xref: /freebsd/sys/kern/vfs_subr.c (revision 822923447e454b30d310cb46903c9ddeca9f0a7a)
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  * 4. 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  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ddb.h"
45 #include "opt_mac.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/conf.h>
52 #include <sys/dirent.h>
53 #include <sys/event.h>
54 #include <sys/eventhandler.h>
55 #include <sys/extattr.h>
56 #include <sys/file.h>
57 #include <sys/fcntl.h>
58 #include <sys/kdb.h>
59 #include <sys/kernel.h>
60 #include <sys/kthread.h>
61 #include <sys/mac.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/namei.h>
65 #include <sys/reboot.h>
66 #include <sys/sleepqueue.h>
67 #include <sys/stat.h>
68 #include <sys/sysctl.h>
69 #include <sys/syslog.h>
70 #include <sys/vmmeter.h>
71 #include <sys/vnode.h>
72 
73 #include <machine/stdarg.h>
74 
75 #include <vm/vm.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/pmap.h>
79 #include <vm/vm_map.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_kern.h>
82 #include <vm/uma.h>
83 
84 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
85 
86 static void	delmntque(struct vnode *vp);
87 static void	insmntque(struct vnode *vp, struct mount *mp);
88 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
89 		    int slpflag, int slptimeo);
90 static void	syncer_shutdown(void *arg, int howto);
91 static int	vtryrecycle(struct vnode *vp);
92 static void	vbusy(struct vnode *vp);
93 static void	vdropl(struct vnode *vp);
94 static void	vinactive(struct vnode *, struct thread *);
95 static void	v_incr_usecount(struct vnode *);
96 static void	v_decr_usecount(struct vnode *);
97 static void	v_decr_useonly(struct vnode *);
98 static void	vfree(struct vnode *);
99 static void	vnlru_free(int);
100 static void	vdestroy(struct vnode *);
101 static void	vgonel(struct vnode *);
102 static void	vfs_knllock(void *arg);
103 static void	vfs_knlunlock(void *arg);
104 static int	vfs_knllocked(void *arg);
105 
106 
107 /*
108  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
109  * build.  Without mpsafevm the buffer cache can not run Giant free.
110  */
111 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__) || \
112 	defined(__ia64__)
113 int mpsafe_vfs = 1;
114 #else
115 int mpsafe_vfs;
116 #endif
117 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
118 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
119     "MPSAFE VFS");
120 
121 /*
122  * Number of vnodes in existence.  Increased whenever getnewvnode()
123  * allocates a new vnode, never decreased.
124  */
125 static unsigned long	numvnodes;
126 
127 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
128 
129 /*
130  * Conversion tables for conversion from vnode types to inode formats
131  * and back.
132  */
133 enum vtype iftovt_tab[16] = {
134 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
135 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
136 };
137 int vttoif_tab[9] = {
138 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
139 	S_IFSOCK, S_IFIFO, S_IFMT,
140 };
141 
142 /*
143  * List of vnodes that are ready for recycling.
144  */
145 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
146 
147 /*
148  * Free vnode target.  Free vnodes may simply be files which have been stat'd
149  * but not read.  This is somewhat common, and a small cache of such files
150  * should be kept to avoid recreation costs.
151  */
152 static u_long wantfreevnodes;
153 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
154 /* Number of vnodes in the free list. */
155 static u_long freevnodes;
156 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
157 
158 /*
159  * Various variables used for debugging the new implementation of
160  * reassignbuf().
161  * XXX these are probably of (very) limited utility now.
162  */
163 static int reassignbufcalls;
164 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
165 
166 /*
167  * Cache for the mount type id assigned to NFS.  This is used for
168  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
169  */
170 int	nfs_mount_type = -1;
171 
172 /* To keep more than one thread at a time from running vfs_getnewfsid */
173 static struct mtx mntid_mtx;
174 
175 /*
176  * Lock for any access to the following:
177  *	vnode_free_list
178  *	numvnodes
179  *	freevnodes
180  */
181 static struct mtx vnode_free_list_mtx;
182 
183 /* Publicly exported FS */
184 struct nfs_public nfs_pub;
185 
186 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
187 static uma_zone_t vnode_zone;
188 static uma_zone_t vnodepoll_zone;
189 
190 /* Set to 1 to print out reclaim of active vnodes */
191 int	prtactive;
192 
193 /*
194  * The workitem queue.
195  *
196  * It is useful to delay writes of file data and filesystem metadata
197  * for tens of seconds so that quickly created and deleted files need
198  * not waste disk bandwidth being created and removed. To realize this,
199  * we append vnodes to a "workitem" queue. When running with a soft
200  * updates implementation, most pending metadata dependencies should
201  * not wait for more than a few seconds. Thus, mounted on block devices
202  * are delayed only about a half the time that file data is delayed.
203  * Similarly, directory updates are more critical, so are only delayed
204  * about a third the time that file data is delayed. Thus, there are
205  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
206  * one each second (driven off the filesystem syncer process). The
207  * syncer_delayno variable indicates the next queue that is to be processed.
208  * Items that need to be processed soon are placed in this queue:
209  *
210  *	syncer_workitem_pending[syncer_delayno]
211  *
212  * A delay of fifteen seconds is done by placing the request fifteen
213  * entries later in the queue:
214  *
215  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
216  *
217  */
218 static int syncer_delayno;
219 static long syncer_mask;
220 LIST_HEAD(synclist, bufobj);
221 static struct synclist *syncer_workitem_pending;
222 /*
223  * The sync_mtx protects:
224  *	bo->bo_synclist
225  *	sync_vnode_count
226  *	syncer_delayno
227  *	syncer_state
228  *	syncer_workitem_pending
229  *	syncer_worklist_len
230  *	rushjob
231  */
232 static struct mtx sync_mtx;
233 
234 #define SYNCER_MAXDELAY		32
235 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
236 static int syncdelay = 30;		/* max time to delay syncing data */
237 static int filedelay = 30;		/* time to delay syncing files */
238 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
239 static int dirdelay = 29;		/* time to delay syncing directories */
240 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
241 static int metadelay = 28;		/* time to delay syncing metadata */
242 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
243 static int rushjob;		/* number of slots to run ASAP */
244 static int stat_rush_requests;	/* number of times I/O speeded up */
245 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
246 
247 /*
248  * When shutting down the syncer, run it at four times normal speed.
249  */
250 #define SYNCER_SHUTDOWN_SPEEDUP		4
251 static int sync_vnode_count;
252 static int syncer_worklist_len;
253 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
254     syncer_state;
255 
256 /*
257  * Number of vnodes we want to exist at any one time.  This is mostly used
258  * to size hash tables in vnode-related code.  It is normally not used in
259  * getnewvnode(), as wantfreevnodes is normally nonzero.)
260  *
261  * XXX desiredvnodes is historical cruft and should not exist.
262  */
263 int desiredvnodes;
264 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
265     &desiredvnodes, 0, "Maximum number of vnodes");
266 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
267     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
268 static int vnlru_nowhere;
269 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
270     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
271 
272 /* Hook for calling soft updates. */
273 int (*softdep_process_worklist_hook)(struct mount *);
274 
275 /*
276  * Macros to control when a vnode is freed and recycled.  All require
277  * the vnode interlock.
278  */
279 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
280 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
281 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
282 
283 
284 /*
285  * Initialize the vnode management data structures.
286  */
287 #ifndef	MAXVNODES_MAX
288 #define	MAXVNODES_MAX	100000
289 #endif
290 static void
291 vntblinit(void *dummy __unused)
292 {
293 
294 	/*
295 	 * Desiredvnodes is a function of the physical memory size and
296 	 * the kernel's heap size.  Specifically, desiredvnodes scales
297 	 * in proportion to the physical memory size until two fifths
298 	 * of the kernel's heap size is consumed by vnodes and vm
299 	 * objects.
300 	 */
301 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
302 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
303 	if (desiredvnodes > MAXVNODES_MAX) {
304 		if (bootverbose)
305 			printf("Reducing kern.maxvnodes %d -> %d\n",
306 			    desiredvnodes, MAXVNODES_MAX);
307 		desiredvnodes = MAXVNODES_MAX;
308 	}
309 	wantfreevnodes = desiredvnodes / 4;
310 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
311 	TAILQ_INIT(&vnode_free_list);
312 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
313 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
314 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
316 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
317 	/*
318 	 * Initialize the filesystem syncer.
319 	 */
320 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
321 		&syncer_mask);
322 	syncer_maxdelay = syncer_mask + 1;
323 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
324 }
325 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
326 
327 
328 /*
329  * Mark a mount point as busy. Used to synchronize access and to delay
330  * unmounting. Interlock is not released on failure.
331  */
332 int
333 vfs_busy(mp, flags, interlkp, td)
334 	struct mount *mp;
335 	int flags;
336 	struct mtx *interlkp;
337 	struct thread *td;
338 {
339 	int lkflags;
340 
341 	MNT_ILOCK(mp);
342 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
343 		if (flags & LK_NOWAIT) {
344 			MNT_IUNLOCK(mp);
345 			return (ENOENT);
346 		}
347 		if (interlkp)
348 			mtx_unlock(interlkp);
349 		mp->mnt_kern_flag |= MNTK_MWAIT;
350 		/*
351 		 * Since all busy locks are shared except the exclusive
352 		 * lock granted when unmounting, the only place that a
353 		 * wakeup needs to be done is at the release of the
354 		 * exclusive lock at the end of dounmount.
355 		 */
356 		msleep(mp, MNT_MTX(mp), PVFS|PDROP, "vfs_busy", 0);
357 		if (interlkp)
358 			mtx_lock(interlkp);
359 		return (ENOENT);
360 	}
361 	if (interlkp)
362 		mtx_unlock(interlkp);
363 	lkflags = LK_SHARED | LK_INTERLOCK;
364 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
365 		panic("vfs_busy: unexpected lock failure");
366 	return (0);
367 }
368 
369 /*
370  * Free a busy filesystem.
371  */
372 void
373 vfs_unbusy(mp, td)
374 	struct mount *mp;
375 	struct thread *td;
376 {
377 
378 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
379 }
380 
381 /*
382  * Lookup a mount point by filesystem identifier.
383  */
384 struct mount *
385 vfs_getvfs(fsid)
386 	fsid_t *fsid;
387 {
388 	struct mount *mp;
389 
390 	mtx_lock(&mountlist_mtx);
391 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
392 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
393 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
394 			mtx_unlock(&mountlist_mtx);
395 			return (mp);
396 		}
397 	}
398 	mtx_unlock(&mountlist_mtx);
399 	return ((struct mount *) 0);
400 }
401 
402 /*
403  * Check if a user can access priveledged mount options.
404  */
405 int
406 vfs_suser(struct mount *mp, struct thread *td)
407 {
408 	int error;
409 
410 	if ((mp->mnt_flag & MNT_USER) == 0 ||
411 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
412 		if ((error = suser(td)) != 0)
413 			return (error);
414 	}
415 	return (0);
416 }
417 
418 /*
419  * Get a new unique fsid.  Try to make its val[0] unique, since this value
420  * will be used to create fake device numbers for stat().  Also try (but
421  * not so hard) make its val[0] unique mod 2^16, since some emulators only
422  * support 16-bit device numbers.  We end up with unique val[0]'s for the
423  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
424  *
425  * Keep in mind that several mounts may be running in parallel.  Starting
426  * the search one past where the previous search terminated is both a
427  * micro-optimization and a defense against returning the same fsid to
428  * different mounts.
429  */
430 void
431 vfs_getnewfsid(mp)
432 	struct mount *mp;
433 {
434 	static u_int16_t mntid_base;
435 	fsid_t tfsid;
436 	int mtype;
437 
438 	mtx_lock(&mntid_mtx);
439 	mtype = mp->mnt_vfc->vfc_typenum;
440 	tfsid.val[1] = mtype;
441 	mtype = (mtype & 0xFF) << 24;
442 	for (;;) {
443 		tfsid.val[0] = makedev(255,
444 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
445 		mntid_base++;
446 		if (vfs_getvfs(&tfsid) == NULL)
447 			break;
448 	}
449 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
450 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
451 	mtx_unlock(&mntid_mtx);
452 }
453 
454 /*
455  * Knob to control the precision of file timestamps:
456  *
457  *   0 = seconds only; nanoseconds zeroed.
458  *   1 = seconds and nanoseconds, accurate within 1/HZ.
459  *   2 = seconds and nanoseconds, truncated to microseconds.
460  * >=3 = seconds and nanoseconds, maximum precision.
461  */
462 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
463 
464 static int timestamp_precision = TSP_SEC;
465 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
466     &timestamp_precision, 0, "");
467 
468 /*
469  * Get a current timestamp.
470  */
471 void
472 vfs_timestamp(tsp)
473 	struct timespec *tsp;
474 {
475 	struct timeval tv;
476 
477 	switch (timestamp_precision) {
478 	case TSP_SEC:
479 		tsp->tv_sec = time_second;
480 		tsp->tv_nsec = 0;
481 		break;
482 	case TSP_HZ:
483 		getnanotime(tsp);
484 		break;
485 	case TSP_USEC:
486 		microtime(&tv);
487 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
488 		break;
489 	case TSP_NSEC:
490 	default:
491 		nanotime(tsp);
492 		break;
493 	}
494 }
495 
496 /*
497  * Set vnode attributes to VNOVAL
498  */
499 void
500 vattr_null(vap)
501 	struct vattr *vap;
502 {
503 
504 	vap->va_type = VNON;
505 	vap->va_size = VNOVAL;
506 	vap->va_bytes = VNOVAL;
507 	vap->va_mode = VNOVAL;
508 	vap->va_nlink = VNOVAL;
509 	vap->va_uid = VNOVAL;
510 	vap->va_gid = VNOVAL;
511 	vap->va_fsid = VNOVAL;
512 	vap->va_fileid = VNOVAL;
513 	vap->va_blocksize = VNOVAL;
514 	vap->va_rdev = VNOVAL;
515 	vap->va_atime.tv_sec = VNOVAL;
516 	vap->va_atime.tv_nsec = VNOVAL;
517 	vap->va_mtime.tv_sec = VNOVAL;
518 	vap->va_mtime.tv_nsec = VNOVAL;
519 	vap->va_ctime.tv_sec = VNOVAL;
520 	vap->va_ctime.tv_nsec = VNOVAL;
521 	vap->va_birthtime.tv_sec = VNOVAL;
522 	vap->va_birthtime.tv_nsec = VNOVAL;
523 	vap->va_flags = VNOVAL;
524 	vap->va_gen = VNOVAL;
525 	vap->va_vaflags = 0;
526 }
527 
528 /*
529  * This routine is called when we have too many vnodes.  It attempts
530  * to free <count> vnodes and will potentially free vnodes that still
531  * have VM backing store (VM backing store is typically the cause
532  * of a vnode blowout so we want to do this).  Therefore, this operation
533  * is not considered cheap.
534  *
535  * A number of conditions may prevent a vnode from being reclaimed.
536  * the buffer cache may have references on the vnode, a directory
537  * vnode may still have references due to the namei cache representing
538  * underlying files, or the vnode may be in active use.   It is not
539  * desireable to reuse such vnodes.  These conditions may cause the
540  * number of vnodes to reach some minimum value regardless of what
541  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
542  */
543 static int
544 vlrureclaim(struct mount *mp)
545 {
546 	struct thread *td;
547 	struct vnode *vp;
548 	int done;
549 	int trigger;
550 	int usevnodes;
551 	int count;
552 
553 	/*
554 	 * Calculate the trigger point, don't allow user
555 	 * screwups to blow us up.   This prevents us from
556 	 * recycling vnodes with lots of resident pages.  We
557 	 * aren't trying to free memory, we are trying to
558 	 * free vnodes.
559 	 */
560 	usevnodes = desiredvnodes;
561 	if (usevnodes <= 0)
562 		usevnodes = 1;
563 	trigger = cnt.v_page_count * 2 / usevnodes;
564 	done = 0;
565 	td = curthread;
566 	vn_start_write(NULL, &mp, V_WAIT);
567 	MNT_ILOCK(mp);
568 	count = mp->mnt_nvnodelistsize / 10 + 1;
569 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
570 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
571 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
572 		--count;
573 		if (!VI_TRYLOCK(vp))
574 			goto next_iter;
575 		/*
576 		 * If it's been deconstructed already, it's still
577 		 * referenced, or it exceeds the trigger, skip it.
578 		 */
579 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
580 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
581 		    vp->v_object->resident_page_count > trigger)) {
582 			VI_UNLOCK(vp);
583 			goto next_iter;
584 		}
585 		MNT_IUNLOCK(mp);
586 		vholdl(vp);
587 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
588 			vdrop(vp);
589 			goto next_iter_mntunlocked;
590 		}
591 		VI_LOCK(vp);
592 		/*
593 		 * v_usecount may have been bumped after VOP_LOCK() dropped
594 		 * the vnode interlock and before it was locked again.
595 		 *
596 		 * It is not necessary to recheck VI_DOOMED because it can
597 		 * only be set by another thread that holds both the vnode
598 		 * lock and vnode interlock.  If another thread has the
599 		 * vnode lock before we get to VOP_LOCK() and obtains the
600 		 * vnode interlock after VOP_LOCK() drops the vnode
601 		 * interlock, the other thread will be unable to drop the
602 		 * vnode lock before our VOP_LOCK() call fails.
603 		 */
604 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
605 		    (vp->v_object != NULL &&
606 		    vp->v_object->resident_page_count > trigger)) {
607 			VOP_UNLOCK(vp, LK_INTERLOCK, td);
608 			goto next_iter_mntunlocked;
609 		}
610 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
611 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
612 		vgonel(vp);
613 		VOP_UNLOCK(vp, 0, td);
614 		vdropl(vp);
615 		done++;
616 next_iter_mntunlocked:
617 		if ((count % 256) != 0)
618 			goto relock_mnt;
619 		goto yield;
620 next_iter:
621 		if ((count % 256) != 0)
622 			continue;
623 		MNT_IUNLOCK(mp);
624 yield:
625 		uio_yield();
626 relock_mnt:
627 		MNT_ILOCK(mp);
628 	}
629 	MNT_IUNLOCK(mp);
630 	vn_finished_write(mp);
631 	return done;
632 }
633 
634 /*
635  * Attempt to keep the free list at wantfreevnodes length.
636  */
637 static void
638 vnlru_free(int count)
639 {
640 	struct vnode *vp;
641 
642 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
643 	for (; count > 0; count--) {
644 		vp = TAILQ_FIRST(&vnode_free_list);
645 		/*
646 		 * The list can be modified while the free_list_mtx
647 		 * has been dropped and vp could be NULL here.
648 		 */
649 		if (!vp)
650 			break;
651 		VNASSERT(vp->v_op != NULL, vp,
652 		    ("vnlru_free: vnode already reclaimed."));
653 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
654 		/*
655 		 * Don't recycle if we can't get the interlock.
656 		 */
657 		if (!VI_TRYLOCK(vp)) {
658 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
659 			continue;
660 		}
661 		VNASSERT(VCANRECYCLE(vp), vp,
662 		    ("vp inconsistent on freelist"));
663 		freevnodes--;
664 		vp->v_iflag &= ~VI_FREE;
665 		vholdl(vp);
666 		mtx_unlock(&vnode_free_list_mtx);
667 		VI_UNLOCK(vp);
668 		vtryrecycle(vp);
669 		/*
670 		 * If the recycled succeeded this vdrop will actually free
671 		 * the vnode.  If not it will simply place it back on
672 		 * the free list.
673 		 */
674 		vdrop(vp);
675 		mtx_lock(&vnode_free_list_mtx);
676 	}
677 }
678 /*
679  * Attempt to recycle vnodes in a context that is always safe to block.
680  * Calling vlrurecycle() from the bowels of filesystem code has some
681  * interesting deadlock problems.
682  */
683 static struct proc *vnlruproc;
684 static int vnlruproc_sig;
685 
686 static void
687 vnlru_proc(void)
688 {
689 	struct mount *mp, *nmp;
690 	int done;
691 	struct proc *p = vnlruproc;
692 	struct thread *td = FIRST_THREAD_IN_PROC(p);
693 
694 	mtx_lock(&Giant);
695 
696 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
697 	    SHUTDOWN_PRI_FIRST);
698 
699 	for (;;) {
700 		kthread_suspend_check(p);
701 		mtx_lock(&vnode_free_list_mtx);
702 		if (freevnodes > wantfreevnodes)
703 			vnlru_free(freevnodes - wantfreevnodes);
704 		if (numvnodes <= desiredvnodes * 9 / 10) {
705 			vnlruproc_sig = 0;
706 			wakeup(&vnlruproc_sig);
707 			msleep(vnlruproc, &vnode_free_list_mtx,
708 			    PVFS|PDROP, "vlruwt", hz);
709 			continue;
710 		}
711 		mtx_unlock(&vnode_free_list_mtx);
712 		done = 0;
713 		mtx_lock(&mountlist_mtx);
714 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
715 			int vfsunlocked;
716 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
717 				nmp = TAILQ_NEXT(mp, mnt_list);
718 				continue;
719 			}
720 			if (!VFS_NEEDSGIANT(mp)) {
721 				mtx_unlock(&Giant);
722 				vfsunlocked = 1;
723 			} else
724 				vfsunlocked = 0;
725 			done += vlrureclaim(mp);
726 			if (vfsunlocked)
727 				mtx_lock(&Giant);
728 			mtx_lock(&mountlist_mtx);
729 			nmp = TAILQ_NEXT(mp, mnt_list);
730 			vfs_unbusy(mp, td);
731 		}
732 		mtx_unlock(&mountlist_mtx);
733 		if (done == 0) {
734 #if 0
735 			/* These messages are temporary debugging aids */
736 			if (vnlru_nowhere < 5)
737 				printf("vnlru process getting nowhere..\n");
738 			else if (vnlru_nowhere == 5)
739 				printf("vnlru process messages stopped.\n");
740 #endif
741 			vnlru_nowhere++;
742 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
743 		} else
744 			uio_yield();
745 	}
746 }
747 
748 static struct kproc_desc vnlru_kp = {
749 	"vnlru",
750 	vnlru_proc,
751 	&vnlruproc
752 };
753 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
754 
755 /*
756  * Routines having to do with the management of the vnode table.
757  */
758 
759 static void
760 vdestroy(struct vnode *vp)
761 {
762 	struct bufobj *bo;
763 
764 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
765 	mtx_lock(&vnode_free_list_mtx);
766 	numvnodes--;
767 	mtx_unlock(&vnode_free_list_mtx);
768 	bo = &vp->v_bufobj;
769 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
770 	    ("cleaned vnode still on the free list."));
771 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
772 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
773 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
774 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
775 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
776 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
777 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
778 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
779 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
780 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
781 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
782 	VI_UNLOCK(vp);
783 #ifdef MAC
784 	mac_destroy_vnode(vp);
785 #endif
786 	if (vp->v_pollinfo != NULL) {
787 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
788 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
789 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
790 	}
791 #ifdef INVARIANTS
792 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
793 	vp->v_op = NULL;
794 #endif
795 	lockdestroy(vp->v_vnlock);
796 	mtx_destroy(&vp->v_interlock);
797 	uma_zfree(vnode_zone, vp);
798 }
799 
800 /*
801  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
802  * before we actually vgone().  This function must be called with the vnode
803  * held to prevent the vnode from being returned to the free list midway
804  * through vgone().
805  */
806 static int
807 vtryrecycle(struct vnode *vp)
808 {
809 	struct thread *td = curthread;
810 	struct mount *vnmp;
811 
812 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
813 	VNASSERT(vp->v_holdcnt, vp,
814 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
815 	/*
816 	 * This vnode may found and locked via some other list, if so we
817 	 * can't recycle it yet.
818 	 */
819 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
820 		return (EWOULDBLOCK);
821 	/*
822 	 * Don't recycle if its filesystem is being suspended.
823 	 */
824 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
825 		VOP_UNLOCK(vp, 0, td);
826 		return (EBUSY);
827 	}
828 	/*
829 	 * If we got this far, we need to acquire the interlock and see if
830 	 * anyone picked up this vnode from another list.  If not, we will
831 	 * mark it with DOOMED via vgonel() so that anyone who does find it
832 	 * will skip over it.
833 	 */
834 	VI_LOCK(vp);
835 	if (vp->v_usecount) {
836 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
837 		vn_finished_write(vnmp);
838 		return (EBUSY);
839 	}
840 	if ((vp->v_iflag & VI_DOOMED) == 0)
841 		vgonel(vp);
842 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
843 	vn_finished_write(vnmp);
844 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
845 	return (0);
846 }
847 
848 /*
849  * Return the next vnode from the free list.
850  */
851 int
852 getnewvnode(tag, mp, vops, vpp)
853 	const char *tag;
854 	struct mount *mp;
855 	struct vop_vector *vops;
856 	struct vnode **vpp;
857 {
858 	struct vnode *vp = NULL;
859 	struct bufobj *bo;
860 
861 	mtx_lock(&vnode_free_list_mtx);
862 	/*
863 	 * Lend our context to reclaim vnodes if they've exceeded the max.
864 	 */
865 	if (freevnodes > wantfreevnodes)
866 		vnlru_free(1);
867 	/*
868 	 * Wait for available vnodes.
869 	 */
870 	if (numvnodes > desiredvnodes) {
871 		if (vnlruproc_sig == 0) {
872 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
873 			wakeup(vnlruproc);
874 		}
875 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
876 		    "vlruwk", hz);
877 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
878 		if (numvnodes > desiredvnodes) {
879 			mtx_unlock(&vnode_free_list_mtx);
880 			return (ENFILE);
881 		}
882 #endif
883 	}
884 	numvnodes++;
885 	mtx_unlock(&vnode_free_list_mtx);
886 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
887 	/*
888 	 * Setup locks.
889 	 */
890 	vp->v_vnlock = &vp->v_lock;
891 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
892 	/*
893 	 * By default, don't allow shared locks unless filesystems
894 	 * opt-in.
895 	 */
896 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
897 	/*
898 	 * Initialize bufobj.
899 	 */
900 	bo = &vp->v_bufobj;
901 	bo->__bo_vnode = vp;
902 	bo->bo_mtx = &vp->v_interlock;
903 	bo->bo_ops = &buf_ops_bio;
904 	bo->bo_private = vp;
905 	TAILQ_INIT(&bo->bo_clean.bv_hd);
906 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
907 	/*
908 	 * Initialize namecache.
909 	 */
910 	LIST_INIT(&vp->v_cache_src);
911 	TAILQ_INIT(&vp->v_cache_dst);
912 	/*
913 	 * Finalize various vnode identity bits.
914 	 */
915 	vp->v_type = VNON;
916 	vp->v_tag = tag;
917 	vp->v_op = vops;
918 	v_incr_usecount(vp);
919 	vp->v_data = 0;
920 #ifdef MAC
921 	mac_init_vnode(vp);
922 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
923 		mac_associate_vnode_singlelabel(mp, vp);
924 	else if (mp == NULL)
925 		printf("NULL mp in getnewvnode()\n");
926 #endif
927 	delmntque(vp);
928 	if (mp != NULL) {
929 		insmntque(vp, mp);
930 		bo->bo_bsize = mp->mnt_stat.f_iosize;
931 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
932 			vp->v_vflag |= VV_NOKNOTE;
933 	}
934 
935 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
936 	*vpp = vp;
937 	return (0);
938 }
939 
940 /*
941  * Delete from old mount point vnode list, if on one.
942  */
943 static void
944 delmntque(struct vnode *vp)
945 {
946 	struct mount *mp;
947 
948 	if (vp->v_mount == NULL)
949 		return;
950 	mp = vp->v_mount;
951 	MNT_ILOCK(mp);
952 	vp->v_mount = NULL;
953 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
954 		("bad mount point vnode list size"));
955 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
956 	mp->mnt_nvnodelistsize--;
957 	MNT_IUNLOCK(mp);
958 }
959 
960 /*
961  * Insert into list of vnodes for the new mount point, if available.
962  */
963 static void
964 insmntque(struct vnode *vp, struct mount *mp)
965 {
966 
967 	vp->v_mount = mp;
968 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
969 	MNT_ILOCK(vp->v_mount);
970 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
971 	mp->mnt_nvnodelistsize++;
972 	MNT_IUNLOCK(vp->v_mount);
973 }
974 
975 /*
976  * Flush out and invalidate all buffers associated with a bufobj
977  * Called with the underlying object locked.
978  */
979 int
980 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, int slptimeo)
981 {
982 	int error;
983 
984 	BO_LOCK(bo);
985 	if (flags & V_SAVE) {
986 		error = bufobj_wwait(bo, slpflag, slptimeo);
987 		if (error) {
988 			BO_UNLOCK(bo);
989 			return (error);
990 		}
991 		if (bo->bo_dirty.bv_cnt > 0) {
992 			BO_UNLOCK(bo);
993 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
994 				return (error);
995 			/*
996 			 * XXX We could save a lock/unlock if this was only
997 			 * enabled under INVARIANTS
998 			 */
999 			BO_LOCK(bo);
1000 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1001 				panic("vinvalbuf: dirty bufs");
1002 		}
1003 	}
1004 	/*
1005 	 * If you alter this loop please notice that interlock is dropped and
1006 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1007 	 * no race conditions occur from this.
1008 	 */
1009 	do {
1010 		error = flushbuflist(&bo->bo_clean,
1011 		    flags, bo, slpflag, slptimeo);
1012 		if (error == 0)
1013 			error = flushbuflist(&bo->bo_dirty,
1014 			    flags, bo, slpflag, slptimeo);
1015 		if (error != 0 && error != EAGAIN) {
1016 			BO_UNLOCK(bo);
1017 			return (error);
1018 		}
1019 	} while (error != 0);
1020 
1021 	/*
1022 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1023 	 * have write I/O in-progress but if there is a VM object then the
1024 	 * VM object can also have read-I/O in-progress.
1025 	 */
1026 	do {
1027 		bufobj_wwait(bo, 0, 0);
1028 		BO_UNLOCK(bo);
1029 		if (bo->bo_object != NULL) {
1030 			VM_OBJECT_LOCK(bo->bo_object);
1031 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1032 			VM_OBJECT_UNLOCK(bo->bo_object);
1033 		}
1034 		BO_LOCK(bo);
1035 	} while (bo->bo_numoutput > 0);
1036 	BO_UNLOCK(bo);
1037 
1038 	/*
1039 	 * Destroy the copy in the VM cache, too.
1040 	 */
1041 	if (bo->bo_object != NULL) {
1042 		VM_OBJECT_LOCK(bo->bo_object);
1043 		vm_object_page_remove(bo->bo_object, 0, 0,
1044 			(flags & V_SAVE) ? TRUE : FALSE);
1045 		VM_OBJECT_UNLOCK(bo->bo_object);
1046 	}
1047 
1048 #ifdef INVARIANTS
1049 	BO_LOCK(bo);
1050 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1051 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1052 		panic("vinvalbuf: flush failed");
1053 	BO_UNLOCK(bo);
1054 #endif
1055 	return (0);
1056 }
1057 
1058 /*
1059  * Flush out and invalidate all buffers associated with a vnode.
1060  * Called with the underlying object locked.
1061  */
1062 int
1063 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, int slptimeo)
1064 {
1065 
1066 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1067 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1068 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1069 }
1070 
1071 /*
1072  * Flush out buffers on the specified list.
1073  *
1074  */
1075 static int
1076 flushbuflist(bufv, flags, bo, slpflag, slptimeo)
1077 	struct bufv *bufv;
1078 	int flags;
1079 	struct bufobj *bo;
1080 	int slpflag, slptimeo;
1081 {
1082 	struct buf *bp, *nbp;
1083 	int retval, error;
1084 	daddr_t lblkno;
1085 	b_xflags_t xflags;
1086 
1087 	ASSERT_BO_LOCKED(bo);
1088 
1089 	retval = 0;
1090 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1091 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1092 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1093 			continue;
1094 		}
1095 		lblkno = 0;
1096 		xflags = 0;
1097 		if (nbp != NULL) {
1098 			lblkno = nbp->b_lblkno;
1099 			xflags = nbp->b_xflags &
1100 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1101 		}
1102 		retval = EAGAIN;
1103 		error = BUF_TIMELOCK(bp,
1104 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1105 		    "flushbuf", slpflag, slptimeo);
1106 		if (error) {
1107 			BO_LOCK(bo);
1108 			return (error != ENOLCK ? error : EAGAIN);
1109 		}
1110 		KASSERT(bp->b_bufobj == bo,
1111 	            ("bp %p wrong b_bufobj %p should be %p",
1112 		    bp, bp->b_bufobj, bo));
1113 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1114 			BUF_UNLOCK(bp);
1115 			BO_LOCK(bo);
1116 			return (EAGAIN);
1117 		}
1118 		/*
1119 		 * XXX Since there are no node locks for NFS, I
1120 		 * believe there is a slight chance that a delayed
1121 		 * write will occur while sleeping just above, so
1122 		 * check for it.
1123 		 */
1124 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1125 		    (flags & V_SAVE)) {
1126 			bremfree(bp);
1127 			bp->b_flags |= B_ASYNC;
1128 			bwrite(bp);
1129 			BO_LOCK(bo);
1130 			return (EAGAIN);	/* XXX: why not loop ? */
1131 		}
1132 		bremfree(bp);
1133 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1134 		bp->b_flags &= ~B_ASYNC;
1135 		brelse(bp);
1136 		BO_LOCK(bo);
1137 		if (nbp != NULL &&
1138 		    (nbp->b_bufobj != bo ||
1139 		     nbp->b_lblkno != lblkno ||
1140 		     (nbp->b_xflags &
1141 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1142 			break;			/* nbp invalid */
1143 	}
1144 	return (retval);
1145 }
1146 
1147 /*
1148  * Truncate a file's buffer and pages to a specified length.  This
1149  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1150  * sync activity.
1151  */
1152 int
1153 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize)
1154 {
1155 	struct buf *bp, *nbp;
1156 	int anyfreed;
1157 	int trunclbn;
1158 	struct bufobj *bo;
1159 
1160 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1161 	/*
1162 	 * Round up to the *next* lbn.
1163 	 */
1164 	trunclbn = (length + blksize - 1) / blksize;
1165 
1166 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1167 restart:
1168 	VI_LOCK(vp);
1169 	bo = &vp->v_bufobj;
1170 	anyfreed = 1;
1171 	for (;anyfreed;) {
1172 		anyfreed = 0;
1173 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1174 			if (bp->b_lblkno < trunclbn)
1175 				continue;
1176 			if (BUF_LOCK(bp,
1177 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1178 			    VI_MTX(vp)) == ENOLCK)
1179 				goto restart;
1180 
1181 			bremfree(bp);
1182 			bp->b_flags |= (B_INVAL | B_RELBUF);
1183 			bp->b_flags &= ~B_ASYNC;
1184 			brelse(bp);
1185 			anyfreed = 1;
1186 
1187 			if (nbp != NULL &&
1188 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1189 			    (nbp->b_vp != vp) ||
1190 			    (nbp->b_flags & B_DELWRI))) {
1191 				goto restart;
1192 			}
1193 			VI_LOCK(vp);
1194 		}
1195 
1196 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1197 			if (bp->b_lblkno < trunclbn)
1198 				continue;
1199 			if (BUF_LOCK(bp,
1200 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1201 			    VI_MTX(vp)) == ENOLCK)
1202 				goto restart;
1203 			bremfree(bp);
1204 			bp->b_flags |= (B_INVAL | B_RELBUF);
1205 			bp->b_flags &= ~B_ASYNC;
1206 			brelse(bp);
1207 			anyfreed = 1;
1208 			if (nbp != NULL &&
1209 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1210 			    (nbp->b_vp != vp) ||
1211 			    (nbp->b_flags & B_DELWRI) == 0)) {
1212 				goto restart;
1213 			}
1214 			VI_LOCK(vp);
1215 		}
1216 	}
1217 
1218 	if (length > 0) {
1219 restartsync:
1220 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1221 			if (bp->b_lblkno > 0)
1222 				continue;
1223 			/*
1224 			 * Since we hold the vnode lock this should only
1225 			 * fail if we're racing with the buf daemon.
1226 			 */
1227 			if (BUF_LOCK(bp,
1228 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1229 			    VI_MTX(vp)) == ENOLCK) {
1230 				goto restart;
1231 			}
1232 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1233 			    ("buf(%p) on dirty queue without DELWRI", bp));
1234 
1235 			bremfree(bp);
1236 			bawrite(bp);
1237 			VI_LOCK(vp);
1238 			goto restartsync;
1239 		}
1240 	}
1241 
1242 	bufobj_wwait(bo, 0, 0);
1243 	VI_UNLOCK(vp);
1244 	vnode_pager_setsize(vp, length);
1245 
1246 	return (0);
1247 }
1248 
1249 /*
1250  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1251  * 		 a vnode.
1252  *
1253  *	NOTE: We have to deal with the special case of a background bitmap
1254  *	buffer, a situation where two buffers will have the same logical
1255  *	block offset.  We want (1) only the foreground buffer to be accessed
1256  *	in a lookup and (2) must differentiate between the foreground and
1257  *	background buffer in the splay tree algorithm because the splay
1258  *	tree cannot normally handle multiple entities with the same 'index'.
1259  *	We accomplish this by adding differentiating flags to the splay tree's
1260  *	numerical domain.
1261  */
1262 static
1263 struct buf *
1264 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1265 {
1266 	struct buf dummy;
1267 	struct buf *lefttreemax, *righttreemin, *y;
1268 
1269 	if (root == NULL)
1270 		return (NULL);
1271 	lefttreemax = righttreemin = &dummy;
1272 	for (;;) {
1273 		if (lblkno < root->b_lblkno ||
1274 		    (lblkno == root->b_lblkno &&
1275 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1276 			if ((y = root->b_left) == NULL)
1277 				break;
1278 			if (lblkno < y->b_lblkno) {
1279 				/* Rotate right. */
1280 				root->b_left = y->b_right;
1281 				y->b_right = root;
1282 				root = y;
1283 				if ((y = root->b_left) == NULL)
1284 					break;
1285 			}
1286 			/* Link into the new root's right tree. */
1287 			righttreemin->b_left = root;
1288 			righttreemin = root;
1289 		} else if (lblkno > root->b_lblkno ||
1290 		    (lblkno == root->b_lblkno &&
1291 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1292 			if ((y = root->b_right) == NULL)
1293 				break;
1294 			if (lblkno > y->b_lblkno) {
1295 				/* Rotate left. */
1296 				root->b_right = y->b_left;
1297 				y->b_left = root;
1298 				root = y;
1299 				if ((y = root->b_right) == NULL)
1300 					break;
1301 			}
1302 			/* Link into the new root's left tree. */
1303 			lefttreemax->b_right = root;
1304 			lefttreemax = root;
1305 		} else {
1306 			break;
1307 		}
1308 		root = y;
1309 	}
1310 	/* Assemble the new root. */
1311 	lefttreemax->b_right = root->b_left;
1312 	righttreemin->b_left = root->b_right;
1313 	root->b_left = dummy.b_right;
1314 	root->b_right = dummy.b_left;
1315 	return (root);
1316 }
1317 
1318 static void
1319 buf_vlist_remove(struct buf *bp)
1320 {
1321 	struct buf *root;
1322 	struct bufv *bv;
1323 
1324 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1325 	ASSERT_BO_LOCKED(bp->b_bufobj);
1326 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1327 	    (BX_VNDIRTY|BX_VNCLEAN),
1328 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1329 	if (bp->b_xflags & BX_VNDIRTY)
1330 		bv = &bp->b_bufobj->bo_dirty;
1331 	else
1332 		bv = &bp->b_bufobj->bo_clean;
1333 	if (bp != bv->bv_root) {
1334 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1335 		KASSERT(root == bp, ("splay lookup failed in remove"));
1336 	}
1337 	if (bp->b_left == NULL) {
1338 		root = bp->b_right;
1339 	} else {
1340 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1341 		root->b_right = bp->b_right;
1342 	}
1343 	bv->bv_root = root;
1344 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1345 	bv->bv_cnt--;
1346 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1347 }
1348 
1349 /*
1350  * Add the buffer to the sorted clean or dirty block list using a
1351  * splay tree algorithm.
1352  *
1353  * NOTE: xflags is passed as a constant, optimizing this inline function!
1354  */
1355 static void
1356 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1357 {
1358 	struct buf *root;
1359 	struct bufv *bv;
1360 
1361 	ASSERT_BO_LOCKED(bo);
1362 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1363 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1364 	bp->b_xflags |= xflags;
1365 	if (xflags & BX_VNDIRTY)
1366 		bv = &bo->bo_dirty;
1367 	else
1368 		bv = &bo->bo_clean;
1369 
1370 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1371 	if (root == NULL) {
1372 		bp->b_left = NULL;
1373 		bp->b_right = NULL;
1374 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1375 	} else if (bp->b_lblkno < root->b_lblkno ||
1376 	    (bp->b_lblkno == root->b_lblkno &&
1377 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1378 		bp->b_left = root->b_left;
1379 		bp->b_right = root;
1380 		root->b_left = NULL;
1381 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1382 	} else {
1383 		bp->b_right = root->b_right;
1384 		bp->b_left = root;
1385 		root->b_right = NULL;
1386 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1387 	}
1388 	bv->bv_cnt++;
1389 	bv->bv_root = bp;
1390 }
1391 
1392 /*
1393  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1394  * shadow buffers used in background bitmap writes.
1395  *
1396  * This code isn't quite efficient as it could be because we are maintaining
1397  * two sorted lists and do not know which list the block resides in.
1398  *
1399  * During a "make buildworld" the desired buffer is found at one of
1400  * the roots more than 60% of the time.  Thus, checking both roots
1401  * before performing either splay eliminates unnecessary splays on the
1402  * first tree splayed.
1403  */
1404 struct buf *
1405 gbincore(struct bufobj *bo, daddr_t lblkno)
1406 {
1407 	struct buf *bp;
1408 
1409 	ASSERT_BO_LOCKED(bo);
1410 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1411 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1412 		return (bp);
1413 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1414 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1415 		return (bp);
1416 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1417 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1418 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1419 			return (bp);
1420 	}
1421 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1422 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1423 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1424 			return (bp);
1425 	}
1426 	return (NULL);
1427 }
1428 
1429 /*
1430  * Associate a buffer with a vnode.
1431  */
1432 void
1433 bgetvp(struct vnode *vp, struct buf *bp)
1434 {
1435 
1436 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1437 
1438 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1439 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1440 	    ("bgetvp: bp already attached! %p", bp));
1441 
1442 	ASSERT_VI_LOCKED(vp, "bgetvp");
1443 	vholdl(vp);
1444 	bp->b_vp = vp;
1445 	bp->b_bufobj = &vp->v_bufobj;
1446 	/*
1447 	 * Insert onto list for new vnode.
1448 	 */
1449 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1450 }
1451 
1452 /*
1453  * Disassociate a buffer from a vnode.
1454  */
1455 void
1456 brelvp(struct buf *bp)
1457 {
1458 	struct bufobj *bo;
1459 	struct vnode *vp;
1460 
1461 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1462 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1463 
1464 	/*
1465 	 * Delete from old vnode list, if on one.
1466 	 */
1467 	vp = bp->b_vp;		/* XXX */
1468 	bo = bp->b_bufobj;
1469 	BO_LOCK(bo);
1470 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1471 		buf_vlist_remove(bp);
1472 	else
1473 		panic("brelvp: Buffer %p not on queue.", bp);
1474 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1475 		bo->bo_flag &= ~BO_ONWORKLST;
1476 		mtx_lock(&sync_mtx);
1477 		LIST_REMOVE(bo, bo_synclist);
1478  		syncer_worklist_len--;
1479 		mtx_unlock(&sync_mtx);
1480 	}
1481 	bp->b_vp = NULL;
1482 	bp->b_bufobj = NULL;
1483 	vdropl(vp);
1484 }
1485 
1486 /*
1487  * Add an item to the syncer work queue.
1488  */
1489 static void
1490 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1491 {
1492 	int slot;
1493 
1494 	ASSERT_BO_LOCKED(bo);
1495 
1496 	mtx_lock(&sync_mtx);
1497 	if (bo->bo_flag & BO_ONWORKLST)
1498 		LIST_REMOVE(bo, bo_synclist);
1499 	else {
1500 		bo->bo_flag |= BO_ONWORKLST;
1501  		syncer_worklist_len++;
1502 	}
1503 
1504 	if (delay > syncer_maxdelay - 2)
1505 		delay = syncer_maxdelay - 2;
1506 	slot = (syncer_delayno + delay) & syncer_mask;
1507 
1508 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1509 	mtx_unlock(&sync_mtx);
1510 }
1511 
1512 static int
1513 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1514 {
1515 	int error, len;
1516 
1517 	mtx_lock(&sync_mtx);
1518 	len = syncer_worklist_len - sync_vnode_count;
1519 	mtx_unlock(&sync_mtx);
1520 	error = SYSCTL_OUT(req, &len, sizeof(len));
1521 	return (error);
1522 }
1523 
1524 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1525     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1526 
1527 static struct proc *updateproc;
1528 static void sched_sync(void);
1529 static struct kproc_desc up_kp = {
1530 	"syncer",
1531 	sched_sync,
1532 	&updateproc
1533 };
1534 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1535 
1536 static int
1537 sync_vnode(struct bufobj *bo, struct thread *td)
1538 {
1539 	struct vnode *vp;
1540 	struct mount *mp;
1541 
1542 	vp = bo->__bo_vnode; 	/* XXX */
1543 	if (VOP_ISLOCKED(vp, NULL) != 0)
1544 		return (1);
1545 	if (VI_TRYLOCK(vp) == 0)
1546 		return (1);
1547 	/*
1548 	 * We use vhold in case the vnode does not
1549 	 * successfully sync.  vhold prevents the vnode from
1550 	 * going away when we unlock the sync_mtx so that
1551 	 * we can acquire the vnode interlock.
1552 	 */
1553 	vholdl(vp);
1554 	mtx_unlock(&sync_mtx);
1555 	VI_UNLOCK(vp);
1556 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1557 		vdrop(vp);
1558 		mtx_lock(&sync_mtx);
1559 		return (1);
1560 	}
1561 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1562 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1563 	VOP_UNLOCK(vp, 0, td);
1564 	vn_finished_write(mp);
1565 	VI_LOCK(vp);
1566 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1567 		/*
1568 		 * Put us back on the worklist.  The worklist
1569 		 * routine will remove us from our current
1570 		 * position and then add us back in at a later
1571 		 * position.
1572 		 */
1573 		vn_syncer_add_to_worklist(bo, syncdelay);
1574 	}
1575 	vdropl(vp);
1576 	mtx_lock(&sync_mtx);
1577 	return (0);
1578 }
1579 
1580 /*
1581  * System filesystem synchronizer daemon.
1582  */
1583 static void
1584 sched_sync(void)
1585 {
1586 	struct synclist *next;
1587 	struct synclist *slp;
1588 	struct bufobj *bo;
1589 	long starttime;
1590 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1591 	static int dummychan;
1592 	int last_work_seen;
1593 	int net_worklist_len;
1594 	int syncer_final_iter;
1595 	int first_printf;
1596 	int error;
1597 
1598 	mtx_lock(&Giant);
1599 	last_work_seen = 0;
1600 	syncer_final_iter = 0;
1601 	first_printf = 1;
1602 	syncer_state = SYNCER_RUNNING;
1603 	starttime = time_uptime;
1604 	td->td_pflags |= TDP_NORUNNINGBUF;
1605 
1606 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1607 	    SHUTDOWN_PRI_LAST);
1608 
1609 	for (;;) {
1610 		mtx_lock(&sync_mtx);
1611 		if (syncer_state == SYNCER_FINAL_DELAY &&
1612 		    syncer_final_iter == 0) {
1613 			mtx_unlock(&sync_mtx);
1614 			kthread_suspend_check(td->td_proc);
1615 			mtx_lock(&sync_mtx);
1616 		}
1617 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1618 		if (syncer_state != SYNCER_RUNNING &&
1619 		    starttime != time_uptime) {
1620 			if (first_printf) {
1621 				printf("\nSyncing disks, vnodes remaining...");
1622 				first_printf = 0;
1623 			}
1624 			printf("%d ", net_worklist_len);
1625 		}
1626 		starttime = time_uptime;
1627 
1628 		/*
1629 		 * Push files whose dirty time has expired.  Be careful
1630 		 * of interrupt race on slp queue.
1631 		 *
1632 		 * Skip over empty worklist slots when shutting down.
1633 		 */
1634 		do {
1635 			slp = &syncer_workitem_pending[syncer_delayno];
1636 			syncer_delayno += 1;
1637 			if (syncer_delayno == syncer_maxdelay)
1638 				syncer_delayno = 0;
1639 			next = &syncer_workitem_pending[syncer_delayno];
1640 			/*
1641 			 * If the worklist has wrapped since the
1642 			 * it was emptied of all but syncer vnodes,
1643 			 * switch to the FINAL_DELAY state and run
1644 			 * for one more second.
1645 			 */
1646 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1647 			    net_worklist_len == 0 &&
1648 			    last_work_seen == syncer_delayno) {
1649 				syncer_state = SYNCER_FINAL_DELAY;
1650 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1651 			}
1652 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1653 		    syncer_worklist_len > 0);
1654 
1655 		/*
1656 		 * Keep track of the last time there was anything
1657 		 * on the worklist other than syncer vnodes.
1658 		 * Return to the SHUTTING_DOWN state if any
1659 		 * new work appears.
1660 		 */
1661 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1662 			last_work_seen = syncer_delayno;
1663 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1664 			syncer_state = SYNCER_SHUTTING_DOWN;
1665 		while ((bo = LIST_FIRST(slp)) != NULL) {
1666 			error = sync_vnode(bo, td);
1667 			if (error == 1) {
1668 				LIST_REMOVE(bo, bo_synclist);
1669 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1670 				continue;
1671 			}
1672 		}
1673 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1674 			syncer_final_iter--;
1675 		mtx_unlock(&sync_mtx);
1676 
1677 		/*
1678 		 * Do soft update processing.
1679 		 */
1680 		if (softdep_process_worklist_hook != NULL)
1681 			(*softdep_process_worklist_hook)(NULL);
1682 
1683 		/*
1684 		 * The variable rushjob allows the kernel to speed up the
1685 		 * processing of the filesystem syncer process. A rushjob
1686 		 * value of N tells the filesystem syncer to process the next
1687 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1688 		 * is used by the soft update code to speed up the filesystem
1689 		 * syncer process when the incore state is getting so far
1690 		 * ahead of the disk that the kernel memory pool is being
1691 		 * threatened with exhaustion.
1692 		 */
1693 		mtx_lock(&sync_mtx);
1694 		if (rushjob > 0) {
1695 			rushjob -= 1;
1696 			mtx_unlock(&sync_mtx);
1697 			continue;
1698 		}
1699 		mtx_unlock(&sync_mtx);
1700 		/*
1701 		 * Just sleep for a short period if time between
1702 		 * iterations when shutting down to allow some I/O
1703 		 * to happen.
1704 		 *
1705 		 * If it has taken us less than a second to process the
1706 		 * current work, then wait. Otherwise start right over
1707 		 * again. We can still lose time if any single round
1708 		 * takes more than two seconds, but it does not really
1709 		 * matter as we are just trying to generally pace the
1710 		 * filesystem activity.
1711 		 */
1712 		if (syncer_state != SYNCER_RUNNING)
1713 			tsleep(&dummychan, PPAUSE, "syncfnl",
1714 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1715 		else if (time_uptime == starttime)
1716 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1717 	}
1718 }
1719 
1720 /*
1721  * Request the syncer daemon to speed up its work.
1722  * We never push it to speed up more than half of its
1723  * normal turn time, otherwise it could take over the cpu.
1724  */
1725 int
1726 speedup_syncer()
1727 {
1728 	struct thread *td;
1729 	int ret = 0;
1730 
1731 	td = FIRST_THREAD_IN_PROC(updateproc);
1732 	sleepq_remove(td, &lbolt);
1733 	mtx_lock(&sync_mtx);
1734 	if (rushjob < syncdelay / 2) {
1735 		rushjob += 1;
1736 		stat_rush_requests += 1;
1737 		ret = 1;
1738 	}
1739 	mtx_unlock(&sync_mtx);
1740 	return (ret);
1741 }
1742 
1743 /*
1744  * Tell the syncer to speed up its work and run though its work
1745  * list several times, then tell it to shut down.
1746  */
1747 static void
1748 syncer_shutdown(void *arg, int howto)
1749 {
1750 	struct thread *td;
1751 
1752 	if (howto & RB_NOSYNC)
1753 		return;
1754 	td = FIRST_THREAD_IN_PROC(updateproc);
1755 	sleepq_remove(td, &lbolt);
1756 	mtx_lock(&sync_mtx);
1757 	syncer_state = SYNCER_SHUTTING_DOWN;
1758 	rushjob = 0;
1759 	mtx_unlock(&sync_mtx);
1760 	kproc_shutdown(arg, howto);
1761 }
1762 
1763 /*
1764  * Reassign a buffer from one vnode to another.
1765  * Used to assign file specific control information
1766  * (indirect blocks) to the vnode to which they belong.
1767  */
1768 void
1769 reassignbuf(struct buf *bp)
1770 {
1771 	struct vnode *vp;
1772 	struct bufobj *bo;
1773 	int delay;
1774 #ifdef INVARIANTS
1775 	struct bufv *bv;
1776 #endif
1777 
1778 	vp = bp->b_vp;
1779 	bo = bp->b_bufobj;
1780 	++reassignbufcalls;
1781 
1782 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1783 	    bp, bp->b_vp, bp->b_flags);
1784 	/*
1785 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1786 	 * is not fully linked in.
1787 	 */
1788 	if (bp->b_flags & B_PAGING)
1789 		panic("cannot reassign paging buffer");
1790 
1791 	/*
1792 	 * Delete from old vnode list, if on one.
1793 	 */
1794 	VI_LOCK(vp);
1795 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1796 		buf_vlist_remove(bp);
1797 	else
1798 		panic("reassignbuf: Buffer %p not on queue.", bp);
1799 	/*
1800 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1801 	 * of clean buffers.
1802 	 */
1803 	if (bp->b_flags & B_DELWRI) {
1804 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1805 			switch (vp->v_type) {
1806 			case VDIR:
1807 				delay = dirdelay;
1808 				break;
1809 			case VCHR:
1810 				delay = metadelay;
1811 				break;
1812 			default:
1813 				delay = filedelay;
1814 			}
1815 			vn_syncer_add_to_worklist(bo, delay);
1816 		}
1817 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1818 	} else {
1819 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1820 
1821 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1822 			mtx_lock(&sync_mtx);
1823 			LIST_REMOVE(bo, bo_synclist);
1824  			syncer_worklist_len--;
1825 			mtx_unlock(&sync_mtx);
1826 			bo->bo_flag &= ~BO_ONWORKLST;
1827 		}
1828 	}
1829 #ifdef INVARIANTS
1830 	bv = &bo->bo_clean;
1831 	bp = TAILQ_FIRST(&bv->bv_hd);
1832 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1833 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1834 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1835 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1836 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1837 	bv = &bo->bo_dirty;
1838 	bp = TAILQ_FIRST(&bv->bv_hd);
1839 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1840 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1841 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1842 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1843 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1844 #endif
1845 	VI_UNLOCK(vp);
1846 }
1847 
1848 /*
1849  * Increment the use and hold counts on the vnode, taking care to reference
1850  * the driver's usecount if this is a chardev.  The vholdl() will remove
1851  * the vnode from the free list if it is presently free.  Requires the
1852  * vnode interlock and returns with it held.
1853  */
1854 static void
1855 v_incr_usecount(struct vnode *vp)
1856 {
1857 
1858 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1859 	    vp, vp->v_holdcnt, vp->v_usecount);
1860 	vp->v_usecount++;
1861 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1862 		dev_lock();
1863 		vp->v_rdev->si_usecount++;
1864 		dev_unlock();
1865 	}
1866 	vholdl(vp);
1867 }
1868 
1869 /*
1870  * Decrement the vnode use and hold count along with the driver's usecount
1871  * if this is a chardev.  The vdropl() below releases the vnode interlock
1872  * as it may free the vnode.
1873  */
1874 static void
1875 v_decr_usecount(struct vnode *vp)
1876 {
1877 
1878 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1879 	    vp, vp->v_holdcnt, vp->v_usecount);
1880 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1881 	VNASSERT(vp->v_usecount > 0, vp,
1882 	    ("v_decr_usecount: negative usecount"));
1883 	vp->v_usecount--;
1884 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1885 		dev_lock();
1886 		vp->v_rdev->si_usecount--;
1887 		dev_unlock();
1888 	}
1889 	vdropl(vp);
1890 }
1891 
1892 /*
1893  * Decrement only the use count and driver use count.  This is intended to
1894  * be paired with a follow on vdropl() to release the remaining hold count.
1895  * In this way we may vgone() a vnode with a 0 usecount without risk of
1896  * having it end up on a free list because the hold count is kept above 0.
1897  */
1898 static void
1899 v_decr_useonly(struct vnode *vp)
1900 {
1901 
1902 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1903 	    vp, vp->v_holdcnt, vp->v_usecount);
1904 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1905 	VNASSERT(vp->v_usecount > 0, vp,
1906 	    ("v_decr_useonly: negative usecount"));
1907 	vp->v_usecount--;
1908 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1909 		dev_lock();
1910 		vp->v_rdev->si_usecount--;
1911 		dev_unlock();
1912 	}
1913 }
1914 
1915 /*
1916  * Grab a particular vnode from the free list, increment its
1917  * reference count and lock it. The vnode lock bit is set if the
1918  * vnode is being eliminated in vgone. The process is awakened
1919  * when the transition is completed, and an error returned to
1920  * indicate that the vnode is no longer usable (possibly having
1921  * been changed to a new filesystem type).
1922  */
1923 int
1924 vget(vp, flags, td)
1925 	struct vnode *vp;
1926 	int flags;
1927 	struct thread *td;
1928 {
1929 	int oweinact;
1930 	int oldflags;
1931 	int error;
1932 
1933 	error = 0;
1934 	oldflags = flags;
1935 	oweinact = 0;
1936 	if ((flags & LK_INTERLOCK) == 0)
1937 		VI_LOCK(vp);
1938 	/*
1939 	 * If the inactive call was deferred because vput() was called
1940 	 * with a shared lock, we have to do it here before another thread
1941 	 * gets a reference to data that should be dead.
1942 	 */
1943 	if (vp->v_iflag & VI_OWEINACT) {
1944 		if (flags & LK_NOWAIT) {
1945 			VI_UNLOCK(vp);
1946 			return (EBUSY);
1947 		}
1948 		flags &= ~LK_TYPE_MASK;
1949 		flags |= LK_EXCLUSIVE;
1950 		oweinact = 1;
1951 	}
1952 	v_incr_usecount(vp);
1953 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1954 		VI_LOCK(vp);
1955 		/*
1956 		 * must expand vrele here because we do not want
1957 		 * to call VOP_INACTIVE if the reference count
1958 		 * drops back to zero since it was never really
1959 		 * active.
1960 		 */
1961 		v_decr_usecount(vp);
1962 		return (error);
1963 	}
1964 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
1965 		panic("vget: vn_lock failed to return ENOENT\n");
1966 	if (oweinact) {
1967 		VI_LOCK(vp);
1968 		if (vp->v_iflag & VI_OWEINACT)
1969 			vinactive(vp, td);
1970 		VI_UNLOCK(vp);
1971 		if ((oldflags & LK_TYPE_MASK) == 0)
1972 			VOP_UNLOCK(vp, 0, td);
1973 	}
1974 	return (0);
1975 }
1976 
1977 /*
1978  * Increase the reference count of a vnode.
1979  */
1980 void
1981 vref(struct vnode *vp)
1982 {
1983 
1984 	VI_LOCK(vp);
1985 	v_incr_usecount(vp);
1986 	VI_UNLOCK(vp);
1987 }
1988 
1989 /*
1990  * Return reference count of a vnode.
1991  *
1992  * The results of this call are only guaranteed when some mechanism other
1993  * than the VI lock is used to stop other processes from gaining references
1994  * to the vnode.  This may be the case if the caller holds the only reference.
1995  * This is also useful when stale data is acceptable as race conditions may
1996  * be accounted for by some other means.
1997  */
1998 int
1999 vrefcnt(struct vnode *vp)
2000 {
2001 	int usecnt;
2002 
2003 	VI_LOCK(vp);
2004 	usecnt = vp->v_usecount;
2005 	VI_UNLOCK(vp);
2006 
2007 	return (usecnt);
2008 }
2009 
2010 
2011 /*
2012  * Vnode put/release.
2013  * If count drops to zero, call inactive routine and return to freelist.
2014  */
2015 void
2016 vrele(vp)
2017 	struct vnode *vp;
2018 {
2019 	struct thread *td = curthread;	/* XXX */
2020 
2021 	KASSERT(vp != NULL, ("vrele: null vp"));
2022 
2023 	VI_LOCK(vp);
2024 
2025 	/* Skip this v_writecount check if we're going to panic below. */
2026 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2027 	    ("vrele: missed vn_close"));
2028 
2029 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2030 	    vp->v_usecount == 1)) {
2031 		v_decr_usecount(vp);
2032 		return;
2033 	}
2034 	if (vp->v_usecount != 1) {
2035 #ifdef DIAGNOSTIC
2036 		vprint("vrele: negative ref count", vp);
2037 #endif
2038 		VI_UNLOCK(vp);
2039 		panic("vrele: negative ref cnt");
2040 	}
2041 	/*
2042 	 * We want to hold the vnode until the inactive finishes to
2043 	 * prevent vgone() races.  We drop the use count here and the
2044 	 * hold count below when we're done.
2045 	 */
2046 	v_decr_useonly(vp);
2047 	/*
2048 	 * We must call VOP_INACTIVE with the node locked. Mark
2049 	 * as VI_DOINGINACT to avoid recursion.
2050 	 */
2051 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2052 		VI_LOCK(vp);
2053 		vinactive(vp, td);
2054 		VOP_UNLOCK(vp, 0, td);
2055 	} else
2056 		VI_LOCK(vp);
2057 	vdropl(vp);
2058 }
2059 
2060 /*
2061  * Release an already locked vnode.  This give the same effects as
2062  * unlock+vrele(), but takes less time and avoids releasing and
2063  * re-aquiring the lock (as vrele() aquires the lock internally.)
2064  */
2065 void
2066 vput(vp)
2067 	struct vnode *vp;
2068 {
2069 	struct thread *td = curthread;	/* XXX */
2070 	int error;
2071 
2072 	KASSERT(vp != NULL, ("vput: null vp"));
2073 	ASSERT_VOP_LOCKED(vp, "vput");
2074 	VI_LOCK(vp);
2075 	/* Skip this v_writecount check if we're going to panic below. */
2076 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2077 	    ("vput: missed vn_close"));
2078 	error = 0;
2079 
2080 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2081 	    vp->v_usecount == 1)) {
2082 		VOP_UNLOCK(vp, 0, td);
2083 		v_decr_usecount(vp);
2084 		return;
2085 	}
2086 
2087 	if (vp->v_usecount != 1) {
2088 #ifdef DIAGNOSTIC
2089 		vprint("vput: negative ref count", vp);
2090 #endif
2091 		panic("vput: negative ref cnt");
2092 	}
2093 	/*
2094 	 * We want to hold the vnode until the inactive finishes to
2095 	 * prevent vgone() races.  We drop the use count here and the
2096 	 * hold count below when we're done.
2097 	 */
2098 	v_decr_useonly(vp);
2099 	vp->v_iflag |= VI_OWEINACT;
2100 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2101 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2102 		VI_LOCK(vp);
2103 		if (error)
2104 			goto done;
2105 	}
2106 	if (vp->v_iflag & VI_OWEINACT)
2107 		vinactive(vp, td);
2108 	VOP_UNLOCK(vp, 0, td);
2109 done:
2110 	vdropl(vp);
2111 }
2112 
2113 /*
2114  * Somebody doesn't want the vnode recycled.
2115  */
2116 void
2117 vhold(struct vnode *vp)
2118 {
2119 
2120 	VI_LOCK(vp);
2121 	vholdl(vp);
2122 	VI_UNLOCK(vp);
2123 }
2124 
2125 void
2126 vholdl(struct vnode *vp)
2127 {
2128 
2129 	vp->v_holdcnt++;
2130 	if (VSHOULDBUSY(vp))
2131 		vbusy(vp);
2132 }
2133 
2134 /*
2135  * Note that there is one less who cares about this vnode.  vdrop() is the
2136  * opposite of vhold().
2137  */
2138 void
2139 vdrop(struct vnode *vp)
2140 {
2141 
2142 	VI_LOCK(vp);
2143 	vdropl(vp);
2144 }
2145 
2146 /*
2147  * Drop the hold count of the vnode.  If this is the last reference to
2148  * the vnode we will free it if it has been vgone'd otherwise it is
2149  * placed on the free list.
2150  */
2151 static void
2152 vdropl(struct vnode *vp)
2153 {
2154 
2155 	if (vp->v_holdcnt <= 0)
2156 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2157 	vp->v_holdcnt--;
2158 	if (vp->v_holdcnt == 0) {
2159 		if (vp->v_iflag & VI_DOOMED) {
2160 			vdestroy(vp);
2161 			return;
2162 		} else
2163 			vfree(vp);
2164 	}
2165 	VI_UNLOCK(vp);
2166 }
2167 
2168 /*
2169  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2170  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2171  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2172  * failed lock upgrade.
2173  */
2174 static void
2175 vinactive(struct vnode *vp, struct thread *td)
2176 {
2177 
2178 	ASSERT_VOP_LOCKED(vp, "vinactive");
2179 	ASSERT_VI_LOCKED(vp, "vinactive");
2180 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2181 	    ("vinactive: recursed on VI_DOINGINACT"));
2182 	vp->v_iflag |= VI_DOINGINACT;
2183 	vp->v_iflag &= ~VI_OWEINACT;
2184 	VI_UNLOCK(vp);
2185 	VOP_INACTIVE(vp, td);
2186 	VI_LOCK(vp);
2187 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2188 	    ("vinactive: lost VI_DOINGINACT"));
2189 	vp->v_iflag &= ~VI_DOINGINACT;
2190 }
2191 
2192 /*
2193  * Remove any vnodes in the vnode table belonging to mount point mp.
2194  *
2195  * If FORCECLOSE is not specified, there should not be any active ones,
2196  * return error if any are found (nb: this is a user error, not a
2197  * system error). If FORCECLOSE is specified, detach any active vnodes
2198  * that are found.
2199  *
2200  * If WRITECLOSE is set, only flush out regular file vnodes open for
2201  * writing.
2202  *
2203  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2204  *
2205  * `rootrefs' specifies the base reference count for the root vnode
2206  * of this filesystem. The root vnode is considered busy if its
2207  * v_usecount exceeds this value. On a successful return, vflush(, td)
2208  * will call vrele() on the root vnode exactly rootrefs times.
2209  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2210  * be zero.
2211  */
2212 #ifdef DIAGNOSTIC
2213 static int busyprt = 0;		/* print out busy vnodes */
2214 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2215 #endif
2216 
2217 int
2218 vflush(mp, rootrefs, flags, td)
2219 	struct mount *mp;
2220 	int rootrefs;
2221 	int flags;
2222 	struct thread *td;
2223 {
2224 	struct vnode *vp, *nvp, *rootvp = NULL;
2225 	struct vattr vattr;
2226 	int busy = 0, error;
2227 
2228 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2229 	if (rootrefs > 0) {
2230 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2231 		    ("vflush: bad args"));
2232 		/*
2233 		 * Get the filesystem root vnode. We can vput() it
2234 		 * immediately, since with rootrefs > 0, it won't go away.
2235 		 */
2236 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2237 			return (error);
2238 		vput(rootvp);
2239 
2240 	}
2241 	MNT_ILOCK(mp);
2242 loop:
2243 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2244 
2245 		VI_LOCK(vp);
2246 		vholdl(vp);
2247 		MNT_IUNLOCK(mp);
2248 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2249 		if (error) {
2250 			vdrop(vp);
2251 			MNT_ILOCK(mp);
2252 			goto loop;
2253 		}
2254 		/*
2255 		 * Skip over a vnodes marked VV_SYSTEM.
2256 		 */
2257 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2258 			VOP_UNLOCK(vp, 0, td);
2259 			vdrop(vp);
2260 			MNT_ILOCK(mp);
2261 			continue;
2262 		}
2263 		/*
2264 		 * If WRITECLOSE is set, flush out unlinked but still open
2265 		 * files (even if open only for reading) and regular file
2266 		 * vnodes open for writing.
2267 		 */
2268 		if (flags & WRITECLOSE) {
2269 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2270 			VI_LOCK(vp);
2271 
2272 			if ((vp->v_type == VNON ||
2273 			    (error == 0 && vattr.va_nlink > 0)) &&
2274 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2275 				VOP_UNLOCK(vp, 0, td);
2276 				vdropl(vp);
2277 				MNT_ILOCK(mp);
2278 				continue;
2279 			}
2280 		} else
2281 			VI_LOCK(vp);
2282 		/*
2283 		 * With v_usecount == 0, all we need to do is clear out the
2284 		 * vnode data structures and we are done.
2285 		 *
2286 		 * If FORCECLOSE is set, forcibly close the vnode.
2287 		 */
2288 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2289 			VNASSERT(vp->v_usecount == 0 ||
2290 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2291 			    ("device VNODE %p is FORCECLOSED", vp));
2292 			vgonel(vp);
2293 		} else {
2294 			busy++;
2295 #ifdef DIAGNOSTIC
2296 			if (busyprt)
2297 				vprint("vflush: busy vnode", vp);
2298 #endif
2299 		}
2300 		VOP_UNLOCK(vp, 0, td);
2301 		vdropl(vp);
2302 		MNT_ILOCK(mp);
2303 	}
2304 	MNT_IUNLOCK(mp);
2305 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2306 		/*
2307 		 * If just the root vnode is busy, and if its refcount
2308 		 * is equal to `rootrefs', then go ahead and kill it.
2309 		 */
2310 		VI_LOCK(rootvp);
2311 		KASSERT(busy > 0, ("vflush: not busy"));
2312 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2313 		    ("vflush: usecount %d < rootrefs %d",
2314 		     rootvp->v_usecount, rootrefs));
2315 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2316 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2317 			vgone(rootvp);
2318 			VOP_UNLOCK(rootvp, 0, td);
2319 			busy = 0;
2320 		} else
2321 			VI_UNLOCK(rootvp);
2322 	}
2323 	if (busy)
2324 		return (EBUSY);
2325 	for (; rootrefs > 0; rootrefs--)
2326 		vrele(rootvp);
2327 	return (0);
2328 }
2329 
2330 /*
2331  * Recycle an unused vnode to the front of the free list.
2332  */
2333 int
2334 vrecycle(struct vnode *vp, struct thread *td)
2335 {
2336 	int recycled;
2337 
2338 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2339 	recycled = 0;
2340 	VI_LOCK(vp);
2341 	if (vp->v_usecount == 0) {
2342 		recycled = 1;
2343 		vgonel(vp);
2344 	}
2345 	VI_UNLOCK(vp);
2346 	return (recycled);
2347 }
2348 
2349 /*
2350  * Eliminate all activity associated with a vnode
2351  * in preparation for reuse.
2352  */
2353 void
2354 vgone(struct vnode *vp)
2355 {
2356 	VI_LOCK(vp);
2357 	vgonel(vp);
2358 	VI_UNLOCK(vp);
2359 }
2360 
2361 /*
2362  * vgone, with the vp interlock held.
2363  */
2364 void
2365 vgonel(struct vnode *vp)
2366 {
2367 	struct thread *td;
2368 	int oweinact;
2369 	int active;
2370 
2371 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2372 	ASSERT_VOP_LOCKED(vp, "vgonel");
2373 	ASSERT_VI_LOCKED(vp, "vgonel");
2374 #if 0
2375 	/* XXX Need to fix ttyvp before I enable this. */
2376 	VNASSERT(vp->v_holdcnt, vp,
2377 	    ("vgonel: vp %p has no reference.", vp));
2378 #endif
2379 	td = curthread;
2380 
2381 	/*
2382 	 * Don't vgonel if we're already doomed.
2383 	 */
2384 	if (vp->v_iflag & VI_DOOMED)
2385 		return;
2386 	vp->v_iflag |= VI_DOOMED;
2387 	/*
2388 	 * Check to see if the vnode is in use.  If so, we have to call
2389 	 * VOP_CLOSE() and VOP_INACTIVE().
2390 	 */
2391 	active = vp->v_usecount;
2392 	oweinact = (vp->v_iflag & VI_OWEINACT);
2393 	VI_UNLOCK(vp);
2394 	/*
2395 	 * Clean out any buffers associated with the vnode.
2396 	 * If the flush fails, just toss the buffers.
2397 	 */
2398 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2399 		(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2400 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2401 		vinvalbuf(vp, 0, td, 0, 0);
2402 
2403 	/*
2404 	 * If purging an active vnode, it must be closed and
2405 	 * deactivated before being reclaimed.
2406 	 */
2407 	if (active)
2408 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2409 	if (oweinact || active) {
2410 		VI_LOCK(vp);
2411 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2412 			vinactive(vp, td);
2413 		VI_UNLOCK(vp);
2414 	}
2415 	/*
2416 	 * Reclaim the vnode.
2417 	 */
2418 	if (VOP_RECLAIM(vp, td))
2419 		panic("vgone: cannot reclaim");
2420 	VNASSERT(vp->v_object == NULL, vp,
2421 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2422 	/*
2423 	 * Delete from old mount point vnode list.
2424 	 */
2425 	delmntque(vp);
2426 	cache_purge(vp);
2427 	/*
2428 	 * Done with purge, reset to the standard lock and invalidate
2429 	 * the vnode.
2430 	 */
2431 	VI_LOCK(vp);
2432 	vp->v_vnlock = &vp->v_lock;
2433 	vp->v_op = &dead_vnodeops;
2434 	vp->v_tag = "none";
2435 	vp->v_type = VBAD;
2436 }
2437 
2438 /*
2439  * Calculate the total number of references to a special device.
2440  */
2441 int
2442 vcount(vp)
2443 	struct vnode *vp;
2444 {
2445 	int count;
2446 
2447 	dev_lock();
2448 	count = vp->v_rdev->si_usecount;
2449 	dev_unlock();
2450 	return (count);
2451 }
2452 
2453 /*
2454  * Same as above, but using the struct cdev *as argument
2455  */
2456 int
2457 count_dev(dev)
2458 	struct cdev *dev;
2459 {
2460 	int count;
2461 
2462 	dev_lock();
2463 	count = dev->si_usecount;
2464 	dev_unlock();
2465 	return(count);
2466 }
2467 
2468 /*
2469  * Print out a description of a vnode.
2470  */
2471 static char *typename[] =
2472 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2473 
2474 void
2475 vn_printf(struct vnode *vp, const char *fmt, ...)
2476 {
2477 	va_list ap;
2478 	char buf[96];
2479 
2480 	va_start(ap, fmt);
2481 	vprintf(fmt, ap);
2482 	va_end(ap);
2483 	printf("%p: ", (void *)vp);
2484 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2485 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2486 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2487 	buf[0] = '\0';
2488 	buf[1] = '\0';
2489 	if (vp->v_vflag & VV_ROOT)
2490 		strcat(buf, "|VV_ROOT");
2491 	if (vp->v_vflag & VV_TEXT)
2492 		strcat(buf, "|VV_TEXT");
2493 	if (vp->v_vflag & VV_SYSTEM)
2494 		strcat(buf, "|VV_SYSTEM");
2495 	if (vp->v_iflag & VI_DOOMED)
2496 		strcat(buf, "|VI_DOOMED");
2497 	if (vp->v_iflag & VI_FREE)
2498 		strcat(buf, "|VI_FREE");
2499 	printf("    flags (%s)\n", buf + 1);
2500 	if (mtx_owned(VI_MTX(vp)))
2501 		printf(" VI_LOCKed");
2502 	if (vp->v_object != NULL)
2503 		printf("    v_object %p ref %d pages %d\n",
2504 		    vp->v_object, vp->v_object->ref_count,
2505 		    vp->v_object->resident_page_count);
2506 	printf("    ");
2507 	lockmgr_printinfo(vp->v_vnlock);
2508 	printf("\n");
2509 	if (vp->v_data != NULL)
2510 		VOP_PRINT(vp);
2511 }
2512 
2513 #ifdef DDB
2514 #include <ddb/ddb.h>
2515 /*
2516  * List all of the locked vnodes in the system.
2517  * Called when debugging the kernel.
2518  */
2519 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2520 {
2521 	struct mount *mp, *nmp;
2522 	struct vnode *vp;
2523 
2524 	/*
2525 	 * Note: because this is DDB, we can't obey the locking semantics
2526 	 * for these structures, which means we could catch an inconsistent
2527 	 * state and dereference a nasty pointer.  Not much to be done
2528 	 * about that.
2529 	 */
2530 	printf("Locked vnodes\n");
2531 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2532 		nmp = TAILQ_NEXT(mp, mnt_list);
2533 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2534 			if (VOP_ISLOCKED(vp, NULL))
2535 				vprint("", vp);
2536 		}
2537 		nmp = TAILQ_NEXT(mp, mnt_list);
2538 	}
2539 }
2540 #endif
2541 
2542 /*
2543  * Fill in a struct xvfsconf based on a struct vfsconf.
2544  */
2545 static void
2546 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2547 {
2548 
2549 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2550 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2551 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2552 	xvfsp->vfc_flags = vfsp->vfc_flags;
2553 	/*
2554 	 * These are unused in userland, we keep them
2555 	 * to not break binary compatibility.
2556 	 */
2557 	xvfsp->vfc_vfsops = NULL;
2558 	xvfsp->vfc_next = NULL;
2559 }
2560 
2561 /*
2562  * Top level filesystem related information gathering.
2563  */
2564 static int
2565 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2566 {
2567 	struct vfsconf *vfsp;
2568 	struct xvfsconf xvfsp;
2569 	int error;
2570 
2571 	error = 0;
2572 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2573 		bzero(&xvfsp, sizeof(xvfsp));
2574 		vfsconf2x(vfsp, &xvfsp);
2575 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2576 		if (error)
2577 			break;
2578 	}
2579 	return (error);
2580 }
2581 
2582 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2583     "S,xvfsconf", "List of all configured filesystems");
2584 
2585 #ifndef BURN_BRIDGES
2586 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2587 
2588 static int
2589 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2590 {
2591 	int *name = (int *)arg1 - 1;	/* XXX */
2592 	u_int namelen = arg2 + 1;	/* XXX */
2593 	struct vfsconf *vfsp;
2594 	struct xvfsconf xvfsp;
2595 
2596 	printf("WARNING: userland calling deprecated sysctl, "
2597 	    "please rebuild world\n");
2598 
2599 #if 1 || defined(COMPAT_PRELITE2)
2600 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2601 	if (namelen == 1)
2602 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2603 #endif
2604 
2605 	switch (name[1]) {
2606 	case VFS_MAXTYPENUM:
2607 		if (namelen != 2)
2608 			return (ENOTDIR);
2609 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2610 	case VFS_CONF:
2611 		if (namelen != 3)
2612 			return (ENOTDIR);	/* overloaded */
2613 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2614 			if (vfsp->vfc_typenum == name[2])
2615 				break;
2616 		if (vfsp == NULL)
2617 			return (EOPNOTSUPP);
2618 		bzero(&xvfsp, sizeof(xvfsp));
2619 		vfsconf2x(vfsp, &xvfsp);
2620 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2621 	}
2622 	return (EOPNOTSUPP);
2623 }
2624 
2625 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2626 	vfs_sysctl, "Generic filesystem");
2627 
2628 #if 1 || defined(COMPAT_PRELITE2)
2629 
2630 static int
2631 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2632 {
2633 	int error;
2634 	struct vfsconf *vfsp;
2635 	struct ovfsconf ovfs;
2636 
2637 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2638 		bzero(&ovfs, sizeof(ovfs));
2639 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2640 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2641 		ovfs.vfc_index = vfsp->vfc_typenum;
2642 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2643 		ovfs.vfc_flags = vfsp->vfc_flags;
2644 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2645 		if (error)
2646 			return error;
2647 	}
2648 	return 0;
2649 }
2650 
2651 #endif /* 1 || COMPAT_PRELITE2 */
2652 #endif /* !BURN_BRIDGES */
2653 
2654 #define KINFO_VNODESLOP		10
2655 #ifdef notyet
2656 /*
2657  * Dump vnode list (via sysctl).
2658  */
2659 /* ARGSUSED */
2660 static int
2661 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2662 {
2663 	struct xvnode *xvn;
2664 	struct thread *td = req->td;
2665 	struct mount *mp;
2666 	struct vnode *vp;
2667 	int error, len, n;
2668 
2669 	/*
2670 	 * Stale numvnodes access is not fatal here.
2671 	 */
2672 	req->lock = 0;
2673 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2674 	if (!req->oldptr)
2675 		/* Make an estimate */
2676 		return (SYSCTL_OUT(req, 0, len));
2677 
2678 	error = sysctl_wire_old_buffer(req, 0);
2679 	if (error != 0)
2680 		return (error);
2681 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2682 	n = 0;
2683 	mtx_lock(&mountlist_mtx);
2684 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2685 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2686 			continue;
2687 		MNT_ILOCK(mp);
2688 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2689 			if (n == len)
2690 				break;
2691 			vref(vp);
2692 			xvn[n].xv_size = sizeof *xvn;
2693 			xvn[n].xv_vnode = vp;
2694 			xvn[n].xv_id = 0;	/* XXX compat */
2695 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2696 			XV_COPY(usecount);
2697 			XV_COPY(writecount);
2698 			XV_COPY(holdcnt);
2699 			XV_COPY(mount);
2700 			XV_COPY(numoutput);
2701 			XV_COPY(type);
2702 #undef XV_COPY
2703 			xvn[n].xv_flag = vp->v_vflag;
2704 
2705 			switch (vp->v_type) {
2706 			case VREG:
2707 			case VDIR:
2708 			case VLNK:
2709 				break;
2710 			case VBLK:
2711 			case VCHR:
2712 				if (vp->v_rdev == NULL) {
2713 					vrele(vp);
2714 					continue;
2715 				}
2716 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2717 				break;
2718 			case VSOCK:
2719 				xvn[n].xv_socket = vp->v_socket;
2720 				break;
2721 			case VFIFO:
2722 				xvn[n].xv_fifo = vp->v_fifoinfo;
2723 				break;
2724 			case VNON:
2725 			case VBAD:
2726 			default:
2727 				/* shouldn't happen? */
2728 				vrele(vp);
2729 				continue;
2730 			}
2731 			vrele(vp);
2732 			++n;
2733 		}
2734 		MNT_IUNLOCK(mp);
2735 		mtx_lock(&mountlist_mtx);
2736 		vfs_unbusy(mp, td);
2737 		if (n == len)
2738 			break;
2739 	}
2740 	mtx_unlock(&mountlist_mtx);
2741 
2742 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2743 	free(xvn, M_TEMP);
2744 	return (error);
2745 }
2746 
2747 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2748 	0, 0, sysctl_vnode, "S,xvnode", "");
2749 #endif
2750 
2751 /*
2752  * Unmount all filesystems. The list is traversed in reverse order
2753  * of mounting to avoid dependencies.
2754  */
2755 void
2756 vfs_unmountall()
2757 {
2758 	struct mount *mp;
2759 	struct thread *td;
2760 	int error;
2761 
2762 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2763 	td = curthread;
2764 	/*
2765 	 * Since this only runs when rebooting, it is not interlocked.
2766 	 */
2767 	while(!TAILQ_EMPTY(&mountlist)) {
2768 		mp = TAILQ_LAST(&mountlist, mntlist);
2769 		error = dounmount(mp, MNT_FORCE, td);
2770 		if (error) {
2771 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2772 			/*
2773 			 * XXX: Due to the way in which we mount the root
2774 			 * file system off of devfs, devfs will generate a
2775 			 * "busy" warning when we try to unmount it before
2776 			 * the root.  Don't print a warning as a result in
2777 			 * order to avoid false positive errors that may
2778 			 * cause needless upset.
2779 			 */
2780 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2781 				printf("unmount of %s failed (",
2782 				    mp->mnt_stat.f_mntonname);
2783 				if (error == EBUSY)
2784 					printf("BUSY)\n");
2785 				else
2786 					printf("%d)\n", error);
2787 			}
2788 		} else {
2789 			/* The unmount has removed mp from the mountlist */
2790 		}
2791 	}
2792 }
2793 
2794 /*
2795  * perform msync on all vnodes under a mount point
2796  * the mount point must be locked.
2797  */
2798 void
2799 vfs_msync(struct mount *mp, int flags)
2800 {
2801 	struct vnode *vp, *nvp;
2802 	struct vm_object *obj;
2803 	int tries;
2804 
2805 	tries = 5;
2806 	MNT_ILOCK(mp);
2807 loop:
2808 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
2809 		if (vp->v_mount != mp) {
2810 			if (--tries > 0)
2811 				goto loop;
2812 			break;
2813 		}
2814 
2815 		VI_LOCK(vp);
2816 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2817 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2818 			MNT_IUNLOCK(mp);
2819 			if (!vget(vp,
2820 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2821 			    curthread)) {
2822 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2823 					vput(vp);
2824 					MNT_ILOCK(mp);
2825 					continue;
2826 				}
2827 
2828 				obj = vp->v_object;
2829 				if (obj != NULL) {
2830 					VM_OBJECT_LOCK(obj);
2831 					vm_object_page_clean(obj, 0, 0,
2832 					    flags == MNT_WAIT ?
2833 					    OBJPC_SYNC : OBJPC_NOSYNC);
2834 					VM_OBJECT_UNLOCK(obj);
2835 				}
2836 				vput(vp);
2837 			}
2838 			MNT_ILOCK(mp);
2839 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2840 				if (--tries > 0)
2841 					goto loop;
2842 				break;
2843 			}
2844 		} else
2845 			VI_UNLOCK(vp);
2846 	}
2847 	MNT_IUNLOCK(mp);
2848 }
2849 
2850 /*
2851  * Mark a vnode as free, putting it up for recycling.
2852  */
2853 static void
2854 vfree(struct vnode *vp)
2855 {
2856 
2857 	CTR1(KTR_VFS, "vfree vp %p", vp);
2858 	ASSERT_VI_LOCKED(vp, "vfree");
2859 	mtx_lock(&vnode_free_list_mtx);
2860 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2861 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2862 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2863 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2864 	    ("vfree: Freeing doomed vnode"));
2865 	if (vp->v_iflag & VI_AGE) {
2866 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2867 	} else {
2868 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2869 	}
2870 	freevnodes++;
2871 	vp->v_iflag &= ~VI_AGE;
2872 	vp->v_iflag |= VI_FREE;
2873 	mtx_unlock(&vnode_free_list_mtx);
2874 }
2875 
2876 /*
2877  * Opposite of vfree() - mark a vnode as in use.
2878  */
2879 static void
2880 vbusy(struct vnode *vp)
2881 {
2882 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2883 	ASSERT_VI_LOCKED(vp, "vbusy");
2884 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2885 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2886 
2887 	mtx_lock(&vnode_free_list_mtx);
2888 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2889 	freevnodes--;
2890 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2891 	mtx_unlock(&vnode_free_list_mtx);
2892 }
2893 
2894 /*
2895  * Initalize per-vnode helper structure to hold poll-related state.
2896  */
2897 void
2898 v_addpollinfo(struct vnode *vp)
2899 {
2900 	struct vpollinfo *vi;
2901 
2902 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2903 	if (vp->v_pollinfo != NULL) {
2904 		uma_zfree(vnodepoll_zone, vi);
2905 		return;
2906 	}
2907 	vp->v_pollinfo = vi;
2908 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2909 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
2910 	    vfs_knlunlock, vfs_knllocked);
2911 }
2912 
2913 /*
2914  * Record a process's interest in events which might happen to
2915  * a vnode.  Because poll uses the historic select-style interface
2916  * internally, this routine serves as both the ``check for any
2917  * pending events'' and the ``record my interest in future events''
2918  * functions.  (These are done together, while the lock is held,
2919  * to avoid race conditions.)
2920  */
2921 int
2922 vn_pollrecord(vp, td, events)
2923 	struct vnode *vp;
2924 	struct thread *td;
2925 	short events;
2926 {
2927 
2928 	if (vp->v_pollinfo == NULL)
2929 		v_addpollinfo(vp);
2930 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2931 	if (vp->v_pollinfo->vpi_revents & events) {
2932 		/*
2933 		 * This leaves events we are not interested
2934 		 * in available for the other process which
2935 		 * which presumably had requested them
2936 		 * (otherwise they would never have been
2937 		 * recorded).
2938 		 */
2939 		events &= vp->v_pollinfo->vpi_revents;
2940 		vp->v_pollinfo->vpi_revents &= ~events;
2941 
2942 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2943 		return events;
2944 	}
2945 	vp->v_pollinfo->vpi_events |= events;
2946 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2947 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2948 	return 0;
2949 }
2950 
2951 /*
2952  * Routine to create and manage a filesystem syncer vnode.
2953  */
2954 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2955 static int	sync_fsync(struct  vop_fsync_args *);
2956 static int	sync_inactive(struct  vop_inactive_args *);
2957 static int	sync_reclaim(struct  vop_reclaim_args *);
2958 
2959 static struct vop_vector sync_vnodeops = {
2960 	.vop_bypass =	VOP_EOPNOTSUPP,
2961 	.vop_close =	sync_close,		/* close */
2962 	.vop_fsync =	sync_fsync,		/* fsync */
2963 	.vop_inactive =	sync_inactive,	/* inactive */
2964 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2965 	.vop_lock =	vop_stdlock,	/* lock */
2966 	.vop_unlock =	vop_stdunlock,	/* unlock */
2967 	.vop_islocked =	vop_stdislocked,	/* islocked */
2968 };
2969 
2970 /*
2971  * Create a new filesystem syncer vnode for the specified mount point.
2972  */
2973 int
2974 vfs_allocate_syncvnode(mp)
2975 	struct mount *mp;
2976 {
2977 	struct vnode *vp;
2978 	static long start, incr, next;
2979 	int error;
2980 
2981 	/* Allocate a new vnode */
2982 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2983 		mp->mnt_syncer = NULL;
2984 		return (error);
2985 	}
2986 	vp->v_type = VNON;
2987 	/*
2988 	 * Place the vnode onto the syncer worklist. We attempt to
2989 	 * scatter them about on the list so that they will go off
2990 	 * at evenly distributed times even if all the filesystems
2991 	 * are mounted at once.
2992 	 */
2993 	next += incr;
2994 	if (next == 0 || next > syncer_maxdelay) {
2995 		start /= 2;
2996 		incr /= 2;
2997 		if (start == 0) {
2998 			start = syncer_maxdelay / 2;
2999 			incr = syncer_maxdelay;
3000 		}
3001 		next = start;
3002 	}
3003 	VI_LOCK(vp);
3004 	vn_syncer_add_to_worklist(&vp->v_bufobj,
3005 	    syncdelay > 0 ? next % syncdelay : 0);
3006 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3007 	mtx_lock(&sync_mtx);
3008 	sync_vnode_count++;
3009 	mtx_unlock(&sync_mtx);
3010 	VI_UNLOCK(vp);
3011 	mp->mnt_syncer = vp;
3012 	return (0);
3013 }
3014 
3015 /*
3016  * Do a lazy sync of the filesystem.
3017  */
3018 static int
3019 sync_fsync(ap)
3020 	struct vop_fsync_args /* {
3021 		struct vnode *a_vp;
3022 		struct ucred *a_cred;
3023 		int a_waitfor;
3024 		struct thread *a_td;
3025 	} */ *ap;
3026 {
3027 	struct vnode *syncvp = ap->a_vp;
3028 	struct mount *mp = syncvp->v_mount;
3029 	struct thread *td = ap->a_td;
3030 	int error, asyncflag;
3031 	struct bufobj *bo;
3032 
3033 	/*
3034 	 * We only need to do something if this is a lazy evaluation.
3035 	 */
3036 	if (ap->a_waitfor != MNT_LAZY)
3037 		return (0);
3038 
3039 	/*
3040 	 * Move ourselves to the back of the sync list.
3041 	 */
3042 	bo = &syncvp->v_bufobj;
3043 	BO_LOCK(bo);
3044 	vn_syncer_add_to_worklist(bo, syncdelay);
3045 	BO_UNLOCK(bo);
3046 
3047 	/*
3048 	 * Walk the list of vnodes pushing all that are dirty and
3049 	 * not already on the sync list.
3050 	 */
3051 	mtx_lock(&mountlist_mtx);
3052 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3053 		mtx_unlock(&mountlist_mtx);
3054 		return (0);
3055 	}
3056 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3057 		vfs_unbusy(mp, td);
3058 		return (0);
3059 	}
3060 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3061 	mp->mnt_flag &= ~MNT_ASYNC;
3062 	vfs_msync(mp, MNT_NOWAIT);
3063 	error = VFS_SYNC(mp, MNT_LAZY, td);
3064 	if (asyncflag)
3065 		mp->mnt_flag |= MNT_ASYNC;
3066 	vn_finished_write(mp);
3067 	vfs_unbusy(mp, td);
3068 	return (error);
3069 }
3070 
3071 /*
3072  * The syncer vnode is no referenced.
3073  */
3074 static int
3075 sync_inactive(ap)
3076 	struct vop_inactive_args /* {
3077 		struct vnode *a_vp;
3078 		struct thread *a_td;
3079 	} */ *ap;
3080 {
3081 
3082 	vgone(ap->a_vp);
3083 	return (0);
3084 }
3085 
3086 /*
3087  * The syncer vnode is no longer needed and is being decommissioned.
3088  *
3089  * Modifications to the worklist must be protected by sync_mtx.
3090  */
3091 static int
3092 sync_reclaim(ap)
3093 	struct vop_reclaim_args /* {
3094 		struct vnode *a_vp;
3095 	} */ *ap;
3096 {
3097 	struct vnode *vp = ap->a_vp;
3098 	struct bufobj *bo;
3099 
3100 	VI_LOCK(vp);
3101 	bo = &vp->v_bufobj;
3102 	vp->v_mount->mnt_syncer = NULL;
3103 	if (bo->bo_flag & BO_ONWORKLST) {
3104 		mtx_lock(&sync_mtx);
3105 		LIST_REMOVE(bo, bo_synclist);
3106  		syncer_worklist_len--;
3107 		sync_vnode_count--;
3108 		mtx_unlock(&sync_mtx);
3109 		bo->bo_flag &= ~BO_ONWORKLST;
3110 	}
3111 	VI_UNLOCK(vp);
3112 
3113 	return (0);
3114 }
3115 
3116 /*
3117  * Check if vnode represents a disk device
3118  */
3119 int
3120 vn_isdisk(vp, errp)
3121 	struct vnode *vp;
3122 	int *errp;
3123 {
3124 	int error;
3125 
3126 	error = 0;
3127 	dev_lock();
3128 	if (vp->v_type != VCHR)
3129 		error = ENOTBLK;
3130 	else if (vp->v_rdev == NULL)
3131 		error = ENXIO;
3132 	else if (vp->v_rdev->si_devsw == NULL)
3133 		error = ENXIO;
3134 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3135 		error = ENOTBLK;
3136 	dev_unlock();
3137 	if (errp != NULL)
3138 		*errp = error;
3139 	return (error == 0);
3140 }
3141 
3142 /*
3143  * Common filesystem object access control check routine.  Accepts a
3144  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3145  * and optional call-by-reference privused argument allowing vaccess()
3146  * to indicate to the caller whether privilege was used to satisfy the
3147  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3148  */
3149 int
3150 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3151 	enum vtype type;
3152 	mode_t file_mode;
3153 	uid_t file_uid;
3154 	gid_t file_gid;
3155 	mode_t acc_mode;
3156 	struct ucred *cred;
3157 	int *privused;
3158 {
3159 	mode_t dac_granted;
3160 #ifdef CAPABILITIES
3161 	mode_t cap_granted;
3162 #endif
3163 
3164 	/*
3165 	 * Look for a normal, non-privileged way to access the file/directory
3166 	 * as requested.  If it exists, go with that.
3167 	 */
3168 
3169 	if (privused != NULL)
3170 		*privused = 0;
3171 
3172 	dac_granted = 0;
3173 
3174 	/* Check the owner. */
3175 	if (cred->cr_uid == file_uid) {
3176 		dac_granted |= VADMIN;
3177 		if (file_mode & S_IXUSR)
3178 			dac_granted |= VEXEC;
3179 		if (file_mode & S_IRUSR)
3180 			dac_granted |= VREAD;
3181 		if (file_mode & S_IWUSR)
3182 			dac_granted |= (VWRITE | VAPPEND);
3183 
3184 		if ((acc_mode & dac_granted) == acc_mode)
3185 			return (0);
3186 
3187 		goto privcheck;
3188 	}
3189 
3190 	/* Otherwise, check the groups (first match) */
3191 	if (groupmember(file_gid, cred)) {
3192 		if (file_mode & S_IXGRP)
3193 			dac_granted |= VEXEC;
3194 		if (file_mode & S_IRGRP)
3195 			dac_granted |= VREAD;
3196 		if (file_mode & S_IWGRP)
3197 			dac_granted |= (VWRITE | VAPPEND);
3198 
3199 		if ((acc_mode & dac_granted) == acc_mode)
3200 			return (0);
3201 
3202 		goto privcheck;
3203 	}
3204 
3205 	/* Otherwise, check everyone else. */
3206 	if (file_mode & S_IXOTH)
3207 		dac_granted |= VEXEC;
3208 	if (file_mode & S_IROTH)
3209 		dac_granted |= VREAD;
3210 	if (file_mode & S_IWOTH)
3211 		dac_granted |= (VWRITE | VAPPEND);
3212 	if ((acc_mode & dac_granted) == acc_mode)
3213 		return (0);
3214 
3215 privcheck:
3216 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3217 		/* XXX audit: privilege used */
3218 		if (privused != NULL)
3219 			*privused = 1;
3220 		return (0);
3221 	}
3222 
3223 #ifdef CAPABILITIES
3224 	/*
3225 	 * Build a capability mask to determine if the set of capabilities
3226 	 * satisfies the requirements when combined with the granted mask
3227 	 * from above.
3228 	 * For each capability, if the capability is required, bitwise
3229 	 * or the request type onto the cap_granted mask.
3230 	 */
3231 	cap_granted = 0;
3232 
3233 	if (type == VDIR) {
3234 		/*
3235 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3236 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3237 		 */
3238 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3239 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3240 			cap_granted |= VEXEC;
3241 	} else {
3242 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3243 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3244 			cap_granted |= VEXEC;
3245 	}
3246 
3247 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3248 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3249 		cap_granted |= VREAD;
3250 
3251 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3252 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3253 		cap_granted |= (VWRITE | VAPPEND);
3254 
3255 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3256 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3257 		cap_granted |= VADMIN;
3258 
3259 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3260 		/* XXX audit: privilege used */
3261 		if (privused != NULL)
3262 			*privused = 1;
3263 		return (0);
3264 	}
3265 #endif
3266 
3267 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3268 }
3269 
3270 /*
3271  * Credential check based on process requesting service, and per-attribute
3272  * permissions.
3273  */
3274 int
3275 extattr_check_cred(struct vnode *vp, int attrnamespace,
3276     struct ucred *cred, struct thread *td, int access)
3277 {
3278 
3279 	/*
3280 	 * Kernel-invoked always succeeds.
3281 	 */
3282 	if (cred == NOCRED)
3283 		return (0);
3284 
3285 	/*
3286 	 * Do not allow privileged processes in jail to directly
3287 	 * manipulate system attributes.
3288 	 *
3289 	 * XXX What capability should apply here?
3290 	 * Probably CAP_SYS_SETFFLAG.
3291 	 */
3292 	switch (attrnamespace) {
3293 	case EXTATTR_NAMESPACE_SYSTEM:
3294 		/* Potentially should be: return (EPERM); */
3295 		return (suser_cred(cred, 0));
3296 	case EXTATTR_NAMESPACE_USER:
3297 		return (VOP_ACCESS(vp, access, cred, td));
3298 	default:
3299 		return (EPERM);
3300 	}
3301 }
3302 
3303 #ifdef DEBUG_VFS_LOCKS
3304 /*
3305  * This only exists to supress warnings from unlocked specfs accesses.  It is
3306  * no longer ok to have an unlocked VFS.
3307  */
3308 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3309 
3310 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3311 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3312 
3313 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3314 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3315 
3316 int vfs_badlock_print = 1;	/* Print lock violations. */
3317 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3318 
3319 #ifdef KDB
3320 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3321 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3322 #endif
3323 
3324 static void
3325 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3326 {
3327 
3328 #ifdef KDB
3329 	if (vfs_badlock_backtrace)
3330 		kdb_backtrace();
3331 #endif
3332 	if (vfs_badlock_print)
3333 		printf("%s: %p %s\n", str, (void *)vp, msg);
3334 	if (vfs_badlock_ddb)
3335 		kdb_enter("lock violation");
3336 }
3337 
3338 void
3339 assert_vi_locked(struct vnode *vp, const char *str)
3340 {
3341 
3342 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3343 		vfs_badlock("interlock is not locked but should be", str, vp);
3344 }
3345 
3346 void
3347 assert_vi_unlocked(struct vnode *vp, const char *str)
3348 {
3349 
3350 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3351 		vfs_badlock("interlock is locked but should not be", str, vp);
3352 }
3353 
3354 void
3355 assert_vop_locked(struct vnode *vp, const char *str)
3356 {
3357 
3358 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3359 		vfs_badlock("is not locked but should be", str, vp);
3360 }
3361 
3362 void
3363 assert_vop_unlocked(struct vnode *vp, const char *str)
3364 {
3365 
3366 	if (vp && !IGNORE_LOCK(vp) &&
3367 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3368 		vfs_badlock("is locked but should not be", str, vp);
3369 }
3370 
3371 void
3372 assert_vop_elocked(struct vnode *vp, const char *str)
3373 {
3374 
3375 	if (vp && !IGNORE_LOCK(vp) &&
3376 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3377 		vfs_badlock("is not exclusive locked but should be", str, vp);
3378 }
3379 
3380 #if 0
3381 void
3382 assert_vop_elocked_other(struct vnode *vp, const char *str)
3383 {
3384 
3385 	if (vp && !IGNORE_LOCK(vp) &&
3386 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3387 		vfs_badlock("is not exclusive locked by another thread",
3388 		    str, vp);
3389 }
3390 
3391 void
3392 assert_vop_slocked(struct vnode *vp, const char *str)
3393 {
3394 
3395 	if (vp && !IGNORE_LOCK(vp) &&
3396 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3397 		vfs_badlock("is not locked shared but should be", str, vp);
3398 }
3399 #endif /* 0 */
3400 #endif /* DEBUG_VFS_LOCKS */
3401 
3402 void
3403 vop_rename_pre(void *ap)
3404 {
3405 	struct vop_rename_args *a = ap;
3406 
3407 #ifdef DEBUG_VFS_LOCKS
3408 	if (a->a_tvp)
3409 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3410 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3411 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3412 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3413 
3414 	/* Check the source (from). */
3415 	if (a->a_tdvp != a->a_fdvp)
3416 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3417 	if (a->a_tvp != a->a_fvp)
3418 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3419 
3420 	/* Check the target. */
3421 	if (a->a_tvp)
3422 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3423 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3424 #endif
3425 	if (a->a_tdvp != a->a_fdvp)
3426 		vhold(a->a_fdvp);
3427 	if (a->a_tvp != a->a_fvp)
3428 		vhold(a->a_fvp);
3429 	vhold(a->a_tdvp);
3430 	if (a->a_tvp)
3431 		vhold(a->a_tvp);
3432 }
3433 
3434 void
3435 vop_strategy_pre(void *ap)
3436 {
3437 #ifdef DEBUG_VFS_LOCKS
3438 	struct vop_strategy_args *a;
3439 	struct buf *bp;
3440 
3441 	a = ap;
3442 	bp = a->a_bp;
3443 
3444 	/*
3445 	 * Cluster ops lock their component buffers but not the IO container.
3446 	 */
3447 	if ((bp->b_flags & B_CLUSTER) != 0)
3448 		return;
3449 
3450 	if (BUF_REFCNT(bp) < 1) {
3451 		if (vfs_badlock_print)
3452 			printf(
3453 			    "VOP_STRATEGY: bp is not locked but should be\n");
3454 		if (vfs_badlock_ddb)
3455 			kdb_enter("lock violation");
3456 	}
3457 #endif
3458 }
3459 
3460 void
3461 vop_lookup_pre(void *ap)
3462 {
3463 #ifdef DEBUG_VFS_LOCKS
3464 	struct vop_lookup_args *a;
3465 	struct vnode *dvp;
3466 
3467 	a = ap;
3468 	dvp = a->a_dvp;
3469 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3470 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3471 #endif
3472 }
3473 
3474 void
3475 vop_lookup_post(void *ap, int rc)
3476 {
3477 #ifdef DEBUG_VFS_LOCKS
3478 	struct vop_lookup_args *a;
3479 	struct vnode *dvp;
3480 	struct vnode *vp;
3481 
3482 	a = ap;
3483 	dvp = a->a_dvp;
3484 	vp = *(a->a_vpp);
3485 
3486 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3487 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3488 
3489 	if (!rc)
3490 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3491 #endif
3492 }
3493 
3494 void
3495 vop_lock_pre(void *ap)
3496 {
3497 #ifdef DEBUG_VFS_LOCKS
3498 	struct vop_lock_args *a = ap;
3499 
3500 	if ((a->a_flags & LK_INTERLOCK) == 0)
3501 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3502 	else
3503 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3504 #endif
3505 }
3506 
3507 void
3508 vop_lock_post(void *ap, int rc)
3509 {
3510 #ifdef DEBUG_VFS_LOCKS
3511 	struct vop_lock_args *a = ap;
3512 
3513 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3514 	if (rc == 0)
3515 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3516 #endif
3517 }
3518 
3519 void
3520 vop_unlock_pre(void *ap)
3521 {
3522 #ifdef DEBUG_VFS_LOCKS
3523 	struct vop_unlock_args *a = ap;
3524 
3525 	if (a->a_flags & LK_INTERLOCK)
3526 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3527 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3528 #endif
3529 }
3530 
3531 void
3532 vop_unlock_post(void *ap, int rc)
3533 {
3534 #ifdef DEBUG_VFS_LOCKS
3535 	struct vop_unlock_args *a = ap;
3536 
3537 	if (a->a_flags & LK_INTERLOCK)
3538 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3539 #endif
3540 }
3541 
3542 void
3543 vop_create_post(void *ap, int rc)
3544 {
3545 	struct vop_create_args *a = ap;
3546 
3547 	if (!rc)
3548 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3549 }
3550 
3551 void
3552 vop_link_post(void *ap, int rc)
3553 {
3554 	struct vop_link_args *a = ap;
3555 
3556 	if (!rc) {
3557 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3558 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3559 	}
3560 }
3561 
3562 void
3563 vop_mkdir_post(void *ap, int rc)
3564 {
3565 	struct vop_mkdir_args *a = ap;
3566 
3567 	if (!rc)
3568 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3569 }
3570 
3571 void
3572 vop_mknod_post(void *ap, int rc)
3573 {
3574 	struct vop_mknod_args *a = ap;
3575 
3576 	if (!rc)
3577 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3578 }
3579 
3580 void
3581 vop_remove_post(void *ap, int rc)
3582 {
3583 	struct vop_remove_args *a = ap;
3584 
3585 	if (!rc) {
3586 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3587 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3588 	}
3589 }
3590 
3591 void
3592 vop_rename_post(void *ap, int rc)
3593 {
3594 	struct vop_rename_args *a = ap;
3595 
3596 	if (!rc) {
3597 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3598 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3599 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3600 		if (a->a_tvp)
3601 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3602 	}
3603 	if (a->a_tdvp != a->a_fdvp)
3604 		vdrop(a->a_fdvp);
3605 	if (a->a_tvp != a->a_fvp)
3606 		vdrop(a->a_fvp);
3607 	vdrop(a->a_tdvp);
3608 	if (a->a_tvp)
3609 		vdrop(a->a_tvp);
3610 }
3611 
3612 void
3613 vop_rmdir_post(void *ap, int rc)
3614 {
3615 	struct vop_rmdir_args *a = ap;
3616 
3617 	if (!rc) {
3618 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3619 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3620 	}
3621 }
3622 
3623 void
3624 vop_setattr_post(void *ap, int rc)
3625 {
3626 	struct vop_setattr_args *a = ap;
3627 
3628 	if (!rc)
3629 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3630 }
3631 
3632 void
3633 vop_symlink_post(void *ap, int rc)
3634 {
3635 	struct vop_symlink_args *a = ap;
3636 
3637 	if (!rc)
3638 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3639 }
3640 
3641 static struct knlist fs_knlist;
3642 
3643 static void
3644 vfs_event_init(void *arg)
3645 {
3646 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3647 }
3648 /* XXX - correct order? */
3649 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3650 
3651 void
3652 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3653 {
3654 
3655 	KNOTE_UNLOCKED(&fs_knlist, event);
3656 }
3657 
3658 static int	filt_fsattach(struct knote *kn);
3659 static void	filt_fsdetach(struct knote *kn);
3660 static int	filt_fsevent(struct knote *kn, long hint);
3661 
3662 struct filterops fs_filtops =
3663 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3664 
3665 static int
3666 filt_fsattach(struct knote *kn)
3667 {
3668 
3669 	kn->kn_flags |= EV_CLEAR;
3670 	knlist_add(&fs_knlist, kn, 0);
3671 	return (0);
3672 }
3673 
3674 static void
3675 filt_fsdetach(struct knote *kn)
3676 {
3677 
3678 	knlist_remove(&fs_knlist, kn, 0);
3679 }
3680 
3681 static int
3682 filt_fsevent(struct knote *kn, long hint)
3683 {
3684 
3685 	kn->kn_fflags |= hint;
3686 	return (kn->kn_fflags != 0);
3687 }
3688 
3689 static int
3690 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3691 {
3692 	struct vfsidctl vc;
3693 	int error;
3694 	struct mount *mp;
3695 
3696 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3697 	if (error)
3698 		return (error);
3699 	if (vc.vc_vers != VFS_CTL_VERS1)
3700 		return (EINVAL);
3701 	mp = vfs_getvfs(&vc.vc_fsid);
3702 	if (mp == NULL)
3703 		return (ENOENT);
3704 	/* ensure that a specific sysctl goes to the right filesystem. */
3705 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3706 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3707 		return (EINVAL);
3708 	}
3709 	VCTLTOREQ(&vc, req);
3710 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3711 }
3712 
3713 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3714         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3715 
3716 /*
3717  * Function to initialize a va_filerev field sensibly.
3718  * XXX: Wouldn't a random number make a lot more sense ??
3719  */
3720 u_quad_t
3721 init_va_filerev(void)
3722 {
3723 	struct bintime bt;
3724 
3725 	getbinuptime(&bt);
3726 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3727 }
3728 
3729 static int	filt_vfsread(struct knote *kn, long hint);
3730 static int	filt_vfswrite(struct knote *kn, long hint);
3731 static int	filt_vfsvnode(struct knote *kn, long hint);
3732 static void	filt_vfsdetach(struct knote *kn);
3733 static struct filterops vfsread_filtops =
3734 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3735 static struct filterops vfswrite_filtops =
3736 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3737 static struct filterops vfsvnode_filtops =
3738 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3739 
3740 static void
3741 vfs_knllock(void *arg)
3742 {
3743 	struct vnode *vp = arg;
3744 
3745 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
3746 }
3747 
3748 static void
3749 vfs_knlunlock(void *arg)
3750 {
3751 	struct vnode *vp = arg;
3752 
3753 	VOP_UNLOCK(vp, 0, curthread);
3754 }
3755 
3756 static int
3757 vfs_knllocked(void *arg)
3758 {
3759 	struct vnode *vp = arg;
3760 
3761 	return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE);
3762 }
3763 
3764 int
3765 vfs_kqfilter(struct vop_kqfilter_args *ap)
3766 {
3767 	struct vnode *vp = ap->a_vp;
3768 	struct knote *kn = ap->a_kn;
3769 	struct knlist *knl;
3770 
3771 	switch (kn->kn_filter) {
3772 	case EVFILT_READ:
3773 		kn->kn_fop = &vfsread_filtops;
3774 		break;
3775 	case EVFILT_WRITE:
3776 		kn->kn_fop = &vfswrite_filtops;
3777 		break;
3778 	case EVFILT_VNODE:
3779 		kn->kn_fop = &vfsvnode_filtops;
3780 		break;
3781 	default:
3782 		return (EINVAL);
3783 	}
3784 
3785 	kn->kn_hook = (caddr_t)vp;
3786 
3787 	if (vp->v_pollinfo == NULL)
3788 		v_addpollinfo(vp);
3789 	if (vp->v_pollinfo == NULL)
3790 		return (ENOMEM);
3791 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3792 	knlist_add(knl, kn, 0);
3793 
3794 	return (0);
3795 }
3796 
3797 /*
3798  * Detach knote from vnode
3799  */
3800 static void
3801 filt_vfsdetach(struct knote *kn)
3802 {
3803 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3804 
3805 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3806 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3807 }
3808 
3809 /*ARGSUSED*/
3810 static int
3811 filt_vfsread(struct knote *kn, long hint)
3812 {
3813 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3814 	struct vattr va;
3815 
3816 	/*
3817 	 * filesystem is gone, so set the EOF flag and schedule
3818 	 * the knote for deletion.
3819 	 */
3820 	if (hint == NOTE_REVOKE) {
3821 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3822 		return (1);
3823 	}
3824 
3825 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3826 		return (0);
3827 
3828 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3829 	return (kn->kn_data != 0);
3830 }
3831 
3832 /*ARGSUSED*/
3833 static int
3834 filt_vfswrite(struct knote *kn, long hint)
3835 {
3836 	/*
3837 	 * filesystem is gone, so set the EOF flag and schedule
3838 	 * the knote for deletion.
3839 	 */
3840 	if (hint == NOTE_REVOKE)
3841 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3842 
3843 	kn->kn_data = 0;
3844 	return (1);
3845 }
3846 
3847 static int
3848 filt_vfsvnode(struct knote *kn, long hint)
3849 {
3850 	if (kn->kn_sfflags & hint)
3851 		kn->kn_fflags |= hint;
3852 	if (hint == NOTE_REVOKE) {
3853 		kn->kn_flags |= EV_EOF;
3854 		return (1);
3855 	}
3856 	return (kn->kn_fflags != 0);
3857 }
3858 
3859 int
3860 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3861 {
3862 	int error;
3863 
3864 	if (dp->d_reclen > ap->a_uio->uio_resid)
3865 		return (ENAMETOOLONG);
3866 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3867 	if (error) {
3868 		if (ap->a_ncookies != NULL) {
3869 			if (ap->a_cookies != NULL)
3870 				free(ap->a_cookies, M_TEMP);
3871 			ap->a_cookies = NULL;
3872 			*ap->a_ncookies = 0;
3873 		}
3874 		return (error);
3875 	}
3876 	if (ap->a_ncookies == NULL)
3877 		return (0);
3878 	*ap->a_cookies = realloc(*ap->a_cookies,
3879 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
3880 	(*ap->a_cookies)[*ap->a_ncookies] = off;
3881 	return (0);
3882 }
3883 
3884 /*
3885  * Mark for update the access time of the file if the filesystem
3886  * supports VA_MARK_ATIME.  This functionality is used by execve
3887  * and mmap, so we want to avoid the synchronous I/O implied by
3888  * directly setting va_atime for the sake of efficiency.
3889  */
3890 void
3891 vfs_mark_atime(struct vnode *vp, struct thread *td)
3892 {
3893 	struct vattr atimeattr;
3894 
3895 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
3896 		VATTR_NULL(&atimeattr);
3897 		atimeattr.va_vaflags |= VA_MARK_ATIME;
3898 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
3899 	}
3900 }
3901