xref: /freebsd/sys/kern/vfs_subr.c (revision d429ea332342fcb98d27a350d0c4944bf9aec3f9)
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/event.h>
53 #include <sys/eventhandler.h>
54 #include <sys/extattr.h>
55 #include <sys/file.h>
56 #include <sys/fcntl.h>
57 #include <sys/kdb.h>
58 #include <sys/kernel.h>
59 #include <sys/kthread.h>
60 #include <sys/mac.h>
61 #include <sys/malloc.h>
62 #include <sys/mount.h>
63 #include <sys/namei.h>
64 #include <sys/reboot.h>
65 #include <sys/sleepqueue.h>
66 #include <sys/stat.h>
67 #include <sys/sysctl.h>
68 #include <sys/syslog.h>
69 #include <sys/vmmeter.h>
70 #include <sys/vnode.h>
71 
72 #include <machine/stdarg.h>
73 
74 #include <vm/vm.h>
75 #include <vm/vm_object.h>
76 #include <vm/vm_extern.h>
77 #include <vm/pmap.h>
78 #include <vm/vm_map.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_kern.h>
81 #include <vm/uma.h>
82 
83 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
84 
85 static void	delmntque(struct vnode *vp);
86 static void	insmntque(struct vnode *vp, struct mount *mp);
87 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
88 		    int slpflag, int slptimeo);
89 static void	syncer_shutdown(void *arg, int howto);
90 static int	vtryrecycle(struct vnode *vp);
91 static void	vbusy(struct vnode *vp);
92 static void	vdropl(struct vnode *vp);
93 static void	vinactive(struct vnode *, struct thread *);
94 static void	v_incr_usecount(struct vnode *);
95 static void	v_decr_usecount(struct vnode *);
96 static void	v_decr_useonly(struct vnode *);
97 static void	vfree(struct vnode *);
98 static void	vnlru_free(int);
99 static void	vdestroy(struct vnode *);
100 static void	vgonel(struct vnode *);
101 
102 /*
103  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
104  * build.  Without mpsafevm the buffer cache can not run Giant free.
105  */
106 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__)
107 int mpsafe_vfs = 1;
108 #else
109 int mpsafe_vfs;
110 #endif
111 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
112 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
113     "MPSAFE VFS");
114 
115 /*
116  * Number of vnodes in existence.  Increased whenever getnewvnode()
117  * allocates a new vnode, never decreased.
118  */
119 static unsigned long	numvnodes;
120 
121 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
122 
123 /*
124  * Conversion tables for conversion from vnode types to inode formats
125  * and back.
126  */
127 enum vtype iftovt_tab[16] = {
128 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
129 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
130 };
131 int vttoif_tab[9] = {
132 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
133 	S_IFSOCK, S_IFIFO, S_IFMT,
134 };
135 
136 /*
137  * List of vnodes that are ready for recycling.
138  */
139 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
140 
141 /*
142  * Free vnode target.  Free vnodes may simply be files which have been stat'd
143  * but not read.  This is somewhat common, and a small cache of such files
144  * should be kept to avoid recreation costs.
145  */
146 static u_long wantfreevnodes;
147 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
148 /* Number of vnodes in the free list. */
149 static u_long freevnodes;
150 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
151 
152 /*
153  * Various variables used for debugging the new implementation of
154  * reassignbuf().
155  * XXX these are probably of (very) limited utility now.
156  */
157 static int reassignbufcalls;
158 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
159 
160 /*
161  * Cache for the mount type id assigned to NFS.  This is used for
162  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
163  */
164 int	nfs_mount_type = -1;
165 
166 /* To keep more than one thread at a time from running vfs_getnewfsid */
167 static struct mtx mntid_mtx;
168 
169 /*
170  * Lock for any access to the following:
171  *	vnode_free_list
172  *	numvnodes
173  *	freevnodes
174  */
175 static struct mtx vnode_free_list_mtx;
176 
177 /* Publicly exported FS */
178 struct nfs_public nfs_pub;
179 
180 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
181 static uma_zone_t vnode_zone;
182 static uma_zone_t vnodepoll_zone;
183 
184 /* Set to 1 to print out reclaim of active vnodes */
185 int	prtactive;
186 
187 /*
188  * The workitem queue.
189  *
190  * It is useful to delay writes of file data and filesystem metadata
191  * for tens of seconds so that quickly created and deleted files need
192  * not waste disk bandwidth being created and removed. To realize this,
193  * we append vnodes to a "workitem" queue. When running with a soft
194  * updates implementation, most pending metadata dependencies should
195  * not wait for more than a few seconds. Thus, mounted on block devices
196  * are delayed only about a half the time that file data is delayed.
197  * Similarly, directory updates are more critical, so are only delayed
198  * about a third the time that file data is delayed. Thus, there are
199  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
200  * one each second (driven off the filesystem syncer process). The
201  * syncer_delayno variable indicates the next queue that is to be processed.
202  * Items that need to be processed soon are placed in this queue:
203  *
204  *	syncer_workitem_pending[syncer_delayno]
205  *
206  * A delay of fifteen seconds is done by placing the request fifteen
207  * entries later in the queue:
208  *
209  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
210  *
211  */
212 static int syncer_delayno;
213 static long syncer_mask;
214 LIST_HEAD(synclist, bufobj);
215 static struct synclist *syncer_workitem_pending;
216 /*
217  * The sync_mtx protects:
218  *	bo->bo_synclist
219  *	sync_vnode_count
220  *	syncer_delayno
221  *	syncer_state
222  *	syncer_workitem_pending
223  *	syncer_worklist_len
224  *	rushjob
225  */
226 static struct mtx sync_mtx;
227 
228 #define SYNCER_MAXDELAY		32
229 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
230 static int syncdelay = 30;		/* max time to delay syncing data */
231 static int filedelay = 30;		/* time to delay syncing files */
232 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
233 static int dirdelay = 29;		/* time to delay syncing directories */
234 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
235 static int metadelay = 28;		/* time to delay syncing metadata */
236 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
237 static int rushjob;		/* number of slots to run ASAP */
238 static int stat_rush_requests;	/* number of times I/O speeded up */
239 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
240 
241 /*
242  * When shutting down the syncer, run it at four times normal speed.
243  */
244 #define SYNCER_SHUTDOWN_SPEEDUP		4
245 static int sync_vnode_count;
246 static int syncer_worklist_len;
247 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
248     syncer_state;
249 
250 /*
251  * Number of vnodes we want to exist at any one time.  This is mostly used
252  * to size hash tables in vnode-related code.  It is normally not used in
253  * getnewvnode(), as wantfreevnodes is normally nonzero.)
254  *
255  * XXX desiredvnodes is historical cruft and should not exist.
256  */
257 int desiredvnodes;
258 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
259     &desiredvnodes, 0, "Maximum number of vnodes");
260 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
261     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
262 static int vnlru_nowhere;
263 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
264     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
265 
266 /* Hook for calling soft updates. */
267 int (*softdep_process_worklist_hook)(struct mount *);
268 
269 /*
270  * Macros to control when a vnode is freed and recycled.  All require
271  * the vnode interlock.
272  */
273 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
274 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
275 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
276 
277 
278 /*
279  * Initialize the vnode management data structures.
280  */
281 #ifndef	MAXVNODES_MAX
282 #define	MAXVNODES_MAX	100000
283 #endif
284 static void
285 vntblinit(void *dummy __unused)
286 {
287 
288 	/*
289 	 * Desiredvnodes is a function of the physical memory size and
290 	 * the kernel's heap size.  Specifically, desiredvnodes scales
291 	 * in proportion to the physical memory size until two fifths
292 	 * of the kernel's heap size is consumed by vnodes and vm
293 	 * objects.
294 	 */
295 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
296 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
297 	if (desiredvnodes > MAXVNODES_MAX) {
298 		if (bootverbose)
299 			printf("Reducing kern.maxvnodes %d -> %d\n",
300 			    desiredvnodes, MAXVNODES_MAX);
301 		desiredvnodes = MAXVNODES_MAX;
302 	}
303 	wantfreevnodes = desiredvnodes / 4;
304 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
305 	TAILQ_INIT(&vnode_free_list);
306 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
307 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
308 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
309 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
310 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
311 	/*
312 	 * Initialize the filesystem syncer.
313 	 */
314 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
315 		&syncer_mask);
316 	syncer_maxdelay = syncer_mask + 1;
317 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
318 }
319 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
320 
321 
322 /*
323  * Mark a mount point as busy. Used to synchronize access and to delay
324  * unmounting. Interlock is not released on failure.
325  */
326 int
327 vfs_busy(mp, flags, interlkp, td)
328 	struct mount *mp;
329 	int flags;
330 	struct mtx *interlkp;
331 	struct thread *td;
332 {
333 	int lkflags;
334 
335 	MNT_ILOCK(mp);
336 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
337 		if (flags & LK_NOWAIT) {
338 			MNT_IUNLOCK(mp);
339 			return (ENOENT);
340 		}
341 		if (interlkp)
342 			mtx_unlock(interlkp);
343 		mp->mnt_kern_flag |= MNTK_MWAIT;
344 		/*
345 		 * Since all busy locks are shared except the exclusive
346 		 * lock granted when unmounting, the only place that a
347 		 * wakeup needs to be done is at the release of the
348 		 * exclusive lock at the end of dounmount.
349 		 */
350 		msleep(mp, MNT_MTX(mp), PVFS|PDROP, "vfs_busy", 0);
351 		if (interlkp)
352 			mtx_lock(interlkp);
353 		return (ENOENT);
354 	}
355 	if (interlkp)
356 		mtx_unlock(interlkp);
357 	lkflags = LK_SHARED | LK_INTERLOCK;
358 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
359 		panic("vfs_busy: unexpected lock failure");
360 	return (0);
361 }
362 
363 /*
364  * Free a busy filesystem.
365  */
366 void
367 vfs_unbusy(mp, td)
368 	struct mount *mp;
369 	struct thread *td;
370 {
371 
372 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
373 }
374 
375 /*
376  * Lookup a mount point by filesystem identifier.
377  */
378 struct mount *
379 vfs_getvfs(fsid)
380 	fsid_t *fsid;
381 {
382 	struct mount *mp;
383 
384 	mtx_lock(&mountlist_mtx);
385 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
386 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
387 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
388 			mtx_unlock(&mountlist_mtx);
389 			return (mp);
390 		}
391 	}
392 	mtx_unlock(&mountlist_mtx);
393 	return ((struct mount *) 0);
394 }
395 
396 /*
397  * Check if a user can access priveledged mount options.
398  */
399 int
400 vfs_suser(struct mount *mp, struct thread *td)
401 {
402 	int error;
403 
404 	if ((mp->mnt_flag & MNT_USER) == 0 ||
405 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
406 		if ((error = suser(td)) != 0)
407 			return (error);
408 	}
409 	return (0);
410 }
411 
412 /*
413  * Get a new unique fsid.  Try to make its val[0] unique, since this value
414  * will be used to create fake device numbers for stat().  Also try (but
415  * not so hard) make its val[0] unique mod 2^16, since some emulators only
416  * support 16-bit device numbers.  We end up with unique val[0]'s for the
417  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
418  *
419  * Keep in mind that several mounts may be running in parallel.  Starting
420  * the search one past where the previous search terminated is both a
421  * micro-optimization and a defense against returning the same fsid to
422  * different mounts.
423  */
424 void
425 vfs_getnewfsid(mp)
426 	struct mount *mp;
427 {
428 	static u_int16_t mntid_base;
429 	fsid_t tfsid;
430 	int mtype;
431 
432 	mtx_lock(&mntid_mtx);
433 	mtype = mp->mnt_vfc->vfc_typenum;
434 	tfsid.val[1] = mtype;
435 	mtype = (mtype & 0xFF) << 24;
436 	for (;;) {
437 		tfsid.val[0] = makedev(255,
438 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
439 		mntid_base++;
440 		if (vfs_getvfs(&tfsid) == NULL)
441 			break;
442 	}
443 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
444 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
445 	mtx_unlock(&mntid_mtx);
446 }
447 
448 /*
449  * Knob to control the precision of file timestamps:
450  *
451  *   0 = seconds only; nanoseconds zeroed.
452  *   1 = seconds and nanoseconds, accurate within 1/HZ.
453  *   2 = seconds and nanoseconds, truncated to microseconds.
454  * >=3 = seconds and nanoseconds, maximum precision.
455  */
456 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
457 
458 static int timestamp_precision = TSP_SEC;
459 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
460     &timestamp_precision, 0, "");
461 
462 /*
463  * Get a current timestamp.
464  */
465 void
466 vfs_timestamp(tsp)
467 	struct timespec *tsp;
468 {
469 	struct timeval tv;
470 
471 	switch (timestamp_precision) {
472 	case TSP_SEC:
473 		tsp->tv_sec = time_second;
474 		tsp->tv_nsec = 0;
475 		break;
476 	case TSP_HZ:
477 		getnanotime(tsp);
478 		break;
479 	case TSP_USEC:
480 		microtime(&tv);
481 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
482 		break;
483 	case TSP_NSEC:
484 	default:
485 		nanotime(tsp);
486 		break;
487 	}
488 }
489 
490 /*
491  * Set vnode attributes to VNOVAL
492  */
493 void
494 vattr_null(vap)
495 	struct vattr *vap;
496 {
497 
498 	vap->va_type = VNON;
499 	vap->va_size = VNOVAL;
500 	vap->va_bytes = VNOVAL;
501 	vap->va_mode = VNOVAL;
502 	vap->va_nlink = VNOVAL;
503 	vap->va_uid = VNOVAL;
504 	vap->va_gid = VNOVAL;
505 	vap->va_fsid = VNOVAL;
506 	vap->va_fileid = VNOVAL;
507 	vap->va_blocksize = VNOVAL;
508 	vap->va_rdev = VNOVAL;
509 	vap->va_atime.tv_sec = VNOVAL;
510 	vap->va_atime.tv_nsec = VNOVAL;
511 	vap->va_mtime.tv_sec = VNOVAL;
512 	vap->va_mtime.tv_nsec = VNOVAL;
513 	vap->va_ctime.tv_sec = VNOVAL;
514 	vap->va_ctime.tv_nsec = VNOVAL;
515 	vap->va_birthtime.tv_sec = VNOVAL;
516 	vap->va_birthtime.tv_nsec = VNOVAL;
517 	vap->va_flags = VNOVAL;
518 	vap->va_gen = VNOVAL;
519 	vap->va_vaflags = 0;
520 }
521 
522 /*
523  * This routine is called when we have too many vnodes.  It attempts
524  * to free <count> vnodes and will potentially free vnodes that still
525  * have VM backing store (VM backing store is typically the cause
526  * of a vnode blowout so we want to do this).  Therefore, this operation
527  * is not considered cheap.
528  *
529  * A number of conditions may prevent a vnode from being reclaimed.
530  * the buffer cache may have references on the vnode, a directory
531  * vnode may still have references due to the namei cache representing
532  * underlying files, or the vnode may be in active use.   It is not
533  * desireable to reuse such vnodes.  These conditions may cause the
534  * number of vnodes to reach some minimum value regardless of what
535  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
536  */
537 static int
538 vlrureclaim(struct mount *mp)
539 {
540 	struct thread *td;
541 	struct vnode *vp;
542 	int done;
543 	int trigger;
544 	int usevnodes;
545 	int count;
546 
547 	/*
548 	 * Calculate the trigger point, don't allow user
549 	 * screwups to blow us up.   This prevents us from
550 	 * recycling vnodes with lots of resident pages.  We
551 	 * aren't trying to free memory, we are trying to
552 	 * free vnodes.
553 	 */
554 	usevnodes = desiredvnodes;
555 	if (usevnodes <= 0)
556 		usevnodes = 1;
557 	trigger = cnt.v_page_count * 2 / usevnodes;
558 	done = 0;
559 	td = curthread;
560 	vn_start_write(NULL, &mp, V_WAIT);
561 	MNT_ILOCK(mp);
562 	count = mp->mnt_nvnodelistsize / 10 + 1;
563 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
564 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
565 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
566 		--count;
567 		if (!VI_TRYLOCK(vp))
568 			continue;
569 		/*
570 		 * If it's been deconstructed already, it's still
571 		 * referenced, or it exceeds the trigger, skip it.
572 		 */
573 		if ((vp->v_iflag & VI_DOOMED) != 0 || vp->v_usecount ||
574 		    !LIST_EMPTY(&(vp)->v_cache_src) || (vp->v_object != NULL &&
575 		    vp->v_object->resident_page_count > trigger)) {
576 			VI_UNLOCK(vp);
577 			continue;
578 		}
579 		MNT_IUNLOCK(mp);
580 		vholdl(vp);
581 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
582 			vdrop(vp);
583 			MNT_ILOCK(mp);
584 			continue;
585 		}
586 		VI_LOCK(vp);
587 		vgonel(vp);
588 		VOP_UNLOCK(vp, 0, td);
589 		vdropl(vp);
590 		done++;
591 		MNT_ILOCK(mp);
592 	}
593 	MNT_IUNLOCK(mp);
594 	vn_finished_write(mp);
595 	return done;
596 }
597 
598 /*
599  * Attempt to keep the free list at wantfreevnodes length.
600  */
601 static void
602 vnlru_free(int count)
603 {
604 	struct vnode *vp;
605 
606 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
607 	for (; count > 0; count--) {
608 		vp = TAILQ_FIRST(&vnode_free_list);
609 		/*
610 		 * The list can be modified while the free_list_mtx
611 		 * has been dropped and vp could be NULL here.
612 		 */
613 		if (!vp)
614 			break;
615 		VNASSERT(vp->v_op != NULL, vp,
616 		    ("vnlru_free: vnode already reclaimed."));
617 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
618 		/*
619 		 * Don't recycle if we can't get the interlock.
620 		 */
621 		if (!VI_TRYLOCK(vp)) {
622 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
623 			continue;
624 		}
625 		VNASSERT(VCANRECYCLE(vp), vp,
626 		    ("vp inconsistent on freelist"));
627 		freevnodes--;
628 		vp->v_iflag &= ~VI_FREE;
629 		vholdl(vp);
630 		mtx_unlock(&vnode_free_list_mtx);
631 		VI_UNLOCK(vp);
632 		vtryrecycle(vp);
633 		/*
634 		 * If the recycled succeeded this vdrop will actually free
635 		 * the vnode.  If not it will simply place it back on
636 		 * the free list.
637 		 */
638 		vdrop(vp);
639 		mtx_lock(&vnode_free_list_mtx);
640 	}
641 }
642 /*
643  * Attempt to recycle vnodes in a context that is always safe to block.
644  * Calling vlrurecycle() from the bowels of filesystem code has some
645  * interesting deadlock problems.
646  */
647 static struct proc *vnlruproc;
648 static int vnlruproc_sig;
649 
650 static void
651 vnlru_proc(void)
652 {
653 	struct mount *mp, *nmp;
654 	int done;
655 	struct proc *p = vnlruproc;
656 	struct thread *td = FIRST_THREAD_IN_PROC(p);
657 
658 	mtx_lock(&Giant);
659 
660 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
661 	    SHUTDOWN_PRI_FIRST);
662 
663 	for (;;) {
664 		kthread_suspend_check(p);
665 		mtx_lock(&vnode_free_list_mtx);
666 		if (freevnodes > wantfreevnodes)
667 			vnlru_free(freevnodes - wantfreevnodes);
668 		if (numvnodes <= desiredvnodes * 9 / 10) {
669 			vnlruproc_sig = 0;
670 			wakeup(&vnlruproc_sig);
671 			msleep(vnlruproc, &vnode_free_list_mtx,
672 			    PVFS|PDROP, "vlruwt", hz);
673 			continue;
674 		}
675 		mtx_unlock(&vnode_free_list_mtx);
676 		done = 0;
677 		mtx_lock(&mountlist_mtx);
678 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
679 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
680 				nmp = TAILQ_NEXT(mp, mnt_list);
681 				continue;
682 			}
683 			done += vlrureclaim(mp);
684 			mtx_lock(&mountlist_mtx);
685 			nmp = TAILQ_NEXT(mp, mnt_list);
686 			vfs_unbusy(mp, td);
687 		}
688 		mtx_unlock(&mountlist_mtx);
689 		if (done == 0) {
690 #if 0
691 			/* These messages are temporary debugging aids */
692 			if (vnlru_nowhere < 5)
693 				printf("vnlru process getting nowhere..\n");
694 			else if (vnlru_nowhere == 5)
695 				printf("vnlru process messages stopped.\n");
696 #endif
697 			vnlru_nowhere++;
698 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
699 		}
700 	}
701 }
702 
703 static struct kproc_desc vnlru_kp = {
704 	"vnlru",
705 	vnlru_proc,
706 	&vnlruproc
707 };
708 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
709 
710 /*
711  * Routines having to do with the management of the vnode table.
712  */
713 
714 static void
715 vdestroy(struct vnode *vp)
716 {
717 	struct bufobj *bo;
718 
719 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
720 	mtx_lock(&vnode_free_list_mtx);
721 	numvnodes--;
722 	mtx_unlock(&vnode_free_list_mtx);
723 	bo = &vp->v_bufobj;
724 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
725 	    ("cleaned vnode still on the free list."));
726 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
727 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
728 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
729 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
730 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
731 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
732 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
733 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
734 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
735 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
736 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
737 #ifdef MAC
738 	mac_destroy_vnode(vp);
739 #endif
740 	if (vp->v_pollinfo != NULL) {
741 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
742 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
743 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
744 	}
745 #ifdef INVARIANTS
746 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
747 	vp->v_op = NULL;
748 #endif
749 	lockdestroy(vp->v_vnlock);
750 	mtx_destroy(&vp->v_interlock);
751 	uma_zfree(vnode_zone, vp);
752 }
753 
754 /*
755  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
756  * before we actually vgone().  This function must be called with the vnode
757  * held to prevent the vnode from being returned to the free list midway
758  * through vgone().
759  */
760 static int
761 vtryrecycle(struct vnode *vp)
762 {
763 	struct thread *td = curthread;
764 	struct mount *vnmp;
765 
766 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
767 	VNASSERT(vp->v_holdcnt, vp,
768 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
769 	/*
770 	 * This vnode may found and locked via some other list, if so we
771 	 * can't recycle it yet.
772 	 */
773 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
774 		return (EWOULDBLOCK);
775 	/*
776 	 * Don't recycle if its filesystem is being suspended.
777 	 */
778 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
779 		VOP_UNLOCK(vp, 0, td);
780 		return (EBUSY);
781 	}
782 	/*
783 	 * If we got this far, we need to acquire the interlock and see if
784 	 * anyone picked up this vnode from another list.  If not, we will
785 	 * mark it with DOOMED via vgonel() so that anyone who does find it
786 	 * will skip over it.
787 	 */
788 	VI_LOCK(vp);
789 	if (vp->v_usecount) {
790 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
791 		vn_finished_write(vnmp);
792 		return (EBUSY);
793 	}
794 	if ((vp->v_iflag & VI_DOOMED) == 0)
795 		vgonel(vp);
796 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
797 	vn_finished_write(vnmp);
798 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
799 	return (0);
800 }
801 
802 /*
803  * Return the next vnode from the free list.
804  */
805 int
806 getnewvnode(tag, mp, vops, vpp)
807 	const char *tag;
808 	struct mount *mp;
809 	struct vop_vector *vops;
810 	struct vnode **vpp;
811 {
812 	struct vnode *vp = NULL;
813 	struct bufobj *bo;
814 
815 	mtx_lock(&vnode_free_list_mtx);
816 	/*
817 	 * Lend our context to reclaim vnodes if they've exceeded the max.
818 	 */
819 	if (freevnodes > wantfreevnodes)
820 		vnlru_free(1);
821 	/*
822 	 * Wait for available vnodes.
823 	 */
824 	if (numvnodes > desiredvnodes) {
825 		if (vnlruproc_sig == 0) {
826 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
827 			wakeup(vnlruproc);
828 		}
829 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
830 		    "vlruwk", hz);
831 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
832 		if (numvnodes > desiredvnodes) {
833 			mtx_unlock(&vnode_free_list_mtx);
834 			return (ENFILE);
835 		}
836 #endif
837 	}
838 	numvnodes++;
839 	mtx_unlock(&vnode_free_list_mtx);
840 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
841 	/*
842 	 * Setup locks.
843 	 */
844 	vp->v_vnlock = &vp->v_lock;
845 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
846 	/*
847 	 * By default, don't allow shared locks unless filesystems
848 	 * opt-in.
849 	 */
850 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
851 	/*
852 	 * Initialize bufobj.
853 	 */
854 	bo = &vp->v_bufobj;
855 	bo->__bo_vnode = vp;
856 	bo->bo_mtx = &vp->v_interlock;
857 	bo->bo_ops = &buf_ops_bio;
858 	bo->bo_private = vp;
859 	TAILQ_INIT(&bo->bo_clean.bv_hd);
860 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
861 	/*
862 	 * Initialize namecache.
863 	 */
864 	LIST_INIT(&vp->v_cache_src);
865 	TAILQ_INIT(&vp->v_cache_dst);
866 	/*
867 	 * Finalize various vnode identity bits.
868 	 */
869 	vp->v_type = VNON;
870 	vp->v_tag = tag;
871 	vp->v_op = vops;
872 	v_incr_usecount(vp);
873 	vp->v_data = 0;
874 #ifdef MAC
875 	mac_init_vnode(vp);
876 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
877 		mac_associate_vnode_singlelabel(mp, vp);
878 	else if (mp == NULL)
879 		printf("NULL mp in getnewvnode()\n");
880 #endif
881 	delmntque(vp);
882 	if (mp != NULL) {
883 		insmntque(vp, mp);
884 		bo->bo_bsize = mp->mnt_stat.f_iosize;
885 	}
886 
887 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
888 	*vpp = vp;
889 	return (0);
890 }
891 
892 /*
893  * Delete from old mount point vnode list, if on one.
894  */
895 static void
896 delmntque(struct vnode *vp)
897 {
898 	struct mount *mp;
899 
900 	if (vp->v_mount == NULL)
901 		return;
902 	mp = vp->v_mount;
903 	MNT_ILOCK(mp);
904 	vp->v_mount = NULL;
905 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
906 		("bad mount point vnode list size"));
907 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
908 	mp->mnt_nvnodelistsize--;
909 	MNT_IUNLOCK(mp);
910 }
911 
912 /*
913  * Insert into list of vnodes for the new mount point, if available.
914  */
915 static void
916 insmntque(struct vnode *vp, struct mount *mp)
917 {
918 
919 	vp->v_mount = mp;
920 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
921 	MNT_ILOCK(vp->v_mount);
922 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
923 	mp->mnt_nvnodelistsize++;
924 	MNT_IUNLOCK(vp->v_mount);
925 }
926 
927 /*
928  * Flush out and invalidate all buffers associated with a bufobj
929  * Called with the underlying object locked.
930  */
931 int
932 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, int slptimeo)
933 {
934 	int error;
935 
936 	BO_LOCK(bo);
937 	if (flags & V_SAVE) {
938 		error = bufobj_wwait(bo, slpflag, slptimeo);
939 		if (error) {
940 			BO_UNLOCK(bo);
941 			return (error);
942 		}
943 		if (bo->bo_dirty.bv_cnt > 0) {
944 			BO_UNLOCK(bo);
945 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
946 				return (error);
947 			/*
948 			 * XXX We could save a lock/unlock if this was only
949 			 * enabled under INVARIANTS
950 			 */
951 			BO_LOCK(bo);
952 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
953 				panic("vinvalbuf: dirty bufs");
954 		}
955 	}
956 	/*
957 	 * If you alter this loop please notice that interlock is dropped and
958 	 * reacquired in flushbuflist.  Special care is needed to ensure that
959 	 * no race conditions occur from this.
960 	 */
961 	do {
962 		error = flushbuflist(&bo->bo_clean,
963 		    flags, bo, slpflag, slptimeo);
964 		if (error == 0)
965 			error = flushbuflist(&bo->bo_dirty,
966 			    flags, bo, slpflag, slptimeo);
967 		if (error != 0 && error != EAGAIN) {
968 			BO_UNLOCK(bo);
969 			return (error);
970 		}
971 	} while (error != 0);
972 
973 	/*
974 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
975 	 * have write I/O in-progress but if there is a VM object then the
976 	 * VM object can also have read-I/O in-progress.
977 	 */
978 	do {
979 		bufobj_wwait(bo, 0, 0);
980 		BO_UNLOCK(bo);
981 		if (bo->bo_object != NULL) {
982 			VM_OBJECT_LOCK(bo->bo_object);
983 			vm_object_pip_wait(bo->bo_object, "bovlbx");
984 			VM_OBJECT_UNLOCK(bo->bo_object);
985 		}
986 		BO_LOCK(bo);
987 	} while (bo->bo_numoutput > 0);
988 	BO_UNLOCK(bo);
989 
990 	/*
991 	 * Destroy the copy in the VM cache, too.
992 	 */
993 	if (bo->bo_object != NULL) {
994 		VM_OBJECT_LOCK(bo->bo_object);
995 		vm_object_page_remove(bo->bo_object, 0, 0,
996 			(flags & V_SAVE) ? TRUE : FALSE);
997 		VM_OBJECT_UNLOCK(bo->bo_object);
998 	}
999 
1000 #ifdef INVARIANTS
1001 	BO_LOCK(bo);
1002 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1003 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1004 		panic("vinvalbuf: flush failed");
1005 	BO_UNLOCK(bo);
1006 #endif
1007 	return (0);
1008 }
1009 
1010 /*
1011  * Flush out and invalidate all buffers associated with a vnode.
1012  * Called with the underlying object locked.
1013  */
1014 int
1015 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, int slptimeo)
1016 {
1017 
1018 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1019 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1020 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1021 }
1022 
1023 /*
1024  * Flush out buffers on the specified list.
1025  *
1026  */
1027 static int
1028 flushbuflist(bufv, flags, bo, slpflag, slptimeo)
1029 	struct bufv *bufv;
1030 	int flags;
1031 	struct bufobj *bo;
1032 	int slpflag, slptimeo;
1033 {
1034 	struct buf *bp, *nbp;
1035 	int retval, error;
1036 
1037 	ASSERT_BO_LOCKED(bo);
1038 
1039 	retval = 0;
1040 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1041 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1042 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1043 			continue;
1044 		}
1045 		retval = EAGAIN;
1046 		error = BUF_TIMELOCK(bp,
1047 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1048 		    "flushbuf", slpflag, slptimeo);
1049 		if (error) {
1050 			BO_LOCK(bo);
1051 			return (error != ENOLCK ? error : EAGAIN);
1052 		}
1053 		KASSERT(bp->b_bufobj == bo,
1054 	            ("bp %p wrong b_bufobj %p should be %p",
1055 		    bp, bp->b_bufobj, bo));
1056 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1057 			BUF_UNLOCK(bp);
1058 			BO_LOCK(bo);
1059 			return (EAGAIN);
1060 		}
1061 		/*
1062 		 * XXX Since there are no node locks for NFS, I
1063 		 * believe there is a slight chance that a delayed
1064 		 * write will occur while sleeping just above, so
1065 		 * check for it.
1066 		 */
1067 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1068 		    (flags & V_SAVE)) {
1069 			bremfree(bp);
1070 			bp->b_flags |= B_ASYNC;
1071 			bwrite(bp);
1072 			BO_LOCK(bo);
1073 			return (EAGAIN);	/* XXX: why not loop ? */
1074 		}
1075 		bremfree(bp);
1076 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1077 		bp->b_flags &= ~B_ASYNC;
1078 		brelse(bp);
1079 		BO_LOCK(bo);
1080 	}
1081 	return (retval);
1082 }
1083 
1084 /*
1085  * Truncate a file's buffer and pages to a specified length.  This
1086  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1087  * sync activity.
1088  */
1089 int
1090 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize)
1091 {
1092 	struct buf *bp, *nbp;
1093 	int anyfreed;
1094 	int trunclbn;
1095 	struct bufobj *bo;
1096 
1097 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1098 	/*
1099 	 * Round up to the *next* lbn.
1100 	 */
1101 	trunclbn = (length + blksize - 1) / blksize;
1102 
1103 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1104 restart:
1105 	VI_LOCK(vp);
1106 	bo = &vp->v_bufobj;
1107 	anyfreed = 1;
1108 	for (;anyfreed;) {
1109 		anyfreed = 0;
1110 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1111 			if (bp->b_lblkno < trunclbn)
1112 				continue;
1113 			if (BUF_LOCK(bp,
1114 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1115 			    VI_MTX(vp)) == ENOLCK)
1116 				goto restart;
1117 
1118 			bremfree(bp);
1119 			bp->b_flags |= (B_INVAL | B_RELBUF);
1120 			bp->b_flags &= ~B_ASYNC;
1121 			brelse(bp);
1122 			anyfreed = 1;
1123 
1124 			if (nbp != NULL &&
1125 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1126 			    (nbp->b_vp != vp) ||
1127 			    (nbp->b_flags & B_DELWRI))) {
1128 				goto restart;
1129 			}
1130 			VI_LOCK(vp);
1131 		}
1132 
1133 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1134 			if (bp->b_lblkno < trunclbn)
1135 				continue;
1136 			if (BUF_LOCK(bp,
1137 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1138 			    VI_MTX(vp)) == ENOLCK)
1139 				goto restart;
1140 			bremfree(bp);
1141 			bp->b_flags |= (B_INVAL | B_RELBUF);
1142 			bp->b_flags &= ~B_ASYNC;
1143 			brelse(bp);
1144 			anyfreed = 1;
1145 			if (nbp != NULL &&
1146 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1147 			    (nbp->b_vp != vp) ||
1148 			    (nbp->b_flags & B_DELWRI) == 0)) {
1149 				goto restart;
1150 			}
1151 			VI_LOCK(vp);
1152 		}
1153 	}
1154 
1155 	if (length > 0) {
1156 restartsync:
1157 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1158 			if (bp->b_lblkno > 0)
1159 				continue;
1160 			/*
1161 			 * Since we hold the vnode lock this should only
1162 			 * fail if we're racing with the buf daemon.
1163 			 */
1164 			if (BUF_LOCK(bp,
1165 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1166 			    VI_MTX(vp)) == ENOLCK) {
1167 				goto restart;
1168 			}
1169 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1170 			    ("buf(%p) on dirty queue without DELWRI", bp));
1171 
1172 			bremfree(bp);
1173 			bawrite(bp);
1174 			VI_LOCK(vp);
1175 			goto restartsync;
1176 		}
1177 	}
1178 
1179 	bufobj_wwait(bo, 0, 0);
1180 	VI_UNLOCK(vp);
1181 	vnode_pager_setsize(vp, length);
1182 
1183 	return (0);
1184 }
1185 
1186 /*
1187  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1188  * 		 a vnode.
1189  *
1190  *	NOTE: We have to deal with the special case of a background bitmap
1191  *	buffer, a situation where two buffers will have the same logical
1192  *	block offset.  We want (1) only the foreground buffer to be accessed
1193  *	in a lookup and (2) must differentiate between the foreground and
1194  *	background buffer in the splay tree algorithm because the splay
1195  *	tree cannot normally handle multiple entities with the same 'index'.
1196  *	We accomplish this by adding differentiating flags to the splay tree's
1197  *	numerical domain.
1198  */
1199 static
1200 struct buf *
1201 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1202 {
1203 	struct buf dummy;
1204 	struct buf *lefttreemax, *righttreemin, *y;
1205 
1206 	if (root == NULL)
1207 		return (NULL);
1208 	lefttreemax = righttreemin = &dummy;
1209 	for (;;) {
1210 		if (lblkno < root->b_lblkno ||
1211 		    (lblkno == root->b_lblkno &&
1212 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1213 			if ((y = root->b_left) == NULL)
1214 				break;
1215 			if (lblkno < y->b_lblkno) {
1216 				/* Rotate right. */
1217 				root->b_left = y->b_right;
1218 				y->b_right = root;
1219 				root = y;
1220 				if ((y = root->b_left) == NULL)
1221 					break;
1222 			}
1223 			/* Link into the new root's right tree. */
1224 			righttreemin->b_left = root;
1225 			righttreemin = root;
1226 		} else if (lblkno > root->b_lblkno ||
1227 		    (lblkno == root->b_lblkno &&
1228 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1229 			if ((y = root->b_right) == NULL)
1230 				break;
1231 			if (lblkno > y->b_lblkno) {
1232 				/* Rotate left. */
1233 				root->b_right = y->b_left;
1234 				y->b_left = root;
1235 				root = y;
1236 				if ((y = root->b_right) == NULL)
1237 					break;
1238 			}
1239 			/* Link into the new root's left tree. */
1240 			lefttreemax->b_right = root;
1241 			lefttreemax = root;
1242 		} else {
1243 			break;
1244 		}
1245 		root = y;
1246 	}
1247 	/* Assemble the new root. */
1248 	lefttreemax->b_right = root->b_left;
1249 	righttreemin->b_left = root->b_right;
1250 	root->b_left = dummy.b_right;
1251 	root->b_right = dummy.b_left;
1252 	return (root);
1253 }
1254 
1255 static void
1256 buf_vlist_remove(struct buf *bp)
1257 {
1258 	struct buf *root;
1259 	struct bufv *bv;
1260 
1261 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1262 	ASSERT_BO_LOCKED(bp->b_bufobj);
1263 	if (bp->b_xflags & BX_VNDIRTY)
1264 		bv = &bp->b_bufobj->bo_dirty;
1265 	else
1266 		bv = &bp->b_bufobj->bo_clean;
1267 	if (bp != bv->bv_root) {
1268 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1269 		KASSERT(root == bp, ("splay lookup failed in remove"));
1270 	}
1271 	if (bp->b_left == NULL) {
1272 		root = bp->b_right;
1273 	} else {
1274 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1275 		root->b_right = bp->b_right;
1276 	}
1277 	bv->bv_root = root;
1278 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1279 	bv->bv_cnt--;
1280 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1281 }
1282 
1283 /*
1284  * Add the buffer to the sorted clean or dirty block list using a
1285  * splay tree algorithm.
1286  *
1287  * NOTE: xflags is passed as a constant, optimizing this inline function!
1288  */
1289 static void
1290 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1291 {
1292 	struct buf *root;
1293 	struct bufv *bv;
1294 
1295 	ASSERT_BO_LOCKED(bo);
1296 	bp->b_xflags |= xflags;
1297 	if (xflags & BX_VNDIRTY)
1298 		bv = &bo->bo_dirty;
1299 	else
1300 		bv = &bo->bo_clean;
1301 
1302 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1303 	if (root == NULL) {
1304 		bp->b_left = NULL;
1305 		bp->b_right = NULL;
1306 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1307 	} else if (bp->b_lblkno < root->b_lblkno ||
1308 	    (bp->b_lblkno == root->b_lblkno &&
1309 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1310 		bp->b_left = root->b_left;
1311 		bp->b_right = root;
1312 		root->b_left = NULL;
1313 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1314 	} else {
1315 		bp->b_right = root->b_right;
1316 		bp->b_left = root;
1317 		root->b_right = NULL;
1318 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1319 	}
1320 	bv->bv_cnt++;
1321 	bv->bv_root = bp;
1322 }
1323 
1324 /*
1325  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1326  * shadow buffers used in background bitmap writes.
1327  *
1328  * This code isn't quite efficient as it could be because we are maintaining
1329  * two sorted lists and do not know which list the block resides in.
1330  *
1331  * During a "make buildworld" the desired buffer is found at one of
1332  * the roots more than 60% of the time.  Thus, checking both roots
1333  * before performing either splay eliminates unnecessary splays on the
1334  * first tree splayed.
1335  */
1336 struct buf *
1337 gbincore(struct bufobj *bo, daddr_t lblkno)
1338 {
1339 	struct buf *bp;
1340 
1341 	ASSERT_BO_LOCKED(bo);
1342 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1343 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1344 		return (bp);
1345 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1346 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1347 		return (bp);
1348 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1349 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1350 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1351 			return (bp);
1352 	}
1353 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1354 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1355 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1356 			return (bp);
1357 	}
1358 	return (NULL);
1359 }
1360 
1361 /*
1362  * Associate a buffer with a vnode.
1363  */
1364 void
1365 bgetvp(struct vnode *vp, struct buf *bp)
1366 {
1367 
1368 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1369 
1370 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1371 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1372 	    ("bgetvp: bp already attached! %p", bp));
1373 
1374 	ASSERT_VI_LOCKED(vp, "bgetvp");
1375 	vholdl(vp);
1376 	bp->b_vp = vp;
1377 	bp->b_bufobj = &vp->v_bufobj;
1378 	/*
1379 	 * Insert onto list for new vnode.
1380 	 */
1381 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1382 }
1383 
1384 /*
1385  * Disassociate a buffer from a vnode.
1386  */
1387 void
1388 brelvp(struct buf *bp)
1389 {
1390 	struct bufobj *bo;
1391 	struct vnode *vp;
1392 
1393 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1394 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1395 
1396 	/*
1397 	 * Delete from old vnode list, if on one.
1398 	 */
1399 	vp = bp->b_vp;		/* XXX */
1400 	bo = bp->b_bufobj;
1401 	BO_LOCK(bo);
1402 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1403 		buf_vlist_remove(bp);
1404 	else
1405 		panic("brelvp: Buffer %p not on queue.", bp);
1406 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1407 		bo->bo_flag &= ~BO_ONWORKLST;
1408 		mtx_lock(&sync_mtx);
1409 		LIST_REMOVE(bo, bo_synclist);
1410  		syncer_worklist_len--;
1411 		mtx_unlock(&sync_mtx);
1412 	}
1413 	bp->b_vp = NULL;
1414 	bp->b_bufobj = NULL;
1415 	vdropl(vp);
1416 }
1417 
1418 /*
1419  * Add an item to the syncer work queue.
1420  */
1421 static void
1422 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1423 {
1424 	int slot;
1425 
1426 	ASSERT_BO_LOCKED(bo);
1427 
1428 	mtx_lock(&sync_mtx);
1429 	if (bo->bo_flag & BO_ONWORKLST)
1430 		LIST_REMOVE(bo, bo_synclist);
1431 	else {
1432 		bo->bo_flag |= BO_ONWORKLST;
1433  		syncer_worklist_len++;
1434 	}
1435 
1436 	if (delay > syncer_maxdelay - 2)
1437 		delay = syncer_maxdelay - 2;
1438 	slot = (syncer_delayno + delay) & syncer_mask;
1439 
1440 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1441 	mtx_unlock(&sync_mtx);
1442 }
1443 
1444 static int
1445 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1446 {
1447 	int error, len;
1448 
1449 	mtx_lock(&sync_mtx);
1450 	len = syncer_worklist_len - sync_vnode_count;
1451 	mtx_unlock(&sync_mtx);
1452 	error = SYSCTL_OUT(req, &len, sizeof(len));
1453 	return (error);
1454 }
1455 
1456 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1457     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1458 
1459 struct  proc *updateproc;
1460 static void sched_sync(void);
1461 static struct kproc_desc up_kp = {
1462 	"syncer",
1463 	sched_sync,
1464 	&updateproc
1465 };
1466 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1467 
1468 static int
1469 sync_vnode(struct bufobj *bo, struct thread *td)
1470 {
1471 	struct vnode *vp;
1472 	struct mount *mp;
1473 
1474 	vp = bo->__bo_vnode; 	/* XXX */
1475 	if (VOP_ISLOCKED(vp, NULL) != 0)
1476 		return (1);
1477 	if (VI_TRYLOCK(vp) == 0)
1478 		return (1);
1479 	/*
1480 	 * We use vhold in case the vnode does not
1481 	 * successfully sync.  vhold prevents the vnode from
1482 	 * going away when we unlock the sync_mtx so that
1483 	 * we can acquire the vnode interlock.
1484 	 */
1485 	vholdl(vp);
1486 	mtx_unlock(&sync_mtx);
1487 	VI_UNLOCK(vp);
1488 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1489 		vdrop(vp);
1490 		mtx_lock(&sync_mtx);
1491 		return (1);
1492 	}
1493 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1494 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1495 	VOP_UNLOCK(vp, 0, td);
1496 	vn_finished_write(mp);
1497 	VI_LOCK(vp);
1498 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1499 		/*
1500 		 * Put us back on the worklist.  The worklist
1501 		 * routine will remove us from our current
1502 		 * position and then add us back in at a later
1503 		 * position.
1504 		 */
1505 		vn_syncer_add_to_worklist(bo, syncdelay);
1506 	}
1507 	vdropl(vp);
1508 	mtx_lock(&sync_mtx);
1509 	return (0);
1510 }
1511 
1512 /*
1513  * System filesystem synchronizer daemon.
1514  */
1515 static void
1516 sched_sync(void)
1517 {
1518 	struct synclist *next;
1519 	struct synclist *slp;
1520 	struct bufobj *bo;
1521 	long starttime;
1522 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1523 	static int dummychan;
1524 	int last_work_seen;
1525 	int net_worklist_len;
1526 	int syncer_final_iter;
1527 	int first_printf;
1528 	int error;
1529 
1530 	mtx_lock(&Giant);
1531 	last_work_seen = 0;
1532 	syncer_final_iter = 0;
1533 	first_printf = 1;
1534 	syncer_state = SYNCER_RUNNING;
1535 	starttime = time_second;
1536 
1537 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1538 	    SHUTDOWN_PRI_LAST);
1539 
1540 	for (;;) {
1541 		mtx_lock(&sync_mtx);
1542 		if (syncer_state == SYNCER_FINAL_DELAY &&
1543 		    syncer_final_iter == 0) {
1544 			mtx_unlock(&sync_mtx);
1545 			kthread_suspend_check(td->td_proc);
1546 			mtx_lock(&sync_mtx);
1547 		}
1548 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1549 		if (syncer_state != SYNCER_RUNNING &&
1550 		    starttime != time_second) {
1551 			if (first_printf) {
1552 				printf("\nSyncing disks, vnodes remaining...");
1553 				first_printf = 0;
1554 			}
1555 			printf("%d ", net_worklist_len);
1556 		}
1557 		starttime = time_second;
1558 
1559 		/*
1560 		 * Push files whose dirty time has expired.  Be careful
1561 		 * of interrupt race on slp queue.
1562 		 *
1563 		 * Skip over empty worklist slots when shutting down.
1564 		 */
1565 		do {
1566 			slp = &syncer_workitem_pending[syncer_delayno];
1567 			syncer_delayno += 1;
1568 			if (syncer_delayno == syncer_maxdelay)
1569 				syncer_delayno = 0;
1570 			next = &syncer_workitem_pending[syncer_delayno];
1571 			/*
1572 			 * If the worklist has wrapped since the
1573 			 * it was emptied of all but syncer vnodes,
1574 			 * switch to the FINAL_DELAY state and run
1575 			 * for one more second.
1576 			 */
1577 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1578 			    net_worklist_len == 0 &&
1579 			    last_work_seen == syncer_delayno) {
1580 				syncer_state = SYNCER_FINAL_DELAY;
1581 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1582 			}
1583 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1584 		    syncer_worklist_len > 0);
1585 
1586 		/*
1587 		 * Keep track of the last time there was anything
1588 		 * on the worklist other than syncer vnodes.
1589 		 * Return to the SHUTTING_DOWN state if any
1590 		 * new work appears.
1591 		 */
1592 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1593 			last_work_seen = syncer_delayno;
1594 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1595 			syncer_state = SYNCER_SHUTTING_DOWN;
1596 		while ((bo = LIST_FIRST(slp)) != NULL) {
1597 			error = sync_vnode(bo, td);
1598 			if (error == 1) {
1599 				LIST_REMOVE(bo, bo_synclist);
1600 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1601 				continue;
1602 			}
1603 		}
1604 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1605 			syncer_final_iter--;
1606 		mtx_unlock(&sync_mtx);
1607 
1608 		/*
1609 		 * Do soft update processing.
1610 		 */
1611 		if (softdep_process_worklist_hook != NULL)
1612 			(*softdep_process_worklist_hook)(NULL);
1613 
1614 		/*
1615 		 * The variable rushjob allows the kernel to speed up the
1616 		 * processing of the filesystem syncer process. A rushjob
1617 		 * value of N tells the filesystem syncer to process the next
1618 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1619 		 * is used by the soft update code to speed up the filesystem
1620 		 * syncer process when the incore state is getting so far
1621 		 * ahead of the disk that the kernel memory pool is being
1622 		 * threatened with exhaustion.
1623 		 */
1624 		mtx_lock(&sync_mtx);
1625 		if (rushjob > 0) {
1626 			rushjob -= 1;
1627 			mtx_unlock(&sync_mtx);
1628 			continue;
1629 		}
1630 		mtx_unlock(&sync_mtx);
1631 		/*
1632 		 * Just sleep for a short period if time between
1633 		 * iterations when shutting down to allow some I/O
1634 		 * to happen.
1635 		 *
1636 		 * If it has taken us less than a second to process the
1637 		 * current work, then wait. Otherwise start right over
1638 		 * again. We can still lose time if any single round
1639 		 * takes more than two seconds, but it does not really
1640 		 * matter as we are just trying to generally pace the
1641 		 * filesystem activity.
1642 		 */
1643 		if (syncer_state != SYNCER_RUNNING)
1644 			tsleep(&dummychan, PPAUSE, "syncfnl",
1645 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1646 		else if (time_second == starttime)
1647 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1648 	}
1649 }
1650 
1651 /*
1652  * Request the syncer daemon to speed up its work.
1653  * We never push it to speed up more than half of its
1654  * normal turn time, otherwise it could take over the cpu.
1655  */
1656 int
1657 speedup_syncer()
1658 {
1659 	struct thread *td;
1660 	int ret = 0;
1661 
1662 	td = FIRST_THREAD_IN_PROC(updateproc);
1663 	sleepq_remove(td, &lbolt);
1664 	mtx_lock(&sync_mtx);
1665 	if (rushjob < syncdelay / 2) {
1666 		rushjob += 1;
1667 		stat_rush_requests += 1;
1668 		ret = 1;
1669 	}
1670 	mtx_unlock(&sync_mtx);
1671 	return (ret);
1672 }
1673 
1674 /*
1675  * Tell the syncer to speed up its work and run though its work
1676  * list several times, then tell it to shut down.
1677  */
1678 static void
1679 syncer_shutdown(void *arg, int howto)
1680 {
1681 	struct thread *td;
1682 
1683 	if (howto & RB_NOSYNC)
1684 		return;
1685 	td = FIRST_THREAD_IN_PROC(updateproc);
1686 	sleepq_remove(td, &lbolt);
1687 	mtx_lock(&sync_mtx);
1688 	syncer_state = SYNCER_SHUTTING_DOWN;
1689 	rushjob = 0;
1690 	mtx_unlock(&sync_mtx);
1691 	kproc_shutdown(arg, howto);
1692 }
1693 
1694 /*
1695  * Reassign a buffer from one vnode to another.
1696  * Used to assign file specific control information
1697  * (indirect blocks) to the vnode to which they belong.
1698  */
1699 void
1700 reassignbuf(struct buf *bp)
1701 {
1702 	struct vnode *vp;
1703 	struct bufobj *bo;
1704 	int delay;
1705 #ifdef INVARIANTS
1706 	struct bufv *bv;
1707 #endif
1708 
1709 	vp = bp->b_vp;
1710 	bo = bp->b_bufobj;
1711 	++reassignbufcalls;
1712 
1713 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1714 	    bp, bp->b_vp, bp->b_flags);
1715 	/*
1716 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1717 	 * is not fully linked in.
1718 	 */
1719 	if (bp->b_flags & B_PAGING)
1720 		panic("cannot reassign paging buffer");
1721 
1722 	/*
1723 	 * Delete from old vnode list, if on one.
1724 	 */
1725 	VI_LOCK(vp);
1726 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1727 		buf_vlist_remove(bp);
1728 	else
1729 		panic("reassignbuf: Buffer %p not on queue.", bp);
1730 	/*
1731 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1732 	 * of clean buffers.
1733 	 */
1734 	if (bp->b_flags & B_DELWRI) {
1735 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1736 			switch (vp->v_type) {
1737 			case VDIR:
1738 				delay = dirdelay;
1739 				break;
1740 			case VCHR:
1741 				delay = metadelay;
1742 				break;
1743 			default:
1744 				delay = filedelay;
1745 			}
1746 			vn_syncer_add_to_worklist(bo, delay);
1747 		}
1748 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1749 	} else {
1750 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1751 
1752 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1753 			mtx_lock(&sync_mtx);
1754 			LIST_REMOVE(bo, bo_synclist);
1755  			syncer_worklist_len--;
1756 			mtx_unlock(&sync_mtx);
1757 			bo->bo_flag &= ~BO_ONWORKLST;
1758 		}
1759 	}
1760 	VI_UNLOCK(vp);
1761 #ifdef INVARIANTS
1762 	bv = &bo->bo_clean;
1763 	bp = TAILQ_FIRST(&bv->bv_hd);
1764 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1765 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1766 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1767 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1768 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1769 	bv = &bo->bo_dirty;
1770 	bp = TAILQ_FIRST(&bv->bv_hd);
1771 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1772 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1773 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1774 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1775 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1776 #endif
1777 }
1778 
1779 /*
1780  * Increment the use and hold counts on the vnode, taking care to reference
1781  * the driver's usecount if this is a chardev.  The vholdl() will remove
1782  * the vnode from the free list if it is presently free.  Requires the
1783  * vnode interlock and returns with it held.
1784  */
1785 static void
1786 v_incr_usecount(struct vnode *vp)
1787 {
1788 
1789 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1790 	    vp, vp->v_holdcnt, vp->v_usecount);
1791 	vp->v_usecount++;
1792 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1793 		dev_lock();
1794 		vp->v_rdev->si_usecount++;
1795 		dev_unlock();
1796 	}
1797 	vholdl(vp);
1798 }
1799 
1800 /*
1801  * Decrement the vnode use and hold count along with the driver's usecount
1802  * if this is a chardev.  The vdropl() below releases the vnode interlock
1803  * as it may free the vnode.
1804  */
1805 static void
1806 v_decr_usecount(struct vnode *vp)
1807 {
1808 
1809 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1810 	    vp, vp->v_holdcnt, vp->v_usecount);
1811 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1812 	VNASSERT(vp->v_usecount > 0, vp,
1813 	    ("v_decr_usecount: negative usecount"));
1814 	vp->v_usecount--;
1815 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1816 		dev_lock();
1817 		vp->v_rdev->si_usecount--;
1818 		dev_unlock();
1819 	}
1820 	vdropl(vp);
1821 }
1822 
1823 /*
1824  * Decrement only the use count and driver use count.  This is intended to
1825  * be paired with a follow on vdropl() to release the remaining hold count.
1826  * In this way we may vgone() a vnode with a 0 usecount without risk of
1827  * having it end up on a free list because the hold count is kept above 0.
1828  */
1829 static void
1830 v_decr_useonly(struct vnode *vp)
1831 {
1832 
1833 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1834 	    vp, vp->v_holdcnt, vp->v_usecount);
1835 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1836 	VNASSERT(vp->v_usecount > 0, vp,
1837 	    ("v_decr_useonly: negative usecount"));
1838 	vp->v_usecount--;
1839 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1840 		dev_lock();
1841 		vp->v_rdev->si_usecount--;
1842 		dev_unlock();
1843 	}
1844 }
1845 
1846 /*
1847  * Grab a particular vnode from the free list, increment its
1848  * reference count and lock it. The vnode lock bit is set if the
1849  * vnode is being eliminated in vgone. The process is awakened
1850  * when the transition is completed, and an error returned to
1851  * indicate that the vnode is no longer usable (possibly having
1852  * been changed to a new filesystem type).
1853  */
1854 int
1855 vget(vp, flags, td)
1856 	struct vnode *vp;
1857 	int flags;
1858 	struct thread *td;
1859 {
1860 	int oweinact;
1861 	int oldflags;
1862 	int error;
1863 
1864 	error = 0;
1865 	oldflags = flags;
1866 	oweinact = 0;
1867 	if ((flags & LK_INTERLOCK) == 0)
1868 		VI_LOCK(vp);
1869 	/*
1870 	 * If the inactive call was deferred because vput() was called
1871 	 * with a shared lock, we have to do it here before another thread
1872 	 * gets a reference to data that should be dead.
1873 	 */
1874 	if (vp->v_iflag & VI_OWEINACT) {
1875 		if (flags & LK_NOWAIT) {
1876 			VI_UNLOCK(vp);
1877 			return (EBUSY);
1878 		}
1879 		flags &= ~LK_TYPE_MASK;
1880 		flags |= LK_EXCLUSIVE;
1881 		oweinact = 1;
1882 	}
1883 	v_incr_usecount(vp);
1884 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1885 		VI_LOCK(vp);
1886 		/*
1887 		 * must expand vrele here because we do not want
1888 		 * to call VOP_INACTIVE if the reference count
1889 		 * drops back to zero since it was never really
1890 		 * active.
1891 		 */
1892 		v_decr_usecount(vp);
1893 		return (error);
1894 	}
1895 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
1896 		panic("vget: vn_lock failed to return ENOENT\n");
1897 	if (oweinact) {
1898 		VI_LOCK(vp);
1899 		if (vp->v_iflag & VI_OWEINACT)
1900 			vinactive(vp, td);
1901 		VI_UNLOCK(vp);
1902 		if ((oldflags & LK_TYPE_MASK) == 0)
1903 			VOP_UNLOCK(vp, 0, td);
1904 	}
1905 	return (0);
1906 }
1907 
1908 /*
1909  * Increase the reference count of a vnode.
1910  */
1911 void
1912 vref(struct vnode *vp)
1913 {
1914 
1915 	VI_LOCK(vp);
1916 	v_incr_usecount(vp);
1917 	VI_UNLOCK(vp);
1918 }
1919 
1920 /*
1921  * Return reference count of a vnode.
1922  *
1923  * The results of this call are only guaranteed when some mechanism other
1924  * than the VI lock is used to stop other processes from gaining references
1925  * to the vnode.  This may be the case if the caller holds the only reference.
1926  * This is also useful when stale data is acceptable as race conditions may
1927  * be accounted for by some other means.
1928  */
1929 int
1930 vrefcnt(struct vnode *vp)
1931 {
1932 	int usecnt;
1933 
1934 	VI_LOCK(vp);
1935 	usecnt = vp->v_usecount;
1936 	VI_UNLOCK(vp);
1937 
1938 	return (usecnt);
1939 }
1940 
1941 
1942 /*
1943  * Vnode put/release.
1944  * If count drops to zero, call inactive routine and return to freelist.
1945  */
1946 void
1947 vrele(vp)
1948 	struct vnode *vp;
1949 {
1950 	struct thread *td = curthread;	/* XXX */
1951 
1952 	KASSERT(vp != NULL, ("vrele: null vp"));
1953 
1954 	VI_LOCK(vp);
1955 
1956 	/* Skip this v_writecount check if we're going to panic below. */
1957 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
1958 	    ("vrele: missed vn_close"));
1959 
1960 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1961 	    vp->v_usecount == 1)) {
1962 		v_decr_usecount(vp);
1963 		return;
1964 	}
1965 	if (vp->v_usecount != 1) {
1966 #ifdef DIAGNOSTIC
1967 		vprint("vrele: negative ref count", vp);
1968 #endif
1969 		VI_UNLOCK(vp);
1970 		panic("vrele: negative ref cnt");
1971 	}
1972 	/*
1973 	 * We want to hold the vnode until the inactive finishes to
1974 	 * prevent vgone() races.  We drop the use count here and the
1975 	 * hold count below when we're done.
1976 	 */
1977 	v_decr_useonly(vp);
1978 	/*
1979 	 * We must call VOP_INACTIVE with the node locked. Mark
1980 	 * as VI_DOINGINACT to avoid recursion.
1981 	 */
1982 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
1983 		VI_LOCK(vp);
1984 		vinactive(vp, td);
1985 		VOP_UNLOCK(vp, 0, td);
1986 	} else
1987 		VI_LOCK(vp);
1988 	vdropl(vp);
1989 }
1990 
1991 /*
1992  * Release an already locked vnode.  This give the same effects as
1993  * unlock+vrele(), but takes less time and avoids releasing and
1994  * re-aquiring the lock (as vrele() aquires the lock internally.)
1995  */
1996 void
1997 vput(vp)
1998 	struct vnode *vp;
1999 {
2000 	struct thread *td = curthread;	/* XXX */
2001 	int error;
2002 
2003 	KASSERT(vp != NULL, ("vput: null vp"));
2004 	ASSERT_VOP_LOCKED(vp, "vput");
2005 	VI_LOCK(vp);
2006 	/* Skip this v_writecount check if we're going to panic below. */
2007 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2008 	    ("vput: missed vn_close"));
2009 	error = 0;
2010 
2011 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2012 	    vp->v_usecount == 1)) {
2013 		VOP_UNLOCK(vp, 0, td);
2014 		v_decr_usecount(vp);
2015 		return;
2016 	}
2017 
2018 	if (vp->v_usecount != 1) {
2019 #ifdef DIAGNOSTIC
2020 		vprint("vput: negative ref count", vp);
2021 #endif
2022 		panic("vput: negative ref cnt");
2023 	}
2024 	/*
2025 	 * We want to hold the vnode until the inactive finishes to
2026 	 * prevent vgone() races.  We drop the use count here and the
2027 	 * hold count below when we're done.
2028 	 */
2029 	v_decr_useonly(vp);
2030 	vp->v_iflag |= VI_OWEINACT;
2031 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2032 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2033 		VI_LOCK(vp);
2034 		if (error)
2035 			goto done;
2036 	}
2037 	if (vp->v_iflag & VI_OWEINACT)
2038 		vinactive(vp, td);
2039 	VOP_UNLOCK(vp, 0, td);
2040 done:
2041 	vdropl(vp);
2042 }
2043 
2044 /*
2045  * Somebody doesn't want the vnode recycled.
2046  */
2047 void
2048 vhold(struct vnode *vp)
2049 {
2050 
2051 	VI_LOCK(vp);
2052 	vholdl(vp);
2053 	VI_UNLOCK(vp);
2054 }
2055 
2056 void
2057 vholdl(struct vnode *vp)
2058 {
2059 
2060 	vp->v_holdcnt++;
2061 	if (VSHOULDBUSY(vp))
2062 		vbusy(vp);
2063 }
2064 
2065 /*
2066  * Note that there is one less who cares about this vnode.  vdrop() is the
2067  * opposite of vhold().
2068  */
2069 void
2070 vdrop(struct vnode *vp)
2071 {
2072 
2073 	VI_LOCK(vp);
2074 	vdropl(vp);
2075 }
2076 
2077 /*
2078  * Drop the hold count of the vnode.  If this is the last reference to
2079  * the vnode we will free it if it has been vgone'd otherwise it is
2080  * placed on the free list.
2081  */
2082 static void
2083 vdropl(struct vnode *vp)
2084 {
2085 
2086 	if (vp->v_holdcnt <= 0)
2087 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2088 	vp->v_holdcnt--;
2089 	if (vp->v_holdcnt == 0) {
2090 		if (vp->v_iflag & VI_DOOMED) {
2091 			vdestroy(vp);
2092 			return;
2093 		} else
2094 			vfree(vp);
2095 	}
2096 	VI_UNLOCK(vp);
2097 }
2098 
2099 /*
2100  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2101  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2102  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2103  * failed lock upgrade.
2104  */
2105 static void
2106 vinactive(struct vnode *vp, struct thread *td)
2107 {
2108 
2109 	ASSERT_VOP_LOCKED(vp, "vinactive");
2110 	ASSERT_VI_LOCKED(vp, "vinactive");
2111 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2112 	    ("vinactive: recursed on VI_DOINGINACT"));
2113 	vp->v_iflag |= VI_DOINGINACT;
2114 	vp->v_iflag &= ~VI_OWEINACT;
2115 	VI_UNLOCK(vp);
2116 	VOP_INACTIVE(vp, td);
2117 	VI_LOCK(vp);
2118 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2119 	    ("vinactive: lost VI_DOINGINACT"));
2120 	vp->v_iflag &= ~VI_DOINGINACT;
2121 }
2122 
2123 /*
2124  * Remove any vnodes in the vnode table belonging to mount point mp.
2125  *
2126  * If FORCECLOSE is not specified, there should not be any active ones,
2127  * return error if any are found (nb: this is a user error, not a
2128  * system error). If FORCECLOSE is specified, detach any active vnodes
2129  * that are found.
2130  *
2131  * If WRITECLOSE is set, only flush out regular file vnodes open for
2132  * writing.
2133  *
2134  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2135  *
2136  * `rootrefs' specifies the base reference count for the root vnode
2137  * of this filesystem. The root vnode is considered busy if its
2138  * v_usecount exceeds this value. On a successful return, vflush(, td)
2139  * will call vrele() on the root vnode exactly rootrefs times.
2140  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2141  * be zero.
2142  */
2143 #ifdef DIAGNOSTIC
2144 static int busyprt = 0;		/* print out busy vnodes */
2145 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2146 #endif
2147 
2148 int
2149 vflush(mp, rootrefs, flags, td)
2150 	struct mount *mp;
2151 	int rootrefs;
2152 	int flags;
2153 	struct thread *td;
2154 {
2155 	struct vnode *vp, *nvp, *rootvp = NULL;
2156 	struct vattr vattr;
2157 	int busy = 0, error;
2158 
2159 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2160 	if (rootrefs > 0) {
2161 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2162 		    ("vflush: bad args"));
2163 		/*
2164 		 * Get the filesystem root vnode. We can vput() it
2165 		 * immediately, since with rootrefs > 0, it won't go away.
2166 		 */
2167 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2168 			return (error);
2169 		vput(rootvp);
2170 
2171 	}
2172 	MNT_ILOCK(mp);
2173 loop:
2174 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2175 
2176 		VI_LOCK(vp);
2177 		vholdl(vp);
2178 		MNT_IUNLOCK(mp);
2179 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2180 		if (error) {
2181 			vdrop(vp);
2182 			MNT_ILOCK(mp);
2183 			goto loop;
2184 		}
2185 		/*
2186 		 * Skip over a vnodes marked VV_SYSTEM.
2187 		 */
2188 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2189 			VOP_UNLOCK(vp, 0, td);
2190 			vdrop(vp);
2191 			MNT_ILOCK(mp);
2192 			continue;
2193 		}
2194 		/*
2195 		 * If WRITECLOSE is set, flush out unlinked but still open
2196 		 * files (even if open only for reading) and regular file
2197 		 * vnodes open for writing.
2198 		 */
2199 		if (flags & WRITECLOSE) {
2200 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2201 			VI_LOCK(vp);
2202 
2203 			if ((vp->v_type == VNON ||
2204 			    (error == 0 && vattr.va_nlink > 0)) &&
2205 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2206 				VOP_UNLOCK(vp, 0, td);
2207 				vdropl(vp);
2208 				MNT_ILOCK(mp);
2209 				continue;
2210 			}
2211 		} else
2212 			VI_LOCK(vp);
2213 		/*
2214 		 * With v_usecount == 0, all we need to do is clear out the
2215 		 * vnode data structures and we are done.
2216 		 *
2217 		 * If FORCECLOSE is set, forcibly close the vnode.
2218 		 */
2219 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2220 			VNASSERT(vp->v_usecount == 0 ||
2221 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2222 			    ("device VNODE %p is FORCECLOSED", vp));
2223 			vgonel(vp);
2224 		} else {
2225 			busy++;
2226 #ifdef DIAGNOSTIC
2227 			if (busyprt)
2228 				vprint("vflush: busy vnode", vp);
2229 #endif
2230 		}
2231 		VOP_UNLOCK(vp, 0, td);
2232 		vdropl(vp);
2233 		MNT_ILOCK(mp);
2234 	}
2235 	MNT_IUNLOCK(mp);
2236 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2237 		/*
2238 		 * If just the root vnode is busy, and if its refcount
2239 		 * is equal to `rootrefs', then go ahead and kill it.
2240 		 */
2241 		VI_LOCK(rootvp);
2242 		KASSERT(busy > 0, ("vflush: not busy"));
2243 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2244 		    ("vflush: usecount %d < rootrefs %d",
2245 		     rootvp->v_usecount, rootrefs));
2246 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2247 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2248 			vgone(rootvp);
2249 			VOP_UNLOCK(rootvp, 0, td);
2250 			busy = 0;
2251 		} else
2252 			VI_UNLOCK(rootvp);
2253 	}
2254 	if (busy)
2255 		return (EBUSY);
2256 	for (; rootrefs > 0; rootrefs--)
2257 		vrele(rootvp);
2258 	return (0);
2259 }
2260 
2261 /*
2262  * Recycle an unused vnode to the front of the free list.
2263  */
2264 int
2265 vrecycle(struct vnode *vp, struct thread *td)
2266 {
2267 	int recycled;
2268 
2269 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2270 	recycled = 0;
2271 	VI_LOCK(vp);
2272 	if (vp->v_usecount == 0) {
2273 		recycled = 1;
2274 		vgonel(vp);
2275 	}
2276 	VI_UNLOCK(vp);
2277 	return (recycled);
2278 }
2279 
2280 /*
2281  * Eliminate all activity associated with a vnode
2282  * in preparation for reuse.
2283  */
2284 void
2285 vgone(struct vnode *vp)
2286 {
2287 	VI_LOCK(vp);
2288 	vgonel(vp);
2289 	VI_UNLOCK(vp);
2290 }
2291 
2292 /*
2293  * vgone, with the vp interlock held.
2294  */
2295 void
2296 vgonel(struct vnode *vp)
2297 {
2298 	struct thread *td;
2299 	int oweinact;
2300 	int active;
2301 
2302 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2303 	ASSERT_VOP_LOCKED(vp, "vgonel");
2304 	ASSERT_VI_LOCKED(vp, "vgonel");
2305 #if 0
2306 	/* XXX Need to fix ttyvp before I enable this. */
2307 	VNASSERT(vp->v_holdcnt, vp,
2308 	    ("vgonel: vp %p has no reference.", vp));
2309 #endif
2310 	td = curthread;
2311 
2312 	/*
2313 	 * Don't vgonel if we're already doomed.
2314 	 */
2315 	if (vp->v_iflag & VI_DOOMED) {
2316 		VI_UNLOCK(vp);
2317 		return;
2318 	}
2319 	vp->v_iflag |= VI_DOOMED;
2320 	/*
2321 	 * Check to see if the vnode is in use.  If so, we have to call
2322 	 * VOP_CLOSE() and VOP_INACTIVE().
2323 	 */
2324 	active = vp->v_usecount;
2325 	oweinact = (vp->v_iflag & VI_OWEINACT);
2326 	VI_UNLOCK(vp);
2327 	/*
2328 	 * Clean out any buffers associated with the vnode.
2329 	 * If the flush fails, just toss the buffers.
2330 	 */
2331 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2332 		(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2333 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2334 		vinvalbuf(vp, 0, td, 0, 0);
2335 
2336 	/*
2337 	 * If purging an active vnode, it must be closed and
2338 	 * deactivated before being reclaimed.
2339 	 */
2340 	if (active)
2341 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2342 	if (oweinact || active) {
2343 		VI_LOCK(vp);
2344 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2345 			vinactive(vp, td);
2346 		VI_UNLOCK(vp);
2347 	}
2348 	/*
2349 	 * Reclaim the vnode.
2350 	 */
2351 	if (VOP_RECLAIM(vp, td))
2352 		panic("vgone: cannot reclaim");
2353 	VNASSERT(vp->v_object == NULL, vp,
2354 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2355 	/*
2356 	 * Delete from old mount point vnode list.
2357 	 */
2358 	delmntque(vp);
2359 	cache_purge(vp);
2360 	/*
2361 	 * Done with purge, reset to the standard lock and invalidate
2362 	 * the vnode.
2363 	 */
2364 	VI_LOCK(vp);
2365 	vp->v_vnlock = &vp->v_lock;
2366 	vp->v_op = &dead_vnodeops;
2367 	vp->v_tag = "none";
2368 	vp->v_type = VBAD;
2369 }
2370 
2371 /*
2372  * Calculate the total number of references to a special device.
2373  */
2374 int
2375 vcount(vp)
2376 	struct vnode *vp;
2377 {
2378 	int count;
2379 
2380 	dev_lock();
2381 	count = vp->v_rdev->si_usecount;
2382 	dev_unlock();
2383 	return (count);
2384 }
2385 
2386 /*
2387  * Same as above, but using the struct cdev *as argument
2388  */
2389 int
2390 count_dev(dev)
2391 	struct cdev *dev;
2392 {
2393 	int count;
2394 
2395 	dev_lock();
2396 	count = dev->si_usecount;
2397 	dev_unlock();
2398 	return(count);
2399 }
2400 
2401 /*
2402  * Print out a description of a vnode.
2403  */
2404 static char *typename[] =
2405 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2406 
2407 void
2408 vn_printf(struct vnode *vp, const char *fmt, ...)
2409 {
2410 	va_list ap;
2411 	char buf[96];
2412 
2413 	va_start(ap, fmt);
2414 	vprintf(fmt, ap);
2415 	va_end(ap);
2416 	printf("%p: ", (void *)vp);
2417 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2418 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2419 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2420 	buf[0] = '\0';
2421 	buf[1] = '\0';
2422 	if (vp->v_vflag & VV_ROOT)
2423 		strcat(buf, "|VV_ROOT");
2424 	if (vp->v_vflag & VV_TEXT)
2425 		strcat(buf, "|VV_TEXT");
2426 	if (vp->v_vflag & VV_SYSTEM)
2427 		strcat(buf, "|VV_SYSTEM");
2428 	if (vp->v_iflag & VI_DOOMED)
2429 		strcat(buf, "|VI_DOOMED");
2430 	if (vp->v_iflag & VI_FREE)
2431 		strcat(buf, "|VI_FREE");
2432 	printf("    flags (%s)\n", buf + 1);
2433 	if (mtx_owned(VI_MTX(vp)))
2434 		printf(" VI_LOCKed");
2435 	if (vp->v_object != NULL)
2436 		printf("    v_object %p ref %d pages %d\n",
2437 		    vp->v_object, vp->v_object->ref_count,
2438 		    vp->v_object->resident_page_count);
2439 	printf("    ");
2440 	lockmgr_printinfo(vp->v_vnlock);
2441 	printf("\n");
2442 	if (vp->v_data != NULL)
2443 		VOP_PRINT(vp);
2444 }
2445 
2446 #ifdef DDB
2447 #include <ddb/ddb.h>
2448 /*
2449  * List all of the locked vnodes in the system.
2450  * Called when debugging the kernel.
2451  */
2452 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2453 {
2454 	struct mount *mp, *nmp;
2455 	struct vnode *vp;
2456 
2457 	/*
2458 	 * Note: because this is DDB, we can't obey the locking semantics
2459 	 * for these structures, which means we could catch an inconsistent
2460 	 * state and dereference a nasty pointer.  Not much to be done
2461 	 * about that.
2462 	 */
2463 	printf("Locked vnodes\n");
2464 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2465 		nmp = TAILQ_NEXT(mp, mnt_list);
2466 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2467 			if (VOP_ISLOCKED(vp, NULL))
2468 				vprint("", vp);
2469 		}
2470 		nmp = TAILQ_NEXT(mp, mnt_list);
2471 	}
2472 }
2473 #endif
2474 
2475 /*
2476  * Fill in a struct xvfsconf based on a struct vfsconf.
2477  */
2478 static void
2479 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2480 {
2481 
2482 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2483 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2484 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2485 	xvfsp->vfc_flags = vfsp->vfc_flags;
2486 	/*
2487 	 * These are unused in userland, we keep them
2488 	 * to not break binary compatibility.
2489 	 */
2490 	xvfsp->vfc_vfsops = NULL;
2491 	xvfsp->vfc_next = NULL;
2492 }
2493 
2494 /*
2495  * Top level filesystem related information gathering.
2496  */
2497 static int
2498 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2499 {
2500 	struct vfsconf *vfsp;
2501 	struct xvfsconf xvfsp;
2502 	int error;
2503 
2504 	error = 0;
2505 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2506 		bzero(&xvfsp, sizeof(xvfsp));
2507 		vfsconf2x(vfsp, &xvfsp);
2508 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2509 		if (error)
2510 			break;
2511 	}
2512 	return (error);
2513 }
2514 
2515 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2516     "S,xvfsconf", "List of all configured filesystems");
2517 
2518 #ifndef BURN_BRIDGES
2519 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2520 
2521 static int
2522 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2523 {
2524 	int *name = (int *)arg1 - 1;	/* XXX */
2525 	u_int namelen = arg2 + 1;	/* XXX */
2526 	struct vfsconf *vfsp;
2527 	struct xvfsconf xvfsp;
2528 
2529 	printf("WARNING: userland calling deprecated sysctl, "
2530 	    "please rebuild world\n");
2531 
2532 #if 1 || defined(COMPAT_PRELITE2)
2533 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2534 	if (namelen == 1)
2535 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2536 #endif
2537 
2538 	switch (name[1]) {
2539 	case VFS_MAXTYPENUM:
2540 		if (namelen != 2)
2541 			return (ENOTDIR);
2542 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2543 	case VFS_CONF:
2544 		if (namelen != 3)
2545 			return (ENOTDIR);	/* overloaded */
2546 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2547 			if (vfsp->vfc_typenum == name[2])
2548 				break;
2549 		if (vfsp == NULL)
2550 			return (EOPNOTSUPP);
2551 		bzero(&xvfsp, sizeof(xvfsp));
2552 		vfsconf2x(vfsp, &xvfsp);
2553 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2554 	}
2555 	return (EOPNOTSUPP);
2556 }
2557 
2558 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2559 	vfs_sysctl, "Generic filesystem");
2560 
2561 #if 1 || defined(COMPAT_PRELITE2)
2562 
2563 static int
2564 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2565 {
2566 	int error;
2567 	struct vfsconf *vfsp;
2568 	struct ovfsconf ovfs;
2569 
2570 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2571 		bzero(&ovfs, sizeof(ovfs));
2572 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2573 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2574 		ovfs.vfc_index = vfsp->vfc_typenum;
2575 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2576 		ovfs.vfc_flags = vfsp->vfc_flags;
2577 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2578 		if (error)
2579 			return error;
2580 	}
2581 	return 0;
2582 }
2583 
2584 #endif /* 1 || COMPAT_PRELITE2 */
2585 #endif /* !BURN_BRIDGES */
2586 
2587 #define KINFO_VNODESLOP		10
2588 #ifdef notyet
2589 /*
2590  * Dump vnode list (via sysctl).
2591  */
2592 /* ARGSUSED */
2593 static int
2594 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2595 {
2596 	struct xvnode *xvn;
2597 	struct thread *td = req->td;
2598 	struct mount *mp;
2599 	struct vnode *vp;
2600 	int error, len, n;
2601 
2602 	/*
2603 	 * Stale numvnodes access is not fatal here.
2604 	 */
2605 	req->lock = 0;
2606 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2607 	if (!req->oldptr)
2608 		/* Make an estimate */
2609 		return (SYSCTL_OUT(req, 0, len));
2610 
2611 	error = sysctl_wire_old_buffer(req, 0);
2612 	if (error != 0)
2613 		return (error);
2614 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2615 	n = 0;
2616 	mtx_lock(&mountlist_mtx);
2617 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2618 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2619 			continue;
2620 		MNT_ILOCK(mp);
2621 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2622 			if (n == len)
2623 				break;
2624 			vref(vp);
2625 			xvn[n].xv_size = sizeof *xvn;
2626 			xvn[n].xv_vnode = vp;
2627 			xvn[n].xv_id = 0;	/* XXX compat */
2628 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2629 			XV_COPY(usecount);
2630 			XV_COPY(writecount);
2631 			XV_COPY(holdcnt);
2632 			XV_COPY(mount);
2633 			XV_COPY(numoutput);
2634 			XV_COPY(type);
2635 #undef XV_COPY
2636 			xvn[n].xv_flag = vp->v_vflag;
2637 
2638 			switch (vp->v_type) {
2639 			case VREG:
2640 			case VDIR:
2641 			case VLNK:
2642 				break;
2643 			case VBLK:
2644 			case VCHR:
2645 				if (vp->v_rdev == NULL) {
2646 					vrele(vp);
2647 					continue;
2648 				}
2649 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2650 				break;
2651 			case VSOCK:
2652 				xvn[n].xv_socket = vp->v_socket;
2653 				break;
2654 			case VFIFO:
2655 				xvn[n].xv_fifo = vp->v_fifoinfo;
2656 				break;
2657 			case VNON:
2658 			case VBAD:
2659 			default:
2660 				/* shouldn't happen? */
2661 				vrele(vp);
2662 				continue;
2663 			}
2664 			vrele(vp);
2665 			++n;
2666 		}
2667 		MNT_IUNLOCK(mp);
2668 		mtx_lock(&mountlist_mtx);
2669 		vfs_unbusy(mp, td);
2670 		if (n == len)
2671 			break;
2672 	}
2673 	mtx_unlock(&mountlist_mtx);
2674 
2675 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2676 	free(xvn, M_TEMP);
2677 	return (error);
2678 }
2679 
2680 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2681 	0, 0, sysctl_vnode, "S,xvnode", "");
2682 #endif
2683 
2684 /*
2685  * Unmount all filesystems. The list is traversed in reverse order
2686  * of mounting to avoid dependencies.
2687  */
2688 void
2689 vfs_unmountall()
2690 {
2691 	struct mount *mp;
2692 	struct thread *td;
2693 	int error;
2694 
2695 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2696 	td = curthread;
2697 	/*
2698 	 * Since this only runs when rebooting, it is not interlocked.
2699 	 */
2700 	while(!TAILQ_EMPTY(&mountlist)) {
2701 		mp = TAILQ_LAST(&mountlist, mntlist);
2702 		error = dounmount(mp, MNT_FORCE, td);
2703 		if (error) {
2704 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2705 			printf("unmount of %s failed (",
2706 			    mp->mnt_stat.f_mntonname);
2707 			if (error == EBUSY)
2708 				printf("BUSY)\n");
2709 			else
2710 				printf("%d)\n", error);
2711 		} else {
2712 			/* The unmount has removed mp from the mountlist */
2713 		}
2714 	}
2715 }
2716 
2717 /*
2718  * perform msync on all vnodes under a mount point
2719  * the mount point must be locked.
2720  */
2721 void
2722 vfs_msync(struct mount *mp, int flags)
2723 {
2724 	struct vnode *vp, *nvp;
2725 	struct vm_object *obj;
2726 	int tries;
2727 
2728 	tries = 5;
2729 	MNT_ILOCK(mp);
2730 loop:
2731 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
2732 		if (vp->v_mount != mp) {
2733 			if (--tries > 0)
2734 				goto loop;
2735 			break;
2736 		}
2737 
2738 		VI_LOCK(vp);
2739 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2740 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2741 			MNT_IUNLOCK(mp);
2742 			if (!vget(vp,
2743 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2744 			    curthread)) {
2745 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2746 					vput(vp);
2747 					MNT_ILOCK(mp);
2748 					continue;
2749 				}
2750 
2751 				obj = vp->v_object;
2752 				if (obj != NULL) {
2753 					VM_OBJECT_LOCK(obj);
2754 					vm_object_page_clean(obj, 0, 0,
2755 					    flags == MNT_WAIT ?
2756 					    OBJPC_SYNC : OBJPC_NOSYNC);
2757 					VM_OBJECT_UNLOCK(obj);
2758 				}
2759 				vput(vp);
2760 			}
2761 			MNT_ILOCK(mp);
2762 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2763 				if (--tries > 0)
2764 					goto loop;
2765 				break;
2766 			}
2767 		} else
2768 			VI_UNLOCK(vp);
2769 	}
2770 	MNT_IUNLOCK(mp);
2771 }
2772 
2773 /*
2774  * Mark a vnode as free, putting it up for recycling.
2775  */
2776 static void
2777 vfree(struct vnode *vp)
2778 {
2779 
2780 	CTR1(KTR_VFS, "vfree vp %p", vp);
2781 	ASSERT_VI_LOCKED(vp, "vfree");
2782 	mtx_lock(&vnode_free_list_mtx);
2783 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2784 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2785 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2786 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2787 	    ("vfree: Freeing doomed vnode"));
2788 	if (vp->v_iflag & VI_AGE) {
2789 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2790 	} else {
2791 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2792 	}
2793 	freevnodes++;
2794 	vp->v_iflag &= ~VI_AGE;
2795 	vp->v_iflag |= VI_FREE;
2796 	mtx_unlock(&vnode_free_list_mtx);
2797 }
2798 
2799 /*
2800  * Opposite of vfree() - mark a vnode as in use.
2801  */
2802 static void
2803 vbusy(struct vnode *vp)
2804 {
2805 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2806 	ASSERT_VI_LOCKED(vp, "vbusy");
2807 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2808 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2809 
2810 	mtx_lock(&vnode_free_list_mtx);
2811 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2812 	freevnodes--;
2813 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2814 	mtx_unlock(&vnode_free_list_mtx);
2815 }
2816 
2817 /*
2818  * Initalize per-vnode helper structure to hold poll-related state.
2819  */
2820 void
2821 v_addpollinfo(struct vnode *vp)
2822 {
2823 	struct vpollinfo *vi;
2824 
2825 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2826 	if (vp->v_pollinfo != NULL) {
2827 		uma_zfree(vnodepoll_zone, vi);
2828 		return;
2829 	}
2830 	vp->v_pollinfo = vi;
2831 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2832 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note,
2833 	    &vp->v_pollinfo->vpi_lock);
2834 }
2835 
2836 /*
2837  * Record a process's interest in events which might happen to
2838  * a vnode.  Because poll uses the historic select-style interface
2839  * internally, this routine serves as both the ``check for any
2840  * pending events'' and the ``record my interest in future events''
2841  * functions.  (These are done together, while the lock is held,
2842  * to avoid race conditions.)
2843  */
2844 int
2845 vn_pollrecord(vp, td, events)
2846 	struct vnode *vp;
2847 	struct thread *td;
2848 	short events;
2849 {
2850 
2851 	if (vp->v_pollinfo == NULL)
2852 		v_addpollinfo(vp);
2853 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2854 	if (vp->v_pollinfo->vpi_revents & events) {
2855 		/*
2856 		 * This leaves events we are not interested
2857 		 * in available for the other process which
2858 		 * which presumably had requested them
2859 		 * (otherwise they would never have been
2860 		 * recorded).
2861 		 */
2862 		events &= vp->v_pollinfo->vpi_revents;
2863 		vp->v_pollinfo->vpi_revents &= ~events;
2864 
2865 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2866 		return events;
2867 	}
2868 	vp->v_pollinfo->vpi_events |= events;
2869 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2870 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2871 	return 0;
2872 }
2873 
2874 /*
2875  * Routine to create and manage a filesystem syncer vnode.
2876  */
2877 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2878 static int	sync_fsync(struct  vop_fsync_args *);
2879 static int	sync_inactive(struct  vop_inactive_args *);
2880 static int	sync_reclaim(struct  vop_reclaim_args *);
2881 
2882 static struct vop_vector sync_vnodeops = {
2883 	.vop_bypass =	VOP_EOPNOTSUPP,
2884 	.vop_close =	sync_close,		/* close */
2885 	.vop_fsync =	sync_fsync,		/* fsync */
2886 	.vop_inactive =	sync_inactive,	/* inactive */
2887 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2888 	.vop_lock =	vop_stdlock,	/* lock */
2889 	.vop_unlock =	vop_stdunlock,	/* unlock */
2890 	.vop_islocked =	vop_stdislocked,	/* islocked */
2891 };
2892 
2893 /*
2894  * Create a new filesystem syncer vnode for the specified mount point.
2895  */
2896 int
2897 vfs_allocate_syncvnode(mp)
2898 	struct mount *mp;
2899 {
2900 	struct vnode *vp;
2901 	static long start, incr, next;
2902 	int error;
2903 
2904 	/* Allocate a new vnode */
2905 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2906 		mp->mnt_syncer = NULL;
2907 		return (error);
2908 	}
2909 	vp->v_type = VNON;
2910 	/*
2911 	 * Place the vnode onto the syncer worklist. We attempt to
2912 	 * scatter them about on the list so that they will go off
2913 	 * at evenly distributed times even if all the filesystems
2914 	 * are mounted at once.
2915 	 */
2916 	next += incr;
2917 	if (next == 0 || next > syncer_maxdelay) {
2918 		start /= 2;
2919 		incr /= 2;
2920 		if (start == 0) {
2921 			start = syncer_maxdelay / 2;
2922 			incr = syncer_maxdelay;
2923 		}
2924 		next = start;
2925 	}
2926 	VI_LOCK(vp);
2927 	vn_syncer_add_to_worklist(&vp->v_bufobj,
2928 	    syncdelay > 0 ? next % syncdelay : 0);
2929 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
2930 	mtx_lock(&sync_mtx);
2931 	sync_vnode_count++;
2932 	mtx_unlock(&sync_mtx);
2933 	VI_UNLOCK(vp);
2934 	mp->mnt_syncer = vp;
2935 	return (0);
2936 }
2937 
2938 /*
2939  * Do a lazy sync of the filesystem.
2940  */
2941 static int
2942 sync_fsync(ap)
2943 	struct vop_fsync_args /* {
2944 		struct vnode *a_vp;
2945 		struct ucred *a_cred;
2946 		int a_waitfor;
2947 		struct thread *a_td;
2948 	} */ *ap;
2949 {
2950 	struct vnode *syncvp = ap->a_vp;
2951 	struct mount *mp = syncvp->v_mount;
2952 	struct thread *td = ap->a_td;
2953 	int error, asyncflag;
2954 	struct bufobj *bo;
2955 
2956 	/*
2957 	 * We only need to do something if this is a lazy evaluation.
2958 	 */
2959 	if (ap->a_waitfor != MNT_LAZY)
2960 		return (0);
2961 
2962 	/*
2963 	 * Move ourselves to the back of the sync list.
2964 	 */
2965 	bo = &syncvp->v_bufobj;
2966 	BO_LOCK(bo);
2967 	vn_syncer_add_to_worklist(bo, syncdelay);
2968 	BO_UNLOCK(bo);
2969 
2970 	/*
2971 	 * Walk the list of vnodes pushing all that are dirty and
2972 	 * not already on the sync list.
2973 	 */
2974 	mtx_lock(&mountlist_mtx);
2975 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
2976 		mtx_unlock(&mountlist_mtx);
2977 		return (0);
2978 	}
2979 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
2980 		vfs_unbusy(mp, td);
2981 		return (0);
2982 	}
2983 	asyncflag = mp->mnt_flag & MNT_ASYNC;
2984 	mp->mnt_flag &= ~MNT_ASYNC;
2985 	vfs_msync(mp, MNT_NOWAIT);
2986 	error = VFS_SYNC(mp, MNT_LAZY, td);
2987 	if (asyncflag)
2988 		mp->mnt_flag |= MNT_ASYNC;
2989 	vn_finished_write(mp);
2990 	vfs_unbusy(mp, td);
2991 	return (error);
2992 }
2993 
2994 /*
2995  * The syncer vnode is no referenced.
2996  */
2997 static int
2998 sync_inactive(ap)
2999 	struct vop_inactive_args /* {
3000 		struct vnode *a_vp;
3001 		struct thread *a_td;
3002 	} */ *ap;
3003 {
3004 
3005 	vgone(ap->a_vp);
3006 	return (0);
3007 }
3008 
3009 /*
3010  * The syncer vnode is no longer needed and is being decommissioned.
3011  *
3012  * Modifications to the worklist must be protected by sync_mtx.
3013  */
3014 static int
3015 sync_reclaim(ap)
3016 	struct vop_reclaim_args /* {
3017 		struct vnode *a_vp;
3018 	} */ *ap;
3019 {
3020 	struct vnode *vp = ap->a_vp;
3021 	struct bufobj *bo;
3022 
3023 	VI_LOCK(vp);
3024 	bo = &vp->v_bufobj;
3025 	vp->v_mount->mnt_syncer = NULL;
3026 	if (bo->bo_flag & BO_ONWORKLST) {
3027 		mtx_lock(&sync_mtx);
3028 		LIST_REMOVE(bo, bo_synclist);
3029  		syncer_worklist_len--;
3030 		sync_vnode_count--;
3031 		mtx_unlock(&sync_mtx);
3032 		bo->bo_flag &= ~BO_ONWORKLST;
3033 	}
3034 	VI_UNLOCK(vp);
3035 
3036 	return (0);
3037 }
3038 
3039 /*
3040  * Check if vnode represents a disk device
3041  */
3042 int
3043 vn_isdisk(vp, errp)
3044 	struct vnode *vp;
3045 	int *errp;
3046 {
3047 	int error;
3048 
3049 	error = 0;
3050 	dev_lock();
3051 	if (vp->v_type != VCHR)
3052 		error = ENOTBLK;
3053 	else if (vp->v_rdev == NULL)
3054 		error = ENXIO;
3055 	else if (vp->v_rdev->si_devsw == NULL)
3056 		error = ENXIO;
3057 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3058 		error = ENOTBLK;
3059 	dev_unlock();
3060 	if (errp != NULL)
3061 		*errp = error;
3062 	return (error == 0);
3063 }
3064 
3065 /*
3066  * Common filesystem object access control check routine.  Accepts a
3067  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3068  * and optional call-by-reference privused argument allowing vaccess()
3069  * to indicate to the caller whether privilege was used to satisfy the
3070  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3071  */
3072 int
3073 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3074 	enum vtype type;
3075 	mode_t file_mode;
3076 	uid_t file_uid;
3077 	gid_t file_gid;
3078 	mode_t acc_mode;
3079 	struct ucred *cred;
3080 	int *privused;
3081 {
3082 	mode_t dac_granted;
3083 #ifdef CAPABILITIES
3084 	mode_t cap_granted;
3085 #endif
3086 
3087 	/*
3088 	 * Look for a normal, non-privileged way to access the file/directory
3089 	 * as requested.  If it exists, go with that.
3090 	 */
3091 
3092 	if (privused != NULL)
3093 		*privused = 0;
3094 
3095 	dac_granted = 0;
3096 
3097 	/* Check the owner. */
3098 	if (cred->cr_uid == file_uid) {
3099 		dac_granted |= VADMIN;
3100 		if (file_mode & S_IXUSR)
3101 			dac_granted |= VEXEC;
3102 		if (file_mode & S_IRUSR)
3103 			dac_granted |= VREAD;
3104 		if (file_mode & S_IWUSR)
3105 			dac_granted |= (VWRITE | VAPPEND);
3106 
3107 		if ((acc_mode & dac_granted) == acc_mode)
3108 			return (0);
3109 
3110 		goto privcheck;
3111 	}
3112 
3113 	/* Otherwise, check the groups (first match) */
3114 	if (groupmember(file_gid, cred)) {
3115 		if (file_mode & S_IXGRP)
3116 			dac_granted |= VEXEC;
3117 		if (file_mode & S_IRGRP)
3118 			dac_granted |= VREAD;
3119 		if (file_mode & S_IWGRP)
3120 			dac_granted |= (VWRITE | VAPPEND);
3121 
3122 		if ((acc_mode & dac_granted) == acc_mode)
3123 			return (0);
3124 
3125 		goto privcheck;
3126 	}
3127 
3128 	/* Otherwise, check everyone else. */
3129 	if (file_mode & S_IXOTH)
3130 		dac_granted |= VEXEC;
3131 	if (file_mode & S_IROTH)
3132 		dac_granted |= VREAD;
3133 	if (file_mode & S_IWOTH)
3134 		dac_granted |= (VWRITE | VAPPEND);
3135 	if ((acc_mode & dac_granted) == acc_mode)
3136 		return (0);
3137 
3138 privcheck:
3139 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3140 		/* XXX audit: privilege used */
3141 		if (privused != NULL)
3142 			*privused = 1;
3143 		return (0);
3144 	}
3145 
3146 #ifdef CAPABILITIES
3147 	/*
3148 	 * Build a capability mask to determine if the set of capabilities
3149 	 * satisfies the requirements when combined with the granted mask
3150 	 * from above.
3151 	 * For each capability, if the capability is required, bitwise
3152 	 * or the request type onto the cap_granted mask.
3153 	 */
3154 	cap_granted = 0;
3155 
3156 	if (type == VDIR) {
3157 		/*
3158 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3159 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3160 		 */
3161 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3162 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3163 			cap_granted |= VEXEC;
3164 	} else {
3165 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3166 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3167 			cap_granted |= VEXEC;
3168 	}
3169 
3170 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3171 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3172 		cap_granted |= VREAD;
3173 
3174 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3175 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3176 		cap_granted |= (VWRITE | VAPPEND);
3177 
3178 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3179 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3180 		cap_granted |= VADMIN;
3181 
3182 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3183 		/* XXX audit: privilege used */
3184 		if (privused != NULL)
3185 			*privused = 1;
3186 		return (0);
3187 	}
3188 #endif
3189 
3190 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3191 }
3192 
3193 /*
3194  * Credential check based on process requesting service, and per-attribute
3195  * permissions.
3196  */
3197 int
3198 extattr_check_cred(struct vnode *vp, int attrnamespace,
3199     struct ucred *cred, struct thread *td, int access)
3200 {
3201 
3202 	/*
3203 	 * Kernel-invoked always succeeds.
3204 	 */
3205 	if (cred == NOCRED)
3206 		return (0);
3207 
3208 	/*
3209 	 * Do not allow privileged processes in jail to directly
3210 	 * manipulate system attributes.
3211 	 *
3212 	 * XXX What capability should apply here?
3213 	 * Probably CAP_SYS_SETFFLAG.
3214 	 */
3215 	switch (attrnamespace) {
3216 	case EXTATTR_NAMESPACE_SYSTEM:
3217 		/* Potentially should be: return (EPERM); */
3218 		return (suser_cred(cred, 0));
3219 	case EXTATTR_NAMESPACE_USER:
3220 		return (VOP_ACCESS(vp, access, cred, td));
3221 	default:
3222 		return (EPERM);
3223 	}
3224 }
3225 
3226 #ifdef DEBUG_VFS_LOCKS
3227 /*
3228  * This only exists to supress warnings from unlocked specfs accesses.  It is
3229  * no longer ok to have an unlocked VFS.
3230  */
3231 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3232 
3233 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3234 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3235 
3236 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3237 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3238 
3239 int vfs_badlock_print = 1;	/* Print lock violations. */
3240 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3241 
3242 #ifdef KDB
3243 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3244 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3245 #endif
3246 
3247 static void
3248 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3249 {
3250 
3251 #ifdef KDB
3252 	if (vfs_badlock_backtrace)
3253 		kdb_backtrace();
3254 #endif
3255 	if (vfs_badlock_print)
3256 		printf("%s: %p %s\n", str, (void *)vp, msg);
3257 	if (vfs_badlock_ddb)
3258 		kdb_enter("lock violation");
3259 }
3260 
3261 void
3262 assert_vi_locked(struct vnode *vp, const char *str)
3263 {
3264 
3265 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3266 		vfs_badlock("interlock is not locked but should be", str, vp);
3267 }
3268 
3269 void
3270 assert_vi_unlocked(struct vnode *vp, const char *str)
3271 {
3272 
3273 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3274 		vfs_badlock("interlock is locked but should not be", str, vp);
3275 }
3276 
3277 void
3278 assert_vop_locked(struct vnode *vp, const char *str)
3279 {
3280 
3281 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3282 		vfs_badlock("is not locked but should be", str, vp);
3283 }
3284 
3285 void
3286 assert_vop_unlocked(struct vnode *vp, const char *str)
3287 {
3288 
3289 	if (vp && !IGNORE_LOCK(vp) &&
3290 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3291 		vfs_badlock("is locked but should not be", str, vp);
3292 }
3293 
3294 void
3295 assert_vop_elocked(struct vnode *vp, const char *str)
3296 {
3297 
3298 	if (vp && !IGNORE_LOCK(vp) &&
3299 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3300 		vfs_badlock("is not exclusive locked but should be", str, vp);
3301 }
3302 
3303 #if 0
3304 void
3305 assert_vop_elocked_other(struct vnode *vp, const char *str)
3306 {
3307 
3308 	if (vp && !IGNORE_LOCK(vp) &&
3309 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3310 		vfs_badlock("is not exclusive locked by another thread",
3311 		    str, vp);
3312 }
3313 
3314 void
3315 assert_vop_slocked(struct vnode *vp, const char *str)
3316 {
3317 
3318 	if (vp && !IGNORE_LOCK(vp) &&
3319 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3320 		vfs_badlock("is not locked shared but should be", str, vp);
3321 }
3322 #endif /* 0 */
3323 #endif /* DEBUG_VFS_LOCKS */
3324 
3325 void
3326 vop_rename_pre(void *ap)
3327 {
3328 	struct vop_rename_args *a = ap;
3329 
3330 #ifdef DEBUG_VFS_LOCKS
3331 	if (a->a_tvp)
3332 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3333 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3334 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3335 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3336 
3337 	/* Check the source (from). */
3338 	if (a->a_tdvp != a->a_fdvp)
3339 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3340 	if (a->a_tvp != a->a_fvp)
3341 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3342 
3343 	/* Check the target. */
3344 	if (a->a_tvp)
3345 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3346 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3347 #endif
3348 	if (a->a_tdvp != a->a_fdvp)
3349 		vholdl(a->a_fdvp);
3350 	if (a->a_tvp != a->a_fvp)
3351 		vhold(a->a_fvp);
3352 	vhold(a->a_tdvp);
3353 	if (a->a_tvp)
3354 		vhold(a->a_tvp);
3355 }
3356 
3357 void
3358 vop_strategy_pre(void *ap)
3359 {
3360 #ifdef DEBUG_VFS_LOCKS
3361 	struct vop_strategy_args *a;
3362 	struct buf *bp;
3363 
3364 	a = ap;
3365 	bp = a->a_bp;
3366 
3367 	/*
3368 	 * Cluster ops lock their component buffers but not the IO container.
3369 	 */
3370 	if ((bp->b_flags & B_CLUSTER) != 0)
3371 		return;
3372 
3373 	if (BUF_REFCNT(bp) < 1) {
3374 		if (vfs_badlock_print)
3375 			printf(
3376 			    "VOP_STRATEGY: bp is not locked but should be\n");
3377 		if (vfs_badlock_ddb)
3378 			kdb_enter("lock violation");
3379 	}
3380 #endif
3381 }
3382 
3383 void
3384 vop_lookup_pre(void *ap)
3385 {
3386 #ifdef DEBUG_VFS_LOCKS
3387 	struct vop_lookup_args *a;
3388 	struct vnode *dvp;
3389 
3390 	a = ap;
3391 	dvp = a->a_dvp;
3392 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3393 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3394 #endif
3395 }
3396 
3397 void
3398 vop_lookup_post(void *ap, int rc)
3399 {
3400 #ifdef DEBUG_VFS_LOCKS
3401 	struct vop_lookup_args *a;
3402 	struct vnode *dvp;
3403 	struct vnode *vp;
3404 
3405 	a = ap;
3406 	dvp = a->a_dvp;
3407 	vp = *(a->a_vpp);
3408 
3409 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3410 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3411 
3412 	if (!rc)
3413 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3414 #endif
3415 }
3416 
3417 void
3418 vop_lock_pre(void *ap)
3419 {
3420 #ifdef DEBUG_VFS_LOCKS
3421 	struct vop_lock_args *a = ap;
3422 
3423 	if ((a->a_flags & LK_INTERLOCK) == 0)
3424 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3425 	else
3426 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3427 #endif
3428 }
3429 
3430 void
3431 vop_lock_post(void *ap, int rc)
3432 {
3433 #ifdef DEBUG_VFS_LOCKS
3434 	struct vop_lock_args *a = ap;
3435 
3436 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3437 	if (rc == 0)
3438 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3439 #endif
3440 }
3441 
3442 void
3443 vop_unlock_pre(void *ap)
3444 {
3445 #ifdef DEBUG_VFS_LOCKS
3446 	struct vop_unlock_args *a = ap;
3447 
3448 	if (a->a_flags & LK_INTERLOCK)
3449 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3450 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3451 #endif
3452 }
3453 
3454 void
3455 vop_unlock_post(void *ap, int rc)
3456 {
3457 #ifdef DEBUG_VFS_LOCKS
3458 	struct vop_unlock_args *a = ap;
3459 
3460 	if (a->a_flags & LK_INTERLOCK)
3461 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3462 #endif
3463 }
3464 
3465 void
3466 vop_create_post(void *ap, int rc)
3467 {
3468 	struct vop_create_args *a = ap;
3469 
3470 	if (!rc)
3471 		VFS_SEND_KNOTE(a->a_dvp, NOTE_WRITE);
3472 }
3473 
3474 void
3475 vop_link_post(void *ap, int rc)
3476 {
3477 	struct vop_link_args *a = ap;
3478 
3479 	if (!rc) {
3480 		VFS_SEND_KNOTE(a->a_vp, NOTE_LINK);
3481 		VFS_SEND_KNOTE(a->a_tdvp, NOTE_WRITE);
3482 	}
3483 }
3484 
3485 void
3486 vop_mkdir_post(void *ap, int rc)
3487 {
3488 	struct vop_mkdir_args *a = ap;
3489 
3490 	if (!rc)
3491 		VFS_SEND_KNOTE(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3492 }
3493 
3494 void
3495 vop_mknod_post(void *ap, int rc)
3496 {
3497 	struct vop_mknod_args *a = ap;
3498 
3499 	if (!rc)
3500 		VFS_SEND_KNOTE(a->a_dvp, NOTE_WRITE);
3501 }
3502 
3503 void
3504 vop_remove_post(void *ap, int rc)
3505 {
3506 	struct vop_remove_args *a = ap;
3507 
3508 	if (!rc) {
3509 		VFS_SEND_KNOTE(a->a_dvp, NOTE_WRITE);
3510 		VFS_SEND_KNOTE(a->a_vp, NOTE_DELETE);
3511 	}
3512 }
3513 
3514 void
3515 vop_rename_post(void *ap, int rc)
3516 {
3517 	struct vop_rename_args *a = ap;
3518 
3519 	if (!rc) {
3520 		VFS_SEND_KNOTE(a->a_fdvp, NOTE_WRITE);
3521 		VFS_SEND_KNOTE(a->a_tdvp, NOTE_WRITE);
3522 		VFS_SEND_KNOTE(a->a_fvp, NOTE_RENAME);
3523 		if (a->a_tvp)
3524 			VFS_SEND_KNOTE(a->a_tvp, NOTE_DELETE);
3525 	}
3526 	if (a->a_tdvp != a->a_fdvp)
3527 		vdrop(a->a_fdvp);
3528 	if (a->a_tvp != a->a_fvp)
3529 		vdrop(a->a_fvp);
3530 	vdrop(a->a_tdvp);
3531 	if (a->a_tvp)
3532 		vdrop(a->a_tvp);
3533 }
3534 
3535 void
3536 vop_rmdir_post(void *ap, int rc)
3537 {
3538 	struct vop_rmdir_args *a = ap;
3539 
3540 	if (!rc) {
3541 		VFS_SEND_KNOTE(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3542 		VFS_SEND_KNOTE(a->a_vp, NOTE_DELETE);
3543 	}
3544 }
3545 
3546 void
3547 vop_setattr_post(void *ap, int rc)
3548 {
3549 	struct vop_setattr_args *a = ap;
3550 
3551 	if (!rc)
3552 		VFS_SEND_KNOTE(a->a_vp, NOTE_ATTRIB);
3553 }
3554 
3555 void
3556 vop_symlink_post(void *ap, int rc)
3557 {
3558 	struct vop_symlink_args *a = ap;
3559 
3560 	if (!rc)
3561 		VFS_SEND_KNOTE(a->a_dvp, NOTE_WRITE);
3562 }
3563 
3564 static struct knlist fs_knlist;
3565 
3566 static void
3567 vfs_event_init(void *arg)
3568 {
3569 	knlist_init(&fs_knlist, NULL);
3570 }
3571 /* XXX - correct order? */
3572 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3573 
3574 void
3575 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3576 {
3577 
3578 	KNOTE_UNLOCKED(&fs_knlist, event);
3579 }
3580 
3581 static int	filt_fsattach(struct knote *kn);
3582 static void	filt_fsdetach(struct knote *kn);
3583 static int	filt_fsevent(struct knote *kn, long hint);
3584 
3585 struct filterops fs_filtops =
3586 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3587 
3588 static int
3589 filt_fsattach(struct knote *kn)
3590 {
3591 
3592 	kn->kn_flags |= EV_CLEAR;
3593 	knlist_add(&fs_knlist, kn, 0);
3594 	return (0);
3595 }
3596 
3597 static void
3598 filt_fsdetach(struct knote *kn)
3599 {
3600 
3601 	knlist_remove(&fs_knlist, kn, 0);
3602 }
3603 
3604 static int
3605 filt_fsevent(struct knote *kn, long hint)
3606 {
3607 
3608 	kn->kn_fflags |= hint;
3609 	return (kn->kn_fflags != 0);
3610 }
3611 
3612 static int
3613 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3614 {
3615 	struct vfsidctl vc;
3616 	int error;
3617 	struct mount *mp;
3618 
3619 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3620 	if (error)
3621 		return (error);
3622 	if (vc.vc_vers != VFS_CTL_VERS1)
3623 		return (EINVAL);
3624 	mp = vfs_getvfs(&vc.vc_fsid);
3625 	if (mp == NULL)
3626 		return (ENOENT);
3627 	/* ensure that a specific sysctl goes to the right filesystem. */
3628 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3629 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3630 		return (EINVAL);
3631 	}
3632 	VCTLTOREQ(&vc, req);
3633 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3634 }
3635 
3636 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3637         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3638 
3639 /*
3640  * Function to initialize a va_filerev field sensibly.
3641  * XXX: Wouldn't a random number make a lot more sense ??
3642  */
3643 u_quad_t
3644 init_va_filerev(void)
3645 {
3646 	struct bintime bt;
3647 
3648 	getbinuptime(&bt);
3649 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3650 }
3651 
3652 static int	filt_vfsread(struct knote *kn, long hint);
3653 static int	filt_vfswrite(struct knote *kn, long hint);
3654 static int	filt_vfsvnode(struct knote *kn, long hint);
3655 static void	filt_vfsdetach(struct knote *kn);
3656 
3657 static struct filterops vfsread_filtops =
3658 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3659 static struct filterops vfswrite_filtops =
3660 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3661 static struct filterops vfsvnode_filtops =
3662 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3663 
3664 int
3665 vfs_kqfilter(struct vop_kqfilter_args *ap)
3666 {
3667 	struct vnode *vp = ap->a_vp;
3668 	struct knote *kn = ap->a_kn;
3669 
3670 	switch (kn->kn_filter) {
3671 	case EVFILT_READ:
3672 		kn->kn_fop = &vfsread_filtops;
3673 		break;
3674 	case EVFILT_WRITE:
3675 		kn->kn_fop = &vfswrite_filtops;
3676 		break;
3677 	case EVFILT_VNODE:
3678 		kn->kn_fop = &vfsvnode_filtops;
3679 		break;
3680 	default:
3681 		return (1);
3682 	}
3683 
3684 	kn->kn_hook = (caddr_t)vp;
3685 
3686 	if (vp->v_pollinfo == NULL)
3687 		v_addpollinfo(vp);
3688 	if (vp->v_pollinfo == NULL)
3689 		return (ENOMEM);
3690 	knlist_add(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3691 
3692 	return (0);
3693 }
3694 
3695 /*
3696  * Detach knote from vnode
3697  */
3698 static void
3699 filt_vfsdetach(struct knote *kn)
3700 {
3701 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3702 
3703 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3704 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3705 }
3706 
3707 /*ARGSUSED*/
3708 static int
3709 filt_vfsread(struct knote *kn, long hint)
3710 {
3711 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3712 	struct vattr va;
3713 
3714 	/*
3715 	 * filesystem is gone, so set the EOF flag and schedule
3716 	 * the knote for deletion.
3717 	 */
3718 	if (hint == NOTE_REVOKE) {
3719 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3720 		return (1);
3721 	}
3722 
3723 	vn_lock(vp, LK_SHARED | LK_RETRY, curthread);
3724 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3725 		return (0);
3726 	if (VOP_UNLOCK(vp, 0, curthread))
3727 		return (0);
3728 
3729 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3730 	return (kn->kn_data != 0);
3731 }
3732 
3733 /*ARGSUSED*/
3734 static int
3735 filt_vfswrite(struct knote *kn, long hint)
3736 {
3737 	/*
3738 	 * filesystem is gone, so set the EOF flag and schedule
3739 	 * the knote for deletion.
3740 	 */
3741 	if (hint == NOTE_REVOKE)
3742 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3743 
3744 	kn->kn_data = 0;
3745 	return (1);
3746 }
3747 
3748 static int
3749 filt_vfsvnode(struct knote *kn, long hint)
3750 {
3751 	if (kn->kn_sfflags & hint)
3752 		kn->kn_fflags |= hint;
3753 	if (hint == NOTE_REVOKE) {
3754 		kn->kn_flags |= EV_EOF;
3755 		return (1);
3756 	}
3757 	return (kn->kn_fflags != 0);
3758 }
3759