xref: /freebsd/sys/kern/vfs_subr.c (revision c128b2d129a8e305b673ef5e3da76af1fb91ae60)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ddb.h"
45 #include "opt_mac.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/conf.h>
52 #include <sys/dirent.h>
53 #include <sys/event.h>
54 #include <sys/eventhandler.h>
55 #include <sys/extattr.h>
56 #include <sys/file.h>
57 #include <sys/fcntl.h>
58 #include <sys/kdb.h>
59 #include <sys/kernel.h>
60 #include <sys/kthread.h>
61 #include <sys/mac.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/namei.h>
65 #include <sys/reboot.h>
66 #include <sys/sleepqueue.h>
67 #include <sys/stat.h>
68 #include <sys/sysctl.h>
69 #include <sys/syslog.h>
70 #include <sys/vmmeter.h>
71 #include <sys/vnode.h>
72 
73 #include <machine/stdarg.h>
74 
75 #include <vm/vm.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/pmap.h>
79 #include <vm/vm_map.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_kern.h>
82 #include <vm/uma.h>
83 
84 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure");
85 
86 static void	delmntque(struct vnode *vp);
87 static void	insmntque(struct vnode *vp, struct mount *mp);
88 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
89 		    int slpflag, int slptimeo);
90 static void	syncer_shutdown(void *arg, int howto);
91 static int	vtryrecycle(struct vnode *vp);
92 static void	vbusy(struct vnode *vp);
93 static void	vdropl(struct vnode *vp);
94 static void	vinactive(struct vnode *, struct thread *);
95 static void	v_incr_usecount(struct vnode *);
96 static void	v_decr_usecount(struct vnode *);
97 static void	v_decr_useonly(struct vnode *);
98 static void	v_upgrade_usecount(struct vnode *);
99 static void	vfree(struct vnode *);
100 static void	vnlru_free(int);
101 static void	vdestroy(struct vnode *);
102 static void	vgonel(struct vnode *);
103 static void	vfs_knllock(void *arg);
104 static void	vfs_knlunlock(void *arg);
105 static int	vfs_knllocked(void *arg);
106 
107 
108 /*
109  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
110  * build.  Without mpsafevm the buffer cache can not run Giant free.
111  */
112 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__) || \
113 	defined(__ia64__) || defined(__sparc64__)
114 int mpsafe_vfs = 1;
115 #else
116 int mpsafe_vfs;
117 #endif
118 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
119 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
120     "MPSAFE VFS");
121 
122 /*
123  * Number of vnodes in existence.  Increased whenever getnewvnode()
124  * allocates a new vnode, never decreased.
125  */
126 static unsigned long	numvnodes;
127 
128 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
129 
130 /*
131  * Conversion tables for conversion from vnode types to inode formats
132  * and back.
133  */
134 enum vtype iftovt_tab[16] = {
135 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
136 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
137 };
138 int vttoif_tab[10] = {
139 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
140 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
141 };
142 
143 /*
144  * List of vnodes that are ready for recycling.
145  */
146 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
147 
148 /*
149  * Free vnode target.  Free vnodes may simply be files which have been stat'd
150  * but not read.  This is somewhat common, and a small cache of such files
151  * should be kept to avoid recreation costs.
152  */
153 static u_long wantfreevnodes;
154 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
155 /* Number of vnodes in the free list. */
156 static u_long freevnodes;
157 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
158 
159 /*
160  * Various variables used for debugging the new implementation of
161  * reassignbuf().
162  * XXX these are probably of (very) limited utility now.
163  */
164 static int reassignbufcalls;
165 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
166 
167 /*
168  * Cache for the mount type id assigned to NFS.  This is used for
169  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
170  */
171 int	nfs_mount_type = -1;
172 
173 /* To keep more than one thread at a time from running vfs_getnewfsid */
174 static struct mtx mntid_mtx;
175 
176 /*
177  * Lock for any access to the following:
178  *	vnode_free_list
179  *	numvnodes
180  *	freevnodes
181  */
182 static struct mtx vnode_free_list_mtx;
183 
184 /* Publicly exported FS */
185 struct nfs_public nfs_pub;
186 
187 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
188 static uma_zone_t vnode_zone;
189 static uma_zone_t vnodepoll_zone;
190 
191 /* Set to 1 to print out reclaim of active vnodes */
192 int	prtactive;
193 
194 /*
195  * The workitem queue.
196  *
197  * It is useful to delay writes of file data and filesystem metadata
198  * for tens of seconds so that quickly created and deleted files need
199  * not waste disk bandwidth being created and removed. To realize this,
200  * we append vnodes to a "workitem" queue. When running with a soft
201  * updates implementation, most pending metadata dependencies should
202  * not wait for more than a few seconds. Thus, mounted on block devices
203  * are delayed only about a half the time that file data is delayed.
204  * Similarly, directory updates are more critical, so are only delayed
205  * about a third the time that file data is delayed. Thus, there are
206  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
207  * one each second (driven off the filesystem syncer process). The
208  * syncer_delayno variable indicates the next queue that is to be processed.
209  * Items that need to be processed soon are placed in this queue:
210  *
211  *	syncer_workitem_pending[syncer_delayno]
212  *
213  * A delay of fifteen seconds is done by placing the request fifteen
214  * entries later in the queue:
215  *
216  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
217  *
218  */
219 static int syncer_delayno;
220 static long syncer_mask;
221 LIST_HEAD(synclist, bufobj);
222 static struct synclist *syncer_workitem_pending;
223 /*
224  * The sync_mtx protects:
225  *	bo->bo_synclist
226  *	sync_vnode_count
227  *	syncer_delayno
228  *	syncer_state
229  *	syncer_workitem_pending
230  *	syncer_worklist_len
231  *	rushjob
232  */
233 static struct mtx sync_mtx;
234 
235 #define SYNCER_MAXDELAY		32
236 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
237 static int syncdelay = 30;		/* max time to delay syncing data */
238 static int filedelay = 30;		/* time to delay syncing files */
239 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
240 static int dirdelay = 29;		/* time to delay syncing directories */
241 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
242 static int metadelay = 28;		/* time to delay syncing metadata */
243 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
244 static int rushjob;		/* number of slots to run ASAP */
245 static int stat_rush_requests;	/* number of times I/O speeded up */
246 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
247 
248 /*
249  * When shutting down the syncer, run it at four times normal speed.
250  */
251 #define SYNCER_SHUTDOWN_SPEEDUP		4
252 static int sync_vnode_count;
253 static int syncer_worklist_len;
254 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
255     syncer_state;
256 
257 /*
258  * Number of vnodes we want to exist at any one time.  This is mostly used
259  * to size hash tables in vnode-related code.  It is normally not used in
260  * getnewvnode(), as wantfreevnodes is normally nonzero.)
261  *
262  * XXX desiredvnodes is historical cruft and should not exist.
263  */
264 int desiredvnodes;
265 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
266     &desiredvnodes, 0, "Maximum number of vnodes");
267 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
268     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
269 static int vnlru_nowhere;
270 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
271     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
272 
273 /*
274  * Macros to control when a vnode is freed and recycled.  All require
275  * the vnode interlock.
276  */
277 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
278 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
279 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
280 
281 
282 /*
283  * Initialize the vnode management data structures.
284  */
285 #ifndef	MAXVNODES_MAX
286 #define	MAXVNODES_MAX	100000
287 #endif
288 static void
289 vntblinit(void *dummy __unused)
290 {
291 
292 	/*
293 	 * Desiredvnodes is a function of the physical memory size and
294 	 * the kernel's heap size.  Specifically, desiredvnodes scales
295 	 * in proportion to the physical memory size until two fifths
296 	 * of the kernel's heap size is consumed by vnodes and vm
297 	 * objects.
298 	 */
299 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
300 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
301 	if (desiredvnodes > MAXVNODES_MAX) {
302 		if (bootverbose)
303 			printf("Reducing kern.maxvnodes %d -> %d\n",
304 			    desiredvnodes, MAXVNODES_MAX);
305 		desiredvnodes = MAXVNODES_MAX;
306 	}
307 	wantfreevnodes = desiredvnodes / 4;
308 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
309 	TAILQ_INIT(&vnode_free_list);
310 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
311 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
312 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
314 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 	/*
316 	 * Initialize the filesystem syncer.
317 	 */
318 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
319 		&syncer_mask);
320 	syncer_maxdelay = syncer_mask + 1;
321 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
322 }
323 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
324 
325 
326 /*
327  * Mark a mount point as busy. Used to synchronize access and to delay
328  * unmounting. Interlock is not released on failure.
329  */
330 int
331 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp,
332     struct thread *td)
333 {
334 	int lkflags;
335 
336 	MNT_ILOCK(mp);
337 	MNT_REF(mp);
338 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
339 		if (flags & LK_NOWAIT) {
340 			MNT_REL(mp);
341 			MNT_IUNLOCK(mp);
342 			return (ENOENT);
343 		}
344 		if (interlkp)
345 			mtx_unlock(interlkp);
346 		mp->mnt_kern_flag |= MNTK_MWAIT;
347 		/*
348 		 * Since all busy locks are shared except the exclusive
349 		 * lock granted when unmounting, the only place that a
350 		 * wakeup needs to be done is at the release of the
351 		 * exclusive lock at the end of dounmount.
352 		 */
353 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
354 		MNT_REL(mp);
355 		MNT_IUNLOCK(mp);
356 		if (interlkp)
357 			mtx_lock(interlkp);
358 		return (ENOENT);
359 	}
360 	if (interlkp)
361 		mtx_unlock(interlkp);
362 	lkflags = LK_SHARED | LK_INTERLOCK;
363 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
364 		panic("vfs_busy: unexpected lock failure");
365 	vfs_rel(mp);
366 	return (0);
367 }
368 
369 /*
370  * Free a busy filesystem.
371  */
372 void
373 vfs_unbusy(struct mount *mp, struct thread *td)
374 {
375 
376 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
377 }
378 
379 /*
380  * Lookup a mount point by filesystem identifier.
381  */
382 struct mount *
383 vfs_getvfs(fsid_t *fsid)
384 {
385 	struct mount *mp;
386 
387 	mtx_lock(&mountlist_mtx);
388 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
389 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
390 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
391 			mtx_unlock(&mountlist_mtx);
392 			return (mp);
393 		}
394 	}
395 	mtx_unlock(&mountlist_mtx);
396 	return ((struct mount *) 0);
397 }
398 
399 /*
400  * Check if a user can access priveledged mount options.
401  */
402 int
403 vfs_suser(struct mount *mp, struct thread *td)
404 {
405 	int error;
406 
407 	if ((mp->mnt_flag & MNT_USER) == 0 ||
408 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
409 		if ((error = suser(td)) != 0)
410 			return (error);
411 	}
412 	return (0);
413 }
414 
415 /*
416  * Get a new unique fsid.  Try to make its val[0] unique, since this value
417  * will be used to create fake device numbers for stat().  Also try (but
418  * not so hard) make its val[0] unique mod 2^16, since some emulators only
419  * support 16-bit device numbers.  We end up with unique val[0]'s for the
420  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
421  *
422  * Keep in mind that several mounts may be running in parallel.  Starting
423  * the search one past where the previous search terminated is both a
424  * micro-optimization and a defense against returning the same fsid to
425  * different mounts.
426  */
427 void
428 vfs_getnewfsid(struct mount *mp)
429 {
430 	static u_int16_t mntid_base;
431 	fsid_t tfsid;
432 	int mtype;
433 
434 	mtx_lock(&mntid_mtx);
435 	mtype = mp->mnt_vfc->vfc_typenum;
436 	tfsid.val[1] = mtype;
437 	mtype = (mtype & 0xFF) << 24;
438 	for (;;) {
439 		tfsid.val[0] = makedev(255,
440 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
441 		mntid_base++;
442 		if (vfs_getvfs(&tfsid) == NULL)
443 			break;
444 	}
445 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
446 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
447 	mtx_unlock(&mntid_mtx);
448 }
449 
450 /*
451  * Knob to control the precision of file timestamps:
452  *
453  *   0 = seconds only; nanoseconds zeroed.
454  *   1 = seconds and nanoseconds, accurate within 1/HZ.
455  *   2 = seconds and nanoseconds, truncated to microseconds.
456  * >=3 = seconds and nanoseconds, maximum precision.
457  */
458 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
459 
460 static int timestamp_precision = TSP_SEC;
461 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
462     &timestamp_precision, 0, "");
463 
464 /*
465  * Get a current timestamp.
466  */
467 void
468 vfs_timestamp(struct timespec *tsp)
469 {
470 	struct timeval tv;
471 
472 	switch (timestamp_precision) {
473 	case TSP_SEC:
474 		tsp->tv_sec = time_second;
475 		tsp->tv_nsec = 0;
476 		break;
477 	case TSP_HZ:
478 		getnanotime(tsp);
479 		break;
480 	case TSP_USEC:
481 		microtime(&tv);
482 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
483 		break;
484 	case TSP_NSEC:
485 	default:
486 		nanotime(tsp);
487 		break;
488 	}
489 }
490 
491 /*
492  * Set vnode attributes to VNOVAL
493  */
494 void
495 vattr_null(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 != 0) {
564 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
565 		while (vp != NULL && vp->v_type == VMARKER)
566 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
567 		if (vp == NULL)
568 			break;
569 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
570 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
571 		--count;
572 		if (!VI_TRYLOCK(vp))
573 			goto next_iter;
574 		/*
575 		 * If it's been deconstructed already, it's still
576 		 * referenced, or it exceeds the trigger, skip it.
577 		 */
578 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
579 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
580 		    vp->v_object->resident_page_count > trigger)) {
581 			VI_UNLOCK(vp);
582 			goto next_iter;
583 		}
584 		MNT_IUNLOCK(mp);
585 		vholdl(vp);
586 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
587 			vdrop(vp);
588 			goto next_iter_mntunlocked;
589 		}
590 		VI_LOCK(vp);
591 		/*
592 		 * v_usecount may have been bumped after VOP_LOCK() dropped
593 		 * the vnode interlock and before it was locked again.
594 		 *
595 		 * It is not necessary to recheck VI_DOOMED because it can
596 		 * only be set by another thread that holds both the vnode
597 		 * lock and vnode interlock.  If another thread has the
598 		 * vnode lock before we get to VOP_LOCK() and obtains the
599 		 * vnode interlock after VOP_LOCK() drops the vnode
600 		 * interlock, the other thread will be unable to drop the
601 		 * vnode lock before our VOP_LOCK() call fails.
602 		 */
603 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
604 		    (vp->v_object != NULL &&
605 		    vp->v_object->resident_page_count > trigger)) {
606 			VOP_UNLOCK(vp, LK_INTERLOCK, td);
607 			goto next_iter_mntunlocked;
608 		}
609 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
610 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
611 		vgonel(vp);
612 		VOP_UNLOCK(vp, 0, td);
613 		vdropl(vp);
614 		done++;
615 next_iter_mntunlocked:
616 		if ((count % 256) != 0)
617 			goto relock_mnt;
618 		goto yield;
619 next_iter:
620 		if ((count % 256) != 0)
621 			continue;
622 		MNT_IUNLOCK(mp);
623 yield:
624 		uio_yield();
625 relock_mnt:
626 		MNT_ILOCK(mp);
627 	}
628 	MNT_IUNLOCK(mp);
629 	vn_finished_write(mp);
630 	return done;
631 }
632 
633 /*
634  * Attempt to keep the free list at wantfreevnodes length.
635  */
636 static void
637 vnlru_free(int count)
638 {
639 	struct vnode *vp;
640 
641 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
642 	for (; count > 0; count--) {
643 		vp = TAILQ_FIRST(&vnode_free_list);
644 		/*
645 		 * The list can be modified while the free_list_mtx
646 		 * has been dropped and vp could be NULL here.
647 		 */
648 		if (!vp)
649 			break;
650 		VNASSERT(vp->v_op != NULL, vp,
651 		    ("vnlru_free: vnode already reclaimed."));
652 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
653 		/*
654 		 * Don't recycle if we can't get the interlock.
655 		 */
656 		if (!VI_TRYLOCK(vp)) {
657 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
658 			continue;
659 		}
660 		VNASSERT(VCANRECYCLE(vp), vp,
661 		    ("vp inconsistent on freelist"));
662 		freevnodes--;
663 		vp->v_iflag &= ~VI_FREE;
664 		vholdl(vp);
665 		mtx_unlock(&vnode_free_list_mtx);
666 		VI_UNLOCK(vp);
667 		vtryrecycle(vp);
668 		/*
669 		 * If the recycled succeeded this vdrop will actually free
670 		 * the vnode.  If not it will simply place it back on
671 		 * the free list.
672 		 */
673 		vdrop(vp);
674 		mtx_lock(&vnode_free_list_mtx);
675 	}
676 }
677 /*
678  * Attempt to recycle vnodes in a context that is always safe to block.
679  * Calling vlrurecycle() from the bowels of filesystem code has some
680  * interesting deadlock problems.
681  */
682 static struct proc *vnlruproc;
683 static int vnlruproc_sig;
684 
685 static void
686 vnlru_proc(void)
687 {
688 	struct mount *mp, *nmp;
689 	int done;
690 	struct proc *p = vnlruproc;
691 	struct thread *td = FIRST_THREAD_IN_PROC(p);
692 
693 	mtx_lock(&Giant);
694 
695 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
696 	    SHUTDOWN_PRI_FIRST);
697 
698 	for (;;) {
699 		kthread_suspend_check(p);
700 		mtx_lock(&vnode_free_list_mtx);
701 		if (freevnodes > wantfreevnodes)
702 			vnlru_free(freevnodes - wantfreevnodes);
703 		if (numvnodes <= desiredvnodes * 9 / 10) {
704 			vnlruproc_sig = 0;
705 			wakeup(&vnlruproc_sig);
706 			msleep(vnlruproc, &vnode_free_list_mtx,
707 			    PVFS|PDROP, "vlruwt", hz);
708 			continue;
709 		}
710 		mtx_unlock(&vnode_free_list_mtx);
711 		done = 0;
712 		mtx_lock(&mountlist_mtx);
713 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
714 			int vfsunlocked;
715 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
716 				nmp = TAILQ_NEXT(mp, mnt_list);
717 				continue;
718 			}
719 			if (!VFS_NEEDSGIANT(mp)) {
720 				mtx_unlock(&Giant);
721 				vfsunlocked = 1;
722 			} else
723 				vfsunlocked = 0;
724 			done += vlrureclaim(mp);
725 			if (vfsunlocked)
726 				mtx_lock(&Giant);
727 			mtx_lock(&mountlist_mtx);
728 			nmp = TAILQ_NEXT(mp, mnt_list);
729 			vfs_unbusy(mp, td);
730 		}
731 		mtx_unlock(&mountlist_mtx);
732 		if (done == 0) {
733 #if 0
734 			/* These messages are temporary debugging aids */
735 			if (vnlru_nowhere < 5)
736 				printf("vnlru process getting nowhere..\n");
737 			else if (vnlru_nowhere == 5)
738 				printf("vnlru process messages stopped.\n");
739 #endif
740 			vnlru_nowhere++;
741 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
742 		} else
743 			uio_yield();
744 	}
745 }
746 
747 static struct kproc_desc vnlru_kp = {
748 	"vnlru",
749 	vnlru_proc,
750 	&vnlruproc
751 };
752 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
753 
754 /*
755  * Routines having to do with the management of the vnode table.
756  */
757 
758 static void
759 vdestroy(struct vnode *vp)
760 {
761 	struct bufobj *bo;
762 
763 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
764 	mtx_lock(&vnode_free_list_mtx);
765 	numvnodes--;
766 	mtx_unlock(&vnode_free_list_mtx);
767 	bo = &vp->v_bufobj;
768 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
769 	    ("cleaned vnode still on the free list."));
770 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
771 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
772 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
773 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
774 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
775 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
776 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
777 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
778 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
779 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
780 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
781 	VI_UNLOCK(vp);
782 #ifdef MAC
783 	mac_destroy_vnode(vp);
784 #endif
785 	if (vp->v_pollinfo != NULL) {
786 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
787 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
788 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
789 	}
790 #ifdef INVARIANTS
791 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
792 	vp->v_op = NULL;
793 #endif
794 	lockdestroy(vp->v_vnlock);
795 	mtx_destroy(&vp->v_interlock);
796 	uma_zfree(vnode_zone, vp);
797 }
798 
799 /*
800  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
801  * before we actually vgone().  This function must be called with the vnode
802  * held to prevent the vnode from being returned to the free list midway
803  * through vgone().
804  */
805 static int
806 vtryrecycle(struct vnode *vp)
807 {
808 	struct thread *td = curthread;
809 	struct mount *vnmp;
810 
811 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
812 	VNASSERT(vp->v_holdcnt, vp,
813 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
814 	/*
815 	 * This vnode may found and locked via some other list, if so we
816 	 * can't recycle it yet.
817 	 */
818 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
819 		return (EWOULDBLOCK);
820 	/*
821 	 * Don't recycle if its filesystem is being suspended.
822 	 */
823 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
824 		VOP_UNLOCK(vp, 0, td);
825 		return (EBUSY);
826 	}
827 	/*
828 	 * If we got this far, we need to acquire the interlock and see if
829 	 * anyone picked up this vnode from another list.  If not, we will
830 	 * mark it with DOOMED via vgonel() so that anyone who does find it
831 	 * will skip over it.
832 	 */
833 	VI_LOCK(vp);
834 	if (vp->v_usecount) {
835 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
836 		vn_finished_write(vnmp);
837 		return (EBUSY);
838 	}
839 	if ((vp->v_iflag & VI_DOOMED) == 0)
840 		vgonel(vp);
841 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
842 	vn_finished_write(vnmp);
843 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
844 	return (0);
845 }
846 
847 /*
848  * Return the next vnode from the free list.
849  */
850 int
851 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
852     struct vnode **vpp)
853 {
854 	struct vnode *vp = NULL;
855 	struct bufobj *bo;
856 
857 	mtx_lock(&vnode_free_list_mtx);
858 	/*
859 	 * Lend our context to reclaim vnodes if they've exceeded the max.
860 	 */
861 	if (freevnodes > wantfreevnodes)
862 		vnlru_free(1);
863 	/*
864 	 * Wait for available vnodes.
865 	 */
866 	if (numvnodes > desiredvnodes) {
867 		if (vnlruproc_sig == 0) {
868 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
869 			wakeup(vnlruproc);
870 		}
871 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
872 		    "vlruwk", hz);
873 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
874 		if (numvnodes > desiredvnodes) {
875 			mtx_unlock(&vnode_free_list_mtx);
876 			return (ENFILE);
877 		}
878 #endif
879 	}
880 	numvnodes++;
881 	mtx_unlock(&vnode_free_list_mtx);
882 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
883 	/*
884 	 * Setup locks.
885 	 */
886 	vp->v_vnlock = &vp->v_lock;
887 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
888 	/*
889 	 * By default, don't allow shared locks unless filesystems
890 	 * opt-in.
891 	 */
892 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
893 	/*
894 	 * Initialize bufobj.
895 	 */
896 	bo = &vp->v_bufobj;
897 	bo->__bo_vnode = vp;
898 	bo->bo_mtx = &vp->v_interlock;
899 	bo->bo_ops = &buf_ops_bio;
900 	bo->bo_private = vp;
901 	TAILQ_INIT(&bo->bo_clean.bv_hd);
902 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
903 	/*
904 	 * Initialize namecache.
905 	 */
906 	LIST_INIT(&vp->v_cache_src);
907 	TAILQ_INIT(&vp->v_cache_dst);
908 	/*
909 	 * Finalize various vnode identity bits.
910 	 */
911 	vp->v_type = VNON;
912 	vp->v_tag = tag;
913 	vp->v_op = vops;
914 	v_incr_usecount(vp);
915 	vp->v_data = 0;
916 #ifdef MAC
917 	mac_init_vnode(vp);
918 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
919 		mac_associate_vnode_singlelabel(mp, vp);
920 	else if (mp == NULL)
921 		printf("NULL mp in getnewvnode()\n");
922 #endif
923 	if (mp != NULL) {
924 		insmntque(vp, mp);
925 		bo->bo_bsize = mp->mnt_stat.f_iosize;
926 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
927 			vp->v_vflag |= VV_NOKNOTE;
928 	}
929 
930 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
931 	*vpp = vp;
932 	return (0);
933 }
934 
935 /*
936  * Delete from old mount point vnode list, if on one.
937  */
938 static void
939 delmntque(struct vnode *vp)
940 {
941 	struct mount *mp;
942 
943 	mp = vp->v_mount;
944 	if (mp == NULL)
945 		return;
946 	MNT_ILOCK(mp);
947 	vp->v_mount = NULL;
948 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
949 		("bad mount point vnode list size"));
950 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
951 	mp->mnt_nvnodelistsize--;
952 	MNT_REL(mp);
953 	MNT_IUNLOCK(mp);
954 }
955 
956 /*
957  * Insert into list of vnodes for the new mount point, if available.
958  */
959 static void
960 insmntque(struct vnode *vp, struct mount *mp)
961 {
962 
963 	vp->v_mount = mp;
964 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
965 	MNT_ILOCK(mp);
966 	MNT_REF(mp);
967 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
968 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
969 		("neg mount point vnode list size"));
970 	mp->mnt_nvnodelistsize++;
971 	MNT_IUNLOCK(mp);
972 }
973 
974 /*
975  * Flush out and invalidate all buffers associated with a bufobj
976  * Called with the underlying object locked.
977  */
978 int
979 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag,
980     int slptimeo)
981 {
982 	int error;
983 
984 	BO_LOCK(bo);
985 	if (flags & V_SAVE) {
986 		error = bufobj_wwait(bo, slpflag, slptimeo);
987 		if (error) {
988 			BO_UNLOCK(bo);
989 			return (error);
990 		}
991 		if (bo->bo_dirty.bv_cnt > 0) {
992 			BO_UNLOCK(bo);
993 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
994 				return (error);
995 			/*
996 			 * XXX We could save a lock/unlock if this was only
997 			 * enabled under INVARIANTS
998 			 */
999 			BO_LOCK(bo);
1000 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1001 				panic("vinvalbuf: dirty bufs");
1002 		}
1003 	}
1004 	/*
1005 	 * If you alter this loop please notice that interlock is dropped and
1006 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1007 	 * no race conditions occur from this.
1008 	 */
1009 	do {
1010 		error = flushbuflist(&bo->bo_clean,
1011 		    flags, bo, slpflag, slptimeo);
1012 		if (error == 0)
1013 			error = flushbuflist(&bo->bo_dirty,
1014 			    flags, bo, slpflag, slptimeo);
1015 		if (error != 0 && error != EAGAIN) {
1016 			BO_UNLOCK(bo);
1017 			return (error);
1018 		}
1019 	} while (error != 0);
1020 
1021 	/*
1022 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1023 	 * have write I/O in-progress but if there is a VM object then the
1024 	 * VM object can also have read-I/O in-progress.
1025 	 */
1026 	do {
1027 		bufobj_wwait(bo, 0, 0);
1028 		BO_UNLOCK(bo);
1029 		if (bo->bo_object != NULL) {
1030 			VM_OBJECT_LOCK(bo->bo_object);
1031 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1032 			VM_OBJECT_UNLOCK(bo->bo_object);
1033 		}
1034 		BO_LOCK(bo);
1035 	} while (bo->bo_numoutput > 0);
1036 	BO_UNLOCK(bo);
1037 
1038 	/*
1039 	 * Destroy the copy in the VM cache, too.
1040 	 */
1041 	if (bo->bo_object != NULL) {
1042 		VM_OBJECT_LOCK(bo->bo_object);
1043 		vm_object_page_remove(bo->bo_object, 0, 0,
1044 			(flags & V_SAVE) ? TRUE : FALSE);
1045 		VM_OBJECT_UNLOCK(bo->bo_object);
1046 	}
1047 
1048 #ifdef INVARIANTS
1049 	BO_LOCK(bo);
1050 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1051 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1052 		panic("vinvalbuf: flush failed");
1053 	BO_UNLOCK(bo);
1054 #endif
1055 	return (0);
1056 }
1057 
1058 /*
1059  * Flush out and invalidate all buffers associated with a vnode.
1060  * Called with the underlying object locked.
1061  */
1062 int
1063 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag,
1064     int slptimeo)
1065 {
1066 
1067 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1068 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1069 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1070 }
1071 
1072 /*
1073  * Flush out buffers on the specified list.
1074  *
1075  */
1076 static int
1077 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1078     int slptimeo)
1079 {
1080 	struct buf *bp, *nbp;
1081 	int retval, error;
1082 	daddr_t lblkno;
1083 	b_xflags_t xflags;
1084 
1085 	ASSERT_BO_LOCKED(bo);
1086 
1087 	retval = 0;
1088 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1089 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1090 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1091 			continue;
1092 		}
1093 		lblkno = 0;
1094 		xflags = 0;
1095 		if (nbp != NULL) {
1096 			lblkno = nbp->b_lblkno;
1097 			xflags = nbp->b_xflags &
1098 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1099 		}
1100 		retval = EAGAIN;
1101 		error = BUF_TIMELOCK(bp,
1102 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1103 		    "flushbuf", slpflag, slptimeo);
1104 		if (error) {
1105 			BO_LOCK(bo);
1106 			return (error != ENOLCK ? error : EAGAIN);
1107 		}
1108 		KASSERT(bp->b_bufobj == bo,
1109 	            ("bp %p wrong b_bufobj %p should be %p",
1110 		    bp, bp->b_bufobj, bo));
1111 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1112 			BUF_UNLOCK(bp);
1113 			BO_LOCK(bo);
1114 			return (EAGAIN);
1115 		}
1116 		/*
1117 		 * XXX Since there are no node locks for NFS, I
1118 		 * believe there is a slight chance that a delayed
1119 		 * write will occur while sleeping just above, so
1120 		 * check for it.
1121 		 */
1122 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1123 		    (flags & V_SAVE)) {
1124 			bremfree(bp);
1125 			bp->b_flags |= B_ASYNC;
1126 			bwrite(bp);
1127 			BO_LOCK(bo);
1128 			return (EAGAIN);	/* XXX: why not loop ? */
1129 		}
1130 		bremfree(bp);
1131 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1132 		bp->b_flags &= ~B_ASYNC;
1133 		brelse(bp);
1134 		BO_LOCK(bo);
1135 		if (nbp != NULL &&
1136 		    (nbp->b_bufobj != bo ||
1137 		     nbp->b_lblkno != lblkno ||
1138 		     (nbp->b_xflags &
1139 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1140 			break;			/* nbp invalid */
1141 	}
1142 	return (retval);
1143 }
1144 
1145 /*
1146  * Truncate a file's buffer and pages to a specified length.  This
1147  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1148  * sync activity.
1149  */
1150 int
1151 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1152     off_t length, int blksize)
1153 {
1154 	struct buf *bp, *nbp;
1155 	int anyfreed;
1156 	int trunclbn;
1157 	struct bufobj *bo;
1158 
1159 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1160 	/*
1161 	 * Round up to the *next* lbn.
1162 	 */
1163 	trunclbn = (length + blksize - 1) / blksize;
1164 
1165 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1166 restart:
1167 	VI_LOCK(vp);
1168 	bo = &vp->v_bufobj;
1169 	anyfreed = 1;
1170 	for (;anyfreed;) {
1171 		anyfreed = 0;
1172 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1173 			if (bp->b_lblkno < trunclbn)
1174 				continue;
1175 			if (BUF_LOCK(bp,
1176 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1177 			    VI_MTX(vp)) == ENOLCK)
1178 				goto restart;
1179 
1180 			bremfree(bp);
1181 			bp->b_flags |= (B_INVAL | B_RELBUF);
1182 			bp->b_flags &= ~B_ASYNC;
1183 			brelse(bp);
1184 			anyfreed = 1;
1185 
1186 			if (nbp != NULL &&
1187 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1188 			    (nbp->b_vp != vp) ||
1189 			    (nbp->b_flags & B_DELWRI))) {
1190 				goto restart;
1191 			}
1192 			VI_LOCK(vp);
1193 		}
1194 
1195 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1196 			if (bp->b_lblkno < trunclbn)
1197 				continue;
1198 			if (BUF_LOCK(bp,
1199 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1200 			    VI_MTX(vp)) == ENOLCK)
1201 				goto restart;
1202 			bremfree(bp);
1203 			bp->b_flags |= (B_INVAL | B_RELBUF);
1204 			bp->b_flags &= ~B_ASYNC;
1205 			brelse(bp);
1206 			anyfreed = 1;
1207 			if (nbp != NULL &&
1208 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1209 			    (nbp->b_vp != vp) ||
1210 			    (nbp->b_flags & B_DELWRI) == 0)) {
1211 				goto restart;
1212 			}
1213 			VI_LOCK(vp);
1214 		}
1215 	}
1216 
1217 	if (length > 0) {
1218 restartsync:
1219 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1220 			if (bp->b_lblkno > 0)
1221 				continue;
1222 			/*
1223 			 * Since we hold the vnode lock this should only
1224 			 * fail if we're racing with the buf daemon.
1225 			 */
1226 			if (BUF_LOCK(bp,
1227 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1228 			    VI_MTX(vp)) == ENOLCK) {
1229 				goto restart;
1230 			}
1231 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1232 			    ("buf(%p) on dirty queue without DELWRI", bp));
1233 
1234 			bremfree(bp);
1235 			bawrite(bp);
1236 			VI_LOCK(vp);
1237 			goto restartsync;
1238 		}
1239 	}
1240 
1241 	bufobj_wwait(bo, 0, 0);
1242 	VI_UNLOCK(vp);
1243 	vnode_pager_setsize(vp, length);
1244 
1245 	return (0);
1246 }
1247 
1248 /*
1249  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1250  * 		 a vnode.
1251  *
1252  *	NOTE: We have to deal with the special case of a background bitmap
1253  *	buffer, a situation where two buffers will have the same logical
1254  *	block offset.  We want (1) only the foreground buffer to be accessed
1255  *	in a lookup and (2) must differentiate between the foreground and
1256  *	background buffer in the splay tree algorithm because the splay
1257  *	tree cannot normally handle multiple entities with the same 'index'.
1258  *	We accomplish this by adding differentiating flags to the splay tree's
1259  *	numerical domain.
1260  */
1261 static
1262 struct buf *
1263 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1264 {
1265 	struct buf dummy;
1266 	struct buf *lefttreemax, *righttreemin, *y;
1267 
1268 	if (root == NULL)
1269 		return (NULL);
1270 	lefttreemax = righttreemin = &dummy;
1271 	for (;;) {
1272 		if (lblkno < root->b_lblkno ||
1273 		    (lblkno == root->b_lblkno &&
1274 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1275 			if ((y = root->b_left) == NULL)
1276 				break;
1277 			if (lblkno < y->b_lblkno) {
1278 				/* Rotate right. */
1279 				root->b_left = y->b_right;
1280 				y->b_right = root;
1281 				root = y;
1282 				if ((y = root->b_left) == NULL)
1283 					break;
1284 			}
1285 			/* Link into the new root's right tree. */
1286 			righttreemin->b_left = root;
1287 			righttreemin = root;
1288 		} else if (lblkno > root->b_lblkno ||
1289 		    (lblkno == root->b_lblkno &&
1290 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1291 			if ((y = root->b_right) == NULL)
1292 				break;
1293 			if (lblkno > y->b_lblkno) {
1294 				/* Rotate left. */
1295 				root->b_right = y->b_left;
1296 				y->b_left = root;
1297 				root = y;
1298 				if ((y = root->b_right) == NULL)
1299 					break;
1300 			}
1301 			/* Link into the new root's left tree. */
1302 			lefttreemax->b_right = root;
1303 			lefttreemax = root;
1304 		} else {
1305 			break;
1306 		}
1307 		root = y;
1308 	}
1309 	/* Assemble the new root. */
1310 	lefttreemax->b_right = root->b_left;
1311 	righttreemin->b_left = root->b_right;
1312 	root->b_left = dummy.b_right;
1313 	root->b_right = dummy.b_left;
1314 	return (root);
1315 }
1316 
1317 static void
1318 buf_vlist_remove(struct buf *bp)
1319 {
1320 	struct buf *root;
1321 	struct bufv *bv;
1322 
1323 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1324 	ASSERT_BO_LOCKED(bp->b_bufobj);
1325 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1326 	    (BX_VNDIRTY|BX_VNCLEAN),
1327 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1328 	if (bp->b_xflags & BX_VNDIRTY)
1329 		bv = &bp->b_bufobj->bo_dirty;
1330 	else
1331 		bv = &bp->b_bufobj->bo_clean;
1332 	if (bp != bv->bv_root) {
1333 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1334 		KASSERT(root == bp, ("splay lookup failed in remove"));
1335 	}
1336 	if (bp->b_left == NULL) {
1337 		root = bp->b_right;
1338 	} else {
1339 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1340 		root->b_right = bp->b_right;
1341 	}
1342 	bv->bv_root = root;
1343 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1344 	bv->bv_cnt--;
1345 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1346 }
1347 
1348 /*
1349  * Add the buffer to the sorted clean or dirty block list using a
1350  * splay tree algorithm.
1351  *
1352  * NOTE: xflags is passed as a constant, optimizing this inline function!
1353  */
1354 static void
1355 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1356 {
1357 	struct buf *root;
1358 	struct bufv *bv;
1359 
1360 	ASSERT_BO_LOCKED(bo);
1361 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1362 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1363 	bp->b_xflags |= xflags;
1364 	if (xflags & BX_VNDIRTY)
1365 		bv = &bo->bo_dirty;
1366 	else
1367 		bv = &bo->bo_clean;
1368 
1369 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1370 	if (root == NULL) {
1371 		bp->b_left = NULL;
1372 		bp->b_right = NULL;
1373 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1374 	} else if (bp->b_lblkno < root->b_lblkno ||
1375 	    (bp->b_lblkno == root->b_lblkno &&
1376 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1377 		bp->b_left = root->b_left;
1378 		bp->b_right = root;
1379 		root->b_left = NULL;
1380 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1381 	} else {
1382 		bp->b_right = root->b_right;
1383 		bp->b_left = root;
1384 		root->b_right = NULL;
1385 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1386 	}
1387 	bv->bv_cnt++;
1388 	bv->bv_root = bp;
1389 }
1390 
1391 /*
1392  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1393  * shadow buffers used in background bitmap writes.
1394  *
1395  * This code isn't quite efficient as it could be because we are maintaining
1396  * two sorted lists and do not know which list the block resides in.
1397  *
1398  * During a "make buildworld" the desired buffer is found at one of
1399  * the roots more than 60% of the time.  Thus, checking both roots
1400  * before performing either splay eliminates unnecessary splays on the
1401  * first tree splayed.
1402  */
1403 struct buf *
1404 gbincore(struct bufobj *bo, daddr_t lblkno)
1405 {
1406 	struct buf *bp;
1407 
1408 	ASSERT_BO_LOCKED(bo);
1409 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1410 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1411 		return (bp);
1412 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1413 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1414 		return (bp);
1415 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1416 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1417 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1418 			return (bp);
1419 	}
1420 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1421 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1422 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1423 			return (bp);
1424 	}
1425 	return (NULL);
1426 }
1427 
1428 /*
1429  * Associate a buffer with a vnode.
1430  */
1431 void
1432 bgetvp(struct vnode *vp, struct buf *bp)
1433 {
1434 
1435 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1436 
1437 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1438 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1439 	    ("bgetvp: bp already attached! %p", bp));
1440 
1441 	ASSERT_VI_LOCKED(vp, "bgetvp");
1442 	vholdl(vp);
1443 	bp->b_vp = vp;
1444 	bp->b_bufobj = &vp->v_bufobj;
1445 	/*
1446 	 * Insert onto list for new vnode.
1447 	 */
1448 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1449 }
1450 
1451 /*
1452  * Disassociate a buffer from a vnode.
1453  */
1454 void
1455 brelvp(struct buf *bp)
1456 {
1457 	struct bufobj *bo;
1458 	struct vnode *vp;
1459 
1460 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1461 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1462 
1463 	/*
1464 	 * Delete from old vnode list, if on one.
1465 	 */
1466 	vp = bp->b_vp;		/* XXX */
1467 	bo = bp->b_bufobj;
1468 	BO_LOCK(bo);
1469 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1470 		buf_vlist_remove(bp);
1471 	else
1472 		panic("brelvp: Buffer %p not on queue.", bp);
1473 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1474 		bo->bo_flag &= ~BO_ONWORKLST;
1475 		mtx_lock(&sync_mtx);
1476 		LIST_REMOVE(bo, bo_synclist);
1477  		syncer_worklist_len--;
1478 		mtx_unlock(&sync_mtx);
1479 	}
1480 	bp->b_vp = NULL;
1481 	bp->b_bufobj = NULL;
1482 	vdropl(vp);
1483 }
1484 
1485 /*
1486  * Add an item to the syncer work queue.
1487  */
1488 static void
1489 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1490 {
1491 	int slot;
1492 
1493 	ASSERT_BO_LOCKED(bo);
1494 
1495 	mtx_lock(&sync_mtx);
1496 	if (bo->bo_flag & BO_ONWORKLST)
1497 		LIST_REMOVE(bo, bo_synclist);
1498 	else {
1499 		bo->bo_flag |= BO_ONWORKLST;
1500  		syncer_worklist_len++;
1501 	}
1502 
1503 	if (delay > syncer_maxdelay - 2)
1504 		delay = syncer_maxdelay - 2;
1505 	slot = (syncer_delayno + delay) & syncer_mask;
1506 
1507 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1508 	mtx_unlock(&sync_mtx);
1509 }
1510 
1511 static int
1512 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1513 {
1514 	int error, len;
1515 
1516 	mtx_lock(&sync_mtx);
1517 	len = syncer_worklist_len - sync_vnode_count;
1518 	mtx_unlock(&sync_mtx);
1519 	error = SYSCTL_OUT(req, &len, sizeof(len));
1520 	return (error);
1521 }
1522 
1523 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1524     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1525 
1526 static struct proc *updateproc;
1527 static void sched_sync(void);
1528 static struct kproc_desc up_kp = {
1529 	"syncer",
1530 	sched_sync,
1531 	&updateproc
1532 };
1533 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1534 
1535 static int
1536 sync_vnode(struct bufobj *bo, struct thread *td)
1537 {
1538 	struct vnode *vp;
1539 	struct mount *mp;
1540 
1541 	vp = bo->__bo_vnode; 	/* XXX */
1542 	if (VOP_ISLOCKED(vp, NULL) != 0)
1543 		return (1);
1544 	if (VI_TRYLOCK(vp) == 0)
1545 		return (1);
1546 	/*
1547 	 * We use vhold in case the vnode does not
1548 	 * successfully sync.  vhold prevents the vnode from
1549 	 * going away when we unlock the sync_mtx so that
1550 	 * we can acquire the vnode interlock.
1551 	 */
1552 	vholdl(vp);
1553 	mtx_unlock(&sync_mtx);
1554 	VI_UNLOCK(vp);
1555 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1556 		vdrop(vp);
1557 		mtx_lock(&sync_mtx);
1558 		return (1);
1559 	}
1560 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1561 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1562 	VOP_UNLOCK(vp, 0, td);
1563 	vn_finished_write(mp);
1564 	VI_LOCK(vp);
1565 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1566 		/*
1567 		 * Put us back on the worklist.  The worklist
1568 		 * routine will remove us from our current
1569 		 * position and then add us back in at a later
1570 		 * position.
1571 		 */
1572 		vn_syncer_add_to_worklist(bo, syncdelay);
1573 	}
1574 	vdropl(vp);
1575 	mtx_lock(&sync_mtx);
1576 	return (0);
1577 }
1578 
1579 /*
1580  * System filesystem synchronizer daemon.
1581  */
1582 static void
1583 sched_sync(void)
1584 {
1585 	struct synclist *next;
1586 	struct synclist *slp;
1587 	struct bufobj *bo;
1588 	long starttime;
1589 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1590 	static int dummychan;
1591 	int last_work_seen;
1592 	int net_worklist_len;
1593 	int syncer_final_iter;
1594 	int first_printf;
1595 	int error;
1596 
1597 	mtx_lock(&Giant);
1598 	last_work_seen = 0;
1599 	syncer_final_iter = 0;
1600 	first_printf = 1;
1601 	syncer_state = SYNCER_RUNNING;
1602 	starttime = time_uptime;
1603 	td->td_pflags |= TDP_NORUNNINGBUF;
1604 
1605 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1606 	    SHUTDOWN_PRI_LAST);
1607 
1608 	for (;;) {
1609 		mtx_lock(&sync_mtx);
1610 		if (syncer_state == SYNCER_FINAL_DELAY &&
1611 		    syncer_final_iter == 0) {
1612 			mtx_unlock(&sync_mtx);
1613 			kthread_suspend_check(td->td_proc);
1614 			mtx_lock(&sync_mtx);
1615 		}
1616 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1617 		if (syncer_state != SYNCER_RUNNING &&
1618 		    starttime != time_uptime) {
1619 			if (first_printf) {
1620 				printf("\nSyncing disks, vnodes remaining...");
1621 				first_printf = 0;
1622 			}
1623 			printf("%d ", net_worklist_len);
1624 		}
1625 		starttime = time_uptime;
1626 
1627 		/*
1628 		 * Push files whose dirty time has expired.  Be careful
1629 		 * of interrupt race on slp queue.
1630 		 *
1631 		 * Skip over empty worklist slots when shutting down.
1632 		 */
1633 		do {
1634 			slp = &syncer_workitem_pending[syncer_delayno];
1635 			syncer_delayno += 1;
1636 			if (syncer_delayno == syncer_maxdelay)
1637 				syncer_delayno = 0;
1638 			next = &syncer_workitem_pending[syncer_delayno];
1639 			/*
1640 			 * If the worklist has wrapped since the
1641 			 * it was emptied of all but syncer vnodes,
1642 			 * switch to the FINAL_DELAY state and run
1643 			 * for one more second.
1644 			 */
1645 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1646 			    net_worklist_len == 0 &&
1647 			    last_work_seen == syncer_delayno) {
1648 				syncer_state = SYNCER_FINAL_DELAY;
1649 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1650 			}
1651 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1652 		    syncer_worklist_len > 0);
1653 
1654 		/*
1655 		 * Keep track of the last time there was anything
1656 		 * on the worklist other than syncer vnodes.
1657 		 * Return to the SHUTTING_DOWN state if any
1658 		 * new work appears.
1659 		 */
1660 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1661 			last_work_seen = syncer_delayno;
1662 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1663 			syncer_state = SYNCER_SHUTTING_DOWN;
1664 		while ((bo = LIST_FIRST(slp)) != NULL) {
1665 			error = sync_vnode(bo, td);
1666 			if (error == 1) {
1667 				LIST_REMOVE(bo, bo_synclist);
1668 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1669 				continue;
1670 			}
1671 		}
1672 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1673 			syncer_final_iter--;
1674 		mtx_unlock(&sync_mtx);
1675 		/*
1676 		 * The variable rushjob allows the kernel to speed up the
1677 		 * processing of the filesystem syncer process. A rushjob
1678 		 * value of N tells the filesystem syncer to process the next
1679 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1680 		 * is used by the soft update code to speed up the filesystem
1681 		 * syncer process when the incore state is getting so far
1682 		 * ahead of the disk that the kernel memory pool is being
1683 		 * threatened with exhaustion.
1684 		 */
1685 		mtx_lock(&sync_mtx);
1686 		if (rushjob > 0) {
1687 			rushjob -= 1;
1688 			mtx_unlock(&sync_mtx);
1689 			continue;
1690 		}
1691 		mtx_unlock(&sync_mtx);
1692 		/*
1693 		 * Just sleep for a short period if time between
1694 		 * iterations when shutting down to allow some I/O
1695 		 * to happen.
1696 		 *
1697 		 * If it has taken us less than a second to process the
1698 		 * current work, then wait. Otherwise start right over
1699 		 * again. We can still lose time if any single round
1700 		 * takes more than two seconds, but it does not really
1701 		 * matter as we are just trying to generally pace the
1702 		 * filesystem activity.
1703 		 */
1704 		if (syncer_state != SYNCER_RUNNING)
1705 			tsleep(&dummychan, PPAUSE, "syncfnl",
1706 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1707 		else if (time_uptime == starttime)
1708 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1709 	}
1710 }
1711 
1712 /*
1713  * Request the syncer daemon to speed up its work.
1714  * We never push it to speed up more than half of its
1715  * normal turn time, otherwise it could take over the cpu.
1716  */
1717 int
1718 speedup_syncer()
1719 {
1720 	struct thread *td;
1721 	int ret = 0;
1722 
1723 	td = FIRST_THREAD_IN_PROC(updateproc);
1724 	sleepq_remove(td, &lbolt);
1725 	mtx_lock(&sync_mtx);
1726 	if (rushjob < syncdelay / 2) {
1727 		rushjob += 1;
1728 		stat_rush_requests += 1;
1729 		ret = 1;
1730 	}
1731 	mtx_unlock(&sync_mtx);
1732 	return (ret);
1733 }
1734 
1735 /*
1736  * Tell the syncer to speed up its work and run though its work
1737  * list several times, then tell it to shut down.
1738  */
1739 static void
1740 syncer_shutdown(void *arg, int howto)
1741 {
1742 	struct thread *td;
1743 
1744 	if (howto & RB_NOSYNC)
1745 		return;
1746 	td = FIRST_THREAD_IN_PROC(updateproc);
1747 	sleepq_remove(td, &lbolt);
1748 	mtx_lock(&sync_mtx);
1749 	syncer_state = SYNCER_SHUTTING_DOWN;
1750 	rushjob = 0;
1751 	mtx_unlock(&sync_mtx);
1752 	kproc_shutdown(arg, howto);
1753 }
1754 
1755 /*
1756  * Reassign a buffer from one vnode to another.
1757  * Used to assign file specific control information
1758  * (indirect blocks) to the vnode to which they belong.
1759  */
1760 void
1761 reassignbuf(struct buf *bp)
1762 {
1763 	struct vnode *vp;
1764 	struct bufobj *bo;
1765 	int delay;
1766 #ifdef INVARIANTS
1767 	struct bufv *bv;
1768 #endif
1769 
1770 	vp = bp->b_vp;
1771 	bo = bp->b_bufobj;
1772 	++reassignbufcalls;
1773 
1774 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1775 	    bp, bp->b_vp, bp->b_flags);
1776 	/*
1777 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1778 	 * is not fully linked in.
1779 	 */
1780 	if (bp->b_flags & B_PAGING)
1781 		panic("cannot reassign paging buffer");
1782 
1783 	/*
1784 	 * Delete from old vnode list, if on one.
1785 	 */
1786 	VI_LOCK(vp);
1787 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1788 		buf_vlist_remove(bp);
1789 	else
1790 		panic("reassignbuf: Buffer %p not on queue.", bp);
1791 	/*
1792 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1793 	 * of clean buffers.
1794 	 */
1795 	if (bp->b_flags & B_DELWRI) {
1796 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1797 			switch (vp->v_type) {
1798 			case VDIR:
1799 				delay = dirdelay;
1800 				break;
1801 			case VCHR:
1802 				delay = metadelay;
1803 				break;
1804 			default:
1805 				delay = filedelay;
1806 			}
1807 			vn_syncer_add_to_worklist(bo, delay);
1808 		}
1809 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1810 	} else {
1811 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1812 
1813 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1814 			mtx_lock(&sync_mtx);
1815 			LIST_REMOVE(bo, bo_synclist);
1816  			syncer_worklist_len--;
1817 			mtx_unlock(&sync_mtx);
1818 			bo->bo_flag &= ~BO_ONWORKLST;
1819 		}
1820 	}
1821 #ifdef INVARIANTS
1822 	bv = &bo->bo_clean;
1823 	bp = TAILQ_FIRST(&bv->bv_hd);
1824 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1825 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1826 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1827 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1828 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1829 	bv = &bo->bo_dirty;
1830 	bp = TAILQ_FIRST(&bv->bv_hd);
1831 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1832 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1833 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1834 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1835 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1836 #endif
1837 	VI_UNLOCK(vp);
1838 }
1839 
1840 /*
1841  * Increment the use and hold counts on the vnode, taking care to reference
1842  * the driver's usecount if this is a chardev.  The vholdl() will remove
1843  * the vnode from the free list if it is presently free.  Requires the
1844  * vnode interlock and returns with it held.
1845  */
1846 static void
1847 v_incr_usecount(struct vnode *vp)
1848 {
1849 
1850 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1851 	    vp, vp->v_holdcnt, vp->v_usecount);
1852 	vp->v_usecount++;
1853 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1854 		dev_lock();
1855 		vp->v_rdev->si_usecount++;
1856 		dev_unlock();
1857 	}
1858 	vholdl(vp);
1859 }
1860 
1861 /*
1862  * Turn a holdcnt into a use+holdcnt such that only one call to
1863  * v_decr_usecount is needed.
1864  */
1865 static void
1866 v_upgrade_usecount(struct vnode *vp)
1867 {
1868 
1869 	CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n",
1870 	    vp, vp->v_holdcnt, vp->v_usecount);
1871 	vp->v_usecount++;
1872 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1873 		dev_lock();
1874 		vp->v_rdev->si_usecount++;
1875 		dev_unlock();
1876 	}
1877 }
1878 
1879 /*
1880  * Decrement the vnode use and hold count along with the driver's usecount
1881  * if this is a chardev.  The vdropl() below releases the vnode interlock
1882  * as it may free the vnode.
1883  */
1884 static void
1885 v_decr_usecount(struct vnode *vp)
1886 {
1887 
1888 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1889 	    vp, vp->v_holdcnt, vp->v_usecount);
1890 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1891 	VNASSERT(vp->v_usecount > 0, vp,
1892 	    ("v_decr_usecount: negative usecount"));
1893 	vp->v_usecount--;
1894 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1895 		dev_lock();
1896 		vp->v_rdev->si_usecount--;
1897 		dev_unlock();
1898 	}
1899 	vdropl(vp);
1900 }
1901 
1902 /*
1903  * Decrement only the use count and driver use count.  This is intended to
1904  * be paired with a follow on vdropl() to release the remaining hold count.
1905  * In this way we may vgone() a vnode with a 0 usecount without risk of
1906  * having it end up on a free list because the hold count is kept above 0.
1907  */
1908 static void
1909 v_decr_useonly(struct vnode *vp)
1910 {
1911 
1912 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1913 	    vp, vp->v_holdcnt, vp->v_usecount);
1914 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1915 	VNASSERT(vp->v_usecount > 0, vp,
1916 	    ("v_decr_useonly: negative usecount"));
1917 	vp->v_usecount--;
1918 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1919 		dev_lock();
1920 		vp->v_rdev->si_usecount--;
1921 		dev_unlock();
1922 	}
1923 }
1924 
1925 /*
1926  * Grab a particular vnode from the free list, increment its
1927  * reference count and lock it. The vnode lock bit is set if the
1928  * vnode is being eliminated in vgone. The process is awakened
1929  * when the transition is completed, and an error returned to
1930  * indicate that the vnode is no longer usable (possibly having
1931  * been changed to a new filesystem type).
1932  */
1933 int
1934 vget(struct vnode *vp, int flags, struct thread *td)
1935 {
1936 	int oweinact;
1937 	int oldflags;
1938 	int error;
1939 
1940 	error = 0;
1941 	oldflags = flags;
1942 	oweinact = 0;
1943 	VFS_ASSERT_GIANT(vp->v_mount);
1944 	if ((flags & LK_INTERLOCK) == 0)
1945 		VI_LOCK(vp);
1946 	/*
1947 	 * If the inactive call was deferred because vput() was called
1948 	 * with a shared lock, we have to do it here before another thread
1949 	 * gets a reference to data that should be dead.
1950 	 */
1951 	if (vp->v_iflag & VI_OWEINACT) {
1952 		if (flags & LK_NOWAIT) {
1953 			VI_UNLOCK(vp);
1954 			return (EBUSY);
1955 		}
1956 		flags &= ~LK_TYPE_MASK;
1957 		flags |= LK_EXCLUSIVE;
1958 		oweinact = 1;
1959 	}
1960 	vholdl(vp);
1961 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1962 		vdrop(vp);
1963 		return (error);
1964 	}
1965 	VI_LOCK(vp);
1966 	/* Upgrade our holdcnt to a usecount. */
1967 	v_upgrade_usecount(vp);
1968 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
1969 		panic("vget: vn_lock failed to return ENOENT\n");
1970 	if (oweinact) {
1971 		if (vp->v_iflag & VI_OWEINACT)
1972 			vinactive(vp, td);
1973 		VI_UNLOCK(vp);
1974 		if ((oldflags & LK_TYPE_MASK) == 0)
1975 			VOP_UNLOCK(vp, 0, td);
1976 	} else
1977 		VI_UNLOCK(vp);
1978 	return (0);
1979 }
1980 
1981 /*
1982  * Increase the reference count of a vnode.
1983  */
1984 void
1985 vref(struct vnode *vp)
1986 {
1987 
1988 	VI_LOCK(vp);
1989 	v_incr_usecount(vp);
1990 	VI_UNLOCK(vp);
1991 }
1992 
1993 /*
1994  * Return reference count of a vnode.
1995  *
1996  * The results of this call are only guaranteed when some mechanism other
1997  * than the VI lock is used to stop other processes from gaining references
1998  * to the vnode.  This may be the case if the caller holds the only reference.
1999  * This is also useful when stale data is acceptable as race conditions may
2000  * be accounted for by some other means.
2001  */
2002 int
2003 vrefcnt(struct vnode *vp)
2004 {
2005 	int usecnt;
2006 
2007 	VI_LOCK(vp);
2008 	usecnt = vp->v_usecount;
2009 	VI_UNLOCK(vp);
2010 
2011 	return (usecnt);
2012 }
2013 
2014 
2015 /*
2016  * Vnode put/release.
2017  * If count drops to zero, call inactive routine and return to freelist.
2018  */
2019 void
2020 vrele(struct vnode *vp)
2021 {
2022 	struct thread *td = curthread;	/* XXX */
2023 
2024 	KASSERT(vp != NULL, ("vrele: null vp"));
2025 	VFS_ASSERT_GIANT(vp->v_mount);
2026 
2027 	VI_LOCK(vp);
2028 
2029 	/* Skip this v_writecount check if we're going to panic below. */
2030 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2031 	    ("vrele: missed vn_close"));
2032 
2033 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2034 	    vp->v_usecount == 1)) {
2035 		v_decr_usecount(vp);
2036 		return;
2037 	}
2038 	if (vp->v_usecount != 1) {
2039 #ifdef DIAGNOSTIC
2040 		vprint("vrele: negative ref count", vp);
2041 #endif
2042 		VI_UNLOCK(vp);
2043 		panic("vrele: negative ref cnt");
2044 	}
2045 	/*
2046 	 * We want to hold the vnode until the inactive finishes to
2047 	 * prevent vgone() races.  We drop the use count here and the
2048 	 * hold count below when we're done.
2049 	 */
2050 	v_decr_useonly(vp);
2051 	/*
2052 	 * We must call VOP_INACTIVE with the node locked. Mark
2053 	 * as VI_DOINGINACT to avoid recursion.
2054 	 */
2055 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2056 		VI_LOCK(vp);
2057 		vinactive(vp, td);
2058 		VOP_UNLOCK(vp, 0, td);
2059 	} else
2060 		VI_LOCK(vp);
2061 	vdropl(vp);
2062 }
2063 
2064 /*
2065  * Release an already locked vnode.  This give the same effects as
2066  * unlock+vrele(), but takes less time and avoids releasing and
2067  * re-aquiring the lock (as vrele() aquires the lock internally.)
2068  */
2069 void
2070 vput(struct vnode *vp)
2071 {
2072 	struct thread *td = curthread;	/* XXX */
2073 	int error;
2074 
2075 	KASSERT(vp != NULL, ("vput: null vp"));
2076 	ASSERT_VOP_LOCKED(vp, "vput");
2077 	VFS_ASSERT_GIANT(vp->v_mount);
2078 	VI_LOCK(vp);
2079 	/* Skip this v_writecount check if we're going to panic below. */
2080 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2081 	    ("vput: missed vn_close"));
2082 	error = 0;
2083 
2084 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2085 	    vp->v_usecount == 1)) {
2086 		VOP_UNLOCK(vp, 0, td);
2087 		v_decr_usecount(vp);
2088 		return;
2089 	}
2090 
2091 	if (vp->v_usecount != 1) {
2092 #ifdef DIAGNOSTIC
2093 		vprint("vput: negative ref count", vp);
2094 #endif
2095 		panic("vput: negative ref cnt");
2096 	}
2097 	/*
2098 	 * We want to hold the vnode until the inactive finishes to
2099 	 * prevent vgone() races.  We drop the use count here and the
2100 	 * hold count below when we're done.
2101 	 */
2102 	v_decr_useonly(vp);
2103 	vp->v_iflag |= VI_OWEINACT;
2104 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2105 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2106 		VI_LOCK(vp);
2107 		if (error)
2108 			goto done;
2109 	}
2110 	if (vp->v_iflag & VI_OWEINACT)
2111 		vinactive(vp, td);
2112 	VOP_UNLOCK(vp, 0, td);
2113 done:
2114 	vdropl(vp);
2115 }
2116 
2117 /*
2118  * Somebody doesn't want the vnode recycled.
2119  */
2120 void
2121 vhold(struct vnode *vp)
2122 {
2123 
2124 	VI_LOCK(vp);
2125 	vholdl(vp);
2126 	VI_UNLOCK(vp);
2127 }
2128 
2129 void
2130 vholdl(struct vnode *vp)
2131 {
2132 
2133 	vp->v_holdcnt++;
2134 	if (VSHOULDBUSY(vp))
2135 		vbusy(vp);
2136 }
2137 
2138 /*
2139  * Note that there is one less who cares about this vnode.  vdrop() is the
2140  * opposite of vhold().
2141  */
2142 void
2143 vdrop(struct vnode *vp)
2144 {
2145 
2146 	VI_LOCK(vp);
2147 	vdropl(vp);
2148 }
2149 
2150 /*
2151  * Drop the hold count of the vnode.  If this is the last reference to
2152  * the vnode we will free it if it has been vgone'd otherwise it is
2153  * placed on the free list.
2154  */
2155 static void
2156 vdropl(struct vnode *vp)
2157 {
2158 
2159 	if (vp->v_holdcnt <= 0)
2160 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2161 	vp->v_holdcnt--;
2162 	if (vp->v_holdcnt == 0) {
2163 		if (vp->v_iflag & VI_DOOMED) {
2164 			vdestroy(vp);
2165 			return;
2166 		} else
2167 			vfree(vp);
2168 	}
2169 	VI_UNLOCK(vp);
2170 }
2171 
2172 /*
2173  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2174  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2175  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2176  * failed lock upgrade.
2177  */
2178 static void
2179 vinactive(struct vnode *vp, struct thread *td)
2180 {
2181 
2182 	ASSERT_VOP_LOCKED(vp, "vinactive");
2183 	ASSERT_VI_LOCKED(vp, "vinactive");
2184 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2185 	    ("vinactive: recursed on VI_DOINGINACT"));
2186 	vp->v_iflag |= VI_DOINGINACT;
2187 	vp->v_iflag &= ~VI_OWEINACT;
2188 	VI_UNLOCK(vp);
2189 	VOP_INACTIVE(vp, td);
2190 	VI_LOCK(vp);
2191 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2192 	    ("vinactive: lost VI_DOINGINACT"));
2193 	vp->v_iflag &= ~VI_DOINGINACT;
2194 }
2195 
2196 /*
2197  * Remove any vnodes in the vnode table belonging to mount point mp.
2198  *
2199  * If FORCECLOSE is not specified, there should not be any active ones,
2200  * return error if any are found (nb: this is a user error, not a
2201  * system error). If FORCECLOSE is specified, detach any active vnodes
2202  * that are found.
2203  *
2204  * If WRITECLOSE is set, only flush out regular file vnodes open for
2205  * writing.
2206  *
2207  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2208  *
2209  * `rootrefs' specifies the base reference count for the root vnode
2210  * of this filesystem. The root vnode is considered busy if its
2211  * v_usecount exceeds this value. On a successful return, vflush(, td)
2212  * will call vrele() on the root vnode exactly rootrefs times.
2213  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2214  * be zero.
2215  */
2216 #ifdef DIAGNOSTIC
2217 static int busyprt = 0;		/* print out busy vnodes */
2218 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2219 #endif
2220 
2221 int
2222 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td)
2223 {
2224 	struct vnode *vp, *mvp, *rootvp = NULL;
2225 	struct vattr vattr;
2226 	int busy = 0, error;
2227 
2228 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2229 	if (rootrefs > 0) {
2230 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2231 		    ("vflush: bad args"));
2232 		/*
2233 		 * Get the filesystem root vnode. We can vput() it
2234 		 * immediately, since with rootrefs > 0, it won't go away.
2235 		 */
2236 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2237 			return (error);
2238 		vput(rootvp);
2239 
2240 	}
2241 	MNT_ILOCK(mp);
2242 loop:
2243 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2244 
2245 		VI_LOCK(vp);
2246 		vholdl(vp);
2247 		MNT_IUNLOCK(mp);
2248 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2249 		if (error) {
2250 			vdrop(vp);
2251 			MNT_ILOCK(mp);
2252 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2253 			goto loop;
2254 		}
2255 		/*
2256 		 * Skip over a vnodes marked VV_SYSTEM.
2257 		 */
2258 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2259 			VOP_UNLOCK(vp, 0, td);
2260 			vdrop(vp);
2261 			MNT_ILOCK(mp);
2262 			continue;
2263 		}
2264 		/*
2265 		 * If WRITECLOSE is set, flush out unlinked but still open
2266 		 * files (even if open only for reading) and regular file
2267 		 * vnodes open for writing.
2268 		 */
2269 		if (flags & WRITECLOSE) {
2270 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2271 			VI_LOCK(vp);
2272 
2273 			if ((vp->v_type == VNON ||
2274 			    (error == 0 && vattr.va_nlink > 0)) &&
2275 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2276 				VOP_UNLOCK(vp, 0, td);
2277 				vdropl(vp);
2278 				MNT_ILOCK(mp);
2279 				continue;
2280 			}
2281 		} else
2282 			VI_LOCK(vp);
2283 		/*
2284 		 * With v_usecount == 0, all we need to do is clear out the
2285 		 * vnode data structures and we are done.
2286 		 *
2287 		 * If FORCECLOSE is set, forcibly close the vnode.
2288 		 */
2289 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2290 			VNASSERT(vp->v_usecount == 0 ||
2291 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2292 			    ("device VNODE %p is FORCECLOSED", vp));
2293 			vgonel(vp);
2294 		} else {
2295 			busy++;
2296 #ifdef DIAGNOSTIC
2297 			if (busyprt)
2298 				vprint("vflush: busy vnode", vp);
2299 #endif
2300 		}
2301 		VOP_UNLOCK(vp, 0, td);
2302 		vdropl(vp);
2303 		MNT_ILOCK(mp);
2304 	}
2305 	MNT_IUNLOCK(mp);
2306 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2307 		/*
2308 		 * If just the root vnode is busy, and if its refcount
2309 		 * is equal to `rootrefs', then go ahead and kill it.
2310 		 */
2311 		VI_LOCK(rootvp);
2312 		KASSERT(busy > 0, ("vflush: not busy"));
2313 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2314 		    ("vflush: usecount %d < rootrefs %d",
2315 		     rootvp->v_usecount, rootrefs));
2316 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2317 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2318 			vgone(rootvp);
2319 			VOP_UNLOCK(rootvp, 0, td);
2320 			busy = 0;
2321 		} else
2322 			VI_UNLOCK(rootvp);
2323 	}
2324 	if (busy)
2325 		return (EBUSY);
2326 	for (; rootrefs > 0; rootrefs--)
2327 		vrele(rootvp);
2328 	return (0);
2329 }
2330 
2331 /*
2332  * Recycle an unused vnode to the front of the free list.
2333  */
2334 int
2335 vrecycle(struct vnode *vp, struct thread *td)
2336 {
2337 	int recycled;
2338 
2339 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2340 	recycled = 0;
2341 	VI_LOCK(vp);
2342 	if (vp->v_usecount == 0) {
2343 		recycled = 1;
2344 		vgonel(vp);
2345 	}
2346 	VI_UNLOCK(vp);
2347 	return (recycled);
2348 }
2349 
2350 /*
2351  * Eliminate all activity associated with a vnode
2352  * in preparation for reuse.
2353  */
2354 void
2355 vgone(struct vnode *vp)
2356 {
2357 	VI_LOCK(vp);
2358 	vgonel(vp);
2359 	VI_UNLOCK(vp);
2360 }
2361 
2362 /*
2363  * vgone, with the vp interlock held.
2364  */
2365 void
2366 vgonel(struct vnode *vp)
2367 {
2368 	struct thread *td;
2369 	int oweinact;
2370 	int active;
2371 
2372 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2373 	ASSERT_VOP_LOCKED(vp, "vgonel");
2374 	ASSERT_VI_LOCKED(vp, "vgonel");
2375 #if 0
2376 	/* XXX Need to fix ttyvp before I enable this. */
2377 	VNASSERT(vp->v_holdcnt, vp,
2378 	    ("vgonel: vp %p has no reference.", vp));
2379 #endif
2380 	td = curthread;
2381 
2382 	/*
2383 	 * Don't vgonel if we're already doomed.
2384 	 */
2385 	if (vp->v_iflag & VI_DOOMED)
2386 		return;
2387 	vp->v_iflag |= VI_DOOMED;
2388 	/*
2389 	 * Check to see if the vnode is in use.  If so, we have to call
2390 	 * VOP_CLOSE() and VOP_INACTIVE().
2391 	 */
2392 	active = vp->v_usecount;
2393 	oweinact = (vp->v_iflag & VI_OWEINACT);
2394 	VI_UNLOCK(vp);
2395 	/*
2396 	 * Clean out any buffers associated with the vnode.
2397 	 * If the flush fails, just toss the buffers.
2398 	 */
2399 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2400 		(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2401 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2402 		vinvalbuf(vp, 0, td, 0, 0);
2403 
2404 	/*
2405 	 * If purging an active vnode, it must be closed and
2406 	 * deactivated before being reclaimed.
2407 	 */
2408 	if (active)
2409 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2410 	if (oweinact || active) {
2411 		VI_LOCK(vp);
2412 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2413 			vinactive(vp, td);
2414 		VI_UNLOCK(vp);
2415 	}
2416 	/*
2417 	 * Reclaim the vnode.
2418 	 */
2419 	if (VOP_RECLAIM(vp, td))
2420 		panic("vgone: cannot reclaim");
2421 	VNASSERT(vp->v_object == NULL, vp,
2422 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2423 	/*
2424 	 * Delete from old mount point vnode list.
2425 	 */
2426 	delmntque(vp);
2427 	cache_purge(vp);
2428 	/*
2429 	 * Done with purge, reset to the standard lock and invalidate
2430 	 * the vnode.
2431 	 */
2432 	VI_LOCK(vp);
2433 	vp->v_vnlock = &vp->v_lock;
2434 	vp->v_op = &dead_vnodeops;
2435 	vp->v_tag = "none";
2436 	vp->v_type = VBAD;
2437 }
2438 
2439 /*
2440  * Calculate the total number of references to a special device.
2441  */
2442 int
2443 vcount(struct vnode *vp)
2444 {
2445 	int count;
2446 
2447 	dev_lock();
2448 	count = vp->v_rdev->si_usecount;
2449 	dev_unlock();
2450 	return (count);
2451 }
2452 
2453 /*
2454  * Same as above, but using the struct cdev *as argument
2455  */
2456 int
2457 count_dev(struct cdev *dev)
2458 {
2459 	int count;
2460 
2461 	dev_lock();
2462 	count = dev->si_usecount;
2463 	dev_unlock();
2464 	return(count);
2465 }
2466 
2467 /*
2468  * Print out a description of a vnode.
2469  */
2470 static char *typename[] =
2471 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2472  "VMARKER"};
2473 
2474 void
2475 vn_printf(struct vnode *vp, const char *fmt, ...)
2476 {
2477 	va_list ap;
2478 	char buf[96];
2479 
2480 	va_start(ap, fmt);
2481 	vprintf(fmt, ap);
2482 	va_end(ap);
2483 	printf("%p: ", (void *)vp);
2484 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2485 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2486 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2487 	buf[0] = '\0';
2488 	buf[1] = '\0';
2489 	if (vp->v_vflag & VV_ROOT)
2490 		strcat(buf, "|VV_ROOT");
2491 	if (vp->v_vflag & VV_TEXT)
2492 		strcat(buf, "|VV_TEXT");
2493 	if (vp->v_vflag & VV_SYSTEM)
2494 		strcat(buf, "|VV_SYSTEM");
2495 	if (vp->v_iflag & VI_DOOMED)
2496 		strcat(buf, "|VI_DOOMED");
2497 	if (vp->v_iflag & VI_FREE)
2498 		strcat(buf, "|VI_FREE");
2499 	printf("    flags (%s)\n", buf + 1);
2500 	if (mtx_owned(VI_MTX(vp)))
2501 		printf(" VI_LOCKed");
2502 	if (vp->v_object != NULL)
2503 		printf("    v_object %p ref %d pages %d\n",
2504 		    vp->v_object, vp->v_object->ref_count,
2505 		    vp->v_object->resident_page_count);
2506 	printf("    ");
2507 	lockmgr_printinfo(vp->v_vnlock);
2508 	printf("\n");
2509 	if (vp->v_data != NULL)
2510 		VOP_PRINT(vp);
2511 }
2512 
2513 #ifdef DDB
2514 #include <ddb/ddb.h>
2515 /*
2516  * List all of the locked vnodes in the system.
2517  * Called when debugging the kernel.
2518  */
2519 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2520 {
2521 	struct mount *mp, *nmp;
2522 	struct vnode *vp;
2523 
2524 	/*
2525 	 * Note: because this is DDB, we can't obey the locking semantics
2526 	 * for these structures, which means we could catch an inconsistent
2527 	 * state and dereference a nasty pointer.  Not much to be done
2528 	 * about that.
2529 	 */
2530 	printf("Locked vnodes\n");
2531 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2532 		nmp = TAILQ_NEXT(mp, mnt_list);
2533 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2534 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL))
2535 				vprint("", vp);
2536 		}
2537 		nmp = TAILQ_NEXT(mp, mnt_list);
2538 	}
2539 }
2540 #endif
2541 
2542 /*
2543  * Fill in a struct xvfsconf based on a struct vfsconf.
2544  */
2545 static void
2546 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2547 {
2548 
2549 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2550 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2551 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2552 	xvfsp->vfc_flags = vfsp->vfc_flags;
2553 	/*
2554 	 * These are unused in userland, we keep them
2555 	 * to not break binary compatibility.
2556 	 */
2557 	xvfsp->vfc_vfsops = NULL;
2558 	xvfsp->vfc_next = NULL;
2559 }
2560 
2561 /*
2562  * Top level filesystem related information gathering.
2563  */
2564 static int
2565 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2566 {
2567 	struct vfsconf *vfsp;
2568 	struct xvfsconf xvfsp;
2569 	int error;
2570 
2571 	error = 0;
2572 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2573 		bzero(&xvfsp, sizeof(xvfsp));
2574 		vfsconf2x(vfsp, &xvfsp);
2575 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2576 		if (error)
2577 			break;
2578 	}
2579 	return (error);
2580 }
2581 
2582 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2583     "S,xvfsconf", "List of all configured filesystems");
2584 
2585 #ifndef BURN_BRIDGES
2586 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2587 
2588 static int
2589 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2590 {
2591 	int *name = (int *)arg1 - 1;	/* XXX */
2592 	u_int namelen = arg2 + 1;	/* XXX */
2593 	struct vfsconf *vfsp;
2594 	struct xvfsconf xvfsp;
2595 
2596 	printf("WARNING: userland calling deprecated sysctl, "
2597 	    "please rebuild world\n");
2598 
2599 #if 1 || defined(COMPAT_PRELITE2)
2600 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2601 	if (namelen == 1)
2602 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2603 #endif
2604 
2605 	switch (name[1]) {
2606 	case VFS_MAXTYPENUM:
2607 		if (namelen != 2)
2608 			return (ENOTDIR);
2609 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2610 	case VFS_CONF:
2611 		if (namelen != 3)
2612 			return (ENOTDIR);	/* overloaded */
2613 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2614 			if (vfsp->vfc_typenum == name[2])
2615 				break;
2616 		if (vfsp == NULL)
2617 			return (EOPNOTSUPP);
2618 		bzero(&xvfsp, sizeof(xvfsp));
2619 		vfsconf2x(vfsp, &xvfsp);
2620 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2621 	}
2622 	return (EOPNOTSUPP);
2623 }
2624 
2625 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2626 	vfs_sysctl, "Generic filesystem");
2627 
2628 #if 1 || defined(COMPAT_PRELITE2)
2629 
2630 static int
2631 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2632 {
2633 	int error;
2634 	struct vfsconf *vfsp;
2635 	struct ovfsconf ovfs;
2636 
2637 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2638 		bzero(&ovfs, sizeof(ovfs));
2639 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2640 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2641 		ovfs.vfc_index = vfsp->vfc_typenum;
2642 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2643 		ovfs.vfc_flags = vfsp->vfc_flags;
2644 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2645 		if (error)
2646 			return error;
2647 	}
2648 	return 0;
2649 }
2650 
2651 #endif /* 1 || COMPAT_PRELITE2 */
2652 #endif /* !BURN_BRIDGES */
2653 
2654 #define KINFO_VNODESLOP		10
2655 #ifdef notyet
2656 /*
2657  * Dump vnode list (via sysctl).
2658  */
2659 /* ARGSUSED */
2660 static int
2661 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2662 {
2663 	struct xvnode *xvn;
2664 	struct thread *td = req->td;
2665 	struct mount *mp;
2666 	struct vnode *vp;
2667 	int error, len, n;
2668 
2669 	/*
2670 	 * Stale numvnodes access is not fatal here.
2671 	 */
2672 	req->lock = 0;
2673 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2674 	if (!req->oldptr)
2675 		/* Make an estimate */
2676 		return (SYSCTL_OUT(req, 0, len));
2677 
2678 	error = sysctl_wire_old_buffer(req, 0);
2679 	if (error != 0)
2680 		return (error);
2681 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2682 	n = 0;
2683 	mtx_lock(&mountlist_mtx);
2684 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2685 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2686 			continue;
2687 		MNT_ILOCK(mp);
2688 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2689 			if (n == len)
2690 				break;
2691 			vref(vp);
2692 			xvn[n].xv_size = sizeof *xvn;
2693 			xvn[n].xv_vnode = vp;
2694 			xvn[n].xv_id = 0;	/* XXX compat */
2695 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2696 			XV_COPY(usecount);
2697 			XV_COPY(writecount);
2698 			XV_COPY(holdcnt);
2699 			XV_COPY(mount);
2700 			XV_COPY(numoutput);
2701 			XV_COPY(type);
2702 #undef XV_COPY
2703 			xvn[n].xv_flag = vp->v_vflag;
2704 
2705 			switch (vp->v_type) {
2706 			case VREG:
2707 			case VDIR:
2708 			case VLNK:
2709 				break;
2710 			case VBLK:
2711 			case VCHR:
2712 				if (vp->v_rdev == NULL) {
2713 					vrele(vp);
2714 					continue;
2715 				}
2716 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2717 				break;
2718 			case VSOCK:
2719 				xvn[n].xv_socket = vp->v_socket;
2720 				break;
2721 			case VFIFO:
2722 				xvn[n].xv_fifo = vp->v_fifoinfo;
2723 				break;
2724 			case VNON:
2725 			case VBAD:
2726 			default:
2727 				/* shouldn't happen? */
2728 				vrele(vp);
2729 				continue;
2730 			}
2731 			vrele(vp);
2732 			++n;
2733 		}
2734 		MNT_IUNLOCK(mp);
2735 		mtx_lock(&mountlist_mtx);
2736 		vfs_unbusy(mp, td);
2737 		if (n == len)
2738 			break;
2739 	}
2740 	mtx_unlock(&mountlist_mtx);
2741 
2742 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2743 	free(xvn, M_TEMP);
2744 	return (error);
2745 }
2746 
2747 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2748 	0, 0, sysctl_vnode, "S,xvnode", "");
2749 #endif
2750 
2751 /*
2752  * Unmount all filesystems. The list is traversed in reverse order
2753  * of mounting to avoid dependencies.
2754  */
2755 void
2756 vfs_unmountall(void)
2757 {
2758 	struct mount *mp;
2759 	struct thread *td;
2760 	int error;
2761 
2762 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2763 	td = curthread;
2764 	/*
2765 	 * Since this only runs when rebooting, it is not interlocked.
2766 	 */
2767 	while(!TAILQ_EMPTY(&mountlist)) {
2768 		mp = TAILQ_LAST(&mountlist, mntlist);
2769 		error = dounmount(mp, MNT_FORCE, td);
2770 		if (error) {
2771 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2772 			/*
2773 			 * XXX: Due to the way in which we mount the root
2774 			 * file system off of devfs, devfs will generate a
2775 			 * "busy" warning when we try to unmount it before
2776 			 * the root.  Don't print a warning as a result in
2777 			 * order to avoid false positive errors that may
2778 			 * cause needless upset.
2779 			 */
2780 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2781 				printf("unmount of %s failed (",
2782 				    mp->mnt_stat.f_mntonname);
2783 				if (error == EBUSY)
2784 					printf("BUSY)\n");
2785 				else
2786 					printf("%d)\n", error);
2787 			}
2788 		} else {
2789 			/* The unmount has removed mp from the mountlist */
2790 		}
2791 	}
2792 }
2793 
2794 /*
2795  * perform msync on all vnodes under a mount point
2796  * the mount point must be locked.
2797  */
2798 void
2799 vfs_msync(struct mount *mp, int flags)
2800 {
2801 	struct vnode *vp, *mvp;
2802 	struct vm_object *obj;
2803 
2804 	(void) vn_start_write(NULL, &mp, V_WAIT);
2805 	MNT_ILOCK(mp);
2806 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2807 		VI_LOCK(vp);
2808 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2809 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2810 			MNT_IUNLOCK(mp);
2811 			if (!vget(vp,
2812 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2813 			    curthread)) {
2814 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2815 					vput(vp);
2816 					MNT_ILOCK(mp);
2817 					continue;
2818 				}
2819 
2820 				obj = vp->v_object;
2821 				if (obj != NULL) {
2822 					VM_OBJECT_LOCK(obj);
2823 					vm_object_page_clean(obj, 0, 0,
2824 					    flags == MNT_WAIT ?
2825 					    OBJPC_SYNC : OBJPC_NOSYNC);
2826 					VM_OBJECT_UNLOCK(obj);
2827 				}
2828 				vput(vp);
2829 			}
2830 			MNT_ILOCK(mp);
2831 		} else
2832 			VI_UNLOCK(vp);
2833 	}
2834 	MNT_IUNLOCK(mp);
2835 	vn_finished_write(mp);
2836 }
2837 
2838 /*
2839  * Mark a vnode as free, putting it up for recycling.
2840  */
2841 static void
2842 vfree(struct vnode *vp)
2843 {
2844 
2845 	CTR1(KTR_VFS, "vfree vp %p", vp);
2846 	ASSERT_VI_LOCKED(vp, "vfree");
2847 	mtx_lock(&vnode_free_list_mtx);
2848 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2849 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2850 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2851 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2852 	    ("vfree: Freeing doomed vnode"));
2853 	if (vp->v_iflag & VI_AGE) {
2854 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2855 	} else {
2856 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2857 	}
2858 	freevnodes++;
2859 	vp->v_iflag &= ~VI_AGE;
2860 	vp->v_iflag |= VI_FREE;
2861 	mtx_unlock(&vnode_free_list_mtx);
2862 }
2863 
2864 /*
2865  * Opposite of vfree() - mark a vnode as in use.
2866  */
2867 static void
2868 vbusy(struct vnode *vp)
2869 {
2870 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2871 	ASSERT_VI_LOCKED(vp, "vbusy");
2872 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2873 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2874 
2875 	mtx_lock(&vnode_free_list_mtx);
2876 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2877 	freevnodes--;
2878 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2879 	mtx_unlock(&vnode_free_list_mtx);
2880 }
2881 
2882 /*
2883  * Initalize per-vnode helper structure to hold poll-related state.
2884  */
2885 void
2886 v_addpollinfo(struct vnode *vp)
2887 {
2888 	struct vpollinfo *vi;
2889 
2890 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2891 	if (vp->v_pollinfo != NULL) {
2892 		uma_zfree(vnodepoll_zone, vi);
2893 		return;
2894 	}
2895 	vp->v_pollinfo = vi;
2896 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2897 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
2898 	    vfs_knlunlock, vfs_knllocked);
2899 }
2900 
2901 /*
2902  * Record a process's interest in events which might happen to
2903  * a vnode.  Because poll uses the historic select-style interface
2904  * internally, this routine serves as both the ``check for any
2905  * pending events'' and the ``record my interest in future events''
2906  * functions.  (These are done together, while the lock is held,
2907  * to avoid race conditions.)
2908  */
2909 int
2910 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
2911 {
2912 
2913 	if (vp->v_pollinfo == NULL)
2914 		v_addpollinfo(vp);
2915 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2916 	if (vp->v_pollinfo->vpi_revents & events) {
2917 		/*
2918 		 * This leaves events we are not interested
2919 		 * in available for the other process which
2920 		 * which presumably had requested them
2921 		 * (otherwise they would never have been
2922 		 * recorded).
2923 		 */
2924 		events &= vp->v_pollinfo->vpi_revents;
2925 		vp->v_pollinfo->vpi_revents &= ~events;
2926 
2927 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2928 		return events;
2929 	}
2930 	vp->v_pollinfo->vpi_events |= events;
2931 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2932 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2933 	return 0;
2934 }
2935 
2936 /*
2937  * Routine to create and manage a filesystem syncer vnode.
2938  */
2939 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2940 static int	sync_fsync(struct  vop_fsync_args *);
2941 static int	sync_inactive(struct  vop_inactive_args *);
2942 static int	sync_reclaim(struct  vop_reclaim_args *);
2943 
2944 static struct vop_vector sync_vnodeops = {
2945 	.vop_bypass =	VOP_EOPNOTSUPP,
2946 	.vop_close =	sync_close,		/* close */
2947 	.vop_fsync =	sync_fsync,		/* fsync */
2948 	.vop_inactive =	sync_inactive,	/* inactive */
2949 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2950 	.vop_lock =	vop_stdlock,	/* lock */
2951 	.vop_unlock =	vop_stdunlock,	/* unlock */
2952 	.vop_islocked =	vop_stdislocked,	/* islocked */
2953 };
2954 
2955 /*
2956  * Create a new filesystem syncer vnode for the specified mount point.
2957  */
2958 int
2959 vfs_allocate_syncvnode(struct mount *mp)
2960 {
2961 	struct vnode *vp;
2962 	static long start, incr, next;
2963 	int error;
2964 
2965 	/* Allocate a new vnode */
2966 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2967 		mp->mnt_syncer = NULL;
2968 		return (error);
2969 	}
2970 	vp->v_type = VNON;
2971 	/*
2972 	 * Place the vnode onto the syncer worklist. We attempt to
2973 	 * scatter them about on the list so that they will go off
2974 	 * at evenly distributed times even if all the filesystems
2975 	 * are mounted at once.
2976 	 */
2977 	next += incr;
2978 	if (next == 0 || next > syncer_maxdelay) {
2979 		start /= 2;
2980 		incr /= 2;
2981 		if (start == 0) {
2982 			start = syncer_maxdelay / 2;
2983 			incr = syncer_maxdelay;
2984 		}
2985 		next = start;
2986 	}
2987 	VI_LOCK(vp);
2988 	vn_syncer_add_to_worklist(&vp->v_bufobj,
2989 	    syncdelay > 0 ? next % syncdelay : 0);
2990 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
2991 	mtx_lock(&sync_mtx);
2992 	sync_vnode_count++;
2993 	mtx_unlock(&sync_mtx);
2994 	VI_UNLOCK(vp);
2995 	mp->mnt_syncer = vp;
2996 	return (0);
2997 }
2998 
2999 /*
3000  * Do a lazy sync of the filesystem.
3001  */
3002 static int
3003 sync_fsync(struct vop_fsync_args *ap)
3004 {
3005 	struct vnode *syncvp = ap->a_vp;
3006 	struct mount *mp = syncvp->v_mount;
3007 	struct thread *td = ap->a_td;
3008 	int error, asyncflag;
3009 	struct bufobj *bo;
3010 
3011 	/*
3012 	 * We only need to do something if this is a lazy evaluation.
3013 	 */
3014 	if (ap->a_waitfor != MNT_LAZY)
3015 		return (0);
3016 
3017 	/*
3018 	 * Move ourselves to the back of the sync list.
3019 	 */
3020 	bo = &syncvp->v_bufobj;
3021 	BO_LOCK(bo);
3022 	vn_syncer_add_to_worklist(bo, syncdelay);
3023 	BO_UNLOCK(bo);
3024 
3025 	/*
3026 	 * Walk the list of vnodes pushing all that are dirty and
3027 	 * not already on the sync list.
3028 	 */
3029 	mtx_lock(&mountlist_mtx);
3030 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3031 		mtx_unlock(&mountlist_mtx);
3032 		return (0);
3033 	}
3034 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3035 		vfs_unbusy(mp, td);
3036 		return (0);
3037 	}
3038 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3039 	mp->mnt_flag &= ~MNT_ASYNC;
3040 	vfs_msync(mp, MNT_NOWAIT);
3041 	error = VFS_SYNC(mp, MNT_LAZY, td);
3042 	if (asyncflag)
3043 		mp->mnt_flag |= MNT_ASYNC;
3044 	vn_finished_write(mp);
3045 	vfs_unbusy(mp, td);
3046 	return (error);
3047 }
3048 
3049 /*
3050  * The syncer vnode is no referenced.
3051  */
3052 static int
3053 sync_inactive(struct vop_inactive_args *ap)
3054 {
3055 
3056 	vgone(ap->a_vp);
3057 	return (0);
3058 }
3059 
3060 /*
3061  * The syncer vnode is no longer needed and is being decommissioned.
3062  *
3063  * Modifications to the worklist must be protected by sync_mtx.
3064  */
3065 static int
3066 sync_reclaim(struct vop_reclaim_args *ap)
3067 {
3068 	struct vnode *vp = ap->a_vp;
3069 	struct bufobj *bo;
3070 
3071 	VI_LOCK(vp);
3072 	bo = &vp->v_bufobj;
3073 	vp->v_mount->mnt_syncer = NULL;
3074 	if (bo->bo_flag & BO_ONWORKLST) {
3075 		mtx_lock(&sync_mtx);
3076 		LIST_REMOVE(bo, bo_synclist);
3077  		syncer_worklist_len--;
3078 		sync_vnode_count--;
3079 		mtx_unlock(&sync_mtx);
3080 		bo->bo_flag &= ~BO_ONWORKLST;
3081 	}
3082 	VI_UNLOCK(vp);
3083 
3084 	return (0);
3085 }
3086 
3087 /*
3088  * Check if vnode represents a disk device
3089  */
3090 int
3091 vn_isdisk(struct vnode *vp, int *errp)
3092 {
3093 	int error;
3094 
3095 	error = 0;
3096 	dev_lock();
3097 	if (vp->v_type != VCHR)
3098 		error = ENOTBLK;
3099 	else if (vp->v_rdev == NULL)
3100 		error = ENXIO;
3101 	else if (vp->v_rdev->si_devsw == NULL)
3102 		error = ENXIO;
3103 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3104 		error = ENOTBLK;
3105 	dev_unlock();
3106 	if (errp != NULL)
3107 		*errp = error;
3108 	return (error == 0);
3109 }
3110 
3111 /*
3112  * Common filesystem object access control check routine.  Accepts a
3113  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3114  * and optional call-by-reference privused argument allowing vaccess()
3115  * to indicate to the caller whether privilege was used to satisfy the
3116  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3117  */
3118 int
3119 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3120     mode_t acc_mode, struct ucred *cred, int *privused)
3121 {
3122 	mode_t dac_granted;
3123 #ifdef CAPABILITIES
3124 	mode_t cap_granted;
3125 #endif
3126 
3127 	/*
3128 	 * Look for a normal, non-privileged way to access the file/directory
3129 	 * as requested.  If it exists, go with that.
3130 	 */
3131 
3132 	if (privused != NULL)
3133 		*privused = 0;
3134 
3135 	dac_granted = 0;
3136 
3137 	/* Check the owner. */
3138 	if (cred->cr_uid == file_uid) {
3139 		dac_granted |= VADMIN;
3140 		if (file_mode & S_IXUSR)
3141 			dac_granted |= VEXEC;
3142 		if (file_mode & S_IRUSR)
3143 			dac_granted |= VREAD;
3144 		if (file_mode & S_IWUSR)
3145 			dac_granted |= (VWRITE | VAPPEND);
3146 
3147 		if ((acc_mode & dac_granted) == acc_mode)
3148 			return (0);
3149 
3150 		goto privcheck;
3151 	}
3152 
3153 	/* Otherwise, check the groups (first match) */
3154 	if (groupmember(file_gid, cred)) {
3155 		if (file_mode & S_IXGRP)
3156 			dac_granted |= VEXEC;
3157 		if (file_mode & S_IRGRP)
3158 			dac_granted |= VREAD;
3159 		if (file_mode & S_IWGRP)
3160 			dac_granted |= (VWRITE | VAPPEND);
3161 
3162 		if ((acc_mode & dac_granted) == acc_mode)
3163 			return (0);
3164 
3165 		goto privcheck;
3166 	}
3167 
3168 	/* Otherwise, check everyone else. */
3169 	if (file_mode & S_IXOTH)
3170 		dac_granted |= VEXEC;
3171 	if (file_mode & S_IROTH)
3172 		dac_granted |= VREAD;
3173 	if (file_mode & S_IWOTH)
3174 		dac_granted |= (VWRITE | VAPPEND);
3175 	if ((acc_mode & dac_granted) == acc_mode)
3176 		return (0);
3177 
3178 privcheck:
3179 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3180 		/* XXX audit: privilege used */
3181 		if (privused != NULL)
3182 			*privused = 1;
3183 		return (0);
3184 	}
3185 
3186 #ifdef CAPABILITIES
3187 	/*
3188 	 * Build a capability mask to determine if the set of capabilities
3189 	 * satisfies the requirements when combined with the granted mask
3190 	 * from above.
3191 	 * For each capability, if the capability is required, bitwise
3192 	 * or the request type onto the cap_granted mask.
3193 	 */
3194 	cap_granted = 0;
3195 
3196 	if (type == VDIR) {
3197 		/*
3198 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3199 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3200 		 */
3201 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3202 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3203 			cap_granted |= VEXEC;
3204 	} else {
3205 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3206 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3207 			cap_granted |= VEXEC;
3208 	}
3209 
3210 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3211 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3212 		cap_granted |= VREAD;
3213 
3214 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3215 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3216 		cap_granted |= (VWRITE | VAPPEND);
3217 
3218 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3219 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3220 		cap_granted |= VADMIN;
3221 
3222 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3223 		/* XXX audit: privilege used */
3224 		if (privused != NULL)
3225 			*privused = 1;
3226 		return (0);
3227 	}
3228 #endif
3229 
3230 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3231 }
3232 
3233 /*
3234  * Credential check based on process requesting service, and per-attribute
3235  * permissions.
3236  */
3237 int
3238 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3239     struct thread *td, int access)
3240 {
3241 
3242 	/*
3243 	 * Kernel-invoked always succeeds.
3244 	 */
3245 	if (cred == NOCRED)
3246 		return (0);
3247 
3248 	/*
3249 	 * Do not allow privileged processes in jail to directly
3250 	 * manipulate system attributes.
3251 	 *
3252 	 * XXX What capability should apply here?
3253 	 * Probably CAP_SYS_SETFFLAG.
3254 	 */
3255 	switch (attrnamespace) {
3256 	case EXTATTR_NAMESPACE_SYSTEM:
3257 		/* Potentially should be: return (EPERM); */
3258 		return (suser_cred(cred, 0));
3259 	case EXTATTR_NAMESPACE_USER:
3260 		return (VOP_ACCESS(vp, access, cred, td));
3261 	default:
3262 		return (EPERM);
3263 	}
3264 }
3265 
3266 #ifdef DEBUG_VFS_LOCKS
3267 /*
3268  * This only exists to supress warnings from unlocked specfs accesses.  It is
3269  * no longer ok to have an unlocked VFS.
3270  */
3271 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3272 
3273 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3274 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3275 
3276 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3277 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3278 
3279 int vfs_badlock_print = 1;	/* Print lock violations. */
3280 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3281 
3282 #ifdef KDB
3283 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3284 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3285 #endif
3286 
3287 static void
3288 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3289 {
3290 
3291 #ifdef KDB
3292 	if (vfs_badlock_backtrace)
3293 		kdb_backtrace();
3294 #endif
3295 	if (vfs_badlock_print)
3296 		printf("%s: %p %s\n", str, (void *)vp, msg);
3297 	if (vfs_badlock_ddb)
3298 		kdb_enter("lock violation");
3299 }
3300 
3301 void
3302 assert_vi_locked(struct vnode *vp, const char *str)
3303 {
3304 
3305 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3306 		vfs_badlock("interlock is not locked but should be", str, vp);
3307 }
3308 
3309 void
3310 assert_vi_unlocked(struct vnode *vp, const char *str)
3311 {
3312 
3313 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3314 		vfs_badlock("interlock is locked but should not be", str, vp);
3315 }
3316 
3317 void
3318 assert_vop_locked(struct vnode *vp, const char *str)
3319 {
3320 
3321 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3322 		vfs_badlock("is not locked but should be", str, vp);
3323 }
3324 
3325 void
3326 assert_vop_unlocked(struct vnode *vp, const char *str)
3327 {
3328 
3329 	if (vp && !IGNORE_LOCK(vp) &&
3330 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3331 		vfs_badlock("is locked but should not be", str, vp);
3332 }
3333 
3334 void
3335 assert_vop_elocked(struct vnode *vp, const char *str)
3336 {
3337 
3338 	if (vp && !IGNORE_LOCK(vp) &&
3339 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3340 		vfs_badlock("is not exclusive locked but should be", str, vp);
3341 }
3342 
3343 #if 0
3344 void
3345 assert_vop_elocked_other(struct vnode *vp, const char *str)
3346 {
3347 
3348 	if (vp && !IGNORE_LOCK(vp) &&
3349 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3350 		vfs_badlock("is not exclusive locked by another thread",
3351 		    str, vp);
3352 }
3353 
3354 void
3355 assert_vop_slocked(struct vnode *vp, const char *str)
3356 {
3357 
3358 	if (vp && !IGNORE_LOCK(vp) &&
3359 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3360 		vfs_badlock("is not locked shared but should be", str, vp);
3361 }
3362 #endif /* 0 */
3363 #endif /* DEBUG_VFS_LOCKS */
3364 
3365 void
3366 vop_rename_pre(void *ap)
3367 {
3368 	struct vop_rename_args *a = ap;
3369 
3370 #ifdef DEBUG_VFS_LOCKS
3371 	if (a->a_tvp)
3372 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3373 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3374 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3375 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3376 
3377 	/* Check the source (from). */
3378 	if (a->a_tdvp != a->a_fdvp)
3379 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3380 	if (a->a_tvp != a->a_fvp)
3381 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3382 
3383 	/* Check the target. */
3384 	if (a->a_tvp)
3385 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3386 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3387 #endif
3388 	if (a->a_tdvp != a->a_fdvp)
3389 		vhold(a->a_fdvp);
3390 	if (a->a_tvp != a->a_fvp)
3391 		vhold(a->a_fvp);
3392 	vhold(a->a_tdvp);
3393 	if (a->a_tvp)
3394 		vhold(a->a_tvp);
3395 }
3396 
3397 void
3398 vop_strategy_pre(void *ap)
3399 {
3400 #ifdef DEBUG_VFS_LOCKS
3401 	struct vop_strategy_args *a;
3402 	struct buf *bp;
3403 
3404 	a = ap;
3405 	bp = a->a_bp;
3406 
3407 	/*
3408 	 * Cluster ops lock their component buffers but not the IO container.
3409 	 */
3410 	if ((bp->b_flags & B_CLUSTER) != 0)
3411 		return;
3412 
3413 	if (BUF_REFCNT(bp) < 1) {
3414 		if (vfs_badlock_print)
3415 			printf(
3416 			    "VOP_STRATEGY: bp is not locked but should be\n");
3417 		if (vfs_badlock_ddb)
3418 			kdb_enter("lock violation");
3419 	}
3420 #endif
3421 }
3422 
3423 void
3424 vop_lookup_pre(void *ap)
3425 {
3426 #ifdef DEBUG_VFS_LOCKS
3427 	struct vop_lookup_args *a;
3428 	struct vnode *dvp;
3429 
3430 	a = ap;
3431 	dvp = a->a_dvp;
3432 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3433 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3434 #endif
3435 }
3436 
3437 void
3438 vop_lookup_post(void *ap, int rc)
3439 {
3440 #ifdef DEBUG_VFS_LOCKS
3441 	struct vop_lookup_args *a;
3442 	struct vnode *dvp;
3443 	struct vnode *vp;
3444 
3445 	a = ap;
3446 	dvp = a->a_dvp;
3447 	vp = *(a->a_vpp);
3448 
3449 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3450 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3451 
3452 	if (!rc)
3453 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3454 #endif
3455 }
3456 
3457 void
3458 vop_lock_pre(void *ap)
3459 {
3460 #ifdef DEBUG_VFS_LOCKS
3461 	struct vop_lock_args *a = ap;
3462 
3463 	if ((a->a_flags & LK_INTERLOCK) == 0)
3464 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3465 	else
3466 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3467 #endif
3468 }
3469 
3470 void
3471 vop_lock_post(void *ap, int rc)
3472 {
3473 #ifdef DEBUG_VFS_LOCKS
3474 	struct vop_lock_args *a = ap;
3475 
3476 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3477 	if (rc == 0)
3478 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3479 #endif
3480 }
3481 
3482 void
3483 vop_unlock_pre(void *ap)
3484 {
3485 #ifdef DEBUG_VFS_LOCKS
3486 	struct vop_unlock_args *a = ap;
3487 
3488 	if (a->a_flags & LK_INTERLOCK)
3489 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3490 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3491 #endif
3492 }
3493 
3494 void
3495 vop_unlock_post(void *ap, int rc)
3496 {
3497 #ifdef DEBUG_VFS_LOCKS
3498 	struct vop_unlock_args *a = ap;
3499 
3500 	if (a->a_flags & LK_INTERLOCK)
3501 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3502 #endif
3503 }
3504 
3505 void
3506 vop_create_post(void *ap, int rc)
3507 {
3508 	struct vop_create_args *a = ap;
3509 
3510 	if (!rc)
3511 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3512 }
3513 
3514 void
3515 vop_link_post(void *ap, int rc)
3516 {
3517 	struct vop_link_args *a = ap;
3518 
3519 	if (!rc) {
3520 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3521 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3522 	}
3523 }
3524 
3525 void
3526 vop_mkdir_post(void *ap, int rc)
3527 {
3528 	struct vop_mkdir_args *a = ap;
3529 
3530 	if (!rc)
3531 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3532 }
3533 
3534 void
3535 vop_mknod_post(void *ap, int rc)
3536 {
3537 	struct vop_mknod_args *a = ap;
3538 
3539 	if (!rc)
3540 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3541 }
3542 
3543 void
3544 vop_remove_post(void *ap, int rc)
3545 {
3546 	struct vop_remove_args *a = ap;
3547 
3548 	if (!rc) {
3549 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3550 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3551 	}
3552 }
3553 
3554 void
3555 vop_rename_post(void *ap, int rc)
3556 {
3557 	struct vop_rename_args *a = ap;
3558 
3559 	if (!rc) {
3560 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3561 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3562 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3563 		if (a->a_tvp)
3564 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3565 	}
3566 	if (a->a_tdvp != a->a_fdvp)
3567 		vdrop(a->a_fdvp);
3568 	if (a->a_tvp != a->a_fvp)
3569 		vdrop(a->a_fvp);
3570 	vdrop(a->a_tdvp);
3571 	if (a->a_tvp)
3572 		vdrop(a->a_tvp);
3573 }
3574 
3575 void
3576 vop_rmdir_post(void *ap, int rc)
3577 {
3578 	struct vop_rmdir_args *a = ap;
3579 
3580 	if (!rc) {
3581 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3582 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3583 	}
3584 }
3585 
3586 void
3587 vop_setattr_post(void *ap, int rc)
3588 {
3589 	struct vop_setattr_args *a = ap;
3590 
3591 	if (!rc)
3592 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3593 }
3594 
3595 void
3596 vop_symlink_post(void *ap, int rc)
3597 {
3598 	struct vop_symlink_args *a = ap;
3599 
3600 	if (!rc)
3601 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3602 }
3603 
3604 static struct knlist fs_knlist;
3605 
3606 static void
3607 vfs_event_init(void *arg)
3608 {
3609 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3610 }
3611 /* XXX - correct order? */
3612 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3613 
3614 void
3615 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3616 {
3617 
3618 	KNOTE_UNLOCKED(&fs_knlist, event);
3619 }
3620 
3621 static int	filt_fsattach(struct knote *kn);
3622 static void	filt_fsdetach(struct knote *kn);
3623 static int	filt_fsevent(struct knote *kn, long hint);
3624 
3625 struct filterops fs_filtops =
3626 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3627 
3628 static int
3629 filt_fsattach(struct knote *kn)
3630 {
3631 
3632 	kn->kn_flags |= EV_CLEAR;
3633 	knlist_add(&fs_knlist, kn, 0);
3634 	return (0);
3635 }
3636 
3637 static void
3638 filt_fsdetach(struct knote *kn)
3639 {
3640 
3641 	knlist_remove(&fs_knlist, kn, 0);
3642 }
3643 
3644 static int
3645 filt_fsevent(struct knote *kn, long hint)
3646 {
3647 
3648 	kn->kn_fflags |= hint;
3649 	return (kn->kn_fflags != 0);
3650 }
3651 
3652 static int
3653 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3654 {
3655 	struct vfsidctl vc;
3656 	int error;
3657 	struct mount *mp;
3658 
3659 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3660 	if (error)
3661 		return (error);
3662 	if (vc.vc_vers != VFS_CTL_VERS1)
3663 		return (EINVAL);
3664 	mp = vfs_getvfs(&vc.vc_fsid);
3665 	if (mp == NULL)
3666 		return (ENOENT);
3667 	/* ensure that a specific sysctl goes to the right filesystem. */
3668 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3669 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3670 		return (EINVAL);
3671 	}
3672 	VCTLTOREQ(&vc, req);
3673 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3674 }
3675 
3676 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3677         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3678 
3679 /*
3680  * Function to initialize a va_filerev field sensibly.
3681  * XXX: Wouldn't a random number make a lot more sense ??
3682  */
3683 u_quad_t
3684 init_va_filerev(void)
3685 {
3686 	struct bintime bt;
3687 
3688 	getbinuptime(&bt);
3689 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3690 }
3691 
3692 static int	filt_vfsread(struct knote *kn, long hint);
3693 static int	filt_vfswrite(struct knote *kn, long hint);
3694 static int	filt_vfsvnode(struct knote *kn, long hint);
3695 static void	filt_vfsdetach(struct knote *kn);
3696 static struct filterops vfsread_filtops =
3697 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3698 static struct filterops vfswrite_filtops =
3699 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3700 static struct filterops vfsvnode_filtops =
3701 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3702 
3703 static void
3704 vfs_knllock(void *arg)
3705 {
3706 	struct vnode *vp = arg;
3707 
3708 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
3709 }
3710 
3711 static void
3712 vfs_knlunlock(void *arg)
3713 {
3714 	struct vnode *vp = arg;
3715 
3716 	VOP_UNLOCK(vp, 0, curthread);
3717 }
3718 
3719 static int
3720 vfs_knllocked(void *arg)
3721 {
3722 	struct vnode *vp = arg;
3723 
3724 	return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE);
3725 }
3726 
3727 int
3728 vfs_kqfilter(struct vop_kqfilter_args *ap)
3729 {
3730 	struct vnode *vp = ap->a_vp;
3731 	struct knote *kn = ap->a_kn;
3732 	struct knlist *knl;
3733 
3734 	switch (kn->kn_filter) {
3735 	case EVFILT_READ:
3736 		kn->kn_fop = &vfsread_filtops;
3737 		break;
3738 	case EVFILT_WRITE:
3739 		kn->kn_fop = &vfswrite_filtops;
3740 		break;
3741 	case EVFILT_VNODE:
3742 		kn->kn_fop = &vfsvnode_filtops;
3743 		break;
3744 	default:
3745 		return (EINVAL);
3746 	}
3747 
3748 	kn->kn_hook = (caddr_t)vp;
3749 
3750 	if (vp->v_pollinfo == NULL)
3751 		v_addpollinfo(vp);
3752 	if (vp->v_pollinfo == NULL)
3753 		return (ENOMEM);
3754 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3755 	knlist_add(knl, kn, 0);
3756 
3757 	return (0);
3758 }
3759 
3760 /*
3761  * Detach knote from vnode
3762  */
3763 static void
3764 filt_vfsdetach(struct knote *kn)
3765 {
3766 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3767 
3768 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3769 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3770 }
3771 
3772 /*ARGSUSED*/
3773 static int
3774 filt_vfsread(struct knote *kn, long hint)
3775 {
3776 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3777 	struct vattr va;
3778 
3779 	/*
3780 	 * filesystem is gone, so set the EOF flag and schedule
3781 	 * the knote for deletion.
3782 	 */
3783 	if (hint == NOTE_REVOKE) {
3784 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3785 		return (1);
3786 	}
3787 
3788 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3789 		return (0);
3790 
3791 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3792 	return (kn->kn_data != 0);
3793 }
3794 
3795 /*ARGSUSED*/
3796 static int
3797 filt_vfswrite(struct knote *kn, long hint)
3798 {
3799 	/*
3800 	 * filesystem is gone, so set the EOF flag and schedule
3801 	 * the knote for deletion.
3802 	 */
3803 	if (hint == NOTE_REVOKE)
3804 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3805 
3806 	kn->kn_data = 0;
3807 	return (1);
3808 }
3809 
3810 static int
3811 filt_vfsvnode(struct knote *kn, long hint)
3812 {
3813 	if (kn->kn_sfflags & hint)
3814 		kn->kn_fflags |= hint;
3815 	if (hint == NOTE_REVOKE) {
3816 		kn->kn_flags |= EV_EOF;
3817 		return (1);
3818 	}
3819 	return (kn->kn_fflags != 0);
3820 }
3821 
3822 int
3823 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3824 {
3825 	int error;
3826 
3827 	if (dp->d_reclen > ap->a_uio->uio_resid)
3828 		return (ENAMETOOLONG);
3829 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3830 	if (error) {
3831 		if (ap->a_ncookies != NULL) {
3832 			if (ap->a_cookies != NULL)
3833 				free(ap->a_cookies, M_TEMP);
3834 			ap->a_cookies = NULL;
3835 			*ap->a_ncookies = 0;
3836 		}
3837 		return (error);
3838 	}
3839 	if (ap->a_ncookies == NULL)
3840 		return (0);
3841 
3842 	KASSERT(ap->a_cookies,
3843 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
3844 
3845 	*ap->a_cookies = realloc(*ap->a_cookies,
3846 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
3847 	(*ap->a_cookies)[*ap->a_ncookies] = off;
3848 	return (0);
3849 }
3850 
3851 /*
3852  * Mark for update the access time of the file if the filesystem
3853  * supports VA_MARK_ATIME.  This functionality is used by execve
3854  * and mmap, so we want to avoid the synchronous I/O implied by
3855  * directly setting va_atime for the sake of efficiency.
3856  */
3857 void
3858 vfs_mark_atime(struct vnode *vp, struct thread *td)
3859 {
3860 	struct vattr atimeattr;
3861 
3862 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
3863 		VATTR_NULL(&atimeattr);
3864 		atimeattr.va_vaflags |= VA_MARK_ATIME;
3865 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
3866 	}
3867 }
3868