt;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
if (i > 0)
1994-05-24 10:09:53 +00:00
return;
for (i = 0; i <= fdp->fd_lastfile; i++) {
fp = fdp->fd_ofiles[i];
if (fp != NULL) {
FILEDESC_XLOCK(fdp);
fdp->fd_ofiles[i] = NULL;
FILEDESC_XUNLOCK(fdp);
(void) closef(fp, td);
}
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
/* XXX This should happen earlier. */
mtx_lock(&fdesc_mtx);
td->td_proc->p_fd = NULL;
mtx_unlock(&fdesc_mtx);
1994-05-24 10:09:53 +00:00
if (fdp->fd_nfiles > NDFILE)
free(fdp->fd_ofiles, M_FILEDESC);
if (NDSLOTS(fdp->fd_nfiles) > NDSLOTS(NDFILE))
free(fdp->fd_map, M_FILEDESC);
fdp->fd_nfiles = 0;
cdir = fdp->fd_cdir;
fdp->fd_cdir = NULL;
rdir = fdp->fd_rdir;
fdp->fd_rdir = NULL;
jdir = fdp->fd_jdir;
fdp->fd_jdir = NULL;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
if (cdir) {
locked = VFS_LOCK_GIANT(cdir->v_mount);
vrele(cdir);
VFS_UNLOCK_GIANT(locked);
}
if (rdir) {
locked = VFS_LOCK_GIANT(rdir->v_mount);
vrele(rdir);
VFS_UNLOCK_GIANT(locked);
}
if (jdir) {
locked = VFS_LOCK_GIANT(jdir->v_mount);
vrele(jdir);
VFS_UNLOCK_GIANT(locked);
}
fddrop(fdp);
1994-05-24 10:09:53 +00:00
}
/*
2000-01-21 02:52:54 +00:00
* For setugid programs, we don't want to people to use that setugidness
* to generate error messages which write to a file which otherwise would
* otherwise be off-limits to the process. We check for filesystems where
* the vnode can change out from under us after execve (like [lin]procfs).
*
* Since setugidsafety calls this only for fd 0, 1 and 2, this check is
2005-03-08 00:58:50 +00:00
* sufficient. We also don't check for setugidness since we know we are.
*/
static int
is_unsafe(struct file *fp)
{
if (fp->f_type == DTYPE_VNODE) {
struct vnode *vp = fp->f_vnode;
if ((vp->v_vflag & VV_PROCDEP) != 0)
return (1);
}
return (0);
}
/*
* Make this setguid thing safe, if at all possible.
*/
void
setugidsafety(struct thread *td)
{
struct filedesc *fdp;
2003-01-01 01:05:54 +00:00
int i;
/* Certain daemons might not have file descriptors. */
fdp = td->td_proc->p_fd;
if (fdp == NULL)
return;
/*
* Note: fdp->fd_ofiles may be reallocated out from under us while
* we are blocked in a close. Be careful!
*/
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
for (i = 0; i <= fdp->fd_lastfile; i++) {
if (i > 2)
break;
if (fdp->fd_ofiles[i] && is_unsafe(fdp->fd_ofiles[i])) {
struct file *fp;
knote_fdclose(td, i);
/*
* NULL-out descriptor prior to close to avoid
* a race while close blocks.
*/
fp = fdp->fd_ofiles[i];
fdp->fd_ofiles[i] = NULL;
fdp->fd_ofileflags[i] = 0;
fdunused(fdp, i);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
(void) closef(fp, td);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
}
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
}
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* If a specific file object occupies a specific file descriptor, close the
* file descriptor entry and drop a reference on the file object. This is a
* convenience function to handle a subsequent error in a function that calls
* falloc() that handles the race that another thread might have closed the
* file descriptor out from under the thread creating the file object.
*/
void
fdclose(struct filedesc *fdp, struct file *fp, int idx, struct thread *td)
{
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
if (fdp->fd_ofiles[idx] == fp) {
fdp->fd_ofiles[idx] = NULL;
fdunused(fdp, idx);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
fdrop(fp, td);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
} else
FILEDESC_XUNLOCK(fdp);
}
/*
* Close any files on exec?
*/
void
fdcloseexec(struct thread *td)
{
struct filedesc *fdp;
struct file *fp;
2003-01-01 01:05:54 +00:00
int i;
/* Certain daemons might not have file descriptors. */
fdp = td->td_proc->p_fd;
if (fdp == NULL)
return;
/*
* We cannot cache fd_ofiles or fd_ofileflags since operations
* may block and rip them out from under us.
*/
2012-06-14 12:37:41 +00:00
FILEDESC_XLOCK(fdp);
for (i = 0; i <= fdp->fd_lastfile; i++) {
2012-06-14 12:37:41 +00:00
fp = fdp->fd_ofiles[i];
if (fp != NULL && (fp->f_type == DTYPE_MQUEUE ||
(fdp->fd_ofileflags[i] & UF_EXCLOSE))) {
fdp->fd_ofiles[i] = NULL;
fdp->fd_ofileflags[i] = 0;
fdunused(fdp, i);
(void) closefp(fdp, i, fp, td, 0);
/* closefp() drops the FILEDESC lock. */
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
}
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
}
/*
* It is unsafe for set[ug]id processes to be started with file
* descriptors 0..2 closed, as these descriptors are given implicit
* significance in the Standard C library. fdcheckstd() will create a
* descriptor referencing /dev/null for each of stdin, stdout, and
* stderr that is not already open.
*/
int
fdcheckstd(struct thread *td)
{
struct filedesc *fdp;
register_t retval, save;
int i, error, devnull;
fdp = td->td_proc->p_fd;
if (fdp == NULL)
return (0);
KASSERT(fdp->fd_refcnt == 1, ("the fdtable should not be shared"));
devnull = -1;
error = 0;
for (i = 0; i < 3; i++) {
if (fdp->fd_ofiles[i] != NULL)
continue;
if (devnull < 0) {
save = td->td_retval[0];
error = kern_open(td, "/dev/null", UIO_SYSSPACE,
O_RDWR, 0);
devnull = td->td_retval[0];
td->td_retval[0] = save;
if (error)
break;
KASSERT(devnull == i, ("oof, we didn't get our fd"));
} else {
- Change falloc() to acquire an fd from the process table last so that it can do it w/o needing to hold the filelist_lock sx lock. - fdalloc() doesn't need Giant to call free() anymore. It also doesn't need to drop and reacquire the filedesc lock around free() now as a result. - Try to make the code that copies fd tables when extending the fd table in fdalloc() a bit more readable by performing assignments in separate statements. This is still a bit ugly though. - Use max() instead of an if statement so to figure out the starting point in the search-for-a-free-fd loop in fdalloc() so it reads better next to the min() in the previous line. - Don't grow nfiles in steps up to the size needed if we dup2() to some really large number. Go ahead and double 'nfiles' in a loop prior to doing the malloc(). - malloc() doesn't need Giant now. - Use malloc() and free() instead of MALLOC() and FREE() in fdalloc(). - Check to see if the size we are going to grow to is too big, not if the current size of the fd table is too big in the loop in fdalloc(). This means if we are out of space or if dup2() requests too high of a fd, then we will return an error before we go off and try to allocate some huge table and copy the existing table into it. - Move all of the logic for dup'ing a file descriptor into do_dup() instead of putting some of it in do_dup() and duplicating other parts in four different places. This makes dup(), dup2(), and fcntl(F_DUPFD) basically wrappers of do_dup now. fcntl() still has an extra check since it uses a different error return value in one case then the other functions. - Add a KASSERT() for an assertion that may not always be true where the fdcheckstd() function assumes that falloc() returns the fd requested and not some other fd. I think that the assertion is always true because we are always single-threaded when we get to this point, but if one was using rfork() and another process sharing the fd table were playing with the fd table, there might could be a problem. - To handle the problem of a file descriptor we are dup()'ing being closed out from under us in dup() in general, do_dup() now obtains a reference on the file in question before calling fdalloc(). If after the call to fdalloc() the file for the fd we are dup'ing is a different file, then we drop our reference on the original file and return EBADF. This race was only handled in the dup2() case before and would just retry the operation. The error return allows the user to know they are being stupid since they have a locking bug in their app instead of dup'ing some other descriptor and returning it to them. Tested on: i386, alpha, sparc64
2002-09-03 20:16:31 +00:00
error = do_dup(td, DUP_FIXED, devnull, i, &retval);
if (error != 0)
break;
}
}
return (error);
}
1994-05-24 10:09:53 +00:00
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Internal form of close. Decrement reference count on file structure.
* Note: td may be NULL when closing a file that was being passed in a
* message.
*
* XXXRW: Giant is not required for the caller, but often will be held; this
* makes it moderately likely the Giant will be recursed in the VFS case.
1994-05-24 10:09:53 +00:00
*/
int
closef(struct file *fp, struct thread *td)
1994-05-24 10:09:53 +00:00
{
struct vnode *vp;
struct flock lf;
struct filedesc_to_leader *fdtol;
struct filedesc *fdp;
struct file *fp_object;
1994-05-24 10:09:53 +00:00
/*
* POSIX record locking dictates that any close releases ALL
* locks owned by this process. This is handled by setting
* a flag in the unlock to free ONLY locks obeying POSIX
* semantics, and not to free BSD-style file locks.
* If the descriptor was in a message, POSIX-style locks
* aren't passed with the descriptor, and the thread pointer
* will be NULL. Callers should be careful only to pass a
* NULL thread pointer when there really is no owning
* context that might have locks, or the locks will be
* leaked.
*
* If this is a capability, we do lock processing under the underlying
* node, not the capability itself.
1994-05-24 10:09:53 +00:00
*/
(void)cap_funwrap(fp, 0, &fp_object);
2012-06-14 12:37:41 +00:00
if (fp_object->f_type == DTYPE_VNODE && td != NULL) {
int vfslocked;
vp = fp_object->f_vnode;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
if ((td->td_proc->p_leader->p_flag & P_ADVLOCK) != 0) {
lf.l_whence = SEEK_SET;
lf.l_start = 0;
lf.l_len = 0;
lf.l_type = F_UNLCK;
(void) VOP_ADVLOCK(vp, (caddr_t)td->td_proc->p_leader,
2012-06-14 12:37:41 +00:00
F_UNLCK, &lf, F_POSIX);
}
fdtol = td->td_proc->p_fdtol;
if (fdtol != NULL) {
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Handle special case where file descriptor table is
* shared between multiple process leaders.
*/
fdp = td->td_proc->p_fd;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
for (fdtol = fdtol->fdl_next;
fdtol != td->td_proc->p_fdtol;
fdtol = fdtol->fdl_next) {
if ((fdtol->fdl_leader->p_flag &
P_ADVLOCK) == 0)
continue;
fdtol->fdl_holdcount++;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
lf.l_whence = SEEK_SET;
lf.l_start = 0;
lf.l_len = 0;
lf.l_type = F_UNLCK;
vp = fp_object->f_vnode;
(void) VOP_ADVLOCK(vp,
2012-06-14 12:37:41 +00:00
(caddr_t)fdtol->fdl_leader, F_UNLCK, &lf,
F_POSIX);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
fdtol->fdl_holdcount--;
if (fdtol->fdl_holdcount == 0 &&
fdtol->fdl_wakeup != 0) {
fdtol->fdl_wakeup = 0;
wakeup(fdtol);
}
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
}
VFS_UNLOCK_GIANT(vfslocked);
1994-05-24 10:09:53 +00:00
}
return (fdrop(fp, td));
}
/*
* Initialize the file pointer with the specified properties.
*
* The ops are set with release semantics to be certain that the flags, type,
* and data are visible when ops is. This is to prevent ops methods from being
* called with bad data.
*/
void
finit(struct file *fp, u_int flag, short type, void *data, struct fileops *ops)
{
fp->f_data = data;
fp->f_flag = flag;
fp->f_type = type;
atomic_store_rel_ptr((volatile uintptr_t *)&fp->f_ops, (uintptr_t)ops);
}
struct file *
fget_unlocked(struct filedesc *fdp, int fd)
{
struct file *fp;
u_int count;
if (fd < 0 || fd >= fdp->fd_nfiles)
return (NULL);
/*
* Fetch the descriptor locklessly. We avoid fdrop() races by
* never raising a refcount above 0. To accomplish this we have
* to use a cmpset loop rather than an atomic_add. The descriptor
* must be re-verified once we acquire a reference to be certain
* that the identity is still correct and we did not lose a race
* due to preemption.
*/
for (;;) {
fp = fdp->fd_ofiles[fd];
if (fp == NULL)
break;
count = fp->f_count;
if (count == 0)
continue;
/*
* Use an acquire barrier to prevent caching of fd_ofiles
* so it is refreshed for verification.
*/
if (atomic_cmpset_acq_int(&fp->f_count, count, count + 1) != 1)
continue;
if (fp == fdp->fd_ofiles[fd])
break;
fdrop(fp, curthread);
}
return (fp);
}
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Extract the file pointer associated with the specified descriptor for the
* current user process.
*
* If the descriptor doesn't exist or doesn't match 'flags', EBADF is
* returned.
*
* If the FGET_GETCAP flag is set, the capability itself will be returned.
* Calling _fget() with FGET_GETCAP on a non-capability will return EINVAL.
* Otherwise, if the file is a capability, its rights will be checked against
* the capability rights mask, and if successful, the object will be unwrapped.
*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* If an error occured the non-zero error is returned and *fpp is set to
* NULL. Otherwise *fpp is held and set and zero is returned. Caller is
* responsible for fdrop().
*/
#define FGET_GETCAP 0x00000001
static __inline int
_fget(struct thread *td, int fd, struct file **fpp, int flags,
cap_rights_t needrights, cap_rights_t *haverightsp, u_char *maxprotp,
int fget_flags)
{
struct filedesc *fdp;
struct file *fp;
#ifdef CAPABILITIES
struct file *fp_fromcap;
#endif
int error;
*fpp = NULL;
if (td == NULL || (fdp = td->td_proc->p_fd) == NULL)
return (EBADF);
if ((fp = fget_unlocked(fdp, fd)) == NULL)
return (EBADF);
if (fp->f_ops == &badfileops) {
fdrop(fp, td);
return (EBADF);
}
#ifdef CAPABILITIES
/*
* If this is a capability, what rights does it have?
*/
if (haverightsp != NULL) {
if (fp->f_type == DTYPE_CAPABILITY)
*haverightsp = cap_rights(fp);
else
*haverightsp = CAP_MASK_VALID;
}
/*
* If a capability has been requested, return the capability directly.
* Otherwise, check capability rights, extract the underlying object,
* and check its access flags.
*/
if (fget_flags & FGET_GETCAP) {
if (fp->f_type != DTYPE_CAPABILITY) {
fdrop(fp, td);
return (EINVAL);
}
} else {
if (maxprotp == NULL)
error = cap_funwrap(fp, needrights, &fp_fromcap);
else
error = cap_funwrap_mmap(fp, needrights, maxprotp,
&fp_fromcap);
if (error != 0) {
fdrop(fp, td);
return (error);
}
/*
* If we've unwrapped a file, drop the original capability
* and hold the new descriptor. fp after this point refers to
* the actual (unwrapped) object, not the capability.
*/
if (fp != fp_fromcap) {
fhold(fp_fromcap);
fdrop(fp, td);
fp = fp_fromcap;
}
}
#else /* !CAPABILITIES */
KASSERT(fp->f_type != DTYPE_CAPABILITY,
("%s: saw capability", __func__));
if (maxprotp != NULL)
*maxprotp = VM_PROT_ALL;
#endif /* CAPABILITIES */
/*
* FREAD and FWRITE failure return EBADF as per POSIX.
*/
error = 0;
switch (flags) {
case FREAD:
case FWRITE:
if ((fp->f_flag & flags) == 0)
error = EBADF;
break;
case FEXEC:
if ((fp->f_flag & (FREAD | FEXEC)) == 0 ||
((fp->f_flag & FWRITE) != 0))
error = EBADF;
break;
case 0:
break;
default:
KASSERT(0, ("wrong flags"));
}
if (error != 0) {
fdrop(fp, td);
return (error);
}
*fpp = fp;
return (0);
}
int
fget(struct thread *td, int fd, cap_rights_t rights, struct file **fpp)
{
return(_fget(td, fd, fpp, 0, rights, NULL, NULL, 0));
}
int
fget_mmap(struct thread *td, int fd, cap_rights_t rights, u_char *maxprotp,
struct file **fpp)
{
return (_fget(td, fd, fpp, 0, rights, NULL, maxprotp, 0));
}
int
fget_read(struct thread *td, int fd, cap_rights_t rights, struct file **fpp)
{
return(_fget(td, fd, fpp, FREAD, rights, NULL, NULL, 0));
}
int
fget_write(struct thread *td, int fd, cap_rights_t rights, struct file **fpp)
{
return (_fget(td, fd, fpp, FWRITE, rights, NULL, NULL, 0));
}
/*
* Unlike the other fget() calls, which accept and check capability rights
* but never return capabilities, fgetcap() returns the capability but doesn't
* check capability rights.
*/
int
fgetcap(struct thread *td, int fd, struct file **fpp)
{
return (_fget(td, fd, fpp, 0, 0, NULL, NULL, FGET_GETCAP));
}
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Like fget() but loads the underlying vnode, or returns an error if the
* descriptor does not represent a vnode. Note that pipes use vnodes but
* never have VM objects. The returned vnode will be vref()'d.
*
* XXX: what about the unused flags ?
*/
static __inline int
_fgetvp(struct thread *td, int fd, int flags, cap_rights_t needrights,
cap_rights_t *haverightsp, struct vnode **vpp)
{
struct file *fp;
int error;
*vpp = NULL;
if ((error = _fget(td, fd, &fp, flags, needrights, haverightsp,
NULL, 0)) != 0)
return (error);
if (fp->f_vnode == NULL) {
error = EINVAL;
} else {
*vpp = fp->f_vnode;
vref(*vpp);
}
fdrop(fp, td);
return (error);
}
int
fgetvp(struct thread *td, int fd, cap_rights_t rights, struct vnode **vpp)
{
return (_fgetvp(td, fd, 0, rights, NULL, vpp));
}
int
fgetvp_rights(struct thread *td, int fd, cap_rights_t need, cap_rights_t *have,
struct vnode **vpp)
{
return (_fgetvp(td, fd, 0, need, have, vpp));
}
int
fgetvp_read(struct thread *td, int fd, cap_rights_t rights, struct vnode **vpp)
{
return (_fgetvp(td, fd, FREAD, rights, NULL, vpp));
}
int
fgetvp_exec(struct thread *td, int fd, cap_rights_t rights, struct vnode **vpp)
{
return (_fgetvp(td, fd, FEXEC, rights, NULL, vpp));
}
#ifdef notyet
int
fgetvp_write(struct thread *td, int fd, cap_rights_t rights,
struct vnode **vpp)
{
return (_fgetvp(td, fd, FWRITE, rights, NULL, vpp));
}
#endif
/*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* Like fget() but loads the underlying socket, or returns an error if the
* descriptor does not represent a socket.
*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* We bump the ref count on the returned socket. XXX Also obtain the SX lock
* in the future.
*
* Note: fgetsock() and fputsock() are deprecated, as consumers should rely
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* on their file descriptor reference to prevent the socket from being free'd
* during use.
*/
int
fgetsock(struct thread *td, int fd, cap_rights_t rights, struct socket **spp,
u_int *fflagp)
{
struct file *fp;
int error;
*spp = NULL;
if (fflagp != NULL)
*fflagp = 0;
if ((error = _fget(td, fd, &fp, 0, rights, NULL, NULL, 0)) != 0)
return (error);
if (fp->f_type != DTYPE_SOCKET) {
error = ENOTSOCK;
} else {
*spp = fp->f_data;
if (fflagp)
*fflagp = fp->f_flag;
SOCK_LOCK(*spp);
soref(*spp);
SOCK_UNLOCK(*spp);
}
fdrop(fp, td);
return (error);
}
/*
* Drop the reference count on the socket and XXX release the SX lock in the
* future. The last reference closes the socket.
*
* Note: fputsock() is deprecated, see comment for fgetsock().
*/
void
fputsock(struct socket *so)
{
Push acquisition of the accept mutex out of sofree() into the caller (sorele()/sotryfree()): - This permits the caller to acquire the accept mutex before the socket mutex, avoiding sofree() having to drop the socket mutex and re-order, which could lead to races permitting more than one thread to enter sofree() after a socket is ready to be free'd. - This also covers clearing of the so_pcb weak socket reference from the protocol to the socket, preventing races in clearing and evaluation of the reference such that sofree() might be called more than once on the same socket. This appears to close a race I was able to easily trigger by repeatedly opening and resetting TCP connections to a host, in which the tcp_close() code called as a result of the RST raced with the close() of the accepted socket in the user process resulting in simultaneous attempts to de-allocate the same socket. The new locking increases the overhead for operations that may potentially free the socket, so we will want to revise the synchronization strategy here as we normalize the reference counting model for sockets. The use of the accept mutex in freeing of sockets that are not listen sockets is primarily motivated by the potential need to remove the socket from the incomplete connection queue on its parent (listen) socket, so cleaning up the reference model here may allow us to substantially weaken the synchronization requirements. RELENG_5_3 candidate. MFC after: 3 days Reviewed by: dwhite Discussed with: gnn, dwhite, green Reported by: Marc UBM Bocklet <ubm at u-boot-man dot de> Reported by: Vlad <marchenko at gmail dot com>
2004-10-18 22:19:43 +00:00
ACCEPT_LOCK();
SOCK_LOCK(so);
CURVNET_SET(so->so_vnet);
sorele(so);
CURVNET_RESTORE();
}
/*
* Handle the last reference to a file being closed.
*
* No special capability handling here, as the capability's fo_close will run
* instead of the object here, and perform any necessary drop on the object.
*/
int
_fdrop(struct file *fp, struct thread *td)
{
int error;
error = 0;
if (fp->f_count != 0)
panic("fdrop: count %d", fp->f_count);
if (fp->f_ops != &badfileops)
error = fo_close(fp, td);
atomic_subtract_int(&openfiles, 1);
crfree(fp->f_cred);
free(fp->f_advice, M_FADVISE);
uma_zfree(file_zone, fp);
1994-05-24 10:09:53 +00:00
return (error);
}
/*
* Apply an advisory lock on a file descriptor.
*
* Just attempt to get a record lock of the requested type on the entire file
* (l_whence = SEEK_SET, l_start = 0, l_len = 0).
1994-05-24 10:09:53 +00:00
*/
#ifndef _SYS_SYSPROTO_H_
1994-05-24 10:09:53 +00:00
struct flock_args {
int fd;
int how;
};
#endif
1994-05-24 10:09:53 +00:00
/* ARGSUSED */
int
sys_flock(struct thread *td, struct flock_args *uap)
1994-05-24 10:09:53 +00:00
{
struct file *fp;
1994-05-24 10:09:53 +00:00
struct vnode *vp;
struct flock lf;
int vfslocked;
int error;
if ((error = fget(td, uap->fd, CAP_FLOCK, &fp)) != 0)
return (error);
if (fp->f_type != DTYPE_VNODE) {
fdrop(fp, td);
return (EOPNOTSUPP);
}
vp = fp->f_vnode;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1994-05-24 10:09:53 +00:00
lf.l_whence = SEEK_SET;
lf.l_start = 0;
lf.l_len = 0;
if (uap->how & LOCK_UN) {
lf.l_type = F_UNLCK;
atomic_clear_int(&fp->f_flag, FHASLOCK);
error = VOP_ADVLOCK(vp, (caddr_t)fp, F_UNLCK, &lf, F_FLOCK);
goto done2;
1994-05-24 10:09:53 +00:00
}
if (uap->how & LOCK_EX)
lf.l_type = F_WRLCK;
else if (uap->how & LOCK_SH)
lf.l_type = F_RDLCK;
else {
error = EBADF;
goto done2;
}
atomic_set_int(&fp->f_flag, FHASLOCK);
error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf,
(uap->how & LOCK_NB) ? F_FLOCK : F_FLOCK | F_WAIT);
done2:
fdrop(fp, td);
VFS_UNLOCK_GIANT(vfslocked);
return (error);
1994-05-24 10:09:53 +00:00
}
/*
* Duplicate the specified descriptor to a free descriptor.
*/
int
dupfdopen(struct thread *td, struct filedesc *fdp, int dfd, int mode, int openerror, int *indxp)
1994-05-24 10:09:53 +00:00
{
struct file *fp;
int error, indx;
1995-05-30 08:16:23 +00:00
KASSERT(openerror == ENODEV || openerror == ENXIO,
("unexpected error %d in %s", openerror, __func__));
1994-05-24 10:09:53 +00:00
/*
* If the to-be-dup'd fd number is greater than the allowed number
* of file descriptors, or the fd to be dup'd has already been
* closed, then reject.
1994-05-24 10:09:53 +00:00
*/
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
if ((fp = fget_locked(fdp, dfd)) == NULL) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
1994-05-24 10:09:53 +00:00
return (EBADF);
}
1994-05-24 10:09:53 +00:00
error = fdalloc(td, 0, &indx);
if (error != 0) {
FILEDESC_XUNLOCK(fdp);
return (error);
}
1994-05-24 10:09:53 +00:00
/*
* There are two cases of interest here.
*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* For ENODEV simply dup (dfd) to file descriptor (indx) and return.
1994-05-24 10:09:53 +00:00
*
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
* For ENXIO steal away the file structure from (dfd) and store it in
* (indx). (dfd) is effectively closed by this operation.
1994-05-24 10:09:53 +00:00
*/
switch (openerror) {
1994-05-24 10:09:53 +00:00
case ENODEV:
/*
* Check that the mode the file is being opened for is a
* subset of the mode of the existing descriptor.
*/
if (((mode & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
fdunused(fdp, indx);
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
1994-05-24 10:09:53 +00:00
return (EACCES);
}
fdp->fd_ofiles[indx] = fp;
1994-05-24 10:09:53 +00:00
fdp->fd_ofileflags[indx] = fdp->fd_ofileflags[dfd];
fhold(fp);
break;
1994-05-24 10:09:53 +00:00
case ENXIO:
/*
* Steal away the file pointer from dfd and stuff it into indx.
1994-05-24 10:09:53 +00:00
*/
fdp->fd_ofiles[indx] = fp;
1994-05-24 10:09:53 +00:00
fdp->fd_ofiles[dfd] = NULL;
fdp->fd_ofileflags[indx] = fdp->fd_ofileflags[dfd];
fdp->fd_ofileflags[dfd] = 0;
fdunused(fdp, dfd);
break;
1994-05-24 10:09:53 +00:00
}
FILEDESC_XUNLOCK(fdp);
*indxp = indx;
return (0);
1994-05-24 10:09:53 +00:00
}
/*
* Scan all active processes and prisons to see if any of them have a current
* or root directory of `olddp'. If so, replace them with the new mount point.
*/
void
mountcheckdirs(struct vnode *olddp, struct vnode *newdp)
{
struct filedesc *fdp;
struct prison *pr;
struct proc *p;
int nrele;
if (vrefcnt(olddp) == 1)
return;
nrele = 0;
sx_slock(&allproc_lock);
FOREACH_PROC_IN_SYSTEM(p) {
fdp = fdhold(p);
if (fdp == NULL)
continue;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
if (fdp->fd_cdir == olddp) {
vref(newdp);
fdp->fd_cdir = newdp;
nrele++;
}
if (fdp->fd_rdir == olddp) {
vref(newdp);
fdp->fd_rdir = newdp;
nrele++;
}
if (fdp->fd_jdir == olddp) {
vref(newdp);
fdp->fd_jdir = newdp;
nrele++;
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
fddrop(fdp);
}
sx_sunlock(&allproc_lock);
if (rootvnode == olddp) {
vref(newdp);
rootvnode = newdp;
nrele++;
}
mtx_lock(&prison0.pr_mtx);
if (prison0.pr_root == olddp) {
vref(newdp);
prison0.pr_root = newdp;
nrele++;
}
mtx_unlock(&prison0.pr_mtx);
sx_slock(&allprison_lock);
TAILQ_FOREACH(pr, &allprison, pr_list) {
mtx_lock(&pr->pr_mtx);
if (pr->pr_root == olddp) {
vref(newdp);
pr->pr_root = newdp;
nrele++;
}
mtx_unlock(&pr->pr_mtx);
}
sx_sunlock(&allprison_lock);
while (nrele--)
vrele(olddp);
}
struct filedesc_to_leader *
filedesc_to_leader_alloc(struct filedesc_to_leader *old, struct filedesc *fdp, struct proc *leader)
{
struct filedesc_to_leader *fdtol;
2004-01-11 19:39:14 +00:00
fdtol = malloc(sizeof(struct filedesc_to_leader),
M_FILEDESC_TO_LEADER,
M_WAITOK);
fdtol->fdl_refcount = 1;
fdtol->fdl_holdcount = 0;
fdtol->fdl_wakeup = 0;
fdtol->fdl_leader = leader;
if (old != NULL) {
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XLOCK(fdp);
fdtol->fdl_next = old->fdl_next;
fdtol->fdl_prev = old;
old->fdl_next = fdtol;
fdtol->fdl_next->fdl_prev = fdtol;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_XUNLOCK(fdp);
} else {
fdtol->fdl_next = fdtol;
fdtol->fdl_prev = fdtol;
}
return (fdtol);
}
/*
* Get file structures globally.
*/
static int
sysctl_kern_file(SYSCTL_HANDLER_ARGS)
{
struct xfile xf;
struct filedesc *fdp;
struct file *fp;
struct proc *p;
int error, n;
error = sysctl_wire_old_buffer(req, 0);
if (error != 0)
return (error);
if (req->oldptr == NULL) {
n = 0;
sx_slock(&allproc_lock);
FOREACH_PROC_IN_SYSTEM(p) {
if (p->p_state == PRS_NEW)
continue;
fdp = fdhold(p);
if (fdp == NULL)
continue;
/* overestimates sparse tables. */
if (fdp->fd_lastfile > 0)
n += fdp->fd_lastfile;
fddrop(fdp);
}
sx_sunlock(&allproc_lock);
return (SYSCTL_OUT(req, 0, n * sizeof(xf)));
}
error = 0;
bzero(&xf, sizeof(xf));
xf.xf_size = sizeof(xf);
sx_slock(&allproc_lock);
FOREACH_PROC_IN_SYSTEM(p) {
PROC_LOCK(p);
if (p->p_state == PRS_NEW) {
PROC_UNLOCK(p);
continue;
}
if (p_cansee(req->td, p) != 0) {
PROC_UNLOCK(p);
continue;
}
xf.xf_pid = p->p_pid;
xf.xf_uid = p->p_ucred->cr_uid;
PROC_UNLOCK(p);
fdp = fdhold(p);
if (fdp == NULL)
continue;
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_SLOCK(fdp);
for (n = 0; fdp->fd_refcnt > 0 && n < fdp->fd_nfiles; ++n) {
if ((fp = fdp->fd_ofiles[n]) == NULL)
continue;
xf.xf_fd = n;
xf.xf_file = fp;
xf.xf_data = fp->f_data;
xf.xf_vnode = fp->f_vnode;
xf.xf_type = fp->f_type;
xf.xf_count = fp->f_count;
xf.xf_msgcount = 0;
xf.xf_offset = foffset_get(fp);
xf.xf_flag = fp->f_flag;
error = SYSCTL_OUT(req, &xf, sizeof(xf));
if (error)
break;
}
Replace custom file descriptor array sleep lock constructed using a mutex and flags with an sxlock. This leads to a significant and measurable performance improvement as a result of access to shared locking for frequent lookup operations, reduced general overhead, and reduced overhead in the event of contention. All of these are imported for threaded applications where simultaneous access to a shared file descriptor array occurs frequently. Kris has reported 2x-4x transaction rate improvements on 8-core MySQL benchmarks; smaller improvements can be expected for many workloads as a result of reduced overhead. - Generally eliminate the distinction between "fast" and regular acquisisition of the filedesc lock; the plan is that they will now all be fast. Change all locking instances to either shared or exclusive locks. - Correct a bug (pointed out by kib) in fdfree() where previously msleep() was called without the mutex held; sx_sleep() is now always called with the sxlock held exclusively. - Universally hold the struct file lock over changes to struct file, rather than the filedesc lock or no lock. Always update the f_ops field last. A further memory barrier is required here in the future (discussed with jhb). - Improve locking and reference management in linux_at(), which fails to properly acquire vnode references before using vnode pointers. Annotate improper use of vn_fullpath(), which will be replaced at a future date. In fcntl(), we conservatively acquire an exclusive lock, even though in some cases a shared lock may be sufficient, which should be revisited. The dropping of the filedesc lock in fdgrowtable() is no longer required as the sxlock can be held over the sleep operation; we should consider removing that (pointed out by attilio). Tested by: kris Discussed with: jhb, kris, attilio, jeff
2007-04-04 09:11:34 +00:00
FILEDESC_SUNLOCK(fdp);
fddrop(fdp);
if (error)
break;
}
sx_sunlock(&allproc_lock);
return (error);
}
SYSCTL_PROC(_kern, KERN_FILE, file, CTLTYPE_OPAQUE|CTLFLAG_RD,
0, 0, sysctl_kern_file, "S,xfile", "Entire file table");
#ifdef KINFO_OFILE_SIZE
CTASSERT(sizeof(struct kinfo_ofile) == KINFO_OFILE_SIZE);
#endif
#ifdef COMPAT_FREEBSD7
static int
export_vnode_for_osysctl(struct vnode *vp, int type,
struct kinfo_ofile *kif, struct filedesc *fdp, struct sysctl_req *req)
{
int error;
char *fullpath, *freepath;
int vfslocked;
bzero(kif, sizeof(*kif));
kif->kf_structsize = sizeof(*kif);
vref(vp);
kif->kf_fd = type;
kif->kf_type = KF_TYPE_VNODE;
/* This function only handles directories. */
if (vp->v_type != VDIR) {
vrele(vp);
return (ENOTDIR);
}
kif->kf_vnode_type = KF_VTYPE_VDIR;
/*
* This is not a true file descriptor, so we set a bogus refcount
* and offset to indicate these fields should be ignored.
*/
kif->kf_ref_count = -1;
kif->kf_offset = -1;
freepath = NULL;
fullpath = "-";
FILEDESC_SUNLOCK(fdp);
vn_fullpath(curthread, vp, &fullpath, &freepath);
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
vrele(vp);
VFS_UNLOCK_GIANT(vfslocked);
strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
if (freepath != NULL)
free(freepath, M_TEMP);
error = SYSCTL_OUT(req, kif, sizeof(*kif));
FILEDESC_SLOCK(fdp);
return (error);
}
/*
* Get per-process file descriptors for use by procstat(1), et al.
*/
static int
sysctl_kern_proc_ofiledesc(SYSCTL_HANDLER_ARGS)
{
char *fullpath, *freepath;
struct kinfo_ofile *kif;
struct filedesc *fdp;
int error, i, *name;
struct shmfd *shmfd;
struct socket *so;
struct vnode *vp;
struct file *fp;
struct proc *p;
struct tty *tp;
int vfslocked;
name = (int *)arg1;
if ((p = pfind((pid_t)name[0])) == NULL)
return (ESRCH);
if ((error = p_candebug(curthread, p))) {
PROC_UNLOCK(p);
return (error);
}
fdp = fdhold(p);
PROC_UNLOCK(p);
if (fdp == NULL)
return (ENOENT);
kif = malloc(sizeof(*kif), M_TEMP, M_WAITOK);
FILEDESC_SLOCK(fdp);
if (fdp->fd_cdir != NULL)
export_vnode_for_osysctl(fdp->fd_cdir, KF_FD_TYPE_CWD, kif,
fdp, req);
if (fdp->fd_rdir != NULL)
export_vnode_for_osysctl(fdp->fd_rdir, KF_FD_TYPE_ROOT, kif,
fdp, req);
if (fdp->fd_jdir != NULL)
export_vnode_for_osysctl(fdp->fd_jdir, KF_FD_TYPE_JAIL, kif,
fdp, req);
for (i = 0; i < fdp->fd_nfiles; i++) {
if ((fp = fdp->fd_ofiles[i]) == NULL)
continue;
bzero(kif, sizeof(*kif));
kif->kf_structsize = sizeof(*kif);
vp = NULL;
so = NULL;
tp = NULL;
shmfd = NULL;
kif->kf_fd = i;
#ifdef CAPABILITIES
/*
* When reporting a capability, most fields will be from the
* underlying object, but do mark as a capability. With
* ofiledesc, we don't have a field to export the cap_rights_t,
* but we do with the new filedesc.
*/
if (fp->f_type == DTYPE_CAPABILITY) {
kif->kf_flags |= KF_FLAG_CAPABILITY;
(void)cap_funwrap(fp, 0, &fp);
}
#else
KASSERT(fp->f_type != DTYPE_CAPABILITY,
("sysctl_kern_proc_ofiledesc: saw capability"));
#endif
switch (fp->f_type) {
case DTYPE_VNODE:
kif->kf_type = KF_TYPE_VNODE;
vp = fp->f_vnode;
break;
case DTYPE_SOCKET:
kif->kf_type = KF_TYPE_SOCKET;
so = fp->f_data;
break;
case DTYPE_PIPE:
kif->kf_type = KF_TYPE_PIPE;
break;
case DTYPE_FIFO:
kif->kf_type = KF_TYPE_FIFO;
vp = fp->f_vnode;
break;
case DTYPE_KQUEUE:
kif->kf_type = KF_TYPE_KQUEUE;
break;
case DTYPE_CRYPTO:
kif->kf_type = KF_TYPE_CRYPTO;
break;
case DTYPE_MQUEUE:
kif->kf_type = KF_TYPE_MQUEUE;
break;
case DTYPE_SHM:
kif->kf_type = KF_TYPE_SHM;
shmfd = fp->f_data;
break;
case DTYPE_SEM:
kif->kf_type = KF_TYPE_SEM;
break;
case DTYPE_PTS:
kif->kf_type = KF_TYPE_PTS;
tp = fp->f_data;
break;
#ifdef PROCDESC
case DTYPE_PROCDESC:
kif->kf_type = KF_TYPE_PROCDESC;
break;
#endif
default:
kif->kf_type = KF_TYPE_UNKNOWN;
break;
}
kif->kf_ref_count = fp->f_count;
if (fp->f_flag & FREAD)
kif->kf_flags |= KF_FLAG_READ;
if (fp->f_flag & FWRITE)
kif->kf_flags |= KF_FLAG_WRITE;
if (fp->f_flag & FAPPEND)
kif->kf_flags |= KF_FLAG_APPEND;
if (fp->f_flag & FASYNC)
kif->kf_flags |= KF_FLAG_ASYNC;
if (fp->f_flag & FFSYNC)
kif->kf_flags |= KF_FLAG_FSYNC;
if (fp->f_flag & FNONBLOCK)
kif->kf_flags |= KF_FLAG_NONBLOCK;
if (fp->f_flag & O_DIRECT)
kif->kf_flags |= KF_FLAG_DIRECT;
if (fp->f_flag & FHASLOCK)
kif->kf_flags |= KF_FLAG_HASLOCK;
kif->kf_offset = foffset_get(fp);
if (vp != NULL) {
vref(vp);
switch (vp->v_type) {
case VNON:
kif->kf_vnode_type = KF_VTYPE_VNON;
break;
case VREG:
kif->kf_vnode_type = KF_VTYPE_VREG;
break;
case VDIR:
kif->kf_vnode_type = KF_VTYPE_VDIR;
break;
case VBLK:
kif->kf_vnode_type = KF_VTYPE_VBLK;
break;
case VCHR:
kif->kf_vnode_type = KF_VTYPE_VCHR;
break;
case VLNK:
kif->kf_vnode_type = KF_VTYPE_VLNK;
break;
case VSOCK:
kif->kf_vnode_type = KF_VTYPE_VSOCK;
break;
case VFIFO:
kif->kf_vnode_type = KF_VTYPE_VFIFO;
break;
case VBAD:
kif->kf_vnode_type = KF_VTYPE_VBAD;
break;
default:
kif->kf_vnode_type = KF_VTYPE_UNKNOWN;
break;
}
/*
* It is OK to drop the filedesc lock here as we will
* re-validate and re-evaluate its properties when
* the loop continues.
*/
freepath = NULL;
fullpath = "-";
FILEDESC_SUNLOCK(fdp);
vn_fullpath(curthread, vp, &fullpath, &freepath);
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
vrele(vp);
VFS_UNLOCK_GIANT(vfslocked);
strlcpy(kif->kf_path, fullpath,
sizeof(kif->kf_path));
if (freepath != NULL)
free(freepath, M_TEMP);
FILEDESC_SLOCK(fdp);
}
if (so != NULL) {
struct sockaddr *sa;
if (so->so_proto->pr_usrreqs->pru_sockaddr(so, &sa)
== 0 && sa->sa_len <= sizeof(kif->kf_sa_local)) {
bcopy(sa, &kif->kf_sa_local, sa->sa_len);
free(sa, M_SONAME);
}
if (so->so_proto->pr_usrreqs->pru_peeraddr(so, &sa)
== 0 && sa->sa_len <= sizeof(kif->kf_sa_peer)) {
bcopy(sa, &kif->kf_sa_peer, sa->sa_len);
free(sa, M_SONAME);
}
kif->kf_sock_domain =
so->so_proto->pr_domain->dom_family;
kif->kf_sock_type = so->so_type;
kif->kf_sock_protocol = so->so_proto->pr_protocol;
}
if (tp != NULL) {
strlcpy(kif->kf_path, tty_devname(tp),
sizeof(kif->kf_path));
}
if (shmfd != NULL)
shm_path(shmfd, kif->kf_path, sizeof(kif->kf_path));
error = SYSCTL_OUT(req, kif, sizeof(*kif));
if (error)
break;
}
FILEDESC_SUNLOCK(fdp);
fddrop(fdp);
free(kif, M_TEMP);
return (0);
}
static SYSCTL_NODE(_kern_proc, KERN_PROC_OFILEDESC, ofiledesc, CTLFLAG_RD,
sysctl_kern_proc_ofiledesc, "Process ofiledesc entries");
#endif /* COMPAT_FREEBSD7 */
#ifdef KINFO_FILE_SIZE
CTASSERT(sizeof(struct kinfo_file) == KINFO_FILE_SIZE);
#endif
static int
export_fd_for_sysctl(void *data, int type, int fd, int fflags, int refcnt,
int64_t offset, int fd_is_cap, cap_rights_t fd_cap_rights,
struct kinfo_file *kif, struct sysctl_req *req)
{
struct {
int fflag;
int kf_fflag;
} fflags_table[] = {
{ FAPPEND, KF_FLAG_APPEND },
{ FASYNC, KF_FLAG_ASYNC },
{ FFSYNC, KF_FLAG_FSYNC },
{ FHASLOCK, KF_FLAG_HASLOCK },
{ FNONBLOCK, KF_FLAG_NONBLOCK },
{ FREAD, KF_FLAG_READ },
{ FWRITE, KF_FLAG_WRITE },
{ O_CREAT, KF_FLAG_CREAT },
{ O_DIRECT, KF_FLAG_DIRECT },
{ O_EXCL, KF_FLAG_EXCL },
{ O_EXEC, KF_FLAG_EXEC },
{ O_EXLOCK, KF_FLAG_EXLOCK },
{ O_NOFOLLOW, KF_FLAG_NOFOLLOW },
{ O_SHLOCK, KF_FLAG_SHLOCK },
{ O_TRUNC, KF_FLAG_TRUNC }
};
#define NFFLAGS (sizeof(fflags_table) / sizeof(*fflags_table))
struct vnode *vp;
int error, vfslocked;
unsigned int i;
bzero(kif, sizeof(*kif));
switch (type) {
case KF_TYPE_FIFO:
case KF_TYPE_VNODE:
vp = (struct vnode *)data;
error = fill_vnode_info(vp, kif);
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
vrele(vp);
VFS_UNLOCK_GIANT(vfslocked);
break;
case KF_TYPE_SOCKET:
error = fill_socket_info((struct socket *)data, kif);
break;
case KF_TYPE_PIPE:
error = fill_pipe_info((struct pipe *)data, kif);
break;
case KF_TYPE_PTS:
error = fill_pts_info((struct tty *)data, kif);
break;
case KF_TYPE_PROCDESC:
error = fill_procdesc_info((struct procdesc *)data, kif);
break;
case KF_TYPE_SHM:
error = fill_shm_info((struct file *)data, kif);
break;
default:
error = 0;
}
if (error == 0)
kif->kf_status |= KF_ATTR_VALID;
/*
* Translate file access flags.
*/
for (i = 0; i < NFFLAGS; i++)
if (fflags & fflags_table[i].fflag)
kif->kf_flags |= fflags_table[i].kf_fflag;
if (fd_is_cap)
kif->kf_flags |= KF_FLAG_CAPABILITY;
if (fd_is_cap)
kif->kf_cap_rights = fd_cap_rights;
kif->kf_fd = fd;
kif->kf_type = type;
kif->kf_ref_count = refcnt;
kif->kf_offset = offset;
/* Pack record size down */
kif->kf_structsize = offsetof(struct kinfo_file, kf_path) +
strlen(kif->kf_path) + 1;
kif->kf_structsize = roundup(kif->kf_structsize, sizeof(uint64_t));
error = SYSCTL_OUT(req, kif, kif->kf_structsize);
return (error);
}
/*
* Get per-process file descriptors for use by procstat(1), et al.
*/
static int
sysctl_kern_proc_filedesc(SYSCTL_HANDLER_ARGS)
{
struct file *fp;
struct filedesc *fdp;
struct kinfo_file *kif;
struct proc *p;
struct vnode *cttyvp, *textvp, *tracevp;
size_t oldidx;
int64_t offset;
void *data;
int error, i, *name;
int fd_is_cap, type, refcnt, fflags;
cap_rights_t fd_cap_rights;
name = (int *)arg1;
if ((p = pfind((pid_t)name[0])) == NULL)
return (ESRCH);
if ((error = p_candebug(curthread, p))) {
PROC_UNLOCK(p);
return (error);
}
/* ktrace vnode */
tracevp = p->p_tracevp;
if (tracevp != NULL)
vref(tracevp);
/* text vnode */
textvp = p->p_textvp;
if (textvp != NULL)
vref(textvp);
/* Controlling tty. */
cttyvp = NULL;
if (p->p_pgrp != NULL && p->p_pgrp->pg_session != NULL) {
cttyvp = p->p_pgrp->pg_session->s_ttyvp;
if (cttyvp != NULL)
vref(cttyvp);
}
fdp = fdhold(p);
PROC_UNLOCK(p);
kif = malloc(sizeof(*kif), M_TEMP, M_WAITOK);
if (tracevp != NULL)
export_fd_for_sysctl(tracevp, KF_TYPE_VNODE, KF_FD_TYPE_TRACE,
FREAD | FWRITE, -1, -1, 0, 0, kif, req);
if (textvp != NULL)
export_fd_for_sysctl(textvp, KF_TYPE_VNODE, KF_FD_TYPE_TEXT,
FREAD, -1, -1, 0, 0, kif, req);
if (cttyvp != NULL)
export_fd_for_sysctl(cttyvp, KF_TYPE_VNODE, KF_FD_TYPE_CTTY,
FREAD | FWRITE, -1, -1, 0, 0, kif, req);
if (fdp == NULL)
goto fail;
FILEDESC_SLOCK(fdp);
/* working directory */
if (fdp->fd_cdir != NULL) {
vref(fdp->fd_cdir);
data = fdp->fd_cdir;
FILEDESC_SUNLOCK(fdp);
export_fd_for_sysctl(data, KF_TYPE_VNODE, KF_FD_TYPE_CWD,
FREAD, -1, -1, 0, 0, kif, req);
FILEDESC_SLOCK(fdp);
}
/* root directory */
if (fdp->fd_rdir != NULL) {
vref(fdp->fd_rdir);
data = fdp->fd_rdir;
FILEDESC_SUNLOCK(fdp);
export_fd_for_sysctl(data, KF_TYPE_VNODE, KF_FD_TYPE_ROOT,
FREAD, -1, -1, 0, 0, kif, req);
FILEDESC_SLOCK(fdp);
}
/* jail directory */
if (fdp->fd_jdir != NULL) {
vref(fdp->fd_jdir);
data = fdp->fd_jdir;
FILEDESC_SUNLOCK(fdp);
export_fd_for_sysctl(data, KF_TYPE_VNODE, KF_FD_TYPE_JAIL,
FREAD, -1, -1, 0, 0, kif, req);
FILEDESC_SLOCK(fdp);
}
for (i = 0; i < fdp->fd_nfiles; i++) {
if ((fp = fdp->fd_ofiles[i]) == NULL)
continue;
data = NULL;
fd_is_cap = 0;
fd_cap_rights = 0;
#ifdef CAPABILITIES
/*
* When reporting a capability, most fields will be from the
* underlying object, but do mark as a capability and export
* the capability rights mask.
*/
if (fp->f_type == DTYPE_CAPABILITY) {
fd_is_cap = 1;
fd_cap_rights = cap_rights(fp);
(void)cap_funwrap(fp, 0, &fp);
}
#else /* !CAPABILITIES */
KASSERT(fp->f_type != DTYPE_CAPABILITY,
("sysctl_kern_proc_filedesc: saw capability"));
#endif
switch (fp->f_type) {
case DTYPE_VNODE:
type = KF_TYPE_VNODE;
vref(fp->f_vnode);
data = fp->f_vnode;
break;
case DTYPE_SOCKET:
type = KF_TYPE_SOCKET;
data = fp->f_data;
break;
case DTYPE_PIPE:
type = KF_TYPE_PIPE;
data = fp->f_data;
break;
case DTYPE_FIFO:
type = KF_TYPE_FIFO;
vref(fp->f_vnode);
data = fp->f_vnode;
break;
case DTYPE_KQUEUE:
type = KF_TYPE_KQUEUE;
break;
case DTYPE_CRYPTO:
type = KF_TYPE_CRYPTO;
break;
case DTYPE_MQUEUE:
type = KF_TYPE_MQUEUE;
break;
case DTYPE_SHM:
type = KF_TYPE_SHM;
data = fp;
break;
Rework the lifetime management of the kernel implementation of POSIX semaphores. Specifically, semaphores are now represented as new file descriptor type that is set to close on exec. This removes the need for all of the manual process reference counting (and fork, exec, and exit event handlers) as the normal file descriptor operations handle all of that for us nicely. It is also suggested as one possible implementation in the spec and at least one other OS (OS X) uses this approach. Some bugs that were fixed as a result include: - References to a named semaphore whose name is removed still work after the sem_unlink() operation. Prior to this patch, if a semaphore's name was removed, valid handles from sem_open() would get EINVAL errors from sem_getvalue(), sem_post(), etc. This fixes that. - Unnamed semaphores created with sem_init() were not cleaned up when a process exited or exec'd. They were only cleaned up if the process did an explicit sem_destroy(). This could result in a leak of semaphore objects that could never be cleaned up. - On the other hand, if another process guessed the id (kernel pointer to 'struct ksem' of an unnamed semaphore (created via sem_init)) and had write access to the semaphore based on UID/GID checks, then that other process could manipulate the semaphore via sem_destroy(), sem_post(), sem_wait(), etc. - As part of the permission check (UID/GID), the umask of the proces creating the semaphore was not honored. Thus if your umask denied group read/write access but the explicit mode in the sem_init() call allowed it, the semaphore would be readable/writable by other users in the same group, for example. This includes access via the previous bug. - If the module refused to unload because there were active semaphores, then it might have deregistered one or more of the semaphore system calls before it noticed that there was a problem. I'm not sure if this actually happened as the order that modules are discovered by the kernel linker depends on how the actual .ko file is linked. One can make the order deterministic by using a single module with a mod_event handler that explicitly registers syscalls (and deregisters during unload after any checks). This also fixes a race where even if the sem_module unloaded first it would have destroyed locks that the syscalls might be trying to access if they are still executing when they are unloaded. XXX: By the way, deregistering system calls doesn't do any blocking to drain any threads from the calls. - Some minor fixes to errno values on error. For example, sem_init() isn't documented to return ENFILE or EMFILE if we run out of semaphores the way that sem_open() can. Instead, it should return ENOSPC in that case. Other changes: - Kernel semaphores now use a hash table to manage the namespace of named semaphores nearly in a similar fashion to the POSIX shared memory object file descriptors. Kernel semaphores can now also have names longer than 14 chars (up to MAXPATHLEN) and can include subdirectories in their pathname. - The UID/GID permission checks for access to a named semaphore are now done via vaccess() rather than a home-rolled set of checks. - Now that kernel semaphores have an associated file object, the various MAC checks for POSIX semaphores accept both a file credential and an active credential. There is also a new posixsem_check_stat() since it is possible to fstat() a semaphore file descriptor. - A small set of regression tests (using the ksem API directly) is present in src/tools/regression/posixsem. Reported by: kris (1) Tested by: kris Reviewed by: rwatson (lightly) MFC after: 1 month
2008-06-27 05:39:04 +00:00
case DTYPE_SEM:
type = KF_TYPE_SEM;
Rework the lifetime management of the kernel implementation of POSIX semaphores. Specifically, semaphores are now represented as new file descriptor type that is set to close on exec. This removes the need for all of the manual process reference counting (and fork, exec, and exit event handlers) as the normal file descriptor operations handle all of that for us nicely. It is also suggested as one possible implementation in the spec and at least one other OS (OS X) uses this approach. Some bugs that were fixed as a result include: - References to a named semaphore whose name is removed still work after the sem_unlink() operation. Prior to this patch, if a semaphore's name was removed, valid handles from sem_open() would get EINVAL errors from sem_getvalue(), sem_post(), etc. This fixes that. - Unnamed semaphores created with sem_init() were not cleaned up when a process exited or exec'd. They were only cleaned up if the process did an explicit sem_destroy(). This could result in a leak of semaphore objects that could never be cleaned up. - On the other hand, if another process guessed the id (kernel pointer to 'struct ksem' of an unnamed semaphore (created via sem_init)) and had write access to the semaphore based on UID/GID checks, then that other process could manipulate the semaphore via sem_destroy(), sem_post(), sem_wait(), etc. - As part of the permission check (UID/GID), the umask of the proces creating the semaphore was not honored. Thus if your umask denied group read/write access but the explicit mode in the sem_init() call allowed it, the semaphore would be readable/writable by other users in the same group, for example. This includes access via the previous bug. - If the module refused to unload because there were active semaphores, then it might have deregistered one or more of the semaphore system calls before it noticed that there was a problem. I'm not sure if this actually happened as the order that modules are discovered by the kernel linker depends on how the actual .ko file is linked. One can make the order deterministic by using a single module with a mod_event handler that explicitly registers syscalls (and deregisters during unload after any checks). This also fixes a race where even if the sem_module unloaded first it would have destroyed locks that the syscalls might be trying to access if they are still executing when they are unloaded. XXX: By the way, deregistering system calls doesn't do any blocking to drain any threads from the calls. - Some minor fixes to errno values on error. For example, sem_init() isn't documented to return ENFILE or EMFILE if we run out of semaphores the way that sem_open() can. Instead, it should return ENOSPC in that case. Other changes: - Kernel semaphores now use a hash table to manage the namespace of named semaphores nearly in a similar fashion to the POSIX shared memory object file descriptors. Kernel semaphores can now also have names longer than 14 chars (up to MAXPATHLEN) and can include subdirectories in their pathname. - The UID/GID permission checks for access to a named semaphore are now done via vaccess() rather than a home-rolled set of checks. - Now that kernel semaphores have an associated file object, the various MAC checks for POSIX semaphores accept both a file credential and an active credential. There is also a new posixsem_check_stat() since it is possible to fstat() a semaphore file descriptor. - A small set of regression tests (using the ksem API directly) is present in src/tools/regression/posixsem. Reported by: kris (1) Tested by: kris Reviewed by: rwatson (lightly) MFC after: 1 month
2008-06-27 05:39:04 +00:00
break;
Integrate the new MPSAFE TTY layer to the FreeBSD operating system. The last half year I've been working on a replacement TTY layer for the FreeBSD kernel. The new TTY layer was designed to improve the following: - Improved driver model: The old TTY layer has a driver model that is not abstract enough to make it friendly to use. A good example is the output path, where the device drivers directly access the output buffers. This means that an in-kernel PPP implementation must always convert network buffers into TTY buffers. If a PPP implementation would be built on top of the new TTY layer (still needs a hooks layer, though), it would allow the PPP implementation to directly hand the data to the TTY driver. - Improved hotplugging: With the old TTY layer, it isn't entirely safe to destroy TTY's from the system. This implementation has a two-step destructing design, where the driver first abandons the TTY. After all threads have left the TTY, the TTY layer calls a routine in the driver, which can be used to free resources (unit numbers, etc). The pts(4) driver also implements this feature, which means posix_openpt() will now return PTY's that are created on the fly. - Improved performance: One of the major improvements is the per-TTY mutex, which is expected to improve scalability when compared to the old Giant locking. Another change is the unbuffered copying to userspace, which is both used on TTY device nodes and PTY masters. Upgrading should be quite straightforward. Unlike previous versions, existing kernel configuration files do not need to be changed, except when they reference device drivers that are listed in UPDATING. Obtained from: //depot/projects/mpsafetty/... Approved by: philip (ex-mentor) Discussed: on the lists, at BSDCan, at the DevSummit Sponsored by: Snow B.V., the Netherlands dcons(4) fixed by: kan
2008-08-20 08:31:58 +00:00
case DTYPE_PTS:
type = KF_TYPE_PTS;
data = fp->f_data;
Integrate the new MPSAFE TTY layer to the FreeBSD operating system. The last half year I've been working on a replacement TTY layer for the FreeBSD kernel. The new TTY layer was designed to improve the following: - Improved driver model: The old TTY layer has a driver model that is not abstract enough to make it friendly to use. A good example is the output path, where the device drivers directly access the output buffers. This means that an in-kernel PPP implementation must always convert network buffers into TTY buffers. If a PPP implementation would be built on top of the new TTY layer (still needs a hooks layer, though), it would allow the PPP implementation to directly hand the data to the TTY driver. - Improved hotplugging: With the old TTY layer, it isn't entirely safe to destroy TTY's from the system. This implementation has a two-step destructing design, where the driver first abandons the TTY. After all threads have left the TTY, the TTY layer calls a routine in the driver, which can be used to free resources (unit numbers, etc). The pts(4) driver also implements this feature, which means posix_openpt() will now return PTY's that are created on the fly. - Improved performance: One of the major improvements is the per-TTY mutex, which is expected to improve scalability when compared to the old Giant locking. Another change is the unbuffered copying to userspace, which is both used on TTY device nodes and PTY masters. Upgrading should be quite straightforward. Unlike previous versions, existing kernel configuration files do not need to be changed, except when they reference device drivers that are listed in UPDATING. Obtained from: //depot/projects/mpsafetty/... Approved by: philip (ex-mentor) Discussed: on the lists, at BSDCan, at the DevSummit Sponsored by: Snow B.V., the Netherlands dcons(4) fixed by: kan
2008-08-20 08:31:58 +00:00
break;
#ifdef PROCDESC
case DTYPE_PROCDESC:
type = KF_TYPE_PROCDESC;
data = fp->f_data;
break;
#endif
default:
type = KF_TYPE_UNKNOWN;
break;
}
refcnt = fp->f_count;
fflags = fp->f_flag;
offset = foffset_get(fp);
/*
* Create sysctl entry.
* It is OK to drop the filedesc lock here as we will
* re-validate and re-evaluate its properties when
* the loop continues.
*/
oldidx = req->oldidx;
if (type == KF_TYPE_VNODE || type == KF_TYPE_FIFO)
FILEDESC_SUNLOCK(fdp);
error = export_fd_for_sysctl(data, type, i, fflags, refcnt,
offset, fd_is_cap, fd_cap_rights, kif, req);
if (type == KF_TYPE_VNODE || type == KF_TYPE_FIFO)
FILEDESC_SLOCK(fdp);
if (error) {
if (error == ENOMEM) {
/*
* The hack to keep the ABI of sysctl
* kern.proc.filedesc intact, but not
* to account a partially copied
* kinfo_file into the oldidx.
*/
req->oldidx = oldidx;
error = 0;
}
break;
}
}
FILEDESC_SUNLOCK(fdp);
fail:
if (fdp != NULL)
fddrop(fdp);
free(kif, M_TEMP);
return (error);
}
int
vntype_to_kinfo(int vtype)
{
struct {
int vtype;
int kf_vtype;
} vtypes_table[] = {
{ VBAD, KF_VTYPE_VBAD },
{ VBLK, KF_VTYPE_VBLK },
{ VCHR, KF_VTYPE_VCHR },
{ VDIR, KF_VTYPE_VDIR },
{ VFIFO, KF_VTYPE_VFIFO },
{ VLNK, KF_VTYPE_VLNK },
{ VNON, KF_VTYPE_VNON },
{ VREG, KF_VTYPE_VREG },
{ VSOCK, KF_VTYPE_VSOCK }
};
#define NVTYPES (sizeof(vtypes_table) / sizeof(*vtypes_table))
unsigned int i;
/*
* Perform vtype translation.
*/
for (i = 0; i < NVTYPES; i++)
if (vtypes_table[i].vtype == vtype)
break;
if (i < NVTYPES)
return (vtypes_table[i].kf_vtype);
return (KF_VTYPE_UNKNOWN);
}
static int
fill_vnode_info(struct vnode *vp, struct kinfo_file *kif)
{
struct vattr va;
char *fullpath, *freepath;
int error, vfslocked;
if (vp == NULL)
return (1);
kif->kf_vnode_type = vntype_to_kinfo(vp->v_type);
freepath = NULL;
fullpath = "-";
error = vn_fullpath(curthread, vp, &fullpath, &freepath);
if (error == 0) {
strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
}
if (freepath != NULL)
free(freepath, M_TEMP);
/*
* Retrieve vnode attributes.
*/
va.va_fsid = VNOVAL;
va.va_rdev = NODEV;
vfslocked = VFS_LOCK_GIANT(vp->v_mount);
vn_lock(vp, LK_SHARED | LK_RETRY);
error = VOP_GETATTR(vp, &va, curthread->td_ucred);
VOP_UNLOCK(vp, 0);
VFS_UNLOCK_GIANT(vfslocked);
if (error != 0)
return (error);
if (va.va_fsid != VNOVAL)
kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
else
kif->kf_un.kf_file.kf_file_fsid =
vp->v_mount->mnt_stat.f_fsid.val[0];
kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
kif->kf_un.kf_file.kf_file_size = va.va_size;
kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
return (0);
}
static int
fill_socket_info(struct socket *so, struct kinfo_file *kif)
{
struct sockaddr *sa;
struct inpcb *inpcb;
struct unpcb *unpcb;
int error;
if (so == NULL)
return (1);
kif->kf_sock_domain = so->so_proto->pr_domain->dom_family;
kif->kf_sock_type = so->so_type;
kif->kf_sock_protocol = so->so_proto->pr_protocol;
kif->kf_un.kf_sock.kf_sock_pcb = (uintptr_t)so->so_pcb;
switch(kif->kf_sock_domain) {
case AF_INET:
case AF_INET6:
if (kif->kf_sock_protocol == IPPROTO_TCP) {
if (so->so_pcb != NULL) {
inpcb = (struct inpcb *)(so->so_pcb);
kif->kf_un.kf_sock.kf_sock_inpcb =
(uintptr_t)inpcb->inp_ppcb;
}
}
break;
case AF_UNIX:
if (so->so_pcb != NULL) {
unpcb = (struct unpcb *)(so->so_pcb);
if (unpcb->unp_conn) {
kif->kf_un.kf_sock.kf_sock_unpconn =
(uintptr_t)unpcb->unp_conn;
kif->kf_un.kf_sock.kf_sock_rcv_sb_state =
so->so_rcv.sb_state;
kif->kf_un.kf_sock.kf_sock_snd_sb_state =
so->so_snd.sb_state;
}
}
break;
}
error = so->so_proto->pr_usrreqs->pru_sockaddr(so, &sa);
if (error == 0 && sa->sa_len <= sizeof(kif->kf_sa_local)) {
bcopy(sa, &kif->kf_sa_local, sa->sa_len);
free(sa, M_SONAME);
}
error = so->so_proto->pr_usrreqs->pru_peeraddr(so, &sa);
if (error == 0 && sa->sa_len <= sizeof(kif->kf_sa_peer)) {
bcopy(sa, &kif->kf_sa_peer, sa->sa_len);
free(sa, M_SONAME);
}
strncpy(kif->kf_path, so->so_proto->pr_domain->dom_name,
sizeof(kif->kf_path));
return (0);
}
static int
fill_pts_info(struct tty *tp, struct kinfo_file *kif)
{
if (tp == NULL)
return (1);
kif->kf_un.kf_pts.kf_pts_dev = tty_udev(tp);
strlcpy(kif->kf_path, tty_devname(tp), sizeof(kif->kf_path));
return (0);
}
static int
fill_pipe_info(struct pipe *pi, struct kinfo_file *kif)
{
if (pi == NULL)
return (1);
kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
return (0);
}
static int
fill_procdesc_info(struct procdesc *pdp, struct kinfo_file *kif)
{
if (pdp == NULL)
return (1);
kif->kf_un.kf_proc.kf_pid = pdp->pd_pid;
return (0);
}
static int
fill_shm_info(struct file *fp, struct kinfo_file *kif)
{
struct thread *td;
struct stat sb;
td = curthread;
if (fp->f_data == NULL)
return (1);
if (fo_stat(fp, &sb, td->td_ucred, td) != 0)
return (1);
shm_path(fp->f_data, kif->kf_path, sizeof(kif->kf_path));
kif->kf_un.kf_file.kf_file_mode = sb.st_mode;
kif->kf_un.kf_file.kf_file_size = sb.st_size;
return (0);
}
static SYSCTL_NODE(_kern_proc, KERN_PROC_FILEDESC, filedesc, CTLFLAG_RD,
sysctl_kern_proc_filedesc, "Process filedesc entries");
#ifdef DDB
/*
* For the purposes of debugging, generate a human-readable string for the
* file type.
*/
static const char *
file_type_to_name(short type)
{
switch (type) {
case 0:
return ("zero");
case DTYPE_VNODE:
return ("vnod");
case DTYPE_SOCKET:
return ("sock");
case DTYPE_PIPE:
return ("pipe");
case DTYPE_FIFO:
return ("fifo");
case DTYPE_KQUEUE:
return ("kque");
case DTYPE_CRYPTO:
return ("crpt");
case DTYPE_MQUEUE:
return ("mque");
Add a new file descriptor type for IPC shared memory objects and use it to implement shm_open(2) and shm_unlink(2) in the kernel: - Each shared memory file descriptor is associated with a swap-backed vm object which provides the backing store. Each descriptor starts off with a size of zero, but the size can be altered via ftruncate(2). The shared memory file descriptors also support fstat(2). read(2), write(2), ioctl(2), select(2), poll(2), and kevent(2) are not supported on shared memory file descriptors. - shm_open(2) and shm_unlink(2) are now implemented as system calls that manage shared memory file descriptors. The virtual namespace that maps pathnames to shared memory file descriptors is implemented as a hash table where the hash key is generated via the 32-bit Fowler/Noll/Vo hash of the pathname. - As an extension, the constant 'SHM_ANON' may be specified in place of the path argument to shm_open(2). In this case, an unnamed shared memory file descriptor will be created similar to the IPC_PRIVATE key for shmget(2). Note that the shared memory object can still be shared among processes by sharing the file descriptor via fork(2) or sendmsg(2), but it is unnamed. This effectively serves to implement the getmemfd() idea bandied about the lists several times over the years. - The backing store for shared memory file descriptors are garbage collected when they are not referenced by any open file descriptors or the shm_open(2) virtual namespace. Submitted by: dillon, peter (previous versions) Submitted by: rwatson (I based this on his version) Reviewed by: alc (suggested converting getmemfd() to shm_open())
2008-01-08 21:58:16 +00:00
case DTYPE_SHM:
return ("shm");
Rework the lifetime management of the kernel implementation of POSIX semaphores. Specifically, semaphores are now represented as new file descriptor type that is set to close on exec. This removes the need for all of the manual process reference counting (and fork, exec, and exit event handlers) as the normal file descriptor operations handle all of that for us nicely. It is also suggested as one possible implementation in the spec and at least one other OS (OS X) uses this approach. Some bugs that were fixed as a result include: - References to a named semaphore whose name is removed still work after the sem_unlink() operation. Prior to this patch, if a semaphore's name was removed, valid handles from sem_open() would get EINVAL errors from sem_getvalue(), sem_post(), etc. This fixes that. - Unnamed semaphores created with sem_init() were not cleaned up when a process exited or exec'd. They were only cleaned up if the process did an explicit sem_destroy(). This could result in a leak of semaphore objects that could never be cleaned up. - On the other hand, if another process guessed the id (kernel pointer to 'struct ksem' of an unnamed semaphore (created via sem_init)) and had write access to the semaphore based on UID/GID checks, then that other process could manipulate the semaphore via sem_destroy(), sem_post(), sem_wait(), etc. - As part of the permission check (UID/GID), the umask of the proces creating the semaphore was not honored. Thus if your umask denied group read/write access but the explicit mode in the sem_init() call allowed it, the semaphore would be readable/writable by other users in the same group, for example. This includes access via the previous bug. - If the module refused to unload because there were active semaphores, then it might have deregistered one or more of the semaphore system calls before it noticed that there was a problem. I'm not sure if this actually happened as the order that modules are discovered by the kernel linker depends on how the actual .ko file is linked. One can make the order deterministic by using a single module with a mod_event handler that explicitly registers syscalls (and deregisters during unload after any checks). This also fixes a race where even if the sem_module unloaded first it would have destroyed locks that the syscalls might be trying to access if they are still executing when they are unloaded. XXX: By the way, deregistering system calls doesn't do any blocking to drain any threads from the calls. - Some minor fixes to errno values on error. For example, sem_init() isn't documented to return ENFILE or EMFILE if we run out of semaphores the way that sem_open() can. Instead, it should return ENOSPC in that case. Other changes: - Kernel semaphores now use a hash table to manage the namespace of named semaphores nearly in a similar fashion to the POSIX shared memory object file descriptors. Kernel semaphores can now also have names longer than 14 chars (up to MAXPATHLEN) and can include subdirectories in their pathname. - The UID/GID permission checks for access to a named semaphore are now done via vaccess() rather than a home-rolled set of checks. - Now that kernel semaphores have an associated file object, the various MAC checks for POSIX semaphores accept both a file credential and an active credential. There is also a new posixsem_check_stat() since it is possible to fstat() a semaphore file descriptor. - A small set of regression tests (using the ksem API directly) is present in src/tools/regression/posixsem. Reported by: kris (1) Tested by: kris Reviewed by: rwatson (lightly) MFC after: 1 month
2008-06-27 05:39:04 +00:00
case DTYPE_SEM:
return ("ksem");
default:
return ("unkn");
}
}
/*
* For the purposes of debugging, identify a process (if any, perhaps one of
* many) that references the passed file in its file descriptor array. Return
* NULL if none.
*/
static struct proc *
file_to_first_proc(struct file *fp)
{
struct filedesc *fdp;
struct proc *p;
int n;
FOREACH_PROC_IN_SYSTEM(p) {
if (p->p_state == PRS_NEW)
continue;
fdp = p->p_fd;
if (fdp == NULL)
continue;
for (n = 0; n < fdp->fd_nfiles; n++) {
if (fp == fdp->fd_ofiles[n])
return (p);
}
}
return (NULL);
}
static void
db_print_file(struct file *fp, int header)
{
struct proc *p;
if (header)
db_printf("%8s %4s %8s %8s %4s %5s %6s %8s %5s %12s\n",
"File", "Type", "Data", "Flag", "GCFl", "Count",
"MCount", "Vnode", "FPID", "FCmd");
p = file_to_first_proc(fp);
db_printf("%8p %4s %8p %08x %04x %5d %6d %8p %5d %12s\n", fp,
file_type_to_name(fp->f_type), fp->f_data, fp->f_flag,
0, fp->f_count, 0, fp->f_vnode,
p != NULL ? p->p_pid : -1, p != NULL ? p->p_comm : "-");
}
DB_SHOW_COMMAND(file, db_show_file)
{
struct file *fp;
if (!have_addr) {
db_printf("usage: show file <addr>\n");
return;
}
fp = (struct file *)addr;
db_print_file(fp, 1);
}
DB_SHOW_COMMAND(files, db_show_files)
{
struct filedesc *fdp;
struct file *fp;
struct proc *p;
int header;
int n;
header = 1;
FOREACH_PROC_IN_SYSTEM(p) {
if (p->p_state == PRS_NEW)
continue;
if ((fdp = p->p_fd) == NULL)
continue;
for (n = 0; n < fdp->fd_nfiles; ++n) {
if ((fp = fdp->fd_ofiles[n]) == NULL)
continue;
db_print_file(fp, header);
header = 0;
}
}
}
#endif
2004-01-11 19:39:14 +00:00
SYSCTL_INT(_kern, KERN_MAXFILESPERPROC, maxfilesperproc, CTLFLAG_RW,
&maxfilesperproc, 0, "Maximum files allowed open per process");
SYSCTL_INT(_kern, KERN_MAXFILES, maxfiles, CTLFLAG_RW,
&maxfiles, 0, "Maximum number of files");
2004-01-11 19:39:14 +00:00
SYSCTL_INT(_kern, OID_AUTO, openfiles, CTLFLAG_RD,
__DEVOLATILE(int *, &openfiles), 0, "System-wide number of open files");
/* ARGSUSED*/
static void
filelistinit(void *dummy)
{
file_zone = uma_zcreate("Files", sizeof(struct file), NULL, NULL,
NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
mtx_init(&sigio_lock, "sigio lock", NULL, MTX_DEF);
mtx_init(&fdesc_mtx, "fdesc", NULL, MTX_DEF);
}
SYSINIT(select, SI_SUB_LOCK, SI_ORDER_FIRST, filelistinit, NULL);
/*-------------------------------------------------------------------*/
static int
badfo_readwrite(struct file *fp, struct uio *uio, struct ucred *active_cred,
int flags, struct thread *td)
{
return (EBADF);
}
static int
badfo_truncate(struct file *fp, off_t length, struct ucred *active_cred,
struct thread *td)
{
return (EINVAL);
}
static int
badfo_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
struct thread *td)
{
return (EBADF);
}
static int
badfo_poll(struct file *fp, int events, struct ucred *active_cred,
struct thread *td)
{
return (0);
}
static int
badfo_kqfilter(struct file *fp, struct knote *kn)
{
return (EBADF);
}
static int
badfo_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
struct thread *td)
{
return (EBADF);
}
static int
badfo_close(struct file *fp, struct thread *td)
{
return (EBADF);
}
static int
badfo_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
struct thread *td)
{
return (EBADF);
}
static int
badfo_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
struct thread *td)
{
return (EBADF);
}
struct fileops badfileops = {
.fo_read = badfo_readwrite,
.fo_write = badfo_readwrite,
.fo_truncate = badfo_truncate,
.fo_ioctl = badfo_ioctl,
.fo_poll = badfo_poll,
.fo_kqfilter = badfo_kqfilter,
.fo_stat = badfo_stat,
.fo_close = badfo_close,
.fo_chmod = badfo_chmod,
.fo_chown = badfo_chown,
};
int
invfo_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
struct thread *td)
{
return (EINVAL);
}
int
invfo_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
struct thread *td)
{
return (EINVAL);
}
/*-------------------------------------------------------------------*/
/*
* File Descriptor pseudo-device driver (/dev/fd/).
*
* Opening minor device N dup()s the file (if any) connected to file
* descriptor N belonging to the calling process. Note that this driver
* consists of only the ``open()'' routine, because all subsequent
* references to this file will be direct to the other driver.
*
* XXX: we could give this one a cloning event handler if necessary.
*/
/* ARGSUSED */
static int
fdopen(struct cdev *dev, int mode, int type, struct thread *td)
{
/*
* XXX Kludge: set curthread->td_dupfd to contain the value of the
* the file descriptor being sought for duplication. The error
* return ensures that the vnode for this device will be released
* by vn_open. Open will detect this special error and take the
* actions in dupfdopen below. Other callers of vn_open or VOP_OPEN
* will simply report the error.
*/
td->td_dupfd = dev2unit(dev);
return (ENODEV);
}
static struct cdevsw fildesc_cdevsw = {
.d_version = D_VERSION,
.d_open = fdopen,
.d_name = "FD",
};
static void
fildesc_drvinit(void *unused)
{
struct cdev *dev;
dev = make_dev_credf(MAKEDEV_ETERNAL, &fildesc_cdevsw, 0, NULL,
UID_ROOT, GID_WHEEL, 0666, "fd/0");
make_dev_alias(dev, "stdin");
dev = make_dev_credf(MAKEDEV_ETERNAL, &fildesc_cdevsw, 1, NULL,
UID_ROOT, GID_WHEEL, 0666, "fd/1");
make_dev_alias(dev, "stdout");
dev = make_dev_credf(MAKEDEV_ETERNAL, &fildesc_cdevsw, 2, NULL,
UID_ROOT, GID_WHEEL, 0666, "fd/2");
make_dev_alias(dev, "stderr");
}
SYSINIT(fildescdev, SI_SUB_DRIVERS, SI_ORDER_MIDDLE, fildesc_drvinit, NULL);