xref: /freebsd/sys/kern/vfs_subr.c (revision 3d11b6c8f01e1fca5936a11d6996448467851a94)
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, "subr_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	v_upgrade_usecount(struct vnode *);
99 static void	vfree(struct vnode *);
100 static void	vnlru_free(int);
101 static void	vdestroy(struct vnode *);
102 static void	vgonel(struct vnode *);
103 static void	vfs_knllock(void *arg);
104 static void	vfs_knlunlock(void *arg);
105 static int	vfs_knllocked(void *arg);
106 
107 
108 /*
109  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
110  * build.  Without mpsafevm the buffer cache can not run Giant free.
111  */
112 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__) || \
113 	defined(__ia64__) || defined(__sparc64__)
114 int mpsafe_vfs = 1;
115 #else
116 int mpsafe_vfs;
117 #endif
118 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
119 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
120     "MPSAFE VFS");
121 
122 /*
123  * Number of vnodes in existence.  Increased whenever getnewvnode()
124  * allocates a new vnode, never decreased.
125  */
126 static unsigned long	numvnodes;
127 
128 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
129 
130 /*
131  * Conversion tables for conversion from vnode types to inode formats
132  * and back.
133  */
134 enum vtype iftovt_tab[16] = {
135 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
136 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
137 };
138 int vttoif_tab[10] = {
139 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
140 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
141 };
142 
143 /*
144  * List of vnodes that are ready for recycling.
145  */
146 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
147 
148 /*
149  * Free vnode target.  Free vnodes may simply be files which have been stat'd
150  * but not read.  This is somewhat common, and a small cache of such files
151  * should be kept to avoid recreation costs.
152  */
153 static u_long wantfreevnodes;
154 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
155 /* Number of vnodes in the free list. */
156 static u_long freevnodes;
157 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
158 
159 /*
160  * Various variables used for debugging the new implementation of
161  * reassignbuf().
162  * XXX these are probably of (very) limited utility now.
163  */
164 static int reassignbufcalls;
165 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
166 
167 /*
168  * Cache for the mount type id assigned to NFS.  This is used for
169  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
170  */
171 int	nfs_mount_type = -1;
172 
173 /* To keep more than one thread at a time from running vfs_getnewfsid */
174 static struct mtx mntid_mtx;
175 
176 /*
177  * Lock for any access to the following:
178  *	vnode_free_list
179  *	numvnodes
180  *	freevnodes
181  */
182 static struct mtx vnode_free_list_mtx;
183 
184 /* Publicly exported FS */
185 struct nfs_public nfs_pub;
186 
187 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
188 static uma_zone_t vnode_zone;
189 static uma_zone_t vnodepoll_zone;
190 
191 /* Set to 1 to print out reclaim of active vnodes */
192 int	prtactive;
193 
194 /*
195  * The workitem queue.
196  *
197  * It is useful to delay writes of file data and filesystem metadata
198  * for tens of seconds so that quickly created and deleted files need
199  * not waste disk bandwidth being created and removed. To realize this,
200  * we append vnodes to a "workitem" queue. When running with a soft
201  * updates implementation, most pending metadata dependencies should
202  * not wait for more than a few seconds. Thus, mounted on block devices
203  * are delayed only about a half the time that file data is delayed.
204  * Similarly, directory updates are more critical, so are only delayed
205  * about a third the time that file data is delayed. Thus, there are
206  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
207  * one each second (driven off the filesystem syncer process). The
208  * syncer_delayno variable indicates the next queue that is to be processed.
209  * Items that need to be processed soon are placed in this queue:
210  *
211  *	syncer_workitem_pending[syncer_delayno]
212  *
213  * A delay of fifteen seconds is done by placing the request fifteen
214  * entries later in the queue:
215  *
216  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
217  *
218  */
219 static int syncer_delayno;
220 static long syncer_mask;
221 LIST_HEAD(synclist, bufobj);
222 static struct synclist *syncer_workitem_pending;
223 /*
224  * The sync_mtx protects:
225  *	bo->bo_synclist
226  *	sync_vnode_count
227  *	syncer_delayno
228  *	syncer_state
229  *	syncer_workitem_pending
230  *	syncer_worklist_len
231  *	rushjob
232  */
233 static struct mtx sync_mtx;
234 
235 #define SYNCER_MAXDELAY		32
236 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
237 static int syncdelay = 30;		/* max time to delay syncing data */
238 static int filedelay = 30;		/* time to delay syncing files */
239 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
240 static int dirdelay = 29;		/* time to delay syncing directories */
241 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
242 static int metadelay = 28;		/* time to delay syncing metadata */
243 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
244 static int rushjob;		/* number of slots to run ASAP */
245 static int stat_rush_requests;	/* number of times I/O speeded up */
246 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
247 
248 /*
249  * When shutting down the syncer, run it at four times normal speed.
250  */
251 #define SYNCER_SHUTDOWN_SPEEDUP		4
252 static int sync_vnode_count;
253 static int syncer_worklist_len;
254 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
255     syncer_state;
256 
257 /*
258  * Number of vnodes we want to exist at any one time.  This is mostly used
259  * to size hash tables in vnode-related code.  It is normally not used in
260  * getnewvnode(), as wantfreevnodes is normally nonzero.)
261  *
262  * XXX desiredvnodes is historical cruft and should not exist.
263  */
264 int desiredvnodes;
265 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
266     &desiredvnodes, 0, "Maximum number of vnodes");
267 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
268     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
269 static int vnlru_nowhere;
270 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
271     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
272 
273 /*
274  * Macros to control when a vnode is freed and recycled.  All require
275  * the vnode interlock.
276  */
277 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
278 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
279 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
280 
281 
282 /*
283  * Initialize the vnode management data structures.
284  */
285 #ifndef	MAXVNODES_MAX
286 #define	MAXVNODES_MAX	100000
287 #endif
288 static void
289 vntblinit(void *dummy __unused)
290 {
291 
292 	/*
293 	 * Desiredvnodes is a function of the physical memory size and
294 	 * the kernel's heap size.  Specifically, desiredvnodes scales
295 	 * in proportion to the physical memory size until two fifths
296 	 * of the kernel's heap size is consumed by vnodes and vm
297 	 * objects.
298 	 */
299 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
300 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
301 	if (desiredvnodes > MAXVNODES_MAX) {
302 		if (bootverbose)
303 			printf("Reducing kern.maxvnodes %d -> %d\n",
304 			    desiredvnodes, MAXVNODES_MAX);
305 		desiredvnodes = MAXVNODES_MAX;
306 	}
307 	wantfreevnodes = desiredvnodes / 4;
308 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
309 	TAILQ_INIT(&vnode_free_list);
310 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
311 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
312 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
314 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 	/*
316 	 * Initialize the filesystem syncer.
317 	 */
318 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
319 		&syncer_mask);
320 	syncer_maxdelay = syncer_mask + 1;
321 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
322 }
323 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
324 
325 
326 /*
327  * Mark a mount point as busy. Used to synchronize access and to delay
328  * unmounting. Interlock is not released on failure.
329  */
330 int
331 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp,
332     struct thread *td)
333 {
334 	int lkflags;
335 
336 	MNT_ILOCK(mp);
337 	MNT_REF(mp);
338 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
339 		if (flags & LK_NOWAIT) {
340 			MNT_REL(mp);
341 			MNT_IUNLOCK(mp);
342 			return (ENOENT);
343 		}
344 		if (interlkp)
345 			mtx_unlock(interlkp);
346 		mp->mnt_kern_flag |= MNTK_MWAIT;
347 		/*
348 		 * Since all busy locks are shared except the exclusive
349 		 * lock granted when unmounting, the only place that a
350 		 * wakeup needs to be done is at the release of the
351 		 * exclusive lock at the end of dounmount.
352 		 */
353 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
354 		MNT_REL(mp);
355 		MNT_IUNLOCK(mp);
356 		if (interlkp)
357 			mtx_lock(interlkp);
358 		return (ENOENT);
359 	}
360 	if (interlkp)
361 		mtx_unlock(interlkp);
362 	lkflags = LK_SHARED | LK_INTERLOCK;
363 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
364 		panic("vfs_busy: unexpected lock failure");
365 	return (0);
366 }
367 
368 /*
369  * Free a busy filesystem.
370  */
371 void
372 vfs_unbusy(struct mount *mp, struct thread *td)
373 {
374 
375 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
376 	vfs_rel(mp);
377 }
378 
379 /*
380  * Lookup a mount point by filesystem identifier.
381  */
382 struct mount *
383 vfs_getvfs(fsid_t *fsid)
384 {
385 	struct mount *mp;
386 
387 	mtx_lock(&mountlist_mtx);
388 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
389 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
390 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
391 			vfs_ref(mp);
392 			mtx_unlock(&mountlist_mtx);
393 			return (mp);
394 		}
395 	}
396 	mtx_unlock(&mountlist_mtx);
397 	return ((struct mount *) 0);
398 }
399 
400 /*
401  * Check if a user can access priveledged mount options.
402  */
403 int
404 vfs_suser(struct mount *mp, struct thread *td)
405 {
406 	int error;
407 
408 	if ((mp->mnt_flag & MNT_USER) == 0 ||
409 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
410 		if ((error = suser(td)) != 0)
411 			return (error);
412 	}
413 	return (0);
414 }
415 
416 /*
417  * Get a new unique fsid.  Try to make its val[0] unique, since this value
418  * will be used to create fake device numbers for stat().  Also try (but
419  * not so hard) make its val[0] unique mod 2^16, since some emulators only
420  * support 16-bit device numbers.  We end up with unique val[0]'s for the
421  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
422  *
423  * Keep in mind that several mounts may be running in parallel.  Starting
424  * the search one past where the previous search terminated is both a
425  * micro-optimization and a defense against returning the same fsid to
426  * different mounts.
427  */
428 void
429 vfs_getnewfsid(struct mount *mp)
430 {
431 	static u_int16_t mntid_base;
432 	struct mount *nmp;
433 	fsid_t tfsid;
434 	int mtype;
435 
436 	mtx_lock(&mntid_mtx);
437 	mtype = mp->mnt_vfc->vfc_typenum;
438 	tfsid.val[1] = mtype;
439 	mtype = (mtype & 0xFF) << 24;
440 	for (;;) {
441 		tfsid.val[0] = makedev(255,
442 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
443 		mntid_base++;
444 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
445 			break;
446 		vfs_rel(nmp);
447 	}
448 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
449 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
450 	mtx_unlock(&mntid_mtx);
451 }
452 
453 /*
454  * Knob to control the precision of file timestamps:
455  *
456  *   0 = seconds only; nanoseconds zeroed.
457  *   1 = seconds and nanoseconds, accurate within 1/HZ.
458  *   2 = seconds and nanoseconds, truncated to microseconds.
459  * >=3 = seconds and nanoseconds, maximum precision.
460  */
461 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
462 
463 static int timestamp_precision = TSP_SEC;
464 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
465     &timestamp_precision, 0, "");
466 
467 /*
468  * Get a current timestamp.
469  */
470 void
471 vfs_timestamp(struct timespec *tsp)
472 {
473 	struct timeval tv;
474 
475 	switch (timestamp_precision) {
476 	case TSP_SEC:
477 		tsp->tv_sec = time_second;
478 		tsp->tv_nsec = 0;
479 		break;
480 	case TSP_HZ:
481 		getnanotime(tsp);
482 		break;
483 	case TSP_USEC:
484 		microtime(&tv);
485 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
486 		break;
487 	case TSP_NSEC:
488 	default:
489 		nanotime(tsp);
490 		break;
491 	}
492 }
493 
494 /*
495  * Set vnode attributes to VNOVAL
496  */
497 void
498 vattr_null(struct vattr *vap)
499 {
500 
501 	vap->va_type = VNON;
502 	vap->va_size = VNOVAL;
503 	vap->va_bytes = VNOVAL;
504 	vap->va_mode = VNOVAL;
505 	vap->va_nlink = VNOVAL;
506 	vap->va_uid = VNOVAL;
507 	vap->va_gid = VNOVAL;
508 	vap->va_fsid = VNOVAL;
509 	vap->va_fileid = VNOVAL;
510 	vap->va_blocksize = VNOVAL;
511 	vap->va_rdev = VNOVAL;
512 	vap->va_atime.tv_sec = VNOVAL;
513 	vap->va_atime.tv_nsec = VNOVAL;
514 	vap->va_mtime.tv_sec = VNOVAL;
515 	vap->va_mtime.tv_nsec = VNOVAL;
516 	vap->va_ctime.tv_sec = VNOVAL;
517 	vap->va_ctime.tv_nsec = VNOVAL;
518 	vap->va_birthtime.tv_sec = VNOVAL;
519 	vap->va_birthtime.tv_nsec = VNOVAL;
520 	vap->va_flags = VNOVAL;
521 	vap->va_gen = VNOVAL;
522 	vap->va_vaflags = 0;
523 }
524 
525 /*
526  * This routine is called when we have too many vnodes.  It attempts
527  * to free <count> vnodes and will potentially free vnodes that still
528  * have VM backing store (VM backing store is typically the cause
529  * of a vnode blowout so we want to do this).  Therefore, this operation
530  * is not considered cheap.
531  *
532  * A number of conditions may prevent a vnode from being reclaimed.
533  * the buffer cache may have references on the vnode, a directory
534  * vnode may still have references due to the namei cache representing
535  * underlying files, or the vnode may be in active use.   It is not
536  * desireable to reuse such vnodes.  These conditions may cause the
537  * number of vnodes to reach some minimum value regardless of what
538  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
539  */
540 static int
541 vlrureclaim(struct mount *mp)
542 {
543 	struct thread *td;
544 	struct vnode *vp;
545 	int done;
546 	int trigger;
547 	int usevnodes;
548 	int count;
549 
550 	/*
551 	 * Calculate the trigger point, don't allow user
552 	 * screwups to blow us up.   This prevents us from
553 	 * recycling vnodes with lots of resident pages.  We
554 	 * aren't trying to free memory, we are trying to
555 	 * free vnodes.
556 	 */
557 	usevnodes = desiredvnodes;
558 	if (usevnodes <= 0)
559 		usevnodes = 1;
560 	trigger = cnt.v_page_count * 2 / usevnodes;
561 	done = 0;
562 	td = curthread;
563 	vn_start_write(NULL, &mp, V_WAIT);
564 	MNT_ILOCK(mp);
565 	count = mp->mnt_nvnodelistsize / 10 + 1;
566 	while (count != 0) {
567 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
568 		while (vp != NULL && vp->v_type == VMARKER)
569 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
570 		if (vp == NULL)
571 			break;
572 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
573 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
574 		--count;
575 		if (!VI_TRYLOCK(vp))
576 			goto next_iter;
577 		/*
578 		 * If it's been deconstructed already, it's still
579 		 * referenced, or it exceeds the trigger, skip it.
580 		 */
581 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
582 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
583 		    vp->v_object->resident_page_count > trigger)) {
584 			VI_UNLOCK(vp);
585 			goto next_iter;
586 		}
587 		MNT_IUNLOCK(mp);
588 		vholdl(vp);
589 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
590 			vdrop(vp);
591 			goto next_iter_mntunlocked;
592 		}
593 		VI_LOCK(vp);
594 		/*
595 		 * v_usecount may have been bumped after VOP_LOCK() dropped
596 		 * the vnode interlock and before it was locked again.
597 		 *
598 		 * It is not necessary to recheck VI_DOOMED because it can
599 		 * only be set by another thread that holds both the vnode
600 		 * lock and vnode interlock.  If another thread has the
601 		 * vnode lock before we get to VOP_LOCK() and obtains the
602 		 * vnode interlock after VOP_LOCK() drops the vnode
603 		 * interlock, the other thread will be unable to drop the
604 		 * vnode lock before our VOP_LOCK() call fails.
605 		 */
606 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
607 		    (vp->v_object != NULL &&
608 		    vp->v_object->resident_page_count > trigger)) {
609 			VOP_UNLOCK(vp, LK_INTERLOCK, td);
610 			goto next_iter_mntunlocked;
611 		}
612 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
613 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
614 		vgonel(vp);
615 		VOP_UNLOCK(vp, 0, td);
616 		vdropl(vp);
617 		done++;
618 next_iter_mntunlocked:
619 		if ((count % 256) != 0)
620 			goto relock_mnt;
621 		goto yield;
622 next_iter:
623 		if ((count % 256) != 0)
624 			continue;
625 		MNT_IUNLOCK(mp);
626 yield:
627 		uio_yield();
628 relock_mnt:
629 		MNT_ILOCK(mp);
630 	}
631 	MNT_IUNLOCK(mp);
632 	vn_finished_write(mp);
633 	return done;
634 }
635 
636 /*
637  * Attempt to keep the free list at wantfreevnodes length.
638  */
639 static void
640 vnlru_free(int count)
641 {
642 	struct vnode *vp;
643 	int vfslocked;
644 
645 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
646 	for (; count > 0; count--) {
647 		vp = TAILQ_FIRST(&vnode_free_list);
648 		/*
649 		 * The list can be modified while the free_list_mtx
650 		 * has been dropped and vp could be NULL here.
651 		 */
652 		if (!vp)
653 			break;
654 		VNASSERT(vp->v_op != NULL, vp,
655 		    ("vnlru_free: vnode already reclaimed."));
656 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
657 		/*
658 		 * Don't recycle if we can't get the interlock.
659 		 */
660 		if (!VI_TRYLOCK(vp)) {
661 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
662 			continue;
663 		}
664 		VNASSERT(VCANRECYCLE(vp), vp,
665 		    ("vp inconsistent on freelist"));
666 		freevnodes--;
667 		vp->v_iflag &= ~VI_FREE;
668 		vholdl(vp);
669 		mtx_unlock(&vnode_free_list_mtx);
670 		VI_UNLOCK(vp);
671 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
672 		vtryrecycle(vp);
673 		VFS_UNLOCK_GIANT(vfslocked);
674 		/*
675 		 * If the recycled succeeded this vdrop will actually free
676 		 * the vnode.  If not it will simply place it back on
677 		 * the free list.
678 		 */
679 		vdrop(vp);
680 		mtx_lock(&vnode_free_list_mtx);
681 	}
682 }
683 /*
684  * Attempt to recycle vnodes in a context that is always safe to block.
685  * Calling vlrurecycle() from the bowels of filesystem code has some
686  * interesting deadlock problems.
687  */
688 static struct proc *vnlruproc;
689 static int vnlruproc_sig;
690 
691 static void
692 vnlru_proc(void)
693 {
694 	struct mount *mp, *nmp;
695 	int done;
696 	struct proc *p = vnlruproc;
697 	struct thread *td = FIRST_THREAD_IN_PROC(p);
698 
699 	mtx_lock(&Giant);
700 
701 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
702 	    SHUTDOWN_PRI_FIRST);
703 
704 	for (;;) {
705 		kthread_suspend_check(p);
706 		mtx_lock(&vnode_free_list_mtx);
707 		if (freevnodes > wantfreevnodes)
708 			vnlru_free(freevnodes - wantfreevnodes);
709 		if (numvnodes <= desiredvnodes * 9 / 10) {
710 			vnlruproc_sig = 0;
711 			wakeup(&vnlruproc_sig);
712 			msleep(vnlruproc, &vnode_free_list_mtx,
713 			    PVFS|PDROP, "vlruwt", hz);
714 			continue;
715 		}
716 		mtx_unlock(&vnode_free_list_mtx);
717 		done = 0;
718 		mtx_lock(&mountlist_mtx);
719 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
720 			int vfsunlocked;
721 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
722 				nmp = TAILQ_NEXT(mp, mnt_list);
723 				continue;
724 			}
725 			if (!VFS_NEEDSGIANT(mp)) {
726 				mtx_unlock(&Giant);
727 				vfsunlocked = 1;
728 			} else
729 				vfsunlocked = 0;
730 			done += vlrureclaim(mp);
731 			if (vfsunlocked)
732 				mtx_lock(&Giant);
733 			mtx_lock(&mountlist_mtx);
734 			nmp = TAILQ_NEXT(mp, mnt_list);
735 			vfs_unbusy(mp, td);
736 		}
737 		mtx_unlock(&mountlist_mtx);
738 		if (done == 0) {
739 #if 0
740 			/* These messages are temporary debugging aids */
741 			if (vnlru_nowhere < 5)
742 				printf("vnlru process getting nowhere..\n");
743 			else if (vnlru_nowhere == 5)
744 				printf("vnlru process messages stopped.\n");
745 #endif
746 			vnlru_nowhere++;
747 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
748 		} else
749 			uio_yield();
750 	}
751 }
752 
753 static struct kproc_desc vnlru_kp = {
754 	"vnlru",
755 	vnlru_proc,
756 	&vnlruproc
757 };
758 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
759 
760 /*
761  * Routines having to do with the management of the vnode table.
762  */
763 
764 static void
765 vdestroy(struct vnode *vp)
766 {
767 	struct bufobj *bo;
768 
769 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
770 	mtx_lock(&vnode_free_list_mtx);
771 	numvnodes--;
772 	mtx_unlock(&vnode_free_list_mtx);
773 	bo = &vp->v_bufobj;
774 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
775 	    ("cleaned vnode still on the free list."));
776 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
777 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
778 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
779 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
780 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
781 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
782 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
783 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
784 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
785 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
786 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
787 	VI_UNLOCK(vp);
788 #ifdef MAC
789 	mac_destroy_vnode(vp);
790 #endif
791 	if (vp->v_pollinfo != NULL) {
792 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
793 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
794 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
795 	}
796 #ifdef INVARIANTS
797 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
798 	vp->v_op = NULL;
799 #endif
800 	lockdestroy(vp->v_vnlock);
801 	mtx_destroy(&vp->v_interlock);
802 	uma_zfree(vnode_zone, vp);
803 }
804 
805 /*
806  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
807  * before we actually vgone().  This function must be called with the vnode
808  * held to prevent the vnode from being returned to the free list midway
809  * through vgone().
810  */
811 static int
812 vtryrecycle(struct vnode *vp)
813 {
814 	struct thread *td = curthread;
815 	struct mount *vnmp;
816 
817 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
818 	VNASSERT(vp->v_holdcnt, vp,
819 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
820 	/*
821 	 * This vnode may found and locked via some other list, if so we
822 	 * can't recycle it yet.
823 	 */
824 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
825 		return (EWOULDBLOCK);
826 	/*
827 	 * Don't recycle if its filesystem is being suspended.
828 	 */
829 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
830 		VOP_UNLOCK(vp, 0, td);
831 		return (EBUSY);
832 	}
833 	/*
834 	 * If we got this far, we need to acquire the interlock and see if
835 	 * anyone picked up this vnode from another list.  If not, we will
836 	 * mark it with DOOMED via vgonel() so that anyone who does find it
837 	 * will skip over it.
838 	 */
839 	VI_LOCK(vp);
840 	if (vp->v_usecount) {
841 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
842 		vn_finished_write(vnmp);
843 		return (EBUSY);
844 	}
845 	if ((vp->v_iflag & VI_DOOMED) == 0)
846 		vgonel(vp);
847 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
848 	vn_finished_write(vnmp);
849 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
850 	return (0);
851 }
852 
853 /*
854  * Return the next vnode from the free list.
855  */
856 int
857 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
858     struct vnode **vpp)
859 {
860 	struct vnode *vp = NULL;
861 	struct bufobj *bo;
862 
863 	mtx_lock(&vnode_free_list_mtx);
864 	/*
865 	 * Lend our context to reclaim vnodes if they've exceeded the max.
866 	 */
867 	if (freevnodes > wantfreevnodes)
868 		vnlru_free(1);
869 	/*
870 	 * Wait for available vnodes.
871 	 */
872 	if (numvnodes > desiredvnodes) {
873 		if (vnlruproc_sig == 0) {
874 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
875 			wakeup(vnlruproc);
876 		}
877 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
878 		    "vlruwk", hz);
879 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
880 		if (numvnodes > desiredvnodes) {
881 			mtx_unlock(&vnode_free_list_mtx);
882 			return (ENFILE);
883 		}
884 #endif
885 	}
886 	numvnodes++;
887 	mtx_unlock(&vnode_free_list_mtx);
888 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
889 	/*
890 	 * Setup locks.
891 	 */
892 	vp->v_vnlock = &vp->v_lock;
893 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
894 	/*
895 	 * By default, don't allow shared locks unless filesystems
896 	 * opt-in.
897 	 */
898 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
899 	/*
900 	 * Initialize bufobj.
901 	 */
902 	bo = &vp->v_bufobj;
903 	bo->__bo_vnode = vp;
904 	bo->bo_mtx = &vp->v_interlock;
905 	bo->bo_ops = &buf_ops_bio;
906 	bo->bo_private = vp;
907 	TAILQ_INIT(&bo->bo_clean.bv_hd);
908 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
909 	/*
910 	 * Initialize namecache.
911 	 */
912 	LIST_INIT(&vp->v_cache_src);
913 	TAILQ_INIT(&vp->v_cache_dst);
914 	/*
915 	 * Finalize various vnode identity bits.
916 	 */
917 	vp->v_type = VNON;
918 	vp->v_tag = tag;
919 	vp->v_op = vops;
920 	v_incr_usecount(vp);
921 	vp->v_data = 0;
922 #ifdef MAC
923 	mac_init_vnode(vp);
924 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
925 		mac_associate_vnode_singlelabel(mp, vp);
926 	else if (mp == NULL)
927 		printf("NULL mp in getnewvnode()\n");
928 #endif
929 	if (mp != NULL) {
930 		insmntque(vp, mp);
931 		bo->bo_bsize = mp->mnt_stat.f_iosize;
932 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
933 			vp->v_vflag |= VV_NOKNOTE;
934 	}
935 
936 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
937 	*vpp = vp;
938 	return (0);
939 }
940 
941 /*
942  * Delete from old mount point vnode list, if on one.
943  */
944 static void
945 delmntque(struct vnode *vp)
946 {
947 	struct mount *mp;
948 
949 	mp = vp->v_mount;
950 	if (mp == NULL)
951 		return;
952 	MNT_ILOCK(mp);
953 	vp->v_mount = NULL;
954 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
955 		("bad mount point vnode list size"));
956 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
957 	mp->mnt_nvnodelistsize--;
958 	MNT_REL(mp);
959 	MNT_IUNLOCK(mp);
960 }
961 
962 /*
963  * Insert into list of vnodes for the new mount point, if available.
964  */
965 static void
966 insmntque(struct vnode *vp, struct mount *mp)
967 {
968 
969 	vp->v_mount = mp;
970 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
971 	MNT_ILOCK(mp);
972 	MNT_REF(mp);
973 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
974 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
975 		("neg mount point vnode list size"));
976 	mp->mnt_nvnodelistsize++;
977 	MNT_IUNLOCK(mp);
978 }
979 
980 /*
981  * Flush out and invalidate all buffers associated with a bufobj
982  * Called with the underlying object locked.
983  */
984 int
985 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag,
986     int slptimeo)
987 {
988 	int error;
989 
990 	BO_LOCK(bo);
991 	if (flags & V_SAVE) {
992 		error = bufobj_wwait(bo, slpflag, slptimeo);
993 		if (error) {
994 			BO_UNLOCK(bo);
995 			return (error);
996 		}
997 		if (bo->bo_dirty.bv_cnt > 0) {
998 			BO_UNLOCK(bo);
999 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
1000 				return (error);
1001 			/*
1002 			 * XXX We could save a lock/unlock if this was only
1003 			 * enabled under INVARIANTS
1004 			 */
1005 			BO_LOCK(bo);
1006 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1007 				panic("vinvalbuf: dirty bufs");
1008 		}
1009 	}
1010 	/*
1011 	 * If you alter this loop please notice that interlock is dropped and
1012 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1013 	 * no race conditions occur from this.
1014 	 */
1015 	do {
1016 		error = flushbuflist(&bo->bo_clean,
1017 		    flags, bo, slpflag, slptimeo);
1018 		if (error == 0)
1019 			error = flushbuflist(&bo->bo_dirty,
1020 			    flags, bo, slpflag, slptimeo);
1021 		if (error != 0 && error != EAGAIN) {
1022 			BO_UNLOCK(bo);
1023 			return (error);
1024 		}
1025 	} while (error != 0);
1026 
1027 	/*
1028 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1029 	 * have write I/O in-progress but if there is a VM object then the
1030 	 * VM object can also have read-I/O in-progress.
1031 	 */
1032 	do {
1033 		bufobj_wwait(bo, 0, 0);
1034 		BO_UNLOCK(bo);
1035 		if (bo->bo_object != NULL) {
1036 			VM_OBJECT_LOCK(bo->bo_object);
1037 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1038 			VM_OBJECT_UNLOCK(bo->bo_object);
1039 		}
1040 		BO_LOCK(bo);
1041 	} while (bo->bo_numoutput > 0);
1042 	BO_UNLOCK(bo);
1043 
1044 	/*
1045 	 * Destroy the copy in the VM cache, too.
1046 	 */
1047 	if (bo->bo_object != NULL) {
1048 		VM_OBJECT_LOCK(bo->bo_object);
1049 		vm_object_page_remove(bo->bo_object, 0, 0,
1050 			(flags & V_SAVE) ? TRUE : FALSE);
1051 		VM_OBJECT_UNLOCK(bo->bo_object);
1052 	}
1053 
1054 #ifdef INVARIANTS
1055 	BO_LOCK(bo);
1056 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1057 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1058 		panic("vinvalbuf: flush failed");
1059 	BO_UNLOCK(bo);
1060 #endif
1061 	return (0);
1062 }
1063 
1064 /*
1065  * Flush out and invalidate all buffers associated with a vnode.
1066  * Called with the underlying object locked.
1067  */
1068 int
1069 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag,
1070     int slptimeo)
1071 {
1072 
1073 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1074 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1075 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1076 }
1077 
1078 /*
1079  * Flush out buffers on the specified list.
1080  *
1081  */
1082 static int
1083 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1084     int slptimeo)
1085 {
1086 	struct buf *bp, *nbp;
1087 	int retval, error;
1088 	daddr_t lblkno;
1089 	b_xflags_t xflags;
1090 
1091 	ASSERT_BO_LOCKED(bo);
1092 
1093 	retval = 0;
1094 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1095 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1096 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1097 			continue;
1098 		}
1099 		lblkno = 0;
1100 		xflags = 0;
1101 		if (nbp != NULL) {
1102 			lblkno = nbp->b_lblkno;
1103 			xflags = nbp->b_xflags &
1104 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1105 		}
1106 		retval = EAGAIN;
1107 		error = BUF_TIMELOCK(bp,
1108 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1109 		    "flushbuf", slpflag, slptimeo);
1110 		if (error) {
1111 			BO_LOCK(bo);
1112 			return (error != ENOLCK ? error : EAGAIN);
1113 		}
1114 		KASSERT(bp->b_bufobj == bo,
1115 	            ("bp %p wrong b_bufobj %p should be %p",
1116 		    bp, bp->b_bufobj, bo));
1117 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1118 			BUF_UNLOCK(bp);
1119 			BO_LOCK(bo);
1120 			return (EAGAIN);
1121 		}
1122 		/*
1123 		 * XXX Since there are no node locks for NFS, I
1124 		 * believe there is a slight chance that a delayed
1125 		 * write will occur while sleeping just above, so
1126 		 * check for it.
1127 		 */
1128 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1129 		    (flags & V_SAVE)) {
1130 			bremfree(bp);
1131 			bp->b_flags |= B_ASYNC;
1132 			bwrite(bp);
1133 			BO_LOCK(bo);
1134 			return (EAGAIN);	/* XXX: why not loop ? */
1135 		}
1136 		bremfree(bp);
1137 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1138 		bp->b_flags &= ~B_ASYNC;
1139 		brelse(bp);
1140 		BO_LOCK(bo);
1141 		if (nbp != NULL &&
1142 		    (nbp->b_bufobj != bo ||
1143 		     nbp->b_lblkno != lblkno ||
1144 		     (nbp->b_xflags &
1145 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1146 			break;			/* nbp invalid */
1147 	}
1148 	return (retval);
1149 }
1150 
1151 /*
1152  * Truncate a file's buffer and pages to a specified length.  This
1153  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1154  * sync activity.
1155  */
1156 int
1157 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1158     off_t length, int blksize)
1159 {
1160 	struct buf *bp, *nbp;
1161 	int anyfreed;
1162 	int trunclbn;
1163 	struct bufobj *bo;
1164 
1165 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1166 	/*
1167 	 * Round up to the *next* lbn.
1168 	 */
1169 	trunclbn = (length + blksize - 1) / blksize;
1170 
1171 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1172 restart:
1173 	VI_LOCK(vp);
1174 	bo = &vp->v_bufobj;
1175 	anyfreed = 1;
1176 	for (;anyfreed;) {
1177 		anyfreed = 0;
1178 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1179 			if (bp->b_lblkno < trunclbn)
1180 				continue;
1181 			if (BUF_LOCK(bp,
1182 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1183 			    VI_MTX(vp)) == ENOLCK)
1184 				goto restart;
1185 
1186 			bremfree(bp);
1187 			bp->b_flags |= (B_INVAL | B_RELBUF);
1188 			bp->b_flags &= ~B_ASYNC;
1189 			brelse(bp);
1190 			anyfreed = 1;
1191 
1192 			if (nbp != NULL &&
1193 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1194 			    (nbp->b_vp != vp) ||
1195 			    (nbp->b_flags & B_DELWRI))) {
1196 				goto restart;
1197 			}
1198 			VI_LOCK(vp);
1199 		}
1200 
1201 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1202 			if (bp->b_lblkno < trunclbn)
1203 				continue;
1204 			if (BUF_LOCK(bp,
1205 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1206 			    VI_MTX(vp)) == ENOLCK)
1207 				goto restart;
1208 			bremfree(bp);
1209 			bp->b_flags |= (B_INVAL | B_RELBUF);
1210 			bp->b_flags &= ~B_ASYNC;
1211 			brelse(bp);
1212 			anyfreed = 1;
1213 			if (nbp != NULL &&
1214 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1215 			    (nbp->b_vp != vp) ||
1216 			    (nbp->b_flags & B_DELWRI) == 0)) {
1217 				goto restart;
1218 			}
1219 			VI_LOCK(vp);
1220 		}
1221 	}
1222 
1223 	if (length > 0) {
1224 restartsync:
1225 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1226 			if (bp->b_lblkno > 0)
1227 				continue;
1228 			/*
1229 			 * Since we hold the vnode lock this should only
1230 			 * fail if we're racing with the buf daemon.
1231 			 */
1232 			if (BUF_LOCK(bp,
1233 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1234 			    VI_MTX(vp)) == ENOLCK) {
1235 				goto restart;
1236 			}
1237 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1238 			    ("buf(%p) on dirty queue without DELWRI", bp));
1239 
1240 			bremfree(bp);
1241 			bawrite(bp);
1242 			VI_LOCK(vp);
1243 			goto restartsync;
1244 		}
1245 	}
1246 
1247 	bufobj_wwait(bo, 0, 0);
1248 	VI_UNLOCK(vp);
1249 	vnode_pager_setsize(vp, length);
1250 
1251 	return (0);
1252 }
1253 
1254 /*
1255  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1256  * 		 a vnode.
1257  *
1258  *	NOTE: We have to deal with the special case of a background bitmap
1259  *	buffer, a situation where two buffers will have the same logical
1260  *	block offset.  We want (1) only the foreground buffer to be accessed
1261  *	in a lookup and (2) must differentiate between the foreground and
1262  *	background buffer in the splay tree algorithm because the splay
1263  *	tree cannot normally handle multiple entities with the same 'index'.
1264  *	We accomplish this by adding differentiating flags to the splay tree's
1265  *	numerical domain.
1266  */
1267 static
1268 struct buf *
1269 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1270 {
1271 	struct buf dummy;
1272 	struct buf *lefttreemax, *righttreemin, *y;
1273 
1274 	if (root == NULL)
1275 		return (NULL);
1276 	lefttreemax = righttreemin = &dummy;
1277 	for (;;) {
1278 		if (lblkno < root->b_lblkno ||
1279 		    (lblkno == root->b_lblkno &&
1280 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1281 			if ((y = root->b_left) == NULL)
1282 				break;
1283 			if (lblkno < y->b_lblkno) {
1284 				/* Rotate right. */
1285 				root->b_left = y->b_right;
1286 				y->b_right = root;
1287 				root = y;
1288 				if ((y = root->b_left) == NULL)
1289 					break;
1290 			}
1291 			/* Link into the new root's right tree. */
1292 			righttreemin->b_left = root;
1293 			righttreemin = root;
1294 		} else if (lblkno > root->b_lblkno ||
1295 		    (lblkno == root->b_lblkno &&
1296 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1297 			if ((y = root->b_right) == NULL)
1298 				break;
1299 			if (lblkno > y->b_lblkno) {
1300 				/* Rotate left. */
1301 				root->b_right = y->b_left;
1302 				y->b_left = root;
1303 				root = y;
1304 				if ((y = root->b_right) == NULL)
1305 					break;
1306 			}
1307 			/* Link into the new root's left tree. */
1308 			lefttreemax->b_right = root;
1309 			lefttreemax = root;
1310 		} else {
1311 			break;
1312 		}
1313 		root = y;
1314 	}
1315 	/* Assemble the new root. */
1316 	lefttreemax->b_right = root->b_left;
1317 	righttreemin->b_left = root->b_right;
1318 	root->b_left = dummy.b_right;
1319 	root->b_right = dummy.b_left;
1320 	return (root);
1321 }
1322 
1323 static void
1324 buf_vlist_remove(struct buf *bp)
1325 {
1326 	struct buf *root;
1327 	struct bufv *bv;
1328 
1329 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1330 	ASSERT_BO_LOCKED(bp->b_bufobj);
1331 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1332 	    (BX_VNDIRTY|BX_VNCLEAN),
1333 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1334 	if (bp->b_xflags & BX_VNDIRTY)
1335 		bv = &bp->b_bufobj->bo_dirty;
1336 	else
1337 		bv = &bp->b_bufobj->bo_clean;
1338 	if (bp != bv->bv_root) {
1339 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1340 		KASSERT(root == bp, ("splay lookup failed in remove"));
1341 	}
1342 	if (bp->b_left == NULL) {
1343 		root = bp->b_right;
1344 	} else {
1345 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1346 		root->b_right = bp->b_right;
1347 	}
1348 	bv->bv_root = root;
1349 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1350 	bv->bv_cnt--;
1351 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1352 }
1353 
1354 /*
1355  * Add the buffer to the sorted clean or dirty block list using a
1356  * splay tree algorithm.
1357  *
1358  * NOTE: xflags is passed as a constant, optimizing this inline function!
1359  */
1360 static void
1361 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1362 {
1363 	struct buf *root;
1364 	struct bufv *bv;
1365 
1366 	ASSERT_BO_LOCKED(bo);
1367 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1368 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1369 	bp->b_xflags |= xflags;
1370 	if (xflags & BX_VNDIRTY)
1371 		bv = &bo->bo_dirty;
1372 	else
1373 		bv = &bo->bo_clean;
1374 
1375 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1376 	if (root == NULL) {
1377 		bp->b_left = NULL;
1378 		bp->b_right = NULL;
1379 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1380 	} else if (bp->b_lblkno < root->b_lblkno ||
1381 	    (bp->b_lblkno == root->b_lblkno &&
1382 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1383 		bp->b_left = root->b_left;
1384 		bp->b_right = root;
1385 		root->b_left = NULL;
1386 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1387 	} else {
1388 		bp->b_right = root->b_right;
1389 		bp->b_left = root;
1390 		root->b_right = NULL;
1391 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1392 	}
1393 	bv->bv_cnt++;
1394 	bv->bv_root = bp;
1395 }
1396 
1397 /*
1398  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1399  * shadow buffers used in background bitmap writes.
1400  *
1401  * This code isn't quite efficient as it could be because we are maintaining
1402  * two sorted lists and do not know which list the block resides in.
1403  *
1404  * During a "make buildworld" the desired buffer is found at one of
1405  * the roots more than 60% of the time.  Thus, checking both roots
1406  * before performing either splay eliminates unnecessary splays on the
1407  * first tree splayed.
1408  */
1409 struct buf *
1410 gbincore(struct bufobj *bo, daddr_t lblkno)
1411 {
1412 	struct buf *bp;
1413 
1414 	ASSERT_BO_LOCKED(bo);
1415 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1416 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1417 		return (bp);
1418 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1419 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1420 		return (bp);
1421 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1422 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1423 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1424 			return (bp);
1425 	}
1426 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1427 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1428 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1429 			return (bp);
1430 	}
1431 	return (NULL);
1432 }
1433 
1434 /*
1435  * Associate a buffer with a vnode.
1436  */
1437 void
1438 bgetvp(struct vnode *vp, struct buf *bp)
1439 {
1440 
1441 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1442 
1443 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1444 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1445 	    ("bgetvp: bp already attached! %p", bp));
1446 
1447 	ASSERT_VI_LOCKED(vp, "bgetvp");
1448 	vholdl(vp);
1449 	if (VFS_NEEDSGIANT(vp->v_mount))
1450 		bp->b_flags |= B_NEEDSGIANT;
1451 	bp->b_vp = vp;
1452 	bp->b_bufobj = &vp->v_bufobj;
1453 	/*
1454 	 * Insert onto list for new vnode.
1455 	 */
1456 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1457 }
1458 
1459 /*
1460  * Disassociate a buffer from a vnode.
1461  */
1462 void
1463 brelvp(struct buf *bp)
1464 {
1465 	struct bufobj *bo;
1466 	struct vnode *vp;
1467 
1468 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1469 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1470 
1471 	/*
1472 	 * Delete from old vnode list, if on one.
1473 	 */
1474 	vp = bp->b_vp;		/* XXX */
1475 	bo = bp->b_bufobj;
1476 	BO_LOCK(bo);
1477 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1478 		buf_vlist_remove(bp);
1479 	else
1480 		panic("brelvp: Buffer %p not on queue.", bp);
1481 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1482 		bo->bo_flag &= ~BO_ONWORKLST;
1483 		mtx_lock(&sync_mtx);
1484 		LIST_REMOVE(bo, bo_synclist);
1485  		syncer_worklist_len--;
1486 		mtx_unlock(&sync_mtx);
1487 	}
1488 	bp->b_flags &= ~B_NEEDSGIANT;
1489 	bp->b_vp = NULL;
1490 	bp->b_bufobj = NULL;
1491 	vdropl(vp);
1492 }
1493 
1494 /*
1495  * Add an item to the syncer work queue.
1496  */
1497 static void
1498 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1499 {
1500 	int slot;
1501 
1502 	ASSERT_BO_LOCKED(bo);
1503 
1504 	mtx_lock(&sync_mtx);
1505 	if (bo->bo_flag & BO_ONWORKLST)
1506 		LIST_REMOVE(bo, bo_synclist);
1507 	else {
1508 		bo->bo_flag |= BO_ONWORKLST;
1509  		syncer_worklist_len++;
1510 	}
1511 
1512 	if (delay > syncer_maxdelay - 2)
1513 		delay = syncer_maxdelay - 2;
1514 	slot = (syncer_delayno + delay) & syncer_mask;
1515 
1516 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1517 	mtx_unlock(&sync_mtx);
1518 }
1519 
1520 static int
1521 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1522 {
1523 	int error, len;
1524 
1525 	mtx_lock(&sync_mtx);
1526 	len = syncer_worklist_len - sync_vnode_count;
1527 	mtx_unlock(&sync_mtx);
1528 	error = SYSCTL_OUT(req, &len, sizeof(len));
1529 	return (error);
1530 }
1531 
1532 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1533     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1534 
1535 static struct proc *updateproc;
1536 static void sched_sync(void);
1537 static struct kproc_desc up_kp = {
1538 	"syncer",
1539 	sched_sync,
1540 	&updateproc
1541 };
1542 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1543 
1544 static int
1545 sync_vnode(struct bufobj *bo, struct thread *td)
1546 {
1547 	struct vnode *vp;
1548 	struct mount *mp;
1549 
1550 	vp = bo->__bo_vnode; 	/* XXX */
1551 	if (VOP_ISLOCKED(vp, NULL) != 0)
1552 		return (1);
1553 	if (VI_TRYLOCK(vp) == 0)
1554 		return (1);
1555 	/*
1556 	 * We use vhold in case the vnode does not
1557 	 * successfully sync.  vhold prevents the vnode from
1558 	 * going away when we unlock the sync_mtx so that
1559 	 * we can acquire the vnode interlock.
1560 	 */
1561 	vholdl(vp);
1562 	mtx_unlock(&sync_mtx);
1563 	VI_UNLOCK(vp);
1564 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1565 		vdrop(vp);
1566 		mtx_lock(&sync_mtx);
1567 		return (1);
1568 	}
1569 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1570 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1571 	VOP_UNLOCK(vp, 0, td);
1572 	vn_finished_write(mp);
1573 	VI_LOCK(vp);
1574 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1575 		/*
1576 		 * Put us back on the worklist.  The worklist
1577 		 * routine will remove us from our current
1578 		 * position and then add us back in at a later
1579 		 * position.
1580 		 */
1581 		vn_syncer_add_to_worklist(bo, syncdelay);
1582 	}
1583 	vdropl(vp);
1584 	mtx_lock(&sync_mtx);
1585 	return (0);
1586 }
1587 
1588 /*
1589  * System filesystem synchronizer daemon.
1590  */
1591 static void
1592 sched_sync(void)
1593 {
1594 	struct synclist *next;
1595 	struct synclist *slp;
1596 	struct bufobj *bo;
1597 	long starttime;
1598 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1599 	static int dummychan;
1600 	int last_work_seen;
1601 	int net_worklist_len;
1602 	int syncer_final_iter;
1603 	int first_printf;
1604 	int error;
1605 
1606 	mtx_lock(&Giant);
1607 	last_work_seen = 0;
1608 	syncer_final_iter = 0;
1609 	first_printf = 1;
1610 	syncer_state = SYNCER_RUNNING;
1611 	starttime = time_uptime;
1612 	td->td_pflags |= TDP_NORUNNINGBUF;
1613 
1614 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1615 	    SHUTDOWN_PRI_LAST);
1616 
1617 	for (;;) {
1618 		mtx_lock(&sync_mtx);
1619 		if (syncer_state == SYNCER_FINAL_DELAY &&
1620 		    syncer_final_iter == 0) {
1621 			mtx_unlock(&sync_mtx);
1622 			kthread_suspend_check(td->td_proc);
1623 			mtx_lock(&sync_mtx);
1624 		}
1625 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1626 		if (syncer_state != SYNCER_RUNNING &&
1627 		    starttime != time_uptime) {
1628 			if (first_printf) {
1629 				printf("\nSyncing disks, vnodes remaining...");
1630 				first_printf = 0;
1631 			}
1632 			printf("%d ", net_worklist_len);
1633 		}
1634 		starttime = time_uptime;
1635 
1636 		/*
1637 		 * Push files whose dirty time has expired.  Be careful
1638 		 * of interrupt race on slp queue.
1639 		 *
1640 		 * Skip over empty worklist slots when shutting down.
1641 		 */
1642 		do {
1643 			slp = &syncer_workitem_pending[syncer_delayno];
1644 			syncer_delayno += 1;
1645 			if (syncer_delayno == syncer_maxdelay)
1646 				syncer_delayno = 0;
1647 			next = &syncer_workitem_pending[syncer_delayno];
1648 			/*
1649 			 * If the worklist has wrapped since the
1650 			 * it was emptied of all but syncer vnodes,
1651 			 * switch to the FINAL_DELAY state and run
1652 			 * for one more second.
1653 			 */
1654 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1655 			    net_worklist_len == 0 &&
1656 			    last_work_seen == syncer_delayno) {
1657 				syncer_state = SYNCER_FINAL_DELAY;
1658 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1659 			}
1660 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1661 		    syncer_worklist_len > 0);
1662 
1663 		/*
1664 		 * Keep track of the last time there was anything
1665 		 * on the worklist other than syncer vnodes.
1666 		 * Return to the SHUTTING_DOWN state if any
1667 		 * new work appears.
1668 		 */
1669 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1670 			last_work_seen = syncer_delayno;
1671 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1672 			syncer_state = SYNCER_SHUTTING_DOWN;
1673 		while ((bo = LIST_FIRST(slp)) != NULL) {
1674 			error = sync_vnode(bo, td);
1675 			if (error == 1) {
1676 				LIST_REMOVE(bo, bo_synclist);
1677 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1678 				continue;
1679 			}
1680 		}
1681 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1682 			syncer_final_iter--;
1683 		mtx_unlock(&sync_mtx);
1684 		/*
1685 		 * The variable rushjob allows the kernel to speed up the
1686 		 * processing of the filesystem syncer process. A rushjob
1687 		 * value of N tells the filesystem syncer to process the next
1688 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1689 		 * is used by the soft update code to speed up the filesystem
1690 		 * syncer process when the incore state is getting so far
1691 		 * ahead of the disk that the kernel memory pool is being
1692 		 * threatened with exhaustion.
1693 		 */
1694 		mtx_lock(&sync_mtx);
1695 		if (rushjob > 0) {
1696 			rushjob -= 1;
1697 			mtx_unlock(&sync_mtx);
1698 			continue;
1699 		}
1700 		mtx_unlock(&sync_mtx);
1701 		/*
1702 		 * Just sleep for a short period if time between
1703 		 * iterations when shutting down to allow some I/O
1704 		 * to happen.
1705 		 *
1706 		 * If it has taken us less than a second to process the
1707 		 * current work, then wait. Otherwise start right over
1708 		 * again. We can still lose time if any single round
1709 		 * takes more than two seconds, but it does not really
1710 		 * matter as we are just trying to generally pace the
1711 		 * filesystem activity.
1712 		 */
1713 		if (syncer_state != SYNCER_RUNNING)
1714 			tsleep(&dummychan, PPAUSE, "syncfnl",
1715 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1716 		else if (time_uptime == starttime)
1717 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1718 	}
1719 }
1720 
1721 /*
1722  * Request the syncer daemon to speed up its work.
1723  * We never push it to speed up more than half of its
1724  * normal turn time, otherwise it could take over the cpu.
1725  */
1726 int
1727 speedup_syncer()
1728 {
1729 	struct thread *td;
1730 	int ret = 0;
1731 
1732 	td = FIRST_THREAD_IN_PROC(updateproc);
1733 	sleepq_remove(td, &lbolt);
1734 	mtx_lock(&sync_mtx);
1735 	if (rushjob < syncdelay / 2) {
1736 		rushjob += 1;
1737 		stat_rush_requests += 1;
1738 		ret = 1;
1739 	}
1740 	mtx_unlock(&sync_mtx);
1741 	return (ret);
1742 }
1743 
1744 /*
1745  * Tell the syncer to speed up its work and run though its work
1746  * list several times, then tell it to shut down.
1747  */
1748 static void
1749 syncer_shutdown(void *arg, int howto)
1750 {
1751 	struct thread *td;
1752 
1753 	if (howto & RB_NOSYNC)
1754 		return;
1755 	td = FIRST_THREAD_IN_PROC(updateproc);
1756 	sleepq_remove(td, &lbolt);
1757 	mtx_lock(&sync_mtx);
1758 	syncer_state = SYNCER_SHUTTING_DOWN;
1759 	rushjob = 0;
1760 	mtx_unlock(&sync_mtx);
1761 	kproc_shutdown(arg, howto);
1762 }
1763 
1764 /*
1765  * Reassign a buffer from one vnode to another.
1766  * Used to assign file specific control information
1767  * (indirect blocks) to the vnode to which they belong.
1768  */
1769 void
1770 reassignbuf(struct buf *bp)
1771 {
1772 	struct vnode *vp;
1773 	struct bufobj *bo;
1774 	int delay;
1775 #ifdef INVARIANTS
1776 	struct bufv *bv;
1777 #endif
1778 
1779 	vp = bp->b_vp;
1780 	bo = bp->b_bufobj;
1781 	++reassignbufcalls;
1782 
1783 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1784 	    bp, bp->b_vp, bp->b_flags);
1785 	/*
1786 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1787 	 * is not fully linked in.
1788 	 */
1789 	if (bp->b_flags & B_PAGING)
1790 		panic("cannot reassign paging buffer");
1791 
1792 	/*
1793 	 * Delete from old vnode list, if on one.
1794 	 */
1795 	VI_LOCK(vp);
1796 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1797 		buf_vlist_remove(bp);
1798 	else
1799 		panic("reassignbuf: Buffer %p not on queue.", bp);
1800 	/*
1801 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1802 	 * of clean buffers.
1803 	 */
1804 	if (bp->b_flags & B_DELWRI) {
1805 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1806 			switch (vp->v_type) {
1807 			case VDIR:
1808 				delay = dirdelay;
1809 				break;
1810 			case VCHR:
1811 				delay = metadelay;
1812 				break;
1813 			default:
1814 				delay = filedelay;
1815 			}
1816 			vn_syncer_add_to_worklist(bo, delay);
1817 		}
1818 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1819 	} else {
1820 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1821 
1822 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1823 			mtx_lock(&sync_mtx);
1824 			LIST_REMOVE(bo, bo_synclist);
1825  			syncer_worklist_len--;
1826 			mtx_unlock(&sync_mtx);
1827 			bo->bo_flag &= ~BO_ONWORKLST;
1828 		}
1829 	}
1830 #ifdef INVARIANTS
1831 	bv = &bo->bo_clean;
1832 	bp = TAILQ_FIRST(&bv->bv_hd);
1833 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1834 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1835 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1836 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1837 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1838 	bv = &bo->bo_dirty;
1839 	bp = TAILQ_FIRST(&bv->bv_hd);
1840 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1841 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1842 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1843 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1844 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1845 #endif
1846 	VI_UNLOCK(vp);
1847 }
1848 
1849 /*
1850  * Increment the use and hold counts on the vnode, taking care to reference
1851  * the driver's usecount if this is a chardev.  The vholdl() will remove
1852  * the vnode from the free list if it is presently free.  Requires the
1853  * vnode interlock and returns with it held.
1854  */
1855 static void
1856 v_incr_usecount(struct vnode *vp)
1857 {
1858 
1859 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1860 	    vp, vp->v_holdcnt, vp->v_usecount);
1861 	vp->v_usecount++;
1862 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1863 		dev_lock();
1864 		vp->v_rdev->si_usecount++;
1865 		dev_unlock();
1866 	}
1867 	vholdl(vp);
1868 }
1869 
1870 /*
1871  * Turn a holdcnt into a use+holdcnt such that only one call to
1872  * v_decr_usecount is needed.
1873  */
1874 static void
1875 v_upgrade_usecount(struct vnode *vp)
1876 {
1877 
1878 	CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n",
1879 	    vp, vp->v_holdcnt, vp->v_usecount);
1880 	vp->v_usecount++;
1881 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1882 		dev_lock();
1883 		vp->v_rdev->si_usecount++;
1884 		dev_unlock();
1885 	}
1886 }
1887 
1888 /*
1889  * Decrement the vnode use and hold count along with the driver's usecount
1890  * if this is a chardev.  The vdropl() below releases the vnode interlock
1891  * as it may free the vnode.
1892  */
1893 static void
1894 v_decr_usecount(struct vnode *vp)
1895 {
1896 
1897 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1898 	    vp, vp->v_holdcnt, vp->v_usecount);
1899 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1900 	VNASSERT(vp->v_usecount > 0, vp,
1901 	    ("v_decr_usecount: negative usecount"));
1902 	vp->v_usecount--;
1903 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1904 		dev_lock();
1905 		vp->v_rdev->si_usecount--;
1906 		dev_unlock();
1907 	}
1908 	vdropl(vp);
1909 }
1910 
1911 /*
1912  * Decrement only the use count and driver use count.  This is intended to
1913  * be paired with a follow on vdropl() to release the remaining hold count.
1914  * In this way we may vgone() a vnode with a 0 usecount without risk of
1915  * having it end up on a free list because the hold count is kept above 0.
1916  */
1917 static void
1918 v_decr_useonly(struct vnode *vp)
1919 {
1920 
1921 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1922 	    vp, vp->v_holdcnt, vp->v_usecount);
1923 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1924 	VNASSERT(vp->v_usecount > 0, vp,
1925 	    ("v_decr_useonly: negative usecount"));
1926 	vp->v_usecount--;
1927 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1928 		dev_lock();
1929 		vp->v_rdev->si_usecount--;
1930 		dev_unlock();
1931 	}
1932 }
1933 
1934 /*
1935  * Grab a particular vnode from the free list, increment its
1936  * reference count and lock it. The vnode lock bit is set if the
1937  * vnode is being eliminated in vgone. The process is awakened
1938  * when the transition is completed, and an error returned to
1939  * indicate that the vnode is no longer usable (possibly having
1940  * been changed to a new filesystem type).
1941  */
1942 int
1943 vget(struct vnode *vp, int flags, struct thread *td)
1944 {
1945 	int oweinact;
1946 	int oldflags;
1947 	int error;
1948 
1949 	error = 0;
1950 	oldflags = flags;
1951 	oweinact = 0;
1952 	VFS_ASSERT_GIANT(vp->v_mount);
1953 	if ((flags & LK_INTERLOCK) == 0)
1954 		VI_LOCK(vp);
1955 	/*
1956 	 * If the inactive call was deferred because vput() was called
1957 	 * with a shared lock, we have to do it here before another thread
1958 	 * gets a reference to data that should be dead.
1959 	 */
1960 	if (vp->v_iflag & VI_OWEINACT) {
1961 		if (flags & LK_NOWAIT) {
1962 			VI_UNLOCK(vp);
1963 			return (EBUSY);
1964 		}
1965 		flags &= ~LK_TYPE_MASK;
1966 		flags |= LK_EXCLUSIVE;
1967 		oweinact = 1;
1968 	}
1969 	vholdl(vp);
1970 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1971 		vdrop(vp);
1972 		return (error);
1973 	}
1974 	VI_LOCK(vp);
1975 	/* Upgrade our holdcnt to a usecount. */
1976 	v_upgrade_usecount(vp);
1977 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
1978 		panic("vget: vn_lock failed to return ENOENT\n");
1979 	if (oweinact) {
1980 		if (vp->v_iflag & VI_OWEINACT)
1981 			vinactive(vp, td);
1982 		VI_UNLOCK(vp);
1983 		if ((oldflags & LK_TYPE_MASK) == 0)
1984 			VOP_UNLOCK(vp, 0, td);
1985 	} else
1986 		VI_UNLOCK(vp);
1987 	return (0);
1988 }
1989 
1990 /*
1991  * Increase the reference count of a vnode.
1992  */
1993 void
1994 vref(struct vnode *vp)
1995 {
1996 
1997 	VI_LOCK(vp);
1998 	v_incr_usecount(vp);
1999 	VI_UNLOCK(vp);
2000 }
2001 
2002 /*
2003  * Return reference count of a vnode.
2004  *
2005  * The results of this call are only guaranteed when some mechanism other
2006  * than the VI lock is used to stop other processes from gaining references
2007  * to the vnode.  This may be the case if the caller holds the only reference.
2008  * This is also useful when stale data is acceptable as race conditions may
2009  * be accounted for by some other means.
2010  */
2011 int
2012 vrefcnt(struct vnode *vp)
2013 {
2014 	int usecnt;
2015 
2016 	VI_LOCK(vp);
2017 	usecnt = vp->v_usecount;
2018 	VI_UNLOCK(vp);
2019 
2020 	return (usecnt);
2021 }
2022 
2023 
2024 /*
2025  * Vnode put/release.
2026  * If count drops to zero, call inactive routine and return to freelist.
2027  */
2028 void
2029 vrele(struct vnode *vp)
2030 {
2031 	struct thread *td = curthread;	/* XXX */
2032 
2033 	KASSERT(vp != NULL, ("vrele: null vp"));
2034 	VFS_ASSERT_GIANT(vp->v_mount);
2035 
2036 	VI_LOCK(vp);
2037 
2038 	/* Skip this v_writecount check if we're going to panic below. */
2039 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2040 	    ("vrele: missed vn_close"));
2041 
2042 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2043 	    vp->v_usecount == 1)) {
2044 		v_decr_usecount(vp);
2045 		return;
2046 	}
2047 	if (vp->v_usecount != 1) {
2048 #ifdef DIAGNOSTIC
2049 		vprint("vrele: negative ref count", vp);
2050 #endif
2051 		VI_UNLOCK(vp);
2052 		panic("vrele: negative ref cnt");
2053 	}
2054 	/*
2055 	 * We want to hold the vnode until the inactive finishes to
2056 	 * prevent vgone() races.  We drop the use count here and the
2057 	 * hold count below when we're done.
2058 	 */
2059 	v_decr_useonly(vp);
2060 	/*
2061 	 * We must call VOP_INACTIVE with the node locked. Mark
2062 	 * as VI_DOINGINACT to avoid recursion.
2063 	 */
2064 	vp->v_iflag |= VI_OWEINACT;
2065 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2066 		VI_LOCK(vp);
2067 		if (vp->v_usecount > 0)
2068 			vp->v_iflag &= ~VI_OWEINACT;
2069 		if (vp->v_iflag & VI_OWEINACT)
2070 			vinactive(vp, td);
2071 		VOP_UNLOCK(vp, 0, td);
2072 	} else {
2073 		VI_LOCK(vp);
2074 		if (vp->v_usecount > 0)
2075 			vp->v_iflag &= ~VI_OWEINACT;
2076 	}
2077 	vdropl(vp);
2078 }
2079 
2080 /*
2081  * Release an already locked vnode.  This give the same effects as
2082  * unlock+vrele(), but takes less time and avoids releasing and
2083  * re-aquiring the lock (as vrele() aquires the lock internally.)
2084  */
2085 void
2086 vput(struct vnode *vp)
2087 {
2088 	struct thread *td = curthread;	/* XXX */
2089 	int error;
2090 
2091 	KASSERT(vp != NULL, ("vput: null vp"));
2092 	ASSERT_VOP_LOCKED(vp, "vput");
2093 	VFS_ASSERT_GIANT(vp->v_mount);
2094 	VI_LOCK(vp);
2095 	/* Skip this v_writecount check if we're going to panic below. */
2096 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2097 	    ("vput: missed vn_close"));
2098 	error = 0;
2099 
2100 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2101 	    vp->v_usecount == 1)) {
2102 		VOP_UNLOCK(vp, 0, td);
2103 		v_decr_usecount(vp);
2104 		return;
2105 	}
2106 
2107 	if (vp->v_usecount != 1) {
2108 #ifdef DIAGNOSTIC
2109 		vprint("vput: negative ref count", vp);
2110 #endif
2111 		panic("vput: negative ref cnt");
2112 	}
2113 	/*
2114 	 * We want to hold the vnode until the inactive finishes to
2115 	 * prevent vgone() races.  We drop the use count here and the
2116 	 * hold count below when we're done.
2117 	 */
2118 	v_decr_useonly(vp);
2119 	vp->v_iflag |= VI_OWEINACT;
2120 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2121 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2122 		VI_LOCK(vp);
2123 		if (error) {
2124 			if (vp->v_usecount > 0)
2125 				vp->v_iflag &= ~VI_OWEINACT;
2126 			goto done;
2127 		}
2128 	}
2129 	if (vp->v_usecount > 0)
2130 		vp->v_iflag &= ~VI_OWEINACT;
2131 	if (vp->v_iflag & VI_OWEINACT)
2132 		vinactive(vp, td);
2133 	VOP_UNLOCK(vp, 0, td);
2134 done:
2135 	vdropl(vp);
2136 }
2137 
2138 /*
2139  * Somebody doesn't want the vnode recycled.
2140  */
2141 void
2142 vhold(struct vnode *vp)
2143 {
2144 
2145 	VI_LOCK(vp);
2146 	vholdl(vp);
2147 	VI_UNLOCK(vp);
2148 }
2149 
2150 void
2151 vholdl(struct vnode *vp)
2152 {
2153 
2154 	vp->v_holdcnt++;
2155 	if (VSHOULDBUSY(vp))
2156 		vbusy(vp);
2157 }
2158 
2159 /*
2160  * Note that there is one less who cares about this vnode.  vdrop() is the
2161  * opposite of vhold().
2162  */
2163 void
2164 vdrop(struct vnode *vp)
2165 {
2166 
2167 	VI_LOCK(vp);
2168 	vdropl(vp);
2169 }
2170 
2171 /*
2172  * Drop the hold count of the vnode.  If this is the last reference to
2173  * the vnode we will free it if it has been vgone'd otherwise it is
2174  * placed on the free list.
2175  */
2176 static void
2177 vdropl(struct vnode *vp)
2178 {
2179 
2180 	if (vp->v_holdcnt <= 0)
2181 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2182 	vp->v_holdcnt--;
2183 	if (vp->v_holdcnt == 0) {
2184 		if (vp->v_iflag & VI_DOOMED) {
2185 			vdestroy(vp);
2186 			return;
2187 		} else
2188 			vfree(vp);
2189 	}
2190 	VI_UNLOCK(vp);
2191 }
2192 
2193 /*
2194  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2195  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2196  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2197  * failed lock upgrade.
2198  */
2199 static void
2200 vinactive(struct vnode *vp, struct thread *td)
2201 {
2202 
2203 	ASSERT_VOP_LOCKED(vp, "vinactive");
2204 	ASSERT_VI_LOCKED(vp, "vinactive");
2205 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2206 	    ("vinactive: recursed on VI_DOINGINACT"));
2207 	vp->v_iflag |= VI_DOINGINACT;
2208 	vp->v_iflag &= ~VI_OWEINACT;
2209 	VI_UNLOCK(vp);
2210 	VOP_INACTIVE(vp, td);
2211 	VI_LOCK(vp);
2212 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2213 	    ("vinactive: lost VI_DOINGINACT"));
2214 	vp->v_iflag &= ~VI_DOINGINACT;
2215 }
2216 
2217 /*
2218  * Remove any vnodes in the vnode table belonging to mount point mp.
2219  *
2220  * If FORCECLOSE is not specified, there should not be any active ones,
2221  * return error if any are found (nb: this is a user error, not a
2222  * system error). If FORCECLOSE is specified, detach any active vnodes
2223  * that are found.
2224  *
2225  * If WRITECLOSE is set, only flush out regular file vnodes open for
2226  * writing.
2227  *
2228  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2229  *
2230  * `rootrefs' specifies the base reference count for the root vnode
2231  * of this filesystem. The root vnode is considered busy if its
2232  * v_usecount exceeds this value. On a successful return, vflush(, td)
2233  * will call vrele() on the root vnode exactly rootrefs times.
2234  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2235  * be zero.
2236  */
2237 #ifdef DIAGNOSTIC
2238 static int busyprt = 0;		/* print out busy vnodes */
2239 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2240 #endif
2241 
2242 int
2243 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td)
2244 {
2245 	struct vnode *vp, *mvp, *rootvp = NULL;
2246 	struct vattr vattr;
2247 	int busy = 0, error;
2248 
2249 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2250 	if (rootrefs > 0) {
2251 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2252 		    ("vflush: bad args"));
2253 		/*
2254 		 * Get the filesystem root vnode. We can vput() it
2255 		 * immediately, since with rootrefs > 0, it won't go away.
2256 		 */
2257 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2258 			return (error);
2259 		vput(rootvp);
2260 
2261 	}
2262 	MNT_ILOCK(mp);
2263 loop:
2264 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2265 
2266 		VI_LOCK(vp);
2267 		vholdl(vp);
2268 		MNT_IUNLOCK(mp);
2269 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2270 		if (error) {
2271 			vdrop(vp);
2272 			MNT_ILOCK(mp);
2273 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2274 			goto loop;
2275 		}
2276 		/*
2277 		 * Skip over a vnodes marked VV_SYSTEM.
2278 		 */
2279 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2280 			VOP_UNLOCK(vp, 0, td);
2281 			vdrop(vp);
2282 			MNT_ILOCK(mp);
2283 			continue;
2284 		}
2285 		/*
2286 		 * If WRITECLOSE is set, flush out unlinked but still open
2287 		 * files (even if open only for reading) and regular file
2288 		 * vnodes open for writing.
2289 		 */
2290 		if (flags & WRITECLOSE) {
2291 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2292 			VI_LOCK(vp);
2293 
2294 			if ((vp->v_type == VNON ||
2295 			    (error == 0 && vattr.va_nlink > 0)) &&
2296 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2297 				VOP_UNLOCK(vp, 0, td);
2298 				vdropl(vp);
2299 				MNT_ILOCK(mp);
2300 				continue;
2301 			}
2302 		} else
2303 			VI_LOCK(vp);
2304 		/*
2305 		 * With v_usecount == 0, all we need to do is clear out the
2306 		 * vnode data structures and we are done.
2307 		 *
2308 		 * If FORCECLOSE is set, forcibly close the vnode.
2309 		 */
2310 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2311 			VNASSERT(vp->v_usecount == 0 ||
2312 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2313 			    ("device VNODE %p is FORCECLOSED", vp));
2314 			vgonel(vp);
2315 		} else {
2316 			busy++;
2317 #ifdef DIAGNOSTIC
2318 			if (busyprt)
2319 				vprint("vflush: busy vnode", vp);
2320 #endif
2321 		}
2322 		VOP_UNLOCK(vp, 0, td);
2323 		vdropl(vp);
2324 		MNT_ILOCK(mp);
2325 	}
2326 	MNT_IUNLOCK(mp);
2327 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2328 		/*
2329 		 * If just the root vnode is busy, and if its refcount
2330 		 * is equal to `rootrefs', then go ahead and kill it.
2331 		 */
2332 		VI_LOCK(rootvp);
2333 		KASSERT(busy > 0, ("vflush: not busy"));
2334 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2335 		    ("vflush: usecount %d < rootrefs %d",
2336 		     rootvp->v_usecount, rootrefs));
2337 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2338 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2339 			vgone(rootvp);
2340 			VOP_UNLOCK(rootvp, 0, td);
2341 			busy = 0;
2342 		} else
2343 			VI_UNLOCK(rootvp);
2344 	}
2345 	if (busy)
2346 		return (EBUSY);
2347 	for (; rootrefs > 0; rootrefs--)
2348 		vrele(rootvp);
2349 	return (0);
2350 }
2351 
2352 /*
2353  * Recycle an unused vnode to the front of the free list.
2354  */
2355 int
2356 vrecycle(struct vnode *vp, struct thread *td)
2357 {
2358 	int recycled;
2359 
2360 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2361 	recycled = 0;
2362 	VI_LOCK(vp);
2363 	if (vp->v_usecount == 0) {
2364 		recycled = 1;
2365 		vgonel(vp);
2366 	}
2367 	VI_UNLOCK(vp);
2368 	return (recycled);
2369 }
2370 
2371 /*
2372  * Eliminate all activity associated with a vnode
2373  * in preparation for reuse.
2374  */
2375 void
2376 vgone(struct vnode *vp)
2377 {
2378 	VI_LOCK(vp);
2379 	vgonel(vp);
2380 	VI_UNLOCK(vp);
2381 }
2382 
2383 /*
2384  * vgone, with the vp interlock held.
2385  */
2386 void
2387 vgonel(struct vnode *vp)
2388 {
2389 	struct thread *td;
2390 	int oweinact;
2391 	int active;
2392 	struct mount *mp;
2393 
2394 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2395 	ASSERT_VOP_LOCKED(vp, "vgonel");
2396 	ASSERT_VI_LOCKED(vp, "vgonel");
2397 	VNASSERT(vp->v_holdcnt, vp,
2398 	    ("vgonel: vp %p has no reference.", vp));
2399 	td = curthread;
2400 
2401 	/*
2402 	 * Don't vgonel if we're already doomed.
2403 	 */
2404 	if (vp->v_iflag & VI_DOOMED)
2405 		return;
2406 	vp->v_iflag |= VI_DOOMED;
2407 	/*
2408 	 * Check to see if the vnode is in use.  If so, we have to call
2409 	 * VOP_CLOSE() and VOP_INACTIVE().
2410 	 */
2411 	active = vp->v_usecount;
2412 	oweinact = (vp->v_iflag & VI_OWEINACT);
2413 	VI_UNLOCK(vp);
2414 	/*
2415 	 * Clean out any buffers associated with the vnode.
2416 	 * If the flush fails, just toss the buffers.
2417 	 */
2418 	mp = NULL;
2419 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2420 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
2421 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2422 		vinvalbuf(vp, 0, td, 0, 0);
2423 
2424 	/*
2425 	 * If purging an active vnode, it must be closed and
2426 	 * deactivated before being reclaimed.
2427 	 */
2428 	if (active)
2429 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2430 	if (oweinact || active) {
2431 		VI_LOCK(vp);
2432 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2433 			vinactive(vp, td);
2434 		VI_UNLOCK(vp);
2435 	}
2436 	/*
2437 	 * Reclaim the vnode.
2438 	 */
2439 	if (VOP_RECLAIM(vp, td))
2440 		panic("vgone: cannot reclaim");
2441 	if (mp != NULL)
2442 		vn_finished_secondary_write(mp);
2443 	VNASSERT(vp->v_object == NULL, vp,
2444 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2445 	/*
2446 	 * Delete from old mount point vnode list.
2447 	 */
2448 	delmntque(vp);
2449 	cache_purge(vp);
2450 	/*
2451 	 * Done with purge, reset to the standard lock and invalidate
2452 	 * the vnode.
2453 	 */
2454 	VI_LOCK(vp);
2455 	vp->v_vnlock = &vp->v_lock;
2456 	vp->v_op = &dead_vnodeops;
2457 	vp->v_tag = "none";
2458 	vp->v_type = VBAD;
2459 }
2460 
2461 /*
2462  * Calculate the total number of references to a special device.
2463  */
2464 int
2465 vcount(struct vnode *vp)
2466 {
2467 	int count;
2468 
2469 	dev_lock();
2470 	count = vp->v_rdev->si_usecount;
2471 	dev_unlock();
2472 	return (count);
2473 }
2474 
2475 /*
2476  * Same as above, but using the struct cdev *as argument
2477  */
2478 int
2479 count_dev(struct cdev *dev)
2480 {
2481 	int count;
2482 
2483 	dev_lock();
2484 	count = dev->si_usecount;
2485 	dev_unlock();
2486 	return(count);
2487 }
2488 
2489 /*
2490  * Print out a description of a vnode.
2491  */
2492 static char *typename[] =
2493 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2494  "VMARKER"};
2495 
2496 void
2497 vn_printf(struct vnode *vp, const char *fmt, ...)
2498 {
2499 	va_list ap;
2500 	char buf[96];
2501 
2502 	va_start(ap, fmt);
2503 	vprintf(fmt, ap);
2504 	va_end(ap);
2505 	printf("%p: ", (void *)vp);
2506 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2507 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2508 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2509 	buf[0] = '\0';
2510 	buf[1] = '\0';
2511 	if (vp->v_vflag & VV_ROOT)
2512 		strcat(buf, "|VV_ROOT");
2513 	if (vp->v_vflag & VV_TEXT)
2514 		strcat(buf, "|VV_TEXT");
2515 	if (vp->v_vflag & VV_SYSTEM)
2516 		strcat(buf, "|VV_SYSTEM");
2517 	if (vp->v_iflag & VI_DOOMED)
2518 		strcat(buf, "|VI_DOOMED");
2519 	if (vp->v_iflag & VI_FREE)
2520 		strcat(buf, "|VI_FREE");
2521 	printf("    flags (%s)\n", buf + 1);
2522 	if (mtx_owned(VI_MTX(vp)))
2523 		printf(" VI_LOCKed");
2524 	if (vp->v_object != NULL)
2525 		printf("    v_object %p ref %d pages %d\n",
2526 		    vp->v_object, vp->v_object->ref_count,
2527 		    vp->v_object->resident_page_count);
2528 	printf("    ");
2529 	lockmgr_printinfo(vp->v_vnlock);
2530 	printf("\n");
2531 	if (vp->v_data != NULL)
2532 		VOP_PRINT(vp);
2533 }
2534 
2535 #ifdef DDB
2536 #include <ddb/ddb.h>
2537 /*
2538  * List all of the locked vnodes in the system.
2539  * Called when debugging the kernel.
2540  */
2541 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2542 {
2543 	struct mount *mp, *nmp;
2544 	struct vnode *vp;
2545 
2546 	/*
2547 	 * Note: because this is DDB, we can't obey the locking semantics
2548 	 * for these structures, which means we could catch an inconsistent
2549 	 * state and dereference a nasty pointer.  Not much to be done
2550 	 * about that.
2551 	 */
2552 	printf("Locked vnodes\n");
2553 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2554 		nmp = TAILQ_NEXT(mp, mnt_list);
2555 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2556 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL))
2557 				vprint("", vp);
2558 		}
2559 		nmp = TAILQ_NEXT(mp, mnt_list);
2560 	}
2561 }
2562 #endif
2563 
2564 /*
2565  * Fill in a struct xvfsconf based on a struct vfsconf.
2566  */
2567 static void
2568 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2569 {
2570 
2571 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2572 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2573 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2574 	xvfsp->vfc_flags = vfsp->vfc_flags;
2575 	/*
2576 	 * These are unused in userland, we keep them
2577 	 * to not break binary compatibility.
2578 	 */
2579 	xvfsp->vfc_vfsops = NULL;
2580 	xvfsp->vfc_next = NULL;
2581 }
2582 
2583 /*
2584  * Top level filesystem related information gathering.
2585  */
2586 static int
2587 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2588 {
2589 	struct vfsconf *vfsp;
2590 	struct xvfsconf xvfsp;
2591 	int error;
2592 
2593 	error = 0;
2594 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2595 		bzero(&xvfsp, sizeof(xvfsp));
2596 		vfsconf2x(vfsp, &xvfsp);
2597 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2598 		if (error)
2599 			break;
2600 	}
2601 	return (error);
2602 }
2603 
2604 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2605     "S,xvfsconf", "List of all configured filesystems");
2606 
2607 #ifndef BURN_BRIDGES
2608 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2609 
2610 static int
2611 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2612 {
2613 	int *name = (int *)arg1 - 1;	/* XXX */
2614 	u_int namelen = arg2 + 1;	/* XXX */
2615 	struct vfsconf *vfsp;
2616 	struct xvfsconf xvfsp;
2617 
2618 	printf("WARNING: userland calling deprecated sysctl, "
2619 	    "please rebuild world\n");
2620 
2621 #if 1 || defined(COMPAT_PRELITE2)
2622 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2623 	if (namelen == 1)
2624 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2625 #endif
2626 
2627 	switch (name[1]) {
2628 	case VFS_MAXTYPENUM:
2629 		if (namelen != 2)
2630 			return (ENOTDIR);
2631 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2632 	case VFS_CONF:
2633 		if (namelen != 3)
2634 			return (ENOTDIR);	/* overloaded */
2635 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2636 			if (vfsp->vfc_typenum == name[2])
2637 				break;
2638 		if (vfsp == NULL)
2639 			return (EOPNOTSUPP);
2640 		bzero(&xvfsp, sizeof(xvfsp));
2641 		vfsconf2x(vfsp, &xvfsp);
2642 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2643 	}
2644 	return (EOPNOTSUPP);
2645 }
2646 
2647 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2648 	vfs_sysctl, "Generic filesystem");
2649 
2650 #if 1 || defined(COMPAT_PRELITE2)
2651 
2652 static int
2653 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2654 {
2655 	int error;
2656 	struct vfsconf *vfsp;
2657 	struct ovfsconf ovfs;
2658 
2659 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2660 		bzero(&ovfs, sizeof(ovfs));
2661 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2662 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2663 		ovfs.vfc_index = vfsp->vfc_typenum;
2664 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2665 		ovfs.vfc_flags = vfsp->vfc_flags;
2666 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2667 		if (error)
2668 			return error;
2669 	}
2670 	return 0;
2671 }
2672 
2673 #endif /* 1 || COMPAT_PRELITE2 */
2674 #endif /* !BURN_BRIDGES */
2675 
2676 #define KINFO_VNODESLOP		10
2677 #ifdef notyet
2678 /*
2679  * Dump vnode list (via sysctl).
2680  */
2681 /* ARGSUSED */
2682 static int
2683 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2684 {
2685 	struct xvnode *xvn;
2686 	struct thread *td = req->td;
2687 	struct mount *mp;
2688 	struct vnode *vp;
2689 	int error, len, n;
2690 
2691 	/*
2692 	 * Stale numvnodes access is not fatal here.
2693 	 */
2694 	req->lock = 0;
2695 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2696 	if (!req->oldptr)
2697 		/* Make an estimate */
2698 		return (SYSCTL_OUT(req, 0, len));
2699 
2700 	error = sysctl_wire_old_buffer(req, 0);
2701 	if (error != 0)
2702 		return (error);
2703 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2704 	n = 0;
2705 	mtx_lock(&mountlist_mtx);
2706 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2707 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2708 			continue;
2709 		MNT_ILOCK(mp);
2710 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2711 			if (n == len)
2712 				break;
2713 			vref(vp);
2714 			xvn[n].xv_size = sizeof *xvn;
2715 			xvn[n].xv_vnode = vp;
2716 			xvn[n].xv_id = 0;	/* XXX compat */
2717 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2718 			XV_COPY(usecount);
2719 			XV_COPY(writecount);
2720 			XV_COPY(holdcnt);
2721 			XV_COPY(mount);
2722 			XV_COPY(numoutput);
2723 			XV_COPY(type);
2724 #undef XV_COPY
2725 			xvn[n].xv_flag = vp->v_vflag;
2726 
2727 			switch (vp->v_type) {
2728 			case VREG:
2729 			case VDIR:
2730 			case VLNK:
2731 				break;
2732 			case VBLK:
2733 			case VCHR:
2734 				if (vp->v_rdev == NULL) {
2735 					vrele(vp);
2736 					continue;
2737 				}
2738 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2739 				break;
2740 			case VSOCK:
2741 				xvn[n].xv_socket = vp->v_socket;
2742 				break;
2743 			case VFIFO:
2744 				xvn[n].xv_fifo = vp->v_fifoinfo;
2745 				break;
2746 			case VNON:
2747 			case VBAD:
2748 			default:
2749 				/* shouldn't happen? */
2750 				vrele(vp);
2751 				continue;
2752 			}
2753 			vrele(vp);
2754 			++n;
2755 		}
2756 		MNT_IUNLOCK(mp);
2757 		mtx_lock(&mountlist_mtx);
2758 		vfs_unbusy(mp, td);
2759 		if (n == len)
2760 			break;
2761 	}
2762 	mtx_unlock(&mountlist_mtx);
2763 
2764 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2765 	free(xvn, M_TEMP);
2766 	return (error);
2767 }
2768 
2769 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2770 	0, 0, sysctl_vnode, "S,xvnode", "");
2771 #endif
2772 
2773 /*
2774  * Unmount all filesystems. The list is traversed in reverse order
2775  * of mounting to avoid dependencies.
2776  */
2777 void
2778 vfs_unmountall(void)
2779 {
2780 	struct mount *mp;
2781 	struct thread *td;
2782 	int error;
2783 
2784 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2785 	td = curthread;
2786 	/*
2787 	 * Since this only runs when rebooting, it is not interlocked.
2788 	 */
2789 	while(!TAILQ_EMPTY(&mountlist)) {
2790 		mp = TAILQ_LAST(&mountlist, mntlist);
2791 		error = dounmount(mp, MNT_FORCE, td);
2792 		if (error) {
2793 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2794 			/*
2795 			 * XXX: Due to the way in which we mount the root
2796 			 * file system off of devfs, devfs will generate a
2797 			 * "busy" warning when we try to unmount it before
2798 			 * the root.  Don't print a warning as a result in
2799 			 * order to avoid false positive errors that may
2800 			 * cause needless upset.
2801 			 */
2802 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2803 				printf("unmount of %s failed (",
2804 				    mp->mnt_stat.f_mntonname);
2805 				if (error == EBUSY)
2806 					printf("BUSY)\n");
2807 				else
2808 					printf("%d)\n", error);
2809 			}
2810 		} else {
2811 			/* The unmount has removed mp from the mountlist */
2812 		}
2813 	}
2814 }
2815 
2816 /*
2817  * perform msync on all vnodes under a mount point
2818  * the mount point must be locked.
2819  */
2820 void
2821 vfs_msync(struct mount *mp, int flags)
2822 {
2823 	struct vnode *vp, *mvp;
2824 	struct vm_object *obj;
2825 
2826 	(void) vn_start_write(NULL, &mp, V_WAIT);
2827 	MNT_ILOCK(mp);
2828 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2829 		VI_LOCK(vp);
2830 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2831 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2832 			MNT_IUNLOCK(mp);
2833 			if (!vget(vp,
2834 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2835 			    curthread)) {
2836 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2837 					vput(vp);
2838 					MNT_ILOCK(mp);
2839 					continue;
2840 				}
2841 
2842 				obj = vp->v_object;
2843 				if (obj != NULL) {
2844 					VM_OBJECT_LOCK(obj);
2845 					vm_object_page_clean(obj, 0, 0,
2846 					    flags == MNT_WAIT ?
2847 					    OBJPC_SYNC : OBJPC_NOSYNC);
2848 					VM_OBJECT_UNLOCK(obj);
2849 				}
2850 				vput(vp);
2851 			}
2852 			MNT_ILOCK(mp);
2853 		} else
2854 			VI_UNLOCK(vp);
2855 	}
2856 	MNT_IUNLOCK(mp);
2857 	vn_finished_write(mp);
2858 }
2859 
2860 /*
2861  * Mark a vnode as free, putting it up for recycling.
2862  */
2863 static void
2864 vfree(struct vnode *vp)
2865 {
2866 
2867 	CTR1(KTR_VFS, "vfree vp %p", vp);
2868 	ASSERT_VI_LOCKED(vp, "vfree");
2869 	mtx_lock(&vnode_free_list_mtx);
2870 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2871 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2872 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2873 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2874 	    ("vfree: Freeing doomed vnode"));
2875 	if (vp->v_iflag & VI_AGE) {
2876 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2877 	} else {
2878 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2879 	}
2880 	freevnodes++;
2881 	vp->v_iflag &= ~VI_AGE;
2882 	vp->v_iflag |= VI_FREE;
2883 	mtx_unlock(&vnode_free_list_mtx);
2884 }
2885 
2886 /*
2887  * Opposite of vfree() - mark a vnode as in use.
2888  */
2889 static void
2890 vbusy(struct vnode *vp)
2891 {
2892 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2893 	ASSERT_VI_LOCKED(vp, "vbusy");
2894 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2895 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2896 
2897 	mtx_lock(&vnode_free_list_mtx);
2898 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2899 	freevnodes--;
2900 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2901 	mtx_unlock(&vnode_free_list_mtx);
2902 }
2903 
2904 /*
2905  * Initalize per-vnode helper structure to hold poll-related state.
2906  */
2907 void
2908 v_addpollinfo(struct vnode *vp)
2909 {
2910 	struct vpollinfo *vi;
2911 
2912 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2913 	if (vp->v_pollinfo != NULL) {
2914 		uma_zfree(vnodepoll_zone, vi);
2915 		return;
2916 	}
2917 	vp->v_pollinfo = vi;
2918 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2919 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
2920 	    vfs_knlunlock, vfs_knllocked);
2921 }
2922 
2923 /*
2924  * Record a process's interest in events which might happen to
2925  * a vnode.  Because poll uses the historic select-style interface
2926  * internally, this routine serves as both the ``check for any
2927  * pending events'' and the ``record my interest in future events''
2928  * functions.  (These are done together, while the lock is held,
2929  * to avoid race conditions.)
2930  */
2931 int
2932 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
2933 {
2934 
2935 	if (vp->v_pollinfo == NULL)
2936 		v_addpollinfo(vp);
2937 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2938 	if (vp->v_pollinfo->vpi_revents & events) {
2939 		/*
2940 		 * This leaves events we are not interested
2941 		 * in available for the other process which
2942 		 * which presumably had requested them
2943 		 * (otherwise they would never have been
2944 		 * recorded).
2945 		 */
2946 		events &= vp->v_pollinfo->vpi_revents;
2947 		vp->v_pollinfo->vpi_revents &= ~events;
2948 
2949 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2950 		return events;
2951 	}
2952 	vp->v_pollinfo->vpi_events |= events;
2953 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2954 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2955 	return 0;
2956 }
2957 
2958 /*
2959  * Routine to create and manage a filesystem syncer vnode.
2960  */
2961 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2962 static int	sync_fsync(struct  vop_fsync_args *);
2963 static int	sync_inactive(struct  vop_inactive_args *);
2964 static int	sync_reclaim(struct  vop_reclaim_args *);
2965 
2966 static struct vop_vector sync_vnodeops = {
2967 	.vop_bypass =	VOP_EOPNOTSUPP,
2968 	.vop_close =	sync_close,		/* close */
2969 	.vop_fsync =	sync_fsync,		/* fsync */
2970 	.vop_inactive =	sync_inactive,	/* inactive */
2971 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2972 	.vop_lock =	vop_stdlock,	/* lock */
2973 	.vop_unlock =	vop_stdunlock,	/* unlock */
2974 	.vop_islocked =	vop_stdislocked,	/* islocked */
2975 };
2976 
2977 /*
2978  * Create a new filesystem syncer vnode for the specified mount point.
2979  */
2980 int
2981 vfs_allocate_syncvnode(struct mount *mp)
2982 {
2983 	struct vnode *vp;
2984 	static long start, incr, next;
2985 	int error;
2986 
2987 	/* Allocate a new vnode */
2988 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2989 		mp->mnt_syncer = NULL;
2990 		return (error);
2991 	}
2992 	vp->v_type = VNON;
2993 	/*
2994 	 * Place the vnode onto the syncer worklist. We attempt to
2995 	 * scatter them about on the list so that they will go off
2996 	 * at evenly distributed times even if all the filesystems
2997 	 * are mounted at once.
2998 	 */
2999 	next += incr;
3000 	if (next == 0 || next > syncer_maxdelay) {
3001 		start /= 2;
3002 		incr /= 2;
3003 		if (start == 0) {
3004 			start = syncer_maxdelay / 2;
3005 			incr = syncer_maxdelay;
3006 		}
3007 		next = start;
3008 	}
3009 	VI_LOCK(vp);
3010 	vn_syncer_add_to_worklist(&vp->v_bufobj,
3011 	    syncdelay > 0 ? next % syncdelay : 0);
3012 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3013 	mtx_lock(&sync_mtx);
3014 	sync_vnode_count++;
3015 	mtx_unlock(&sync_mtx);
3016 	VI_UNLOCK(vp);
3017 	mp->mnt_syncer = vp;
3018 	return (0);
3019 }
3020 
3021 /*
3022  * Do a lazy sync of the filesystem.
3023  */
3024 static int
3025 sync_fsync(struct vop_fsync_args *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(struct vop_inactive_args *ap)
3076 {
3077 
3078 	vgone(ap->a_vp);
3079 	return (0);
3080 }
3081 
3082 /*
3083  * The syncer vnode is no longer needed and is being decommissioned.
3084  *
3085  * Modifications to the worklist must be protected by sync_mtx.
3086  */
3087 static int
3088 sync_reclaim(struct vop_reclaim_args *ap)
3089 {
3090 	struct vnode *vp = ap->a_vp;
3091 	struct bufobj *bo;
3092 
3093 	VI_LOCK(vp);
3094 	bo = &vp->v_bufobj;
3095 	vp->v_mount->mnt_syncer = NULL;
3096 	if (bo->bo_flag & BO_ONWORKLST) {
3097 		mtx_lock(&sync_mtx);
3098 		LIST_REMOVE(bo, bo_synclist);
3099  		syncer_worklist_len--;
3100 		sync_vnode_count--;
3101 		mtx_unlock(&sync_mtx);
3102 		bo->bo_flag &= ~BO_ONWORKLST;
3103 	}
3104 	VI_UNLOCK(vp);
3105 
3106 	return (0);
3107 }
3108 
3109 /*
3110  * Check if vnode represents a disk device
3111  */
3112 int
3113 vn_isdisk(struct vnode *vp, int *errp)
3114 {
3115 	int error;
3116 
3117 	error = 0;
3118 	dev_lock();
3119 	if (vp->v_type != VCHR)
3120 		error = ENOTBLK;
3121 	else if (vp->v_rdev == NULL)
3122 		error = ENXIO;
3123 	else if (vp->v_rdev->si_devsw == NULL)
3124 		error = ENXIO;
3125 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3126 		error = ENOTBLK;
3127 	dev_unlock();
3128 	if (errp != NULL)
3129 		*errp = error;
3130 	return (error == 0);
3131 }
3132 
3133 /*
3134  * Common filesystem object access control check routine.  Accepts a
3135  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3136  * and optional call-by-reference privused argument allowing vaccess()
3137  * to indicate to the caller whether privilege was used to satisfy the
3138  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3139  */
3140 int
3141 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3142     mode_t acc_mode, struct ucred *cred, int *privused)
3143 {
3144 	mode_t dac_granted;
3145 #ifdef CAPABILITIES
3146 	mode_t cap_granted;
3147 #endif
3148 
3149 	/*
3150 	 * Look for a normal, non-privileged way to access the file/directory
3151 	 * as requested.  If it exists, go with that.
3152 	 */
3153 
3154 	if (privused != NULL)
3155 		*privused = 0;
3156 
3157 	dac_granted = 0;
3158 
3159 	/* Check the owner. */
3160 	if (cred->cr_uid == file_uid) {
3161 		dac_granted |= VADMIN;
3162 		if (file_mode & S_IXUSR)
3163 			dac_granted |= VEXEC;
3164 		if (file_mode & S_IRUSR)
3165 			dac_granted |= VREAD;
3166 		if (file_mode & S_IWUSR)
3167 			dac_granted |= (VWRITE | VAPPEND);
3168 
3169 		if ((acc_mode & dac_granted) == acc_mode)
3170 			return (0);
3171 
3172 		goto privcheck;
3173 	}
3174 
3175 	/* Otherwise, check the groups (first match) */
3176 	if (groupmember(file_gid, cred)) {
3177 		if (file_mode & S_IXGRP)
3178 			dac_granted |= VEXEC;
3179 		if (file_mode & S_IRGRP)
3180 			dac_granted |= VREAD;
3181 		if (file_mode & S_IWGRP)
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 everyone else. */
3191 	if (file_mode & S_IXOTH)
3192 		dac_granted |= VEXEC;
3193 	if (file_mode & S_IROTH)
3194 		dac_granted |= VREAD;
3195 	if (file_mode & S_IWOTH)
3196 		dac_granted |= (VWRITE | VAPPEND);
3197 	if ((acc_mode & dac_granted) == acc_mode)
3198 		return (0);
3199 
3200 privcheck:
3201 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3202 		/* XXX audit: privilege used */
3203 		if (privused != NULL)
3204 			*privused = 1;
3205 		return (0);
3206 	}
3207 
3208 #ifdef CAPABILITIES
3209 	/*
3210 	 * Build a capability mask to determine if the set of capabilities
3211 	 * satisfies the requirements when combined with the granted mask
3212 	 * from above.
3213 	 * For each capability, if the capability is required, bitwise
3214 	 * or the request type onto the cap_granted mask.
3215 	 */
3216 	cap_granted = 0;
3217 
3218 	if (type == VDIR) {
3219 		/*
3220 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3221 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3222 		 */
3223 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3224 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3225 			cap_granted |= VEXEC;
3226 	} else {
3227 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3228 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3229 			cap_granted |= VEXEC;
3230 	}
3231 
3232 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3233 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3234 		cap_granted |= VREAD;
3235 
3236 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3237 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3238 		cap_granted |= (VWRITE | VAPPEND);
3239 
3240 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3241 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3242 		cap_granted |= VADMIN;
3243 
3244 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3245 		/* XXX audit: privilege used */
3246 		if (privused != NULL)
3247 			*privused = 1;
3248 		return (0);
3249 	}
3250 #endif
3251 
3252 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3253 }
3254 
3255 /*
3256  * Credential check based on process requesting service, and per-attribute
3257  * permissions.
3258  */
3259 int
3260 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3261     struct thread *td, int access)
3262 {
3263 
3264 	/*
3265 	 * Kernel-invoked always succeeds.
3266 	 */
3267 	if (cred == NOCRED)
3268 		return (0);
3269 
3270 	/*
3271 	 * Do not allow privileged processes in jail to directly
3272 	 * manipulate system attributes.
3273 	 *
3274 	 * XXX What capability should apply here?
3275 	 * Probably CAP_SYS_SETFFLAG.
3276 	 */
3277 	switch (attrnamespace) {
3278 	case EXTATTR_NAMESPACE_SYSTEM:
3279 		/* Potentially should be: return (EPERM); */
3280 		return (suser_cred(cred, 0));
3281 	case EXTATTR_NAMESPACE_USER:
3282 		return (VOP_ACCESS(vp, access, cred, td));
3283 	default:
3284 		return (EPERM);
3285 	}
3286 }
3287 
3288 #ifdef DEBUG_VFS_LOCKS
3289 /*
3290  * This only exists to supress warnings from unlocked specfs accesses.  It is
3291  * no longer ok to have an unlocked VFS.
3292  */
3293 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3294 
3295 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3296 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3297 
3298 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3299 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3300 
3301 int vfs_badlock_print = 1;	/* Print lock violations. */
3302 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3303 
3304 #ifdef KDB
3305 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3306 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3307 #endif
3308 
3309 static void
3310 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3311 {
3312 
3313 #ifdef KDB
3314 	if (vfs_badlock_backtrace)
3315 		kdb_backtrace();
3316 #endif
3317 	if (vfs_badlock_print)
3318 		printf("%s: %p %s\n", str, (void *)vp, msg);
3319 	if (vfs_badlock_ddb)
3320 		kdb_enter("lock violation");
3321 }
3322 
3323 void
3324 assert_vi_locked(struct vnode *vp, const char *str)
3325 {
3326 
3327 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3328 		vfs_badlock("interlock is not locked but should be", str, vp);
3329 }
3330 
3331 void
3332 assert_vi_unlocked(struct vnode *vp, const char *str)
3333 {
3334 
3335 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3336 		vfs_badlock("interlock is locked but should not be", str, vp);
3337 }
3338 
3339 void
3340 assert_vop_locked(struct vnode *vp, const char *str)
3341 {
3342 
3343 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3344 		vfs_badlock("is not locked but should be", str, vp);
3345 }
3346 
3347 void
3348 assert_vop_unlocked(struct vnode *vp, const char *str)
3349 {
3350 
3351 	if (vp && !IGNORE_LOCK(vp) &&
3352 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3353 		vfs_badlock("is locked but should not be", str, vp);
3354 }
3355 
3356 void
3357 assert_vop_elocked(struct vnode *vp, const char *str)
3358 {
3359 
3360 	if (vp && !IGNORE_LOCK(vp) &&
3361 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3362 		vfs_badlock("is not exclusive locked but should be", str, vp);
3363 }
3364 
3365 #if 0
3366 void
3367 assert_vop_elocked_other(struct vnode *vp, const char *str)
3368 {
3369 
3370 	if (vp && !IGNORE_LOCK(vp) &&
3371 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3372 		vfs_badlock("is not exclusive locked by another thread",
3373 		    str, vp);
3374 }
3375 
3376 void
3377 assert_vop_slocked(struct vnode *vp, const char *str)
3378 {
3379 
3380 	if (vp && !IGNORE_LOCK(vp) &&
3381 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3382 		vfs_badlock("is not locked shared but should be", str, vp);
3383 }
3384 #endif /* 0 */
3385 #endif /* DEBUG_VFS_LOCKS */
3386 
3387 void
3388 vop_rename_pre(void *ap)
3389 {
3390 	struct vop_rename_args *a = ap;
3391 
3392 #ifdef DEBUG_VFS_LOCKS
3393 	if (a->a_tvp)
3394 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3395 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3396 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3397 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3398 
3399 	/* Check the source (from). */
3400 	if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp)
3401 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3402 	if (a->a_tvp != a->a_fvp)
3403 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3404 
3405 	/* Check the target. */
3406 	if (a->a_tvp)
3407 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3408 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3409 #endif
3410 	if (a->a_tdvp != a->a_fdvp)
3411 		vhold(a->a_fdvp);
3412 	if (a->a_tvp != a->a_fvp)
3413 		vhold(a->a_fvp);
3414 	vhold(a->a_tdvp);
3415 	if (a->a_tvp)
3416 		vhold(a->a_tvp);
3417 }
3418 
3419 void
3420 vop_strategy_pre(void *ap)
3421 {
3422 #ifdef DEBUG_VFS_LOCKS
3423 	struct vop_strategy_args *a;
3424 	struct buf *bp;
3425 
3426 	a = ap;
3427 	bp = a->a_bp;
3428 
3429 	/*
3430 	 * Cluster ops lock their component buffers but not the IO container.
3431 	 */
3432 	if ((bp->b_flags & B_CLUSTER) != 0)
3433 		return;
3434 
3435 	if (BUF_REFCNT(bp) < 1) {
3436 		if (vfs_badlock_print)
3437 			printf(
3438 			    "VOP_STRATEGY: bp is not locked but should be\n");
3439 		if (vfs_badlock_ddb)
3440 			kdb_enter("lock violation");
3441 	}
3442 #endif
3443 }
3444 
3445 void
3446 vop_lookup_pre(void *ap)
3447 {
3448 #ifdef DEBUG_VFS_LOCKS
3449 	struct vop_lookup_args *a;
3450 	struct vnode *dvp;
3451 
3452 	a = ap;
3453 	dvp = a->a_dvp;
3454 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3455 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3456 #endif
3457 }
3458 
3459 void
3460 vop_lookup_post(void *ap, int rc)
3461 {
3462 #ifdef DEBUG_VFS_LOCKS
3463 	struct vop_lookup_args *a;
3464 	struct vnode *dvp;
3465 	struct vnode *vp;
3466 
3467 	a = ap;
3468 	dvp = a->a_dvp;
3469 	vp = *(a->a_vpp);
3470 
3471 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3472 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3473 
3474 	if (!rc)
3475 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3476 #endif
3477 }
3478 
3479 void
3480 vop_lock_pre(void *ap)
3481 {
3482 #ifdef DEBUG_VFS_LOCKS
3483 	struct vop_lock_args *a = ap;
3484 
3485 	if ((a->a_flags & LK_INTERLOCK) == 0)
3486 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3487 	else
3488 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3489 #endif
3490 }
3491 
3492 void
3493 vop_lock_post(void *ap, int rc)
3494 {
3495 #ifdef DEBUG_VFS_LOCKS
3496 	struct vop_lock_args *a = ap;
3497 
3498 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3499 	if (rc == 0)
3500 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3501 #endif
3502 }
3503 
3504 void
3505 vop_unlock_pre(void *ap)
3506 {
3507 #ifdef DEBUG_VFS_LOCKS
3508 	struct vop_unlock_args *a = ap;
3509 
3510 	if (a->a_flags & LK_INTERLOCK)
3511 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3512 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3513 #endif
3514 }
3515 
3516 void
3517 vop_unlock_post(void *ap, int rc)
3518 {
3519 #ifdef DEBUG_VFS_LOCKS
3520 	struct vop_unlock_args *a = ap;
3521 
3522 	if (a->a_flags & LK_INTERLOCK)
3523 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3524 #endif
3525 }
3526 
3527 void
3528 vop_create_post(void *ap, int rc)
3529 {
3530 	struct vop_create_args *a = ap;
3531 
3532 	if (!rc)
3533 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3534 }
3535 
3536 void
3537 vop_link_post(void *ap, int rc)
3538 {
3539 	struct vop_link_args *a = ap;
3540 
3541 	if (!rc) {
3542 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3543 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3544 	}
3545 }
3546 
3547 void
3548 vop_mkdir_post(void *ap, int rc)
3549 {
3550 	struct vop_mkdir_args *a = ap;
3551 
3552 	if (!rc)
3553 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3554 }
3555 
3556 void
3557 vop_mknod_post(void *ap, int rc)
3558 {
3559 	struct vop_mknod_args *a = ap;
3560 
3561 	if (!rc)
3562 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3563 }
3564 
3565 void
3566 vop_remove_post(void *ap, int rc)
3567 {
3568 	struct vop_remove_args *a = ap;
3569 
3570 	if (!rc) {
3571 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3572 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3573 	}
3574 }
3575 
3576 void
3577 vop_rename_post(void *ap, int rc)
3578 {
3579 	struct vop_rename_args *a = ap;
3580 
3581 	if (!rc) {
3582 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3583 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3584 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3585 		if (a->a_tvp)
3586 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3587 	}
3588 	if (a->a_tdvp != a->a_fdvp)
3589 		vdrop(a->a_fdvp);
3590 	if (a->a_tvp != a->a_fvp)
3591 		vdrop(a->a_fvp);
3592 	vdrop(a->a_tdvp);
3593 	if (a->a_tvp)
3594 		vdrop(a->a_tvp);
3595 }
3596 
3597 void
3598 vop_rmdir_post(void *ap, int rc)
3599 {
3600 	struct vop_rmdir_args *a = ap;
3601 
3602 	if (!rc) {
3603 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3604 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3605 	}
3606 }
3607 
3608 void
3609 vop_setattr_post(void *ap, int rc)
3610 {
3611 	struct vop_setattr_args *a = ap;
3612 
3613 	if (!rc)
3614 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3615 }
3616 
3617 void
3618 vop_symlink_post(void *ap, int rc)
3619 {
3620 	struct vop_symlink_args *a = ap;
3621 
3622 	if (!rc)
3623 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3624 }
3625 
3626 static struct knlist fs_knlist;
3627 
3628 static void
3629 vfs_event_init(void *arg)
3630 {
3631 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3632 }
3633 /* XXX - correct order? */
3634 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3635 
3636 void
3637 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3638 {
3639 
3640 	KNOTE_UNLOCKED(&fs_knlist, event);
3641 }
3642 
3643 static int	filt_fsattach(struct knote *kn);
3644 static void	filt_fsdetach(struct knote *kn);
3645 static int	filt_fsevent(struct knote *kn, long hint);
3646 
3647 struct filterops fs_filtops =
3648 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3649 
3650 static int
3651 filt_fsattach(struct knote *kn)
3652 {
3653 
3654 	kn->kn_flags |= EV_CLEAR;
3655 	knlist_add(&fs_knlist, kn, 0);
3656 	return (0);
3657 }
3658 
3659 static void
3660 filt_fsdetach(struct knote *kn)
3661 {
3662 
3663 	knlist_remove(&fs_knlist, kn, 0);
3664 }
3665 
3666 static int
3667 filt_fsevent(struct knote *kn, long hint)
3668 {
3669 
3670 	kn->kn_fflags |= hint;
3671 	return (kn->kn_fflags != 0);
3672 }
3673 
3674 static int
3675 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3676 {
3677 	struct vfsidctl vc;
3678 	int error;
3679 	struct mount *mp;
3680 
3681 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3682 	if (error)
3683 		return (error);
3684 	if (vc.vc_vers != VFS_CTL_VERS1)
3685 		return (EINVAL);
3686 	mp = vfs_getvfs(&vc.vc_fsid);
3687 	if (mp == NULL)
3688 		return (ENOENT);
3689 	/* ensure that a specific sysctl goes to the right filesystem. */
3690 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3691 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3692 		vfs_rel(mp);
3693 		return (EINVAL);
3694 	}
3695 	VCTLTOREQ(&vc, req);
3696 	error = VFS_SYSCTL(mp, vc.vc_op, req);
3697 	vfs_rel(mp);
3698 	return (error);
3699 }
3700 
3701 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3702         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3703 
3704 /*
3705  * Function to initialize a va_filerev field sensibly.
3706  * XXX: Wouldn't a random number make a lot more sense ??
3707  */
3708 u_quad_t
3709 init_va_filerev(void)
3710 {
3711 	struct bintime bt;
3712 
3713 	getbinuptime(&bt);
3714 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3715 }
3716 
3717 static int	filt_vfsread(struct knote *kn, long hint);
3718 static int	filt_vfswrite(struct knote *kn, long hint);
3719 static int	filt_vfsvnode(struct knote *kn, long hint);
3720 static void	filt_vfsdetach(struct knote *kn);
3721 static struct filterops vfsread_filtops =
3722 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3723 static struct filterops vfswrite_filtops =
3724 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3725 static struct filterops vfsvnode_filtops =
3726 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3727 
3728 static void
3729 vfs_knllock(void *arg)
3730 {
3731 	struct vnode *vp = arg;
3732 
3733 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
3734 }
3735 
3736 static void
3737 vfs_knlunlock(void *arg)
3738 {
3739 	struct vnode *vp = arg;
3740 
3741 	VOP_UNLOCK(vp, 0, curthread);
3742 }
3743 
3744 static int
3745 vfs_knllocked(void *arg)
3746 {
3747 	struct vnode *vp = arg;
3748 
3749 	return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE);
3750 }
3751 
3752 int
3753 vfs_kqfilter(struct vop_kqfilter_args *ap)
3754 {
3755 	struct vnode *vp = ap->a_vp;
3756 	struct knote *kn = ap->a_kn;
3757 	struct knlist *knl;
3758 
3759 	switch (kn->kn_filter) {
3760 	case EVFILT_READ:
3761 		kn->kn_fop = &vfsread_filtops;
3762 		break;
3763 	case EVFILT_WRITE:
3764 		kn->kn_fop = &vfswrite_filtops;
3765 		break;
3766 	case EVFILT_VNODE:
3767 		kn->kn_fop = &vfsvnode_filtops;
3768 		break;
3769 	default:
3770 		return (EINVAL);
3771 	}
3772 
3773 	kn->kn_hook = (caddr_t)vp;
3774 
3775 	if (vp->v_pollinfo == NULL)
3776 		v_addpollinfo(vp);
3777 	if (vp->v_pollinfo == NULL)
3778 		return (ENOMEM);
3779 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3780 	knlist_add(knl, kn, 0);
3781 
3782 	return (0);
3783 }
3784 
3785 /*
3786  * Detach knote from vnode
3787  */
3788 static void
3789 filt_vfsdetach(struct knote *kn)
3790 {
3791 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3792 
3793 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3794 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3795 }
3796 
3797 /*ARGSUSED*/
3798 static int
3799 filt_vfsread(struct knote *kn, long hint)
3800 {
3801 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3802 	struct vattr va;
3803 
3804 	/*
3805 	 * filesystem is gone, so set the EOF flag and schedule
3806 	 * the knote for deletion.
3807 	 */
3808 	if (hint == NOTE_REVOKE) {
3809 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3810 		return (1);
3811 	}
3812 
3813 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3814 		return (0);
3815 
3816 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3817 	return (kn->kn_data != 0);
3818 }
3819 
3820 /*ARGSUSED*/
3821 static int
3822 filt_vfswrite(struct knote *kn, long hint)
3823 {
3824 	/*
3825 	 * filesystem is gone, so set the EOF flag and schedule
3826 	 * the knote for deletion.
3827 	 */
3828 	if (hint == NOTE_REVOKE)
3829 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3830 
3831 	kn->kn_data = 0;
3832 	return (1);
3833 }
3834 
3835 static int
3836 filt_vfsvnode(struct knote *kn, long hint)
3837 {
3838 	if (kn->kn_sfflags & hint)
3839 		kn->kn_fflags |= hint;
3840 	if (hint == NOTE_REVOKE) {
3841 		kn->kn_flags |= EV_EOF;
3842 		return (1);
3843 	}
3844 	return (kn->kn_fflags != 0);
3845 }
3846 
3847 int
3848 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3849 {
3850 	int error;
3851 
3852 	if (dp->d_reclen > ap->a_uio->uio_resid)
3853 		return (ENAMETOOLONG);
3854 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3855 	if (error) {
3856 		if (ap->a_ncookies != NULL) {
3857 			if (ap->a_cookies != NULL)
3858 				free(ap->a_cookies, M_TEMP);
3859 			ap->a_cookies = NULL;
3860 			*ap->a_ncookies = 0;
3861 		}
3862 		return (error);
3863 	}
3864 	if (ap->a_ncookies == NULL)
3865 		return (0);
3866 
3867 	KASSERT(ap->a_cookies,
3868 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
3869 
3870 	*ap->a_cookies = realloc(*ap->a_cookies,
3871 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
3872 	(*ap->a_cookies)[*ap->a_ncookies] = off;
3873 	return (0);
3874 }
3875 
3876 /*
3877  * Mark for update the access time of the file if the filesystem
3878  * supports VA_MARK_ATIME.  This functionality is used by execve
3879  * and mmap, so we want to avoid the synchronous I/O implied by
3880  * directly setting va_atime for the sake of efficiency.
3881  */
3882 void
3883 vfs_mark_atime(struct vnode *vp, struct thread *td)
3884 {
3885 	struct vattr atimeattr;
3886 
3887 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
3888 		VATTR_NULL(&atimeattr);
3889 		atimeattr.va_vaflags |= VA_MARK_ATIME;
3890 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
3891 	}
3892 }
3893