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
47 #include <sys/param.h>
48 #include <sys/systm.h>
51 #include <sys/dirent.h>
52 #include <sys/domain.h>
53 #include <sys/eventhandler.h>
54 #include <sys/fcntl.h>
55 #include <sys/kernel.h>
56 #include <sys/kthread.h>
57 #include <sys/malloc.h>
59 #include <sys/mount.h>
61 #include <sys/namei.h>
62 #include <sys/reboot.h>
63 #include <sys/socket.h>
65 #include <sys/sysctl.h>
66 #include <sys/syslog.h>
67 #include <sys/vmmeter.h>
68 #include <sys/vnode.h>
70 #include <machine/limits.h>
73 #include <vm/vm_object.h>
74 #include <vm/vm_extern.h>
75 #include <vm/vm_kern.h>
77 #include <vm/vm_map.h>
78 #include <vm/vm_page.h>
79 #include <vm/vm_pager.h>
80 #include <vm/vnode_pager.h>
83 #include <sys/thread2.h>
88 #define SYNCER_MAXDELAY 32
89 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
90 time_t syncdelay = 30; /* max time to delay syncing data */
91 SYSCTL_INT(_kern, OID_AUTO, syncdelay, CTLFLAG_RW,
92 &syncdelay, 0, "VFS data synchronization delay");
93 time_t filedelay = 30; /* time to delay syncing files */
94 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW,
95 &filedelay, 0, "File synchronization delay");
96 time_t dirdelay = 29; /* time to delay syncing directories */
97 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW,
98 &dirdelay, 0, "Directory synchronization delay");
99 time_t metadelay = 28; /* time to delay syncing metadata */
100 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW,
101 &metadelay, 0, "VFS metadata synchronization delay");
102 static int rushjob; /* number of slots to run ASAP */
103 static int stat_rush_requests; /* number of times I/O speeded up */
104 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW,
105 &stat_rush_requests, 0, "");
107 LIST_HEAD(synclist, vnode);
109 #define SC_FLAG_EXIT (0x1) /* request syncer exit */
110 #define SC_FLAG_DONE (0x2) /* syncer confirm exit */
111 #define SC_FLAG_BIOOPS_ALL (0x4) /* do bufops_sync(NULL) */
115 struct lwkt_token sc_token;
116 struct thread *sc_thread;
119 struct synclist *syncer_workitem_pending;
124 static struct syncer_ctx syncer_ctx0;
126 static void syncer_thread(void *);
129 syncer_ctx_init(struct syncer_ctx *ctx, struct mount *mp)
132 lwkt_token_init(&ctx->sc_token, "syncer");
135 ctx->syncer_workitem_pending = hashinit(syncer_maxdelay, M_DEVBUF,
137 ctx->syncer_delayno = 0;
141 * Called from vfsinit()
146 syncer_ctx_init(&syncer_ctx0, NULL);
147 syncer_maxdelay = syncer_ctx0.syncer_mask + 1;
148 syncer_ctx0.sc_flags |= SC_FLAG_BIOOPS_ALL;
150 /* Support schedcpu wakeup of syncer0 */
151 lbolt_syncer = &syncer_ctx0;
154 static struct syncer_ctx *
155 vn_get_syncer(struct vnode *vp) {
157 struct syncer_ctx *ctx;
162 ctx = mp->mnt_syncer_ctx;
170 * The workitem queue.
172 * It is useful to delay writes of file data and filesystem metadata
173 * for tens of seconds so that quickly created and deleted files need
174 * not waste disk bandwidth being created and removed. To realize this,
175 * we append vnodes to a "workitem" queue. When running with a soft
176 * updates implementation, most pending metadata dependencies should
177 * not wait for more than a few seconds. Thus, mounted on block devices
178 * are delayed only about a half the time that file data is delayed.
179 * Similarly, directory updates are more critical, so are only delayed
180 * about a third the time that file data is delayed. Thus, there are
181 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
182 * one each second (driven off the filesystem syncer process). The
183 * syncer_delayno variable indicates the next queue that is to be processed.
184 * Items that need to be processed soon are placed in this queue:
186 * syncer_workitem_pending[syncer_delayno]
188 * A delay of fifteen seconds is done by placing the request fifteen
189 * entries later in the queue:
191 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
196 * Add an item to the syncer work queue.
198 * WARNING: Cannot get vp->v_token here if not already held, we must
199 * depend on the syncer_token (which might already be held by
200 * the caller) to protect v_synclist and VONWORKLST.
205 vn_syncer_add(struct vnode *vp, int delay)
207 struct syncer_ctx *ctx;
210 ctx = vn_get_syncer(vp);
212 lwkt_gettoken(&ctx->sc_token);
214 if (vp->v_flag & VONWORKLST)
215 LIST_REMOVE(vp, v_synclist);
216 if (delay > syncer_maxdelay - 2)
217 delay = syncer_maxdelay - 2;
218 slot = (ctx->syncer_delayno + delay) & ctx->syncer_mask;
220 LIST_INSERT_HEAD(&ctx->syncer_workitem_pending[slot], vp, v_synclist);
221 vsetflags(vp, VONWORKLST);
223 lwkt_reltoken(&ctx->sc_token);
227 * Removes the vnode from the syncer list. Since we might block while
228 * acquiring the syncer_token we have to recheck conditions.
230 * vp->v_token held on call
233 vn_syncer_remove(struct vnode *vp)
235 struct syncer_ctx *ctx;
237 ctx = vn_get_syncer(vp);
239 lwkt_gettoken(&ctx->sc_token);
241 if ((vp->v_flag & VONWORKLST) && RB_EMPTY(&vp->v_rbdirty_tree)) {
242 vclrflags(vp, VONWORKLST);
243 LIST_REMOVE(vp, v_synclist);
246 lwkt_reltoken(&ctx->sc_token);
250 * Create per-filesystem syncer process
253 vn_syncer_thr_create(struct mount *mp)
255 struct syncer_ctx *ctx;
256 static int syncalloc = 0;
259 ctx = kmalloc(sizeof(struct syncer_ctx), M_TEMP, M_WAITOK);
261 syncer_ctx_init(ctx, mp);
262 mp->mnt_syncer_ctx = ctx;
264 rc = kthread_create(syncer_thread, ctx, &ctx->sc_thread,
265 "syncer%d", ++syncalloc);
269 vn_syncer_thr_getctx(struct mount *mp)
271 return (mp->mnt_syncer_ctx);
275 * Stop per-filesystem syncer process
278 vn_syncer_thr_stop(void *ctxp)
280 struct syncer_ctx *ctx;
284 lwkt_gettoken(&ctx->sc_token);
286 /* Signal the syncer process to exit */
287 ctx->sc_flags |= SC_FLAG_EXIT;
290 /* Wait till syncer process exits */
291 while ((ctx->sc_flags & SC_FLAG_DONE) == 0)
292 tsleep(&ctx->sc_flags, 0, "syncexit", hz);
294 lwkt_reltoken(&ctx->sc_token);
296 kfree(ctx->syncer_workitem_pending, M_DEVBUF);
300 struct thread *updatethread;
303 * System filesystem synchronizer daemon.
306 syncer_thread(void *_ctx)
308 struct thread *td = curthread;
309 struct syncer_ctx *ctx = _ctx;
310 struct synclist *slp;
315 int vnodes_synced = 0;
318 * syncer0 runs till system shutdown; per-filesystem syncers are
319 * terminated on filesystem unmount
321 if (ctx == &syncer_ctx0)
322 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
325 kproc_suspend_loop();
327 starttime = time_second;
328 lwkt_gettoken(&ctx->sc_token);
331 * Push files whose dirty time has expired. Be careful
332 * of interrupt race on slp queue.
334 slp = &ctx->syncer_workitem_pending[ctx->syncer_delayno];
335 ctx->syncer_delayno += 1;
336 if (ctx->syncer_delayno == syncer_maxdelay)
337 ctx->syncer_delayno = 0;
339 while ((vp = LIST_FIRST(slp)) != NULL) {
340 if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
341 VOP_FSYNC(vp, MNT_LAZY, 0);
347 * vp is stale but can still be used if we can
348 * verify that it remains at the head of the list.
349 * Be careful not to try to get vp->v_token as
350 * vp can become stale if this blocks.
352 * If the vp is still at the head of the list were
353 * unable to completely flush it and move it to
354 * a later slot to give other vnodes a fair shot.
356 * Note that v_tag VT_VFS vnodes can remain on the
357 * worklist with no dirty blocks, but sync_fsync()
358 * moves it to a later slot so we will never see it
361 * It is possible to race a vnode with no dirty
362 * buffers being removed from the list. If this
363 * occurs we will move the vnode in the synclist
364 * and then the other thread will remove it. Do
365 * not try to remove it here.
367 if (LIST_FIRST(slp) == vp)
368 vn_syncer_add(vp, syncdelay);
371 sc_flags = ctx->sc_flags;
373 /* Exit on unmount */
374 if (sc_flags & SC_FLAG_EXIT)
377 lwkt_reltoken(&ctx->sc_token);
380 * Do sync processing for each mount.
382 if (ctx->sc_mp || sc_flags & SC_FLAG_BIOOPS_ALL)
383 bio_ops_sync(ctx->sc_mp);
386 * The variable rushjob allows the kernel to speed up the
387 * processing of the filesystem syncer process. A rushjob
388 * value of N tells the filesystem syncer to process the next
389 * N seconds worth of work on its queue ASAP. Currently rushjob
390 * is used by the soft update code to speed up the filesystem
391 * syncer process when the incore state is getting so far
392 * ahead of the disk that the kernel memory pool is being
393 * threatened with exhaustion.
395 if (ctx == &syncer_ctx0 && rushjob > 0) {
396 atomic_subtract_int(&rushjob, 1);
400 * If it has taken us less than a second to process the
401 * current work, then wait. Otherwise start right over
402 * again. We can still lose time if any single round
403 * takes more than two seconds, but it does not really
404 * matter as we are just trying to generally pace the
405 * filesystem activity.
407 if (time_second == starttime)
408 tsleep(ctx, 0, "syncer", hz);
412 * Unmount/exit path for per-filesystem syncers; sc_token held
414 ctx->sc_flags |= SC_FLAG_DONE;
415 sc_flagsp = &ctx->sc_flags;
416 lwkt_reltoken(&ctx->sc_token);
423 syncer_thread_start(void) {
424 syncer_thread(&syncer_ctx0);
427 static struct kproc_desc up_kp = {
432 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
435 * Request the syncer daemon to speed up its work.
436 * We never push it to speed up more than half of its
437 * normal turn time, otherwise it could take over the cpu.
443 * Don't bother protecting the test. unsleep_and_wakeup_thread()
444 * will only do something real if the thread is in the right state.
446 wakeup(lbolt_syncer);
447 if (rushjob < syncdelay / 2) {
448 atomic_add_int(&rushjob, 1);
449 stat_rush_requests += 1;
456 * Routine to create and manage a filesystem syncer vnode.
458 static int sync_close(struct vop_close_args *);
459 static int sync_fsync(struct vop_fsync_args *);
460 static int sync_inactive(struct vop_inactive_args *);
461 static int sync_reclaim (struct vop_reclaim_args *);
462 static int sync_print(struct vop_print_args *);
464 static struct vop_ops sync_vnode_vops = {
465 .vop_default = vop_eopnotsupp,
466 .vop_close = sync_close,
467 .vop_fsync = sync_fsync,
468 .vop_inactive = sync_inactive,
469 .vop_reclaim = sync_reclaim,
470 .vop_print = sync_print,
473 static struct vop_ops *sync_vnode_vops_p = &sync_vnode_vops;
475 VNODEOP_SET(sync_vnode_vops);
478 * Create a new filesystem syncer vnode for the specified mount point.
479 * This vnode is placed on the worklist and is responsible for sync'ing
482 * NOTE: read-only mounts are also placed on the worklist. The filesystem
483 * sync code is also responsible for cleaning up vnodes.
486 vfs_allocate_syncvnode(struct mount *mp)
489 static long start, incr, next;
492 /* Allocate a new vnode */
493 error = getspecialvnode(VT_VFS, mp, &sync_vnode_vops_p, &vp, 0, 0);
495 mp->mnt_syncer = NULL;
500 * Place the vnode onto the syncer worklist. We attempt to
501 * scatter them about on the list so that they will go off
502 * at evenly distributed times even if all the filesystems
503 * are mounted at once.
506 if (next == 0 || next > syncer_maxdelay) {
510 start = syncer_maxdelay / 2;
511 incr = syncer_maxdelay;
515 vn_syncer_add(vp, syncdelay > 0 ? next % syncdelay : 0);
518 * The mnt_syncer field inherits the vnode reference, which is
519 * held until later decomissioning.
527 sync_close(struct vop_close_args *ap)
533 * Do a lazy sync of the filesystem.
535 * sync_fsync { struct vnode *a_vp, int a_waitfor }
538 sync_fsync(struct vop_fsync_args *ap)
540 struct vnode *syncvp = ap->a_vp;
541 struct mount *mp = syncvp->v_mount;
545 * We only need to do something if this is a lazy evaluation.
547 if ((ap->a_waitfor & MNT_LAZY) == 0)
551 * Move ourselves to the back of the sync list.
553 vn_syncer_add(syncvp, syncdelay);
556 * Walk the list of vnodes pushing all that are dirty and
557 * not already on the sync list, and freeing vnodes which have
558 * no refs and whos VM objects are empty. vfs_msync() handles
559 * the VM issues and must be called whether the mount is readonly
562 if (vfs_busy(mp, LK_NOWAIT) != 0)
564 if (mp->mnt_flag & MNT_RDONLY) {
565 vfs_msync(mp, MNT_NOWAIT);
567 asyncflag = mp->mnt_flag & MNT_ASYNC;
568 mp->mnt_flag &= ~MNT_ASYNC; /* ZZZ hack */
569 vfs_msync(mp, MNT_NOWAIT);
570 VFS_SYNC(mp, MNT_NOWAIT | MNT_LAZY);
572 mp->mnt_flag |= MNT_ASYNC;
579 * The syncer vnode is no longer referenced.
581 * sync_inactive { struct vnode *a_vp, struct proc *a_p }
584 sync_inactive(struct vop_inactive_args *ap)
586 vgone_vxlocked(ap->a_vp);
591 * The syncer vnode is no longer needed and is being decommissioned.
592 * This can only occur when the last reference has been released on
593 * mp->mnt_syncer, so mp->mnt_syncer had better be NULL.
595 * Modifications to the worklist must be protected with a critical
598 * sync_reclaim { struct vnode *a_vp }
601 sync_reclaim(struct vop_reclaim_args *ap)
603 struct vnode *vp = ap->a_vp;
604 struct syncer_ctx *ctx;
606 ctx = vn_get_syncer(vp);
608 lwkt_gettoken(&ctx->sc_token);
609 KKASSERT(vp->v_mount->mnt_syncer != vp);
610 if (vp->v_flag & VONWORKLST) {
611 LIST_REMOVE(vp, v_synclist);
612 vclrflags(vp, VONWORKLST);
614 lwkt_reltoken(&ctx->sc_token);
620 * Print out a syncer vnode.
622 * sync_print { struct vnode *a_vp }
625 sync_print(struct vop_print_args *ap)
627 struct vnode *vp = ap->a_vp;
629 kprintf("syncer vnode");
630 lockmgr_printinfo(&vp->v_lock);