xref: /dflybsd-src/sys/kern/vfs_sync.c (revision ef94481424b6bf8cbda60b2a9633e25f1513329c)
1 /*
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.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
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.
21  *
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
32  * SUCH DAMAGE.
33  *
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 $
36  */
37 
38 /*
39  * External virtual filesystem routines
40  */
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/buf.h>
45 #include <sys/conf.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>
53 #include <sys/mbuf.h>
54 #include <sys/mount.h>
55 #include <sys/proc.h>
56 #include <sys/namei.h>
57 #include <sys/reboot.h>
58 #include <sys/socket.h>
59 #include <sys/stat.h>
60 #include <sys/sysctl.h>
61 #include <sys/syslog.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 
65 #include <machine/limits.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_kern.h>
71 #include <vm/pmap.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>
76 
77 #include <sys/buf2.h>
78 #include <sys/thread2.h>
79 
80 /*
81  * The workitem queue.
82  */
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, "");
104 
105 LIST_HEAD(synclist, vnode);
106 
107 #define	SC_FLAG_EXIT		(0x1)		/* request syncer exit */
108 #define	SC_FLAG_DONE		(0x2)		/* syncer confirm exit */
109 
110 struct syncer_ctx {
111 	struct mount		*sc_mp;
112 	struct lwkt_token 	sc_token;
113 	struct thread		*sc_thread;
114 	int			sc_flags;
115 	struct synclist 	*syncer_workitem_pending;
116 	long			syncer_mask;
117 	int 			syncer_delayno;
118 	int			syncer_forced;
119 	int			syncer_rushjob;	/* sequence vnodes faster */
120 	int			syncer_trigger;	/* trigger full sync */
121 	long			syncer_count;
122 };
123 
124 static void syncer_thread(void *);
125 
126 static int
127 sysctl_kern_syncdelay(SYSCTL_HANDLER_ARGS)
128 {
129 	int error;
130 	int v = syncdelay;
131 
132 	error = sysctl_handle_int(oidp, &v, 0, req);
133 	if (error || !req->newptr)
134 		return (error);
135 	if (v < 1)
136 		v = 1;
137 	if (v > SYNCER_MAXDELAY)
138 		v = SYNCER_MAXDELAY;
139 	syncdelay = v;
140 
141 	return(0);
142 }
143 
144 /*
145  * The workitem queue.
146  *
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:
160  *
161  *	syncer_workitem_pending[syncer_delayno]
162  *
163  * A delay of fifteen seconds is done by placing the request fifteen
164  * entries later in the queue:
165  *
166  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
167  *
168  */
169 
170 /*
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.
174  */
175 long
176 vn_syncer_count(struct mount *mp)
177 {
178 	struct syncer_ctx *ctx;
179 
180 	ctx = mp->mnt_syncer_ctx;
181 	if (ctx)
182 		return (ctx->syncer_count);
183 	return 0;
184 }
185 
186 /*
187  * Add an item to the syncer work queue.
188  *
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.
192  *
193  * WARNING: The syncer depends on this function not blocking if the caller
194  *	    already holds the syncer token.
195  */
196 void
197 vn_syncer_add(struct vnode *vp, int delay)
198 {
199 	struct syncer_ctx *ctx;
200 	int slot;
201 
202 	ctx = vp->v_mount->mnt_syncer_ctx;
203 	lwkt_gettoken(&ctx->sc_token);
204 
205 	if (vp->v_flag & VONWORKLST) {
206 		LIST_REMOVE(vp, v_synclist);
207 		--ctx->syncer_count;
208 	}
209 	if (delay <= 0) {
210 		slot = -delay & ctx->syncer_mask;
211 	} else {
212 		if (delay > SYNCER_MAXDELAY - 2)
213 			delay = SYNCER_MAXDELAY - 2;
214 		slot = (ctx->syncer_delayno + delay) & ctx->syncer_mask;
215 	}
216 
217 	LIST_INSERT_HEAD(&ctx->syncer_workitem_pending[slot], vp, v_synclist);
218 	vsetflags(vp, VONWORKLST);
219 	++ctx->syncer_count;
220 
221 	lwkt_reltoken(&ctx->sc_token);
222 }
223 
224 /*
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.
228  *
229  * Force removal if force != 0.  This can only occur during a forced unmount.
230  *
231  * vp->v_token held on call
232  */
233 void
234 vn_syncer_remove(struct vnode *vp, int force)
235 {
236 	struct syncer_ctx *ctx;
237 
238 	ctx = vp->v_mount->mnt_syncer_ctx;
239 	lwkt_gettoken(&ctx->sc_token);
240 
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);
245 		--ctx->syncer_count;
246 	} else if (force && (vp->v_flag & VONWORKLST)) {
247 		vclrflags(vp, VONWORKLST);
248 		LIST_REMOVE(vp, v_synclist);
249 		--ctx->syncer_count;
250 	}
251 
252 	lwkt_reltoken(&ctx->sc_token);
253 }
254 
255 /*
256  * vnode must be locked
257  */
258 void
259 vclrisdirty(struct vnode *vp)
260 {
261 	vclrflags(vp, VISDIRTY);
262 	if (vp->v_flag & VONWORKLST)
263 		vn_syncer_remove(vp, 0);
264 }
265 
266 void
267 vclrobjdirty(struct vnode *vp)
268 {
269 	vclrflags(vp, VOBJDIRTY);
270 	if (vp->v_flag & VONWORKLST)
271 		vn_syncer_remove(vp, 0);
272 }
273 
274 /*
275  * vnode must be stable
276  */
277 void
278 vsetisdirty(struct vnode *vp)
279 {
280 	struct syncer_ctx *ctx;
281 
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);
289 	}
290 }
291 
292 void
293 vsetobjdirty(struct vnode *vp)
294 {
295 	struct syncer_ctx *ctx;
296 
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);
304 	}
305 }
306 
307 /*
308  * Create per-filesystem syncer process
309  */
310 void
311 vn_syncer_thr_create(struct mount *mp)
312 {
313 	struct syncer_ctx *ctx;
314 	static int syncalloc = 0;
315 
316 	ctx = kmalloc(sizeof(struct syncer_ctx), M_TEMP, M_WAITOK | M_ZERO);
317 	ctx->sc_mp = mp;
318 	ctx->sc_flags = 0;
319 	ctx->syncer_workitem_pending = hashinit(SYNCER_MAXDELAY, M_DEVBUF,
320 						&ctx->syncer_mask);
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);
326 }
327 
328 /*
329  * Stop per-filesystem syncer process
330  */
331 void
332 vn_syncer_thr_stop(struct mount *mp)
333 {
334 	struct syncer_ctx *ctx;
335 
336 	ctx = mp->mnt_syncer_ctx;
337 	if (ctx == NULL)
338 		return;
339 
340 	lwkt_gettoken(&ctx->sc_token);
341 
342 	/* Signal the syncer process to exit */
343 	ctx->sc_flags |= SC_FLAG_EXIT;
344 	wakeup(ctx);
345 
346 	/* Wait till syncer process exits */
347 	while ((ctx->sc_flags & SC_FLAG_DONE) == 0)
348 		tsleep(&ctx->sc_flags, 0, "syncexit", hz);
349 
350 	mp->mnt_syncer_ctx = NULL;
351 	lwkt_reltoken(&ctx->sc_token);
352 
353 	hashdestroy(ctx->syncer_workitem_pending, M_DEVBUF, ctx->syncer_mask);
354 	kfree(ctx, M_TEMP);
355 }
356 
357 struct  thread *updatethread;
358 
359 /*
360  * System filesystem synchronizer daemon.
361  */
362 static void
363 syncer_thread(void *_ctx)
364 {
365 	struct syncer_ctx *ctx = _ctx;
366 	struct synclist *slp;
367 	struct vnode *vp;
368 	long starttime;
369 	int *sc_flagsp;
370 	int sc_flags;
371 	int vnodes_synced = 0;
372 	int delta;
373 	int dummy = 0;
374 
375 	for (;;) {
376 		kproc_suspend_loop();
377 
378 		starttime = time_uptime;
379 		lwkt_gettoken(&ctx->sc_token);
380 
381 		/*
382 		 * Push files whose dirty time has expired.  Be careful
383 		 * of interrupt race on slp queue.
384 		 *
385 		 * Note that vsyncscan() and vn_syncer_one() can pull items
386 		 * off the same list, so we shift vp's position in the
387 		 * list immediately.
388 		 */
389 		slp = &ctx->syncer_workitem_pending[ctx->syncer_delayno];
390 
391 		/*
392 		 * If syncer_trigger is set (from trigger_syncer(mp)),
393 		 * Immediately do a full filesystem sync.
394 		 */
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);
403 						vput(vp);
404 						vnodes_synced++;
405 					}
406 				}
407 			}
408 		}
409 
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);
415 					vput(vp);
416 					vnodes_synced++;
417 				}
418 			} else {
419 				if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
420 					VOP_FSYNC(vp, MNT_LAZY, 0);
421 					vput(vp);
422 					vnodes_synced++;
423 				}
424 			}
425 		}
426 
427 		/*
428 		 * Increment the slot upon completion.
429 		 */
430 		ctx->syncer_delayno = (ctx->syncer_delayno + 1) &
431 				      ctx->syncer_mask;
432 
433 		sc_flags = ctx->sc_flags;
434 
435 		/* Exit on unmount */
436 		if (sc_flags & SC_FLAG_EXIT)
437 			break;
438 
439 		lwkt_reltoken(&ctx->sc_token);
440 
441 		/*
442 		 * Do sync processing for each mount.
443 		 */
444 		if (ctx->sc_mp)
445 			bio_ops_sync(ctx->sc_mp);
446 
447 		/*
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.
456 		 */
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);
461 			continue;
462 		}
463 		if (delta) {
464 			++ctx->syncer_rushjob;
465 			tsleep(&dummy, 0, "rush", 1);
466 			continue;
467 		}
468 
469 		/*
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.
476 		 */
477 		if (time_uptime == starttime)
478 			tsleep(ctx, 0, "syncer", hz);
479 	}
480 
481 	/*
482 	 * Unmount/exit path for per-filesystem syncers; sc_token held
483 	 */
484 	ctx->sc_flags |= SC_FLAG_DONE;
485 	sc_flagsp = &ctx->sc_flags;
486 	lwkt_reltoken(&ctx->sc_token);
487 	wakeup(sc_flagsp);
488 
489 	kthread_exit();
490 }
491 
492 /*
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.
496  */
497 void
498 vn_syncer_one(struct mount *mp)
499 {
500 	struct syncer_ctx *ctx;
501 	struct synclist *slp;
502 	struct vnode *vp;
503 	int i;
504 	int n = syncdelay;
505 
506 	ctx = mp->mnt_syncer_ctx;
507 	i = ctx->syncer_delayno & ctx->syncer_mask;
508 	cpu_ccfence();
509 
510 	if (lwkt_trytoken(&ctx->sc_token) == 0)
511 		return;
512 
513 	/*
514 	 * Look ahead on our syncer time array.
515 	 */
516 	do {
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);
521 		if (vp)
522 			break;
523 		i = (i + 1) & ctx->syncer_mask;
524 		/* i will be wrong if we stop here but vp is NULL so ok */
525 	} while(--n);
526 
527 	/*
528 	 * Process one vnode, skip the syncer vnode but also stop
529 	 * if the syncer vnode is the only thing on this list.
530 	 */
531 	if (vp) {
532 		vn_syncer_add(vp, retrydelay);
533 		if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
534 			VOP_FSYNC(vp, MNT_LAZY, 0);
535 			vput(vp);
536 		}
537 	}
538 	lwkt_reltoken(&ctx->sc_token);
539 }
540 
541 /*
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.
544  */
545 void
546 speedup_syncer(struct mount *mp)
547 {
548 	/*
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.
551 	 */
552 	atomic_add_int(&rushjob, 1);
553 	++stat_rush_requests;
554 	if (mp && mp->mnt_syncer_ctx)
555 		wakeup(mp->mnt_syncer_ctx);
556 }
557 
558 /*
559  * trigger a full sync
560  */
561 void
562 trigger_syncer(struct mount *mp)
563 {
564 	struct syncer_ctx *ctx;
565 
566 	if (mp && (ctx = mp->mnt_syncer_ctx) != NULL) {
567 		if (ctx->syncer_trigger == 0) {
568 			ctx->syncer_trigger = 1;
569 			wakeup(ctx);
570 		}
571 	}
572 }
573 
574 /*
575  * Routine to create and manage a filesystem syncer vnode.
576  */
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 *);
582 
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,
590 };
591 
592 static struct vop_ops *sync_vnode_vops_p = &sync_vnode_vops;
593 
594 VNODEOP_SET(sync_vnode_vops);
595 
596 /*
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
599  * the filesystem.
600  *
601  * NOTE: read-only mounts are also placed on the worklist.  The filesystem
602  * sync code is also responsible for cleaning up vnodes.
603  */
604 int
605 vfs_allocate_syncvnode(struct mount *mp)
606 {
607 	struct vnode *vp;
608 	static long start, incr, next;
609 	int error;
610 
611 	/* Allocate a new vnode */
612 	error = getspecialvnode(VT_VFS, mp, &sync_vnode_vops_p, &vp, 0, 0);
613 	if (error) {
614 		mp->mnt_syncer = NULL;
615 		return (error);
616 	}
617 	vp->v_type = VNON;
618 	/*
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.
623 	 */
624 	next += incr;
625 	if (next == 0 || next > SYNCER_MAXDELAY) {
626 		start /= 2;
627 		incr /= 2;
628 		if (start == 0) {
629 			start = SYNCER_MAXDELAY / 2;
630 			incr = SYNCER_MAXDELAY;
631 		}
632 		next = start;
633 	}
634 
635 	/*
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
638 	 * thread.
639 	 */
640 	if (mp->mnt_syncer_ctx)
641 		vn_syncer_add(vp, syncdelay > 0 ? next % syncdelay : 0);
642 
643 	/*
644 	 * The mnt_syncer field inherits the vnode reference, which is
645 	 * held until later decomissioning.
646 	 */
647 	mp->mnt_syncer = vp;
648 	vx_unlock(vp);
649 	return (0);
650 }
651 
652 static int
653 sync_close(struct vop_close_args *ap)
654 {
655 	return (0);
656 }
657 
658 /*
659  * Do a lazy sync of the filesystem.
660  *
661  * sync_fsync { struct vnode *a_vp, int a_waitfor }
662  */
663 static int
664 sync_fsync(struct vop_fsync_args *ap)
665 {
666 	struct vnode *syncvp = ap->a_vp;
667 	struct mount *mp = syncvp->v_mount;
668 	int asyncflag;
669 
670 	/*
671 	 * We only need to do something if this is a lazy evaluation.
672 	 */
673 	if ((ap->a_waitfor & MNT_LAZY) == 0)
674 		return (0);
675 
676 	/*
677 	 * Move ourselves to the back of the sync list.
678 	 */
679 	vn_syncer_add(syncvp, syncdelay);
680 
681 	/*
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
686 	 * or not.
687 	 */
688 	if (vfs_busy(mp, LK_NOWAIT) != 0)
689 		return (0);
690 	if (mp->mnt_flag & MNT_RDONLY) {
691 		vfs_msync(mp, MNT_NOWAIT);
692 	} else {
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);
697 		if (asyncflag)
698 			mp->mnt_flag |= MNT_ASYNC;
699 	}
700 	vfs_unbusy(mp);
701 	return (0);
702 }
703 
704 /*
705  * The syncer vnode is no longer referenced.
706  *
707  * sync_inactive { struct vnode *a_vp, struct proc *a_p }
708  */
709 static int
710 sync_inactive(struct vop_inactive_args *ap)
711 {
712 	vgone_vxlocked(ap->a_vp);
713 	return (0);
714 }
715 
716 /*
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.
720  *
721  * Modifications to the worklist must be protected with a critical
722  * section.
723  *
724  *	sync_reclaim { struct vnode *a_vp }
725  */
726 static int
727 sync_reclaim(struct vop_reclaim_args *ap)
728 {
729 	struct vnode *vp = ap->a_vp;
730 	struct syncer_ctx *ctx;
731 
732 	ctx = vp->v_mount->mnt_syncer_ctx;
733 	if (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);
739 			--ctx->syncer_count;
740 		}
741 		lwkt_reltoken(&ctx->sc_token);
742 	} else {
743 		KKASSERT((vp->v_flag & VONWORKLST) == 0);
744 	}
745 
746 	return (0);
747 }
748 
749 /*
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().
755  *
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.
758  *
759  * This is important when a system has millions of vnodes.
760  */
761 int
762 vsyncscan(
763     struct mount *mp,
764     int vmsc_flags,
765     int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
766     void *data
767 ) {
768 	struct syncer_ctx *ctx;
769 	struct synclist *slp;
770 	struct vnode *vp;
771 	int i;
772 	int count;
773 	int lkflags;
774 
775 	if (vmsc_flags & VMSC_NOWAIT)
776 		lkflags = LK_NOWAIT;
777 	else
778 		lkflags = 0;
779 
780 	/*
781 	 * Syncer list context.  This API requires a dedicated syncer thread.
782 	 * (MNTK_THR_SYNC).
783 	 */
784 	KKASSERT(mp->mnt_kern_flag & MNTK_THR_SYNC);
785 	ctx = mp->mnt_syncer_ctx;
786 	lwkt_gettoken(&ctx->sc_token);
787 
788 	/*
789 	 * Setup for loop.  Allow races against the syncer thread but
790 	 * require that the syncer thread no be lazy if we were told
791 	 * not to be lazy.
792 	 */
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];
798 
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);
804 					vput(vp);
805 				}
806 			} else if (vmsc_flags & VMSC_GETVX) {
807 				vx_get(vp);
808 				slowfunc(mp, vp, data);
809 				vx_put(vp);
810 			} else {
811 				vhold(vp);
812 				slowfunc(mp, vp, data);
813 				vdrop(vp);
814 			}
815 
816 			/*
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.
820 			 */
821 			if (LIST_FIRST(slp) == vp)
822 				vn_syncer_add(vp, -(i + syncdelay));
823 		}
824 		i = (i + 1) & ctx->syncer_mask;
825 	}
826 
827 	if ((vmsc_flags & VMSC_NOWAIT) == 0)
828 		--ctx->syncer_forced;
829 	lwkt_reltoken(&ctx->sc_token);
830 	return(0);
831 }
832 
833 /*
834  * Print out a syncer vnode.
835  *
836  *	sync_print { struct vnode *a_vp }
837  */
838 static int
839 sync_print(struct vop_print_args *ap)
840 {
841 	struct vnode *vp = ap->a_vp;
842 
843 	kprintf("syncer vnode");
844 	lockmgr_printinfo(&vp->v_lock);
845 	kprintf("\n");
846 	return (0);
847 }
848 
849