xref: /freebsd/sys/kern/vfs_subr.c (revision 4b50c451720d8b427757a6da1dd2bb4c52cd9e35)
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	destroy_vpollinfo(struct vpollinfo *vi);
118 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
119 		    daddr_t startlbn, daddr_t endlbn);
120 static void	vnlru_recalc(void);
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 u_long __exclusive_cache_line 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 /*
152  * Conversion tables for conversion from vnode types to inode formats
153  * and back.
154  */
155 enum vtype iftovt_tab[16] = {
156 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
157 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
158 };
159 int vttoif_tab[10] = {
160 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
161 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
162 };
163 
164 /*
165  * List of allocates vnodes in the system.
166  */
167 static TAILQ_HEAD(freelst, vnode) vnode_list;
168 static struct vnode *vnode_list_free_marker;
169 
170 /*
171  * "Free" vnode target.  Free vnodes are rarely completely free, but are
172  * just ones that are cheap to recycle.  Usually they are for files which
173  * have been stat'd but not read; these usually have inode and namecache
174  * data attached to them.  This target is the preferred minimum size of a
175  * sub-cache consisting mostly of such files. The system balances the size
176  * of this sub-cache with its complement to try to prevent either from
177  * thrashing while the other is relatively inactive.  The targets express
178  * a preference for the best balance.
179  *
180  * "Above" this target there are 2 further targets (watermarks) related
181  * to recyling of free vnodes.  In the best-operating case, the cache is
182  * exactly full, the free list has size between vlowat and vhiwat above the
183  * free target, and recycling from it and normal use maintains this state.
184  * Sometimes the free list is below vlowat or even empty, but this state
185  * is even better for immediate use provided the cache is not full.
186  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
187  * ones) to reach one of these states.  The watermarks are currently hard-
188  * coded as 4% and 9% of the available space higher.  These and the default
189  * of 25% for wantfreevnodes are too large if the memory size is large.
190  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
191  * whenever vnlru_proc() becomes active.
192  */
193 static u_long wantfreevnodes;
194 static u_long __exclusive_cache_line freevnodes;
195 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
196     &freevnodes, 0, "Number of \"free\" vnodes");
197 
198 static counter_u64_t recycles_count;
199 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
200     "Number of vnodes recycled to meet vnode cache targets");
201 
202 static counter_u64_t recycles_free_count;
203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
204     "Number of free vnodes recycled to meet vnode cache targets");
205 
206 /*
207  * Various variables used for debugging the new implementation of
208  * reassignbuf().
209  * XXX these are probably of (very) limited utility now.
210  */
211 static int reassignbufcalls;
212 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
213     &reassignbufcalls, 0, "Number of calls to reassignbuf");
214 
215 static counter_u64_t deferred_inact;
216 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
217     "Number of times inactive processing was deferred");
218 
219 /* To keep more than one thread at a time from running vfs_getnewfsid */
220 static struct mtx mntid_mtx;
221 
222 /*
223  * Lock for any access to the following:
224  *	vnode_list
225  *	numvnodes
226  *	freevnodes
227  */
228 static struct mtx __exclusive_cache_line vnode_list_mtx;
229 
230 /* Publicly exported FS */
231 struct nfs_public nfs_pub;
232 
233 static uma_zone_t buf_trie_zone;
234 
235 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
236 static uma_zone_t vnode_zone;
237 static uma_zone_t vnodepoll_zone;
238 
239 /*
240  * The workitem queue.
241  *
242  * It is useful to delay writes of file data and filesystem metadata
243  * for tens of seconds so that quickly created and deleted files need
244  * not waste disk bandwidth being created and removed. To realize this,
245  * we append vnodes to a "workitem" queue. When running with a soft
246  * updates implementation, most pending metadata dependencies should
247  * not wait for more than a few seconds. Thus, mounted on block devices
248  * are delayed only about a half the time that file data is delayed.
249  * Similarly, directory updates are more critical, so are only delayed
250  * about a third the time that file data is delayed. Thus, there are
251  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
252  * one each second (driven off the filesystem syncer process). The
253  * syncer_delayno variable indicates the next queue that is to be processed.
254  * Items that need to be processed soon are placed in this queue:
255  *
256  *	syncer_workitem_pending[syncer_delayno]
257  *
258  * A delay of fifteen seconds is done by placing the request fifteen
259  * entries later in the queue:
260  *
261  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
262  *
263  */
264 static int syncer_delayno;
265 static long syncer_mask;
266 LIST_HEAD(synclist, bufobj);
267 static struct synclist *syncer_workitem_pending;
268 /*
269  * The sync_mtx protects:
270  *	bo->bo_synclist
271  *	sync_vnode_count
272  *	syncer_delayno
273  *	syncer_state
274  *	syncer_workitem_pending
275  *	syncer_worklist_len
276  *	rushjob
277  */
278 static struct mtx sync_mtx;
279 static struct cv sync_wakeup;
280 
281 #define SYNCER_MAXDELAY		32
282 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
283 static int syncdelay = 30;		/* max time to delay syncing data */
284 static int filedelay = 30;		/* time to delay syncing files */
285 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
286     "Time to delay syncing files (in seconds)");
287 static int dirdelay = 29;		/* time to delay syncing directories */
288 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
289     "Time to delay syncing directories (in seconds)");
290 static int metadelay = 28;		/* time to delay syncing metadata */
291 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
292     "Time to delay syncing metadata (in seconds)");
293 static int rushjob;		/* number of slots to run ASAP */
294 static int stat_rush_requests;	/* number of times I/O speeded up */
295 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
296     "Number of times I/O speeded up (rush requests)");
297 
298 #define	VDBATCH_SIZE 8
299 struct vdbatch {
300 	u_int index;
301 	struct mtx lock;
302 	struct vnode *tab[VDBATCH_SIZE];
303 };
304 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
305 
306 static void	vdbatch_dequeue(struct vnode *vp);
307 
308 /*
309  * When shutting down the syncer, run it at four times normal speed.
310  */
311 #define SYNCER_SHUTDOWN_SPEEDUP		4
312 static int sync_vnode_count;
313 static int syncer_worklist_len;
314 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
315     syncer_state;
316 
317 /* Target for maximum number of vnodes. */
318 u_long desiredvnodes;
319 static u_long gapvnodes;		/* gap between wanted and desired */
320 static u_long vhiwat;		/* enough extras after expansion */
321 static u_long vlowat;		/* minimal extras before expansion */
322 static u_long vstir;		/* nonzero to stir non-free vnodes */
323 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
324 
325 /*
326  * Note that no attempt is made to sanitize these parameters.
327  */
328 static int
329 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
330 {
331 	u_long val;
332 	int error;
333 
334 	val = desiredvnodes;
335 	error = sysctl_handle_long(oidp, &val, 0, req);
336 	if (error != 0 || req->newptr == NULL)
337 		return (error);
338 
339 	if (val == desiredvnodes)
340 		return (0);
341 	mtx_lock(&vnode_list_mtx);
342 	desiredvnodes = val;
343 	wantfreevnodes = desiredvnodes / 4;
344 	vnlru_recalc();
345 	mtx_unlock(&vnode_list_mtx);
346 	/*
347 	 * XXX There is no protection against multiple threads changing
348 	 * desiredvnodes at the same time. Locking above only helps vnlru and
349 	 * getnewvnode.
350 	 */
351 	vfs_hash_changesize(desiredvnodes);
352 	cache_changesize(desiredvnodes);
353 	return (0);
354 }
355 
356 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
357     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
358     "UL", "Target for maximum number of vnodes");
359 
360 static int
361 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
362 {
363 	u_long val;
364 	int error;
365 
366 	val = wantfreevnodes;
367 	error = sysctl_handle_long(oidp, &val, 0, req);
368 	if (error != 0 || req->newptr == NULL)
369 		return (error);
370 
371 	if (val == wantfreevnodes)
372 		return (0);
373 	mtx_lock(&vnode_list_mtx);
374 	wantfreevnodes = val;
375 	vnlru_recalc();
376 	mtx_unlock(&vnode_list_mtx);
377 	return (0);
378 }
379 
380 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
381     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
382     "UL", "Target for minimum number of \"free\" vnodes");
383 
384 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
385     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
386 static int vnlru_nowhere;
387 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
388     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
389 
390 static int
391 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
392 {
393 	struct vnode *vp;
394 	struct nameidata nd;
395 	char *buf;
396 	unsigned long ndflags;
397 	int error;
398 
399 	if (req->newptr == NULL)
400 		return (EINVAL);
401 	if (req->newlen >= PATH_MAX)
402 		return (E2BIG);
403 
404 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
405 	error = SYSCTL_IN(req, buf, req->newlen);
406 	if (error != 0)
407 		goto out;
408 
409 	buf[req->newlen] = '\0';
410 
411 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
412 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
413 	if ((error = namei(&nd)) != 0)
414 		goto out;
415 	vp = nd.ni_vp;
416 
417 	if (VN_IS_DOOMED(vp)) {
418 		/*
419 		 * This vnode is being recycled.  Return != 0 to let the caller
420 		 * know that the sysctl had no effect.  Return EAGAIN because a
421 		 * subsequent call will likely succeed (since namei will create
422 		 * a new vnode if necessary)
423 		 */
424 		error = EAGAIN;
425 		goto putvnode;
426 	}
427 
428 	counter_u64_add(recycles_count, 1);
429 	vgone(vp);
430 putvnode:
431 	NDFREE(&nd, 0);
432 out:
433 	free(buf, M_TEMP);
434 	return (error);
435 }
436 
437 static int
438 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
439 {
440 	struct thread *td = curthread;
441 	struct vnode *vp;
442 	struct file *fp;
443 	int error;
444 	int fd;
445 
446 	if (req->newptr == NULL)
447 		return (EBADF);
448 
449         error = sysctl_handle_int(oidp, &fd, 0, req);
450         if (error != 0)
451                 return (error);
452 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
453 	if (error != 0)
454 		return (error);
455 	vp = fp->f_vnode;
456 
457 	error = vn_lock(vp, LK_EXCLUSIVE);
458 	if (error != 0)
459 		goto drop;
460 
461 	counter_u64_add(recycles_count, 1);
462 	vgone(vp);
463 	VOP_UNLOCK(vp);
464 drop:
465 	fdrop(fp, td);
466 	return (error);
467 }
468 
469 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
470     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
471     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
472 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
473     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
474     sysctl_ftry_reclaim_vnode, "I",
475     "Try to reclaim a vnode by its file descriptor");
476 
477 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
478 static int vnsz2log;
479 
480 /*
481  * Support for the bufobj clean & dirty pctrie.
482  */
483 static void *
484 buf_trie_alloc(struct pctrie *ptree)
485 {
486 
487 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
488 }
489 
490 static void
491 buf_trie_free(struct pctrie *ptree, void *node)
492 {
493 
494 	uma_zfree(buf_trie_zone, node);
495 }
496 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
497 
498 /*
499  * Initialize the vnode management data structures.
500  *
501  * Reevaluate the following cap on the number of vnodes after the physical
502  * memory size exceeds 512GB.  In the limit, as the physical memory size
503  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
504  */
505 #ifndef	MAXVNODES_MAX
506 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
507 #endif
508 
509 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
510 
511 static struct vnode *
512 vn_alloc_marker(struct mount *mp)
513 {
514 	struct vnode *vp;
515 
516 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
517 	vp->v_type = VMARKER;
518 	vp->v_mount = mp;
519 
520 	return (vp);
521 }
522 
523 static void
524 vn_free_marker(struct vnode *vp)
525 {
526 
527 	MPASS(vp->v_type == VMARKER);
528 	free(vp, M_VNODE_MARKER);
529 }
530 
531 /*
532  * Initialize a vnode as it first enters the zone.
533  */
534 static int
535 vnode_init(void *mem, int size, int flags)
536 {
537 	struct vnode *vp;
538 
539 	vp = mem;
540 	bzero(vp, size);
541 	/*
542 	 * Setup locks.
543 	 */
544 	vp->v_vnlock = &vp->v_lock;
545 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
546 	/*
547 	 * By default, don't allow shared locks unless filesystems opt-in.
548 	 */
549 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
550 	    LK_NOSHARE | LK_IS_VNODE);
551 	/*
552 	 * Initialize bufobj.
553 	 */
554 	bufobj_init(&vp->v_bufobj, vp);
555 	/*
556 	 * Initialize namecache.
557 	 */
558 	LIST_INIT(&vp->v_cache_src);
559 	TAILQ_INIT(&vp->v_cache_dst);
560 	/*
561 	 * Initialize rangelocks.
562 	 */
563 	rangelock_init(&vp->v_rl);
564 
565 	vp->v_dbatchcpu = NOCPU;
566 
567 	mtx_lock(&vnode_list_mtx);
568 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
569 	mtx_unlock(&vnode_list_mtx);
570 	return (0);
571 }
572 
573 /*
574  * Free a vnode when it is cleared from the zone.
575  */
576 static void
577 vnode_fini(void *mem, int size)
578 {
579 	struct vnode *vp;
580 	struct bufobj *bo;
581 
582 	vp = mem;
583 	vdbatch_dequeue(vp);
584 	mtx_lock(&vnode_list_mtx);
585 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
586 	mtx_unlock(&vnode_list_mtx);
587 	rangelock_destroy(&vp->v_rl);
588 	lockdestroy(vp->v_vnlock);
589 	mtx_destroy(&vp->v_interlock);
590 	bo = &vp->v_bufobj;
591 	rw_destroy(BO_LOCKPTR(bo));
592 }
593 
594 /*
595  * Provide the size of NFS nclnode and NFS fh for calculation of the
596  * vnode memory consumption.  The size is specified directly to
597  * eliminate dependency on NFS-private header.
598  *
599  * Other filesystems may use bigger or smaller (like UFS and ZFS)
600  * private inode data, but the NFS-based estimation is ample enough.
601  * Still, we care about differences in the size between 64- and 32-bit
602  * platforms.
603  *
604  * Namecache structure size is heuristically
605  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
606  */
607 #ifdef _LP64
608 #define	NFS_NCLNODE_SZ	(528 + 64)
609 #define	NC_SZ		148
610 #else
611 #define	NFS_NCLNODE_SZ	(360 + 32)
612 #define	NC_SZ		92
613 #endif
614 
615 static void
616 vntblinit(void *dummy __unused)
617 {
618 	struct vdbatch *vd;
619 	int cpu, physvnodes, virtvnodes;
620 	u_int i;
621 
622 	/*
623 	 * Desiredvnodes is a function of the physical memory size and the
624 	 * kernel's heap size.  Generally speaking, it scales with the
625 	 * physical memory size.  The ratio of desiredvnodes to the physical
626 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
627 	 * Thereafter, the
628 	 * marginal ratio of desiredvnodes to the physical memory size is
629 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
630 	 * size.  The memory required by desiredvnodes vnodes and vm objects
631 	 * must not exceed 1/10th of the kernel's heap size.
632 	 */
633 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
634 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
635 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
636 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
637 	desiredvnodes = min(physvnodes, virtvnodes);
638 	if (desiredvnodes > MAXVNODES_MAX) {
639 		if (bootverbose)
640 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
641 			    desiredvnodes, MAXVNODES_MAX);
642 		desiredvnodes = MAXVNODES_MAX;
643 	}
644 	wantfreevnodes = desiredvnodes / 4;
645 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
646 	TAILQ_INIT(&vnode_list);
647 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
648 	/*
649 	 * The lock is taken to appease WITNESS.
650 	 */
651 	mtx_lock(&vnode_list_mtx);
652 	vnlru_recalc();
653 	mtx_unlock(&vnode_list_mtx);
654 	vnode_list_free_marker = vn_alloc_marker(NULL);
655 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
656 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
657 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
658 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
659 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
660 	/*
661 	 * Preallocate enough nodes to support one-per buf so that
662 	 * we can not fail an insert.  reassignbuf() callers can not
663 	 * tolerate the insertion failure.
664 	 */
665 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
666 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
667 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
668 	uma_prealloc(buf_trie_zone, nbuf);
669 
670 	vnodes_created = counter_u64_alloc(M_WAITOK);
671 	recycles_count = counter_u64_alloc(M_WAITOK);
672 	recycles_free_count = counter_u64_alloc(M_WAITOK);
673 	deferred_inact = counter_u64_alloc(M_WAITOK);
674 
675 	/*
676 	 * Initialize the filesystem syncer.
677 	 */
678 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
679 	    &syncer_mask);
680 	syncer_maxdelay = syncer_mask + 1;
681 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
682 	cv_init(&sync_wakeup, "syncer");
683 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
684 		vnsz2log++;
685 	vnsz2log--;
686 
687 	CPU_FOREACH(cpu) {
688 		vd = DPCPU_ID_PTR((cpu), vd);
689 		bzero(vd, sizeof(*vd));
690 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
691 	}
692 }
693 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
694 
695 
696 /*
697  * Mark a mount point as busy. Used to synchronize access and to delay
698  * unmounting. Eventually, mountlist_mtx is not released on failure.
699  *
700  * vfs_busy() is a custom lock, it can block the caller.
701  * vfs_busy() only sleeps if the unmount is active on the mount point.
702  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
703  * vnode belonging to mp.
704  *
705  * Lookup uses vfs_busy() to traverse mount points.
706  * root fs			var fs
707  * / vnode lock		A	/ vnode lock (/var)		D
708  * /var vnode lock	B	/log vnode lock(/var/log)	E
709  * vfs_busy lock	C	vfs_busy lock			F
710  *
711  * Within each file system, the lock order is C->A->B and F->D->E.
712  *
713  * When traversing across mounts, the system follows that lock order:
714  *
715  *        C->A->B
716  *              |
717  *              +->F->D->E
718  *
719  * The lookup() process for namei("/var") illustrates the process:
720  *  VOP_LOOKUP() obtains B while A is held
721  *  vfs_busy() obtains a shared lock on F while A and B are held
722  *  vput() releases lock on B
723  *  vput() releases lock on A
724  *  VFS_ROOT() obtains lock on D while shared lock on F is held
725  *  vfs_unbusy() releases shared lock on F
726  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
727  *    Attempt to lock A (instead of vp_crossmp) while D is held would
728  *    violate the global order, causing deadlocks.
729  *
730  * dounmount() locks B while F is drained.
731  */
732 int
733 vfs_busy(struct mount *mp, int flags)
734 {
735 
736 	MPASS((flags & ~MBF_MASK) == 0);
737 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
738 
739 	if (vfs_op_thread_enter(mp)) {
740 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
741 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
742 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
743 		vfs_mp_count_add_pcpu(mp, ref, 1);
744 		vfs_mp_count_add_pcpu(mp, lockref, 1);
745 		vfs_op_thread_exit(mp);
746 		if (flags & MBF_MNTLSTLOCK)
747 			mtx_unlock(&mountlist_mtx);
748 		return (0);
749 	}
750 
751 	MNT_ILOCK(mp);
752 	vfs_assert_mount_counters(mp);
753 	MNT_REF(mp);
754 	/*
755 	 * If mount point is currently being unmounted, sleep until the
756 	 * mount point fate is decided.  If thread doing the unmounting fails,
757 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
758 	 * that this mount point has survived the unmount attempt and vfs_busy
759 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
760 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
761 	 * about to be really destroyed.  vfs_busy needs to release its
762 	 * reference on the mount point in this case and return with ENOENT,
763 	 * telling the caller that mount mount it tried to busy is no longer
764 	 * valid.
765 	 */
766 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
767 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
768 			MNT_REL(mp);
769 			MNT_IUNLOCK(mp);
770 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
771 			    __func__);
772 			return (ENOENT);
773 		}
774 		if (flags & MBF_MNTLSTLOCK)
775 			mtx_unlock(&mountlist_mtx);
776 		mp->mnt_kern_flag |= MNTK_MWAIT;
777 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
778 		if (flags & MBF_MNTLSTLOCK)
779 			mtx_lock(&mountlist_mtx);
780 		MNT_ILOCK(mp);
781 	}
782 	if (flags & MBF_MNTLSTLOCK)
783 		mtx_unlock(&mountlist_mtx);
784 	mp->mnt_lockref++;
785 	MNT_IUNLOCK(mp);
786 	return (0);
787 }
788 
789 /*
790  * Free a busy filesystem.
791  */
792 void
793 vfs_unbusy(struct mount *mp)
794 {
795 	int c;
796 
797 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
798 
799 	if (vfs_op_thread_enter(mp)) {
800 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
801 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
802 		vfs_mp_count_sub_pcpu(mp, ref, 1);
803 		vfs_op_thread_exit(mp);
804 		return;
805 	}
806 
807 	MNT_ILOCK(mp);
808 	vfs_assert_mount_counters(mp);
809 	MNT_REL(mp);
810 	c = --mp->mnt_lockref;
811 	if (mp->mnt_vfs_ops == 0) {
812 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
813 		MNT_IUNLOCK(mp);
814 		return;
815 	}
816 	if (c < 0)
817 		vfs_dump_mount_counters(mp);
818 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
819 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
820 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
821 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
822 		wakeup(&mp->mnt_lockref);
823 	}
824 	MNT_IUNLOCK(mp);
825 }
826 
827 /*
828  * Lookup a mount point by filesystem identifier.
829  */
830 struct mount *
831 vfs_getvfs(fsid_t *fsid)
832 {
833 	struct mount *mp;
834 
835 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
836 	mtx_lock(&mountlist_mtx);
837 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
838 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
839 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
840 			vfs_ref(mp);
841 			mtx_unlock(&mountlist_mtx);
842 			return (mp);
843 		}
844 	}
845 	mtx_unlock(&mountlist_mtx);
846 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
847 	return ((struct mount *) 0);
848 }
849 
850 /*
851  * Lookup a mount point by filesystem identifier, busying it before
852  * returning.
853  *
854  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
855  * cache for popular filesystem identifiers.  The cache is lockess, using
856  * the fact that struct mount's are never freed.  In worst case we may
857  * get pointer to unmounted or even different filesystem, so we have to
858  * check what we got, and go slow way if so.
859  */
860 struct mount *
861 vfs_busyfs(fsid_t *fsid)
862 {
863 #define	FSID_CACHE_SIZE	256
864 	typedef struct mount * volatile vmp_t;
865 	static vmp_t cache[FSID_CACHE_SIZE];
866 	struct mount *mp;
867 	int error;
868 	uint32_t hash;
869 
870 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
871 	hash = fsid->val[0] ^ fsid->val[1];
872 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
873 	mp = cache[hash];
874 	if (mp == NULL ||
875 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
876 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
877 		goto slow;
878 	if (vfs_busy(mp, 0) != 0) {
879 		cache[hash] = NULL;
880 		goto slow;
881 	}
882 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
883 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
884 		return (mp);
885 	else
886 	    vfs_unbusy(mp);
887 
888 slow:
889 	mtx_lock(&mountlist_mtx);
890 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
891 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
892 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
893 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
894 			if (error) {
895 				cache[hash] = NULL;
896 				mtx_unlock(&mountlist_mtx);
897 				return (NULL);
898 			}
899 			cache[hash] = mp;
900 			return (mp);
901 		}
902 	}
903 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
904 	mtx_unlock(&mountlist_mtx);
905 	return ((struct mount *) 0);
906 }
907 
908 /*
909  * Check if a user can access privileged mount options.
910  */
911 int
912 vfs_suser(struct mount *mp, struct thread *td)
913 {
914 	int error;
915 
916 	if (jailed(td->td_ucred)) {
917 		/*
918 		 * If the jail of the calling thread lacks permission for
919 		 * this type of file system, deny immediately.
920 		 */
921 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
922 			return (EPERM);
923 
924 		/*
925 		 * If the file system was mounted outside the jail of the
926 		 * calling thread, deny immediately.
927 		 */
928 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
929 			return (EPERM);
930 	}
931 
932 	/*
933 	 * If file system supports delegated administration, we don't check
934 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
935 	 * by the file system itself.
936 	 * If this is not the user that did original mount, we check for
937 	 * the PRIV_VFS_MOUNT_OWNER privilege.
938 	 */
939 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
940 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
941 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
942 			return (error);
943 	}
944 	return (0);
945 }
946 
947 /*
948  * Get a new unique fsid.  Try to make its val[0] unique, since this value
949  * will be used to create fake device numbers for stat().  Also try (but
950  * not so hard) make its val[0] unique mod 2^16, since some emulators only
951  * support 16-bit device numbers.  We end up with unique val[0]'s for the
952  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
953  *
954  * Keep in mind that several mounts may be running in parallel.  Starting
955  * the search one past where the previous search terminated is both a
956  * micro-optimization and a defense against returning the same fsid to
957  * different mounts.
958  */
959 void
960 vfs_getnewfsid(struct mount *mp)
961 {
962 	static uint16_t mntid_base;
963 	struct mount *nmp;
964 	fsid_t tfsid;
965 	int mtype;
966 
967 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
968 	mtx_lock(&mntid_mtx);
969 	mtype = mp->mnt_vfc->vfc_typenum;
970 	tfsid.val[1] = mtype;
971 	mtype = (mtype & 0xFF) << 24;
972 	for (;;) {
973 		tfsid.val[0] = makedev(255,
974 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
975 		mntid_base++;
976 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
977 			break;
978 		vfs_rel(nmp);
979 	}
980 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
981 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
982 	mtx_unlock(&mntid_mtx);
983 }
984 
985 /*
986  * Knob to control the precision of file timestamps:
987  *
988  *   0 = seconds only; nanoseconds zeroed.
989  *   1 = seconds and nanoseconds, accurate within 1/HZ.
990  *   2 = seconds and nanoseconds, truncated to microseconds.
991  * >=3 = seconds and nanoseconds, maximum precision.
992  */
993 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
994 
995 static int timestamp_precision = TSP_USEC;
996 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
997     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
998     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
999     "3+: sec + ns (max. precision))");
1000 
1001 /*
1002  * Get a current timestamp.
1003  */
1004 void
1005 vfs_timestamp(struct timespec *tsp)
1006 {
1007 	struct timeval tv;
1008 
1009 	switch (timestamp_precision) {
1010 	case TSP_SEC:
1011 		tsp->tv_sec = time_second;
1012 		tsp->tv_nsec = 0;
1013 		break;
1014 	case TSP_HZ:
1015 		getnanotime(tsp);
1016 		break;
1017 	case TSP_USEC:
1018 		microtime(&tv);
1019 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1020 		break;
1021 	case TSP_NSEC:
1022 	default:
1023 		nanotime(tsp);
1024 		break;
1025 	}
1026 }
1027 
1028 /*
1029  * Set vnode attributes to VNOVAL
1030  */
1031 void
1032 vattr_null(struct vattr *vap)
1033 {
1034 
1035 	vap->va_type = VNON;
1036 	vap->va_size = VNOVAL;
1037 	vap->va_bytes = VNOVAL;
1038 	vap->va_mode = VNOVAL;
1039 	vap->va_nlink = VNOVAL;
1040 	vap->va_uid = VNOVAL;
1041 	vap->va_gid = VNOVAL;
1042 	vap->va_fsid = VNOVAL;
1043 	vap->va_fileid = VNOVAL;
1044 	vap->va_blocksize = VNOVAL;
1045 	vap->va_rdev = VNOVAL;
1046 	vap->va_atime.tv_sec = VNOVAL;
1047 	vap->va_atime.tv_nsec = VNOVAL;
1048 	vap->va_mtime.tv_sec = VNOVAL;
1049 	vap->va_mtime.tv_nsec = VNOVAL;
1050 	vap->va_ctime.tv_sec = VNOVAL;
1051 	vap->va_ctime.tv_nsec = VNOVAL;
1052 	vap->va_birthtime.tv_sec = VNOVAL;
1053 	vap->va_birthtime.tv_nsec = VNOVAL;
1054 	vap->va_flags = VNOVAL;
1055 	vap->va_gen = VNOVAL;
1056 	vap->va_vaflags = 0;
1057 }
1058 
1059 /*
1060  * This routine is called when we have too many vnodes.  It attempts
1061  * to free <count> vnodes and will potentially free vnodes that still
1062  * have VM backing store (VM backing store is typically the cause
1063  * of a vnode blowout so we want to do this).  Therefore, this operation
1064  * is not considered cheap.
1065  *
1066  * A number of conditions may prevent a vnode from being reclaimed.
1067  * the buffer cache may have references on the vnode, a directory
1068  * vnode may still have references due to the namei cache representing
1069  * underlying files, or the vnode may be in active use.   It is not
1070  * desirable to reuse such vnodes.  These conditions may cause the
1071  * number of vnodes to reach some minimum value regardless of what
1072  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1073  *
1074  * @param mp		 Try to reclaim vnodes from this mountpoint
1075  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1076  * 			 entries if this argument is strue
1077  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1078  *			 pages.
1079  * @return		 The number of vnodes that were reclaimed.
1080  */
1081 static int
1082 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
1083 {
1084 	struct vnode *vp;
1085 	int count, done, target;
1086 
1087 	done = 0;
1088 	vn_start_write(NULL, &mp, V_WAIT);
1089 	MNT_ILOCK(mp);
1090 	count = mp->mnt_nvnodelistsize;
1091 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1092 	target = target / 10 + 1;
1093 	while (count != 0 && done < target) {
1094 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1095 		while (vp != NULL && vp->v_type == VMARKER)
1096 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
1097 		if (vp == NULL)
1098 			break;
1099 		/*
1100 		 * XXX LRU is completely broken for non-free vnodes.  First
1101 		 * by calling here in mountpoint order, then by moving
1102 		 * unselected vnodes to the end here, and most grossly by
1103 		 * removing the vlruvp() function that was supposed to
1104 		 * maintain the order.  (This function was born broken
1105 		 * since syncer problems prevented it doing anything.)  The
1106 		 * order is closer to LRC (C = Created).
1107 		 *
1108 		 * LRU reclaiming of vnodes seems to have last worked in
1109 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
1110 		 * Then there was no hold count, and inactive vnodes were
1111 		 * simply put on the free list in LRU order.  The separate
1112 		 * lists also break LRU.  We prefer to reclaim from the
1113 		 * free list for technical reasons.  This tends to thrash
1114 		 * the free list to keep very unrecently used held vnodes.
1115 		 * The problem is mitigated by keeping the free list large.
1116 		 */
1117 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1118 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1119 		--count;
1120 		if (!VI_TRYLOCK(vp))
1121 			goto next_iter;
1122 		/*
1123 		 * If it's been deconstructed already, it's still
1124 		 * referenced, or it exceeds the trigger, skip it.
1125 		 * Also skip free vnodes.  We are trying to make space
1126 		 * to expand the free list, not reduce it.
1127 		 */
1128 		if (vp->v_usecount ||
1129 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1130 		    vp->v_holdcnt == 0 ||
1131 		    VN_IS_DOOMED(vp) || (vp->v_object != NULL &&
1132 		    vp->v_object->resident_page_count > trigger)) {
1133 			VI_UNLOCK(vp);
1134 			goto next_iter;
1135 		}
1136 		MNT_IUNLOCK(mp);
1137 		vholdl(vp);
1138 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1139 			vdrop(vp);
1140 			goto next_iter_mntunlocked;
1141 		}
1142 		VI_LOCK(vp);
1143 		/*
1144 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1145 		 * the vnode interlock and before it was locked again.
1146 		 *
1147 		 * It is not necessary to recheck VIRF_DOOMED because it can
1148 		 * only be set by another thread that holds both the vnode
1149 		 * lock and vnode interlock.  If another thread has the
1150 		 * vnode lock before we get to VOP_LOCK() and obtains the
1151 		 * vnode interlock after VOP_LOCK() drops the vnode
1152 		 * interlock, the other thread will be unable to drop the
1153 		 * vnode lock before our VOP_LOCK() call fails.
1154 		 */
1155 		if (vp->v_usecount ||
1156 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1157 		    (vp->v_object != NULL &&
1158 		    vp->v_object->resident_page_count > trigger)) {
1159 			VOP_UNLOCK(vp);
1160 			vdropl(vp);
1161 			goto next_iter_mntunlocked;
1162 		}
1163 		KASSERT(!VN_IS_DOOMED(vp),
1164 		    ("VIRF_DOOMED unexpectedly detected in vlrureclaim()"));
1165 		counter_u64_add(recycles_count, 1);
1166 		vgonel(vp);
1167 		VOP_UNLOCK(vp);
1168 		vdropl(vp);
1169 		done++;
1170 next_iter_mntunlocked:
1171 		if (!should_yield())
1172 			goto relock_mnt;
1173 		goto yield;
1174 next_iter:
1175 		if (!should_yield())
1176 			continue;
1177 		MNT_IUNLOCK(mp);
1178 yield:
1179 		kern_yield(PRI_USER);
1180 relock_mnt:
1181 		MNT_ILOCK(mp);
1182 	}
1183 	MNT_IUNLOCK(mp);
1184 	vn_finished_write(mp);
1185 	return done;
1186 }
1187 
1188 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1189 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1190     0,
1191     "limit on vnode free requests per call to the vnlru_free routine");
1192 
1193 /*
1194  * Attempt to reduce the free list by the requested amount.
1195  */
1196 static void
1197 vnlru_free_locked(int count, struct vfsops *mnt_op)
1198 {
1199 	struct vnode *vp, *mvp;
1200 	struct mount *mp;
1201 
1202 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1203 	if (count > max_vnlru_free)
1204 		count = max_vnlru_free;
1205 	mvp = vnode_list_free_marker;
1206 restart:
1207 	vp = mvp;
1208 	while (count > 0) {
1209 		vp = TAILQ_NEXT(vp, v_vnodelist);
1210 		if (__predict_false(vp == NULL)) {
1211 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1212 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1213 			break;
1214 		}
1215 		if (__predict_false(vp->v_type == VMARKER))
1216 			continue;
1217 
1218 		/*
1219 		 * Don't recycle if our vnode is from different type
1220 		 * of mount point.  Note that mp is type-safe, the
1221 		 * check does not reach unmapped address even if
1222 		 * vnode is reclaimed.
1223 		 * Don't recycle if we can't get the interlock without
1224 		 * blocking.
1225 		 */
1226 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1227 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1228 			continue;
1229 		}
1230 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1231 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1232 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1233 			VI_UNLOCK(vp);
1234 			continue;
1235 		}
1236 		vholdl(vp);
1237 		count--;
1238 		mtx_unlock(&vnode_list_mtx);
1239 		VI_UNLOCK(vp);
1240 		vtryrecycle(vp);
1241 		vdrop(vp);
1242 		mtx_lock(&vnode_list_mtx);
1243 		goto restart;
1244 	}
1245 }
1246 
1247 void
1248 vnlru_free(int count, struct vfsops *mnt_op)
1249 {
1250 
1251 	mtx_lock(&vnode_list_mtx);
1252 	vnlru_free_locked(count, mnt_op);
1253 	mtx_unlock(&vnode_list_mtx);
1254 }
1255 
1256 static void
1257 vnlru_recalc(void)
1258 {
1259 
1260 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1261 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1262 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1263 	vlowat = vhiwat / 2;
1264 }
1265 
1266 /* XXX some names and initialization are bad for limits and watermarks. */
1267 static int
1268 vspace(void)
1269 {
1270 	u_long rnumvnodes, rfreevnodes;
1271 	int space;
1272 
1273 	rnumvnodes = atomic_load_long(&numvnodes);
1274 	rfreevnodes = atomic_load_long(&freevnodes);
1275 	if (rnumvnodes > desiredvnodes)
1276 		return (0);
1277 	space = desiredvnodes - rnumvnodes;
1278 	if (freevnodes > wantfreevnodes)
1279 		space += rfreevnodes - wantfreevnodes;
1280 	return (space);
1281 }
1282 
1283 /*
1284  * Attempt to recycle vnodes in a context that is always safe to block.
1285  * Calling vlrurecycle() from the bowels of filesystem code has some
1286  * interesting deadlock problems.
1287  */
1288 static struct proc *vnlruproc;
1289 static int vnlruproc_sig;
1290 
1291 static void
1292 vnlru_proc(void)
1293 {
1294 	u_long rnumvnodes, rfreevnodes;
1295 	struct mount *mp, *nmp;
1296 	unsigned long onumvnodes;
1297 	int done, force, trigger, usevnodes, vsp;
1298 	bool reclaim_nc_src;
1299 
1300 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1301 	    SHUTDOWN_PRI_FIRST);
1302 
1303 	force = 0;
1304 	for (;;) {
1305 		kproc_suspend_check(vnlruproc);
1306 		mtx_lock(&vnode_list_mtx);
1307 		rnumvnodes = atomic_load_long(&numvnodes);
1308 		/*
1309 		 * If numvnodes is too large (due to desiredvnodes being
1310 		 * adjusted using its sysctl, or emergency growth), first
1311 		 * try to reduce it by discarding from the free list.
1312 		 */
1313 		if (rnumvnodes > desiredvnodes)
1314 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1315 		/*
1316 		 * Sleep if the vnode cache is in a good state.  This is
1317 		 * when it is not over-full and has space for about a 4%
1318 		 * or 9% expansion (by growing its size or inexcessively
1319 		 * reducing its free list).  Otherwise, try to reclaim
1320 		 * space for a 10% expansion.
1321 		 */
1322 		if (vstir && force == 0) {
1323 			force = 1;
1324 			vstir = 0;
1325 		}
1326 		vsp = vspace();
1327 		if (vsp >= vlowat && force == 0) {
1328 			vnlruproc_sig = 0;
1329 			wakeup(&vnlruproc_sig);
1330 			msleep(vnlruproc, &vnode_list_mtx,
1331 			    PVFS|PDROP, "vlruwt", hz);
1332 			continue;
1333 		}
1334 		mtx_unlock(&vnode_list_mtx);
1335 		done = 0;
1336 		rnumvnodes = atomic_load_long(&numvnodes);
1337 		rfreevnodes = atomic_load_long(&freevnodes);
1338 
1339 		onumvnodes = rnumvnodes;
1340 		/*
1341 		 * Calculate parameters for recycling.  These are the same
1342 		 * throughout the loop to give some semblance of fairness.
1343 		 * The trigger point is to avoid recycling vnodes with lots
1344 		 * of resident pages.  We aren't trying to free memory; we
1345 		 * are trying to recycle or at least free vnodes.
1346 		 */
1347 		if (rnumvnodes <= desiredvnodes)
1348 			usevnodes = rnumvnodes - rfreevnodes;
1349 		else
1350 			usevnodes = rnumvnodes;
1351 		if (usevnodes <= 0)
1352 			usevnodes = 1;
1353 		/*
1354 		 * The trigger value is is chosen to give a conservatively
1355 		 * large value to ensure that it alone doesn't prevent
1356 		 * making progress.  The value can easily be so large that
1357 		 * it is effectively infinite in some congested and
1358 		 * misconfigured cases, and this is necessary.  Normally
1359 		 * it is about 8 to 100 (pages), which is quite large.
1360 		 */
1361 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1362 		if (force < 2)
1363 			trigger = vsmalltrigger;
1364 		reclaim_nc_src = force >= 3;
1365 		mtx_lock(&mountlist_mtx);
1366 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1367 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1368 				nmp = TAILQ_NEXT(mp, mnt_list);
1369 				continue;
1370 			}
1371 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1372 			mtx_lock(&mountlist_mtx);
1373 			nmp = TAILQ_NEXT(mp, mnt_list);
1374 			vfs_unbusy(mp);
1375 		}
1376 		mtx_unlock(&mountlist_mtx);
1377 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1378 			uma_reclaim(UMA_RECLAIM_DRAIN);
1379 		if (done == 0) {
1380 			if (force == 0 || force == 1) {
1381 				force = 2;
1382 				continue;
1383 			}
1384 			if (force == 2) {
1385 				force = 3;
1386 				continue;
1387 			}
1388 			force = 0;
1389 			vnlru_nowhere++;
1390 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1391 		} else
1392 			kern_yield(PRI_USER);
1393 		/*
1394 		 * After becoming active to expand above low water, keep
1395 		 * active until above high water.
1396 		 */
1397 		vsp = vspace();
1398 		force = vsp < vhiwat;
1399 	}
1400 }
1401 
1402 static struct kproc_desc vnlru_kp = {
1403 	"vnlru",
1404 	vnlru_proc,
1405 	&vnlruproc
1406 };
1407 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1408     &vnlru_kp);
1409 
1410 /*
1411  * Routines having to do with the management of the vnode table.
1412  */
1413 
1414 /*
1415  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1416  * before we actually vgone().  This function must be called with the vnode
1417  * held to prevent the vnode from being returned to the free list midway
1418  * through vgone().
1419  */
1420 static int
1421 vtryrecycle(struct vnode *vp)
1422 {
1423 	struct mount *vnmp;
1424 
1425 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1426 	VNASSERT(vp->v_holdcnt, vp,
1427 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1428 	/*
1429 	 * This vnode may found and locked via some other list, if so we
1430 	 * can't recycle it yet.
1431 	 */
1432 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1433 		CTR2(KTR_VFS,
1434 		    "%s: impossible to recycle, vp %p lock is already held",
1435 		    __func__, vp);
1436 		return (EWOULDBLOCK);
1437 	}
1438 	/*
1439 	 * Don't recycle if its filesystem is being suspended.
1440 	 */
1441 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1442 		VOP_UNLOCK(vp);
1443 		CTR2(KTR_VFS,
1444 		    "%s: impossible to recycle, cannot start the write for %p",
1445 		    __func__, vp);
1446 		return (EBUSY);
1447 	}
1448 	/*
1449 	 * If we got this far, we need to acquire the interlock and see if
1450 	 * anyone picked up this vnode from another list.  If not, we will
1451 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1452 	 * will skip over it.
1453 	 */
1454 	VI_LOCK(vp);
1455 	if (vp->v_usecount) {
1456 		VOP_UNLOCK(vp);
1457 		VI_UNLOCK(vp);
1458 		vn_finished_write(vnmp);
1459 		CTR2(KTR_VFS,
1460 		    "%s: impossible to recycle, %p is already referenced",
1461 		    __func__, vp);
1462 		return (EBUSY);
1463 	}
1464 	if (!VN_IS_DOOMED(vp)) {
1465 		counter_u64_add(recycles_free_count, 1);
1466 		vgonel(vp);
1467 	}
1468 	VOP_UNLOCK(vp);
1469 	VI_UNLOCK(vp);
1470 	vn_finished_write(vnmp);
1471 	return (0);
1472 }
1473 
1474 static void
1475 vcheckspace(void)
1476 {
1477 	int vsp;
1478 
1479 	vsp = vspace();
1480 	if (vsp < vlowat && vnlruproc_sig == 0) {
1481 		vnlruproc_sig = 1;
1482 		wakeup(vnlruproc);
1483 	}
1484 }
1485 
1486 /*
1487  * Wait if necessary for space for a new vnode.
1488  */
1489 static int
1490 vn_alloc_wait(int suspended)
1491 {
1492 
1493 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1494 	if (numvnodes >= desiredvnodes) {
1495 		if (suspended) {
1496 			/*
1497 			 * The file system is being suspended.  We cannot
1498 			 * risk a deadlock here, so allow allocation of
1499 			 * another vnode even if this would give too many.
1500 			 */
1501 			return (0);
1502 		}
1503 		if (vnlruproc_sig == 0) {
1504 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1505 			wakeup(vnlruproc);
1506 		}
1507 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS,
1508 		    "vlruwk", hz);
1509 	}
1510 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1511 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1512 		vnlru_free_locked(1, NULL);
1513 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1514 }
1515 
1516 static struct vnode *
1517 vn_alloc(struct mount *mp)
1518 {
1519 	static int cyclecount;
1520 	int error __unused;
1521 
1522 	mtx_lock(&vnode_list_mtx);
1523 	if (numvnodes < desiredvnodes)
1524 		cyclecount = 0;
1525 	else if (cyclecount++ >= freevnodes) {
1526 		cyclecount = 0;
1527 		vstir = 1;
1528 	}
1529 	/*
1530 	 * Grow the vnode cache if it will not be above its target max
1531 	 * after growing.  Otherwise, if the free list is nonempty, try
1532 	 * to reclaim 1 item from it before growing the cache (possibly
1533 	 * above its target max if the reclamation failed or is delayed).
1534 	 * Otherwise, wait for some space.  In all cases, schedule
1535 	 * vnlru_proc() if we are getting short of space.  The watermarks
1536 	 * should be chosen so that we never wait or even reclaim from
1537 	 * the free list to below its target minimum.
1538 	 */
1539 	if (numvnodes + 1 <= desiredvnodes)
1540 		;
1541 	else if (freevnodes > 0)
1542 		vnlru_free_locked(1, NULL);
1543 	else {
1544 		error = vn_alloc_wait(mp != NULL && (mp->mnt_kern_flag &
1545 		    MNTK_SUSPEND));
1546 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1547 		if (error != 0) {
1548 			mtx_unlock(&vnode_list_mtx);
1549 			return (error);
1550 		}
1551 #endif
1552 	}
1553 	vcheckspace();
1554 	atomic_add_long(&numvnodes, 1);
1555 	mtx_unlock(&vnode_list_mtx);
1556 	return (uma_zalloc(vnode_zone, M_WAITOK));
1557 }
1558 
1559 static void
1560 vn_free(struct vnode *vp)
1561 {
1562 
1563 	atomic_subtract_long(&numvnodes, 1);
1564 	uma_zfree(vnode_zone, vp);
1565 }
1566 
1567 /*
1568  * Return the next vnode from the free list.
1569  */
1570 int
1571 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1572     struct vnode **vpp)
1573 {
1574 	struct vnode *vp;
1575 	struct thread *td;
1576 	struct lock_object *lo;
1577 
1578 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1579 
1580 	KASSERT(vops->registered,
1581 	    ("%s: not registered vector op %p\n", __func__, vops));
1582 
1583 	td = curthread;
1584 	if (td->td_vp_reserved != NULL) {
1585 		vp = td->td_vp_reserved;
1586 		td->td_vp_reserved = NULL;
1587 	} else {
1588 		vp = vn_alloc(mp);
1589 	}
1590 	counter_u64_add(vnodes_created, 1);
1591 	/*
1592 	 * Locks are given the generic name "vnode" when created.
1593 	 * Follow the historic practice of using the filesystem
1594 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1595 	 *
1596 	 * Locks live in a witness group keyed on their name. Thus,
1597 	 * when a lock is renamed, it must also move from the witness
1598 	 * group of its old name to the witness group of its new name.
1599 	 *
1600 	 * The change only needs to be made when the vnode moves
1601 	 * from one filesystem type to another. We ensure that each
1602 	 * filesystem use a single static name pointer for its tag so
1603 	 * that we can compare pointers rather than doing a strcmp().
1604 	 */
1605 	lo = &vp->v_vnlock->lock_object;
1606 	if (lo->lo_name != tag) {
1607 		lo->lo_name = tag;
1608 		WITNESS_DESTROY(lo);
1609 		WITNESS_INIT(lo, tag);
1610 	}
1611 	/*
1612 	 * By default, don't allow shared locks unless filesystems opt-in.
1613 	 */
1614 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1615 	/*
1616 	 * Finalize various vnode identity bits.
1617 	 */
1618 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1619 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1620 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1621 	vp->v_type = VNON;
1622 	vp->v_op = vops;
1623 	v_init_counters(vp);
1624 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1625 #ifdef DIAGNOSTIC
1626 	if (mp == NULL && vops != &dead_vnodeops)
1627 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1628 #endif
1629 #ifdef MAC
1630 	mac_vnode_init(vp);
1631 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1632 		mac_vnode_associate_singlelabel(mp, vp);
1633 #endif
1634 	if (mp != NULL) {
1635 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1636 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1637 			vp->v_vflag |= VV_NOKNOTE;
1638 	}
1639 
1640 	/*
1641 	 * For the filesystems which do not use vfs_hash_insert(),
1642 	 * still initialize v_hash to have vfs_hash_index() useful.
1643 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1644 	 * its own hashing.
1645 	 */
1646 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1647 
1648 	*vpp = vp;
1649 	return (0);
1650 }
1651 
1652 void
1653 getnewvnode_reserve(void)
1654 {
1655 	struct thread *td;
1656 
1657 	td = curthread;
1658 	MPASS(td->td_vp_reserved == NULL);
1659 	td->td_vp_reserved = vn_alloc(NULL);
1660 }
1661 
1662 void
1663 getnewvnode_drop_reserve(void)
1664 {
1665 	struct thread *td;
1666 
1667 	td = curthread;
1668 	if (td->td_vp_reserved != NULL) {
1669 		vn_free(td->td_vp_reserved);
1670 		td->td_vp_reserved = NULL;
1671 	}
1672 }
1673 
1674 static void
1675 freevnode(struct vnode *vp)
1676 {
1677 	struct bufobj *bo;
1678 
1679 	/*
1680 	 * The vnode has been marked for destruction, so free it.
1681 	 *
1682 	 * The vnode will be returned to the zone where it will
1683 	 * normally remain until it is needed for another vnode. We
1684 	 * need to cleanup (or verify that the cleanup has already
1685 	 * been done) any residual data left from its current use
1686 	 * so as not to contaminate the freshly allocated vnode.
1687 	 */
1688 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1689 	bo = &vp->v_bufobj;
1690 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1691 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1692 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1693 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1694 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1695 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1696 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1697 	    ("clean blk trie not empty"));
1698 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1699 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1700 	    ("dirty blk trie not empty"));
1701 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1702 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1703 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1704 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1705 	    ("Dangling rangelock waiters"));
1706 	VI_UNLOCK(vp);
1707 #ifdef MAC
1708 	mac_vnode_destroy(vp);
1709 #endif
1710 	if (vp->v_pollinfo != NULL) {
1711 		destroy_vpollinfo(vp->v_pollinfo);
1712 		vp->v_pollinfo = NULL;
1713 	}
1714 #ifdef INVARIANTS
1715 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1716 	vp->v_op = NULL;
1717 #endif
1718 	vp->v_mountedhere = NULL;
1719 	vp->v_unpcb = NULL;
1720 	vp->v_rdev = NULL;
1721 	vp->v_fifoinfo = NULL;
1722 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1723 	vp->v_irflag = 0;
1724 	vp->v_iflag = 0;
1725 	vp->v_vflag = 0;
1726 	bo->bo_flag = 0;
1727 	vn_free(vp);
1728 }
1729 
1730 /*
1731  * Delete from old mount point vnode list, if on one.
1732  */
1733 static void
1734 delmntque(struct vnode *vp)
1735 {
1736 	struct mount *mp;
1737 
1738 	mp = vp->v_mount;
1739 	if (mp == NULL)
1740 		return;
1741 	MNT_ILOCK(mp);
1742 	VI_LOCK(vp);
1743 	if (vp->v_mflag & VMP_LAZYLIST) {
1744 		mtx_lock(&mp->mnt_listmtx);
1745 		if (vp->v_mflag & VMP_LAZYLIST) {
1746 			vp->v_mflag &= ~VMP_LAZYLIST;
1747 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
1748 			mp->mnt_lazyvnodelistsize--;
1749 		}
1750 		mtx_unlock(&mp->mnt_listmtx);
1751 	}
1752 	vp->v_mount = NULL;
1753 	VI_UNLOCK(vp);
1754 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1755 		("bad mount point vnode list size"));
1756 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1757 	mp->mnt_nvnodelistsize--;
1758 	MNT_REL(mp);
1759 	MNT_IUNLOCK(mp);
1760 }
1761 
1762 static void
1763 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1764 {
1765 
1766 	vp->v_data = NULL;
1767 	vp->v_op = &dead_vnodeops;
1768 	vgone(vp);
1769 	vput(vp);
1770 }
1771 
1772 /*
1773  * Insert into list of vnodes for the new mount point, if available.
1774  */
1775 int
1776 insmntque1(struct vnode *vp, struct mount *mp,
1777 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1778 {
1779 
1780 	KASSERT(vp->v_mount == NULL,
1781 		("insmntque: vnode already on per mount vnode list"));
1782 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1783 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1784 
1785 	/*
1786 	 * We acquire the vnode interlock early to ensure that the
1787 	 * vnode cannot be recycled by another process releasing a
1788 	 * holdcnt on it before we get it on both the vnode list
1789 	 * and the active vnode list. The mount mutex protects only
1790 	 * manipulation of the vnode list and the vnode freelist
1791 	 * mutex protects only manipulation of the active vnode list.
1792 	 * Hence the need to hold the vnode interlock throughout.
1793 	 */
1794 	MNT_ILOCK(mp);
1795 	VI_LOCK(vp);
1796 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1797 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1798 	    mp->mnt_nvnodelistsize == 0)) &&
1799 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1800 		VI_UNLOCK(vp);
1801 		MNT_IUNLOCK(mp);
1802 		if (dtr != NULL)
1803 			dtr(vp, dtr_arg);
1804 		return (EBUSY);
1805 	}
1806 	vp->v_mount = mp;
1807 	MNT_REF(mp);
1808 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1809 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1810 		("neg mount point vnode list size"));
1811 	mp->mnt_nvnodelistsize++;
1812 	VI_UNLOCK(vp);
1813 	MNT_IUNLOCK(mp);
1814 	return (0);
1815 }
1816 
1817 int
1818 insmntque(struct vnode *vp, struct mount *mp)
1819 {
1820 
1821 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1822 }
1823 
1824 /*
1825  * Flush out and invalidate all buffers associated with a bufobj
1826  * Called with the underlying object locked.
1827  */
1828 int
1829 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1830 {
1831 	int error;
1832 
1833 	BO_LOCK(bo);
1834 	if (flags & V_SAVE) {
1835 		error = bufobj_wwait(bo, slpflag, slptimeo);
1836 		if (error) {
1837 			BO_UNLOCK(bo);
1838 			return (error);
1839 		}
1840 		if (bo->bo_dirty.bv_cnt > 0) {
1841 			BO_UNLOCK(bo);
1842 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1843 				return (error);
1844 			/*
1845 			 * XXX We could save a lock/unlock if this was only
1846 			 * enabled under INVARIANTS
1847 			 */
1848 			BO_LOCK(bo);
1849 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1850 				panic("vinvalbuf: dirty bufs");
1851 		}
1852 	}
1853 	/*
1854 	 * If you alter this loop please notice that interlock is dropped and
1855 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1856 	 * no race conditions occur from this.
1857 	 */
1858 	do {
1859 		error = flushbuflist(&bo->bo_clean,
1860 		    flags, bo, slpflag, slptimeo);
1861 		if (error == 0 && !(flags & V_CLEANONLY))
1862 			error = flushbuflist(&bo->bo_dirty,
1863 			    flags, bo, slpflag, slptimeo);
1864 		if (error != 0 && error != EAGAIN) {
1865 			BO_UNLOCK(bo);
1866 			return (error);
1867 		}
1868 	} while (error != 0);
1869 
1870 	/*
1871 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1872 	 * have write I/O in-progress but if there is a VM object then the
1873 	 * VM object can also have read-I/O in-progress.
1874 	 */
1875 	do {
1876 		bufobj_wwait(bo, 0, 0);
1877 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1878 			BO_UNLOCK(bo);
1879 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1880 			BO_LOCK(bo);
1881 		}
1882 	} while (bo->bo_numoutput > 0);
1883 	BO_UNLOCK(bo);
1884 
1885 	/*
1886 	 * Destroy the copy in the VM cache, too.
1887 	 */
1888 	if (bo->bo_object != NULL &&
1889 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1890 		VM_OBJECT_WLOCK(bo->bo_object);
1891 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1892 		    OBJPR_CLEANONLY : 0);
1893 		VM_OBJECT_WUNLOCK(bo->bo_object);
1894 	}
1895 
1896 #ifdef INVARIANTS
1897 	BO_LOCK(bo);
1898 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1899 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1900 	    bo->bo_clean.bv_cnt > 0))
1901 		panic("vinvalbuf: flush failed");
1902 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1903 	    bo->bo_dirty.bv_cnt > 0)
1904 		panic("vinvalbuf: flush dirty failed");
1905 	BO_UNLOCK(bo);
1906 #endif
1907 	return (0);
1908 }
1909 
1910 /*
1911  * Flush out and invalidate all buffers associated with a vnode.
1912  * Called with the underlying object locked.
1913  */
1914 int
1915 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1916 {
1917 
1918 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1919 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1920 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1921 		return (0);
1922 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1923 }
1924 
1925 /*
1926  * Flush out buffers on the specified list.
1927  *
1928  */
1929 static int
1930 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1931     int slptimeo)
1932 {
1933 	struct buf *bp, *nbp;
1934 	int retval, error;
1935 	daddr_t lblkno;
1936 	b_xflags_t xflags;
1937 
1938 	ASSERT_BO_WLOCKED(bo);
1939 
1940 	retval = 0;
1941 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1942 		/*
1943 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1944 		 * do not skip any buffers. If we are flushing only V_NORMAL
1945 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1946 		 * flushing only V_ALT buffers then skip buffers not marked
1947 		 * as BX_ALTDATA.
1948 		 */
1949 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1950 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1951 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1952 			continue;
1953 		}
1954 		if (nbp != NULL) {
1955 			lblkno = nbp->b_lblkno;
1956 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1957 		}
1958 		retval = EAGAIN;
1959 		error = BUF_TIMELOCK(bp,
1960 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1961 		    "flushbuf", slpflag, slptimeo);
1962 		if (error) {
1963 			BO_LOCK(bo);
1964 			return (error != ENOLCK ? error : EAGAIN);
1965 		}
1966 		KASSERT(bp->b_bufobj == bo,
1967 		    ("bp %p wrong b_bufobj %p should be %p",
1968 		    bp, bp->b_bufobj, bo));
1969 		/*
1970 		 * XXX Since there are no node locks for NFS, I
1971 		 * believe there is a slight chance that a delayed
1972 		 * write will occur while sleeping just above, so
1973 		 * check for it.
1974 		 */
1975 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1976 		    (flags & V_SAVE)) {
1977 			bremfree(bp);
1978 			bp->b_flags |= B_ASYNC;
1979 			bwrite(bp);
1980 			BO_LOCK(bo);
1981 			return (EAGAIN);	/* XXX: why not loop ? */
1982 		}
1983 		bremfree(bp);
1984 		bp->b_flags |= (B_INVAL | B_RELBUF);
1985 		bp->b_flags &= ~B_ASYNC;
1986 		brelse(bp);
1987 		BO_LOCK(bo);
1988 		if (nbp == NULL)
1989 			break;
1990 		nbp = gbincore(bo, lblkno);
1991 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1992 		    != xflags)
1993 			break;			/* nbp invalid */
1994 	}
1995 	return (retval);
1996 }
1997 
1998 int
1999 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2000 {
2001 	struct buf *bp;
2002 	int error;
2003 	daddr_t lblkno;
2004 
2005 	ASSERT_BO_LOCKED(bo);
2006 
2007 	for (lblkno = startn;;) {
2008 again:
2009 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2010 		if (bp == NULL || bp->b_lblkno >= endn ||
2011 		    bp->b_lblkno < startn)
2012 			break;
2013 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2014 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2015 		if (error != 0) {
2016 			BO_RLOCK(bo);
2017 			if (error == ENOLCK)
2018 				goto again;
2019 			return (error);
2020 		}
2021 		KASSERT(bp->b_bufobj == bo,
2022 		    ("bp %p wrong b_bufobj %p should be %p",
2023 		    bp, bp->b_bufobj, bo));
2024 		lblkno = bp->b_lblkno + 1;
2025 		if ((bp->b_flags & B_MANAGED) == 0)
2026 			bremfree(bp);
2027 		bp->b_flags |= B_RELBUF;
2028 		/*
2029 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2030 		 * pages backing each buffer in the range are unlikely to be
2031 		 * reused.  Dirty buffers will have the hint applied once
2032 		 * they've been written.
2033 		 */
2034 		if ((bp->b_flags & B_VMIO) != 0)
2035 			bp->b_flags |= B_NOREUSE;
2036 		brelse(bp);
2037 		BO_RLOCK(bo);
2038 	}
2039 	return (0);
2040 }
2041 
2042 /*
2043  * Truncate a file's buffer and pages to a specified length.  This
2044  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2045  * sync activity.
2046  */
2047 int
2048 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2049 {
2050 	struct buf *bp, *nbp;
2051 	struct bufobj *bo;
2052 	daddr_t startlbn;
2053 
2054 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2055 	    vp, blksize, (uintmax_t)length);
2056 
2057 	/*
2058 	 * Round up to the *next* lbn.
2059 	 */
2060 	startlbn = howmany(length, blksize);
2061 
2062 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2063 
2064 	bo = &vp->v_bufobj;
2065 restart_unlocked:
2066 	BO_LOCK(bo);
2067 
2068 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2069 		;
2070 
2071 	if (length > 0) {
2072 restartsync:
2073 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2074 			if (bp->b_lblkno > 0)
2075 				continue;
2076 			/*
2077 			 * Since we hold the vnode lock this should only
2078 			 * fail if we're racing with the buf daemon.
2079 			 */
2080 			if (BUF_LOCK(bp,
2081 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2082 			    BO_LOCKPTR(bo)) == ENOLCK)
2083 				goto restart_unlocked;
2084 
2085 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2086 			    ("buf(%p) on dirty queue without DELWRI", bp));
2087 
2088 			bremfree(bp);
2089 			bawrite(bp);
2090 			BO_LOCK(bo);
2091 			goto restartsync;
2092 		}
2093 	}
2094 
2095 	bufobj_wwait(bo, 0, 0);
2096 	BO_UNLOCK(bo);
2097 	vnode_pager_setsize(vp, length);
2098 
2099 	return (0);
2100 }
2101 
2102 /*
2103  * Invalidate the cached pages of a file's buffer within the range of block
2104  * numbers [startlbn, endlbn).
2105  */
2106 void
2107 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2108     int blksize)
2109 {
2110 	struct bufobj *bo;
2111 	off_t start, end;
2112 
2113 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2114 
2115 	start = blksize * startlbn;
2116 	end = blksize * endlbn;
2117 
2118 	bo = &vp->v_bufobj;
2119 	BO_LOCK(bo);
2120 	MPASS(blksize == bo->bo_bsize);
2121 
2122 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2123 		;
2124 
2125 	BO_UNLOCK(bo);
2126 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2127 }
2128 
2129 static int
2130 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2131     daddr_t startlbn, daddr_t endlbn)
2132 {
2133 	struct buf *bp, *nbp;
2134 	bool anyfreed;
2135 
2136 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2137 	ASSERT_BO_LOCKED(bo);
2138 
2139 	do {
2140 		anyfreed = false;
2141 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2142 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2143 				continue;
2144 			if (BUF_LOCK(bp,
2145 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2146 			    BO_LOCKPTR(bo)) == ENOLCK) {
2147 				BO_LOCK(bo);
2148 				return (EAGAIN);
2149 			}
2150 
2151 			bremfree(bp);
2152 			bp->b_flags |= B_INVAL | B_RELBUF;
2153 			bp->b_flags &= ~B_ASYNC;
2154 			brelse(bp);
2155 			anyfreed = true;
2156 
2157 			BO_LOCK(bo);
2158 			if (nbp != NULL &&
2159 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2160 			    nbp->b_vp != vp ||
2161 			    (nbp->b_flags & B_DELWRI) != 0))
2162 				return (EAGAIN);
2163 		}
2164 
2165 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2166 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2167 				continue;
2168 			if (BUF_LOCK(bp,
2169 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2170 			    BO_LOCKPTR(bo)) == ENOLCK) {
2171 				BO_LOCK(bo);
2172 				return (EAGAIN);
2173 			}
2174 			bremfree(bp);
2175 			bp->b_flags |= B_INVAL | B_RELBUF;
2176 			bp->b_flags &= ~B_ASYNC;
2177 			brelse(bp);
2178 			anyfreed = true;
2179 
2180 			BO_LOCK(bo);
2181 			if (nbp != NULL &&
2182 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2183 			    (nbp->b_vp != vp) ||
2184 			    (nbp->b_flags & B_DELWRI) == 0))
2185 				return (EAGAIN);
2186 		}
2187 	} while (anyfreed);
2188 	return (0);
2189 }
2190 
2191 static void
2192 buf_vlist_remove(struct buf *bp)
2193 {
2194 	struct bufv *bv;
2195 
2196 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2197 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2198 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2199 	    (BX_VNDIRTY|BX_VNCLEAN),
2200 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2201 	if (bp->b_xflags & BX_VNDIRTY)
2202 		bv = &bp->b_bufobj->bo_dirty;
2203 	else
2204 		bv = &bp->b_bufobj->bo_clean;
2205 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2206 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2207 	bv->bv_cnt--;
2208 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2209 }
2210 
2211 /*
2212  * Add the buffer to the sorted clean or dirty block list.
2213  *
2214  * NOTE: xflags is passed as a constant, optimizing this inline function!
2215  */
2216 static void
2217 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2218 {
2219 	struct bufv *bv;
2220 	struct buf *n;
2221 	int error;
2222 
2223 	ASSERT_BO_WLOCKED(bo);
2224 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2225 	    ("dead bo %p", bo));
2226 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2227 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2228 	bp->b_xflags |= xflags;
2229 	if (xflags & BX_VNDIRTY)
2230 		bv = &bo->bo_dirty;
2231 	else
2232 		bv = &bo->bo_clean;
2233 
2234 	/*
2235 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2236 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2237 	 * than _ge.
2238 	 */
2239 	if (bv->bv_cnt == 0 ||
2240 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2241 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2242 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2243 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2244 	else
2245 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2246 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2247 	if (error)
2248 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2249 	bv->bv_cnt++;
2250 }
2251 
2252 /*
2253  * Look up a buffer using the buffer tries.
2254  */
2255 struct buf *
2256 gbincore(struct bufobj *bo, daddr_t lblkno)
2257 {
2258 	struct buf *bp;
2259 
2260 	ASSERT_BO_LOCKED(bo);
2261 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2262 	if (bp != NULL)
2263 		return (bp);
2264 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2265 }
2266 
2267 /*
2268  * Associate a buffer with a vnode.
2269  */
2270 void
2271 bgetvp(struct vnode *vp, struct buf *bp)
2272 {
2273 	struct bufobj *bo;
2274 
2275 	bo = &vp->v_bufobj;
2276 	ASSERT_BO_WLOCKED(bo);
2277 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2278 
2279 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2280 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2281 	    ("bgetvp: bp already attached! %p", bp));
2282 
2283 	vhold(vp);
2284 	bp->b_vp = vp;
2285 	bp->b_bufobj = bo;
2286 	/*
2287 	 * Insert onto list for new vnode.
2288 	 */
2289 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2290 }
2291 
2292 /*
2293  * Disassociate a buffer from a vnode.
2294  */
2295 void
2296 brelvp(struct buf *bp)
2297 {
2298 	struct bufobj *bo;
2299 	struct vnode *vp;
2300 
2301 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2302 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2303 
2304 	/*
2305 	 * Delete from old vnode list, if on one.
2306 	 */
2307 	vp = bp->b_vp;		/* XXX */
2308 	bo = bp->b_bufobj;
2309 	BO_LOCK(bo);
2310 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2311 		buf_vlist_remove(bp);
2312 	else
2313 		panic("brelvp: Buffer %p not on queue.", bp);
2314 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2315 		bo->bo_flag &= ~BO_ONWORKLST;
2316 		mtx_lock(&sync_mtx);
2317 		LIST_REMOVE(bo, bo_synclist);
2318 		syncer_worklist_len--;
2319 		mtx_unlock(&sync_mtx);
2320 	}
2321 	bp->b_vp = NULL;
2322 	bp->b_bufobj = NULL;
2323 	BO_UNLOCK(bo);
2324 	vdrop(vp);
2325 }
2326 
2327 /*
2328  * Add an item to the syncer work queue.
2329  */
2330 static void
2331 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2332 {
2333 	int slot;
2334 
2335 	ASSERT_BO_WLOCKED(bo);
2336 
2337 	mtx_lock(&sync_mtx);
2338 	if (bo->bo_flag & BO_ONWORKLST)
2339 		LIST_REMOVE(bo, bo_synclist);
2340 	else {
2341 		bo->bo_flag |= BO_ONWORKLST;
2342 		syncer_worklist_len++;
2343 	}
2344 
2345 	if (delay > syncer_maxdelay - 2)
2346 		delay = syncer_maxdelay - 2;
2347 	slot = (syncer_delayno + delay) & syncer_mask;
2348 
2349 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2350 	mtx_unlock(&sync_mtx);
2351 }
2352 
2353 static int
2354 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2355 {
2356 	int error, len;
2357 
2358 	mtx_lock(&sync_mtx);
2359 	len = syncer_worklist_len - sync_vnode_count;
2360 	mtx_unlock(&sync_mtx);
2361 	error = SYSCTL_OUT(req, &len, sizeof(len));
2362 	return (error);
2363 }
2364 
2365 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2366     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2367     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2368 
2369 static struct proc *updateproc;
2370 static void sched_sync(void);
2371 static struct kproc_desc up_kp = {
2372 	"syncer",
2373 	sched_sync,
2374 	&updateproc
2375 };
2376 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2377 
2378 static int
2379 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2380 {
2381 	struct vnode *vp;
2382 	struct mount *mp;
2383 
2384 	*bo = LIST_FIRST(slp);
2385 	if (*bo == NULL)
2386 		return (0);
2387 	vp = bo2vnode(*bo);
2388 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2389 		return (1);
2390 	/*
2391 	 * We use vhold in case the vnode does not
2392 	 * successfully sync.  vhold prevents the vnode from
2393 	 * going away when we unlock the sync_mtx so that
2394 	 * we can acquire the vnode interlock.
2395 	 */
2396 	vholdl(vp);
2397 	mtx_unlock(&sync_mtx);
2398 	VI_UNLOCK(vp);
2399 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2400 		vdrop(vp);
2401 		mtx_lock(&sync_mtx);
2402 		return (*bo == LIST_FIRST(slp));
2403 	}
2404 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2405 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2406 	VOP_UNLOCK(vp);
2407 	vn_finished_write(mp);
2408 	BO_LOCK(*bo);
2409 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2410 		/*
2411 		 * Put us back on the worklist.  The worklist
2412 		 * routine will remove us from our current
2413 		 * position and then add us back in at a later
2414 		 * position.
2415 		 */
2416 		vn_syncer_add_to_worklist(*bo, syncdelay);
2417 	}
2418 	BO_UNLOCK(*bo);
2419 	vdrop(vp);
2420 	mtx_lock(&sync_mtx);
2421 	return (0);
2422 }
2423 
2424 static int first_printf = 1;
2425 
2426 /*
2427  * System filesystem synchronizer daemon.
2428  */
2429 static void
2430 sched_sync(void)
2431 {
2432 	struct synclist *next, *slp;
2433 	struct bufobj *bo;
2434 	long starttime;
2435 	struct thread *td = curthread;
2436 	int last_work_seen;
2437 	int net_worklist_len;
2438 	int syncer_final_iter;
2439 	int error;
2440 
2441 	last_work_seen = 0;
2442 	syncer_final_iter = 0;
2443 	syncer_state = SYNCER_RUNNING;
2444 	starttime = time_uptime;
2445 	td->td_pflags |= TDP_NORUNNINGBUF;
2446 
2447 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2448 	    SHUTDOWN_PRI_LAST);
2449 
2450 	mtx_lock(&sync_mtx);
2451 	for (;;) {
2452 		if (syncer_state == SYNCER_FINAL_DELAY &&
2453 		    syncer_final_iter == 0) {
2454 			mtx_unlock(&sync_mtx);
2455 			kproc_suspend_check(td->td_proc);
2456 			mtx_lock(&sync_mtx);
2457 		}
2458 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2459 		if (syncer_state != SYNCER_RUNNING &&
2460 		    starttime != time_uptime) {
2461 			if (first_printf) {
2462 				printf("\nSyncing disks, vnodes remaining... ");
2463 				first_printf = 0;
2464 			}
2465 			printf("%d ", net_worklist_len);
2466 		}
2467 		starttime = time_uptime;
2468 
2469 		/*
2470 		 * Push files whose dirty time has expired.  Be careful
2471 		 * of interrupt race on slp queue.
2472 		 *
2473 		 * Skip over empty worklist slots when shutting down.
2474 		 */
2475 		do {
2476 			slp = &syncer_workitem_pending[syncer_delayno];
2477 			syncer_delayno += 1;
2478 			if (syncer_delayno == syncer_maxdelay)
2479 				syncer_delayno = 0;
2480 			next = &syncer_workitem_pending[syncer_delayno];
2481 			/*
2482 			 * If the worklist has wrapped since the
2483 			 * it was emptied of all but syncer vnodes,
2484 			 * switch to the FINAL_DELAY state and run
2485 			 * for one more second.
2486 			 */
2487 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2488 			    net_worklist_len == 0 &&
2489 			    last_work_seen == syncer_delayno) {
2490 				syncer_state = SYNCER_FINAL_DELAY;
2491 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2492 			}
2493 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2494 		    syncer_worklist_len > 0);
2495 
2496 		/*
2497 		 * Keep track of the last time there was anything
2498 		 * on the worklist other than syncer vnodes.
2499 		 * Return to the SHUTTING_DOWN state if any
2500 		 * new work appears.
2501 		 */
2502 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2503 			last_work_seen = syncer_delayno;
2504 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2505 			syncer_state = SYNCER_SHUTTING_DOWN;
2506 		while (!LIST_EMPTY(slp)) {
2507 			error = sync_vnode(slp, &bo, td);
2508 			if (error == 1) {
2509 				LIST_REMOVE(bo, bo_synclist);
2510 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2511 				continue;
2512 			}
2513 
2514 			if (first_printf == 0) {
2515 				/*
2516 				 * Drop the sync mutex, because some watchdog
2517 				 * drivers need to sleep while patting
2518 				 */
2519 				mtx_unlock(&sync_mtx);
2520 				wdog_kern_pat(WD_LASTVAL);
2521 				mtx_lock(&sync_mtx);
2522 			}
2523 
2524 		}
2525 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2526 			syncer_final_iter--;
2527 		/*
2528 		 * The variable rushjob allows the kernel to speed up the
2529 		 * processing of the filesystem syncer process. A rushjob
2530 		 * value of N tells the filesystem syncer to process the next
2531 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2532 		 * is used by the soft update code to speed up the filesystem
2533 		 * syncer process when the incore state is getting so far
2534 		 * ahead of the disk that the kernel memory pool is being
2535 		 * threatened with exhaustion.
2536 		 */
2537 		if (rushjob > 0) {
2538 			rushjob -= 1;
2539 			continue;
2540 		}
2541 		/*
2542 		 * Just sleep for a short period of time between
2543 		 * iterations when shutting down to allow some I/O
2544 		 * to happen.
2545 		 *
2546 		 * If it has taken us less than a second to process the
2547 		 * current work, then wait. Otherwise start right over
2548 		 * again. We can still lose time if any single round
2549 		 * takes more than two seconds, but it does not really
2550 		 * matter as we are just trying to generally pace the
2551 		 * filesystem activity.
2552 		 */
2553 		if (syncer_state != SYNCER_RUNNING ||
2554 		    time_uptime == starttime) {
2555 			thread_lock(td);
2556 			sched_prio(td, PPAUSE);
2557 			thread_unlock(td);
2558 		}
2559 		if (syncer_state != SYNCER_RUNNING)
2560 			cv_timedwait(&sync_wakeup, &sync_mtx,
2561 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2562 		else if (time_uptime == starttime)
2563 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2564 	}
2565 }
2566 
2567 /*
2568  * Request the syncer daemon to speed up its work.
2569  * We never push it to speed up more than half of its
2570  * normal turn time, otherwise it could take over the cpu.
2571  */
2572 int
2573 speedup_syncer(void)
2574 {
2575 	int ret = 0;
2576 
2577 	mtx_lock(&sync_mtx);
2578 	if (rushjob < syncdelay / 2) {
2579 		rushjob += 1;
2580 		stat_rush_requests += 1;
2581 		ret = 1;
2582 	}
2583 	mtx_unlock(&sync_mtx);
2584 	cv_broadcast(&sync_wakeup);
2585 	return (ret);
2586 }
2587 
2588 /*
2589  * Tell the syncer to speed up its work and run though its work
2590  * list several times, then tell it to shut down.
2591  */
2592 static void
2593 syncer_shutdown(void *arg, int howto)
2594 {
2595 
2596 	if (howto & RB_NOSYNC)
2597 		return;
2598 	mtx_lock(&sync_mtx);
2599 	syncer_state = SYNCER_SHUTTING_DOWN;
2600 	rushjob = 0;
2601 	mtx_unlock(&sync_mtx);
2602 	cv_broadcast(&sync_wakeup);
2603 	kproc_shutdown(arg, howto);
2604 }
2605 
2606 void
2607 syncer_suspend(void)
2608 {
2609 
2610 	syncer_shutdown(updateproc, 0);
2611 }
2612 
2613 void
2614 syncer_resume(void)
2615 {
2616 
2617 	mtx_lock(&sync_mtx);
2618 	first_printf = 1;
2619 	syncer_state = SYNCER_RUNNING;
2620 	mtx_unlock(&sync_mtx);
2621 	cv_broadcast(&sync_wakeup);
2622 	kproc_resume(updateproc);
2623 }
2624 
2625 /*
2626  * Reassign a buffer from one vnode to another.
2627  * Used to assign file specific control information
2628  * (indirect blocks) to the vnode to which they belong.
2629  */
2630 void
2631 reassignbuf(struct buf *bp)
2632 {
2633 	struct vnode *vp;
2634 	struct bufobj *bo;
2635 	int delay;
2636 #ifdef INVARIANTS
2637 	struct bufv *bv;
2638 #endif
2639 
2640 	vp = bp->b_vp;
2641 	bo = bp->b_bufobj;
2642 	++reassignbufcalls;
2643 
2644 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2645 	    bp, bp->b_vp, bp->b_flags);
2646 	/*
2647 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2648 	 * is not fully linked in.
2649 	 */
2650 	if (bp->b_flags & B_PAGING)
2651 		panic("cannot reassign paging buffer");
2652 
2653 	/*
2654 	 * Delete from old vnode list, if on one.
2655 	 */
2656 	BO_LOCK(bo);
2657 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2658 		buf_vlist_remove(bp);
2659 	else
2660 		panic("reassignbuf: Buffer %p not on queue.", bp);
2661 	/*
2662 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2663 	 * of clean buffers.
2664 	 */
2665 	if (bp->b_flags & B_DELWRI) {
2666 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2667 			switch (vp->v_type) {
2668 			case VDIR:
2669 				delay = dirdelay;
2670 				break;
2671 			case VCHR:
2672 				delay = metadelay;
2673 				break;
2674 			default:
2675 				delay = filedelay;
2676 			}
2677 			vn_syncer_add_to_worklist(bo, delay);
2678 		}
2679 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2680 	} else {
2681 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2682 
2683 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2684 			mtx_lock(&sync_mtx);
2685 			LIST_REMOVE(bo, bo_synclist);
2686 			syncer_worklist_len--;
2687 			mtx_unlock(&sync_mtx);
2688 			bo->bo_flag &= ~BO_ONWORKLST;
2689 		}
2690 	}
2691 #ifdef INVARIANTS
2692 	bv = &bo->bo_clean;
2693 	bp = TAILQ_FIRST(&bv->bv_hd);
2694 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2695 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2696 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2697 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2698 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2699 	bv = &bo->bo_dirty;
2700 	bp = TAILQ_FIRST(&bv->bv_hd);
2701 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2702 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2703 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2704 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2705 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2706 #endif
2707 	BO_UNLOCK(bo);
2708 }
2709 
2710 static void
2711 v_init_counters(struct vnode *vp)
2712 {
2713 
2714 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2715 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2716 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2717 
2718 	refcount_init(&vp->v_holdcnt, 1);
2719 	refcount_init(&vp->v_usecount, 1);
2720 }
2721 
2722 /*
2723  * Increment si_usecount of the associated device, if any.
2724  */
2725 static void
2726 v_incr_devcount(struct vnode *vp)
2727 {
2728 
2729 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2730 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2731 		dev_lock();
2732 		vp->v_rdev->si_usecount++;
2733 		dev_unlock();
2734 	}
2735 }
2736 
2737 /*
2738  * Decrement si_usecount of the associated device, if any.
2739  */
2740 static void
2741 v_decr_devcount(struct vnode *vp)
2742 {
2743 
2744 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2745 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2746 		dev_lock();
2747 		vp->v_rdev->si_usecount--;
2748 		dev_unlock();
2749 	}
2750 }
2751 
2752 /*
2753  * Grab a particular vnode from the free list, increment its
2754  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2755  * is being destroyed.  Only callers who specify LK_RETRY will
2756  * see doomed vnodes.  If inactive processing was delayed in
2757  * vput try to do it here.
2758  *
2759  * Both holdcnt and usecount can be manipulated using atomics without holding
2760  * any locks except in these cases which require the vnode interlock:
2761  * holdcnt: 1->0 and 0->1
2762  * usecount: 0->1
2763  *
2764  * usecount is permitted to transition 1->0 without the interlock because
2765  * vnode is kept live by holdcnt.
2766  */
2767 static enum vgetstate __always_inline
2768 _vget_prep(struct vnode *vp, bool interlock)
2769 {
2770 	enum vgetstate vs;
2771 
2772 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2773 		vs = VGET_USECOUNT;
2774 	} else {
2775 		if (interlock)
2776 			vholdl(vp);
2777 		else
2778 			vhold(vp);
2779 		vs = VGET_HOLDCNT;
2780 	}
2781 	return (vs);
2782 }
2783 
2784 enum vgetstate
2785 vget_prep(struct vnode *vp)
2786 {
2787 
2788 	return (_vget_prep(vp, false));
2789 }
2790 
2791 int
2792 vget(struct vnode *vp, int flags, struct thread *td)
2793 {
2794 	enum vgetstate vs;
2795 
2796 	MPASS(td == curthread);
2797 
2798 	vs = _vget_prep(vp, (flags & LK_INTERLOCK) != 0);
2799 	return (vget_finish(vp, flags, vs));
2800 }
2801 
2802 int
2803 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2804 {
2805 	int error, oweinact;
2806 
2807 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2808 	    ("%s: invalid lock operation", __func__));
2809 
2810 	if ((flags & LK_INTERLOCK) != 0)
2811 		ASSERT_VI_LOCKED(vp, __func__);
2812 	else
2813 		ASSERT_VI_UNLOCKED(vp, __func__);
2814 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2815 	if (vs == VGET_USECOUNT) {
2816 		VNASSERT(vp->v_usecount > 0, vp,
2817 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2818 		    __func__));
2819 	}
2820 
2821 	if ((error = vn_lock(vp, flags)) != 0) {
2822 		if (vs == VGET_USECOUNT)
2823 			vrele(vp);
2824 		else
2825 			vdrop(vp);
2826 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2827 		    vp);
2828 		return (error);
2829 	}
2830 
2831 	if (vs == VGET_USECOUNT) {
2832 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2833 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2834 		return (0);
2835 	}
2836 
2837 	/*
2838 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2839 	 * the vnode around. Otherwise someone else lended their hold count and
2840 	 * we have to drop ours.
2841 	 */
2842 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2843 #ifdef INVARIANTS
2844 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2845 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2846 #else
2847 		refcount_release(&vp->v_holdcnt);
2848 #endif
2849 		VNODE_REFCOUNT_FENCE_ACQ();
2850 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2851 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2852 		return (0);
2853 	}
2854 
2855 	/*
2856 	 * We don't guarantee that any particular close will
2857 	 * trigger inactive processing so just make a best effort
2858 	 * here at preventing a reference to a removed file.  If
2859 	 * we don't succeed no harm is done.
2860 	 *
2861 	 * Upgrade our holdcnt to a usecount.
2862 	 */
2863 	VI_LOCK(vp);
2864 	/*
2865 	 * See the previous section. By the time we get here we may find
2866 	 * ourselves in the same spot.
2867 	 */
2868 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2869 #ifdef INVARIANTS
2870 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2871 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2872 #else
2873 		refcount_release(&vp->v_holdcnt);
2874 #endif
2875 		VNODE_REFCOUNT_FENCE_ACQ();
2876 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2877 		    ("%s: vnode with usecount and VI_OWEINACT set",
2878 		    __func__));
2879 		VI_UNLOCK(vp);
2880 		return (0);
2881 	}
2882 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
2883 		oweinact = 0;
2884 	} else {
2885 		oweinact = 1;
2886 		vp->v_iflag &= ~VI_OWEINACT;
2887 		VNODE_REFCOUNT_FENCE_REL();
2888 	}
2889 	v_incr_devcount(vp);
2890 	refcount_acquire(&vp->v_usecount);
2891 	if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2892 	    (flags & LK_NOWAIT) == 0)
2893 		vinactive(vp);
2894 	VI_UNLOCK(vp);
2895 	return (0);
2896 }
2897 
2898 /*
2899  * Increase the reference (use) and hold count of a vnode.
2900  * This will also remove the vnode from the free list if it is presently free.
2901  */
2902 void
2903 vref(struct vnode *vp)
2904 {
2905 
2906 	ASSERT_VI_UNLOCKED(vp, __func__);
2907 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2908 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2909 		VNODE_REFCOUNT_FENCE_ACQ();
2910 		VNASSERT(vp->v_holdcnt > 0, vp,
2911 		    ("%s: active vnode not held", __func__));
2912 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2913 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2914 		return;
2915 	}
2916 	VI_LOCK(vp);
2917 	vrefl(vp);
2918 	VI_UNLOCK(vp);
2919 }
2920 
2921 void
2922 vrefl(struct vnode *vp)
2923 {
2924 
2925 	ASSERT_VI_LOCKED(vp, __func__);
2926 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2927 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2928 		VNODE_REFCOUNT_FENCE_ACQ();
2929 		VNASSERT(vp->v_holdcnt > 0, vp,
2930 		    ("%s: active vnode not held", __func__));
2931 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2932 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2933 		return;
2934 	}
2935 	vholdl(vp);
2936 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2937 		vp->v_iflag &= ~VI_OWEINACT;
2938 		VNODE_REFCOUNT_FENCE_REL();
2939 	}
2940 	v_incr_devcount(vp);
2941 	refcount_acquire(&vp->v_usecount);
2942 }
2943 
2944 void
2945 vrefact(struct vnode *vp)
2946 {
2947 
2948 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2949 #ifdef INVARIANTS
2950 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2951 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
2952 #else
2953 	refcount_acquire(&vp->v_usecount);
2954 #endif
2955 }
2956 
2957 /*
2958  * Return reference count of a vnode.
2959  *
2960  * The results of this call are only guaranteed when some mechanism is used to
2961  * stop other processes from gaining references to the vnode.  This may be the
2962  * case if the caller holds the only reference.  This is also useful when stale
2963  * data is acceptable as race conditions may be accounted for by some other
2964  * means.
2965  */
2966 int
2967 vrefcnt(struct vnode *vp)
2968 {
2969 
2970 	return (vp->v_usecount);
2971 }
2972 
2973 void
2974 vlazy(struct vnode *vp)
2975 {
2976 	struct mount *mp;
2977 
2978 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2979 
2980 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
2981 		return;
2982 	mp = vp->v_mount;
2983 	mtx_lock(&mp->mnt_listmtx);
2984 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
2985 		vp->v_mflag |= VMP_LAZYLIST;
2986 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
2987 		mp->mnt_lazyvnodelistsize++;
2988 	}
2989 	mtx_unlock(&mp->mnt_listmtx);
2990 }
2991 
2992 static void
2993 vdefer_inactive(struct vnode *vp)
2994 {
2995 
2996 	ASSERT_VI_LOCKED(vp, __func__);
2997 	VNASSERT(vp->v_iflag & VI_OWEINACT, vp,
2998 	    ("%s: vnode without VI_OWEINACT", __func__));
2999 	if (VN_IS_DOOMED(vp)) {
3000 		vdropl(vp);
3001 		return;
3002 	}
3003 	if (vp->v_iflag & VI_DEFINACT) {
3004 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3005 		vdropl(vp);
3006 		return;
3007 	}
3008 	vlazy(vp);
3009 	vp->v_iflag |= VI_DEFINACT;
3010 	VI_UNLOCK(vp);
3011 	counter_u64_add(deferred_inact, 1);
3012 }
3013 
3014 static void
3015 vdefer_inactive_cond(struct vnode *vp)
3016 {
3017 
3018 	VI_LOCK(vp);
3019 	VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3020 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3021 		vdropl(vp);
3022 		return;
3023 	}
3024 	vdefer_inactive(vp);
3025 }
3026 
3027 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
3028 
3029 /*
3030  * Decrement the use and hold counts for a vnode.
3031  *
3032  * See an explanation near vget() as to why atomic operation is safe.
3033  */
3034 static void
3035 vputx(struct vnode *vp, enum vputx_op func)
3036 {
3037 	int error;
3038 
3039 	KASSERT(vp != NULL, ("vputx: null vp"));
3040 	if (func == VPUTX_VUNREF)
3041 		ASSERT_VOP_LOCKED(vp, "vunref");
3042 	ASSERT_VI_UNLOCKED(vp, __func__);
3043 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
3044 	    ("%s: wrong ref counts", __func__));
3045 
3046 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3047 
3048 	/*
3049 	 * We want to hold the vnode until the inactive finishes to
3050 	 * prevent vgone() races.  We drop the use count here and the
3051 	 * hold count below when we're done.
3052 	 *
3053 	 * If we release the last usecount we take ownership of the hold
3054 	 * count which provides liveness of the vnode, in which case we
3055 	 * have to vdrop.
3056 	 */
3057 	if (!refcount_release(&vp->v_usecount))
3058 		return;
3059 	VI_LOCK(vp);
3060 	v_decr_devcount(vp);
3061 	/*
3062 	 * By the time we got here someone else might have transitioned
3063 	 * the count back to > 0.
3064 	 */
3065 	if (vp->v_usecount > 0) {
3066 		vdropl(vp);
3067 		return;
3068 	}
3069 	if (vp->v_iflag & VI_DOINGINACT) {
3070 		vdropl(vp);
3071 		return;
3072 	}
3073 
3074 	/*
3075 	 * Check if the fs wants to perform inactive processing. Note we
3076 	 * may be only holding the interlock, in which case it is possible
3077 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3078 	 * Since vgone performs inactive on its own there is nothing to do
3079 	 * here but to drop our hold count.
3080 	 */
3081 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3082 	    VOP_NEED_INACTIVE(vp) == 0) {
3083 		vdropl(vp);
3084 		return;
3085 	}
3086 
3087 	/*
3088 	 * We must call VOP_INACTIVE with the node locked. Mark
3089 	 * as VI_DOINGINACT to avoid recursion.
3090 	 */
3091 	vp->v_iflag |= VI_OWEINACT;
3092 	switch (func) {
3093 	case VPUTX_VRELE:
3094 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3095 		VI_LOCK(vp);
3096 		break;
3097 	case VPUTX_VPUT:
3098 		error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT);
3099 		VI_LOCK(vp);
3100 		break;
3101 	case VPUTX_VUNREF:
3102 		error = 0;
3103 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3104 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3105 			VI_LOCK(vp);
3106 		}
3107 		break;
3108 	}
3109 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
3110 	    ("vnode with usecount and VI_OWEINACT set"));
3111 	if (error == 0) {
3112 		if (vp->v_iflag & VI_OWEINACT)
3113 			vinactive(vp);
3114 		if (func != VPUTX_VUNREF)
3115 			VOP_UNLOCK(vp);
3116 		vdropl(vp);
3117 	} else if (vp->v_iflag & VI_OWEINACT) {
3118 		vdefer_inactive(vp);
3119 	} else {
3120 		vdropl(vp);
3121 	}
3122 }
3123 
3124 /*
3125  * Vnode put/release.
3126  * If count drops to zero, call inactive routine and return to freelist.
3127  */
3128 void
3129 vrele(struct vnode *vp)
3130 {
3131 
3132 	vputx(vp, VPUTX_VRELE);
3133 }
3134 
3135 /*
3136  * Release an already locked vnode.  This give the same effects as
3137  * unlock+vrele(), but takes less time and avoids releasing and
3138  * re-aquiring the lock (as vrele() acquires the lock internally.)
3139  *
3140  * It is an invariant that all VOP_* calls operate on a held vnode.
3141  * We may be only having an implicit hold stemming from our usecount,
3142  * which we are about to release. If we unlock the vnode afterwards we
3143  * open a time window where someone else dropped the last usecount and
3144  * proceeded to free the vnode before our unlock finished. For this
3145  * reason we unlock the vnode early. This is a little bit wasteful as
3146  * it may be the vnode is exclusively locked and inactive processing is
3147  * needed, in which case we are adding work.
3148  */
3149 void
3150 vput(struct vnode *vp)
3151 {
3152 
3153 	VOP_UNLOCK(vp);
3154 	vputx(vp, VPUTX_VPUT);
3155 }
3156 
3157 /*
3158  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3159  */
3160 void
3161 vunref(struct vnode *vp)
3162 {
3163 
3164 	vputx(vp, VPUTX_VUNREF);
3165 }
3166 
3167 /*
3168  * Increase the hold count and activate if this is the first reference.
3169  */
3170 static void
3171 vhold_activate(struct vnode *vp)
3172 {
3173 
3174 	ASSERT_VI_LOCKED(vp, __func__);
3175 	VNASSERT(vp->v_holdcnt == 0, vp,
3176 	    ("%s: wrong hold count", __func__));
3177 	VNASSERT(vp->v_op != NULL, vp,
3178 	    ("%s: vnode already reclaimed.", __func__));
3179 	atomic_subtract_long(&freevnodes, 1);
3180 	refcount_acquire(&vp->v_holdcnt);
3181 }
3182 
3183 void
3184 vhold(struct vnode *vp)
3185 {
3186 
3187 	ASSERT_VI_UNLOCKED(vp, __func__);
3188 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3189 	if (refcount_acquire_if_not_zero(&vp->v_holdcnt))
3190 		return;
3191 	VI_LOCK(vp);
3192 	vholdl(vp);
3193 	VI_UNLOCK(vp);
3194 }
3195 
3196 void
3197 vholdl(struct vnode *vp)
3198 {
3199 
3200 	ASSERT_VI_LOCKED(vp, __func__);
3201 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3202 	if (vp->v_holdcnt > 0) {
3203 		refcount_acquire(&vp->v_holdcnt);
3204 		return;
3205 	}
3206 	vhold_activate(vp);
3207 }
3208 
3209 void
3210 vholdnz(struct vnode *vp)
3211 {
3212 
3213 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3214 #ifdef INVARIANTS
3215 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3216 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3217 #else
3218 	atomic_add_int(&vp->v_holdcnt, 1);
3219 #endif
3220 }
3221 
3222 static void __noinline
3223 vdbatch_process(struct vdbatch *vd)
3224 {
3225 	struct vnode *vp;
3226 	int i;
3227 
3228 	mtx_assert(&vd->lock, MA_OWNED);
3229 	MPASS(vd->index == VDBATCH_SIZE);
3230 
3231 	mtx_lock(&vnode_list_mtx);
3232 	for (i = 0; i < VDBATCH_SIZE; i++) {
3233 		vp = vd->tab[i];
3234 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3235 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3236 		MPASS(vp->v_dbatchcpu != NOCPU);
3237 		vp->v_dbatchcpu = NOCPU;
3238 	}
3239 	bzero(vd->tab, sizeof(vd->tab));
3240 	vd->index = 0;
3241 	mtx_unlock(&vnode_list_mtx);
3242 }
3243 
3244 static void
3245 vdbatch_enqueue(struct vnode *vp)
3246 {
3247 	struct vdbatch *vd;
3248 
3249 	ASSERT_VI_LOCKED(vp, __func__);
3250 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3251 	    ("%s: deferring requeue of a doomed vnode", __func__));
3252 
3253 	if (vp->v_dbatchcpu != NOCPU) {
3254 		VI_UNLOCK(vp);
3255 		return;
3256 	}
3257 
3258 	/*
3259 	 * A hack: pin us to the current CPU so that we know what to put in
3260 	 * ->v_dbatchcpu.
3261 	 */
3262 	sched_pin();
3263 	vd = DPCPU_PTR(vd);
3264 	mtx_lock(&vd->lock);
3265 	MPASS(vd->index < VDBATCH_SIZE);
3266 	MPASS(vd->tab[vd->index] == NULL);
3267 	vp->v_dbatchcpu = curcpu;
3268 	vd->tab[vd->index] = vp;
3269 	vd->index++;
3270 	VI_UNLOCK(vp);
3271 	if (vd->index == VDBATCH_SIZE)
3272 		vdbatch_process(vd);
3273 	mtx_unlock(&vd->lock);
3274 	sched_unpin();
3275 }
3276 
3277 /*
3278  * This routine must only be called for vnodes which are about to be
3279  * deallocated. Supporting dequeue for arbitrary vndoes would require
3280  * validating that the locked batch matches.
3281  */
3282 static void
3283 vdbatch_dequeue(struct vnode *vp)
3284 {
3285 	struct vdbatch *vd;
3286 	int i;
3287 	short cpu;
3288 
3289 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3290 	    ("%s: called for a used vnode\n", __func__));
3291 
3292 	cpu = vp->v_dbatchcpu;
3293 	if (cpu == NOCPU)
3294 		return;
3295 
3296 	vd = DPCPU_ID_PTR(cpu, vd);
3297 	mtx_lock(&vd->lock);
3298 	for (i = 0; i < vd->index; i++) {
3299 		if (vd->tab[i] != vp)
3300 			continue;
3301 		vp->v_dbatchcpu = NOCPU;
3302 		vd->index--;
3303 		vd->tab[i] = vd->tab[vd->index];
3304 		vd->tab[vd->index] = NULL;
3305 		break;
3306 	}
3307 	mtx_unlock(&vd->lock);
3308 	/*
3309 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3310 	 */
3311 	MPASS(vp->v_dbatchcpu == NOCPU);
3312 }
3313 
3314 /*
3315  * Drop the hold count of the vnode.  If this is the last reference to
3316  * the vnode we place it on the free list unless it has been vgone'd
3317  * (marked VIRF_DOOMED) in which case we will free it.
3318  *
3319  * Because the vnode vm object keeps a hold reference on the vnode if
3320  * there is at least one resident non-cached page, the vnode cannot
3321  * leave the active list without the page cleanup done.
3322  */
3323 static void
3324 vdrop_deactivate(struct vnode *vp)
3325 {
3326 	struct mount *mp;
3327 
3328 	ASSERT_VI_LOCKED(vp, __func__);
3329 	/*
3330 	 * Mark a vnode as free: remove it from its active list
3331 	 * and put it up for recycling on the freelist.
3332 	 */
3333 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3334 	    ("vdrop: returning doomed vnode"));
3335 	VNASSERT(vp->v_op != NULL, vp,
3336 	    ("vdrop: vnode already reclaimed."));
3337 	VNASSERT(vp->v_holdcnt == 0, vp,
3338 	    ("vdrop: freeing when we shouldn't"));
3339 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3340 	    ("vnode with VI_OWEINACT set"));
3341 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3342 	    ("vnode with VI_DEFINACT set"));
3343 	if (vp->v_mflag & VMP_LAZYLIST) {
3344 		mp = vp->v_mount;
3345 		mtx_lock(&mp->mnt_listmtx);
3346 		vp->v_mflag &= ~VMP_LAZYLIST;
3347 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3348 		mp->mnt_lazyvnodelistsize--;
3349 		mtx_unlock(&mp->mnt_listmtx);
3350 	}
3351 	atomic_add_long(&freevnodes, 1);
3352 	vdbatch_enqueue(vp);
3353 }
3354 
3355 void
3356 vdrop(struct vnode *vp)
3357 {
3358 
3359 	ASSERT_VI_UNLOCKED(vp, __func__);
3360 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3361 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3362 		return;
3363 	VI_LOCK(vp);
3364 	vdropl(vp);
3365 }
3366 
3367 void
3368 vdropl(struct vnode *vp)
3369 {
3370 
3371 	ASSERT_VI_LOCKED(vp, __func__);
3372 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3373 	if (!refcount_release(&vp->v_holdcnt)) {
3374 		VI_UNLOCK(vp);
3375 		return;
3376 	}
3377 	if (VN_IS_DOOMED(vp)) {
3378 		freevnode(vp);
3379 		return;
3380 	}
3381 	vdrop_deactivate(vp);
3382 }
3383 
3384 /*
3385  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3386  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3387  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3388  * failed lock upgrade.
3389  */
3390 void
3391 vinactive(struct vnode *vp)
3392 {
3393 	struct vm_object *obj;
3394 
3395 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3396 	ASSERT_VI_LOCKED(vp, "vinactive");
3397 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3398 	    ("vinactive: recursed on VI_DOINGINACT"));
3399 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3400 	vp->v_iflag |= VI_DOINGINACT;
3401 	vp->v_iflag &= ~VI_OWEINACT;
3402 	VI_UNLOCK(vp);
3403 	/*
3404 	 * Before moving off the active list, we must be sure that any
3405 	 * modified pages are converted into the vnode's dirty
3406 	 * buffers, since these will no longer be checked once the
3407 	 * vnode is on the inactive list.
3408 	 *
3409 	 * The write-out of the dirty pages is asynchronous.  At the
3410 	 * point that VOP_INACTIVE() is called, there could still be
3411 	 * pending I/O and dirty pages in the object.
3412 	 */
3413 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3414 	    vm_object_mightbedirty(obj)) {
3415 		VM_OBJECT_WLOCK(obj);
3416 		vm_object_page_clean(obj, 0, 0, 0);
3417 		VM_OBJECT_WUNLOCK(obj);
3418 	}
3419 	VOP_INACTIVE(vp, curthread);
3420 	VI_LOCK(vp);
3421 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3422 	    ("vinactive: lost VI_DOINGINACT"));
3423 	vp->v_iflag &= ~VI_DOINGINACT;
3424 }
3425 
3426 /*
3427  * Remove any vnodes in the vnode table belonging to mount point mp.
3428  *
3429  * If FORCECLOSE is not specified, there should not be any active ones,
3430  * return error if any are found (nb: this is a user error, not a
3431  * system error). If FORCECLOSE is specified, detach any active vnodes
3432  * that are found.
3433  *
3434  * If WRITECLOSE is set, only flush out regular file vnodes open for
3435  * writing.
3436  *
3437  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3438  *
3439  * `rootrefs' specifies the base reference count for the root vnode
3440  * of this filesystem. The root vnode is considered busy if its
3441  * v_usecount exceeds this value. On a successful return, vflush(, td)
3442  * will call vrele() on the root vnode exactly rootrefs times.
3443  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3444  * be zero.
3445  */
3446 #ifdef DIAGNOSTIC
3447 static int busyprt = 0;		/* print out busy vnodes */
3448 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3449 #endif
3450 
3451 int
3452 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3453 {
3454 	struct vnode *vp, *mvp, *rootvp = NULL;
3455 	struct vattr vattr;
3456 	int busy = 0, error;
3457 
3458 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3459 	    rootrefs, flags);
3460 	if (rootrefs > 0) {
3461 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3462 		    ("vflush: bad args"));
3463 		/*
3464 		 * Get the filesystem root vnode. We can vput() it
3465 		 * immediately, since with rootrefs > 0, it won't go away.
3466 		 */
3467 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3468 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3469 			    __func__, error);
3470 			return (error);
3471 		}
3472 		vput(rootvp);
3473 	}
3474 loop:
3475 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3476 		vholdl(vp);
3477 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3478 		if (error) {
3479 			vdrop(vp);
3480 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3481 			goto loop;
3482 		}
3483 		/*
3484 		 * Skip over a vnodes marked VV_SYSTEM.
3485 		 */
3486 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3487 			VOP_UNLOCK(vp);
3488 			vdrop(vp);
3489 			continue;
3490 		}
3491 		/*
3492 		 * If WRITECLOSE is set, flush out unlinked but still open
3493 		 * files (even if open only for reading) and regular file
3494 		 * vnodes open for writing.
3495 		 */
3496 		if (flags & WRITECLOSE) {
3497 			if (vp->v_object != NULL) {
3498 				VM_OBJECT_WLOCK(vp->v_object);
3499 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3500 				VM_OBJECT_WUNLOCK(vp->v_object);
3501 			}
3502 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3503 			if (error != 0) {
3504 				VOP_UNLOCK(vp);
3505 				vdrop(vp);
3506 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3507 				return (error);
3508 			}
3509 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3510 			VI_LOCK(vp);
3511 
3512 			if ((vp->v_type == VNON ||
3513 			    (error == 0 && vattr.va_nlink > 0)) &&
3514 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3515 				VOP_UNLOCK(vp);
3516 				vdropl(vp);
3517 				continue;
3518 			}
3519 		} else
3520 			VI_LOCK(vp);
3521 		/*
3522 		 * With v_usecount == 0, all we need to do is clear out the
3523 		 * vnode data structures and we are done.
3524 		 *
3525 		 * If FORCECLOSE is set, forcibly close the vnode.
3526 		 */
3527 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3528 			vgonel(vp);
3529 		} else {
3530 			busy++;
3531 #ifdef DIAGNOSTIC
3532 			if (busyprt)
3533 				vn_printf(vp, "vflush: busy vnode ");
3534 #endif
3535 		}
3536 		VOP_UNLOCK(vp);
3537 		vdropl(vp);
3538 	}
3539 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3540 		/*
3541 		 * If just the root vnode is busy, and if its refcount
3542 		 * is equal to `rootrefs', then go ahead and kill it.
3543 		 */
3544 		VI_LOCK(rootvp);
3545 		KASSERT(busy > 0, ("vflush: not busy"));
3546 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3547 		    ("vflush: usecount %d < rootrefs %d",
3548 		     rootvp->v_usecount, rootrefs));
3549 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3550 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3551 			vgone(rootvp);
3552 			VOP_UNLOCK(rootvp);
3553 			busy = 0;
3554 		} else
3555 			VI_UNLOCK(rootvp);
3556 	}
3557 	if (busy) {
3558 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3559 		    busy);
3560 		return (EBUSY);
3561 	}
3562 	for (; rootrefs > 0; rootrefs--)
3563 		vrele(rootvp);
3564 	return (0);
3565 }
3566 
3567 /*
3568  * Recycle an unused vnode to the front of the free list.
3569  */
3570 int
3571 vrecycle(struct vnode *vp)
3572 {
3573 	int recycled;
3574 
3575 	VI_LOCK(vp);
3576 	recycled = vrecyclel(vp);
3577 	VI_UNLOCK(vp);
3578 	return (recycled);
3579 }
3580 
3581 /*
3582  * vrecycle, with the vp interlock held.
3583  */
3584 int
3585 vrecyclel(struct vnode *vp)
3586 {
3587 	int recycled;
3588 
3589 	ASSERT_VOP_ELOCKED(vp, __func__);
3590 	ASSERT_VI_LOCKED(vp, __func__);
3591 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3592 	recycled = 0;
3593 	if (vp->v_usecount == 0) {
3594 		recycled = 1;
3595 		vgonel(vp);
3596 	}
3597 	return (recycled);
3598 }
3599 
3600 /*
3601  * Eliminate all activity associated with a vnode
3602  * in preparation for reuse.
3603  */
3604 void
3605 vgone(struct vnode *vp)
3606 {
3607 	VI_LOCK(vp);
3608 	vgonel(vp);
3609 	VI_UNLOCK(vp);
3610 }
3611 
3612 static void
3613 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3614     struct vnode *lowervp __unused)
3615 {
3616 }
3617 
3618 /*
3619  * Notify upper mounts about reclaimed or unlinked vnode.
3620  */
3621 void
3622 vfs_notify_upper(struct vnode *vp, int event)
3623 {
3624 	static struct vfsops vgonel_vfsops = {
3625 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3626 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3627 	};
3628 	struct mount *mp, *ump, *mmp;
3629 
3630 	mp = vp->v_mount;
3631 	if (mp == NULL)
3632 		return;
3633 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3634 		return;
3635 
3636 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3637 	mmp->mnt_op = &vgonel_vfsops;
3638 	mmp->mnt_kern_flag |= MNTK_MARKER;
3639 	MNT_ILOCK(mp);
3640 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3641 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3642 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3643 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3644 			continue;
3645 		}
3646 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3647 		MNT_IUNLOCK(mp);
3648 		switch (event) {
3649 		case VFS_NOTIFY_UPPER_RECLAIM:
3650 			VFS_RECLAIM_LOWERVP(ump, vp);
3651 			break;
3652 		case VFS_NOTIFY_UPPER_UNLINK:
3653 			VFS_UNLINK_LOWERVP(ump, vp);
3654 			break;
3655 		default:
3656 			KASSERT(0, ("invalid event %d", event));
3657 			break;
3658 		}
3659 		MNT_ILOCK(mp);
3660 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3661 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3662 	}
3663 	free(mmp, M_TEMP);
3664 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3665 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3666 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3667 		wakeup(&mp->mnt_uppers);
3668 	}
3669 	MNT_IUNLOCK(mp);
3670 }
3671 
3672 /*
3673  * vgone, with the vp interlock held.
3674  */
3675 static void
3676 vgonel(struct vnode *vp)
3677 {
3678 	struct thread *td;
3679 	struct mount *mp;
3680 	vm_object_t object;
3681 	bool active, oweinact;
3682 
3683 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3684 	ASSERT_VI_LOCKED(vp, "vgonel");
3685 	VNASSERT(vp->v_holdcnt, vp,
3686 	    ("vgonel: vp %p has no reference.", vp));
3687 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3688 	td = curthread;
3689 
3690 	/*
3691 	 * Don't vgonel if we're already doomed.
3692 	 */
3693 	if (vp->v_irflag & VIRF_DOOMED)
3694 		return;
3695 	vp->v_irflag |= VIRF_DOOMED;
3696 
3697 	/*
3698 	 * Check to see if the vnode is in use.  If so, we have to call
3699 	 * VOP_CLOSE() and VOP_INACTIVE().
3700 	 */
3701 	active = vp->v_usecount > 0;
3702 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3703 	/*
3704 	 * If we need to do inactive VI_OWEINACT will be set.
3705 	 */
3706 	if (vp->v_iflag & VI_DEFINACT) {
3707 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3708 		vp->v_iflag &= ~VI_DEFINACT;
3709 		vdropl(vp);
3710 	} else {
3711 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3712 		VI_UNLOCK(vp);
3713 	}
3714 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3715 
3716 	/*
3717 	 * If purging an active vnode, it must be closed and
3718 	 * deactivated before being reclaimed.
3719 	 */
3720 	if (active)
3721 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3722 	if (oweinact || active) {
3723 		VI_LOCK(vp);
3724 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3725 			vinactive(vp);
3726 		VI_UNLOCK(vp);
3727 	}
3728 	if (vp->v_type == VSOCK)
3729 		vfs_unp_reclaim(vp);
3730 
3731 	/*
3732 	 * Clean out any buffers associated with the vnode.
3733 	 * If the flush fails, just toss the buffers.
3734 	 */
3735 	mp = NULL;
3736 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3737 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3738 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3739 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3740 			;
3741 	}
3742 
3743 	BO_LOCK(&vp->v_bufobj);
3744 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3745 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3746 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3747 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3748 	    ("vp %p bufobj not invalidated", vp));
3749 
3750 	/*
3751 	 * For VMIO bufobj, BO_DEAD is set later, or in
3752 	 * vm_object_terminate() after the object's page queue is
3753 	 * flushed.
3754 	 */
3755 	object = vp->v_bufobj.bo_object;
3756 	if (object == NULL)
3757 		vp->v_bufobj.bo_flag |= BO_DEAD;
3758 	BO_UNLOCK(&vp->v_bufobj);
3759 
3760 	/*
3761 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3762 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3763 	 * should not touch the object borrowed from the lower vnode
3764 	 * (the handle check).
3765 	 */
3766 	if (object != NULL && object->type == OBJT_VNODE &&
3767 	    object->handle == vp)
3768 		vnode_destroy_vobject(vp);
3769 
3770 	/*
3771 	 * Reclaim the vnode.
3772 	 */
3773 	if (VOP_RECLAIM(vp, td))
3774 		panic("vgone: cannot reclaim");
3775 	if (mp != NULL)
3776 		vn_finished_secondary_write(mp);
3777 	VNASSERT(vp->v_object == NULL, vp,
3778 	    ("vop_reclaim left v_object vp=%p", vp));
3779 	/*
3780 	 * Clear the advisory locks and wake up waiting threads.
3781 	 */
3782 	(void)VOP_ADVLOCKPURGE(vp);
3783 	vp->v_lockf = NULL;
3784 	/*
3785 	 * Delete from old mount point vnode list.
3786 	 */
3787 	delmntque(vp);
3788 	cache_purge(vp);
3789 	/*
3790 	 * Done with purge, reset to the standard lock and invalidate
3791 	 * the vnode.
3792 	 */
3793 	VI_LOCK(vp);
3794 	vp->v_vnlock = &vp->v_lock;
3795 	vp->v_op = &dead_vnodeops;
3796 	vp->v_type = VBAD;
3797 }
3798 
3799 /*
3800  * Calculate the total number of references to a special device.
3801  */
3802 int
3803 vcount(struct vnode *vp)
3804 {
3805 	int count;
3806 
3807 	dev_lock();
3808 	count = vp->v_rdev->si_usecount;
3809 	dev_unlock();
3810 	return (count);
3811 }
3812 
3813 /*
3814  * Print out a description of a vnode.
3815  */
3816 static char *typename[] =
3817 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3818  "VMARKER"};
3819 
3820 void
3821 vn_printf(struct vnode *vp, const char *fmt, ...)
3822 {
3823 	va_list ap;
3824 	char buf[256], buf2[16];
3825 	u_long flags;
3826 
3827 	va_start(ap, fmt);
3828 	vprintf(fmt, ap);
3829 	va_end(ap);
3830 	printf("%p: ", (void *)vp);
3831 	printf("type %s\n", typename[vp->v_type]);
3832 	printf("    usecount %d, writecount %d, refcount %d",
3833 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3834 	switch (vp->v_type) {
3835 	case VDIR:
3836 		printf(" mountedhere %p\n", vp->v_mountedhere);
3837 		break;
3838 	case VCHR:
3839 		printf(" rdev %p\n", vp->v_rdev);
3840 		break;
3841 	case VSOCK:
3842 		printf(" socket %p\n", vp->v_unpcb);
3843 		break;
3844 	case VFIFO:
3845 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3846 		break;
3847 	default:
3848 		printf("\n");
3849 		break;
3850 	}
3851 	buf[0] = '\0';
3852 	buf[1] = '\0';
3853 	if (vp->v_irflag & VIRF_DOOMED)
3854 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3855 	flags = vp->v_irflag & ~(VIRF_DOOMED);
3856 	if (flags != 0) {
3857 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3858 		strlcat(buf, buf2, sizeof(buf));
3859 	}
3860 	if (vp->v_vflag & VV_ROOT)
3861 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3862 	if (vp->v_vflag & VV_ISTTY)
3863 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3864 	if (vp->v_vflag & VV_NOSYNC)
3865 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3866 	if (vp->v_vflag & VV_ETERNALDEV)
3867 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3868 	if (vp->v_vflag & VV_CACHEDLABEL)
3869 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3870 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3871 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3872 	if (vp->v_vflag & VV_COPYONWRITE)
3873 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3874 	if (vp->v_vflag & VV_SYSTEM)
3875 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3876 	if (vp->v_vflag & VV_PROCDEP)
3877 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3878 	if (vp->v_vflag & VV_NOKNOTE)
3879 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3880 	if (vp->v_vflag & VV_DELETED)
3881 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3882 	if (vp->v_vflag & VV_MD)
3883 		strlcat(buf, "|VV_MD", sizeof(buf));
3884 	if (vp->v_vflag & VV_FORCEINSMQ)
3885 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3886 	if (vp->v_vflag & VV_READLINK)
3887 		strlcat(buf, "|VV_READLINK", sizeof(buf));
3888 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3889 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3890 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3891 	if (flags != 0) {
3892 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3893 		strlcat(buf, buf2, sizeof(buf));
3894 	}
3895 	if (vp->v_iflag & VI_TEXT_REF)
3896 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
3897 	if (vp->v_iflag & VI_MOUNT)
3898 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3899 	if (vp->v_iflag & VI_DOINGINACT)
3900 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3901 	if (vp->v_iflag & VI_OWEINACT)
3902 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3903 	if (vp->v_iflag & VI_DEFINACT)
3904 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
3905 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
3906 	    VI_OWEINACT | VI_DEFINACT);
3907 	if (flags != 0) {
3908 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3909 		strlcat(buf, buf2, sizeof(buf));
3910 	}
3911 	if (vp->v_mflag & VMP_LAZYLIST)
3912 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
3913 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
3914 	if (flags != 0) {
3915 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
3916 		strlcat(buf, buf2, sizeof(buf));
3917 	}
3918 	printf("    flags (%s)\n", buf + 1);
3919 	if (mtx_owned(VI_MTX(vp)))
3920 		printf(" VI_LOCKed");
3921 	if (vp->v_object != NULL)
3922 		printf("    v_object %p ref %d pages %d "
3923 		    "cleanbuf %d dirtybuf %d\n",
3924 		    vp->v_object, vp->v_object->ref_count,
3925 		    vp->v_object->resident_page_count,
3926 		    vp->v_bufobj.bo_clean.bv_cnt,
3927 		    vp->v_bufobj.bo_dirty.bv_cnt);
3928 	printf("    ");
3929 	lockmgr_printinfo(vp->v_vnlock);
3930 	if (vp->v_data != NULL)
3931 		VOP_PRINT(vp);
3932 }
3933 
3934 #ifdef DDB
3935 /*
3936  * List all of the locked vnodes in the system.
3937  * Called when debugging the kernel.
3938  */
3939 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3940 {
3941 	struct mount *mp;
3942 	struct vnode *vp;
3943 
3944 	/*
3945 	 * Note: because this is DDB, we can't obey the locking semantics
3946 	 * for these structures, which means we could catch an inconsistent
3947 	 * state and dereference a nasty pointer.  Not much to be done
3948 	 * about that.
3949 	 */
3950 	db_printf("Locked vnodes\n");
3951 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3952 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3953 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3954 				vn_printf(vp, "vnode ");
3955 		}
3956 	}
3957 }
3958 
3959 /*
3960  * Show details about the given vnode.
3961  */
3962 DB_SHOW_COMMAND(vnode, db_show_vnode)
3963 {
3964 	struct vnode *vp;
3965 
3966 	if (!have_addr)
3967 		return;
3968 	vp = (struct vnode *)addr;
3969 	vn_printf(vp, "vnode ");
3970 }
3971 
3972 /*
3973  * Show details about the given mount point.
3974  */
3975 DB_SHOW_COMMAND(mount, db_show_mount)
3976 {
3977 	struct mount *mp;
3978 	struct vfsopt *opt;
3979 	struct statfs *sp;
3980 	struct vnode *vp;
3981 	char buf[512];
3982 	uint64_t mflags;
3983 	u_int flags;
3984 
3985 	if (!have_addr) {
3986 		/* No address given, print short info about all mount points. */
3987 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3988 			db_printf("%p %s on %s (%s)\n", mp,
3989 			    mp->mnt_stat.f_mntfromname,
3990 			    mp->mnt_stat.f_mntonname,
3991 			    mp->mnt_stat.f_fstypename);
3992 			if (db_pager_quit)
3993 				break;
3994 		}
3995 		db_printf("\nMore info: show mount <addr>\n");
3996 		return;
3997 	}
3998 
3999 	mp = (struct mount *)addr;
4000 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4001 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4002 
4003 	buf[0] = '\0';
4004 	mflags = mp->mnt_flag;
4005 #define	MNT_FLAG(flag)	do {						\
4006 	if (mflags & (flag)) {						\
4007 		if (buf[0] != '\0')					\
4008 			strlcat(buf, ", ", sizeof(buf));		\
4009 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4010 		mflags &= ~(flag);					\
4011 	}								\
4012 } while (0)
4013 	MNT_FLAG(MNT_RDONLY);
4014 	MNT_FLAG(MNT_SYNCHRONOUS);
4015 	MNT_FLAG(MNT_NOEXEC);
4016 	MNT_FLAG(MNT_NOSUID);
4017 	MNT_FLAG(MNT_NFS4ACLS);
4018 	MNT_FLAG(MNT_UNION);
4019 	MNT_FLAG(MNT_ASYNC);
4020 	MNT_FLAG(MNT_SUIDDIR);
4021 	MNT_FLAG(MNT_SOFTDEP);
4022 	MNT_FLAG(MNT_NOSYMFOLLOW);
4023 	MNT_FLAG(MNT_GJOURNAL);
4024 	MNT_FLAG(MNT_MULTILABEL);
4025 	MNT_FLAG(MNT_ACLS);
4026 	MNT_FLAG(MNT_NOATIME);
4027 	MNT_FLAG(MNT_NOCLUSTERR);
4028 	MNT_FLAG(MNT_NOCLUSTERW);
4029 	MNT_FLAG(MNT_SUJ);
4030 	MNT_FLAG(MNT_EXRDONLY);
4031 	MNT_FLAG(MNT_EXPORTED);
4032 	MNT_FLAG(MNT_DEFEXPORTED);
4033 	MNT_FLAG(MNT_EXPORTANON);
4034 	MNT_FLAG(MNT_EXKERB);
4035 	MNT_FLAG(MNT_EXPUBLIC);
4036 	MNT_FLAG(MNT_LOCAL);
4037 	MNT_FLAG(MNT_QUOTA);
4038 	MNT_FLAG(MNT_ROOTFS);
4039 	MNT_FLAG(MNT_USER);
4040 	MNT_FLAG(MNT_IGNORE);
4041 	MNT_FLAG(MNT_UPDATE);
4042 	MNT_FLAG(MNT_DELEXPORT);
4043 	MNT_FLAG(MNT_RELOAD);
4044 	MNT_FLAG(MNT_FORCE);
4045 	MNT_FLAG(MNT_SNAPSHOT);
4046 	MNT_FLAG(MNT_BYFSID);
4047 #undef MNT_FLAG
4048 	if (mflags != 0) {
4049 		if (buf[0] != '\0')
4050 			strlcat(buf, ", ", sizeof(buf));
4051 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4052 		    "0x%016jx", mflags);
4053 	}
4054 	db_printf("    mnt_flag = %s\n", buf);
4055 
4056 	buf[0] = '\0';
4057 	flags = mp->mnt_kern_flag;
4058 #define	MNT_KERN_FLAG(flag)	do {					\
4059 	if (flags & (flag)) {						\
4060 		if (buf[0] != '\0')					\
4061 			strlcat(buf, ", ", sizeof(buf));		\
4062 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4063 		flags &= ~(flag);					\
4064 	}								\
4065 } while (0)
4066 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4067 	MNT_KERN_FLAG(MNTK_ASYNC);
4068 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4069 	MNT_KERN_FLAG(MNTK_DRAINING);
4070 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4071 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4072 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4073 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4074 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4075 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4076 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4077 	MNT_KERN_FLAG(MNTK_MARKER);
4078 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4079 	MNT_KERN_FLAG(MNTK_NOASYNC);
4080 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4081 	MNT_KERN_FLAG(MNTK_MWAIT);
4082 	MNT_KERN_FLAG(MNTK_SUSPEND);
4083 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4084 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4085 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4086 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4087 #undef MNT_KERN_FLAG
4088 	if (flags != 0) {
4089 		if (buf[0] != '\0')
4090 			strlcat(buf, ", ", sizeof(buf));
4091 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4092 		    "0x%08x", flags);
4093 	}
4094 	db_printf("    mnt_kern_flag = %s\n", buf);
4095 
4096 	db_printf("    mnt_opt = ");
4097 	opt = TAILQ_FIRST(mp->mnt_opt);
4098 	if (opt != NULL) {
4099 		db_printf("%s", opt->name);
4100 		opt = TAILQ_NEXT(opt, link);
4101 		while (opt != NULL) {
4102 			db_printf(", %s", opt->name);
4103 			opt = TAILQ_NEXT(opt, link);
4104 		}
4105 	}
4106 	db_printf("\n");
4107 
4108 	sp = &mp->mnt_stat;
4109 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4110 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4111 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4112 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4113 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4114 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4115 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4116 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4117 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4118 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4119 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4120 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4121 
4122 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4123 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4124 	if (jailed(mp->mnt_cred))
4125 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4126 	db_printf(" }\n");
4127 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4128 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4129 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4130 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4131 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4132 	    mp->mnt_lazyvnodelistsize);
4133 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4134 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4135 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4136 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4137 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4138 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4139 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4140 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4141 	db_printf("    mnt_secondary_accwrites = %d\n",
4142 	    mp->mnt_secondary_accwrites);
4143 	db_printf("    mnt_gjprovider = %s\n",
4144 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4145 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4146 
4147 	db_printf("\n\nList of active vnodes\n");
4148 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4149 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4150 			vn_printf(vp, "vnode ");
4151 			if (db_pager_quit)
4152 				break;
4153 		}
4154 	}
4155 	db_printf("\n\nList of inactive vnodes\n");
4156 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4157 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4158 			vn_printf(vp, "vnode ");
4159 			if (db_pager_quit)
4160 				break;
4161 		}
4162 	}
4163 }
4164 #endif	/* DDB */
4165 
4166 /*
4167  * Fill in a struct xvfsconf based on a struct vfsconf.
4168  */
4169 static int
4170 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4171 {
4172 	struct xvfsconf xvfsp;
4173 
4174 	bzero(&xvfsp, sizeof(xvfsp));
4175 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4176 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4177 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4178 	xvfsp.vfc_flags = vfsp->vfc_flags;
4179 	/*
4180 	 * These are unused in userland, we keep them
4181 	 * to not break binary compatibility.
4182 	 */
4183 	xvfsp.vfc_vfsops = NULL;
4184 	xvfsp.vfc_next = NULL;
4185 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4186 }
4187 
4188 #ifdef COMPAT_FREEBSD32
4189 struct xvfsconf32 {
4190 	uint32_t	vfc_vfsops;
4191 	char		vfc_name[MFSNAMELEN];
4192 	int32_t		vfc_typenum;
4193 	int32_t		vfc_refcount;
4194 	int32_t		vfc_flags;
4195 	uint32_t	vfc_next;
4196 };
4197 
4198 static int
4199 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4200 {
4201 	struct xvfsconf32 xvfsp;
4202 
4203 	bzero(&xvfsp, sizeof(xvfsp));
4204 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4205 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4206 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4207 	xvfsp.vfc_flags = vfsp->vfc_flags;
4208 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4209 }
4210 #endif
4211 
4212 /*
4213  * Top level filesystem related information gathering.
4214  */
4215 static int
4216 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4217 {
4218 	struct vfsconf *vfsp;
4219 	int error;
4220 
4221 	error = 0;
4222 	vfsconf_slock();
4223 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4224 #ifdef COMPAT_FREEBSD32
4225 		if (req->flags & SCTL_MASK32)
4226 			error = vfsconf2x32(req, vfsp);
4227 		else
4228 #endif
4229 			error = vfsconf2x(req, vfsp);
4230 		if (error)
4231 			break;
4232 	}
4233 	vfsconf_sunlock();
4234 	return (error);
4235 }
4236 
4237 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4238     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4239     "S,xvfsconf", "List of all configured filesystems");
4240 
4241 #ifndef BURN_BRIDGES
4242 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4243 
4244 static int
4245 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4246 {
4247 	int *name = (int *)arg1 - 1;	/* XXX */
4248 	u_int namelen = arg2 + 1;	/* XXX */
4249 	struct vfsconf *vfsp;
4250 
4251 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4252 	    "please rebuild world\n");
4253 
4254 #if 1 || defined(COMPAT_PRELITE2)
4255 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4256 	if (namelen == 1)
4257 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4258 #endif
4259 
4260 	switch (name[1]) {
4261 	case VFS_MAXTYPENUM:
4262 		if (namelen != 2)
4263 			return (ENOTDIR);
4264 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4265 	case VFS_CONF:
4266 		if (namelen != 3)
4267 			return (ENOTDIR);	/* overloaded */
4268 		vfsconf_slock();
4269 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4270 			if (vfsp->vfc_typenum == name[2])
4271 				break;
4272 		}
4273 		vfsconf_sunlock();
4274 		if (vfsp == NULL)
4275 			return (EOPNOTSUPP);
4276 #ifdef COMPAT_FREEBSD32
4277 		if (req->flags & SCTL_MASK32)
4278 			return (vfsconf2x32(req, vfsp));
4279 		else
4280 #endif
4281 			return (vfsconf2x(req, vfsp));
4282 	}
4283 	return (EOPNOTSUPP);
4284 }
4285 
4286 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4287     CTLFLAG_MPSAFE, vfs_sysctl,
4288     "Generic filesystem");
4289 
4290 #if 1 || defined(COMPAT_PRELITE2)
4291 
4292 static int
4293 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4294 {
4295 	int error;
4296 	struct vfsconf *vfsp;
4297 	struct ovfsconf ovfs;
4298 
4299 	vfsconf_slock();
4300 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4301 		bzero(&ovfs, sizeof(ovfs));
4302 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4303 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4304 		ovfs.vfc_index = vfsp->vfc_typenum;
4305 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4306 		ovfs.vfc_flags = vfsp->vfc_flags;
4307 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4308 		if (error != 0) {
4309 			vfsconf_sunlock();
4310 			return (error);
4311 		}
4312 	}
4313 	vfsconf_sunlock();
4314 	return (0);
4315 }
4316 
4317 #endif /* 1 || COMPAT_PRELITE2 */
4318 #endif /* !BURN_BRIDGES */
4319 
4320 #define KINFO_VNODESLOP		10
4321 #ifdef notyet
4322 /*
4323  * Dump vnode list (via sysctl).
4324  */
4325 /* ARGSUSED */
4326 static int
4327 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4328 {
4329 	struct xvnode *xvn;
4330 	struct mount *mp;
4331 	struct vnode *vp;
4332 	int error, len, n;
4333 
4334 	/*
4335 	 * Stale numvnodes access is not fatal here.
4336 	 */
4337 	req->lock = 0;
4338 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4339 	if (!req->oldptr)
4340 		/* Make an estimate */
4341 		return (SYSCTL_OUT(req, 0, len));
4342 
4343 	error = sysctl_wire_old_buffer(req, 0);
4344 	if (error != 0)
4345 		return (error);
4346 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4347 	n = 0;
4348 	mtx_lock(&mountlist_mtx);
4349 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4350 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4351 			continue;
4352 		MNT_ILOCK(mp);
4353 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4354 			if (n == len)
4355 				break;
4356 			vref(vp);
4357 			xvn[n].xv_size = sizeof *xvn;
4358 			xvn[n].xv_vnode = vp;
4359 			xvn[n].xv_id = 0;	/* XXX compat */
4360 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4361 			XV_COPY(usecount);
4362 			XV_COPY(writecount);
4363 			XV_COPY(holdcnt);
4364 			XV_COPY(mount);
4365 			XV_COPY(numoutput);
4366 			XV_COPY(type);
4367 #undef XV_COPY
4368 			xvn[n].xv_flag = vp->v_vflag;
4369 
4370 			switch (vp->v_type) {
4371 			case VREG:
4372 			case VDIR:
4373 			case VLNK:
4374 				break;
4375 			case VBLK:
4376 			case VCHR:
4377 				if (vp->v_rdev == NULL) {
4378 					vrele(vp);
4379 					continue;
4380 				}
4381 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4382 				break;
4383 			case VSOCK:
4384 				xvn[n].xv_socket = vp->v_socket;
4385 				break;
4386 			case VFIFO:
4387 				xvn[n].xv_fifo = vp->v_fifoinfo;
4388 				break;
4389 			case VNON:
4390 			case VBAD:
4391 			default:
4392 				/* shouldn't happen? */
4393 				vrele(vp);
4394 				continue;
4395 			}
4396 			vrele(vp);
4397 			++n;
4398 		}
4399 		MNT_IUNLOCK(mp);
4400 		mtx_lock(&mountlist_mtx);
4401 		vfs_unbusy(mp);
4402 		if (n == len)
4403 			break;
4404 	}
4405 	mtx_unlock(&mountlist_mtx);
4406 
4407 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4408 	free(xvn, M_TEMP);
4409 	return (error);
4410 }
4411 
4412 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4413     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4414     "");
4415 #endif
4416 
4417 static void
4418 unmount_or_warn(struct mount *mp)
4419 {
4420 	int error;
4421 
4422 	error = dounmount(mp, MNT_FORCE, curthread);
4423 	if (error != 0) {
4424 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4425 		if (error == EBUSY)
4426 			printf("BUSY)\n");
4427 		else
4428 			printf("%d)\n", error);
4429 	}
4430 }
4431 
4432 /*
4433  * Unmount all filesystems. The list is traversed in reverse order
4434  * of mounting to avoid dependencies.
4435  */
4436 void
4437 vfs_unmountall(void)
4438 {
4439 	struct mount *mp, *tmp;
4440 
4441 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4442 
4443 	/*
4444 	 * Since this only runs when rebooting, it is not interlocked.
4445 	 */
4446 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4447 		vfs_ref(mp);
4448 
4449 		/*
4450 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4451 		 * unmount of the latter.
4452 		 */
4453 		if (mp == rootdevmp)
4454 			continue;
4455 
4456 		unmount_or_warn(mp);
4457 	}
4458 
4459 	if (rootdevmp != NULL)
4460 		unmount_or_warn(rootdevmp);
4461 }
4462 
4463 static void
4464 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4465 {
4466 
4467 	ASSERT_VI_LOCKED(vp, __func__);
4468 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4469 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4470 		vdropl(vp);
4471 		return;
4472 	}
4473 	if (vn_lock(vp, lkflags) == 0) {
4474 		VI_LOCK(vp);
4475 		if ((vp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == VI_OWEINACT)
4476 			vinactive(vp);
4477 		VOP_UNLOCK(vp);
4478 		vdropl(vp);
4479 		return;
4480 	}
4481 	vdefer_inactive_cond(vp);
4482 }
4483 
4484 static int
4485 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4486 {
4487 
4488 	return (vp->v_iflag & VI_DEFINACT);
4489 }
4490 
4491 static void __noinline
4492 vfs_periodic_inactive(struct mount *mp, int flags)
4493 {
4494 	struct vnode *vp, *mvp;
4495 	int lkflags;
4496 
4497 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4498 	if (flags != MNT_WAIT)
4499 		lkflags |= LK_NOWAIT;
4500 
4501 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4502 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4503 			VI_UNLOCK(vp);
4504 			continue;
4505 		}
4506 		vp->v_iflag &= ~VI_DEFINACT;
4507 		vfs_deferred_inactive(vp, lkflags);
4508 	}
4509 }
4510 
4511 static inline bool
4512 vfs_want_msync(struct vnode *vp)
4513 {
4514 	struct vm_object *obj;
4515 
4516 	/*
4517 	 * This test may be performed without any locks held.
4518 	 * We rely on vm_object's type stability.
4519 	 */
4520 	if (vp->v_vflag & VV_NOSYNC)
4521 		return (false);
4522 	obj = vp->v_object;
4523 	return (obj != NULL && vm_object_mightbedirty(obj));
4524 }
4525 
4526 static int
4527 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4528 {
4529 
4530 	if (vp->v_vflag & VV_NOSYNC)
4531 		return (false);
4532 	if (vp->v_iflag & VI_DEFINACT)
4533 		return (true);
4534 	return (vfs_want_msync(vp));
4535 }
4536 
4537 static void __noinline
4538 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4539 {
4540 	struct vnode *vp, *mvp;
4541 	struct vm_object *obj;
4542 	struct thread *td;
4543 	int lkflags, objflags;
4544 	bool seen_defer;
4545 
4546 	td = curthread;
4547 
4548 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4549 	if (flags != MNT_WAIT) {
4550 		lkflags |= LK_NOWAIT;
4551 		objflags = OBJPC_NOSYNC;
4552 	} else {
4553 		objflags = OBJPC_SYNC;
4554 	}
4555 
4556 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4557 		seen_defer = false;
4558 		if (vp->v_iflag & VI_DEFINACT) {
4559 			vp->v_iflag &= ~VI_DEFINACT;
4560 			seen_defer = true;
4561 		}
4562 		if (!vfs_want_msync(vp)) {
4563 			if (seen_defer)
4564 				vfs_deferred_inactive(vp, lkflags);
4565 			else
4566 				VI_UNLOCK(vp);
4567 			continue;
4568 		}
4569 		if (vget(vp, lkflags, td) == 0) {
4570 			obj = vp->v_object;
4571 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4572 				VM_OBJECT_WLOCK(obj);
4573 				vm_object_page_clean(obj, 0, 0, objflags);
4574 				VM_OBJECT_WUNLOCK(obj);
4575 			}
4576 			vput(vp);
4577 			if (seen_defer)
4578 				vdrop(vp);
4579 		} else {
4580 			if (seen_defer)
4581 				vdefer_inactive_cond(vp);
4582 		}
4583 	}
4584 }
4585 
4586 void
4587 vfs_periodic(struct mount *mp, int flags)
4588 {
4589 
4590 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4591 
4592 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4593 		vfs_periodic_inactive(mp, flags);
4594 	else
4595 		vfs_periodic_msync_inactive(mp, flags);
4596 }
4597 
4598 static void
4599 destroy_vpollinfo_free(struct vpollinfo *vi)
4600 {
4601 
4602 	knlist_destroy(&vi->vpi_selinfo.si_note);
4603 	mtx_destroy(&vi->vpi_lock);
4604 	uma_zfree(vnodepoll_zone, vi);
4605 }
4606 
4607 static void
4608 destroy_vpollinfo(struct vpollinfo *vi)
4609 {
4610 
4611 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4612 	seldrain(&vi->vpi_selinfo);
4613 	destroy_vpollinfo_free(vi);
4614 }
4615 
4616 /*
4617  * Initialize per-vnode helper structure to hold poll-related state.
4618  */
4619 void
4620 v_addpollinfo(struct vnode *vp)
4621 {
4622 	struct vpollinfo *vi;
4623 
4624 	if (vp->v_pollinfo != NULL)
4625 		return;
4626 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4627 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4628 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4629 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4630 	VI_LOCK(vp);
4631 	if (vp->v_pollinfo != NULL) {
4632 		VI_UNLOCK(vp);
4633 		destroy_vpollinfo_free(vi);
4634 		return;
4635 	}
4636 	vp->v_pollinfo = vi;
4637 	VI_UNLOCK(vp);
4638 }
4639 
4640 /*
4641  * Record a process's interest in events which might happen to
4642  * a vnode.  Because poll uses the historic select-style interface
4643  * internally, this routine serves as both the ``check for any
4644  * pending events'' and the ``record my interest in future events''
4645  * functions.  (These are done together, while the lock is held,
4646  * to avoid race conditions.)
4647  */
4648 int
4649 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4650 {
4651 
4652 	v_addpollinfo(vp);
4653 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4654 	if (vp->v_pollinfo->vpi_revents & events) {
4655 		/*
4656 		 * This leaves events we are not interested
4657 		 * in available for the other process which
4658 		 * which presumably had requested them
4659 		 * (otherwise they would never have been
4660 		 * recorded).
4661 		 */
4662 		events &= vp->v_pollinfo->vpi_revents;
4663 		vp->v_pollinfo->vpi_revents &= ~events;
4664 
4665 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4666 		return (events);
4667 	}
4668 	vp->v_pollinfo->vpi_events |= events;
4669 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4670 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4671 	return (0);
4672 }
4673 
4674 /*
4675  * Routine to create and manage a filesystem syncer vnode.
4676  */
4677 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4678 static int	sync_fsync(struct  vop_fsync_args *);
4679 static int	sync_inactive(struct  vop_inactive_args *);
4680 static int	sync_reclaim(struct  vop_reclaim_args *);
4681 
4682 static struct vop_vector sync_vnodeops = {
4683 	.vop_bypass =	VOP_EOPNOTSUPP,
4684 	.vop_close =	sync_close,		/* close */
4685 	.vop_fsync =	sync_fsync,		/* fsync */
4686 	.vop_inactive =	sync_inactive,	/* inactive */
4687 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4688 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4689 	.vop_lock1 =	vop_stdlock,	/* lock */
4690 	.vop_unlock =	vop_stdunlock,	/* unlock */
4691 	.vop_islocked =	vop_stdislocked,	/* islocked */
4692 };
4693 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4694 
4695 /*
4696  * Create a new filesystem syncer vnode for the specified mount point.
4697  */
4698 void
4699 vfs_allocate_syncvnode(struct mount *mp)
4700 {
4701 	struct vnode *vp;
4702 	struct bufobj *bo;
4703 	static long start, incr, next;
4704 	int error;
4705 
4706 	/* Allocate a new vnode */
4707 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4708 	if (error != 0)
4709 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4710 	vp->v_type = VNON;
4711 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4712 	vp->v_vflag |= VV_FORCEINSMQ;
4713 	error = insmntque(vp, mp);
4714 	if (error != 0)
4715 		panic("vfs_allocate_syncvnode: insmntque() failed");
4716 	vp->v_vflag &= ~VV_FORCEINSMQ;
4717 	VOP_UNLOCK(vp);
4718 	/*
4719 	 * Place the vnode onto the syncer worklist. We attempt to
4720 	 * scatter them about on the list so that they will go off
4721 	 * at evenly distributed times even if all the filesystems
4722 	 * are mounted at once.
4723 	 */
4724 	next += incr;
4725 	if (next == 0 || next > syncer_maxdelay) {
4726 		start /= 2;
4727 		incr /= 2;
4728 		if (start == 0) {
4729 			start = syncer_maxdelay / 2;
4730 			incr = syncer_maxdelay;
4731 		}
4732 		next = start;
4733 	}
4734 	bo = &vp->v_bufobj;
4735 	BO_LOCK(bo);
4736 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4737 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4738 	mtx_lock(&sync_mtx);
4739 	sync_vnode_count++;
4740 	if (mp->mnt_syncer == NULL) {
4741 		mp->mnt_syncer = vp;
4742 		vp = NULL;
4743 	}
4744 	mtx_unlock(&sync_mtx);
4745 	BO_UNLOCK(bo);
4746 	if (vp != NULL) {
4747 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4748 		vgone(vp);
4749 		vput(vp);
4750 	}
4751 }
4752 
4753 void
4754 vfs_deallocate_syncvnode(struct mount *mp)
4755 {
4756 	struct vnode *vp;
4757 
4758 	mtx_lock(&sync_mtx);
4759 	vp = mp->mnt_syncer;
4760 	if (vp != NULL)
4761 		mp->mnt_syncer = NULL;
4762 	mtx_unlock(&sync_mtx);
4763 	if (vp != NULL)
4764 		vrele(vp);
4765 }
4766 
4767 /*
4768  * Do a lazy sync of the filesystem.
4769  */
4770 static int
4771 sync_fsync(struct vop_fsync_args *ap)
4772 {
4773 	struct vnode *syncvp = ap->a_vp;
4774 	struct mount *mp = syncvp->v_mount;
4775 	int error, save;
4776 	struct bufobj *bo;
4777 
4778 	/*
4779 	 * We only need to do something if this is a lazy evaluation.
4780 	 */
4781 	if (ap->a_waitfor != MNT_LAZY)
4782 		return (0);
4783 
4784 	/*
4785 	 * Move ourselves to the back of the sync list.
4786 	 */
4787 	bo = &syncvp->v_bufobj;
4788 	BO_LOCK(bo);
4789 	vn_syncer_add_to_worklist(bo, syncdelay);
4790 	BO_UNLOCK(bo);
4791 
4792 	/*
4793 	 * Walk the list of vnodes pushing all that are dirty and
4794 	 * not already on the sync list.
4795 	 */
4796 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4797 		return (0);
4798 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4799 		vfs_unbusy(mp);
4800 		return (0);
4801 	}
4802 	save = curthread_pflags_set(TDP_SYNCIO);
4803 	/*
4804 	 * The filesystem at hand may be idle with free vnodes stored in the
4805 	 * batch.  Return them instead of letting them stay there indefinitely.
4806 	 */
4807 	vfs_periodic(mp, MNT_NOWAIT);
4808 	error = VFS_SYNC(mp, MNT_LAZY);
4809 	curthread_pflags_restore(save);
4810 	vn_finished_write(mp);
4811 	vfs_unbusy(mp);
4812 	return (error);
4813 }
4814 
4815 /*
4816  * The syncer vnode is no referenced.
4817  */
4818 static int
4819 sync_inactive(struct vop_inactive_args *ap)
4820 {
4821 
4822 	vgone(ap->a_vp);
4823 	return (0);
4824 }
4825 
4826 /*
4827  * The syncer vnode is no longer needed and is being decommissioned.
4828  *
4829  * Modifications to the worklist must be protected by sync_mtx.
4830  */
4831 static int
4832 sync_reclaim(struct vop_reclaim_args *ap)
4833 {
4834 	struct vnode *vp = ap->a_vp;
4835 	struct bufobj *bo;
4836 
4837 	bo = &vp->v_bufobj;
4838 	BO_LOCK(bo);
4839 	mtx_lock(&sync_mtx);
4840 	if (vp->v_mount->mnt_syncer == vp)
4841 		vp->v_mount->mnt_syncer = NULL;
4842 	if (bo->bo_flag & BO_ONWORKLST) {
4843 		LIST_REMOVE(bo, bo_synclist);
4844 		syncer_worklist_len--;
4845 		sync_vnode_count--;
4846 		bo->bo_flag &= ~BO_ONWORKLST;
4847 	}
4848 	mtx_unlock(&sync_mtx);
4849 	BO_UNLOCK(bo);
4850 
4851 	return (0);
4852 }
4853 
4854 int
4855 vn_need_pageq_flush(struct vnode *vp)
4856 {
4857 	struct vm_object *obj;
4858 	int need;
4859 
4860 	MPASS(mtx_owned(VI_MTX(vp)));
4861 	need = 0;
4862 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4863 	    vm_object_mightbedirty(obj))
4864 		need = 1;
4865 	return (need);
4866 }
4867 
4868 /*
4869  * Check if vnode represents a disk device
4870  */
4871 int
4872 vn_isdisk(struct vnode *vp, int *errp)
4873 {
4874 	int error;
4875 
4876 	if (vp->v_type != VCHR) {
4877 		error = ENOTBLK;
4878 		goto out;
4879 	}
4880 	error = 0;
4881 	dev_lock();
4882 	if (vp->v_rdev == NULL)
4883 		error = ENXIO;
4884 	else if (vp->v_rdev->si_devsw == NULL)
4885 		error = ENXIO;
4886 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4887 		error = ENOTBLK;
4888 	dev_unlock();
4889 out:
4890 	if (errp != NULL)
4891 		*errp = error;
4892 	return (error == 0);
4893 }
4894 
4895 /*
4896  * Common filesystem object access control check routine.  Accepts a
4897  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4898  * and optional call-by-reference privused argument allowing vaccess()
4899  * to indicate to the caller whether privilege was used to satisfy the
4900  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4901  */
4902 int
4903 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4904     accmode_t accmode, struct ucred *cred, int *privused)
4905 {
4906 	accmode_t dac_granted;
4907 	accmode_t priv_granted;
4908 
4909 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4910 	    ("invalid bit in accmode"));
4911 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4912 	    ("VAPPEND without VWRITE"));
4913 
4914 	/*
4915 	 * Look for a normal, non-privileged way to access the file/directory
4916 	 * as requested.  If it exists, go with that.
4917 	 */
4918 
4919 	if (privused != NULL)
4920 		*privused = 0;
4921 
4922 	dac_granted = 0;
4923 
4924 	/* Check the owner. */
4925 	if (cred->cr_uid == file_uid) {
4926 		dac_granted |= VADMIN;
4927 		if (file_mode & S_IXUSR)
4928 			dac_granted |= VEXEC;
4929 		if (file_mode & S_IRUSR)
4930 			dac_granted |= VREAD;
4931 		if (file_mode & S_IWUSR)
4932 			dac_granted |= (VWRITE | VAPPEND);
4933 
4934 		if ((accmode & dac_granted) == accmode)
4935 			return (0);
4936 
4937 		goto privcheck;
4938 	}
4939 
4940 	/* Otherwise, check the groups (first match) */
4941 	if (groupmember(file_gid, cred)) {
4942 		if (file_mode & S_IXGRP)
4943 			dac_granted |= VEXEC;
4944 		if (file_mode & S_IRGRP)
4945 			dac_granted |= VREAD;
4946 		if (file_mode & S_IWGRP)
4947 			dac_granted |= (VWRITE | VAPPEND);
4948 
4949 		if ((accmode & dac_granted) == accmode)
4950 			return (0);
4951 
4952 		goto privcheck;
4953 	}
4954 
4955 	/* Otherwise, check everyone else. */
4956 	if (file_mode & S_IXOTH)
4957 		dac_granted |= VEXEC;
4958 	if (file_mode & S_IROTH)
4959 		dac_granted |= VREAD;
4960 	if (file_mode & S_IWOTH)
4961 		dac_granted |= (VWRITE | VAPPEND);
4962 	if ((accmode & dac_granted) == accmode)
4963 		return (0);
4964 
4965 privcheck:
4966 	/*
4967 	 * Build a privilege mask to determine if the set of privileges
4968 	 * satisfies the requirements when combined with the granted mask
4969 	 * from above.  For each privilege, if the privilege is required,
4970 	 * bitwise or the request type onto the priv_granted mask.
4971 	 */
4972 	priv_granted = 0;
4973 
4974 	if (type == VDIR) {
4975 		/*
4976 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4977 		 * requests, instead of PRIV_VFS_EXEC.
4978 		 */
4979 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4980 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4981 			priv_granted |= VEXEC;
4982 	} else {
4983 		/*
4984 		 * Ensure that at least one execute bit is on. Otherwise,
4985 		 * a privileged user will always succeed, and we don't want
4986 		 * this to happen unless the file really is executable.
4987 		 */
4988 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4989 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4990 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4991 			priv_granted |= VEXEC;
4992 	}
4993 
4994 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4995 	    !priv_check_cred(cred, PRIV_VFS_READ))
4996 		priv_granted |= VREAD;
4997 
4998 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4999 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5000 		priv_granted |= (VWRITE | VAPPEND);
5001 
5002 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5003 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5004 		priv_granted |= VADMIN;
5005 
5006 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5007 		/* XXX audit: privilege used */
5008 		if (privused != NULL)
5009 			*privused = 1;
5010 		return (0);
5011 	}
5012 
5013 	return ((accmode & VADMIN) ? EPERM : EACCES);
5014 }
5015 
5016 /*
5017  * Credential check based on process requesting service, and per-attribute
5018  * permissions.
5019  */
5020 int
5021 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5022     struct thread *td, accmode_t accmode)
5023 {
5024 
5025 	/*
5026 	 * Kernel-invoked always succeeds.
5027 	 */
5028 	if (cred == NOCRED)
5029 		return (0);
5030 
5031 	/*
5032 	 * Do not allow privileged processes in jail to directly manipulate
5033 	 * system attributes.
5034 	 */
5035 	switch (attrnamespace) {
5036 	case EXTATTR_NAMESPACE_SYSTEM:
5037 		/* Potentially should be: return (EPERM); */
5038 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5039 	case EXTATTR_NAMESPACE_USER:
5040 		return (VOP_ACCESS(vp, accmode, cred, td));
5041 	default:
5042 		return (EPERM);
5043 	}
5044 }
5045 
5046 #ifdef DEBUG_VFS_LOCKS
5047 /*
5048  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5049  * no longer ok to have an unlocked VFS.
5050  */
5051 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5052 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5053 
5054 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5055 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5056     "Drop into debugger on lock violation");
5057 
5058 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5059 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5060     0, "Check for interlock across VOPs");
5061 
5062 int vfs_badlock_print = 1;	/* Print lock violations. */
5063 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5064     0, "Print lock violations");
5065 
5066 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5067 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5068     0, "Print vnode details on lock violations");
5069 
5070 #ifdef KDB
5071 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5072 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5073     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5074 #endif
5075 
5076 static void
5077 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5078 {
5079 
5080 #ifdef KDB
5081 	if (vfs_badlock_backtrace)
5082 		kdb_backtrace();
5083 #endif
5084 	if (vfs_badlock_vnode)
5085 		vn_printf(vp, "vnode ");
5086 	if (vfs_badlock_print)
5087 		printf("%s: %p %s\n", str, (void *)vp, msg);
5088 	if (vfs_badlock_ddb)
5089 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5090 }
5091 
5092 void
5093 assert_vi_locked(struct vnode *vp, const char *str)
5094 {
5095 
5096 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5097 		vfs_badlock("interlock is not locked but should be", str, vp);
5098 }
5099 
5100 void
5101 assert_vi_unlocked(struct vnode *vp, const char *str)
5102 {
5103 
5104 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5105 		vfs_badlock("interlock is locked but should not be", str, vp);
5106 }
5107 
5108 void
5109 assert_vop_locked(struct vnode *vp, const char *str)
5110 {
5111 	int locked;
5112 
5113 	if (!IGNORE_LOCK(vp)) {
5114 		locked = VOP_ISLOCKED(vp);
5115 		if (locked == 0 || locked == LK_EXCLOTHER)
5116 			vfs_badlock("is not locked but should be", str, vp);
5117 	}
5118 }
5119 
5120 void
5121 assert_vop_unlocked(struct vnode *vp, const char *str)
5122 {
5123 
5124 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5125 		vfs_badlock("is locked but should not be", str, vp);
5126 }
5127 
5128 void
5129 assert_vop_elocked(struct vnode *vp, const char *str)
5130 {
5131 
5132 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5133 		vfs_badlock("is not exclusive locked but should be", str, vp);
5134 }
5135 #endif /* DEBUG_VFS_LOCKS */
5136 
5137 void
5138 vop_rename_fail(struct vop_rename_args *ap)
5139 {
5140 
5141 	if (ap->a_tvp != NULL)
5142 		vput(ap->a_tvp);
5143 	if (ap->a_tdvp == ap->a_tvp)
5144 		vrele(ap->a_tdvp);
5145 	else
5146 		vput(ap->a_tdvp);
5147 	vrele(ap->a_fdvp);
5148 	vrele(ap->a_fvp);
5149 }
5150 
5151 void
5152 vop_rename_pre(void *ap)
5153 {
5154 	struct vop_rename_args *a = ap;
5155 
5156 #ifdef DEBUG_VFS_LOCKS
5157 	if (a->a_tvp)
5158 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5159 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5160 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5161 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5162 
5163 	/* Check the source (from). */
5164 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5165 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5166 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5167 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5168 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5169 
5170 	/* Check the target. */
5171 	if (a->a_tvp)
5172 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5173 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5174 #endif
5175 	if (a->a_tdvp != a->a_fdvp)
5176 		vhold(a->a_fdvp);
5177 	if (a->a_tvp != a->a_fvp)
5178 		vhold(a->a_fvp);
5179 	vhold(a->a_tdvp);
5180 	if (a->a_tvp)
5181 		vhold(a->a_tvp);
5182 }
5183 
5184 #ifdef DEBUG_VFS_LOCKS
5185 void
5186 vop_strategy_pre(void *ap)
5187 {
5188 	struct vop_strategy_args *a;
5189 	struct buf *bp;
5190 
5191 	a = ap;
5192 	bp = a->a_bp;
5193 
5194 	/*
5195 	 * Cluster ops lock their component buffers but not the IO container.
5196 	 */
5197 	if ((bp->b_flags & B_CLUSTER) != 0)
5198 		return;
5199 
5200 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5201 		if (vfs_badlock_print)
5202 			printf(
5203 			    "VOP_STRATEGY: bp is not locked but should be\n");
5204 		if (vfs_badlock_ddb)
5205 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5206 	}
5207 }
5208 
5209 void
5210 vop_lock_pre(void *ap)
5211 {
5212 	struct vop_lock1_args *a = ap;
5213 
5214 	if ((a->a_flags & LK_INTERLOCK) == 0)
5215 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5216 	else
5217 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5218 }
5219 
5220 void
5221 vop_lock_post(void *ap, int rc)
5222 {
5223 	struct vop_lock1_args *a = ap;
5224 
5225 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5226 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5227 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5228 }
5229 
5230 void
5231 vop_unlock_pre(void *ap)
5232 {
5233 	struct vop_unlock_args *a = ap;
5234 
5235 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5236 }
5237 
5238 void
5239 vop_unlock_post(void *ap, int rc)
5240 {
5241 	return;
5242 }
5243 
5244 void
5245 vop_need_inactive_pre(void *ap)
5246 {
5247 	struct vop_need_inactive_args *a = ap;
5248 
5249 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5250 }
5251 
5252 void
5253 vop_need_inactive_post(void *ap, int rc)
5254 {
5255 	struct vop_need_inactive_args *a = ap;
5256 
5257 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5258 }
5259 #endif
5260 
5261 void
5262 vop_create_post(void *ap, int rc)
5263 {
5264 	struct vop_create_args *a = ap;
5265 
5266 	if (!rc)
5267 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5268 }
5269 
5270 void
5271 vop_deleteextattr_post(void *ap, int rc)
5272 {
5273 	struct vop_deleteextattr_args *a = ap;
5274 
5275 	if (!rc)
5276 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5277 }
5278 
5279 void
5280 vop_link_post(void *ap, int rc)
5281 {
5282 	struct vop_link_args *a = ap;
5283 
5284 	if (!rc) {
5285 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5286 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5287 	}
5288 }
5289 
5290 void
5291 vop_mkdir_post(void *ap, int rc)
5292 {
5293 	struct vop_mkdir_args *a = ap;
5294 
5295 	if (!rc)
5296 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5297 }
5298 
5299 void
5300 vop_mknod_post(void *ap, int rc)
5301 {
5302 	struct vop_mknod_args *a = ap;
5303 
5304 	if (!rc)
5305 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5306 }
5307 
5308 void
5309 vop_reclaim_post(void *ap, int rc)
5310 {
5311 	struct vop_reclaim_args *a = ap;
5312 
5313 	if (!rc)
5314 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5315 }
5316 
5317 void
5318 vop_remove_post(void *ap, int rc)
5319 {
5320 	struct vop_remove_args *a = ap;
5321 
5322 	if (!rc) {
5323 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5324 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5325 	}
5326 }
5327 
5328 void
5329 vop_rename_post(void *ap, int rc)
5330 {
5331 	struct vop_rename_args *a = ap;
5332 	long hint;
5333 
5334 	if (!rc) {
5335 		hint = NOTE_WRITE;
5336 		if (a->a_fdvp == a->a_tdvp) {
5337 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5338 				hint |= NOTE_LINK;
5339 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5340 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5341 		} else {
5342 			hint |= NOTE_EXTEND;
5343 			if (a->a_fvp->v_type == VDIR)
5344 				hint |= NOTE_LINK;
5345 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5346 
5347 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5348 			    a->a_tvp->v_type == VDIR)
5349 				hint &= ~NOTE_LINK;
5350 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5351 		}
5352 
5353 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5354 		if (a->a_tvp)
5355 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5356 	}
5357 	if (a->a_tdvp != a->a_fdvp)
5358 		vdrop(a->a_fdvp);
5359 	if (a->a_tvp != a->a_fvp)
5360 		vdrop(a->a_fvp);
5361 	vdrop(a->a_tdvp);
5362 	if (a->a_tvp)
5363 		vdrop(a->a_tvp);
5364 }
5365 
5366 void
5367 vop_rmdir_post(void *ap, int rc)
5368 {
5369 	struct vop_rmdir_args *a = ap;
5370 
5371 	if (!rc) {
5372 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5373 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5374 	}
5375 }
5376 
5377 void
5378 vop_setattr_post(void *ap, int rc)
5379 {
5380 	struct vop_setattr_args *a = ap;
5381 
5382 	if (!rc)
5383 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5384 }
5385 
5386 void
5387 vop_setextattr_post(void *ap, int rc)
5388 {
5389 	struct vop_setextattr_args *a = ap;
5390 
5391 	if (!rc)
5392 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5393 }
5394 
5395 void
5396 vop_symlink_post(void *ap, int rc)
5397 {
5398 	struct vop_symlink_args *a = ap;
5399 
5400 	if (!rc)
5401 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5402 }
5403 
5404 void
5405 vop_open_post(void *ap, int rc)
5406 {
5407 	struct vop_open_args *a = ap;
5408 
5409 	if (!rc)
5410 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5411 }
5412 
5413 void
5414 vop_close_post(void *ap, int rc)
5415 {
5416 	struct vop_close_args *a = ap;
5417 
5418 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5419 	    !VN_IS_DOOMED(a->a_vp))) {
5420 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5421 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5422 	}
5423 }
5424 
5425 void
5426 vop_read_post(void *ap, int rc)
5427 {
5428 	struct vop_read_args *a = ap;
5429 
5430 	if (!rc)
5431 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5432 }
5433 
5434 void
5435 vop_readdir_post(void *ap, int rc)
5436 {
5437 	struct vop_readdir_args *a = ap;
5438 
5439 	if (!rc)
5440 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5441 }
5442 
5443 static struct knlist fs_knlist;
5444 
5445 static void
5446 vfs_event_init(void *arg)
5447 {
5448 	knlist_init_mtx(&fs_knlist, NULL);
5449 }
5450 /* XXX - correct order? */
5451 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5452 
5453 void
5454 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5455 {
5456 
5457 	KNOTE_UNLOCKED(&fs_knlist, event);
5458 }
5459 
5460 static int	filt_fsattach(struct knote *kn);
5461 static void	filt_fsdetach(struct knote *kn);
5462 static int	filt_fsevent(struct knote *kn, long hint);
5463 
5464 struct filterops fs_filtops = {
5465 	.f_isfd = 0,
5466 	.f_attach = filt_fsattach,
5467 	.f_detach = filt_fsdetach,
5468 	.f_event = filt_fsevent
5469 };
5470 
5471 static int
5472 filt_fsattach(struct knote *kn)
5473 {
5474 
5475 	kn->kn_flags |= EV_CLEAR;
5476 	knlist_add(&fs_knlist, kn, 0);
5477 	return (0);
5478 }
5479 
5480 static void
5481 filt_fsdetach(struct knote *kn)
5482 {
5483 
5484 	knlist_remove(&fs_knlist, kn, 0);
5485 }
5486 
5487 static int
5488 filt_fsevent(struct knote *kn, long hint)
5489 {
5490 
5491 	kn->kn_fflags |= hint;
5492 	return (kn->kn_fflags != 0);
5493 }
5494 
5495 static int
5496 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5497 {
5498 	struct vfsidctl vc;
5499 	int error;
5500 	struct mount *mp;
5501 
5502 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5503 	if (error)
5504 		return (error);
5505 	if (vc.vc_vers != VFS_CTL_VERS1)
5506 		return (EINVAL);
5507 	mp = vfs_getvfs(&vc.vc_fsid);
5508 	if (mp == NULL)
5509 		return (ENOENT);
5510 	/* ensure that a specific sysctl goes to the right filesystem. */
5511 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5512 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5513 		vfs_rel(mp);
5514 		return (EINVAL);
5515 	}
5516 	VCTLTOREQ(&vc, req);
5517 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5518 	vfs_rel(mp);
5519 	return (error);
5520 }
5521 
5522 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5523     NULL, 0, sysctl_vfs_ctl, "",
5524     "Sysctl by fsid");
5525 
5526 /*
5527  * Function to initialize a va_filerev field sensibly.
5528  * XXX: Wouldn't a random number make a lot more sense ??
5529  */
5530 u_quad_t
5531 init_va_filerev(void)
5532 {
5533 	struct bintime bt;
5534 
5535 	getbinuptime(&bt);
5536 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5537 }
5538 
5539 static int	filt_vfsread(struct knote *kn, long hint);
5540 static int	filt_vfswrite(struct knote *kn, long hint);
5541 static int	filt_vfsvnode(struct knote *kn, long hint);
5542 static void	filt_vfsdetach(struct knote *kn);
5543 static struct filterops vfsread_filtops = {
5544 	.f_isfd = 1,
5545 	.f_detach = filt_vfsdetach,
5546 	.f_event = filt_vfsread
5547 };
5548 static struct filterops vfswrite_filtops = {
5549 	.f_isfd = 1,
5550 	.f_detach = filt_vfsdetach,
5551 	.f_event = filt_vfswrite
5552 };
5553 static struct filterops vfsvnode_filtops = {
5554 	.f_isfd = 1,
5555 	.f_detach = filt_vfsdetach,
5556 	.f_event = filt_vfsvnode
5557 };
5558 
5559 static void
5560 vfs_knllock(void *arg)
5561 {
5562 	struct vnode *vp = arg;
5563 
5564 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5565 }
5566 
5567 static void
5568 vfs_knlunlock(void *arg)
5569 {
5570 	struct vnode *vp = arg;
5571 
5572 	VOP_UNLOCK(vp);
5573 }
5574 
5575 static void
5576 vfs_knl_assert_locked(void *arg)
5577 {
5578 #ifdef DEBUG_VFS_LOCKS
5579 	struct vnode *vp = arg;
5580 
5581 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5582 #endif
5583 }
5584 
5585 static void
5586 vfs_knl_assert_unlocked(void *arg)
5587 {
5588 #ifdef DEBUG_VFS_LOCKS
5589 	struct vnode *vp = arg;
5590 
5591 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5592 #endif
5593 }
5594 
5595 int
5596 vfs_kqfilter(struct vop_kqfilter_args *ap)
5597 {
5598 	struct vnode *vp = ap->a_vp;
5599 	struct knote *kn = ap->a_kn;
5600 	struct knlist *knl;
5601 
5602 	switch (kn->kn_filter) {
5603 	case EVFILT_READ:
5604 		kn->kn_fop = &vfsread_filtops;
5605 		break;
5606 	case EVFILT_WRITE:
5607 		kn->kn_fop = &vfswrite_filtops;
5608 		break;
5609 	case EVFILT_VNODE:
5610 		kn->kn_fop = &vfsvnode_filtops;
5611 		break;
5612 	default:
5613 		return (EINVAL);
5614 	}
5615 
5616 	kn->kn_hook = (caddr_t)vp;
5617 
5618 	v_addpollinfo(vp);
5619 	if (vp->v_pollinfo == NULL)
5620 		return (ENOMEM);
5621 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5622 	vhold(vp);
5623 	knlist_add(knl, kn, 0);
5624 
5625 	return (0);
5626 }
5627 
5628 /*
5629  * Detach knote from vnode
5630  */
5631 static void
5632 filt_vfsdetach(struct knote *kn)
5633 {
5634 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5635 
5636 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5637 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5638 	vdrop(vp);
5639 }
5640 
5641 /*ARGSUSED*/
5642 static int
5643 filt_vfsread(struct knote *kn, long hint)
5644 {
5645 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5646 	struct vattr va;
5647 	int res;
5648 
5649 	/*
5650 	 * filesystem is gone, so set the EOF flag and schedule
5651 	 * the knote for deletion.
5652 	 */
5653 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5654 		VI_LOCK(vp);
5655 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5656 		VI_UNLOCK(vp);
5657 		return (1);
5658 	}
5659 
5660 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5661 		return (0);
5662 
5663 	VI_LOCK(vp);
5664 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5665 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5666 	VI_UNLOCK(vp);
5667 	return (res);
5668 }
5669 
5670 /*ARGSUSED*/
5671 static int
5672 filt_vfswrite(struct knote *kn, long hint)
5673 {
5674 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5675 
5676 	VI_LOCK(vp);
5677 
5678 	/*
5679 	 * filesystem is gone, so set the EOF flag and schedule
5680 	 * the knote for deletion.
5681 	 */
5682 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5683 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5684 
5685 	kn->kn_data = 0;
5686 	VI_UNLOCK(vp);
5687 	return (1);
5688 }
5689 
5690 static int
5691 filt_vfsvnode(struct knote *kn, long hint)
5692 {
5693 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5694 	int res;
5695 
5696 	VI_LOCK(vp);
5697 	if (kn->kn_sfflags & hint)
5698 		kn->kn_fflags |= hint;
5699 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5700 		kn->kn_flags |= EV_EOF;
5701 		VI_UNLOCK(vp);
5702 		return (1);
5703 	}
5704 	res = (kn->kn_fflags != 0);
5705 	VI_UNLOCK(vp);
5706 	return (res);
5707 }
5708 
5709 /*
5710  * Returns whether the directory is empty or not.
5711  * If it is empty, the return value is 0; otherwise
5712  * the return value is an error value (which may
5713  * be ENOTEMPTY).
5714  */
5715 int
5716 vfs_emptydir(struct vnode *vp)
5717 {
5718 	struct uio uio;
5719 	struct iovec iov;
5720 	struct dirent *dirent, *dp, *endp;
5721 	int error, eof;
5722 
5723 	error = 0;
5724 	eof = 0;
5725 
5726 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5727 
5728 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5729 	iov.iov_base = dirent;
5730 	iov.iov_len = sizeof(struct dirent);
5731 
5732 	uio.uio_iov = &iov;
5733 	uio.uio_iovcnt = 1;
5734 	uio.uio_offset = 0;
5735 	uio.uio_resid = sizeof(struct dirent);
5736 	uio.uio_segflg = UIO_SYSSPACE;
5737 	uio.uio_rw = UIO_READ;
5738 	uio.uio_td = curthread;
5739 
5740 	while (eof == 0 && error == 0) {
5741 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5742 		    NULL, NULL);
5743 		if (error != 0)
5744 			break;
5745 		endp = (void *)((uint8_t *)dirent +
5746 		    sizeof(struct dirent) - uio.uio_resid);
5747 		for (dp = dirent; dp < endp;
5748 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5749 			if (dp->d_type == DT_WHT)
5750 				continue;
5751 			if (dp->d_namlen == 0)
5752 				continue;
5753 			if (dp->d_type != DT_DIR &&
5754 			    dp->d_type != DT_UNKNOWN) {
5755 				error = ENOTEMPTY;
5756 				break;
5757 			}
5758 			if (dp->d_namlen > 2) {
5759 				error = ENOTEMPTY;
5760 				break;
5761 			}
5762 			if (dp->d_namlen == 1 &&
5763 			    dp->d_name[0] != '.') {
5764 				error = ENOTEMPTY;
5765 				break;
5766 			}
5767 			if (dp->d_namlen == 2 &&
5768 			    dp->d_name[1] != '.') {
5769 				error = ENOTEMPTY;
5770 				break;
5771 			}
5772 			uio.uio_resid = sizeof(struct dirent);
5773 		}
5774 	}
5775 	free(dirent, M_TEMP);
5776 	return (error);
5777 }
5778 
5779 int
5780 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5781 {
5782 	int error;
5783 
5784 	if (dp->d_reclen > ap->a_uio->uio_resid)
5785 		return (ENAMETOOLONG);
5786 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5787 	if (error) {
5788 		if (ap->a_ncookies != NULL) {
5789 			if (ap->a_cookies != NULL)
5790 				free(ap->a_cookies, M_TEMP);
5791 			ap->a_cookies = NULL;
5792 			*ap->a_ncookies = 0;
5793 		}
5794 		return (error);
5795 	}
5796 	if (ap->a_ncookies == NULL)
5797 		return (0);
5798 
5799 	KASSERT(ap->a_cookies,
5800 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5801 
5802 	*ap->a_cookies = realloc(*ap->a_cookies,
5803 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5804 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5805 	*ap->a_ncookies += 1;
5806 	return (0);
5807 }
5808 
5809 /*
5810  * Mark for update the access time of the file if the filesystem
5811  * supports VOP_MARKATIME.  This functionality is used by execve and
5812  * mmap, so we want to avoid the I/O implied by directly setting
5813  * va_atime for the sake of efficiency.
5814  */
5815 void
5816 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5817 {
5818 	struct mount *mp;
5819 
5820 	mp = vp->v_mount;
5821 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5822 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5823 		(void)VOP_MARKATIME(vp);
5824 }
5825 
5826 /*
5827  * The purpose of this routine is to remove granularity from accmode_t,
5828  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5829  * VADMIN and VAPPEND.
5830  *
5831  * If it returns 0, the caller is supposed to continue with the usual
5832  * access checks using 'accmode' as modified by this routine.  If it
5833  * returns nonzero value, the caller is supposed to return that value
5834  * as errno.
5835  *
5836  * Note that after this routine runs, accmode may be zero.
5837  */
5838 int
5839 vfs_unixify_accmode(accmode_t *accmode)
5840 {
5841 	/*
5842 	 * There is no way to specify explicit "deny" rule using
5843 	 * file mode or POSIX.1e ACLs.
5844 	 */
5845 	if (*accmode & VEXPLICIT_DENY) {
5846 		*accmode = 0;
5847 		return (0);
5848 	}
5849 
5850 	/*
5851 	 * None of these can be translated into usual access bits.
5852 	 * Also, the common case for NFSv4 ACLs is to not contain
5853 	 * either of these bits. Caller should check for VWRITE
5854 	 * on the containing directory instead.
5855 	 */
5856 	if (*accmode & (VDELETE_CHILD | VDELETE))
5857 		return (EPERM);
5858 
5859 	if (*accmode & VADMIN_PERMS) {
5860 		*accmode &= ~VADMIN_PERMS;
5861 		*accmode |= VADMIN;
5862 	}
5863 
5864 	/*
5865 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5866 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5867 	 */
5868 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5869 
5870 	return (0);
5871 }
5872 
5873 /*
5874  * Clear out a doomed vnode (if any) and replace it with a new one as long
5875  * as the fs is not being unmounted. Return the root vnode to the caller.
5876  */
5877 static int __noinline
5878 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
5879 {
5880 	struct vnode *vp;
5881 	int error;
5882 
5883 restart:
5884 	if (mp->mnt_rootvnode != NULL) {
5885 		MNT_ILOCK(mp);
5886 		vp = mp->mnt_rootvnode;
5887 		if (vp != NULL) {
5888 			if (!VN_IS_DOOMED(vp)) {
5889 				vrefact(vp);
5890 				MNT_IUNLOCK(mp);
5891 				error = vn_lock(vp, flags);
5892 				if (error == 0) {
5893 					*vpp = vp;
5894 					return (0);
5895 				}
5896 				vrele(vp);
5897 				goto restart;
5898 			}
5899 			/*
5900 			 * Clear the old one.
5901 			 */
5902 			mp->mnt_rootvnode = NULL;
5903 		}
5904 		MNT_IUNLOCK(mp);
5905 		if (vp != NULL) {
5906 			/*
5907 			 * Paired with a fence in vfs_op_thread_exit().
5908 			 */
5909 			atomic_thread_fence_acq();
5910 			vfs_op_barrier_wait(mp);
5911 			vrele(vp);
5912 		}
5913 	}
5914 	error = VFS_CACHEDROOT(mp, flags, vpp);
5915 	if (error != 0)
5916 		return (error);
5917 	if (mp->mnt_vfs_ops == 0) {
5918 		MNT_ILOCK(mp);
5919 		if (mp->mnt_vfs_ops != 0) {
5920 			MNT_IUNLOCK(mp);
5921 			return (0);
5922 		}
5923 		if (mp->mnt_rootvnode == NULL) {
5924 			vrefact(*vpp);
5925 			mp->mnt_rootvnode = *vpp;
5926 		} else {
5927 			if (mp->mnt_rootvnode != *vpp) {
5928 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
5929 					panic("%s: mismatch between vnode returned "
5930 					    " by VFS_CACHEDROOT and the one cached "
5931 					    " (%p != %p)",
5932 					    __func__, *vpp, mp->mnt_rootvnode);
5933 				}
5934 			}
5935 		}
5936 		MNT_IUNLOCK(mp);
5937 	}
5938 	return (0);
5939 }
5940 
5941 int
5942 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
5943 {
5944 	struct vnode *vp;
5945 	int error;
5946 
5947 	if (!vfs_op_thread_enter(mp))
5948 		return (vfs_cache_root_fallback(mp, flags, vpp));
5949 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
5950 	if (vp == NULL || VN_IS_DOOMED(vp)) {
5951 		vfs_op_thread_exit(mp);
5952 		return (vfs_cache_root_fallback(mp, flags, vpp));
5953 	}
5954 	vrefact(vp);
5955 	vfs_op_thread_exit(mp);
5956 	error = vn_lock(vp, flags);
5957 	if (error != 0) {
5958 		vrele(vp);
5959 		return (vfs_cache_root_fallback(mp, flags, vpp));
5960 	}
5961 	*vpp = vp;
5962 	return (0);
5963 }
5964 
5965 struct vnode *
5966 vfs_cache_root_clear(struct mount *mp)
5967 {
5968 	struct vnode *vp;
5969 
5970 	/*
5971 	 * ops > 0 guarantees there is nobody who can see this vnode
5972 	 */
5973 	MPASS(mp->mnt_vfs_ops > 0);
5974 	vp = mp->mnt_rootvnode;
5975 	mp->mnt_rootvnode = NULL;
5976 	return (vp);
5977 }
5978 
5979 void
5980 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
5981 {
5982 
5983 	MPASS(mp->mnt_vfs_ops > 0);
5984 	vrefact(vp);
5985 	mp->mnt_rootvnode = vp;
5986 }
5987 
5988 /*
5989  * These are helper functions for filesystems to traverse all
5990  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5991  *
5992  * This interface replaces MNT_VNODE_FOREACH.
5993  */
5994 
5995 
5996 struct vnode *
5997 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5998 {
5999 	struct vnode *vp;
6000 
6001 	if (should_yield())
6002 		kern_yield(PRI_USER);
6003 	MNT_ILOCK(mp);
6004 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6005 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6006 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6007 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6008 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6009 			continue;
6010 		VI_LOCK(vp);
6011 		if (VN_IS_DOOMED(vp)) {
6012 			VI_UNLOCK(vp);
6013 			continue;
6014 		}
6015 		break;
6016 	}
6017 	if (vp == NULL) {
6018 		__mnt_vnode_markerfree_all(mvp, mp);
6019 		/* MNT_IUNLOCK(mp); -- done in above function */
6020 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6021 		return (NULL);
6022 	}
6023 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6024 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6025 	MNT_IUNLOCK(mp);
6026 	return (vp);
6027 }
6028 
6029 struct vnode *
6030 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6031 {
6032 	struct vnode *vp;
6033 
6034 	*mvp = vn_alloc_marker(mp);
6035 	MNT_ILOCK(mp);
6036 	MNT_REF(mp);
6037 
6038 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6039 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6040 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6041 			continue;
6042 		VI_LOCK(vp);
6043 		if (VN_IS_DOOMED(vp)) {
6044 			VI_UNLOCK(vp);
6045 			continue;
6046 		}
6047 		break;
6048 	}
6049 	if (vp == NULL) {
6050 		MNT_REL(mp);
6051 		MNT_IUNLOCK(mp);
6052 		vn_free_marker(*mvp);
6053 		*mvp = NULL;
6054 		return (NULL);
6055 	}
6056 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6057 	MNT_IUNLOCK(mp);
6058 	return (vp);
6059 }
6060 
6061 void
6062 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6063 {
6064 
6065 	if (*mvp == NULL) {
6066 		MNT_IUNLOCK(mp);
6067 		return;
6068 	}
6069 
6070 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6071 
6072 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6073 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6074 	MNT_REL(mp);
6075 	MNT_IUNLOCK(mp);
6076 	vn_free_marker(*mvp);
6077 	*mvp = NULL;
6078 }
6079 
6080 /*
6081  * These are helper functions for filesystems to traverse their
6082  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6083  */
6084 static void
6085 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6086 {
6087 
6088 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6089 
6090 	MNT_ILOCK(mp);
6091 	MNT_REL(mp);
6092 	MNT_IUNLOCK(mp);
6093 	vn_free_marker(*mvp);
6094 	*mvp = NULL;
6095 }
6096 
6097 /*
6098  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6099  * conventional lock order during mnt_vnode_next_lazy iteration.
6100  *
6101  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6102  * The list lock is dropped and reacquired.  On success, both locks are held.
6103  * On failure, the mount vnode list lock is held but the vnode interlock is
6104  * not, and the procedure may have yielded.
6105  */
6106 static bool
6107 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6108     struct vnode *vp)
6109 {
6110 	const struct vnode *tmp;
6111 	bool held, ret;
6112 
6113 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6114 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6115 	    ("%s: bad marker", __func__));
6116 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6117 	    ("%s: inappropriate vnode", __func__));
6118 	ASSERT_VI_UNLOCKED(vp, __func__);
6119 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6120 
6121 	ret = false;
6122 
6123 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6124 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6125 
6126 	/*
6127 	 * Use a hold to prevent vp from disappearing while the mount vnode
6128 	 * list lock is dropped and reacquired.  Normally a hold would be
6129 	 * acquired with vhold(), but that might try to acquire the vnode
6130 	 * interlock, which would be a LOR with the mount vnode list lock.
6131 	 */
6132 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
6133 	mtx_unlock(&mp->mnt_listmtx);
6134 	if (!held)
6135 		goto abort;
6136 	VI_LOCK(vp);
6137 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
6138 		vdropl(vp);
6139 		goto abort;
6140 	}
6141 	mtx_lock(&mp->mnt_listmtx);
6142 
6143 	/*
6144 	 * Determine whether the vnode is still the next one after the marker,
6145 	 * excepting any other markers.  If the vnode has not been doomed by
6146 	 * vgone() then the hold should have ensured that it remained on the
6147 	 * lazy list.  If it has been doomed but is still on the lazy list,
6148 	 * don't abort, but rather skip over it (avoid spinning on doomed
6149 	 * vnodes).
6150 	 */
6151 	tmp = mvp;
6152 	do {
6153 		tmp = TAILQ_NEXT(tmp, v_lazylist);
6154 	} while (tmp != NULL && tmp->v_type == VMARKER);
6155 	if (tmp != vp) {
6156 		mtx_unlock(&mp->mnt_listmtx);
6157 		VI_UNLOCK(vp);
6158 		goto abort;
6159 	}
6160 
6161 	ret = true;
6162 	goto out;
6163 abort:
6164 	maybe_yield();
6165 	mtx_lock(&mp->mnt_listmtx);
6166 out:
6167 	if (ret)
6168 		ASSERT_VI_LOCKED(vp, __func__);
6169 	else
6170 		ASSERT_VI_UNLOCKED(vp, __func__);
6171 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6172 	return (ret);
6173 }
6174 
6175 static struct vnode *
6176 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6177     void *cbarg)
6178 {
6179 	struct vnode *vp, *nvp;
6180 
6181 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6182 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6183 restart:
6184 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6185 	while (vp != NULL) {
6186 		if (vp->v_type == VMARKER) {
6187 			vp = TAILQ_NEXT(vp, v_lazylist);
6188 			continue;
6189 		}
6190 		/*
6191 		 * See if we want to process the vnode. Note we may encounter a
6192 		 * long string of vnodes we don't care about and hog the list
6193 		 * as a result. Check for it and requeue the marker.
6194 		 */
6195 		if (VN_IS_DOOMED(vp) || !cb(vp, cbarg)) {
6196 			if (!should_yield()) {
6197 				vp = TAILQ_NEXT(vp, v_lazylist);
6198 				continue;
6199 			}
6200 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6201 			    v_lazylist);
6202 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6203 			    v_lazylist);
6204 			mtx_unlock(&mp->mnt_listmtx);
6205 			kern_yield(PRI_USER);
6206 			mtx_lock(&mp->mnt_listmtx);
6207 			goto restart;
6208 		}
6209 		/*
6210 		 * Try-lock because this is the wrong lock order.  If that does
6211 		 * not succeed, drop the mount vnode list lock and try to
6212 		 * reacquire it and the vnode interlock in the right order.
6213 		 */
6214 		if (!VI_TRYLOCK(vp) &&
6215 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6216 			goto restart;
6217 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6218 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6219 		    ("alien vnode on the lazy list %p %p", vp, mp));
6220 		if (vp->v_mount == mp && !VN_IS_DOOMED(vp))
6221 			break;
6222 		nvp = TAILQ_NEXT(vp, v_lazylist);
6223 		VI_UNLOCK(vp);
6224 		vp = nvp;
6225 	}
6226 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6227 
6228 	/* Check if we are done */
6229 	if (vp == NULL) {
6230 		mtx_unlock(&mp->mnt_listmtx);
6231 		mnt_vnode_markerfree_lazy(mvp, mp);
6232 		return (NULL);
6233 	}
6234 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6235 	mtx_unlock(&mp->mnt_listmtx);
6236 	ASSERT_VI_LOCKED(vp, "lazy iter");
6237 	return (vp);
6238 }
6239 
6240 struct vnode *
6241 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6242     void *cbarg)
6243 {
6244 
6245 	if (should_yield())
6246 		kern_yield(PRI_USER);
6247 	mtx_lock(&mp->mnt_listmtx);
6248 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6249 }
6250 
6251 struct vnode *
6252 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6253     void *cbarg)
6254 {
6255 	struct vnode *vp;
6256 
6257 	*mvp = vn_alloc_marker(mp);
6258 	MNT_ILOCK(mp);
6259 	MNT_REF(mp);
6260 	MNT_IUNLOCK(mp);
6261 
6262 	mtx_lock(&mp->mnt_listmtx);
6263 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6264 	if (vp == NULL) {
6265 		mtx_unlock(&mp->mnt_listmtx);
6266 		mnt_vnode_markerfree_lazy(mvp, mp);
6267 		return (NULL);
6268 	}
6269 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6270 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6271 }
6272 
6273 void
6274 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6275 {
6276 
6277 	if (*mvp == NULL)
6278 		return;
6279 
6280 	mtx_lock(&mp->mnt_listmtx);
6281 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6282 	mtx_unlock(&mp->mnt_listmtx);
6283 	mnt_vnode_markerfree_lazy(mvp, mp);
6284 }
6285