2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95
35 * $FreeBSD: src/sys/kern/vfs_subr.c,v 1.249.2.30 2003/04/04 20:35:57 tegge Exp $
39 * External virtual filesystem routines
42 #include <sys/param.h>
43 #include <sys/systm.h>
46 #include <sys/dirent.h>
47 #include <sys/domain.h>
48 #include <sys/eventhandler.h>
49 #include <sys/fcntl.h>
50 #include <sys/kernel.h>
51 #include <sys/kthread.h>
52 #include <sys/malloc.h>
54 #include <sys/mount.h>
56 #include <sys/namei.h>
57 #include <sys/reboot.h>
58 #include <sys/socket.h>
60 #include <sys/sysctl.h>
61 #include <sys/syslog.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
65 #include <machine/limits.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_kern.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_page.h>
74 #include <vm/vm_pager.h>
75 #include <vm/vnode_pager.h>
78 #include <sys/thread2.h>
83 #define SYNCER_MAXDELAY 32
84 static int sysctl_kern_syncdelay(SYSCTL_HANDLER_ARGS);
85 time_t syncdelay = 30; /* max time to delay syncing data */
86 SYSCTL_PROC(_kern, OID_AUTO, syncdelay, CTLTYPE_INT | CTLFLAG_RW, 0, 0,
87 sysctl_kern_syncdelay, "I", "VFS data synchronization delay");
88 time_t filedelay = 30; /* time to delay syncing files */
89 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW,
90 &filedelay, 0, "File synchronization delay");
91 time_t dirdelay = 29; /* time to delay syncing directories */
92 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW,
93 &dirdelay, 0, "Directory synchronization delay");
94 time_t metadelay = 28; /* time to delay syncing metadata */
95 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW,
96 &metadelay, 0, "VFS metadata synchronization delay");
97 time_t retrydelay = 1; /* retry delay after failure */
98 SYSCTL_INT(_kern, OID_AUTO, retrydelay, CTLFLAG_RW,
99 &retrydelay, 0, "VFS retry synchronization delay");
100 static int rushjob; /* number of slots to run ASAP */
101 static int stat_rush_requests; /* number of times I/O speeded up */
102 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW,
103 &stat_rush_requests, 0, "");
105 LIST_HEAD(synclist, vnode);
107 #define SC_FLAG_EXIT (0x1) /* request syncer exit */
108 #define SC_FLAG_DONE (0x2) /* syncer confirm exit */
112 struct lwkt_token sc_token;
113 struct thread *sc_thread;
115 struct synclist *syncer_workitem_pending;
119 int syncer_rushjob; /* sequence vnodes faster */
120 int syncer_trigger; /* trigger full sync */
124 static void syncer_thread(void *);
127 sysctl_kern_syncdelay(SYSCTL_HANDLER_ARGS)
132 error = sysctl_handle_int(oidp, &v, 0, req);
133 if (error || !req->newptr)
137 if (v > SYNCER_MAXDELAY)
145 * The workitem queue.
147 * It is useful to delay writes of file data and filesystem metadata
148 * for tens of seconds so that quickly created and deleted files need
149 * not waste disk bandwidth being created and removed. To realize this,
150 * we append vnodes to a "workitem" queue. When running with a soft
151 * updates implementation, most pending metadata dependencies should
152 * not wait for more than a few seconds. Thus, mounted on block devices
153 * are delayed only about a half the time that file data is delayed.
154 * Similarly, directory updates are more critical, so are only delayed
155 * about a third the time that file data is delayed. Thus, there are
156 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
157 * one each second (driven off the filesystem syncer process). The
158 * syncer_delayno variable indicates the next queue that is to be processed.
159 * Items that need to be processed soon are placed in this queue:
161 * syncer_workitem_pending[syncer_delayno]
163 * A delay of fifteen seconds is done by placing the request fifteen
164 * entries later in the queue:
166 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
171 * Return the number of vnodes on the syncer's timed list. This will
172 * include the syncer vnode (mp->mnt_syncer) so if used, a minimum
173 * value of 1 will be returned.
176 vn_syncer_count(struct mount *mp)
178 struct syncer_ctx *ctx;
180 ctx = mp->mnt_syncer_ctx;
182 return (ctx->syncer_count);
187 * Add an item to the syncer work queue.
189 * WARNING: Cannot get vp->v_token here if not already held, we must
190 * depend on the syncer_token (which might already be held by
191 * the caller) to protect v_synclist and VONWORKLST.
193 * WARNING: The syncer depends on this function not blocking if the caller
194 * already holds the syncer token.
197 vn_syncer_add(struct vnode *vp, int delay)
199 struct syncer_ctx *ctx;
202 ctx = vp->v_mount->mnt_syncer_ctx;
203 lwkt_gettoken(&ctx->sc_token);
205 if (vp->v_flag & VONWORKLST) {
206 LIST_REMOVE(vp, v_synclist);
210 slot = -delay & ctx->syncer_mask;
212 if (delay > SYNCER_MAXDELAY - 2)
213 delay = SYNCER_MAXDELAY - 2;
214 slot = (ctx->syncer_delayno + delay) & ctx->syncer_mask;
217 LIST_INSERT_HEAD(&ctx->syncer_workitem_pending[slot], vp, v_synclist);
218 vsetflags(vp, VONWORKLST);
221 lwkt_reltoken(&ctx->sc_token);
225 * Removes the vnode from the syncer list. Since we might block while
226 * acquiring the syncer_token we have to [re]check conditions to determine
227 * that it is ok to remove the vnode.
229 * Force removal if force != 0. This can only occur during a forced unmount.
231 * vp->v_token held on call
234 vn_syncer_remove(struct vnode *vp, int force)
236 struct syncer_ctx *ctx;
238 ctx = vp->v_mount->mnt_syncer_ctx;
239 lwkt_gettoken(&ctx->sc_token);
241 if ((vp->v_flag & (VISDIRTY | VONWORKLST | VOBJDIRTY)) == VONWORKLST &&
242 RB_EMPTY(&vp->v_rbdirty_tree)) {
243 vclrflags(vp, VONWORKLST);
244 LIST_REMOVE(vp, v_synclist);
246 } else if (force && (vp->v_flag & VONWORKLST)) {
247 vclrflags(vp, VONWORKLST);
248 LIST_REMOVE(vp, v_synclist);
252 lwkt_reltoken(&ctx->sc_token);
256 * vnode must be locked
259 vclrisdirty(struct vnode *vp)
261 vclrflags(vp, VISDIRTY);
262 if (vp->v_flag & VONWORKLST)
263 vn_syncer_remove(vp, 0);
267 vclrobjdirty(struct vnode *vp)
269 vclrflags(vp, VOBJDIRTY);
270 if (vp->v_flag & VONWORKLST)
271 vn_syncer_remove(vp, 0);
275 * vnode must be stable
278 vsetisdirty(struct vnode *vp)
280 struct syncer_ctx *ctx;
282 if ((vp->v_flag & VISDIRTY) == 0) {
283 ctx = vp->v_mount->mnt_syncer_ctx;
284 vsetflags(vp, VISDIRTY);
285 lwkt_gettoken(&ctx->sc_token);
286 if ((vp->v_flag & VONWORKLST) == 0)
287 vn_syncer_add(vp, syncdelay);
288 lwkt_reltoken(&ctx->sc_token);
293 vsetobjdirty(struct vnode *vp)
295 struct syncer_ctx *ctx;
297 if ((vp->v_flag & VOBJDIRTY) == 0) {
298 ctx = vp->v_mount->mnt_syncer_ctx;
299 vsetflags(vp, VOBJDIRTY);
300 lwkt_gettoken(&ctx->sc_token);
301 if ((vp->v_flag & VONWORKLST) == 0)
302 vn_syncer_add(vp, syncdelay);
303 lwkt_reltoken(&ctx->sc_token);
308 * Create per-filesystem syncer process
311 vn_syncer_thr_create(struct mount *mp)
313 struct syncer_ctx *ctx;
314 static int syncalloc = 0;
316 ctx = kmalloc(sizeof(struct syncer_ctx), M_TEMP, M_WAITOK | M_ZERO);
319 ctx->syncer_workitem_pending = hashinit(SYNCER_MAXDELAY, M_DEVBUF,
321 ctx->syncer_delayno = 0;
322 lwkt_token_init(&ctx->sc_token, "syncer");
323 mp->mnt_syncer_ctx = ctx;
324 kthread_create(syncer_thread, ctx, &ctx->sc_thread,
325 "syncer%d", ++syncalloc & 0x7FFFFFFF);
329 * Stop per-filesystem syncer process
332 vn_syncer_thr_stop(struct mount *mp)
334 struct syncer_ctx *ctx;
336 ctx = mp->mnt_syncer_ctx;
340 lwkt_gettoken(&ctx->sc_token);
342 /* Signal the syncer process to exit */
343 ctx->sc_flags |= SC_FLAG_EXIT;
346 /* Wait till syncer process exits */
347 while ((ctx->sc_flags & SC_FLAG_DONE) == 0)
348 tsleep(&ctx->sc_flags, 0, "syncexit", hz);
350 mp->mnt_syncer_ctx = NULL;
351 lwkt_reltoken(&ctx->sc_token);
353 hashdestroy(ctx->syncer_workitem_pending, M_DEVBUF, ctx->syncer_mask);
357 struct thread *updatethread;
360 * System filesystem synchronizer daemon.
363 syncer_thread(void *_ctx)
365 struct syncer_ctx *ctx = _ctx;
366 struct synclist *slp;
371 int vnodes_synced = 0;
376 kproc_suspend_loop();
378 starttime = time_uptime;
379 lwkt_gettoken(&ctx->sc_token);
382 * Push files whose dirty time has expired. Be careful
383 * of interrupt race on slp queue.
385 * Note that vsyncscan() and vn_syncer_one() can pull items
386 * off the same list, so we shift vp's position in the
389 slp = &ctx->syncer_workitem_pending[ctx->syncer_delayno];
392 * If syncer_trigger is set (from trigger_syncer(mp)),
393 * Immediately do a full filesystem sync.
395 if (ctx->syncer_trigger) {
396 ctx->syncer_trigger = 0;
397 if (ctx->sc_mp && ctx->sc_mp->mnt_syncer) {
398 vp = ctx->sc_mp->mnt_syncer;
399 if (vp->v_flag & VONWORKLST) {
400 vn_syncer_add(vp, retrydelay);
401 if (vget(vp, LK_EXCLUSIVE) == 0) {
402 VOP_FSYNC(vp, MNT_LAZY, 0);
410 while ((vp = LIST_FIRST(slp)) != NULL) {
411 vn_syncer_add(vp, retrydelay);
412 if (ctx->syncer_forced) {
413 if (vget(vp, LK_EXCLUSIVE) == 0) {
414 VOP_FSYNC(vp, MNT_NOWAIT, 0);
419 if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
420 VOP_FSYNC(vp, MNT_LAZY, 0);
428 * Increment the slot upon completion.
430 ctx->syncer_delayno = (ctx->syncer_delayno + 1) &
433 sc_flags = ctx->sc_flags;
435 /* Exit on unmount */
436 if (sc_flags & SC_FLAG_EXIT)
439 lwkt_reltoken(&ctx->sc_token);
442 * Do sync processing for each mount.
445 bio_ops_sync(ctx->sc_mp);
448 * The variable rushjob allows the kernel to speed up the
449 * processing of the filesystem syncer process. A rushjob
450 * value of N tells the filesystem syncer to process the next
451 * N seconds worth of work on its queue ASAP. Currently rushjob
452 * is used by the soft update code to speed up the filesystem
453 * syncer process when the incore state is getting so far
454 * ahead of the disk that the kernel memory pool is being
455 * threatened with exhaustion.
457 delta = rushjob - ctx->syncer_rushjob;
458 if ((u_int)delta > syncdelay / 2) {
459 ctx->syncer_rushjob = rushjob - syncdelay / 2;
460 tsleep(&dummy, 0, "rush", 1);
464 ++ctx->syncer_rushjob;
465 tsleep(&dummy, 0, "rush", 1);
470 * If it has taken us less than a second to process the
471 * current work, then wait. Otherwise start right over
472 * again. We can still lose time if any single round
473 * takes more than two seconds, but it does not really
474 * matter as we are just trying to generally pace the
475 * filesystem activity.
477 if (time_uptime == starttime)
478 tsleep(ctx, 0, "syncer", hz);
482 * Unmount/exit path for per-filesystem syncers; sc_token held
484 ctx->sc_flags |= SC_FLAG_DONE;
485 sc_flagsp = &ctx->sc_flags;
486 lwkt_reltoken(&ctx->sc_token);
493 * This allows a filesystem to pro-actively request that a dirty
494 * vnode be fsync()d. This routine does not guarantee that one
495 * will actually be fsynced.
498 vn_syncer_one(struct mount *mp)
500 struct syncer_ctx *ctx;
501 struct synclist *slp;
506 ctx = mp->mnt_syncer_ctx;
507 i = ctx->syncer_delayno & ctx->syncer_mask;
510 if (lwkt_trytoken(&ctx->sc_token) == 0)
514 * Look ahead on our syncer time array.
517 slp = &ctx->syncer_workitem_pending[i];
518 vp = LIST_FIRST(slp);
519 if (vp && vp->v_type == VNON)
520 vp = LIST_NEXT(vp, v_synclist);
523 i = (i + 1) & ctx->syncer_mask;
524 /* i will be wrong if we stop here but vp is NULL so ok */
528 * Process one vnode, skip the syncer vnode but also stop
529 * if the syncer vnode is the only thing on this list.
532 vn_syncer_add(vp, retrydelay);
533 if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
534 VOP_FSYNC(vp, MNT_LAZY, 0);
538 lwkt_reltoken(&ctx->sc_token);
542 * Request that the syncer daemon for a specific mount speed up its work.
543 * If mp is NULL the caller generally wants to speed up all syncers.
546 speedup_syncer(struct mount *mp)
549 * Don't bother protecting the test. unsleep_and_wakeup_thread()
550 * will only do something real if the thread is in the right state.
552 atomic_add_int(&rushjob, 1);
553 ++stat_rush_requests;
554 if (mp && mp->mnt_syncer_ctx)
555 wakeup(mp->mnt_syncer_ctx);
559 * trigger a full sync
562 trigger_syncer(struct mount *mp)
564 struct syncer_ctx *ctx;
566 if (mp && (ctx = mp->mnt_syncer_ctx) != NULL) {
567 if (ctx->syncer_trigger == 0) {
568 ctx->syncer_trigger = 1;
575 * Routine to create and manage a filesystem syncer vnode.
577 static int sync_close(struct vop_close_args *);
578 static int sync_fsync(struct vop_fsync_args *);
579 static int sync_inactive(struct vop_inactive_args *);
580 static int sync_reclaim (struct vop_reclaim_args *);
581 static int sync_print(struct vop_print_args *);
583 static struct vop_ops sync_vnode_vops = {
584 .vop_default = vop_eopnotsupp,
585 .vop_close = sync_close,
586 .vop_fsync = sync_fsync,
587 .vop_inactive = sync_inactive,
588 .vop_reclaim = sync_reclaim,
589 .vop_print = sync_print,
592 static struct vop_ops *sync_vnode_vops_p = &sync_vnode_vops;
594 VNODEOP_SET(sync_vnode_vops);
597 * Create a new filesystem syncer vnode for the specified mount point.
598 * This vnode is placed on the worklist and is responsible for sync'ing
601 * NOTE: read-only mounts are also placed on the worklist. The filesystem
602 * sync code is also responsible for cleaning up vnodes.
605 vfs_allocate_syncvnode(struct mount *mp)
608 static long start, incr, next;
611 /* Allocate a new vnode */
612 error = getspecialvnode(VT_VFS, mp, &sync_vnode_vops_p, &vp, 0, 0);
614 mp->mnt_syncer = NULL;
619 * Place the vnode onto the syncer worklist. We attempt to
620 * scatter them about on the list so that they will go off
621 * at evenly distributed times even if all the filesystems
622 * are mounted at once.
625 if (next == 0 || next > SYNCER_MAXDELAY) {
629 start = SYNCER_MAXDELAY / 2;
630 incr = SYNCER_MAXDELAY;
636 * Only put the syncer vnode onto the syncer list if we have a
637 * syncer thread. Some VFS's (aka NULLFS) don't need a syncer
640 if (mp->mnt_syncer_ctx)
641 vn_syncer_add(vp, syncdelay > 0 ? next % syncdelay : 0);
644 * The mnt_syncer field inherits the vnode reference, which is
645 * held until later decomissioning.
653 sync_close(struct vop_close_args *ap)
659 * Do a lazy sync of the filesystem.
661 * sync_fsync { struct vnode *a_vp, int a_waitfor }
664 sync_fsync(struct vop_fsync_args *ap)
666 struct vnode *syncvp = ap->a_vp;
667 struct mount *mp = syncvp->v_mount;
671 * We only need to do something if this is a lazy evaluation.
673 if ((ap->a_waitfor & MNT_LAZY) == 0)
677 * Move ourselves to the back of the sync list.
679 vn_syncer_add(syncvp, syncdelay);
682 * Walk the list of vnodes pushing all that are dirty and
683 * not already on the sync list, and freeing vnodes which have
684 * no refs and whos VM objects are empty. vfs_msync() handles
685 * the VM issues and must be called whether the mount is readonly
688 if (vfs_busy(mp, LK_NOWAIT) != 0)
690 if (mp->mnt_flag & MNT_RDONLY) {
691 vfs_msync(mp, MNT_NOWAIT);
693 asyncflag = mp->mnt_flag & MNT_ASYNC;
694 mp->mnt_flag &= ~MNT_ASYNC; /* ZZZ hack */
695 vfs_msync(mp, MNT_NOWAIT);
696 VFS_SYNC(mp, MNT_NOWAIT | MNT_LAZY);
698 mp->mnt_flag |= MNT_ASYNC;
705 * The syncer vnode is no longer referenced.
707 * sync_inactive { struct vnode *a_vp, struct proc *a_p }
710 sync_inactive(struct vop_inactive_args *ap)
712 vgone_vxlocked(ap->a_vp);
717 * The syncer vnode is no longer needed and is being decommissioned.
718 * This can only occur when the last reference has been released on
719 * mp->mnt_syncer, so mp->mnt_syncer had better be NULL.
721 * Modifications to the worklist must be protected with a critical
724 * sync_reclaim { struct vnode *a_vp }
727 sync_reclaim(struct vop_reclaim_args *ap)
729 struct vnode *vp = ap->a_vp;
730 struct syncer_ctx *ctx;
732 ctx = vp->v_mount->mnt_syncer_ctx;
734 lwkt_gettoken(&ctx->sc_token);
735 KKASSERT(vp->v_mount->mnt_syncer != vp);
736 if (vp->v_flag & VONWORKLST) {
737 LIST_REMOVE(vp, v_synclist);
738 vclrflags(vp, VONWORKLST);
741 lwkt_reltoken(&ctx->sc_token);
743 KKASSERT((vp->v_flag & VONWORKLST) == 0);
750 * This is very similar to vmntvnodescan() but it only scans the
751 * vnodes on the syncer list. VFS's which support faster VFS_SYNC
752 * operations use the VISDIRTY flag on the vnode to ensure that vnodes
753 * with dirty inodes are added to the syncer in addition to vnodes
754 * with dirty buffers, and can use this function instead of nmntvnodescan().
756 * This scan does not issue VOP_FSYNC()s. The supplied callback is intended
757 * to synchronize the file in the manner intended by the VFS using it.
759 * This is important when a system has millions of vnodes.
765 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
768 struct syncer_ctx *ctx;
769 struct synclist *slp;
775 if (vmsc_flags & VMSC_NOWAIT)
781 * Syncer list context. This API requires a dedicated syncer thread.
784 KKASSERT(mp->mnt_kern_flag & MNTK_THR_SYNC);
785 ctx = mp->mnt_syncer_ctx;
786 lwkt_gettoken(&ctx->sc_token);
789 * Setup for loop. Allow races against the syncer thread but
790 * require that the syncer thread no be lazy if we were told
793 i = ctx->syncer_delayno & ctx->syncer_mask;
794 if ((vmsc_flags & VMSC_NOWAIT) == 0)
795 ++ctx->syncer_forced;
796 for (count = 0; count <= ctx->syncer_mask; ++count) {
797 slp = &ctx->syncer_workitem_pending[i];
799 while ((vp = LIST_FIRST(slp)) != NULL) {
800 KKASSERT(vp->v_mount == mp);
801 if (vmsc_flags & VMSC_GETVP) {
802 if (vget(vp, LK_EXCLUSIVE | lkflags) == 0) {
803 slowfunc(mp, vp, data);
806 } else if (vmsc_flags & VMSC_GETVX) {
808 slowfunc(mp, vp, data);
812 slowfunc(mp, vp, data);
817 * vp could be invalid. However, if vp is still at
818 * the head of the list it is clearly valid and we
819 * can safely move it.
821 if (LIST_FIRST(slp) == vp)
822 vn_syncer_add(vp, -(i + syncdelay));
824 i = (i + 1) & ctx->syncer_mask;
827 if ((vmsc_flags & VMSC_NOWAIT) == 0)
828 --ctx->syncer_forced;
829 lwkt_reltoken(&ctx->sc_token);
834 * Print out a syncer vnode.
836 * sync_print { struct vnode *a_vp }
839 sync_print(struct vop_print_args *ap)
841 struct vnode *vp = ap->a_vp;
843 kprintf("syncer vnode");
844 lockmgr_printinfo(&vp->v_lock);