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. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95
39 * $FreeBSD: src/sys/kern/vfs_subr.c,v 1.249.2.30 2003/04/04 20:35:57 tegge Exp $
43 * External virtual filesystem routines
46 #include <sys/param.h>
47 #include <sys/systm.h>
50 #include <sys/dirent.h>
51 #include <sys/domain.h>
52 #include <sys/eventhandler.h>
53 #include <sys/fcntl.h>
54 #include <sys/kernel.h>
55 #include <sys/kthread.h>
56 #include <sys/malloc.h>
58 #include <sys/mount.h>
60 #include <sys/namei.h>
61 #include <sys/reboot.h>
62 #include <sys/socket.h>
64 #include <sys/sysctl.h>
65 #include <sys/syslog.h>
66 #include <sys/vmmeter.h>
67 #include <sys/vnode.h>
69 #include <machine/limits.h>
72 #include <vm/vm_object.h>
73 #include <vm/vm_extern.h>
74 #include <vm/vm_kern.h>
76 #include <vm/vm_map.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_pager.h>
79 #include <vm/vnode_pager.h>
82 #include <sys/thread2.h>
87 #define SYNCER_MAXDELAY 32
88 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
89 time_t syncdelay = 30; /* max time to delay syncing data */
90 SYSCTL_INT(_kern, OID_AUTO, syncdelay, CTLFLAG_RW,
91 &syncdelay, 0, "VFS data synchronization delay");
92 time_t filedelay = 30; /* time to delay syncing files */
93 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW,
94 &filedelay, 0, "File synchronization delay");
95 time_t dirdelay = 29; /* time to delay syncing directories */
96 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW,
97 &dirdelay, 0, "Directory synchronization delay");
98 time_t metadelay = 28; /* time to delay syncing metadata */
99 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW,
100 &metadelay, 0, "VFS metadata synchronization delay");
101 static int rushjob; /* number of slots to run ASAP */
102 static int stat_rush_requests; /* number of times I/O speeded up */
103 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW,
104 &stat_rush_requests, 0, "");
106 LIST_HEAD(synclist, vnode);
108 #define SC_FLAG_EXIT (0x1) /* request syncer exit */
109 #define SC_FLAG_DONE (0x2) /* syncer confirm exit */
110 #define SC_FLAG_BIOOPS_ALL (0x4) /* do bufops_sync(NULL) */
114 struct lwkt_token sc_token;
115 struct thread *sc_thread;
118 struct synclist *syncer_workitem_pending;
123 static struct syncer_ctx syncer_ctx0;
125 static void syncer_thread(void *);
128 syncer_ctx_init(struct syncer_ctx *ctx, struct mount *mp)
131 lwkt_token_init(&ctx->sc_token, "syncer");
134 ctx->syncer_workitem_pending = hashinit(syncer_maxdelay, M_DEVBUF,
136 ctx->syncer_delayno = 0;
140 * Called from vfsinit()
145 syncer_ctx_init(&syncer_ctx0, NULL);
146 syncer_maxdelay = syncer_ctx0.syncer_mask + 1;
147 syncer_ctx0.sc_flags |= SC_FLAG_BIOOPS_ALL;
149 /* Support schedcpu wakeup of syncer0 */
150 lbolt_syncer = &syncer_ctx0;
153 static struct syncer_ctx *
154 vn_get_syncer(struct vnode *vp) {
156 struct syncer_ctx *ctx;
161 ctx = mp->mnt_syncer_ctx;
169 * The workitem queue.
171 * It is useful to delay writes of file data and filesystem metadata
172 * for tens of seconds so that quickly created and deleted files need
173 * not waste disk bandwidth being created and removed. To realize this,
174 * we append vnodes to a "workitem" queue. When running with a soft
175 * updates implementation, most pending metadata dependencies should
176 * not wait for more than a few seconds. Thus, mounted on block devices
177 * are delayed only about a half the time that file data is delayed.
178 * Similarly, directory updates are more critical, so are only delayed
179 * about a third the time that file data is delayed. Thus, there are
180 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
181 * one each second (driven off the filesystem syncer process). The
182 * syncer_delayno variable indicates the next queue that is to be processed.
183 * Items that need to be processed soon are placed in this queue:
185 * syncer_workitem_pending[syncer_delayno]
187 * A delay of fifteen seconds is done by placing the request fifteen
188 * entries later in the queue:
190 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
195 * Add an item to the syncer work queue.
197 * WARNING: Cannot get vp->v_token here if not already held, we must
198 * depend on the syncer_token (which might already be held by
199 * the caller) to protect v_synclist and VONWORKLST.
204 vn_syncer_add(struct vnode *vp, int delay)
206 struct syncer_ctx *ctx;
209 ctx = vn_get_syncer(vp);
211 lwkt_gettoken(&ctx->sc_token);
213 if (vp->v_flag & VONWORKLST)
214 LIST_REMOVE(vp, v_synclist);
215 if (delay > syncer_maxdelay - 2)
216 delay = syncer_maxdelay - 2;
217 slot = (ctx->syncer_delayno + delay) & ctx->syncer_mask;
219 LIST_INSERT_HEAD(&ctx->syncer_workitem_pending[slot], vp, v_synclist);
220 vsetflags(vp, VONWORKLST);
222 lwkt_reltoken(&ctx->sc_token);
226 * Removes the vnode from the syncer list. Since we might block while
227 * acquiring the syncer_token we have to recheck conditions.
229 * vp->v_token held on call
232 vn_syncer_remove(struct vnode *vp)
234 struct syncer_ctx *ctx;
236 ctx = vn_get_syncer(vp);
238 lwkt_gettoken(&ctx->sc_token);
240 if ((vp->v_flag & VONWORKLST) && RB_EMPTY(&vp->v_rbdirty_tree)) {
241 vclrflags(vp, VONWORKLST);
242 LIST_REMOVE(vp, v_synclist);
245 lwkt_reltoken(&ctx->sc_token);
249 * Create per-filesystem syncer process
252 vn_syncer_thr_create(struct mount *mp)
254 struct syncer_ctx *ctx;
255 static int syncalloc = 0;
258 ctx = kmalloc(sizeof(struct syncer_ctx), M_TEMP, M_WAITOK);
260 syncer_ctx_init(ctx, mp);
261 mp->mnt_syncer_ctx = ctx;
263 rc = kthread_create(syncer_thread, ctx, &ctx->sc_thread,
264 "syncer%d", ++syncalloc);
268 vn_syncer_thr_getctx(struct mount *mp)
270 return (mp->mnt_syncer_ctx);
274 * Stop per-filesystem syncer process
277 vn_syncer_thr_stop(void *ctxp)
279 struct syncer_ctx *ctx;
283 lwkt_gettoken(&ctx->sc_token);
285 /* Signal the syncer process to exit */
286 ctx->sc_flags |= SC_FLAG_EXIT;
289 /* Wait till syncer process exits */
290 while ((ctx->sc_flags & SC_FLAG_DONE) == 0)
291 tsleep(&ctx->sc_flags, 0, "syncexit", hz);
293 lwkt_reltoken(&ctx->sc_token);
295 hashdestroy(ctx->syncer_workitem_pending, M_DEVBUF, ctx->syncer_mask);
299 struct thread *updatethread;
302 * System filesystem synchronizer daemon.
305 syncer_thread(void *_ctx)
307 struct thread *td = curthread;
308 struct syncer_ctx *ctx = _ctx;
309 struct synclist *slp;
314 int vnodes_synced = 0;
317 * syncer0 runs till system shutdown; per-filesystem syncers are
318 * terminated on filesystem unmount
320 if (ctx == &syncer_ctx0)
321 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
324 kproc_suspend_loop();
326 starttime = time_second;
327 lwkt_gettoken(&ctx->sc_token);
330 * Push files whose dirty time has expired. Be careful
331 * of interrupt race on slp queue.
333 slp = &ctx->syncer_workitem_pending[ctx->syncer_delayno];
334 ctx->syncer_delayno += 1;
335 if (ctx->syncer_delayno == syncer_maxdelay)
336 ctx->syncer_delayno = 0;
338 while ((vp = LIST_FIRST(slp)) != NULL) {
339 if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
340 VOP_FSYNC(vp, MNT_LAZY, 0);
346 * vp is stale but can still be used if we can
347 * verify that it remains at the head of the list.
348 * Be careful not to try to get vp->v_token as
349 * vp can become stale if this blocks.
351 * If the vp is still at the head of the list were
352 * unable to completely flush it and move it to
353 * a later slot to give other vnodes a fair shot.
355 * Note that v_tag VT_VFS vnodes can remain on the
356 * worklist with no dirty blocks, but sync_fsync()
357 * moves it to a later slot so we will never see it
360 * It is possible to race a vnode with no dirty
361 * buffers being removed from the list. If this
362 * occurs we will move the vnode in the synclist
363 * and then the other thread will remove it. Do
364 * not try to remove it here.
366 if (LIST_FIRST(slp) == vp)
367 vn_syncer_add(vp, syncdelay);
370 sc_flags = ctx->sc_flags;
372 /* Exit on unmount */
373 if (sc_flags & SC_FLAG_EXIT)
376 lwkt_reltoken(&ctx->sc_token);
379 * Do sync processing for each mount.
381 if (ctx->sc_mp || sc_flags & SC_FLAG_BIOOPS_ALL)
382 bio_ops_sync(ctx->sc_mp);
385 * The variable rushjob allows the kernel to speed up the
386 * processing of the filesystem syncer process. A rushjob
387 * value of N tells the filesystem syncer to process the next
388 * N seconds worth of work on its queue ASAP. Currently rushjob
389 * is used by the soft update code to speed up the filesystem
390 * syncer process when the incore state is getting so far
391 * ahead of the disk that the kernel memory pool is being
392 * threatened with exhaustion.
394 if (ctx == &syncer_ctx0 && rushjob > 0) {
395 atomic_subtract_int(&rushjob, 1);
399 * If it has taken us less than a second to process the
400 * current work, then wait. Otherwise start right over
401 * again. We can still lose time if any single round
402 * takes more than two seconds, but it does not really
403 * matter as we are just trying to generally pace the
404 * filesystem activity.
406 if (time_second == starttime)
407 tsleep(ctx, 0, "syncer", hz);
411 * Unmount/exit path for per-filesystem syncers; sc_token held
413 ctx->sc_flags |= SC_FLAG_DONE;
414 sc_flagsp = &ctx->sc_flags;
415 lwkt_reltoken(&ctx->sc_token);
422 syncer_thread_start(void) {
423 syncer_thread(&syncer_ctx0);
426 static struct kproc_desc up_kp = {
431 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
434 * Request the syncer daemon to speed up its work.
435 * We never push it to speed up more than half of its
436 * normal turn time, otherwise it could take over the cpu.
442 * Don't bother protecting the test. unsleep_and_wakeup_thread()
443 * will only do something real if the thread is in the right state.
445 wakeup(lbolt_syncer);
446 if (rushjob < syncdelay / 2) {
447 atomic_add_int(&rushjob, 1);
448 stat_rush_requests += 1;
455 * Routine to create and manage a filesystem syncer vnode.
457 static int sync_close(struct vop_close_args *);
458 static int sync_fsync(struct vop_fsync_args *);
459 static int sync_inactive(struct vop_inactive_args *);
460 static int sync_reclaim (struct vop_reclaim_args *);
461 static int sync_print(struct vop_print_args *);
463 static struct vop_ops sync_vnode_vops = {
464 .vop_default = vop_eopnotsupp,
465 .vop_close = sync_close,
466 .vop_fsync = sync_fsync,
467 .vop_inactive = sync_inactive,
468 .vop_reclaim = sync_reclaim,
469 .vop_print = sync_print,
472 static struct vop_ops *sync_vnode_vops_p = &sync_vnode_vops;
474 VNODEOP_SET(sync_vnode_vops);
477 * Create a new filesystem syncer vnode for the specified mount point.
478 * This vnode is placed on the worklist and is responsible for sync'ing
481 * NOTE: read-only mounts are also placed on the worklist. The filesystem
482 * sync code is also responsible for cleaning up vnodes.
485 vfs_allocate_syncvnode(struct mount *mp)
488 static long start, incr, next;
491 /* Allocate a new vnode */
492 error = getspecialvnode(VT_VFS, mp, &sync_vnode_vops_p, &vp, 0, 0);
494 mp->mnt_syncer = NULL;
499 * Place the vnode onto the syncer worklist. We attempt to
500 * scatter them about on the list so that they will go off
501 * at evenly distributed times even if all the filesystems
502 * are mounted at once.
505 if (next == 0 || next > syncer_maxdelay) {
509 start = syncer_maxdelay / 2;
510 incr = syncer_maxdelay;
514 vn_syncer_add(vp, syncdelay > 0 ? next % syncdelay : 0);
517 * The mnt_syncer field inherits the vnode reference, which is
518 * held until later decomissioning.
526 sync_close(struct vop_close_args *ap)
532 * Do a lazy sync of the filesystem.
534 * sync_fsync { struct vnode *a_vp, int a_waitfor }
537 sync_fsync(struct vop_fsync_args *ap)
539 struct vnode *syncvp = ap->a_vp;
540 struct mount *mp = syncvp->v_mount;
544 * We only need to do something if this is a lazy evaluation.
546 if ((ap->a_waitfor & MNT_LAZY) == 0)
550 * Move ourselves to the back of the sync list.
552 vn_syncer_add(syncvp, syncdelay);
555 * Walk the list of vnodes pushing all that are dirty and
556 * not already on the sync list, and freeing vnodes which have
557 * no refs and whos VM objects are empty. vfs_msync() handles
558 * the VM issues and must be called whether the mount is readonly
561 if (vfs_busy(mp, LK_NOWAIT) != 0)
563 if (mp->mnt_flag & MNT_RDONLY) {
564 vfs_msync(mp, MNT_NOWAIT);
566 asyncflag = mp->mnt_flag & MNT_ASYNC;
567 mp->mnt_flag &= ~MNT_ASYNC; /* ZZZ hack */
568 vfs_msync(mp, MNT_NOWAIT);
569 VFS_SYNC(mp, MNT_NOWAIT | MNT_LAZY);
571 mp->mnt_flag |= MNT_ASYNC;
578 * The syncer vnode is no longer referenced.
580 * sync_inactive { struct vnode *a_vp, struct proc *a_p }
583 sync_inactive(struct vop_inactive_args *ap)
585 vgone_vxlocked(ap->a_vp);
590 * The syncer vnode is no longer needed and is being decommissioned.
591 * This can only occur when the last reference has been released on
592 * mp->mnt_syncer, so mp->mnt_syncer had better be NULL.
594 * Modifications to the worklist must be protected with a critical
597 * sync_reclaim { struct vnode *a_vp }
600 sync_reclaim(struct vop_reclaim_args *ap)
602 struct vnode *vp = ap->a_vp;
603 struct syncer_ctx *ctx;
605 ctx = vn_get_syncer(vp);
607 lwkt_gettoken(&ctx->sc_token);
608 KKASSERT(vp->v_mount->mnt_syncer != vp);
609 if (vp->v_flag & VONWORKLST) {
610 LIST_REMOVE(vp, v_synclist);
611 vclrflags(vp, VONWORKLST);
613 lwkt_reltoken(&ctx->sc_token);
619 * Print out a syncer vnode.
621 * sync_print { struct vnode *a_vp }
624 sync_print(struct vop_print_args *ap)
626 struct vnode *vp = ap->a_vp;
628 kprintf("syncer vnode");
629 lockmgr_printinfo(&vp->v_lock);