xref: /freebsd/sys/kern/vfs_subr.c (revision d8010b1175094374c8295d0fba56c2402b32a4da)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/stat.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/uma.h>
99 
100 #ifdef DDB
101 #include <ddb/ddb.h>
102 #endif
103 
104 static void	delmntque(struct vnode *vp);
105 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 		    int slpflag, int slptimeo);
107 static void	syncer_shutdown(void *arg, int howto);
108 static int	vtryrecycle(struct vnode *vp);
109 static void	v_init_counters(struct vnode *);
110 static void	v_incr_devcount(struct vnode *);
111 static void	v_decr_devcount(struct vnode *);
112 static void	vgonel(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_locked(void *arg);
116 static void	vfs_knl_assert_unlocked(void *arg);
117 static void	vnlru_return_batches(struct vfsops *mnt_op);
118 static void	destroy_vpollinfo(struct vpollinfo *vi);
119 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
120 		    daddr_t startlbn, daddr_t endlbn);
121 
122 /*
123  * These fences are intended for cases where some synchronization is
124  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
125  * and v_usecount) updates.  Access to v_iflags is generally synchronized
126  * by the interlock, but we have some internal assertions that check vnode
127  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
128  * for now.
129  */
130 #ifdef INVARIANTS
131 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
132 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
133 #else
134 #define	VNODE_REFCOUNT_FENCE_ACQ()
135 #define	VNODE_REFCOUNT_FENCE_REL()
136 #endif
137 
138 /*
139  * Number of vnodes in existence.  Increased whenever getnewvnode()
140  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141  */
142 static unsigned long	numvnodes;
143 
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145     "Number of vnodes in existence");
146 
147 static counter_u64_t vnodes_created;
148 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
149     "Number of vnodes created by getnewvnode");
150 
151 static u_long mnt_free_list_batch = 128;
152 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
153     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
154 
155 /*
156  * Conversion tables for conversion from vnode types to inode formats
157  * and back.
158  */
159 enum vtype iftovt_tab[16] = {
160 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
161 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
162 };
163 int vttoif_tab[10] = {
164 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
165 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
166 };
167 
168 /*
169  * List of vnodes that are ready for recycling.
170  */
171 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
172 
173 /*
174  * "Free" vnode target.  Free vnodes are rarely completely free, but are
175  * just ones that are cheap to recycle.  Usually they are for files which
176  * have been stat'd but not read; these usually have inode and namecache
177  * data attached to them.  This target is the preferred minimum size of a
178  * sub-cache consisting mostly of such files. The system balances the size
179  * of this sub-cache with its complement to try to prevent either from
180  * thrashing while the other is relatively inactive.  The targets express
181  * a preference for the best balance.
182  *
183  * "Above" this target there are 2 further targets (watermarks) related
184  * to recyling of free vnodes.  In the best-operating case, the cache is
185  * exactly full, the free list has size between vlowat and vhiwat above the
186  * free target, and recycling from it and normal use maintains this state.
187  * Sometimes the free list is below vlowat or even empty, but this state
188  * is even better for immediate use provided the cache is not full.
189  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
190  * ones) to reach one of these states.  The watermarks are currently hard-
191  * coded as 4% and 9% of the available space higher.  These and the default
192  * of 25% for wantfreevnodes are too large if the memory size is large.
193  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
194  * whenever vnlru_proc() becomes active.
195  */
196 static u_long wantfreevnodes;
197 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
198     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
199 static u_long freevnodes;
200 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
201     &freevnodes, 0, "Number of \"free\" vnodes");
202 
203 static counter_u64_t recycles_count;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
205     "Number of vnodes recycled to meet vnode cache targets");
206 
207 /*
208  * Various variables used for debugging the new implementation of
209  * reassignbuf().
210  * XXX these are probably of (very) limited utility now.
211  */
212 static int reassignbufcalls;
213 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
214     &reassignbufcalls, 0, "Number of calls to reassignbuf");
215 
216 static counter_u64_t free_owe_inact;
217 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
218     "Number of times free vnodes kept on active list due to VFS "
219     "owing inactivation");
220 
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
223 
224 /*
225  * Lock for any access to the following:
226  *	vnode_free_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx vnode_free_list_mtx;
231 
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
234 
235 static uma_zone_t buf_trie_zone;
236 
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
240 
241 /*
242  * The workitem queue.
243  *
244  * It is useful to delay writes of file data and filesystem metadata
245  * for tens of seconds so that quickly created and deleted files need
246  * not waste disk bandwidth being created and removed. To realize this,
247  * we append vnodes to a "workitem" queue. When running with a soft
248  * updates implementation, most pending metadata dependencies should
249  * not wait for more than a few seconds. Thus, mounted on block devices
250  * are delayed only about a half the time that file data is delayed.
251  * Similarly, directory updates are more critical, so are only delayed
252  * about a third the time that file data is delayed. Thus, there are
253  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254  * one each second (driven off the filesystem syncer process). The
255  * syncer_delayno variable indicates the next queue that is to be processed.
256  * Items that need to be processed soon are placed in this queue:
257  *
258  *	syncer_workitem_pending[syncer_delayno]
259  *
260  * A delay of fifteen seconds is done by placing the request fifteen
261  * entries later in the queue:
262  *
263  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264  *
265  */
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
270 /*
271  * The sync_mtx protects:
272  *	bo->bo_synclist
273  *	sync_vnode_count
274  *	syncer_delayno
275  *	syncer_state
276  *	syncer_workitem_pending
277  *	syncer_worklist_len
278  *	rushjob
279  */
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
282 
283 #define SYNCER_MAXDELAY		32
284 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285 static int syncdelay = 30;		/* max time to delay syncing data */
286 static int filedelay = 30;		/* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288     "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29;		/* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291     "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28;		/* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294     "Time to delay syncing metadata (in seconds)");
295 static int rushjob;		/* number of slots to run ASAP */
296 static int stat_rush_requests;	/* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298     "Number of times I/O speeded up (rush requests)");
299 
300 /*
301  * When shutting down the syncer, run it at four times normal speed.
302  */
303 #define SYNCER_SHUTDOWN_SPEEDUP		4
304 static int sync_vnode_count;
305 static int syncer_worklist_len;
306 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
307     syncer_state;
308 
309 /* Target for maximum number of vnodes. */
310 int desiredvnodes;
311 static int gapvnodes;		/* gap between wanted and desired */
312 static int vhiwat;		/* enough extras after expansion */
313 static int vlowat;		/* minimal extras before expansion */
314 static int vstir;		/* nonzero to stir non-free vnodes */
315 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
316 
317 static int
318 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
319 {
320 	int error, old_desiredvnodes;
321 
322 	old_desiredvnodes = desiredvnodes;
323 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
324 		return (error);
325 	if (old_desiredvnodes != desiredvnodes) {
326 		wantfreevnodes = desiredvnodes / 4;
327 		/* XXX locking seems to be incomplete. */
328 		vfs_hash_changesize(desiredvnodes);
329 		cache_changesize(desiredvnodes);
330 	}
331 	return (0);
332 }
333 
334 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
335     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
336     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
337 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
338     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
339 static int vnlru_nowhere;
340 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
341     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
342 
343 static int
344 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
345 {
346 	struct vnode *vp;
347 	struct nameidata nd;
348 	char *buf;
349 	unsigned long ndflags;
350 	int error;
351 
352 	if (req->newptr == NULL)
353 		return (EINVAL);
354 	if (req->newlen >= PATH_MAX)
355 		return (E2BIG);
356 
357 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
358 	error = SYSCTL_IN(req, buf, req->newlen);
359 	if (error != 0)
360 		goto out;
361 
362 	buf[req->newlen] = '\0';
363 
364 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
365 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
366 	if ((error = namei(&nd)) != 0)
367 		goto out;
368 	vp = nd.ni_vp;
369 
370 	if (VN_IS_DOOMED(vp)) {
371 		/*
372 		 * This vnode is being recycled.  Return != 0 to let the caller
373 		 * know that the sysctl had no effect.  Return EAGAIN because a
374 		 * subsequent call will likely succeed (since namei will create
375 		 * a new vnode if necessary)
376 		 */
377 		error = EAGAIN;
378 		goto putvnode;
379 	}
380 
381 	counter_u64_add(recycles_count, 1);
382 	vgone(vp);
383 putvnode:
384 	NDFREE(&nd, 0);
385 out:
386 	free(buf, M_TEMP);
387 	return (error);
388 }
389 
390 static int
391 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
392 {
393 	struct thread *td = curthread;
394 	struct vnode *vp;
395 	struct file *fp;
396 	int error;
397 	int fd;
398 
399 	if (req->newptr == NULL)
400 		return (EBADF);
401 
402         error = sysctl_handle_int(oidp, &fd, 0, req);
403         if (error != 0)
404                 return (error);
405 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
406 	if (error != 0)
407 		return (error);
408 	vp = fp->f_vnode;
409 
410 	error = vn_lock(vp, LK_EXCLUSIVE);
411 	if (error != 0)
412 		goto drop;
413 
414 	counter_u64_add(recycles_count, 1);
415 	vgone(vp);
416 	VOP_UNLOCK(vp, 0);
417 drop:
418 	fdrop(fp, td);
419 	return (error);
420 }
421 
422 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
423     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
424     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
425 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
426     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
427     sysctl_ftry_reclaim_vnode, "I",
428     "Try to reclaim a vnode by its file descriptor");
429 
430 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
431 static int vnsz2log;
432 
433 /*
434  * Support for the bufobj clean & dirty pctrie.
435  */
436 static void *
437 buf_trie_alloc(struct pctrie *ptree)
438 {
439 
440 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
441 }
442 
443 static void
444 buf_trie_free(struct pctrie *ptree, void *node)
445 {
446 
447 	uma_zfree(buf_trie_zone, node);
448 }
449 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
450 
451 /*
452  * Initialize the vnode management data structures.
453  *
454  * Reevaluate the following cap on the number of vnodes after the physical
455  * memory size exceeds 512GB.  In the limit, as the physical memory size
456  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
457  */
458 #ifndef	MAXVNODES_MAX
459 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
460 #endif
461 
462 /*
463  * Initialize a vnode as it first enters the zone.
464  */
465 static int
466 vnode_init(void *mem, int size, int flags)
467 {
468 	struct vnode *vp;
469 
470 	vp = mem;
471 	bzero(vp, size);
472 	/*
473 	 * Setup locks.
474 	 */
475 	vp->v_vnlock = &vp->v_lock;
476 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
477 	/*
478 	 * By default, don't allow shared locks unless filesystems opt-in.
479 	 */
480 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
481 	    LK_NOSHARE | LK_IS_VNODE);
482 	/*
483 	 * Initialize bufobj.
484 	 */
485 	bufobj_init(&vp->v_bufobj, vp);
486 	/*
487 	 * Initialize namecache.
488 	 */
489 	LIST_INIT(&vp->v_cache_src);
490 	TAILQ_INIT(&vp->v_cache_dst);
491 	/*
492 	 * Initialize rangelocks.
493 	 */
494 	rangelock_init(&vp->v_rl);
495 	return (0);
496 }
497 
498 /*
499  * Free a vnode when it is cleared from the zone.
500  */
501 static void
502 vnode_fini(void *mem, int size)
503 {
504 	struct vnode *vp;
505 	struct bufobj *bo;
506 
507 	vp = mem;
508 	rangelock_destroy(&vp->v_rl);
509 	lockdestroy(vp->v_vnlock);
510 	mtx_destroy(&vp->v_interlock);
511 	bo = &vp->v_bufobj;
512 	rw_destroy(BO_LOCKPTR(bo));
513 }
514 
515 /*
516  * Provide the size of NFS nclnode and NFS fh for calculation of the
517  * vnode memory consumption.  The size is specified directly to
518  * eliminate dependency on NFS-private header.
519  *
520  * Other filesystems may use bigger or smaller (like UFS and ZFS)
521  * private inode data, but the NFS-based estimation is ample enough.
522  * Still, we care about differences in the size between 64- and 32-bit
523  * platforms.
524  *
525  * Namecache structure size is heuristically
526  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
527  */
528 #ifdef _LP64
529 #define	NFS_NCLNODE_SZ	(528 + 64)
530 #define	NC_SZ		148
531 #else
532 #define	NFS_NCLNODE_SZ	(360 + 32)
533 #define	NC_SZ		92
534 #endif
535 
536 static void
537 vntblinit(void *dummy __unused)
538 {
539 	u_int i;
540 	int physvnodes, virtvnodes;
541 
542 	/*
543 	 * Desiredvnodes is a function of the physical memory size and the
544 	 * kernel's heap size.  Generally speaking, it scales with the
545 	 * physical memory size.  The ratio of desiredvnodes to the physical
546 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
547 	 * Thereafter, the
548 	 * marginal ratio of desiredvnodes to the physical memory size is
549 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
550 	 * size.  The memory required by desiredvnodes vnodes and vm objects
551 	 * must not exceed 1/10th of the kernel's heap size.
552 	 */
553 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
554 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
555 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
556 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
557 	desiredvnodes = min(physvnodes, virtvnodes);
558 	if (desiredvnodes > MAXVNODES_MAX) {
559 		if (bootverbose)
560 			printf("Reducing kern.maxvnodes %d -> %d\n",
561 			    desiredvnodes, MAXVNODES_MAX);
562 		desiredvnodes = MAXVNODES_MAX;
563 	}
564 	wantfreevnodes = desiredvnodes / 4;
565 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
566 	TAILQ_INIT(&vnode_free_list);
567 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
568 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
569 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
570 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
571 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
572 	/*
573 	 * Preallocate enough nodes to support one-per buf so that
574 	 * we can not fail an insert.  reassignbuf() callers can not
575 	 * tolerate the insertion failure.
576 	 */
577 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
578 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
579 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
580 	uma_prealloc(buf_trie_zone, nbuf);
581 
582 	vnodes_created = counter_u64_alloc(M_WAITOK);
583 	recycles_count = counter_u64_alloc(M_WAITOK);
584 	free_owe_inact = counter_u64_alloc(M_WAITOK);
585 
586 	/*
587 	 * Initialize the filesystem syncer.
588 	 */
589 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
590 	    &syncer_mask);
591 	syncer_maxdelay = syncer_mask + 1;
592 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
593 	cv_init(&sync_wakeup, "syncer");
594 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
595 		vnsz2log++;
596 	vnsz2log--;
597 }
598 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
599 
600 
601 /*
602  * Mark a mount point as busy. Used to synchronize access and to delay
603  * unmounting. Eventually, mountlist_mtx is not released on failure.
604  *
605  * vfs_busy() is a custom lock, it can block the caller.
606  * vfs_busy() only sleeps if the unmount is active on the mount point.
607  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
608  * vnode belonging to mp.
609  *
610  * Lookup uses vfs_busy() to traverse mount points.
611  * root fs			var fs
612  * / vnode lock		A	/ vnode lock (/var)		D
613  * /var vnode lock	B	/log vnode lock(/var/log)	E
614  * vfs_busy lock	C	vfs_busy lock			F
615  *
616  * Within each file system, the lock order is C->A->B and F->D->E.
617  *
618  * When traversing across mounts, the system follows that lock order:
619  *
620  *        C->A->B
621  *              |
622  *              +->F->D->E
623  *
624  * The lookup() process for namei("/var") illustrates the process:
625  *  VOP_LOOKUP() obtains B while A is held
626  *  vfs_busy() obtains a shared lock on F while A and B are held
627  *  vput() releases lock on B
628  *  vput() releases lock on A
629  *  VFS_ROOT() obtains lock on D while shared lock on F is held
630  *  vfs_unbusy() releases shared lock on F
631  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
632  *    Attempt to lock A (instead of vp_crossmp) while D is held would
633  *    violate the global order, causing deadlocks.
634  *
635  * dounmount() locks B while F is drained.
636  */
637 int
638 vfs_busy(struct mount *mp, int flags)
639 {
640 
641 	MPASS((flags & ~MBF_MASK) == 0);
642 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
643 
644 	if (vfs_op_thread_enter(mp)) {
645 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
646 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
647 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
648 		vfs_mp_count_add_pcpu(mp, ref, 1);
649 		vfs_mp_count_add_pcpu(mp, lockref, 1);
650 		vfs_op_thread_exit(mp);
651 		if (flags & MBF_MNTLSTLOCK)
652 			mtx_unlock(&mountlist_mtx);
653 		return (0);
654 	}
655 
656 	MNT_ILOCK(mp);
657 	vfs_assert_mount_counters(mp);
658 	MNT_REF(mp);
659 	/*
660 	 * If mount point is currently being unmounted, sleep until the
661 	 * mount point fate is decided.  If thread doing the unmounting fails,
662 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
663 	 * that this mount point has survived the unmount attempt and vfs_busy
664 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
665 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
666 	 * about to be really destroyed.  vfs_busy needs to release its
667 	 * reference on the mount point in this case and return with ENOENT,
668 	 * telling the caller that mount mount it tried to busy is no longer
669 	 * valid.
670 	 */
671 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
672 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
673 			MNT_REL(mp);
674 			MNT_IUNLOCK(mp);
675 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
676 			    __func__);
677 			return (ENOENT);
678 		}
679 		if (flags & MBF_MNTLSTLOCK)
680 			mtx_unlock(&mountlist_mtx);
681 		mp->mnt_kern_flag |= MNTK_MWAIT;
682 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
683 		if (flags & MBF_MNTLSTLOCK)
684 			mtx_lock(&mountlist_mtx);
685 		MNT_ILOCK(mp);
686 	}
687 	if (flags & MBF_MNTLSTLOCK)
688 		mtx_unlock(&mountlist_mtx);
689 	mp->mnt_lockref++;
690 	MNT_IUNLOCK(mp);
691 	return (0);
692 }
693 
694 /*
695  * Free a busy filesystem.
696  */
697 void
698 vfs_unbusy(struct mount *mp)
699 {
700 	int c;
701 
702 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
703 
704 	if (vfs_op_thread_enter(mp)) {
705 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
706 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
707 		vfs_mp_count_sub_pcpu(mp, ref, 1);
708 		vfs_op_thread_exit(mp);
709 		return;
710 	}
711 
712 	MNT_ILOCK(mp);
713 	vfs_assert_mount_counters(mp);
714 	MNT_REL(mp);
715 	c = --mp->mnt_lockref;
716 	if (mp->mnt_vfs_ops == 0) {
717 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
718 		MNT_IUNLOCK(mp);
719 		return;
720 	}
721 	if (c < 0)
722 		vfs_dump_mount_counters(mp);
723 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
724 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
725 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
726 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
727 		wakeup(&mp->mnt_lockref);
728 	}
729 	MNT_IUNLOCK(mp);
730 }
731 
732 /*
733  * Lookup a mount point by filesystem identifier.
734  */
735 struct mount *
736 vfs_getvfs(fsid_t *fsid)
737 {
738 	struct mount *mp;
739 
740 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
741 	mtx_lock(&mountlist_mtx);
742 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
743 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
744 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
745 			vfs_ref(mp);
746 			mtx_unlock(&mountlist_mtx);
747 			return (mp);
748 		}
749 	}
750 	mtx_unlock(&mountlist_mtx);
751 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
752 	return ((struct mount *) 0);
753 }
754 
755 /*
756  * Lookup a mount point by filesystem identifier, busying it before
757  * returning.
758  *
759  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
760  * cache for popular filesystem identifiers.  The cache is lockess, using
761  * the fact that struct mount's are never freed.  In worst case we may
762  * get pointer to unmounted or even different filesystem, so we have to
763  * check what we got, and go slow way if so.
764  */
765 struct mount *
766 vfs_busyfs(fsid_t *fsid)
767 {
768 #define	FSID_CACHE_SIZE	256
769 	typedef struct mount * volatile vmp_t;
770 	static vmp_t cache[FSID_CACHE_SIZE];
771 	struct mount *mp;
772 	int error;
773 	uint32_t hash;
774 
775 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
776 	hash = fsid->val[0] ^ fsid->val[1];
777 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
778 	mp = cache[hash];
779 	if (mp == NULL ||
780 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
781 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
782 		goto slow;
783 	if (vfs_busy(mp, 0) != 0) {
784 		cache[hash] = NULL;
785 		goto slow;
786 	}
787 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
788 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
789 		return (mp);
790 	else
791 	    vfs_unbusy(mp);
792 
793 slow:
794 	mtx_lock(&mountlist_mtx);
795 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
796 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
797 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
798 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
799 			if (error) {
800 				cache[hash] = NULL;
801 				mtx_unlock(&mountlist_mtx);
802 				return (NULL);
803 			}
804 			cache[hash] = mp;
805 			return (mp);
806 		}
807 	}
808 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
809 	mtx_unlock(&mountlist_mtx);
810 	return ((struct mount *) 0);
811 }
812 
813 /*
814  * Check if a user can access privileged mount options.
815  */
816 int
817 vfs_suser(struct mount *mp, struct thread *td)
818 {
819 	int error;
820 
821 	if (jailed(td->td_ucred)) {
822 		/*
823 		 * If the jail of the calling thread lacks permission for
824 		 * this type of file system, deny immediately.
825 		 */
826 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
827 			return (EPERM);
828 
829 		/*
830 		 * If the file system was mounted outside the jail of the
831 		 * calling thread, deny immediately.
832 		 */
833 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
834 			return (EPERM);
835 	}
836 
837 	/*
838 	 * If file system supports delegated administration, we don't check
839 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
840 	 * by the file system itself.
841 	 * If this is not the user that did original mount, we check for
842 	 * the PRIV_VFS_MOUNT_OWNER privilege.
843 	 */
844 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
845 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
846 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
847 			return (error);
848 	}
849 	return (0);
850 }
851 
852 /*
853  * Get a new unique fsid.  Try to make its val[0] unique, since this value
854  * will be used to create fake device numbers for stat().  Also try (but
855  * not so hard) make its val[0] unique mod 2^16, since some emulators only
856  * support 16-bit device numbers.  We end up with unique val[0]'s for the
857  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
858  *
859  * Keep in mind that several mounts may be running in parallel.  Starting
860  * the search one past where the previous search terminated is both a
861  * micro-optimization and a defense against returning the same fsid to
862  * different mounts.
863  */
864 void
865 vfs_getnewfsid(struct mount *mp)
866 {
867 	static uint16_t mntid_base;
868 	struct mount *nmp;
869 	fsid_t tfsid;
870 	int mtype;
871 
872 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
873 	mtx_lock(&mntid_mtx);
874 	mtype = mp->mnt_vfc->vfc_typenum;
875 	tfsid.val[1] = mtype;
876 	mtype = (mtype & 0xFF) << 24;
877 	for (;;) {
878 		tfsid.val[0] = makedev(255,
879 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
880 		mntid_base++;
881 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
882 			break;
883 		vfs_rel(nmp);
884 	}
885 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
886 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
887 	mtx_unlock(&mntid_mtx);
888 }
889 
890 /*
891  * Knob to control the precision of file timestamps:
892  *
893  *   0 = seconds only; nanoseconds zeroed.
894  *   1 = seconds and nanoseconds, accurate within 1/HZ.
895  *   2 = seconds and nanoseconds, truncated to microseconds.
896  * >=3 = seconds and nanoseconds, maximum precision.
897  */
898 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
899 
900 static int timestamp_precision = TSP_USEC;
901 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
902     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
903     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
904     "3+: sec + ns (max. precision))");
905 
906 /*
907  * Get a current timestamp.
908  */
909 void
910 vfs_timestamp(struct timespec *tsp)
911 {
912 	struct timeval tv;
913 
914 	switch (timestamp_precision) {
915 	case TSP_SEC:
916 		tsp->tv_sec = time_second;
917 		tsp->tv_nsec = 0;
918 		break;
919 	case TSP_HZ:
920 		getnanotime(tsp);
921 		break;
922 	case TSP_USEC:
923 		microtime(&tv);
924 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
925 		break;
926 	case TSP_NSEC:
927 	default:
928 		nanotime(tsp);
929 		break;
930 	}
931 }
932 
933 /*
934  * Set vnode attributes to VNOVAL
935  */
936 void
937 vattr_null(struct vattr *vap)
938 {
939 
940 	vap->va_type = VNON;
941 	vap->va_size = VNOVAL;
942 	vap->va_bytes = VNOVAL;
943 	vap->va_mode = VNOVAL;
944 	vap->va_nlink = VNOVAL;
945 	vap->va_uid = VNOVAL;
946 	vap->va_gid = VNOVAL;
947 	vap->va_fsid = VNOVAL;
948 	vap->va_fileid = VNOVAL;
949 	vap->va_blocksize = VNOVAL;
950 	vap->va_rdev = VNOVAL;
951 	vap->va_atime.tv_sec = VNOVAL;
952 	vap->va_atime.tv_nsec = VNOVAL;
953 	vap->va_mtime.tv_sec = VNOVAL;
954 	vap->va_mtime.tv_nsec = VNOVAL;
955 	vap->va_ctime.tv_sec = VNOVAL;
956 	vap->va_ctime.tv_nsec = VNOVAL;
957 	vap->va_birthtime.tv_sec = VNOVAL;
958 	vap->va_birthtime.tv_nsec = VNOVAL;
959 	vap->va_flags = VNOVAL;
960 	vap->va_gen = VNOVAL;
961 	vap->va_vaflags = 0;
962 }
963 
964 /*
965  * This routine is called when we have too many vnodes.  It attempts
966  * to free <count> vnodes and will potentially free vnodes that still
967  * have VM backing store (VM backing store is typically the cause
968  * of a vnode blowout so we want to do this).  Therefore, this operation
969  * is not considered cheap.
970  *
971  * A number of conditions may prevent a vnode from being reclaimed.
972  * the buffer cache may have references on the vnode, a directory
973  * vnode may still have references due to the namei cache representing
974  * underlying files, or the vnode may be in active use.   It is not
975  * desirable to reuse such vnodes.  These conditions may cause the
976  * number of vnodes to reach some minimum value regardless of what
977  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
978  *
979  * @param mp		 Try to reclaim vnodes from this mountpoint
980  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
981  * 			 entries if this argument is strue
982  * @param trigger	 Only reclaim vnodes with fewer than this many resident
983  *			 pages.
984  * @return		 The number of vnodes that were reclaimed.
985  */
986 static int
987 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
988 {
989 	struct vnode *vp;
990 	int count, done, target;
991 
992 	done = 0;
993 	vn_start_write(NULL, &mp, V_WAIT);
994 	MNT_ILOCK(mp);
995 	count = mp->mnt_nvnodelistsize;
996 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
997 	target = target / 10 + 1;
998 	while (count != 0 && done < target) {
999 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1000 		while (vp != NULL && vp->v_type == VMARKER)
1001 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
1002 		if (vp == NULL)
1003 			break;
1004 		/*
1005 		 * XXX LRU is completely broken for non-free vnodes.  First
1006 		 * by calling here in mountpoint order, then by moving
1007 		 * unselected vnodes to the end here, and most grossly by
1008 		 * removing the vlruvp() function that was supposed to
1009 		 * maintain the order.  (This function was born broken
1010 		 * since syncer problems prevented it doing anything.)  The
1011 		 * order is closer to LRC (C = Created).
1012 		 *
1013 		 * LRU reclaiming of vnodes seems to have last worked in
1014 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
1015 		 * Then there was no hold count, and inactive vnodes were
1016 		 * simply put on the free list in LRU order.  The separate
1017 		 * lists also break LRU.  We prefer to reclaim from the
1018 		 * free list for technical reasons.  This tends to thrash
1019 		 * the free list to keep very unrecently used held vnodes.
1020 		 * The problem is mitigated by keeping the free list large.
1021 		 */
1022 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1023 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1024 		--count;
1025 		if (!VI_TRYLOCK(vp))
1026 			goto next_iter;
1027 		/*
1028 		 * If it's been deconstructed already, it's still
1029 		 * referenced, or it exceeds the trigger, skip it.
1030 		 * Also skip free vnodes.  We are trying to make space
1031 		 * to expand the free list, not reduce it.
1032 		 */
1033 		if (vp->v_usecount ||
1034 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1035 		    ((vp->v_iflag & VI_FREE) != 0) ||
1036 		    VN_IS_DOOMED(vp) || (vp->v_object != NULL &&
1037 		    vp->v_object->resident_page_count > trigger)) {
1038 			VI_UNLOCK(vp);
1039 			goto next_iter;
1040 		}
1041 		MNT_IUNLOCK(mp);
1042 		vholdl(vp);
1043 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1044 			vdrop(vp);
1045 			goto next_iter_mntunlocked;
1046 		}
1047 		VI_LOCK(vp);
1048 		/*
1049 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1050 		 * the vnode interlock and before it was locked again.
1051 		 *
1052 		 * It is not necessary to recheck VIRF_DOOMED because it can
1053 		 * only be set by another thread that holds both the vnode
1054 		 * lock and vnode interlock.  If another thread has the
1055 		 * vnode lock before we get to VOP_LOCK() and obtains the
1056 		 * vnode interlock after VOP_LOCK() drops the vnode
1057 		 * interlock, the other thread will be unable to drop the
1058 		 * vnode lock before our VOP_LOCK() call fails.
1059 		 */
1060 		if (vp->v_usecount ||
1061 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1062 		    (vp->v_iflag & VI_FREE) != 0 ||
1063 		    (vp->v_object != NULL &&
1064 		    vp->v_object->resident_page_count > trigger)) {
1065 			VOP_UNLOCK(vp, 0);
1066 			vdropl(vp);
1067 			goto next_iter_mntunlocked;
1068 		}
1069 		KASSERT(!VN_IS_DOOMED(vp),
1070 		    ("VIRF_DOOMED unexpectedly detected in vlrureclaim()"));
1071 		counter_u64_add(recycles_count, 1);
1072 		vgonel(vp);
1073 		VOP_UNLOCK(vp, 0);
1074 		vdropl(vp);
1075 		done++;
1076 next_iter_mntunlocked:
1077 		if (!should_yield())
1078 			goto relock_mnt;
1079 		goto yield;
1080 next_iter:
1081 		if (!should_yield())
1082 			continue;
1083 		MNT_IUNLOCK(mp);
1084 yield:
1085 		kern_yield(PRI_USER);
1086 relock_mnt:
1087 		MNT_ILOCK(mp);
1088 	}
1089 	MNT_IUNLOCK(mp);
1090 	vn_finished_write(mp);
1091 	return done;
1092 }
1093 
1094 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1095 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1096     0,
1097     "limit on vnode free requests per call to the vnlru_free routine");
1098 
1099 /*
1100  * Attempt to reduce the free list by the requested amount.
1101  */
1102 static void
1103 vnlru_free_locked(int count, struct vfsops *mnt_op)
1104 {
1105 	struct vnode *vp;
1106 	struct mount *mp;
1107 	bool tried_batches;
1108 
1109 	tried_batches = false;
1110 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1111 	if (count > max_vnlru_free)
1112 		count = max_vnlru_free;
1113 	for (; count > 0; count--) {
1114 		vp = TAILQ_FIRST(&vnode_free_list);
1115 		/*
1116 		 * The list can be modified while the free_list_mtx
1117 		 * has been dropped and vp could be NULL here.
1118 		 */
1119 		if (vp == NULL) {
1120 			if (tried_batches)
1121 				break;
1122 			mtx_unlock(&vnode_free_list_mtx);
1123 			vnlru_return_batches(mnt_op);
1124 			tried_batches = true;
1125 			mtx_lock(&vnode_free_list_mtx);
1126 			continue;
1127 		}
1128 
1129 		VNASSERT(vp->v_op != NULL, vp,
1130 		    ("vnlru_free: vnode already reclaimed."));
1131 		KASSERT((vp->v_iflag & VI_FREE) != 0,
1132 		    ("Removing vnode not on freelist"));
1133 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1134 		    ("Mangling active vnode"));
1135 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1136 
1137 		/*
1138 		 * Don't recycle if our vnode is from different type
1139 		 * of mount point.  Note that mp is type-safe, the
1140 		 * check does not reach unmapped address even if
1141 		 * vnode is reclaimed.
1142 		 * Don't recycle if we can't get the interlock without
1143 		 * blocking.
1144 		 */
1145 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1146 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1147 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1148 			continue;
1149 		}
1150 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1151 		    vp, ("vp inconsistent on freelist"));
1152 
1153 		/*
1154 		 * The clear of VI_FREE prevents activation of the
1155 		 * vnode.  There is no sense in putting the vnode on
1156 		 * the mount point active list, only to remove it
1157 		 * later during recycling.  Inline the relevant part
1158 		 * of vholdl(), to avoid triggering assertions or
1159 		 * activating.
1160 		 */
1161 		freevnodes--;
1162 		vp->v_iflag &= ~VI_FREE;
1163 		VNODE_REFCOUNT_FENCE_REL();
1164 		refcount_acquire(&vp->v_holdcnt);
1165 
1166 		mtx_unlock(&vnode_free_list_mtx);
1167 		VI_UNLOCK(vp);
1168 		vtryrecycle(vp);
1169 		/*
1170 		 * If the recycled succeeded this vdrop will actually free
1171 		 * the vnode.  If not it will simply place it back on
1172 		 * the free list.
1173 		 */
1174 		vdrop(vp);
1175 		mtx_lock(&vnode_free_list_mtx);
1176 	}
1177 }
1178 
1179 void
1180 vnlru_free(int count, struct vfsops *mnt_op)
1181 {
1182 
1183 	mtx_lock(&vnode_free_list_mtx);
1184 	vnlru_free_locked(count, mnt_op);
1185 	mtx_unlock(&vnode_free_list_mtx);
1186 }
1187 
1188 
1189 /* XXX some names and initialization are bad for limits and watermarks. */
1190 static int
1191 vspace(void)
1192 {
1193 	int space;
1194 
1195 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1196 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1197 	vlowat = vhiwat / 2;
1198 	if (numvnodes > desiredvnodes)
1199 		return (0);
1200 	space = desiredvnodes - numvnodes;
1201 	if (freevnodes > wantfreevnodes)
1202 		space += freevnodes - wantfreevnodes;
1203 	return (space);
1204 }
1205 
1206 static void
1207 vnlru_return_batch_locked(struct mount *mp)
1208 {
1209 	struct vnode *vp;
1210 
1211 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1212 
1213 	if (mp->mnt_tmpfreevnodelistsize == 0)
1214 		return;
1215 
1216 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1217 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1218 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1219 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1220 	}
1221 	mtx_lock(&vnode_free_list_mtx);
1222 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1223 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1224 	mtx_unlock(&vnode_free_list_mtx);
1225 	mp->mnt_tmpfreevnodelistsize = 0;
1226 }
1227 
1228 static void
1229 vnlru_return_batch(struct mount *mp)
1230 {
1231 
1232 	mtx_lock(&mp->mnt_listmtx);
1233 	vnlru_return_batch_locked(mp);
1234 	mtx_unlock(&mp->mnt_listmtx);
1235 }
1236 
1237 static void
1238 vnlru_return_batches(struct vfsops *mnt_op)
1239 {
1240 	struct mount *mp, *nmp;
1241 	bool need_unbusy;
1242 
1243 	mtx_lock(&mountlist_mtx);
1244 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1245 		need_unbusy = false;
1246 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1247 			goto next;
1248 		if (mp->mnt_tmpfreevnodelistsize == 0)
1249 			goto next;
1250 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1251 			vnlru_return_batch(mp);
1252 			need_unbusy = true;
1253 			mtx_lock(&mountlist_mtx);
1254 		}
1255 next:
1256 		nmp = TAILQ_NEXT(mp, mnt_list);
1257 		if (need_unbusy)
1258 			vfs_unbusy(mp);
1259 	}
1260 	mtx_unlock(&mountlist_mtx);
1261 }
1262 
1263 /*
1264  * Attempt to recycle vnodes in a context that is always safe to block.
1265  * Calling vlrurecycle() from the bowels of filesystem code has some
1266  * interesting deadlock problems.
1267  */
1268 static struct proc *vnlruproc;
1269 static int vnlruproc_sig;
1270 
1271 static void
1272 vnlru_proc(void)
1273 {
1274 	struct mount *mp, *nmp;
1275 	unsigned long onumvnodes;
1276 	int done, force, trigger, usevnodes, vsp;
1277 	bool reclaim_nc_src;
1278 
1279 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1280 	    SHUTDOWN_PRI_FIRST);
1281 
1282 	force = 0;
1283 	for (;;) {
1284 		kproc_suspend_check(vnlruproc);
1285 		mtx_lock(&vnode_free_list_mtx);
1286 		/*
1287 		 * If numvnodes is too large (due to desiredvnodes being
1288 		 * adjusted using its sysctl, or emergency growth), first
1289 		 * try to reduce it by discarding from the free list.
1290 		 */
1291 		if (numvnodes > desiredvnodes)
1292 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1293 		/*
1294 		 * Sleep if the vnode cache is in a good state.  This is
1295 		 * when it is not over-full and has space for about a 4%
1296 		 * or 9% expansion (by growing its size or inexcessively
1297 		 * reducing its free list).  Otherwise, try to reclaim
1298 		 * space for a 10% expansion.
1299 		 */
1300 		if (vstir && force == 0) {
1301 			force = 1;
1302 			vstir = 0;
1303 		}
1304 		vsp = vspace();
1305 		if (vsp >= vlowat && force == 0) {
1306 			vnlruproc_sig = 0;
1307 			wakeup(&vnlruproc_sig);
1308 			msleep(vnlruproc, &vnode_free_list_mtx,
1309 			    PVFS|PDROP, "vlruwt", hz);
1310 			continue;
1311 		}
1312 		mtx_unlock(&vnode_free_list_mtx);
1313 		done = 0;
1314 		onumvnodes = numvnodes;
1315 		/*
1316 		 * Calculate parameters for recycling.  These are the same
1317 		 * throughout the loop to give some semblance of fairness.
1318 		 * The trigger point is to avoid recycling vnodes with lots
1319 		 * of resident pages.  We aren't trying to free memory; we
1320 		 * are trying to recycle or at least free vnodes.
1321 		 */
1322 		if (numvnodes <= desiredvnodes)
1323 			usevnodes = numvnodes - freevnodes;
1324 		else
1325 			usevnodes = numvnodes;
1326 		if (usevnodes <= 0)
1327 			usevnodes = 1;
1328 		/*
1329 		 * The trigger value is is chosen to give a conservatively
1330 		 * large value to ensure that it alone doesn't prevent
1331 		 * making progress.  The value can easily be so large that
1332 		 * it is effectively infinite in some congested and
1333 		 * misconfigured cases, and this is necessary.  Normally
1334 		 * it is about 8 to 100 (pages), which is quite large.
1335 		 */
1336 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1337 		if (force < 2)
1338 			trigger = vsmalltrigger;
1339 		reclaim_nc_src = force >= 3;
1340 		mtx_lock(&mountlist_mtx);
1341 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1342 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1343 				nmp = TAILQ_NEXT(mp, mnt_list);
1344 				continue;
1345 			}
1346 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1347 			mtx_lock(&mountlist_mtx);
1348 			nmp = TAILQ_NEXT(mp, mnt_list);
1349 			vfs_unbusy(mp);
1350 		}
1351 		mtx_unlock(&mountlist_mtx);
1352 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1353 			uma_reclaim(UMA_RECLAIM_DRAIN);
1354 		if (done == 0) {
1355 			if (force == 0 || force == 1) {
1356 				force = 2;
1357 				continue;
1358 			}
1359 			if (force == 2) {
1360 				force = 3;
1361 				continue;
1362 			}
1363 			force = 0;
1364 			vnlru_nowhere++;
1365 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1366 		} else
1367 			kern_yield(PRI_USER);
1368 		/*
1369 		 * After becoming active to expand above low water, keep
1370 		 * active until above high water.
1371 		 */
1372 		vsp = vspace();
1373 		force = vsp < vhiwat;
1374 	}
1375 }
1376 
1377 static struct kproc_desc vnlru_kp = {
1378 	"vnlru",
1379 	vnlru_proc,
1380 	&vnlruproc
1381 };
1382 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1383     &vnlru_kp);
1384 
1385 /*
1386  * Routines having to do with the management of the vnode table.
1387  */
1388 
1389 /*
1390  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1391  * before we actually vgone().  This function must be called with the vnode
1392  * held to prevent the vnode from being returned to the free list midway
1393  * through vgone().
1394  */
1395 static int
1396 vtryrecycle(struct vnode *vp)
1397 {
1398 	struct mount *vnmp;
1399 
1400 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1401 	VNASSERT(vp->v_holdcnt, vp,
1402 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1403 	/*
1404 	 * This vnode may found and locked via some other list, if so we
1405 	 * can't recycle it yet.
1406 	 */
1407 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1408 		CTR2(KTR_VFS,
1409 		    "%s: impossible to recycle, vp %p lock is already held",
1410 		    __func__, vp);
1411 		return (EWOULDBLOCK);
1412 	}
1413 	/*
1414 	 * Don't recycle if its filesystem is being suspended.
1415 	 */
1416 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1417 		VOP_UNLOCK(vp, 0);
1418 		CTR2(KTR_VFS,
1419 		    "%s: impossible to recycle, cannot start the write for %p",
1420 		    __func__, vp);
1421 		return (EBUSY);
1422 	}
1423 	/*
1424 	 * If we got this far, we need to acquire the interlock and see if
1425 	 * anyone picked up this vnode from another list.  If not, we will
1426 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1427 	 * will skip over it.
1428 	 */
1429 	VI_LOCK(vp);
1430 	if (vp->v_usecount) {
1431 		VOP_UNLOCK(vp, 0);
1432 		VI_UNLOCK(vp);
1433 		vn_finished_write(vnmp);
1434 		CTR2(KTR_VFS,
1435 		    "%s: impossible to recycle, %p is already referenced",
1436 		    __func__, vp);
1437 		return (EBUSY);
1438 	}
1439 	if (!VN_IS_DOOMED(vp)) {
1440 		counter_u64_add(recycles_count, 1);
1441 		vgonel(vp);
1442 	}
1443 	VOP_UNLOCK(vp, 0);
1444 	VI_UNLOCK(vp);
1445 	vn_finished_write(vnmp);
1446 	return (0);
1447 }
1448 
1449 static void
1450 vcheckspace(void)
1451 {
1452 	int vsp;
1453 
1454 	vsp = vspace();
1455 	if (vsp < vlowat && vnlruproc_sig == 0) {
1456 		vnlruproc_sig = 1;
1457 		wakeup(vnlruproc);
1458 	}
1459 }
1460 
1461 /*
1462  * Wait if necessary for space for a new vnode.
1463  */
1464 static int
1465 getnewvnode_wait(int suspended)
1466 {
1467 
1468 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1469 	if (numvnodes >= desiredvnodes) {
1470 		if (suspended) {
1471 			/*
1472 			 * The file system is being suspended.  We cannot
1473 			 * risk a deadlock here, so allow allocation of
1474 			 * another vnode even if this would give too many.
1475 			 */
1476 			return (0);
1477 		}
1478 		if (vnlruproc_sig == 0) {
1479 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1480 			wakeup(vnlruproc);
1481 		}
1482 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1483 		    "vlruwk", hz);
1484 	}
1485 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1486 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1487 		vnlru_free_locked(1, NULL);
1488 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1489 }
1490 
1491 /*
1492  * This hack is fragile, and probably not needed any more now that the
1493  * watermark handling works.
1494  */
1495 void
1496 getnewvnode_reserve(u_int count)
1497 {
1498 	struct thread *td;
1499 
1500 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1501 	/* XXX no longer so quick, but this part is not racy. */
1502 	mtx_lock(&vnode_free_list_mtx);
1503 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1504 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1505 		    freevnodes - wantfreevnodes), NULL);
1506 	mtx_unlock(&vnode_free_list_mtx);
1507 
1508 	td = curthread;
1509 	/* First try to be quick and racy. */
1510 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1511 		td->td_vp_reserv += count;
1512 		vcheckspace();	/* XXX no longer so quick, but more racy */
1513 		return;
1514 	} else
1515 		atomic_subtract_long(&numvnodes, count);
1516 
1517 	mtx_lock(&vnode_free_list_mtx);
1518 	while (count > 0) {
1519 		if (getnewvnode_wait(0) == 0) {
1520 			count--;
1521 			td->td_vp_reserv++;
1522 			atomic_add_long(&numvnodes, 1);
1523 		}
1524 	}
1525 	vcheckspace();
1526 	mtx_unlock(&vnode_free_list_mtx);
1527 }
1528 
1529 /*
1530  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1531  * misconfgured or changed significantly.  Reducing desiredvnodes below
1532  * the reserved amount should cause bizarre behaviour like reducing it
1533  * below the number of active vnodes -- the system will try to reduce
1534  * numvnodes to match, but should fail, so the subtraction below should
1535  * not overflow.
1536  */
1537 void
1538 getnewvnode_drop_reserve(void)
1539 {
1540 	struct thread *td;
1541 
1542 	td = curthread;
1543 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1544 	td->td_vp_reserv = 0;
1545 }
1546 
1547 /*
1548  * Return the next vnode from the free list.
1549  */
1550 int
1551 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1552     struct vnode **vpp)
1553 {
1554 	struct vnode *vp;
1555 	struct thread *td;
1556 	struct lock_object *lo;
1557 	static int cyclecount;
1558 	int error __unused;
1559 
1560 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1561 	vp = NULL;
1562 	td = curthread;
1563 	if (td->td_vp_reserv > 0) {
1564 		td->td_vp_reserv -= 1;
1565 		goto alloc;
1566 	}
1567 	mtx_lock(&vnode_free_list_mtx);
1568 	if (numvnodes < desiredvnodes)
1569 		cyclecount = 0;
1570 	else if (cyclecount++ >= freevnodes) {
1571 		cyclecount = 0;
1572 		vstir = 1;
1573 	}
1574 	/*
1575 	 * Grow the vnode cache if it will not be above its target max
1576 	 * after growing.  Otherwise, if the free list is nonempty, try
1577 	 * to reclaim 1 item from it before growing the cache (possibly
1578 	 * above its target max if the reclamation failed or is delayed).
1579 	 * Otherwise, wait for some space.  In all cases, schedule
1580 	 * vnlru_proc() if we are getting short of space.  The watermarks
1581 	 * should be chosen so that we never wait or even reclaim from
1582 	 * the free list to below its target minimum.
1583 	 */
1584 	if (numvnodes + 1 <= desiredvnodes)
1585 		;
1586 	else if (freevnodes > 0)
1587 		vnlru_free_locked(1, NULL);
1588 	else {
1589 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1590 		    MNTK_SUSPEND));
1591 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1592 		if (error != 0) {
1593 			mtx_unlock(&vnode_free_list_mtx);
1594 			return (error);
1595 		}
1596 #endif
1597 	}
1598 	vcheckspace();
1599 	atomic_add_long(&numvnodes, 1);
1600 	mtx_unlock(&vnode_free_list_mtx);
1601 alloc:
1602 	counter_u64_add(vnodes_created, 1);
1603 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1604 	/*
1605 	 * Locks are given the generic name "vnode" when created.
1606 	 * Follow the historic practice of using the filesystem
1607 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1608 	 *
1609 	 * Locks live in a witness group keyed on their name. Thus,
1610 	 * when a lock is renamed, it must also move from the witness
1611 	 * group of its old name to the witness group of its new name.
1612 	 *
1613 	 * The change only needs to be made when the vnode moves
1614 	 * from one filesystem type to another. We ensure that each
1615 	 * filesystem use a single static name pointer for its tag so
1616 	 * that we can compare pointers rather than doing a strcmp().
1617 	 */
1618 	lo = &vp->v_vnlock->lock_object;
1619 	if (lo->lo_name != tag) {
1620 		lo->lo_name = tag;
1621 		WITNESS_DESTROY(lo);
1622 		WITNESS_INIT(lo, tag);
1623 	}
1624 	/*
1625 	 * By default, don't allow shared locks unless filesystems opt-in.
1626 	 */
1627 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1628 	/*
1629 	 * Finalize various vnode identity bits.
1630 	 */
1631 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1632 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1633 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1634 	vp->v_type = VNON;
1635 	vp->v_tag = tag;
1636 	vp->v_op = vops;
1637 	v_init_counters(vp);
1638 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1639 #ifdef DIAGNOSTIC
1640 	if (mp == NULL && vops != &dead_vnodeops)
1641 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1642 #endif
1643 #ifdef MAC
1644 	mac_vnode_init(vp);
1645 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1646 		mac_vnode_associate_singlelabel(mp, vp);
1647 #endif
1648 	if (mp != NULL) {
1649 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1650 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1651 			vp->v_vflag |= VV_NOKNOTE;
1652 	}
1653 
1654 	/*
1655 	 * For the filesystems which do not use vfs_hash_insert(),
1656 	 * still initialize v_hash to have vfs_hash_index() useful.
1657 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1658 	 * its own hashing.
1659 	 */
1660 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1661 
1662 	*vpp = vp;
1663 	return (0);
1664 }
1665 
1666 static void
1667 freevnode(struct vnode *vp)
1668 {
1669 	struct bufobj *bo;
1670 
1671 	/*
1672 	 * The vnode has been marked for destruction, so free it.
1673 	 *
1674 	 * The vnode will be returned to the zone where it will
1675 	 * normally remain until it is needed for another vnode. We
1676 	 * need to cleanup (or verify that the cleanup has already
1677 	 * been done) any residual data left from its current use
1678 	 * so as not to contaminate the freshly allocated vnode.
1679 	 */
1680 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1681 	atomic_subtract_long(&numvnodes, 1);
1682 	bo = &vp->v_bufobj;
1683 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
1684 	    ("cleaned vnode still on the free list."));
1685 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1686 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1687 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1688 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1689 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1690 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1691 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1692 	    ("clean blk trie not empty"));
1693 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1694 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1695 	    ("dirty blk trie not empty"));
1696 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1697 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1698 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1699 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1700 	    ("Dangling rangelock waiters"));
1701 	VI_UNLOCK(vp);
1702 #ifdef MAC
1703 	mac_vnode_destroy(vp);
1704 #endif
1705 	if (vp->v_pollinfo != NULL) {
1706 		destroy_vpollinfo(vp->v_pollinfo);
1707 		vp->v_pollinfo = NULL;
1708 	}
1709 #ifdef INVARIANTS
1710 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1711 	vp->v_op = NULL;
1712 #endif
1713 	vp->v_mountedhere = NULL;
1714 	vp->v_unpcb = NULL;
1715 	vp->v_rdev = NULL;
1716 	vp->v_fifoinfo = NULL;
1717 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1718 	vp->v_irflag = 0;
1719 	vp->v_iflag = 0;
1720 	vp->v_vflag = 0;
1721 	bo->bo_flag = 0;
1722 	uma_zfree(vnode_zone, vp);
1723 }
1724 
1725 /*
1726  * Delete from old mount point vnode list, if on one.
1727  */
1728 static void
1729 delmntque(struct vnode *vp)
1730 {
1731 	struct mount *mp;
1732 
1733 	mp = vp->v_mount;
1734 	if (mp == NULL)
1735 		return;
1736 	MNT_ILOCK(mp);
1737 	VI_LOCK(vp);
1738 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1739 	    ("Active vnode list size %d > Vnode list size %d",
1740 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1741 	if (vp->v_iflag & VI_ACTIVE) {
1742 		vp->v_iflag &= ~VI_ACTIVE;
1743 		mtx_lock(&mp->mnt_listmtx);
1744 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1745 		mp->mnt_activevnodelistsize--;
1746 		mtx_unlock(&mp->mnt_listmtx);
1747 	}
1748 	vp->v_mount = NULL;
1749 	VI_UNLOCK(vp);
1750 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1751 		("bad mount point vnode list size"));
1752 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1753 	mp->mnt_nvnodelistsize--;
1754 	MNT_REL(mp);
1755 	MNT_IUNLOCK(mp);
1756 }
1757 
1758 static void
1759 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1760 {
1761 
1762 	vp->v_data = NULL;
1763 	vp->v_op = &dead_vnodeops;
1764 	vgone(vp);
1765 	vput(vp);
1766 }
1767 
1768 /*
1769  * Insert into list of vnodes for the new mount point, if available.
1770  */
1771 int
1772 insmntque1(struct vnode *vp, struct mount *mp,
1773 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1774 {
1775 
1776 	KASSERT(vp->v_mount == NULL,
1777 		("insmntque: vnode already on per mount vnode list"));
1778 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1779 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1780 
1781 	/*
1782 	 * We acquire the vnode interlock early to ensure that the
1783 	 * vnode cannot be recycled by another process releasing a
1784 	 * holdcnt on it before we get it on both the vnode list
1785 	 * and the active vnode list. The mount mutex protects only
1786 	 * manipulation of the vnode list and the vnode freelist
1787 	 * mutex protects only manipulation of the active vnode list.
1788 	 * Hence the need to hold the vnode interlock throughout.
1789 	 */
1790 	MNT_ILOCK(mp);
1791 	VI_LOCK(vp);
1792 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1793 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1794 	    mp->mnt_nvnodelistsize == 0)) &&
1795 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1796 		VI_UNLOCK(vp);
1797 		MNT_IUNLOCK(mp);
1798 		if (dtr != NULL)
1799 			dtr(vp, dtr_arg);
1800 		return (EBUSY);
1801 	}
1802 	vp->v_mount = mp;
1803 	MNT_REF(mp);
1804 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1805 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1806 		("neg mount point vnode list size"));
1807 	mp->mnt_nvnodelistsize++;
1808 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1809 	    ("Activating already active vnode"));
1810 	vp->v_iflag |= VI_ACTIVE;
1811 	mtx_lock(&mp->mnt_listmtx);
1812 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1813 	mp->mnt_activevnodelistsize++;
1814 	mtx_unlock(&mp->mnt_listmtx);
1815 	VI_UNLOCK(vp);
1816 	MNT_IUNLOCK(mp);
1817 	return (0);
1818 }
1819 
1820 int
1821 insmntque(struct vnode *vp, struct mount *mp)
1822 {
1823 
1824 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1825 }
1826 
1827 /*
1828  * Flush out and invalidate all buffers associated with a bufobj
1829  * Called with the underlying object locked.
1830  */
1831 int
1832 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1833 {
1834 	int error;
1835 
1836 	BO_LOCK(bo);
1837 	if (flags & V_SAVE) {
1838 		error = bufobj_wwait(bo, slpflag, slptimeo);
1839 		if (error) {
1840 			BO_UNLOCK(bo);
1841 			return (error);
1842 		}
1843 		if (bo->bo_dirty.bv_cnt > 0) {
1844 			BO_UNLOCK(bo);
1845 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1846 				return (error);
1847 			/*
1848 			 * XXX We could save a lock/unlock if this was only
1849 			 * enabled under INVARIANTS
1850 			 */
1851 			BO_LOCK(bo);
1852 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1853 				panic("vinvalbuf: dirty bufs");
1854 		}
1855 	}
1856 	/*
1857 	 * If you alter this loop please notice that interlock is dropped and
1858 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1859 	 * no race conditions occur from this.
1860 	 */
1861 	do {
1862 		error = flushbuflist(&bo->bo_clean,
1863 		    flags, bo, slpflag, slptimeo);
1864 		if (error == 0 && !(flags & V_CLEANONLY))
1865 			error = flushbuflist(&bo->bo_dirty,
1866 			    flags, bo, slpflag, slptimeo);
1867 		if (error != 0 && error != EAGAIN) {
1868 			BO_UNLOCK(bo);
1869 			return (error);
1870 		}
1871 	} while (error != 0);
1872 
1873 	/*
1874 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1875 	 * have write I/O in-progress but if there is a VM object then the
1876 	 * VM object can also have read-I/O in-progress.
1877 	 */
1878 	do {
1879 		bufobj_wwait(bo, 0, 0);
1880 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1881 			BO_UNLOCK(bo);
1882 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1883 			BO_LOCK(bo);
1884 		}
1885 	} while (bo->bo_numoutput > 0);
1886 	BO_UNLOCK(bo);
1887 
1888 	/*
1889 	 * Destroy the copy in the VM cache, too.
1890 	 */
1891 	if (bo->bo_object != NULL &&
1892 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1893 		VM_OBJECT_WLOCK(bo->bo_object);
1894 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1895 		    OBJPR_CLEANONLY : 0);
1896 		VM_OBJECT_WUNLOCK(bo->bo_object);
1897 	}
1898 
1899 #ifdef INVARIANTS
1900 	BO_LOCK(bo);
1901 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1902 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1903 	    bo->bo_clean.bv_cnt > 0))
1904 		panic("vinvalbuf: flush failed");
1905 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1906 	    bo->bo_dirty.bv_cnt > 0)
1907 		panic("vinvalbuf: flush dirty failed");
1908 	BO_UNLOCK(bo);
1909 #endif
1910 	return (0);
1911 }
1912 
1913 /*
1914  * Flush out and invalidate all buffers associated with a vnode.
1915  * Called with the underlying object locked.
1916  */
1917 int
1918 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1919 {
1920 
1921 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1922 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1923 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1924 		return (0);
1925 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1926 }
1927 
1928 /*
1929  * Flush out buffers on the specified list.
1930  *
1931  */
1932 static int
1933 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1934     int slptimeo)
1935 {
1936 	struct buf *bp, *nbp;
1937 	int retval, error;
1938 	daddr_t lblkno;
1939 	b_xflags_t xflags;
1940 
1941 	ASSERT_BO_WLOCKED(bo);
1942 
1943 	retval = 0;
1944 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1945 		/*
1946 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1947 		 * do not skip any buffers. If we are flushing only V_NORMAL
1948 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1949 		 * flushing only V_ALT buffers then skip buffers not marked
1950 		 * as BX_ALTDATA.
1951 		 */
1952 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1953 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1954 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1955 			continue;
1956 		}
1957 		if (nbp != NULL) {
1958 			lblkno = nbp->b_lblkno;
1959 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1960 		}
1961 		retval = EAGAIN;
1962 		error = BUF_TIMELOCK(bp,
1963 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1964 		    "flushbuf", slpflag, slptimeo);
1965 		if (error) {
1966 			BO_LOCK(bo);
1967 			return (error != ENOLCK ? error : EAGAIN);
1968 		}
1969 		KASSERT(bp->b_bufobj == bo,
1970 		    ("bp %p wrong b_bufobj %p should be %p",
1971 		    bp, bp->b_bufobj, bo));
1972 		/*
1973 		 * XXX Since there are no node locks for NFS, I
1974 		 * believe there is a slight chance that a delayed
1975 		 * write will occur while sleeping just above, so
1976 		 * check for it.
1977 		 */
1978 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1979 		    (flags & V_SAVE)) {
1980 			bremfree(bp);
1981 			bp->b_flags |= B_ASYNC;
1982 			bwrite(bp);
1983 			BO_LOCK(bo);
1984 			return (EAGAIN);	/* XXX: why not loop ? */
1985 		}
1986 		bremfree(bp);
1987 		bp->b_flags |= (B_INVAL | B_RELBUF);
1988 		bp->b_flags &= ~B_ASYNC;
1989 		brelse(bp);
1990 		BO_LOCK(bo);
1991 		if (nbp == NULL)
1992 			break;
1993 		nbp = gbincore(bo, lblkno);
1994 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1995 		    != xflags)
1996 			break;			/* nbp invalid */
1997 	}
1998 	return (retval);
1999 }
2000 
2001 int
2002 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2003 {
2004 	struct buf *bp;
2005 	int error;
2006 	daddr_t lblkno;
2007 
2008 	ASSERT_BO_LOCKED(bo);
2009 
2010 	for (lblkno = startn;;) {
2011 again:
2012 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2013 		if (bp == NULL || bp->b_lblkno >= endn ||
2014 		    bp->b_lblkno < startn)
2015 			break;
2016 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2017 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2018 		if (error != 0) {
2019 			BO_RLOCK(bo);
2020 			if (error == ENOLCK)
2021 				goto again;
2022 			return (error);
2023 		}
2024 		KASSERT(bp->b_bufobj == bo,
2025 		    ("bp %p wrong b_bufobj %p should be %p",
2026 		    bp, bp->b_bufobj, bo));
2027 		lblkno = bp->b_lblkno + 1;
2028 		if ((bp->b_flags & B_MANAGED) == 0)
2029 			bremfree(bp);
2030 		bp->b_flags |= B_RELBUF;
2031 		/*
2032 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2033 		 * pages backing each buffer in the range are unlikely to be
2034 		 * reused.  Dirty buffers will have the hint applied once
2035 		 * they've been written.
2036 		 */
2037 		if ((bp->b_flags & B_VMIO) != 0)
2038 			bp->b_flags |= B_NOREUSE;
2039 		brelse(bp);
2040 		BO_RLOCK(bo);
2041 	}
2042 	return (0);
2043 }
2044 
2045 /*
2046  * Truncate a file's buffer and pages to a specified length.  This
2047  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2048  * sync activity.
2049  */
2050 int
2051 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2052 {
2053 	struct buf *bp, *nbp;
2054 	struct bufobj *bo;
2055 	daddr_t startlbn;
2056 
2057 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2058 	    vp, blksize, (uintmax_t)length);
2059 
2060 	/*
2061 	 * Round up to the *next* lbn.
2062 	 */
2063 	startlbn = howmany(length, blksize);
2064 
2065 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2066 
2067 	bo = &vp->v_bufobj;
2068 restart_unlocked:
2069 	BO_LOCK(bo);
2070 
2071 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2072 		;
2073 
2074 	if (length > 0) {
2075 restartsync:
2076 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2077 			if (bp->b_lblkno > 0)
2078 				continue;
2079 			/*
2080 			 * Since we hold the vnode lock this should only
2081 			 * fail if we're racing with the buf daemon.
2082 			 */
2083 			if (BUF_LOCK(bp,
2084 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2085 			    BO_LOCKPTR(bo)) == ENOLCK)
2086 				goto restart_unlocked;
2087 
2088 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2089 			    ("buf(%p) on dirty queue without DELWRI", bp));
2090 
2091 			bremfree(bp);
2092 			bawrite(bp);
2093 			BO_LOCK(bo);
2094 			goto restartsync;
2095 		}
2096 	}
2097 
2098 	bufobj_wwait(bo, 0, 0);
2099 	BO_UNLOCK(bo);
2100 	vnode_pager_setsize(vp, length);
2101 
2102 	return (0);
2103 }
2104 
2105 /*
2106  * Invalidate the cached pages of a file's buffer within the range of block
2107  * numbers [startlbn, endlbn).
2108  */
2109 void
2110 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2111     int blksize)
2112 {
2113 	struct bufobj *bo;
2114 	off_t start, end;
2115 
2116 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2117 
2118 	start = blksize * startlbn;
2119 	end = blksize * endlbn;
2120 
2121 	bo = &vp->v_bufobj;
2122 	BO_LOCK(bo);
2123 	MPASS(blksize == bo->bo_bsize);
2124 
2125 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2126 		;
2127 
2128 	BO_UNLOCK(bo);
2129 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2130 }
2131 
2132 static int
2133 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2134     daddr_t startlbn, daddr_t endlbn)
2135 {
2136 	struct buf *bp, *nbp;
2137 	bool anyfreed;
2138 
2139 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2140 	ASSERT_BO_LOCKED(bo);
2141 
2142 	do {
2143 		anyfreed = false;
2144 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2145 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2146 				continue;
2147 			if (BUF_LOCK(bp,
2148 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2149 			    BO_LOCKPTR(bo)) == ENOLCK) {
2150 				BO_LOCK(bo);
2151 				return (EAGAIN);
2152 			}
2153 
2154 			bremfree(bp);
2155 			bp->b_flags |= B_INVAL | B_RELBUF;
2156 			bp->b_flags &= ~B_ASYNC;
2157 			brelse(bp);
2158 			anyfreed = true;
2159 
2160 			BO_LOCK(bo);
2161 			if (nbp != NULL &&
2162 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2163 			    nbp->b_vp != vp ||
2164 			    (nbp->b_flags & B_DELWRI) != 0))
2165 				return (EAGAIN);
2166 		}
2167 
2168 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2169 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2170 				continue;
2171 			if (BUF_LOCK(bp,
2172 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2173 			    BO_LOCKPTR(bo)) == ENOLCK) {
2174 				BO_LOCK(bo);
2175 				return (EAGAIN);
2176 			}
2177 			bremfree(bp);
2178 			bp->b_flags |= B_INVAL | B_RELBUF;
2179 			bp->b_flags &= ~B_ASYNC;
2180 			brelse(bp);
2181 			anyfreed = true;
2182 
2183 			BO_LOCK(bo);
2184 			if (nbp != NULL &&
2185 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2186 			    (nbp->b_vp != vp) ||
2187 			    (nbp->b_flags & B_DELWRI) == 0))
2188 				return (EAGAIN);
2189 		}
2190 	} while (anyfreed);
2191 	return (0);
2192 }
2193 
2194 static void
2195 buf_vlist_remove(struct buf *bp)
2196 {
2197 	struct bufv *bv;
2198 
2199 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2200 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2201 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2202 	    (BX_VNDIRTY|BX_VNCLEAN),
2203 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2204 	if (bp->b_xflags & BX_VNDIRTY)
2205 		bv = &bp->b_bufobj->bo_dirty;
2206 	else
2207 		bv = &bp->b_bufobj->bo_clean;
2208 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2209 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2210 	bv->bv_cnt--;
2211 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2212 }
2213 
2214 /*
2215  * Add the buffer to the sorted clean or dirty block list.
2216  *
2217  * NOTE: xflags is passed as a constant, optimizing this inline function!
2218  */
2219 static void
2220 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2221 {
2222 	struct bufv *bv;
2223 	struct buf *n;
2224 	int error;
2225 
2226 	ASSERT_BO_WLOCKED(bo);
2227 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2228 	    ("dead bo %p", bo));
2229 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2230 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2231 	bp->b_xflags |= xflags;
2232 	if (xflags & BX_VNDIRTY)
2233 		bv = &bo->bo_dirty;
2234 	else
2235 		bv = &bo->bo_clean;
2236 
2237 	/*
2238 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2239 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2240 	 * than _ge.
2241 	 */
2242 	if (bv->bv_cnt == 0 ||
2243 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2244 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2245 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2246 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2247 	else
2248 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2249 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2250 	if (error)
2251 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2252 	bv->bv_cnt++;
2253 }
2254 
2255 /*
2256  * Look up a buffer using the buffer tries.
2257  */
2258 struct buf *
2259 gbincore(struct bufobj *bo, daddr_t lblkno)
2260 {
2261 	struct buf *bp;
2262 
2263 	ASSERT_BO_LOCKED(bo);
2264 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2265 	if (bp != NULL)
2266 		return (bp);
2267 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2268 }
2269 
2270 /*
2271  * Associate a buffer with a vnode.
2272  */
2273 void
2274 bgetvp(struct vnode *vp, struct buf *bp)
2275 {
2276 	struct bufobj *bo;
2277 
2278 	bo = &vp->v_bufobj;
2279 	ASSERT_BO_WLOCKED(bo);
2280 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2281 
2282 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2283 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2284 	    ("bgetvp: bp already attached! %p", bp));
2285 
2286 	vhold(vp);
2287 	bp->b_vp = vp;
2288 	bp->b_bufobj = bo;
2289 	/*
2290 	 * Insert onto list for new vnode.
2291 	 */
2292 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2293 }
2294 
2295 /*
2296  * Disassociate a buffer from a vnode.
2297  */
2298 void
2299 brelvp(struct buf *bp)
2300 {
2301 	struct bufobj *bo;
2302 	struct vnode *vp;
2303 
2304 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2305 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2306 
2307 	/*
2308 	 * Delete from old vnode list, if on one.
2309 	 */
2310 	vp = bp->b_vp;		/* XXX */
2311 	bo = bp->b_bufobj;
2312 	BO_LOCK(bo);
2313 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2314 		buf_vlist_remove(bp);
2315 	else
2316 		panic("brelvp: Buffer %p not on queue.", bp);
2317 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2318 		bo->bo_flag &= ~BO_ONWORKLST;
2319 		mtx_lock(&sync_mtx);
2320 		LIST_REMOVE(bo, bo_synclist);
2321 		syncer_worklist_len--;
2322 		mtx_unlock(&sync_mtx);
2323 	}
2324 	bp->b_vp = NULL;
2325 	bp->b_bufobj = NULL;
2326 	BO_UNLOCK(bo);
2327 	vdrop(vp);
2328 }
2329 
2330 /*
2331  * Add an item to the syncer work queue.
2332  */
2333 static void
2334 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2335 {
2336 	int slot;
2337 
2338 	ASSERT_BO_WLOCKED(bo);
2339 
2340 	mtx_lock(&sync_mtx);
2341 	if (bo->bo_flag & BO_ONWORKLST)
2342 		LIST_REMOVE(bo, bo_synclist);
2343 	else {
2344 		bo->bo_flag |= BO_ONWORKLST;
2345 		syncer_worklist_len++;
2346 	}
2347 
2348 	if (delay > syncer_maxdelay - 2)
2349 		delay = syncer_maxdelay - 2;
2350 	slot = (syncer_delayno + delay) & syncer_mask;
2351 
2352 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2353 	mtx_unlock(&sync_mtx);
2354 }
2355 
2356 static int
2357 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2358 {
2359 	int error, len;
2360 
2361 	mtx_lock(&sync_mtx);
2362 	len = syncer_worklist_len - sync_vnode_count;
2363 	mtx_unlock(&sync_mtx);
2364 	error = SYSCTL_OUT(req, &len, sizeof(len));
2365 	return (error);
2366 }
2367 
2368 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2369     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2370 
2371 static struct proc *updateproc;
2372 static void sched_sync(void);
2373 static struct kproc_desc up_kp = {
2374 	"syncer",
2375 	sched_sync,
2376 	&updateproc
2377 };
2378 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2379 
2380 static int
2381 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2382 {
2383 	struct vnode *vp;
2384 	struct mount *mp;
2385 
2386 	*bo = LIST_FIRST(slp);
2387 	if (*bo == NULL)
2388 		return (0);
2389 	vp = bo2vnode(*bo);
2390 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2391 		return (1);
2392 	/*
2393 	 * We use vhold in case the vnode does not
2394 	 * successfully sync.  vhold prevents the vnode from
2395 	 * going away when we unlock the sync_mtx so that
2396 	 * we can acquire the vnode interlock.
2397 	 */
2398 	vholdl(vp);
2399 	mtx_unlock(&sync_mtx);
2400 	VI_UNLOCK(vp);
2401 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2402 		vdrop(vp);
2403 		mtx_lock(&sync_mtx);
2404 		return (*bo == LIST_FIRST(slp));
2405 	}
2406 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2407 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2408 	VOP_UNLOCK(vp, 0);
2409 	vn_finished_write(mp);
2410 	BO_LOCK(*bo);
2411 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2412 		/*
2413 		 * Put us back on the worklist.  The worklist
2414 		 * routine will remove us from our current
2415 		 * position and then add us back in at a later
2416 		 * position.
2417 		 */
2418 		vn_syncer_add_to_worklist(*bo, syncdelay);
2419 	}
2420 	BO_UNLOCK(*bo);
2421 	vdrop(vp);
2422 	mtx_lock(&sync_mtx);
2423 	return (0);
2424 }
2425 
2426 static int first_printf = 1;
2427 
2428 /*
2429  * System filesystem synchronizer daemon.
2430  */
2431 static void
2432 sched_sync(void)
2433 {
2434 	struct synclist *next, *slp;
2435 	struct bufobj *bo;
2436 	long starttime;
2437 	struct thread *td = curthread;
2438 	int last_work_seen;
2439 	int net_worklist_len;
2440 	int syncer_final_iter;
2441 	int error;
2442 
2443 	last_work_seen = 0;
2444 	syncer_final_iter = 0;
2445 	syncer_state = SYNCER_RUNNING;
2446 	starttime = time_uptime;
2447 	td->td_pflags |= TDP_NORUNNINGBUF;
2448 
2449 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2450 	    SHUTDOWN_PRI_LAST);
2451 
2452 	mtx_lock(&sync_mtx);
2453 	for (;;) {
2454 		if (syncer_state == SYNCER_FINAL_DELAY &&
2455 		    syncer_final_iter == 0) {
2456 			mtx_unlock(&sync_mtx);
2457 			kproc_suspend_check(td->td_proc);
2458 			mtx_lock(&sync_mtx);
2459 		}
2460 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2461 		if (syncer_state != SYNCER_RUNNING &&
2462 		    starttime != time_uptime) {
2463 			if (first_printf) {
2464 				printf("\nSyncing disks, vnodes remaining... ");
2465 				first_printf = 0;
2466 			}
2467 			printf("%d ", net_worklist_len);
2468 		}
2469 		starttime = time_uptime;
2470 
2471 		/*
2472 		 * Push files whose dirty time has expired.  Be careful
2473 		 * of interrupt race on slp queue.
2474 		 *
2475 		 * Skip over empty worklist slots when shutting down.
2476 		 */
2477 		do {
2478 			slp = &syncer_workitem_pending[syncer_delayno];
2479 			syncer_delayno += 1;
2480 			if (syncer_delayno == syncer_maxdelay)
2481 				syncer_delayno = 0;
2482 			next = &syncer_workitem_pending[syncer_delayno];
2483 			/*
2484 			 * If the worklist has wrapped since the
2485 			 * it was emptied of all but syncer vnodes,
2486 			 * switch to the FINAL_DELAY state and run
2487 			 * for one more second.
2488 			 */
2489 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2490 			    net_worklist_len == 0 &&
2491 			    last_work_seen == syncer_delayno) {
2492 				syncer_state = SYNCER_FINAL_DELAY;
2493 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2494 			}
2495 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2496 		    syncer_worklist_len > 0);
2497 
2498 		/*
2499 		 * Keep track of the last time there was anything
2500 		 * on the worklist other than syncer vnodes.
2501 		 * Return to the SHUTTING_DOWN state if any
2502 		 * new work appears.
2503 		 */
2504 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2505 			last_work_seen = syncer_delayno;
2506 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2507 			syncer_state = SYNCER_SHUTTING_DOWN;
2508 		while (!LIST_EMPTY(slp)) {
2509 			error = sync_vnode(slp, &bo, td);
2510 			if (error == 1) {
2511 				LIST_REMOVE(bo, bo_synclist);
2512 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2513 				continue;
2514 			}
2515 
2516 			if (first_printf == 0) {
2517 				/*
2518 				 * Drop the sync mutex, because some watchdog
2519 				 * drivers need to sleep while patting
2520 				 */
2521 				mtx_unlock(&sync_mtx);
2522 				wdog_kern_pat(WD_LASTVAL);
2523 				mtx_lock(&sync_mtx);
2524 			}
2525 
2526 		}
2527 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2528 			syncer_final_iter--;
2529 		/*
2530 		 * The variable rushjob allows the kernel to speed up the
2531 		 * processing of the filesystem syncer process. A rushjob
2532 		 * value of N tells the filesystem syncer to process the next
2533 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2534 		 * is used by the soft update code to speed up the filesystem
2535 		 * syncer process when the incore state is getting so far
2536 		 * ahead of the disk that the kernel memory pool is being
2537 		 * threatened with exhaustion.
2538 		 */
2539 		if (rushjob > 0) {
2540 			rushjob -= 1;
2541 			continue;
2542 		}
2543 		/*
2544 		 * Just sleep for a short period of time between
2545 		 * iterations when shutting down to allow some I/O
2546 		 * to happen.
2547 		 *
2548 		 * If it has taken us less than a second to process the
2549 		 * current work, then wait. Otherwise start right over
2550 		 * again. We can still lose time if any single round
2551 		 * takes more than two seconds, but it does not really
2552 		 * matter as we are just trying to generally pace the
2553 		 * filesystem activity.
2554 		 */
2555 		if (syncer_state != SYNCER_RUNNING ||
2556 		    time_uptime == starttime) {
2557 			thread_lock(td);
2558 			sched_prio(td, PPAUSE);
2559 			thread_unlock(td);
2560 		}
2561 		if (syncer_state != SYNCER_RUNNING)
2562 			cv_timedwait(&sync_wakeup, &sync_mtx,
2563 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2564 		else if (time_uptime == starttime)
2565 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2566 	}
2567 }
2568 
2569 /*
2570  * Request the syncer daemon to speed up its work.
2571  * We never push it to speed up more than half of its
2572  * normal turn time, otherwise it could take over the cpu.
2573  */
2574 int
2575 speedup_syncer(void)
2576 {
2577 	int ret = 0;
2578 
2579 	mtx_lock(&sync_mtx);
2580 	if (rushjob < syncdelay / 2) {
2581 		rushjob += 1;
2582 		stat_rush_requests += 1;
2583 		ret = 1;
2584 	}
2585 	mtx_unlock(&sync_mtx);
2586 	cv_broadcast(&sync_wakeup);
2587 	return (ret);
2588 }
2589 
2590 /*
2591  * Tell the syncer to speed up its work and run though its work
2592  * list several times, then tell it to shut down.
2593  */
2594 static void
2595 syncer_shutdown(void *arg, int howto)
2596 {
2597 
2598 	if (howto & RB_NOSYNC)
2599 		return;
2600 	mtx_lock(&sync_mtx);
2601 	syncer_state = SYNCER_SHUTTING_DOWN;
2602 	rushjob = 0;
2603 	mtx_unlock(&sync_mtx);
2604 	cv_broadcast(&sync_wakeup);
2605 	kproc_shutdown(arg, howto);
2606 }
2607 
2608 void
2609 syncer_suspend(void)
2610 {
2611 
2612 	syncer_shutdown(updateproc, 0);
2613 }
2614 
2615 void
2616 syncer_resume(void)
2617 {
2618 
2619 	mtx_lock(&sync_mtx);
2620 	first_printf = 1;
2621 	syncer_state = SYNCER_RUNNING;
2622 	mtx_unlock(&sync_mtx);
2623 	cv_broadcast(&sync_wakeup);
2624 	kproc_resume(updateproc);
2625 }
2626 
2627 /*
2628  * Reassign a buffer from one vnode to another.
2629  * Used to assign file specific control information
2630  * (indirect blocks) to the vnode to which they belong.
2631  */
2632 void
2633 reassignbuf(struct buf *bp)
2634 {
2635 	struct vnode *vp;
2636 	struct bufobj *bo;
2637 	int delay;
2638 #ifdef INVARIANTS
2639 	struct bufv *bv;
2640 #endif
2641 
2642 	vp = bp->b_vp;
2643 	bo = bp->b_bufobj;
2644 	++reassignbufcalls;
2645 
2646 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2647 	    bp, bp->b_vp, bp->b_flags);
2648 	/*
2649 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2650 	 * is not fully linked in.
2651 	 */
2652 	if (bp->b_flags & B_PAGING)
2653 		panic("cannot reassign paging buffer");
2654 
2655 	/*
2656 	 * Delete from old vnode list, if on one.
2657 	 */
2658 	BO_LOCK(bo);
2659 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2660 		buf_vlist_remove(bp);
2661 	else
2662 		panic("reassignbuf: Buffer %p not on queue.", bp);
2663 	/*
2664 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2665 	 * of clean buffers.
2666 	 */
2667 	if (bp->b_flags & B_DELWRI) {
2668 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2669 			switch (vp->v_type) {
2670 			case VDIR:
2671 				delay = dirdelay;
2672 				break;
2673 			case VCHR:
2674 				delay = metadelay;
2675 				break;
2676 			default:
2677 				delay = filedelay;
2678 			}
2679 			vn_syncer_add_to_worklist(bo, delay);
2680 		}
2681 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2682 	} else {
2683 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2684 
2685 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2686 			mtx_lock(&sync_mtx);
2687 			LIST_REMOVE(bo, bo_synclist);
2688 			syncer_worklist_len--;
2689 			mtx_unlock(&sync_mtx);
2690 			bo->bo_flag &= ~BO_ONWORKLST;
2691 		}
2692 	}
2693 #ifdef INVARIANTS
2694 	bv = &bo->bo_clean;
2695 	bp = TAILQ_FIRST(&bv->bv_hd);
2696 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2697 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2698 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2699 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2700 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2701 	bv = &bo->bo_dirty;
2702 	bp = TAILQ_FIRST(&bv->bv_hd);
2703 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2704 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2705 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2706 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2707 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2708 #endif
2709 	BO_UNLOCK(bo);
2710 }
2711 
2712 static void
2713 v_init_counters(struct vnode *vp)
2714 {
2715 
2716 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2717 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2718 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2719 
2720 	refcount_init(&vp->v_holdcnt, 1);
2721 	refcount_init(&vp->v_usecount, 1);
2722 }
2723 
2724 /*
2725  * Increment si_usecount of the associated device, if any.
2726  */
2727 static void
2728 v_incr_devcount(struct vnode *vp)
2729 {
2730 
2731 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2732 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2733 		dev_lock();
2734 		vp->v_rdev->si_usecount++;
2735 		dev_unlock();
2736 	}
2737 }
2738 
2739 /*
2740  * Decrement si_usecount of the associated device, if any.
2741  */
2742 static void
2743 v_decr_devcount(struct vnode *vp)
2744 {
2745 
2746 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2747 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2748 		dev_lock();
2749 		vp->v_rdev->si_usecount--;
2750 		dev_unlock();
2751 	}
2752 }
2753 
2754 /*
2755  * Grab a particular vnode from the free list, increment its
2756  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2757  * is being destroyed.  Only callers who specify LK_RETRY will
2758  * see doomed vnodes.  If inactive processing was delayed in
2759  * vput try to do it here.
2760  *
2761  * Both holdcnt and usecount can be manipulated using atomics without holding
2762  * any locks except in these cases which require the vnode interlock:
2763  * holdcnt: 1->0 and 0->1
2764  * usecount: 0->1
2765  *
2766  * usecount is permitted to transition 1->0 without the interlock because
2767  * vnode is kept live by holdcnt.
2768  */
2769 static enum vgetstate
2770 _vget_prep(struct vnode *vp, bool interlock)
2771 {
2772 	enum vgetstate vs;
2773 
2774 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2775 		vs = VGET_USECOUNT;
2776 	} else {
2777 		_vhold(vp, interlock);
2778 		vs = VGET_HOLDCNT;
2779 	}
2780 	return (vs);
2781 }
2782 
2783 enum vgetstate
2784 vget_prep(struct vnode *vp)
2785 {
2786 
2787 	return (_vget_prep(vp, false));
2788 }
2789 
2790 int
2791 vget(struct vnode *vp, int flags, struct thread *td)
2792 {
2793 	enum vgetstate vs;
2794 
2795 	MPASS(td == curthread);
2796 
2797 	vs = _vget_prep(vp, (flags & LK_INTERLOCK) != 0);
2798 	return (vget_finish(vp, flags, vs));
2799 }
2800 
2801 int
2802 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2803 {
2804 	int error, oweinact;
2805 
2806 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2807 	    ("%s: invalid lock operation", __func__));
2808 
2809 	if ((flags & LK_INTERLOCK) != 0)
2810 		ASSERT_VI_LOCKED(vp, __func__);
2811 	else
2812 		ASSERT_VI_UNLOCKED(vp, __func__);
2813 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2814 	if (vs == VGET_USECOUNT) {
2815 		VNASSERT(vp->v_usecount > 0, vp,
2816 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2817 		    __func__));
2818 	}
2819 
2820 	if ((error = vn_lock(vp, flags)) != 0) {
2821 		if (vs == VGET_USECOUNT)
2822 			vrele(vp);
2823 		else
2824 			vdrop(vp);
2825 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2826 		    vp);
2827 		return (error);
2828 	}
2829 
2830 	if (vs == VGET_USECOUNT) {
2831 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2832 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2833 		return (0);
2834 	}
2835 
2836 	/*
2837 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2838 	 * the vnode around. Otherwise someone else lended their hold count and
2839 	 * we have to drop ours.
2840 	 */
2841 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2842 #ifdef INVARIANTS
2843 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1) - 1;
2844 		VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2845 #else
2846 		refcount_release(&vp->v_holdcnt);
2847 #endif
2848 		VNODE_REFCOUNT_FENCE_ACQ();
2849 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2850 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2851 		return (0);
2852 	}
2853 
2854 	/*
2855 	 * We don't guarantee that any particular close will
2856 	 * trigger inactive processing so just make a best effort
2857 	 * here at preventing a reference to a removed file.  If
2858 	 * we don't succeed no harm is done.
2859 	 *
2860 	 * Upgrade our holdcnt to a usecount.
2861 	 */
2862 	VI_LOCK(vp);
2863 	/*
2864 	 * See the previous section. By the time we get here we may find
2865 	 * ourselves in the same spot.
2866 	 */
2867 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2868 #ifdef INVARIANTS
2869 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1) - 1;
2870 		VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2871 #else
2872 		refcount_release(&vp->v_holdcnt);
2873 #endif
2874 		VNODE_REFCOUNT_FENCE_ACQ();
2875 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2876 		    ("%s: vnode with usecount and VI_OWEINACT set",
2877 		    __func__));
2878 		VI_UNLOCK(vp);
2879 		return (0);
2880 	}
2881 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
2882 		oweinact = 0;
2883 	} else {
2884 		oweinact = 1;
2885 		vp->v_iflag &= ~VI_OWEINACT;
2886 		VNODE_REFCOUNT_FENCE_REL();
2887 	}
2888 	v_incr_devcount(vp);
2889 	refcount_acquire(&vp->v_usecount);
2890 	if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2891 	    (flags & LK_NOWAIT) == 0)
2892 		vinactive(vp, curthread);
2893 	VI_UNLOCK(vp);
2894 	return (0);
2895 }
2896 
2897 /*
2898  * Increase the reference (use) and hold count of a vnode.
2899  * This will also remove the vnode from the free list if it is presently free.
2900  */
2901 void
2902 vref(struct vnode *vp)
2903 {
2904 
2905 	ASSERT_VI_UNLOCKED(vp, __func__);
2906 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2907 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2908 		VNODE_REFCOUNT_FENCE_ACQ();
2909 		VNASSERT(vp->v_holdcnt > 0, vp,
2910 		    ("%s: active vnode not held", __func__));
2911 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2912 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2913 		return;
2914 	}
2915 	VI_LOCK(vp);
2916 	vrefl(vp);
2917 	VI_UNLOCK(vp);
2918 }
2919 
2920 void
2921 vrefl(struct vnode *vp)
2922 {
2923 
2924 	ASSERT_VI_LOCKED(vp, __func__);
2925 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2926 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2927 		VNODE_REFCOUNT_FENCE_ACQ();
2928 		VNASSERT(vp->v_holdcnt > 0, vp,
2929 		    ("%s: active vnode not held", __func__));
2930 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2931 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2932 		return;
2933 	}
2934 	vholdl(vp);
2935 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2936 		vp->v_iflag &= ~VI_OWEINACT;
2937 		VNODE_REFCOUNT_FENCE_REL();
2938 	}
2939 	v_incr_devcount(vp);
2940 	refcount_acquire(&vp->v_usecount);
2941 }
2942 
2943 void
2944 vrefact(struct vnode *vp)
2945 {
2946 
2947 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2948 #ifdef INVARIANTS
2949 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2950 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2951 #else
2952 	refcount_acquire(&vp->v_usecount);
2953 #endif
2954 }
2955 
2956 /*
2957  * Return reference count of a vnode.
2958  *
2959  * The results of this call are only guaranteed when some mechanism is used to
2960  * stop other processes from gaining references to the vnode.  This may be the
2961  * case if the caller holds the only reference.  This is also useful when stale
2962  * data is acceptable as race conditions may be accounted for by some other
2963  * means.
2964  */
2965 int
2966 vrefcnt(struct vnode *vp)
2967 {
2968 
2969 	return (vp->v_usecount);
2970 }
2971 
2972 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
2973 
2974 /*
2975  * Decrement the use and hold counts for a vnode.
2976  *
2977  * See an explanation near vget() as to why atomic operation is safe.
2978  */
2979 static void
2980 vputx(struct vnode *vp, enum vputx_op func)
2981 {
2982 	int error;
2983 
2984 	KASSERT(vp != NULL, ("vputx: null vp"));
2985 	if (func == VPUTX_VUNREF)
2986 		ASSERT_VOP_LOCKED(vp, "vunref");
2987 	ASSERT_VI_UNLOCKED(vp, __func__);
2988 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2989 	    ("%s: wrong ref counts", __func__));
2990 
2991 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2992 
2993 	/*
2994 	 * We want to hold the vnode until the inactive finishes to
2995 	 * prevent vgone() races.  We drop the use count here and the
2996 	 * hold count below when we're done.
2997 	 *
2998 	 * If we release the last usecount we take ownership of the hold
2999 	 * count which provides liveness of the vnode, in which case we
3000 	 * have to vdrop.
3001 	 */
3002 	if (!refcount_release(&vp->v_usecount))
3003 		return;
3004 	VI_LOCK(vp);
3005 	v_decr_devcount(vp);
3006 	/*
3007 	 * By the time we got here someone else might have transitioned
3008 	 * the count back to > 0.
3009 	 */
3010 	if (vp->v_usecount > 0) {
3011 		vdropl(vp);
3012 		return;
3013 	}
3014 	if (vp->v_iflag & VI_DOINGINACT) {
3015 		vdropl(vp);
3016 		return;
3017 	}
3018 
3019 	/*
3020 	 * Check if the fs wants to perform inactive processing. Note we
3021 	 * may be only holding the interlock, in which case it is possible
3022 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3023 	 * Since vgone performs inactive on its own there is nothing to do
3024 	 * here but to drop our hold count.
3025 	 */
3026 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3027 	    VOP_NEED_INACTIVE(vp) == 0) {
3028 		vdropl(vp);
3029 		return;
3030 	}
3031 
3032 	/*
3033 	 * We must call VOP_INACTIVE with the node locked. Mark
3034 	 * as VI_DOINGINACT to avoid recursion.
3035 	 */
3036 	vp->v_iflag |= VI_OWEINACT;
3037 	switch (func) {
3038 	case VPUTX_VRELE:
3039 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3040 		VI_LOCK(vp);
3041 		break;
3042 	case VPUTX_VPUT:
3043 		error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT);
3044 		VI_LOCK(vp);
3045 		break;
3046 	case VPUTX_VUNREF:
3047 		error = 0;
3048 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3049 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3050 			VI_LOCK(vp);
3051 		}
3052 		break;
3053 	}
3054 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
3055 	    ("vnode with usecount and VI_OWEINACT set"));
3056 	if (error == 0) {
3057 		if (vp->v_iflag & VI_OWEINACT)
3058 			vinactive(vp, curthread);
3059 		if (func != VPUTX_VUNREF)
3060 			VOP_UNLOCK(vp, 0);
3061 	}
3062 	vdropl(vp);
3063 }
3064 
3065 /*
3066  * Vnode put/release.
3067  * If count drops to zero, call inactive routine and return to freelist.
3068  */
3069 void
3070 vrele(struct vnode *vp)
3071 {
3072 
3073 	vputx(vp, VPUTX_VRELE);
3074 }
3075 
3076 /*
3077  * Release an already locked vnode.  This give the same effects as
3078  * unlock+vrele(), but takes less time and avoids releasing and
3079  * re-aquiring the lock (as vrele() acquires the lock internally.)
3080  *
3081  * It is an invariant that all VOP_* calls operate on a held vnode.
3082  * We may be only having an implicit hold stemming from our usecount,
3083  * which we are about to release. If we unlock the vnode afterwards we
3084  * open a time window where someone else dropped the last usecount and
3085  * proceeded to free the vnode before our unlock finished. For this
3086  * reason we unlock the vnode early. This is a little bit wasteful as
3087  * it may be the vnode is exclusively locked and inactive processing is
3088  * needed, in which case we are adding work.
3089  */
3090 void
3091 vput(struct vnode *vp)
3092 {
3093 
3094 	VOP_UNLOCK(vp, 0);
3095 	vputx(vp, VPUTX_VPUT);
3096 }
3097 
3098 /*
3099  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3100  */
3101 void
3102 vunref(struct vnode *vp)
3103 {
3104 
3105 	vputx(vp, VPUTX_VUNREF);
3106 }
3107 
3108 /*
3109  * Increase the hold count and activate if this is the first reference.
3110  */
3111 void
3112 _vhold(struct vnode *vp, bool locked)
3113 {
3114 	struct mount *mp;
3115 
3116 	if (locked)
3117 		ASSERT_VI_LOCKED(vp, __func__);
3118 	else
3119 		ASSERT_VI_UNLOCKED(vp, __func__);
3120 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3121 	if (!locked) {
3122 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
3123 			VNODE_REFCOUNT_FENCE_ACQ();
3124 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3125 			    ("_vhold: vnode with holdcnt is free"));
3126 			return;
3127 		}
3128 		VI_LOCK(vp);
3129 	}
3130 	if ((vp->v_iflag & VI_FREE) == 0) {
3131 		refcount_acquire(&vp->v_holdcnt);
3132 		if (!locked)
3133 			VI_UNLOCK(vp);
3134 		return;
3135 	}
3136 	VNASSERT(vp->v_holdcnt == 0, vp,
3137 	    ("%s: wrong hold count", __func__));
3138 	VNASSERT(vp->v_op != NULL, vp,
3139 	    ("%s: vnode already reclaimed.", __func__));
3140 	/*
3141 	 * Remove a vnode from the free list, mark it as in use,
3142 	 * and put it on the active list.
3143 	 */
3144 	VNASSERT(vp->v_mount != NULL, vp,
3145 	    ("_vhold: vnode not on per mount vnode list"));
3146 	mp = vp->v_mount;
3147 	mtx_lock(&mp->mnt_listmtx);
3148 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
3149 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3150 		mp->mnt_tmpfreevnodelistsize--;
3151 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
3152 	} else {
3153 		mtx_lock(&vnode_free_list_mtx);
3154 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3155 		freevnodes--;
3156 		mtx_unlock(&vnode_free_list_mtx);
3157 	}
3158 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3159 	    ("Activating already active vnode"));
3160 	vp->v_iflag &= ~VI_FREE;
3161 	vp->v_iflag |= VI_ACTIVE;
3162 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3163 	mp->mnt_activevnodelistsize++;
3164 	mtx_unlock(&mp->mnt_listmtx);
3165 	refcount_acquire(&vp->v_holdcnt);
3166 	if (!locked)
3167 		VI_UNLOCK(vp);
3168 }
3169 
3170 void
3171 vholdnz(struct vnode *vp)
3172 {
3173 
3174 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3175 #ifdef INVARIANTS
3176 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3177 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
3178 #else
3179 	atomic_add_int(&vp->v_holdcnt, 1);
3180 #endif
3181 }
3182 
3183 /*
3184  * Drop the hold count of the vnode.  If this is the last reference to
3185  * the vnode we place it on the free list unless it has been vgone'd
3186  * (marked VIRF_DOOMED) in which case we will free it.
3187  *
3188  * Because the vnode vm object keeps a hold reference on the vnode if
3189  * there is at least one resident non-cached page, the vnode cannot
3190  * leave the active list without the page cleanup done.
3191  */
3192 void
3193 _vdrop(struct vnode *vp, bool locked)
3194 {
3195 	struct mount *mp;
3196 
3197 	if (locked)
3198 		ASSERT_VI_LOCKED(vp, __func__);
3199 	else
3200 		ASSERT_VI_UNLOCKED(vp, __func__);
3201 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3202 	if (__predict_false((int)vp->v_holdcnt <= 0)) {
3203 		vn_printf(vp, "vdrop: holdcnt %d", vp->v_holdcnt);
3204 		panic("vdrop: wrong holdcnt");
3205 	}
3206 	if (!locked) {
3207 		if (refcount_release_if_not_last(&vp->v_holdcnt))
3208 			return;
3209 		VI_LOCK(vp);
3210 	}
3211 	if (refcount_release(&vp->v_holdcnt) == 0) {
3212 		VI_UNLOCK(vp);
3213 		return;
3214 	}
3215 	if (!VN_IS_DOOMED(vp)) {
3216 		/*
3217 		 * Mark a vnode as free: remove it from its active list
3218 		 * and put it up for recycling on the freelist.
3219 		 */
3220 		VNASSERT(vp->v_op != NULL, vp,
3221 		    ("vdropl: vnode already reclaimed."));
3222 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3223 		    ("vnode already free"));
3224 		VNASSERT(vp->v_holdcnt == 0, vp,
3225 		    ("vdropl: freeing when we shouldn't"));
3226 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
3227 			mp = vp->v_mount;
3228 			if (mp != NULL) {
3229 				mtx_lock(&mp->mnt_listmtx);
3230 				if (vp->v_iflag & VI_ACTIVE) {
3231 					vp->v_iflag &= ~VI_ACTIVE;
3232 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
3233 					    vp, v_actfreelist);
3234 					mp->mnt_activevnodelistsize--;
3235 				}
3236 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3237 				    vp, v_actfreelist);
3238 				mp->mnt_tmpfreevnodelistsize++;
3239 				vp->v_iflag |= VI_FREE;
3240 				vp->v_mflag |= VMP_TMPMNTFREELIST;
3241 				VI_UNLOCK(vp);
3242 				if (mp->mnt_tmpfreevnodelistsize >=
3243 				    mnt_free_list_batch)
3244 					vnlru_return_batch_locked(mp);
3245 				mtx_unlock(&mp->mnt_listmtx);
3246 			} else {
3247 				VNASSERT((vp->v_iflag & VI_ACTIVE) == 0, vp,
3248 				    ("vdropl: active vnode not on per mount "
3249 				    "vnode list"));
3250 				mtx_lock(&vnode_free_list_mtx);
3251 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3252 				    v_actfreelist);
3253 				freevnodes++;
3254 				vp->v_iflag |= VI_FREE;
3255 				VI_UNLOCK(vp);
3256 				mtx_unlock(&vnode_free_list_mtx);
3257 			}
3258 		} else {
3259 			VI_UNLOCK(vp);
3260 			counter_u64_add(free_owe_inact, 1);
3261 		}
3262 		return;
3263 	}
3264 	freevnode(vp);
3265 }
3266 
3267 /*
3268  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3269  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3270  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3271  * failed lock upgrade.
3272  */
3273 void
3274 vinactive(struct vnode *vp, struct thread *td)
3275 {
3276 	struct vm_object *obj;
3277 
3278 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3279 	ASSERT_VI_LOCKED(vp, "vinactive");
3280 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3281 	    ("vinactive: recursed on VI_DOINGINACT"));
3282 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3283 	vp->v_iflag |= VI_DOINGINACT;
3284 	vp->v_iflag &= ~VI_OWEINACT;
3285 	VI_UNLOCK(vp);
3286 	/*
3287 	 * Before moving off the active list, we must be sure that any
3288 	 * modified pages are converted into the vnode's dirty
3289 	 * buffers, since these will no longer be checked once the
3290 	 * vnode is on the inactive list.
3291 	 *
3292 	 * The write-out of the dirty pages is asynchronous.  At the
3293 	 * point that VOP_INACTIVE() is called, there could still be
3294 	 * pending I/O and dirty pages in the object.
3295 	 */
3296 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3297 	    vm_object_mightbedirty(obj)) {
3298 		VM_OBJECT_WLOCK(obj);
3299 		vm_object_page_clean(obj, 0, 0, 0);
3300 		VM_OBJECT_WUNLOCK(obj);
3301 	}
3302 	VOP_INACTIVE(vp, td);
3303 	VI_LOCK(vp);
3304 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3305 	    ("vinactive: lost VI_DOINGINACT"));
3306 	vp->v_iflag &= ~VI_DOINGINACT;
3307 }
3308 
3309 /*
3310  * Remove any vnodes in the vnode table belonging to mount point mp.
3311  *
3312  * If FORCECLOSE is not specified, there should not be any active ones,
3313  * return error if any are found (nb: this is a user error, not a
3314  * system error). If FORCECLOSE is specified, detach any active vnodes
3315  * that are found.
3316  *
3317  * If WRITECLOSE is set, only flush out regular file vnodes open for
3318  * writing.
3319  *
3320  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3321  *
3322  * `rootrefs' specifies the base reference count for the root vnode
3323  * of this filesystem. The root vnode is considered busy if its
3324  * v_usecount exceeds this value. On a successful return, vflush(, td)
3325  * will call vrele() on the root vnode exactly rootrefs times.
3326  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3327  * be zero.
3328  */
3329 #ifdef DIAGNOSTIC
3330 static int busyprt = 0;		/* print out busy vnodes */
3331 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3332 #endif
3333 
3334 int
3335 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3336 {
3337 	struct vnode *vp, *mvp, *rootvp = NULL;
3338 	struct vattr vattr;
3339 	int busy = 0, error;
3340 
3341 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3342 	    rootrefs, flags);
3343 	if (rootrefs > 0) {
3344 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3345 		    ("vflush: bad args"));
3346 		/*
3347 		 * Get the filesystem root vnode. We can vput() it
3348 		 * immediately, since with rootrefs > 0, it won't go away.
3349 		 */
3350 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3351 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3352 			    __func__, error);
3353 			return (error);
3354 		}
3355 		vput(rootvp);
3356 	}
3357 loop:
3358 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3359 		vholdl(vp);
3360 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3361 		if (error) {
3362 			vdrop(vp);
3363 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3364 			goto loop;
3365 		}
3366 		/*
3367 		 * Skip over a vnodes marked VV_SYSTEM.
3368 		 */
3369 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3370 			VOP_UNLOCK(vp, 0);
3371 			vdrop(vp);
3372 			continue;
3373 		}
3374 		/*
3375 		 * If WRITECLOSE is set, flush out unlinked but still open
3376 		 * files (even if open only for reading) and regular file
3377 		 * vnodes open for writing.
3378 		 */
3379 		if (flags & WRITECLOSE) {
3380 			if (vp->v_object != NULL) {
3381 				VM_OBJECT_WLOCK(vp->v_object);
3382 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3383 				VM_OBJECT_WUNLOCK(vp->v_object);
3384 			}
3385 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3386 			if (error != 0) {
3387 				VOP_UNLOCK(vp, 0);
3388 				vdrop(vp);
3389 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3390 				return (error);
3391 			}
3392 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3393 			VI_LOCK(vp);
3394 
3395 			if ((vp->v_type == VNON ||
3396 			    (error == 0 && vattr.va_nlink > 0)) &&
3397 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3398 				VOP_UNLOCK(vp, 0);
3399 				vdropl(vp);
3400 				continue;
3401 			}
3402 		} else
3403 			VI_LOCK(vp);
3404 		/*
3405 		 * With v_usecount == 0, all we need to do is clear out the
3406 		 * vnode data structures and we are done.
3407 		 *
3408 		 * If FORCECLOSE is set, forcibly close the vnode.
3409 		 */
3410 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3411 			vgonel(vp);
3412 		} else {
3413 			busy++;
3414 #ifdef DIAGNOSTIC
3415 			if (busyprt)
3416 				vn_printf(vp, "vflush: busy vnode ");
3417 #endif
3418 		}
3419 		VOP_UNLOCK(vp, 0);
3420 		vdropl(vp);
3421 	}
3422 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3423 		/*
3424 		 * If just the root vnode is busy, and if its refcount
3425 		 * is equal to `rootrefs', then go ahead and kill it.
3426 		 */
3427 		VI_LOCK(rootvp);
3428 		KASSERT(busy > 0, ("vflush: not busy"));
3429 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3430 		    ("vflush: usecount %d < rootrefs %d",
3431 		     rootvp->v_usecount, rootrefs));
3432 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3433 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3434 			vgone(rootvp);
3435 			VOP_UNLOCK(rootvp, 0);
3436 			busy = 0;
3437 		} else
3438 			VI_UNLOCK(rootvp);
3439 	}
3440 	if (busy) {
3441 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3442 		    busy);
3443 		return (EBUSY);
3444 	}
3445 	for (; rootrefs > 0; rootrefs--)
3446 		vrele(rootvp);
3447 	return (0);
3448 }
3449 
3450 /*
3451  * Recycle an unused vnode to the front of the free list.
3452  */
3453 int
3454 vrecycle(struct vnode *vp)
3455 {
3456 	int recycled;
3457 
3458 	VI_LOCK(vp);
3459 	recycled = vrecyclel(vp);
3460 	VI_UNLOCK(vp);
3461 	return (recycled);
3462 }
3463 
3464 /*
3465  * vrecycle, with the vp interlock held.
3466  */
3467 int
3468 vrecyclel(struct vnode *vp)
3469 {
3470 	int recycled;
3471 
3472 	ASSERT_VOP_ELOCKED(vp, __func__);
3473 	ASSERT_VI_LOCKED(vp, __func__);
3474 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3475 	recycled = 0;
3476 	if (vp->v_usecount == 0) {
3477 		recycled = 1;
3478 		vgonel(vp);
3479 	}
3480 	return (recycled);
3481 }
3482 
3483 /*
3484  * Eliminate all activity associated with a vnode
3485  * in preparation for reuse.
3486  */
3487 void
3488 vgone(struct vnode *vp)
3489 {
3490 	VI_LOCK(vp);
3491 	vgonel(vp);
3492 	VI_UNLOCK(vp);
3493 }
3494 
3495 static void
3496 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3497     struct vnode *lowervp __unused)
3498 {
3499 }
3500 
3501 /*
3502  * Notify upper mounts about reclaimed or unlinked vnode.
3503  */
3504 void
3505 vfs_notify_upper(struct vnode *vp, int event)
3506 {
3507 	static struct vfsops vgonel_vfsops = {
3508 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3509 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3510 	};
3511 	struct mount *mp, *ump, *mmp;
3512 
3513 	mp = vp->v_mount;
3514 	if (mp == NULL)
3515 		return;
3516 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3517 		return;
3518 
3519 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3520 	mmp->mnt_op = &vgonel_vfsops;
3521 	mmp->mnt_kern_flag |= MNTK_MARKER;
3522 	MNT_ILOCK(mp);
3523 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3524 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3525 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3526 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3527 			continue;
3528 		}
3529 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3530 		MNT_IUNLOCK(mp);
3531 		switch (event) {
3532 		case VFS_NOTIFY_UPPER_RECLAIM:
3533 			VFS_RECLAIM_LOWERVP(ump, vp);
3534 			break;
3535 		case VFS_NOTIFY_UPPER_UNLINK:
3536 			VFS_UNLINK_LOWERVP(ump, vp);
3537 			break;
3538 		default:
3539 			KASSERT(0, ("invalid event %d", event));
3540 			break;
3541 		}
3542 		MNT_ILOCK(mp);
3543 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3544 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3545 	}
3546 	free(mmp, M_TEMP);
3547 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3548 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3549 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3550 		wakeup(&mp->mnt_uppers);
3551 	}
3552 	MNT_IUNLOCK(mp);
3553 }
3554 
3555 /*
3556  * vgone, with the vp interlock held.
3557  */
3558 static void
3559 vgonel(struct vnode *vp)
3560 {
3561 	struct thread *td;
3562 	struct mount *mp;
3563 	vm_object_t object;
3564 	bool active, oweinact;
3565 
3566 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3567 	ASSERT_VI_LOCKED(vp, "vgonel");
3568 	VNASSERT(vp->v_holdcnt, vp,
3569 	    ("vgonel: vp %p has no reference.", vp));
3570 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3571 	td = curthread;
3572 
3573 	/*
3574 	 * Don't vgonel if we're already doomed.
3575 	 */
3576 	if (vp->v_irflag & VIRF_DOOMED)
3577 		return;
3578 	vp->v_irflag |= VIRF_DOOMED;
3579 
3580 	/*
3581 	 * Check to see if the vnode is in use.  If so, we have to call
3582 	 * VOP_CLOSE() and VOP_INACTIVE().
3583 	 */
3584 	active = vp->v_usecount > 0;
3585 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3586 	VI_UNLOCK(vp);
3587 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3588 
3589 	/*
3590 	 * If purging an active vnode, it must be closed and
3591 	 * deactivated before being reclaimed.
3592 	 */
3593 	if (active)
3594 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3595 	if (oweinact || active) {
3596 		VI_LOCK(vp);
3597 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3598 			vinactive(vp, td);
3599 		VI_UNLOCK(vp);
3600 	}
3601 	if (vp->v_type == VSOCK)
3602 		vfs_unp_reclaim(vp);
3603 
3604 	/*
3605 	 * Clean out any buffers associated with the vnode.
3606 	 * If the flush fails, just toss the buffers.
3607 	 */
3608 	mp = NULL;
3609 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3610 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3611 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3612 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3613 			;
3614 	}
3615 
3616 	BO_LOCK(&vp->v_bufobj);
3617 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3618 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3619 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3620 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3621 	    ("vp %p bufobj not invalidated", vp));
3622 
3623 	/*
3624 	 * For VMIO bufobj, BO_DEAD is set later, or in
3625 	 * vm_object_terminate() after the object's page queue is
3626 	 * flushed.
3627 	 */
3628 	object = vp->v_bufobj.bo_object;
3629 	if (object == NULL)
3630 		vp->v_bufobj.bo_flag |= BO_DEAD;
3631 	BO_UNLOCK(&vp->v_bufobj);
3632 
3633 	/*
3634 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3635 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3636 	 * should not touch the object borrowed from the lower vnode
3637 	 * (the handle check).
3638 	 */
3639 	if (object != NULL && object->type == OBJT_VNODE &&
3640 	    object->handle == vp)
3641 		vnode_destroy_vobject(vp);
3642 
3643 	/*
3644 	 * Reclaim the vnode.
3645 	 */
3646 	if (VOP_RECLAIM(vp, td))
3647 		panic("vgone: cannot reclaim");
3648 	if (mp != NULL)
3649 		vn_finished_secondary_write(mp);
3650 	VNASSERT(vp->v_object == NULL, vp,
3651 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3652 	/*
3653 	 * Clear the advisory locks and wake up waiting threads.
3654 	 */
3655 	(void)VOP_ADVLOCKPURGE(vp);
3656 	vp->v_lockf = NULL;
3657 	/*
3658 	 * Delete from old mount point vnode list.
3659 	 */
3660 	delmntque(vp);
3661 	cache_purge(vp);
3662 	/*
3663 	 * Done with purge, reset to the standard lock and invalidate
3664 	 * the vnode.
3665 	 */
3666 	VI_LOCK(vp);
3667 	vp->v_vnlock = &vp->v_lock;
3668 	vp->v_op = &dead_vnodeops;
3669 	vp->v_tag = "none";
3670 	vp->v_type = VBAD;
3671 }
3672 
3673 /*
3674  * Calculate the total number of references to a special device.
3675  */
3676 int
3677 vcount(struct vnode *vp)
3678 {
3679 	int count;
3680 
3681 	dev_lock();
3682 	count = vp->v_rdev->si_usecount;
3683 	dev_unlock();
3684 	return (count);
3685 }
3686 
3687 /*
3688  * Print out a description of a vnode.
3689  */
3690 static char *typename[] =
3691 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3692  "VMARKER"};
3693 
3694 void
3695 vn_printf(struct vnode *vp, const char *fmt, ...)
3696 {
3697 	va_list ap;
3698 	char buf[256], buf2[16];
3699 	u_long flags;
3700 
3701 	va_start(ap, fmt);
3702 	vprintf(fmt, ap);
3703 	va_end(ap);
3704 	printf("%p: ", (void *)vp);
3705 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3706 	printf("    usecount %d, writecount %d, refcount %d",
3707 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3708 	switch (vp->v_type) {
3709 	case VDIR:
3710 		printf(" mountedhere %p\n", vp->v_mountedhere);
3711 		break;
3712 	case VCHR:
3713 		printf(" rdev %p\n", vp->v_rdev);
3714 		break;
3715 	case VSOCK:
3716 		printf(" socket %p\n", vp->v_unpcb);
3717 		break;
3718 	case VFIFO:
3719 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3720 		break;
3721 	default:
3722 		printf("\n");
3723 		break;
3724 	}
3725 	buf[0] = '\0';
3726 	buf[1] = '\0';
3727 	if (vp->v_irflag & VIRF_DOOMED)
3728 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3729 	flags = vp->v_irflag & ~(VIRF_DOOMED);
3730 	if (flags != 0) {
3731 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3732 		strlcat(buf, buf2, sizeof(buf));
3733 	}
3734 	if (vp->v_vflag & VV_ROOT)
3735 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3736 	if (vp->v_vflag & VV_ISTTY)
3737 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3738 	if (vp->v_vflag & VV_NOSYNC)
3739 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3740 	if (vp->v_vflag & VV_ETERNALDEV)
3741 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3742 	if (vp->v_vflag & VV_CACHEDLABEL)
3743 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3744 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3745 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3746 	if (vp->v_vflag & VV_COPYONWRITE)
3747 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3748 	if (vp->v_vflag & VV_SYSTEM)
3749 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3750 	if (vp->v_vflag & VV_PROCDEP)
3751 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3752 	if (vp->v_vflag & VV_NOKNOTE)
3753 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3754 	if (vp->v_vflag & VV_DELETED)
3755 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3756 	if (vp->v_vflag & VV_MD)
3757 		strlcat(buf, "|VV_MD", sizeof(buf));
3758 	if (vp->v_vflag & VV_FORCEINSMQ)
3759 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3760 	if (vp->v_vflag & VV_READLINK)
3761 		strlcat(buf, "|VV_READLINK", sizeof(buf));
3762 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3763 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3764 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3765 	if (flags != 0) {
3766 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3767 		strlcat(buf, buf2, sizeof(buf));
3768 	}
3769 	if (vp->v_iflag & VI_TEXT_REF)
3770 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
3771 	if (vp->v_iflag & VI_MOUNT)
3772 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3773 	if (vp->v_iflag & VI_FREE)
3774 		strlcat(buf, "|VI_FREE", sizeof(buf));
3775 	if (vp->v_iflag & VI_ACTIVE)
3776 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3777 	if (vp->v_iflag & VI_DOINGINACT)
3778 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3779 	if (vp->v_iflag & VI_OWEINACT)
3780 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3781 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_FREE | VI_ACTIVE |
3782 	    VI_DOINGINACT | VI_OWEINACT);
3783 	if (flags != 0) {
3784 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3785 		strlcat(buf, buf2, sizeof(buf));
3786 	}
3787 	if (vp->v_mflag & VMP_TMPMNTFREELIST)
3788 		strlcat(buf, "|VMP_TMPMNTFREELIST", sizeof(buf));
3789 	flags = vp->v_mflag & ~(VMP_TMPMNTFREELIST);
3790 	if (flags != 0) {
3791 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
3792 		strlcat(buf, buf2, sizeof(buf));
3793 	}
3794 	printf("    flags (%s)\n", buf + 1);
3795 	if (mtx_owned(VI_MTX(vp)))
3796 		printf(" VI_LOCKed");
3797 	if (vp->v_object != NULL)
3798 		printf("    v_object %p ref %d pages %d "
3799 		    "cleanbuf %d dirtybuf %d\n",
3800 		    vp->v_object, vp->v_object->ref_count,
3801 		    vp->v_object->resident_page_count,
3802 		    vp->v_bufobj.bo_clean.bv_cnt,
3803 		    vp->v_bufobj.bo_dirty.bv_cnt);
3804 	printf("    ");
3805 	lockmgr_printinfo(vp->v_vnlock);
3806 	if (vp->v_data != NULL)
3807 		VOP_PRINT(vp);
3808 }
3809 
3810 #ifdef DDB
3811 /*
3812  * List all of the locked vnodes in the system.
3813  * Called when debugging the kernel.
3814  */
3815 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3816 {
3817 	struct mount *mp;
3818 	struct vnode *vp;
3819 
3820 	/*
3821 	 * Note: because this is DDB, we can't obey the locking semantics
3822 	 * for these structures, which means we could catch an inconsistent
3823 	 * state and dereference a nasty pointer.  Not much to be done
3824 	 * about that.
3825 	 */
3826 	db_printf("Locked vnodes\n");
3827 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3828 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3829 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3830 				vn_printf(vp, "vnode ");
3831 		}
3832 	}
3833 }
3834 
3835 /*
3836  * Show details about the given vnode.
3837  */
3838 DB_SHOW_COMMAND(vnode, db_show_vnode)
3839 {
3840 	struct vnode *vp;
3841 
3842 	if (!have_addr)
3843 		return;
3844 	vp = (struct vnode *)addr;
3845 	vn_printf(vp, "vnode ");
3846 }
3847 
3848 /*
3849  * Show details about the given mount point.
3850  */
3851 DB_SHOW_COMMAND(mount, db_show_mount)
3852 {
3853 	struct mount *mp;
3854 	struct vfsopt *opt;
3855 	struct statfs *sp;
3856 	struct vnode *vp;
3857 	char buf[512];
3858 	uint64_t mflags;
3859 	u_int flags;
3860 
3861 	if (!have_addr) {
3862 		/* No address given, print short info about all mount points. */
3863 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3864 			db_printf("%p %s on %s (%s)\n", mp,
3865 			    mp->mnt_stat.f_mntfromname,
3866 			    mp->mnt_stat.f_mntonname,
3867 			    mp->mnt_stat.f_fstypename);
3868 			if (db_pager_quit)
3869 				break;
3870 		}
3871 		db_printf("\nMore info: show mount <addr>\n");
3872 		return;
3873 	}
3874 
3875 	mp = (struct mount *)addr;
3876 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3877 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3878 
3879 	buf[0] = '\0';
3880 	mflags = mp->mnt_flag;
3881 #define	MNT_FLAG(flag)	do {						\
3882 	if (mflags & (flag)) {						\
3883 		if (buf[0] != '\0')					\
3884 			strlcat(buf, ", ", sizeof(buf));		\
3885 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3886 		mflags &= ~(flag);					\
3887 	}								\
3888 } while (0)
3889 	MNT_FLAG(MNT_RDONLY);
3890 	MNT_FLAG(MNT_SYNCHRONOUS);
3891 	MNT_FLAG(MNT_NOEXEC);
3892 	MNT_FLAG(MNT_NOSUID);
3893 	MNT_FLAG(MNT_NFS4ACLS);
3894 	MNT_FLAG(MNT_UNION);
3895 	MNT_FLAG(MNT_ASYNC);
3896 	MNT_FLAG(MNT_SUIDDIR);
3897 	MNT_FLAG(MNT_SOFTDEP);
3898 	MNT_FLAG(MNT_NOSYMFOLLOW);
3899 	MNT_FLAG(MNT_GJOURNAL);
3900 	MNT_FLAG(MNT_MULTILABEL);
3901 	MNT_FLAG(MNT_ACLS);
3902 	MNT_FLAG(MNT_NOATIME);
3903 	MNT_FLAG(MNT_NOCLUSTERR);
3904 	MNT_FLAG(MNT_NOCLUSTERW);
3905 	MNT_FLAG(MNT_SUJ);
3906 	MNT_FLAG(MNT_EXRDONLY);
3907 	MNT_FLAG(MNT_EXPORTED);
3908 	MNT_FLAG(MNT_DEFEXPORTED);
3909 	MNT_FLAG(MNT_EXPORTANON);
3910 	MNT_FLAG(MNT_EXKERB);
3911 	MNT_FLAG(MNT_EXPUBLIC);
3912 	MNT_FLAG(MNT_LOCAL);
3913 	MNT_FLAG(MNT_QUOTA);
3914 	MNT_FLAG(MNT_ROOTFS);
3915 	MNT_FLAG(MNT_USER);
3916 	MNT_FLAG(MNT_IGNORE);
3917 	MNT_FLAG(MNT_UPDATE);
3918 	MNT_FLAG(MNT_DELEXPORT);
3919 	MNT_FLAG(MNT_RELOAD);
3920 	MNT_FLAG(MNT_FORCE);
3921 	MNT_FLAG(MNT_SNAPSHOT);
3922 	MNT_FLAG(MNT_BYFSID);
3923 #undef MNT_FLAG
3924 	if (mflags != 0) {
3925 		if (buf[0] != '\0')
3926 			strlcat(buf, ", ", sizeof(buf));
3927 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3928 		    "0x%016jx", mflags);
3929 	}
3930 	db_printf("    mnt_flag = %s\n", buf);
3931 
3932 	buf[0] = '\0';
3933 	flags = mp->mnt_kern_flag;
3934 #define	MNT_KERN_FLAG(flag)	do {					\
3935 	if (flags & (flag)) {						\
3936 		if (buf[0] != '\0')					\
3937 			strlcat(buf, ", ", sizeof(buf));		\
3938 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3939 		flags &= ~(flag);					\
3940 	}								\
3941 } while (0)
3942 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3943 	MNT_KERN_FLAG(MNTK_ASYNC);
3944 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3945 	MNT_KERN_FLAG(MNTK_DRAINING);
3946 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3947 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3948 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3949 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3950 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3951 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3952 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3953 	MNT_KERN_FLAG(MNTK_MARKER);
3954 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3955 	MNT_KERN_FLAG(MNTK_NOASYNC);
3956 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3957 	MNT_KERN_FLAG(MNTK_MWAIT);
3958 	MNT_KERN_FLAG(MNTK_SUSPEND);
3959 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3960 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3961 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3962 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3963 #undef MNT_KERN_FLAG
3964 	if (flags != 0) {
3965 		if (buf[0] != '\0')
3966 			strlcat(buf, ", ", sizeof(buf));
3967 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3968 		    "0x%08x", flags);
3969 	}
3970 	db_printf("    mnt_kern_flag = %s\n", buf);
3971 
3972 	db_printf("    mnt_opt = ");
3973 	opt = TAILQ_FIRST(mp->mnt_opt);
3974 	if (opt != NULL) {
3975 		db_printf("%s", opt->name);
3976 		opt = TAILQ_NEXT(opt, link);
3977 		while (opt != NULL) {
3978 			db_printf(", %s", opt->name);
3979 			opt = TAILQ_NEXT(opt, link);
3980 		}
3981 	}
3982 	db_printf("\n");
3983 
3984 	sp = &mp->mnt_stat;
3985 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3986 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3987 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3988 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3989 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3990 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3991 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3992 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3993 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3994 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3995 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3996 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3997 
3998 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3999 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4000 	if (jailed(mp->mnt_cred))
4001 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4002 	db_printf(" }\n");
4003 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4004 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4005 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4006 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4007 	db_printf("    mnt_activevnodelistsize = %d\n",
4008 	    mp->mnt_activevnodelistsize);
4009 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4010 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4011 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4012 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4013 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4014 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4015 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4016 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4017 	db_printf("    mnt_secondary_accwrites = %d\n",
4018 	    mp->mnt_secondary_accwrites);
4019 	db_printf("    mnt_gjprovider = %s\n",
4020 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4021 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4022 
4023 	db_printf("\n\nList of active vnodes\n");
4024 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
4025 		if (vp->v_type != VMARKER) {
4026 			vn_printf(vp, "vnode ");
4027 			if (db_pager_quit)
4028 				break;
4029 		}
4030 	}
4031 	db_printf("\n\nList of inactive vnodes\n");
4032 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4033 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
4034 			vn_printf(vp, "vnode ");
4035 			if (db_pager_quit)
4036 				break;
4037 		}
4038 	}
4039 }
4040 #endif	/* DDB */
4041 
4042 /*
4043  * Fill in a struct xvfsconf based on a struct vfsconf.
4044  */
4045 static int
4046 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4047 {
4048 	struct xvfsconf xvfsp;
4049 
4050 	bzero(&xvfsp, sizeof(xvfsp));
4051 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4052 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4053 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4054 	xvfsp.vfc_flags = vfsp->vfc_flags;
4055 	/*
4056 	 * These are unused in userland, we keep them
4057 	 * to not break binary compatibility.
4058 	 */
4059 	xvfsp.vfc_vfsops = NULL;
4060 	xvfsp.vfc_next = NULL;
4061 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4062 }
4063 
4064 #ifdef COMPAT_FREEBSD32
4065 struct xvfsconf32 {
4066 	uint32_t	vfc_vfsops;
4067 	char		vfc_name[MFSNAMELEN];
4068 	int32_t		vfc_typenum;
4069 	int32_t		vfc_refcount;
4070 	int32_t		vfc_flags;
4071 	uint32_t	vfc_next;
4072 };
4073 
4074 static int
4075 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4076 {
4077 	struct xvfsconf32 xvfsp;
4078 
4079 	bzero(&xvfsp, sizeof(xvfsp));
4080 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4081 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4082 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4083 	xvfsp.vfc_flags = vfsp->vfc_flags;
4084 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4085 }
4086 #endif
4087 
4088 /*
4089  * Top level filesystem related information gathering.
4090  */
4091 static int
4092 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4093 {
4094 	struct vfsconf *vfsp;
4095 	int error;
4096 
4097 	error = 0;
4098 	vfsconf_slock();
4099 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4100 #ifdef COMPAT_FREEBSD32
4101 		if (req->flags & SCTL_MASK32)
4102 			error = vfsconf2x32(req, vfsp);
4103 		else
4104 #endif
4105 			error = vfsconf2x(req, vfsp);
4106 		if (error)
4107 			break;
4108 	}
4109 	vfsconf_sunlock();
4110 	return (error);
4111 }
4112 
4113 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4114     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4115     "S,xvfsconf", "List of all configured filesystems");
4116 
4117 #ifndef BURN_BRIDGES
4118 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4119 
4120 static int
4121 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4122 {
4123 	int *name = (int *)arg1 - 1;	/* XXX */
4124 	u_int namelen = arg2 + 1;	/* XXX */
4125 	struct vfsconf *vfsp;
4126 
4127 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4128 	    "please rebuild world\n");
4129 
4130 #if 1 || defined(COMPAT_PRELITE2)
4131 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4132 	if (namelen == 1)
4133 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4134 #endif
4135 
4136 	switch (name[1]) {
4137 	case VFS_MAXTYPENUM:
4138 		if (namelen != 2)
4139 			return (ENOTDIR);
4140 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4141 	case VFS_CONF:
4142 		if (namelen != 3)
4143 			return (ENOTDIR);	/* overloaded */
4144 		vfsconf_slock();
4145 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4146 			if (vfsp->vfc_typenum == name[2])
4147 				break;
4148 		}
4149 		vfsconf_sunlock();
4150 		if (vfsp == NULL)
4151 			return (EOPNOTSUPP);
4152 #ifdef COMPAT_FREEBSD32
4153 		if (req->flags & SCTL_MASK32)
4154 			return (vfsconf2x32(req, vfsp));
4155 		else
4156 #endif
4157 			return (vfsconf2x(req, vfsp));
4158 	}
4159 	return (EOPNOTSUPP);
4160 }
4161 
4162 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4163     CTLFLAG_MPSAFE, vfs_sysctl,
4164     "Generic filesystem");
4165 
4166 #if 1 || defined(COMPAT_PRELITE2)
4167 
4168 static int
4169 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4170 {
4171 	int error;
4172 	struct vfsconf *vfsp;
4173 	struct ovfsconf ovfs;
4174 
4175 	vfsconf_slock();
4176 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4177 		bzero(&ovfs, sizeof(ovfs));
4178 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4179 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4180 		ovfs.vfc_index = vfsp->vfc_typenum;
4181 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4182 		ovfs.vfc_flags = vfsp->vfc_flags;
4183 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4184 		if (error != 0) {
4185 			vfsconf_sunlock();
4186 			return (error);
4187 		}
4188 	}
4189 	vfsconf_sunlock();
4190 	return (0);
4191 }
4192 
4193 #endif /* 1 || COMPAT_PRELITE2 */
4194 #endif /* !BURN_BRIDGES */
4195 
4196 #define KINFO_VNODESLOP		10
4197 #ifdef notyet
4198 /*
4199  * Dump vnode list (via sysctl).
4200  */
4201 /* ARGSUSED */
4202 static int
4203 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4204 {
4205 	struct xvnode *xvn;
4206 	struct mount *mp;
4207 	struct vnode *vp;
4208 	int error, len, n;
4209 
4210 	/*
4211 	 * Stale numvnodes access is not fatal here.
4212 	 */
4213 	req->lock = 0;
4214 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4215 	if (!req->oldptr)
4216 		/* Make an estimate */
4217 		return (SYSCTL_OUT(req, 0, len));
4218 
4219 	error = sysctl_wire_old_buffer(req, 0);
4220 	if (error != 0)
4221 		return (error);
4222 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4223 	n = 0;
4224 	mtx_lock(&mountlist_mtx);
4225 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4226 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4227 			continue;
4228 		MNT_ILOCK(mp);
4229 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4230 			if (n == len)
4231 				break;
4232 			vref(vp);
4233 			xvn[n].xv_size = sizeof *xvn;
4234 			xvn[n].xv_vnode = vp;
4235 			xvn[n].xv_id = 0;	/* XXX compat */
4236 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4237 			XV_COPY(usecount);
4238 			XV_COPY(writecount);
4239 			XV_COPY(holdcnt);
4240 			XV_COPY(mount);
4241 			XV_COPY(numoutput);
4242 			XV_COPY(type);
4243 #undef XV_COPY
4244 			xvn[n].xv_flag = vp->v_vflag;
4245 
4246 			switch (vp->v_type) {
4247 			case VREG:
4248 			case VDIR:
4249 			case VLNK:
4250 				break;
4251 			case VBLK:
4252 			case VCHR:
4253 				if (vp->v_rdev == NULL) {
4254 					vrele(vp);
4255 					continue;
4256 				}
4257 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4258 				break;
4259 			case VSOCK:
4260 				xvn[n].xv_socket = vp->v_socket;
4261 				break;
4262 			case VFIFO:
4263 				xvn[n].xv_fifo = vp->v_fifoinfo;
4264 				break;
4265 			case VNON:
4266 			case VBAD:
4267 			default:
4268 				/* shouldn't happen? */
4269 				vrele(vp);
4270 				continue;
4271 			}
4272 			vrele(vp);
4273 			++n;
4274 		}
4275 		MNT_IUNLOCK(mp);
4276 		mtx_lock(&mountlist_mtx);
4277 		vfs_unbusy(mp);
4278 		if (n == len)
4279 			break;
4280 	}
4281 	mtx_unlock(&mountlist_mtx);
4282 
4283 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4284 	free(xvn, M_TEMP);
4285 	return (error);
4286 }
4287 
4288 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4289     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4290     "");
4291 #endif
4292 
4293 static void
4294 unmount_or_warn(struct mount *mp)
4295 {
4296 	int error;
4297 
4298 	error = dounmount(mp, MNT_FORCE, curthread);
4299 	if (error != 0) {
4300 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4301 		if (error == EBUSY)
4302 			printf("BUSY)\n");
4303 		else
4304 			printf("%d)\n", error);
4305 	}
4306 }
4307 
4308 /*
4309  * Unmount all filesystems. The list is traversed in reverse order
4310  * of mounting to avoid dependencies.
4311  */
4312 void
4313 vfs_unmountall(void)
4314 {
4315 	struct mount *mp, *tmp;
4316 
4317 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4318 
4319 	/*
4320 	 * Since this only runs when rebooting, it is not interlocked.
4321 	 */
4322 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4323 		vfs_ref(mp);
4324 
4325 		/*
4326 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4327 		 * unmount of the latter.
4328 		 */
4329 		if (mp == rootdevmp)
4330 			continue;
4331 
4332 		unmount_or_warn(mp);
4333 	}
4334 
4335 	if (rootdevmp != NULL)
4336 		unmount_or_warn(rootdevmp);
4337 }
4338 
4339 /*
4340  * perform msync on all vnodes under a mount point
4341  * the mount point must be locked.
4342  */
4343 void
4344 vfs_msync(struct mount *mp, int flags)
4345 {
4346 	struct vnode *vp, *mvp;
4347 	struct vm_object *obj;
4348 
4349 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4350 
4351 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4352 		return;
4353 
4354 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4355 		obj = vp->v_object;
4356 		if (obj != NULL && vm_object_mightbedirty(obj) &&
4357 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4358 			if (!vget(vp,
4359 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4360 			    curthread)) {
4361 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4362 					vput(vp);
4363 					continue;
4364 				}
4365 
4366 				obj = vp->v_object;
4367 				if (obj != NULL) {
4368 					VM_OBJECT_WLOCK(obj);
4369 					vm_object_page_clean(obj, 0, 0,
4370 					    flags == MNT_WAIT ?
4371 					    OBJPC_SYNC : OBJPC_NOSYNC);
4372 					VM_OBJECT_WUNLOCK(obj);
4373 				}
4374 				vput(vp);
4375 			}
4376 		} else
4377 			VI_UNLOCK(vp);
4378 	}
4379 }
4380 
4381 static void
4382 destroy_vpollinfo_free(struct vpollinfo *vi)
4383 {
4384 
4385 	knlist_destroy(&vi->vpi_selinfo.si_note);
4386 	mtx_destroy(&vi->vpi_lock);
4387 	uma_zfree(vnodepoll_zone, vi);
4388 }
4389 
4390 static void
4391 destroy_vpollinfo(struct vpollinfo *vi)
4392 {
4393 
4394 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4395 	seldrain(&vi->vpi_selinfo);
4396 	destroy_vpollinfo_free(vi);
4397 }
4398 
4399 /*
4400  * Initialize per-vnode helper structure to hold poll-related state.
4401  */
4402 void
4403 v_addpollinfo(struct vnode *vp)
4404 {
4405 	struct vpollinfo *vi;
4406 
4407 	if (vp->v_pollinfo != NULL)
4408 		return;
4409 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4410 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4411 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4412 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4413 	VI_LOCK(vp);
4414 	if (vp->v_pollinfo != NULL) {
4415 		VI_UNLOCK(vp);
4416 		destroy_vpollinfo_free(vi);
4417 		return;
4418 	}
4419 	vp->v_pollinfo = vi;
4420 	VI_UNLOCK(vp);
4421 }
4422 
4423 /*
4424  * Record a process's interest in events which might happen to
4425  * a vnode.  Because poll uses the historic select-style interface
4426  * internally, this routine serves as both the ``check for any
4427  * pending events'' and the ``record my interest in future events''
4428  * functions.  (These are done together, while the lock is held,
4429  * to avoid race conditions.)
4430  */
4431 int
4432 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4433 {
4434 
4435 	v_addpollinfo(vp);
4436 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4437 	if (vp->v_pollinfo->vpi_revents & events) {
4438 		/*
4439 		 * This leaves events we are not interested
4440 		 * in available for the other process which
4441 		 * which presumably had requested them
4442 		 * (otherwise they would never have been
4443 		 * recorded).
4444 		 */
4445 		events &= vp->v_pollinfo->vpi_revents;
4446 		vp->v_pollinfo->vpi_revents &= ~events;
4447 
4448 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4449 		return (events);
4450 	}
4451 	vp->v_pollinfo->vpi_events |= events;
4452 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4453 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4454 	return (0);
4455 }
4456 
4457 /*
4458  * Routine to create and manage a filesystem syncer vnode.
4459  */
4460 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4461 static int	sync_fsync(struct  vop_fsync_args *);
4462 static int	sync_inactive(struct  vop_inactive_args *);
4463 static int	sync_reclaim(struct  vop_reclaim_args *);
4464 
4465 static struct vop_vector sync_vnodeops = {
4466 	.vop_bypass =	VOP_EOPNOTSUPP,
4467 	.vop_close =	sync_close,		/* close */
4468 	.vop_fsync =	sync_fsync,		/* fsync */
4469 	.vop_inactive =	sync_inactive,	/* inactive */
4470 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4471 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4472 	.vop_lock1 =	vop_stdlock,	/* lock */
4473 	.vop_unlock =	vop_stdunlock,	/* unlock */
4474 	.vop_islocked =	vop_stdislocked,	/* islocked */
4475 };
4476 
4477 /*
4478  * Create a new filesystem syncer vnode for the specified mount point.
4479  */
4480 void
4481 vfs_allocate_syncvnode(struct mount *mp)
4482 {
4483 	struct vnode *vp;
4484 	struct bufobj *bo;
4485 	static long start, incr, next;
4486 	int error;
4487 
4488 	/* Allocate a new vnode */
4489 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4490 	if (error != 0)
4491 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4492 	vp->v_type = VNON;
4493 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4494 	vp->v_vflag |= VV_FORCEINSMQ;
4495 	error = insmntque(vp, mp);
4496 	if (error != 0)
4497 		panic("vfs_allocate_syncvnode: insmntque() failed");
4498 	vp->v_vflag &= ~VV_FORCEINSMQ;
4499 	VOP_UNLOCK(vp, 0);
4500 	/*
4501 	 * Place the vnode onto the syncer worklist. We attempt to
4502 	 * scatter them about on the list so that they will go off
4503 	 * at evenly distributed times even if all the filesystems
4504 	 * are mounted at once.
4505 	 */
4506 	next += incr;
4507 	if (next == 0 || next > syncer_maxdelay) {
4508 		start /= 2;
4509 		incr /= 2;
4510 		if (start == 0) {
4511 			start = syncer_maxdelay / 2;
4512 			incr = syncer_maxdelay;
4513 		}
4514 		next = start;
4515 	}
4516 	bo = &vp->v_bufobj;
4517 	BO_LOCK(bo);
4518 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4519 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4520 	mtx_lock(&sync_mtx);
4521 	sync_vnode_count++;
4522 	if (mp->mnt_syncer == NULL) {
4523 		mp->mnt_syncer = vp;
4524 		vp = NULL;
4525 	}
4526 	mtx_unlock(&sync_mtx);
4527 	BO_UNLOCK(bo);
4528 	if (vp != NULL) {
4529 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4530 		vgone(vp);
4531 		vput(vp);
4532 	}
4533 }
4534 
4535 void
4536 vfs_deallocate_syncvnode(struct mount *mp)
4537 {
4538 	struct vnode *vp;
4539 
4540 	mtx_lock(&sync_mtx);
4541 	vp = mp->mnt_syncer;
4542 	if (vp != NULL)
4543 		mp->mnt_syncer = NULL;
4544 	mtx_unlock(&sync_mtx);
4545 	if (vp != NULL)
4546 		vrele(vp);
4547 }
4548 
4549 /*
4550  * Do a lazy sync of the filesystem.
4551  */
4552 static int
4553 sync_fsync(struct vop_fsync_args *ap)
4554 {
4555 	struct vnode *syncvp = ap->a_vp;
4556 	struct mount *mp = syncvp->v_mount;
4557 	int error, save;
4558 	struct bufobj *bo;
4559 
4560 	/*
4561 	 * We only need to do something if this is a lazy evaluation.
4562 	 */
4563 	if (ap->a_waitfor != MNT_LAZY)
4564 		return (0);
4565 
4566 	/*
4567 	 * Move ourselves to the back of the sync list.
4568 	 */
4569 	bo = &syncvp->v_bufobj;
4570 	BO_LOCK(bo);
4571 	vn_syncer_add_to_worklist(bo, syncdelay);
4572 	BO_UNLOCK(bo);
4573 
4574 	/*
4575 	 * Walk the list of vnodes pushing all that are dirty and
4576 	 * not already on the sync list.
4577 	 */
4578 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4579 		return (0);
4580 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4581 		vfs_unbusy(mp);
4582 		return (0);
4583 	}
4584 	save = curthread_pflags_set(TDP_SYNCIO);
4585 	/*
4586 	 * The filesystem at hand may be idle with free vnodes stored in the
4587 	 * batch.  Return them instead of letting them stay there indefinitely.
4588 	 */
4589 	vnlru_return_batch(mp);
4590 	vfs_msync(mp, MNT_NOWAIT);
4591 	error = VFS_SYNC(mp, MNT_LAZY);
4592 	curthread_pflags_restore(save);
4593 	vn_finished_write(mp);
4594 	vfs_unbusy(mp);
4595 	return (error);
4596 }
4597 
4598 /*
4599  * The syncer vnode is no referenced.
4600  */
4601 static int
4602 sync_inactive(struct vop_inactive_args *ap)
4603 {
4604 
4605 	vgone(ap->a_vp);
4606 	return (0);
4607 }
4608 
4609 /*
4610  * The syncer vnode is no longer needed and is being decommissioned.
4611  *
4612  * Modifications to the worklist must be protected by sync_mtx.
4613  */
4614 static int
4615 sync_reclaim(struct vop_reclaim_args *ap)
4616 {
4617 	struct vnode *vp = ap->a_vp;
4618 	struct bufobj *bo;
4619 
4620 	bo = &vp->v_bufobj;
4621 	BO_LOCK(bo);
4622 	mtx_lock(&sync_mtx);
4623 	if (vp->v_mount->mnt_syncer == vp)
4624 		vp->v_mount->mnt_syncer = NULL;
4625 	if (bo->bo_flag & BO_ONWORKLST) {
4626 		LIST_REMOVE(bo, bo_synclist);
4627 		syncer_worklist_len--;
4628 		sync_vnode_count--;
4629 		bo->bo_flag &= ~BO_ONWORKLST;
4630 	}
4631 	mtx_unlock(&sync_mtx);
4632 	BO_UNLOCK(bo);
4633 
4634 	return (0);
4635 }
4636 
4637 int
4638 vn_need_pageq_flush(struct vnode *vp)
4639 {
4640 	struct vm_object *obj;
4641 	int need;
4642 
4643 	MPASS(mtx_owned(VI_MTX(vp)));
4644 	need = 0;
4645 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4646 	    vm_object_mightbedirty(obj))
4647 		need = 1;
4648 	return (need);
4649 }
4650 
4651 /*
4652  * Check if vnode represents a disk device
4653  */
4654 int
4655 vn_isdisk(struct vnode *vp, int *errp)
4656 {
4657 	int error;
4658 
4659 	if (vp->v_type != VCHR) {
4660 		error = ENOTBLK;
4661 		goto out;
4662 	}
4663 	error = 0;
4664 	dev_lock();
4665 	if (vp->v_rdev == NULL)
4666 		error = ENXIO;
4667 	else if (vp->v_rdev->si_devsw == NULL)
4668 		error = ENXIO;
4669 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4670 		error = ENOTBLK;
4671 	dev_unlock();
4672 out:
4673 	if (errp != NULL)
4674 		*errp = error;
4675 	return (error == 0);
4676 }
4677 
4678 /*
4679  * Common filesystem object access control check routine.  Accepts a
4680  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4681  * and optional call-by-reference privused argument allowing vaccess()
4682  * to indicate to the caller whether privilege was used to satisfy the
4683  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4684  */
4685 int
4686 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4687     accmode_t accmode, struct ucred *cred, int *privused)
4688 {
4689 	accmode_t dac_granted;
4690 	accmode_t priv_granted;
4691 
4692 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4693 	    ("invalid bit in accmode"));
4694 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4695 	    ("VAPPEND without VWRITE"));
4696 
4697 	/*
4698 	 * Look for a normal, non-privileged way to access the file/directory
4699 	 * as requested.  If it exists, go with that.
4700 	 */
4701 
4702 	if (privused != NULL)
4703 		*privused = 0;
4704 
4705 	dac_granted = 0;
4706 
4707 	/* Check the owner. */
4708 	if (cred->cr_uid == file_uid) {
4709 		dac_granted |= VADMIN;
4710 		if (file_mode & S_IXUSR)
4711 			dac_granted |= VEXEC;
4712 		if (file_mode & S_IRUSR)
4713 			dac_granted |= VREAD;
4714 		if (file_mode & S_IWUSR)
4715 			dac_granted |= (VWRITE | VAPPEND);
4716 
4717 		if ((accmode & dac_granted) == accmode)
4718 			return (0);
4719 
4720 		goto privcheck;
4721 	}
4722 
4723 	/* Otherwise, check the groups (first match) */
4724 	if (groupmember(file_gid, cred)) {
4725 		if (file_mode & S_IXGRP)
4726 			dac_granted |= VEXEC;
4727 		if (file_mode & S_IRGRP)
4728 			dac_granted |= VREAD;
4729 		if (file_mode & S_IWGRP)
4730 			dac_granted |= (VWRITE | VAPPEND);
4731 
4732 		if ((accmode & dac_granted) == accmode)
4733 			return (0);
4734 
4735 		goto privcheck;
4736 	}
4737 
4738 	/* Otherwise, check everyone else. */
4739 	if (file_mode & S_IXOTH)
4740 		dac_granted |= VEXEC;
4741 	if (file_mode & S_IROTH)
4742 		dac_granted |= VREAD;
4743 	if (file_mode & S_IWOTH)
4744 		dac_granted |= (VWRITE | VAPPEND);
4745 	if ((accmode & dac_granted) == accmode)
4746 		return (0);
4747 
4748 privcheck:
4749 	/*
4750 	 * Build a privilege mask to determine if the set of privileges
4751 	 * satisfies the requirements when combined with the granted mask
4752 	 * from above.  For each privilege, if the privilege is required,
4753 	 * bitwise or the request type onto the priv_granted mask.
4754 	 */
4755 	priv_granted = 0;
4756 
4757 	if (type == VDIR) {
4758 		/*
4759 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4760 		 * requests, instead of PRIV_VFS_EXEC.
4761 		 */
4762 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4763 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4764 			priv_granted |= VEXEC;
4765 	} else {
4766 		/*
4767 		 * Ensure that at least one execute bit is on. Otherwise,
4768 		 * a privileged user will always succeed, and we don't want
4769 		 * this to happen unless the file really is executable.
4770 		 */
4771 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4772 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4773 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4774 			priv_granted |= VEXEC;
4775 	}
4776 
4777 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4778 	    !priv_check_cred(cred, PRIV_VFS_READ))
4779 		priv_granted |= VREAD;
4780 
4781 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4782 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4783 		priv_granted |= (VWRITE | VAPPEND);
4784 
4785 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4786 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4787 		priv_granted |= VADMIN;
4788 
4789 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4790 		/* XXX audit: privilege used */
4791 		if (privused != NULL)
4792 			*privused = 1;
4793 		return (0);
4794 	}
4795 
4796 	return ((accmode & VADMIN) ? EPERM : EACCES);
4797 }
4798 
4799 /*
4800  * Credential check based on process requesting service, and per-attribute
4801  * permissions.
4802  */
4803 int
4804 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4805     struct thread *td, accmode_t accmode)
4806 {
4807 
4808 	/*
4809 	 * Kernel-invoked always succeeds.
4810 	 */
4811 	if (cred == NOCRED)
4812 		return (0);
4813 
4814 	/*
4815 	 * Do not allow privileged processes in jail to directly manipulate
4816 	 * system attributes.
4817 	 */
4818 	switch (attrnamespace) {
4819 	case EXTATTR_NAMESPACE_SYSTEM:
4820 		/* Potentially should be: return (EPERM); */
4821 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4822 	case EXTATTR_NAMESPACE_USER:
4823 		return (VOP_ACCESS(vp, accmode, cred, td));
4824 	default:
4825 		return (EPERM);
4826 	}
4827 }
4828 
4829 #ifdef DEBUG_VFS_LOCKS
4830 /*
4831  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4832  * no longer ok to have an unlocked VFS.
4833  */
4834 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4835 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4836 
4837 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4838 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4839     "Drop into debugger on lock violation");
4840 
4841 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4842 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4843     0, "Check for interlock across VOPs");
4844 
4845 int vfs_badlock_print = 1;	/* Print lock violations. */
4846 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4847     0, "Print lock violations");
4848 
4849 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4850 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4851     0, "Print vnode details on lock violations");
4852 
4853 #ifdef KDB
4854 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4855 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4856     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4857 #endif
4858 
4859 static void
4860 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4861 {
4862 
4863 #ifdef KDB
4864 	if (vfs_badlock_backtrace)
4865 		kdb_backtrace();
4866 #endif
4867 	if (vfs_badlock_vnode)
4868 		vn_printf(vp, "vnode ");
4869 	if (vfs_badlock_print)
4870 		printf("%s: %p %s\n", str, (void *)vp, msg);
4871 	if (vfs_badlock_ddb)
4872 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4873 }
4874 
4875 void
4876 assert_vi_locked(struct vnode *vp, const char *str)
4877 {
4878 
4879 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4880 		vfs_badlock("interlock is not locked but should be", str, vp);
4881 }
4882 
4883 void
4884 assert_vi_unlocked(struct vnode *vp, const char *str)
4885 {
4886 
4887 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4888 		vfs_badlock("interlock is locked but should not be", str, vp);
4889 }
4890 
4891 void
4892 assert_vop_locked(struct vnode *vp, const char *str)
4893 {
4894 	int locked;
4895 
4896 	if (!IGNORE_LOCK(vp)) {
4897 		locked = VOP_ISLOCKED(vp);
4898 		if (locked == 0 || locked == LK_EXCLOTHER)
4899 			vfs_badlock("is not locked but should be", str, vp);
4900 	}
4901 }
4902 
4903 void
4904 assert_vop_unlocked(struct vnode *vp, const char *str)
4905 {
4906 
4907 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4908 		vfs_badlock("is locked but should not be", str, vp);
4909 }
4910 
4911 void
4912 assert_vop_elocked(struct vnode *vp, const char *str)
4913 {
4914 
4915 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4916 		vfs_badlock("is not exclusive locked but should be", str, vp);
4917 }
4918 #endif /* DEBUG_VFS_LOCKS */
4919 
4920 void
4921 vop_rename_fail(struct vop_rename_args *ap)
4922 {
4923 
4924 	if (ap->a_tvp != NULL)
4925 		vput(ap->a_tvp);
4926 	if (ap->a_tdvp == ap->a_tvp)
4927 		vrele(ap->a_tdvp);
4928 	else
4929 		vput(ap->a_tdvp);
4930 	vrele(ap->a_fdvp);
4931 	vrele(ap->a_fvp);
4932 }
4933 
4934 void
4935 vop_rename_pre(void *ap)
4936 {
4937 	struct vop_rename_args *a = ap;
4938 
4939 #ifdef DEBUG_VFS_LOCKS
4940 	if (a->a_tvp)
4941 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4942 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4943 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4944 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4945 
4946 	/* Check the source (from). */
4947 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4948 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4949 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4950 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4951 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4952 
4953 	/* Check the target. */
4954 	if (a->a_tvp)
4955 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4956 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4957 #endif
4958 	if (a->a_tdvp != a->a_fdvp)
4959 		vhold(a->a_fdvp);
4960 	if (a->a_tvp != a->a_fvp)
4961 		vhold(a->a_fvp);
4962 	vhold(a->a_tdvp);
4963 	if (a->a_tvp)
4964 		vhold(a->a_tvp);
4965 }
4966 
4967 #ifdef DEBUG_VFS_LOCKS
4968 void
4969 vop_strategy_pre(void *ap)
4970 {
4971 	struct vop_strategy_args *a;
4972 	struct buf *bp;
4973 
4974 	a = ap;
4975 	bp = a->a_bp;
4976 
4977 	/*
4978 	 * Cluster ops lock their component buffers but not the IO container.
4979 	 */
4980 	if ((bp->b_flags & B_CLUSTER) != 0)
4981 		return;
4982 
4983 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4984 		if (vfs_badlock_print)
4985 			printf(
4986 			    "VOP_STRATEGY: bp is not locked but should be\n");
4987 		if (vfs_badlock_ddb)
4988 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4989 	}
4990 }
4991 
4992 void
4993 vop_lock_pre(void *ap)
4994 {
4995 	struct vop_lock1_args *a = ap;
4996 
4997 	if ((a->a_flags & LK_INTERLOCK) == 0)
4998 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4999 	else
5000 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5001 }
5002 
5003 void
5004 vop_lock_post(void *ap, int rc)
5005 {
5006 	struct vop_lock1_args *a = ap;
5007 
5008 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5009 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5010 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5011 }
5012 
5013 void
5014 vop_unlock_pre(void *ap)
5015 {
5016 	struct vop_unlock_args *a = ap;
5017 
5018 	if (a->a_flags & LK_INTERLOCK)
5019 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
5020 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5021 }
5022 
5023 void
5024 vop_unlock_post(void *ap, int rc)
5025 {
5026 	struct vop_unlock_args *a = ap;
5027 
5028 	if (a->a_flags & LK_INTERLOCK)
5029 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
5030 }
5031 
5032 void
5033 vop_need_inactive_pre(void *ap)
5034 {
5035 	struct vop_need_inactive_args *a = ap;
5036 
5037 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5038 }
5039 
5040 void
5041 vop_need_inactive_post(void *ap, int rc)
5042 {
5043 	struct vop_need_inactive_args *a = ap;
5044 
5045 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5046 }
5047 #endif
5048 
5049 void
5050 vop_create_post(void *ap, int rc)
5051 {
5052 	struct vop_create_args *a = ap;
5053 
5054 	if (!rc)
5055 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5056 }
5057 
5058 void
5059 vop_deleteextattr_post(void *ap, int rc)
5060 {
5061 	struct vop_deleteextattr_args *a = ap;
5062 
5063 	if (!rc)
5064 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5065 }
5066 
5067 void
5068 vop_link_post(void *ap, int rc)
5069 {
5070 	struct vop_link_args *a = ap;
5071 
5072 	if (!rc) {
5073 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5074 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5075 	}
5076 }
5077 
5078 void
5079 vop_mkdir_post(void *ap, int rc)
5080 {
5081 	struct vop_mkdir_args *a = ap;
5082 
5083 	if (!rc)
5084 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5085 }
5086 
5087 void
5088 vop_mknod_post(void *ap, int rc)
5089 {
5090 	struct vop_mknod_args *a = ap;
5091 
5092 	if (!rc)
5093 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5094 }
5095 
5096 void
5097 vop_reclaim_post(void *ap, int rc)
5098 {
5099 	struct vop_reclaim_args *a = ap;
5100 
5101 	if (!rc)
5102 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5103 }
5104 
5105 void
5106 vop_remove_post(void *ap, int rc)
5107 {
5108 	struct vop_remove_args *a = ap;
5109 
5110 	if (!rc) {
5111 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5112 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5113 	}
5114 }
5115 
5116 void
5117 vop_rename_post(void *ap, int rc)
5118 {
5119 	struct vop_rename_args *a = ap;
5120 	long hint;
5121 
5122 	if (!rc) {
5123 		hint = NOTE_WRITE;
5124 		if (a->a_fdvp == a->a_tdvp) {
5125 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5126 				hint |= NOTE_LINK;
5127 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5128 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5129 		} else {
5130 			hint |= NOTE_EXTEND;
5131 			if (a->a_fvp->v_type == VDIR)
5132 				hint |= NOTE_LINK;
5133 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5134 
5135 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5136 			    a->a_tvp->v_type == VDIR)
5137 				hint &= ~NOTE_LINK;
5138 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5139 		}
5140 
5141 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5142 		if (a->a_tvp)
5143 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5144 	}
5145 	if (a->a_tdvp != a->a_fdvp)
5146 		vdrop(a->a_fdvp);
5147 	if (a->a_tvp != a->a_fvp)
5148 		vdrop(a->a_fvp);
5149 	vdrop(a->a_tdvp);
5150 	if (a->a_tvp)
5151 		vdrop(a->a_tvp);
5152 }
5153 
5154 void
5155 vop_rmdir_post(void *ap, int rc)
5156 {
5157 	struct vop_rmdir_args *a = ap;
5158 
5159 	if (!rc) {
5160 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5161 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5162 	}
5163 }
5164 
5165 void
5166 vop_setattr_post(void *ap, int rc)
5167 {
5168 	struct vop_setattr_args *a = ap;
5169 
5170 	if (!rc)
5171 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5172 }
5173 
5174 void
5175 vop_setextattr_post(void *ap, int rc)
5176 {
5177 	struct vop_setextattr_args *a = ap;
5178 
5179 	if (!rc)
5180 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5181 }
5182 
5183 void
5184 vop_symlink_post(void *ap, int rc)
5185 {
5186 	struct vop_symlink_args *a = ap;
5187 
5188 	if (!rc)
5189 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5190 }
5191 
5192 void
5193 vop_open_post(void *ap, int rc)
5194 {
5195 	struct vop_open_args *a = ap;
5196 
5197 	if (!rc)
5198 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5199 }
5200 
5201 void
5202 vop_close_post(void *ap, int rc)
5203 {
5204 	struct vop_close_args *a = ap;
5205 
5206 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5207 	    !VN_IS_DOOMED(a->a_vp))) {
5208 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5209 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5210 	}
5211 }
5212 
5213 void
5214 vop_read_post(void *ap, int rc)
5215 {
5216 	struct vop_read_args *a = ap;
5217 
5218 	if (!rc)
5219 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5220 }
5221 
5222 void
5223 vop_readdir_post(void *ap, int rc)
5224 {
5225 	struct vop_readdir_args *a = ap;
5226 
5227 	if (!rc)
5228 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5229 }
5230 
5231 static struct knlist fs_knlist;
5232 
5233 static void
5234 vfs_event_init(void *arg)
5235 {
5236 	knlist_init_mtx(&fs_knlist, NULL);
5237 }
5238 /* XXX - correct order? */
5239 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5240 
5241 void
5242 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5243 {
5244 
5245 	KNOTE_UNLOCKED(&fs_knlist, event);
5246 }
5247 
5248 static int	filt_fsattach(struct knote *kn);
5249 static void	filt_fsdetach(struct knote *kn);
5250 static int	filt_fsevent(struct knote *kn, long hint);
5251 
5252 struct filterops fs_filtops = {
5253 	.f_isfd = 0,
5254 	.f_attach = filt_fsattach,
5255 	.f_detach = filt_fsdetach,
5256 	.f_event = filt_fsevent
5257 };
5258 
5259 static int
5260 filt_fsattach(struct knote *kn)
5261 {
5262 
5263 	kn->kn_flags |= EV_CLEAR;
5264 	knlist_add(&fs_knlist, kn, 0);
5265 	return (0);
5266 }
5267 
5268 static void
5269 filt_fsdetach(struct knote *kn)
5270 {
5271 
5272 	knlist_remove(&fs_knlist, kn, 0);
5273 }
5274 
5275 static int
5276 filt_fsevent(struct knote *kn, long hint)
5277 {
5278 
5279 	kn->kn_fflags |= hint;
5280 	return (kn->kn_fflags != 0);
5281 }
5282 
5283 static int
5284 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5285 {
5286 	struct vfsidctl vc;
5287 	int error;
5288 	struct mount *mp;
5289 
5290 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5291 	if (error)
5292 		return (error);
5293 	if (vc.vc_vers != VFS_CTL_VERS1)
5294 		return (EINVAL);
5295 	mp = vfs_getvfs(&vc.vc_fsid);
5296 	if (mp == NULL)
5297 		return (ENOENT);
5298 	/* ensure that a specific sysctl goes to the right filesystem. */
5299 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5300 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5301 		vfs_rel(mp);
5302 		return (EINVAL);
5303 	}
5304 	VCTLTOREQ(&vc, req);
5305 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5306 	vfs_rel(mp);
5307 	return (error);
5308 }
5309 
5310 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5311     NULL, 0, sysctl_vfs_ctl, "",
5312     "Sysctl by fsid");
5313 
5314 /*
5315  * Function to initialize a va_filerev field sensibly.
5316  * XXX: Wouldn't a random number make a lot more sense ??
5317  */
5318 u_quad_t
5319 init_va_filerev(void)
5320 {
5321 	struct bintime bt;
5322 
5323 	getbinuptime(&bt);
5324 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5325 }
5326 
5327 static int	filt_vfsread(struct knote *kn, long hint);
5328 static int	filt_vfswrite(struct knote *kn, long hint);
5329 static int	filt_vfsvnode(struct knote *kn, long hint);
5330 static void	filt_vfsdetach(struct knote *kn);
5331 static struct filterops vfsread_filtops = {
5332 	.f_isfd = 1,
5333 	.f_detach = filt_vfsdetach,
5334 	.f_event = filt_vfsread
5335 };
5336 static struct filterops vfswrite_filtops = {
5337 	.f_isfd = 1,
5338 	.f_detach = filt_vfsdetach,
5339 	.f_event = filt_vfswrite
5340 };
5341 static struct filterops vfsvnode_filtops = {
5342 	.f_isfd = 1,
5343 	.f_detach = filt_vfsdetach,
5344 	.f_event = filt_vfsvnode
5345 };
5346 
5347 static void
5348 vfs_knllock(void *arg)
5349 {
5350 	struct vnode *vp = arg;
5351 
5352 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5353 }
5354 
5355 static void
5356 vfs_knlunlock(void *arg)
5357 {
5358 	struct vnode *vp = arg;
5359 
5360 	VOP_UNLOCK(vp, 0);
5361 }
5362 
5363 static void
5364 vfs_knl_assert_locked(void *arg)
5365 {
5366 #ifdef DEBUG_VFS_LOCKS
5367 	struct vnode *vp = arg;
5368 
5369 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5370 #endif
5371 }
5372 
5373 static void
5374 vfs_knl_assert_unlocked(void *arg)
5375 {
5376 #ifdef DEBUG_VFS_LOCKS
5377 	struct vnode *vp = arg;
5378 
5379 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5380 #endif
5381 }
5382 
5383 int
5384 vfs_kqfilter(struct vop_kqfilter_args *ap)
5385 {
5386 	struct vnode *vp = ap->a_vp;
5387 	struct knote *kn = ap->a_kn;
5388 	struct knlist *knl;
5389 
5390 	switch (kn->kn_filter) {
5391 	case EVFILT_READ:
5392 		kn->kn_fop = &vfsread_filtops;
5393 		break;
5394 	case EVFILT_WRITE:
5395 		kn->kn_fop = &vfswrite_filtops;
5396 		break;
5397 	case EVFILT_VNODE:
5398 		kn->kn_fop = &vfsvnode_filtops;
5399 		break;
5400 	default:
5401 		return (EINVAL);
5402 	}
5403 
5404 	kn->kn_hook = (caddr_t)vp;
5405 
5406 	v_addpollinfo(vp);
5407 	if (vp->v_pollinfo == NULL)
5408 		return (ENOMEM);
5409 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5410 	vhold(vp);
5411 	knlist_add(knl, kn, 0);
5412 
5413 	return (0);
5414 }
5415 
5416 /*
5417  * Detach knote from vnode
5418  */
5419 static void
5420 filt_vfsdetach(struct knote *kn)
5421 {
5422 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5423 
5424 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5425 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5426 	vdrop(vp);
5427 }
5428 
5429 /*ARGSUSED*/
5430 static int
5431 filt_vfsread(struct knote *kn, long hint)
5432 {
5433 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5434 	struct vattr va;
5435 	int res;
5436 
5437 	/*
5438 	 * filesystem is gone, so set the EOF flag and schedule
5439 	 * the knote for deletion.
5440 	 */
5441 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5442 		VI_LOCK(vp);
5443 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5444 		VI_UNLOCK(vp);
5445 		return (1);
5446 	}
5447 
5448 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5449 		return (0);
5450 
5451 	VI_LOCK(vp);
5452 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5453 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5454 	VI_UNLOCK(vp);
5455 	return (res);
5456 }
5457 
5458 /*ARGSUSED*/
5459 static int
5460 filt_vfswrite(struct knote *kn, long hint)
5461 {
5462 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5463 
5464 	VI_LOCK(vp);
5465 
5466 	/*
5467 	 * filesystem is gone, so set the EOF flag and schedule
5468 	 * the knote for deletion.
5469 	 */
5470 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5471 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5472 
5473 	kn->kn_data = 0;
5474 	VI_UNLOCK(vp);
5475 	return (1);
5476 }
5477 
5478 static int
5479 filt_vfsvnode(struct knote *kn, long hint)
5480 {
5481 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5482 	int res;
5483 
5484 	VI_LOCK(vp);
5485 	if (kn->kn_sfflags & hint)
5486 		kn->kn_fflags |= hint;
5487 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5488 		kn->kn_flags |= EV_EOF;
5489 		VI_UNLOCK(vp);
5490 		return (1);
5491 	}
5492 	res = (kn->kn_fflags != 0);
5493 	VI_UNLOCK(vp);
5494 	return (res);
5495 }
5496 
5497 /*
5498  * Returns whether the directory is empty or not.
5499  * If it is empty, the return value is 0; otherwise
5500  * the return value is an error value (which may
5501  * be ENOTEMPTY).
5502  */
5503 int
5504 vfs_emptydir(struct vnode *vp)
5505 {
5506 	struct uio uio;
5507 	struct iovec iov;
5508 	struct dirent *dirent, *dp, *endp;
5509 	int error, eof;
5510 
5511 	error = 0;
5512 	eof = 0;
5513 
5514 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5515 
5516 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5517 	iov.iov_base = dirent;
5518 	iov.iov_len = sizeof(struct dirent);
5519 
5520 	uio.uio_iov = &iov;
5521 	uio.uio_iovcnt = 1;
5522 	uio.uio_offset = 0;
5523 	uio.uio_resid = sizeof(struct dirent);
5524 	uio.uio_segflg = UIO_SYSSPACE;
5525 	uio.uio_rw = UIO_READ;
5526 	uio.uio_td = curthread;
5527 
5528 	while (eof == 0 && error == 0) {
5529 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5530 		    NULL, NULL);
5531 		if (error != 0)
5532 			break;
5533 		endp = (void *)((uint8_t *)dirent +
5534 		    sizeof(struct dirent) - uio.uio_resid);
5535 		for (dp = dirent; dp < endp;
5536 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5537 			if (dp->d_type == DT_WHT)
5538 				continue;
5539 			if (dp->d_namlen == 0)
5540 				continue;
5541 			if (dp->d_type != DT_DIR &&
5542 			    dp->d_type != DT_UNKNOWN) {
5543 				error = ENOTEMPTY;
5544 				break;
5545 			}
5546 			if (dp->d_namlen > 2) {
5547 				error = ENOTEMPTY;
5548 				break;
5549 			}
5550 			if (dp->d_namlen == 1 &&
5551 			    dp->d_name[0] != '.') {
5552 				error = ENOTEMPTY;
5553 				break;
5554 			}
5555 			if (dp->d_namlen == 2 &&
5556 			    dp->d_name[1] != '.') {
5557 				error = ENOTEMPTY;
5558 				break;
5559 			}
5560 			uio.uio_resid = sizeof(struct dirent);
5561 		}
5562 	}
5563 	free(dirent, M_TEMP);
5564 	return (error);
5565 }
5566 
5567 int
5568 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5569 {
5570 	int error;
5571 
5572 	if (dp->d_reclen > ap->a_uio->uio_resid)
5573 		return (ENAMETOOLONG);
5574 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5575 	if (error) {
5576 		if (ap->a_ncookies != NULL) {
5577 			if (ap->a_cookies != NULL)
5578 				free(ap->a_cookies, M_TEMP);
5579 			ap->a_cookies = NULL;
5580 			*ap->a_ncookies = 0;
5581 		}
5582 		return (error);
5583 	}
5584 	if (ap->a_ncookies == NULL)
5585 		return (0);
5586 
5587 	KASSERT(ap->a_cookies,
5588 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5589 
5590 	*ap->a_cookies = realloc(*ap->a_cookies,
5591 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5592 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5593 	*ap->a_ncookies += 1;
5594 	return (0);
5595 }
5596 
5597 /*
5598  * Mark for update the access time of the file if the filesystem
5599  * supports VOP_MARKATIME.  This functionality is used by execve and
5600  * mmap, so we want to avoid the I/O implied by directly setting
5601  * va_atime for the sake of efficiency.
5602  */
5603 void
5604 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5605 {
5606 	struct mount *mp;
5607 
5608 	mp = vp->v_mount;
5609 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5610 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5611 		(void)VOP_MARKATIME(vp);
5612 }
5613 
5614 /*
5615  * The purpose of this routine is to remove granularity from accmode_t,
5616  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5617  * VADMIN and VAPPEND.
5618  *
5619  * If it returns 0, the caller is supposed to continue with the usual
5620  * access checks using 'accmode' as modified by this routine.  If it
5621  * returns nonzero value, the caller is supposed to return that value
5622  * as errno.
5623  *
5624  * Note that after this routine runs, accmode may be zero.
5625  */
5626 int
5627 vfs_unixify_accmode(accmode_t *accmode)
5628 {
5629 	/*
5630 	 * There is no way to specify explicit "deny" rule using
5631 	 * file mode or POSIX.1e ACLs.
5632 	 */
5633 	if (*accmode & VEXPLICIT_DENY) {
5634 		*accmode = 0;
5635 		return (0);
5636 	}
5637 
5638 	/*
5639 	 * None of these can be translated into usual access bits.
5640 	 * Also, the common case for NFSv4 ACLs is to not contain
5641 	 * either of these bits. Caller should check for VWRITE
5642 	 * on the containing directory instead.
5643 	 */
5644 	if (*accmode & (VDELETE_CHILD | VDELETE))
5645 		return (EPERM);
5646 
5647 	if (*accmode & VADMIN_PERMS) {
5648 		*accmode &= ~VADMIN_PERMS;
5649 		*accmode |= VADMIN;
5650 	}
5651 
5652 	/*
5653 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5654 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5655 	 */
5656 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5657 
5658 	return (0);
5659 }
5660 
5661 /*
5662  * Clear out a doomed vnode (if any) and replace it with a new one as long
5663  * as the fs is not being unmounted. Return the root vnode to the caller.
5664  */
5665 static int __noinline
5666 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
5667 {
5668 	struct vnode *vp;
5669 	int error;
5670 
5671 restart:
5672 	if (mp->mnt_rootvnode != NULL) {
5673 		MNT_ILOCK(mp);
5674 		vp = mp->mnt_rootvnode;
5675 		if (vp != NULL) {
5676 			if (!VN_IS_DOOMED(vp)) {
5677 				vrefact(vp);
5678 				MNT_IUNLOCK(mp);
5679 				error = vn_lock(vp, flags);
5680 				if (error == 0) {
5681 					*vpp = vp;
5682 					return (0);
5683 				}
5684 				vrele(vp);
5685 				goto restart;
5686 			}
5687 			/*
5688 			 * Clear the old one.
5689 			 */
5690 			mp->mnt_rootvnode = NULL;
5691 		}
5692 		MNT_IUNLOCK(mp);
5693 		if (vp != NULL) {
5694 			/*
5695 			 * Paired with a fence in vfs_op_thread_exit().
5696 			 */
5697 			atomic_thread_fence_acq();
5698 			vfs_op_barrier_wait(mp);
5699 			vrele(vp);
5700 		}
5701 	}
5702 	error = VFS_CACHEDROOT(mp, flags, vpp);
5703 	if (error != 0)
5704 		return (error);
5705 	if (mp->mnt_vfs_ops == 0) {
5706 		MNT_ILOCK(mp);
5707 		if (mp->mnt_vfs_ops != 0) {
5708 			MNT_IUNLOCK(mp);
5709 			return (0);
5710 		}
5711 		if (mp->mnt_rootvnode == NULL) {
5712 			vrefact(*vpp);
5713 			mp->mnt_rootvnode = *vpp;
5714 		} else {
5715 			if (mp->mnt_rootvnode != *vpp) {
5716 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
5717 					panic("%s: mismatch between vnode returned "
5718 					    " by VFS_CACHEDROOT and the one cached "
5719 					    " (%p != %p)",
5720 					    __func__, *vpp, mp->mnt_rootvnode);
5721 				}
5722 			}
5723 		}
5724 		MNT_IUNLOCK(mp);
5725 	}
5726 	return (0);
5727 }
5728 
5729 int
5730 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
5731 {
5732 	struct vnode *vp;
5733 	int error;
5734 
5735 	if (!vfs_op_thread_enter(mp))
5736 		return (vfs_cache_root_fallback(mp, flags, vpp));
5737 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
5738 	if (vp == NULL || VN_IS_DOOMED(vp)) {
5739 		vfs_op_thread_exit(mp);
5740 		return (vfs_cache_root_fallback(mp, flags, vpp));
5741 	}
5742 	vrefact(vp);
5743 	vfs_op_thread_exit(mp);
5744 	error = vn_lock(vp, flags);
5745 	if (error != 0) {
5746 		vrele(vp);
5747 		return (vfs_cache_root_fallback(mp, flags, vpp));
5748 	}
5749 	*vpp = vp;
5750 	return (0);
5751 }
5752 
5753 struct vnode *
5754 vfs_cache_root_clear(struct mount *mp)
5755 {
5756 	struct vnode *vp;
5757 
5758 	/*
5759 	 * ops > 0 guarantees there is nobody who can see this vnode
5760 	 */
5761 	MPASS(mp->mnt_vfs_ops > 0);
5762 	vp = mp->mnt_rootvnode;
5763 	mp->mnt_rootvnode = NULL;
5764 	return (vp);
5765 }
5766 
5767 void
5768 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
5769 {
5770 
5771 	MPASS(mp->mnt_vfs_ops > 0);
5772 	vrefact(vp);
5773 	mp->mnt_rootvnode = vp;
5774 }
5775 
5776 /*
5777  * These are helper functions for filesystems to traverse all
5778  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5779  *
5780  * This interface replaces MNT_VNODE_FOREACH.
5781  */
5782 
5783 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5784 
5785 struct vnode *
5786 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5787 {
5788 	struct vnode *vp;
5789 
5790 	if (should_yield())
5791 		kern_yield(PRI_USER);
5792 	MNT_ILOCK(mp);
5793 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5794 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5795 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5796 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
5797 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
5798 			continue;
5799 		VI_LOCK(vp);
5800 		if (VN_IS_DOOMED(vp)) {
5801 			VI_UNLOCK(vp);
5802 			continue;
5803 		}
5804 		break;
5805 	}
5806 	if (vp == NULL) {
5807 		__mnt_vnode_markerfree_all(mvp, mp);
5808 		/* MNT_IUNLOCK(mp); -- done in above function */
5809 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5810 		return (NULL);
5811 	}
5812 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5813 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5814 	MNT_IUNLOCK(mp);
5815 	return (vp);
5816 }
5817 
5818 struct vnode *
5819 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5820 {
5821 	struct vnode *vp;
5822 
5823 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5824 	MNT_ILOCK(mp);
5825 	MNT_REF(mp);
5826 	(*mvp)->v_mount = mp;
5827 	(*mvp)->v_type = VMARKER;
5828 
5829 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5830 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
5831 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
5832 			continue;
5833 		VI_LOCK(vp);
5834 		if (VN_IS_DOOMED(vp)) {
5835 			VI_UNLOCK(vp);
5836 			continue;
5837 		}
5838 		break;
5839 	}
5840 	if (vp == NULL) {
5841 		MNT_REL(mp);
5842 		MNT_IUNLOCK(mp);
5843 		free(*mvp, M_VNODE_MARKER);
5844 		*mvp = NULL;
5845 		return (NULL);
5846 	}
5847 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5848 	MNT_IUNLOCK(mp);
5849 	return (vp);
5850 }
5851 
5852 void
5853 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5854 {
5855 
5856 	if (*mvp == NULL) {
5857 		MNT_IUNLOCK(mp);
5858 		return;
5859 	}
5860 
5861 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5862 
5863 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5864 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5865 	MNT_REL(mp);
5866 	MNT_IUNLOCK(mp);
5867 	free(*mvp, M_VNODE_MARKER);
5868 	*mvp = NULL;
5869 }
5870 
5871 /*
5872  * These are helper functions for filesystems to traverse their
5873  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5874  */
5875 static void
5876 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5877 {
5878 
5879 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5880 
5881 	MNT_ILOCK(mp);
5882 	MNT_REL(mp);
5883 	MNT_IUNLOCK(mp);
5884 	free(*mvp, M_VNODE_MARKER);
5885 	*mvp = NULL;
5886 }
5887 
5888 /*
5889  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5890  * conventional lock order during mnt_vnode_next_active iteration.
5891  *
5892  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5893  * The list lock is dropped and reacquired.  On success, both locks are held.
5894  * On failure, the mount vnode list lock is held but the vnode interlock is
5895  * not, and the procedure may have yielded.
5896  */
5897 static bool
5898 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5899     struct vnode *vp)
5900 {
5901 	const struct vnode *tmp;
5902 	bool held, ret;
5903 
5904 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5905 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5906 	    ("%s: bad marker", __func__));
5907 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5908 	    ("%s: inappropriate vnode", __func__));
5909 	ASSERT_VI_UNLOCKED(vp, __func__);
5910 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5911 
5912 	ret = false;
5913 
5914 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5915 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5916 
5917 	/*
5918 	 * Use a hold to prevent vp from disappearing while the mount vnode
5919 	 * list lock is dropped and reacquired.  Normally a hold would be
5920 	 * acquired with vhold(), but that might try to acquire the vnode
5921 	 * interlock, which would be a LOR with the mount vnode list lock.
5922 	 */
5923 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5924 	mtx_unlock(&mp->mnt_listmtx);
5925 	if (!held)
5926 		goto abort;
5927 	VI_LOCK(vp);
5928 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5929 		vdropl(vp);
5930 		goto abort;
5931 	}
5932 	mtx_lock(&mp->mnt_listmtx);
5933 
5934 	/*
5935 	 * Determine whether the vnode is still the next one after the marker,
5936 	 * excepting any other markers.  If the vnode has not been doomed by
5937 	 * vgone() then the hold should have ensured that it remained on the
5938 	 * active list.  If it has been doomed but is still on the active list,
5939 	 * don't abort, but rather skip over it (avoid spinning on doomed
5940 	 * vnodes).
5941 	 */
5942 	tmp = mvp;
5943 	do {
5944 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5945 	} while (tmp != NULL && tmp->v_type == VMARKER);
5946 	if (tmp != vp) {
5947 		mtx_unlock(&mp->mnt_listmtx);
5948 		VI_UNLOCK(vp);
5949 		goto abort;
5950 	}
5951 
5952 	ret = true;
5953 	goto out;
5954 abort:
5955 	maybe_yield();
5956 	mtx_lock(&mp->mnt_listmtx);
5957 out:
5958 	if (ret)
5959 		ASSERT_VI_LOCKED(vp, __func__);
5960 	else
5961 		ASSERT_VI_UNLOCKED(vp, __func__);
5962 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5963 	return (ret);
5964 }
5965 
5966 static struct vnode *
5967 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5968 {
5969 	struct vnode *vp, *nvp;
5970 
5971 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5972 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5973 restart:
5974 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5975 	while (vp != NULL) {
5976 		if (vp->v_type == VMARKER) {
5977 			vp = TAILQ_NEXT(vp, v_actfreelist);
5978 			continue;
5979 		}
5980 		/*
5981 		 * Try-lock because this is the wrong lock order.  If that does
5982 		 * not succeed, drop the mount vnode list lock and try to
5983 		 * reacquire it and the vnode interlock in the right order.
5984 		 */
5985 		if (!VI_TRYLOCK(vp) &&
5986 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5987 			goto restart;
5988 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5989 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5990 		    ("alien vnode on the active list %p %p", vp, mp));
5991 		if (vp->v_mount == mp && !VN_IS_DOOMED(vp))
5992 			break;
5993 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5994 		VI_UNLOCK(vp);
5995 		vp = nvp;
5996 	}
5997 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5998 
5999 	/* Check if we are done */
6000 	if (vp == NULL) {
6001 		mtx_unlock(&mp->mnt_listmtx);
6002 		mnt_vnode_markerfree_active(mvp, mp);
6003 		return (NULL);
6004 	}
6005 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
6006 	mtx_unlock(&mp->mnt_listmtx);
6007 	ASSERT_VI_LOCKED(vp, "active iter");
6008 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
6009 	return (vp);
6010 }
6011 
6012 struct vnode *
6013 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
6014 {
6015 
6016 	if (should_yield())
6017 		kern_yield(PRI_USER);
6018 	mtx_lock(&mp->mnt_listmtx);
6019 	return (mnt_vnode_next_active(mvp, mp));
6020 }
6021 
6022 struct vnode *
6023 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
6024 {
6025 	struct vnode *vp;
6026 
6027 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
6028 	MNT_ILOCK(mp);
6029 	MNT_REF(mp);
6030 	MNT_IUNLOCK(mp);
6031 	(*mvp)->v_type = VMARKER;
6032 	(*mvp)->v_mount = mp;
6033 
6034 	mtx_lock(&mp->mnt_listmtx);
6035 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
6036 	if (vp == NULL) {
6037 		mtx_unlock(&mp->mnt_listmtx);
6038 		mnt_vnode_markerfree_active(mvp, mp);
6039 		return (NULL);
6040 	}
6041 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
6042 	return (mnt_vnode_next_active(mvp, mp));
6043 }
6044 
6045 void
6046 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
6047 {
6048 
6049 	if (*mvp == NULL)
6050 		return;
6051 
6052 	mtx_lock(&mp->mnt_listmtx);
6053 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
6054 	mtx_unlock(&mp->mnt_listmtx);
6055 	mnt_vnode_markerfree_active(mvp, mp);
6056 }
6057