xref: /freebsd/sys/kern/vfs_subr.c (revision 580744621f33383027108364dcadad718df46ffe)
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 static struct vnode *vnode_list_reclaim_marker;
170 
171 /*
172  * "Free" vnode target.  Free vnodes are rarely completely free, but are
173  * just ones that are cheap to recycle.  Usually they are for files which
174  * have been stat'd but not read; these usually have inode and namecache
175  * data attached to them.  This target is the preferred minimum size of a
176  * sub-cache consisting mostly of such files. The system balances the size
177  * of this sub-cache with its complement to try to prevent either from
178  * thrashing while the other is relatively inactive.  The targets express
179  * a preference for the best balance.
180  *
181  * "Above" this target there are 2 further targets (watermarks) related
182  * to recyling of free vnodes.  In the best-operating case, the cache is
183  * exactly full, the free list has size between vlowat and vhiwat above the
184  * free target, and recycling from it and normal use maintains this state.
185  * Sometimes the free list is below vlowat or even empty, but this state
186  * is even better for immediate use provided the cache is not full.
187  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
188  * ones) to reach one of these states.  The watermarks are currently hard-
189  * coded as 4% and 9% of the available space higher.  These and the default
190  * of 25% for wantfreevnodes are too large if the memory size is large.
191  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
192  * whenever vnlru_proc() becomes active.
193  */
194 static long wantfreevnodes;
195 static long __exclusive_cache_line freevnodes;
196 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
197     &freevnodes, 0, "Number of \"free\" vnodes");
198 static long freevnodes_old;
199 
200 static counter_u64_t recycles_count;
201 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
202     "Number of vnodes recycled to meet vnode cache targets");
203 
204 static counter_u64_t recycles_free_count;
205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
206     "Number of free vnodes recycled to meet vnode cache targets");
207 
208 /*
209  * Various variables used for debugging the new implementation of
210  * reassignbuf().
211  * XXX these are probably of (very) limited utility now.
212  */
213 static int reassignbufcalls;
214 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
215     &reassignbufcalls, 0, "Number of calls to reassignbuf");
216 
217 static counter_u64_t deferred_inact;
218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
219     "Number of times inactive processing was deferred");
220 
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
223 
224 /*
225  * Lock for any access to the following:
226  *	vnode_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx __exclusive_cache_line vnode_list_mtx;
231 
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
234 
235 static uma_zone_t buf_trie_zone;
236 
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
240 
241 /*
242  * The workitem queue.
243  *
244  * It is useful to delay writes of file data and filesystem metadata
245  * for tens of seconds so that quickly created and deleted files need
246  * not waste disk bandwidth being created and removed. To realize this,
247  * we append vnodes to a "workitem" queue. When running with a soft
248  * updates implementation, most pending metadata dependencies should
249  * not wait for more than a few seconds. Thus, mounted on block devices
250  * are delayed only about a half the time that file data is delayed.
251  * Similarly, directory updates are more critical, so are only delayed
252  * about a third the time that file data is delayed. Thus, there are
253  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254  * one each second (driven off the filesystem syncer process). The
255  * syncer_delayno variable indicates the next queue that is to be processed.
256  * Items that need to be processed soon are placed in this queue:
257  *
258  *	syncer_workitem_pending[syncer_delayno]
259  *
260  * A delay of fifteen seconds is done by placing the request fifteen
261  * entries later in the queue:
262  *
263  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264  *
265  */
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
270 /*
271  * The sync_mtx protects:
272  *	bo->bo_synclist
273  *	sync_vnode_count
274  *	syncer_delayno
275  *	syncer_state
276  *	syncer_workitem_pending
277  *	syncer_worklist_len
278  *	rushjob
279  */
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
282 
283 #define SYNCER_MAXDELAY		32
284 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285 static int syncdelay = 30;		/* max time to delay syncing data */
286 static int filedelay = 30;		/* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288     "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29;		/* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291     "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28;		/* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294     "Time to delay syncing metadata (in seconds)");
295 static int rushjob;		/* number of slots to run ASAP */
296 static int stat_rush_requests;	/* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298     "Number of times I/O speeded up (rush requests)");
299 
300 #define	VDBATCH_SIZE 8
301 struct vdbatch {
302 	u_int index;
303 	long freevnodes;
304 	struct mtx lock;
305 	struct vnode *tab[VDBATCH_SIZE];
306 };
307 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
308 
309 static void	vdbatch_dequeue(struct vnode *vp);
310 
311 /*
312  * When shutting down the syncer, run it at four times normal speed.
313  */
314 #define SYNCER_SHUTDOWN_SPEEDUP		4
315 static int sync_vnode_count;
316 static int syncer_worklist_len;
317 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
318     syncer_state;
319 
320 /* Target for maximum number of vnodes. */
321 u_long desiredvnodes;
322 static u_long gapvnodes;		/* gap between wanted and desired */
323 static u_long vhiwat;		/* enough extras after expansion */
324 static u_long vlowat;		/* minimal extras before expansion */
325 static u_long vstir;		/* nonzero to stir non-free vnodes */
326 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
327 
328 static u_long vnlru_read_freevnodes(void);
329 
330 /*
331  * Note that no attempt is made to sanitize these parameters.
332  */
333 static int
334 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
335 {
336 	u_long val;
337 	int error;
338 
339 	val = desiredvnodes;
340 	error = sysctl_handle_long(oidp, &val, 0, req);
341 	if (error != 0 || req->newptr == NULL)
342 		return (error);
343 
344 	if (val == desiredvnodes)
345 		return (0);
346 	mtx_lock(&vnode_list_mtx);
347 	desiredvnodes = val;
348 	wantfreevnodes = desiredvnodes / 4;
349 	vnlru_recalc();
350 	mtx_unlock(&vnode_list_mtx);
351 	/*
352 	 * XXX There is no protection against multiple threads changing
353 	 * desiredvnodes at the same time. Locking above only helps vnlru and
354 	 * getnewvnode.
355 	 */
356 	vfs_hash_changesize(desiredvnodes);
357 	cache_changesize(desiredvnodes);
358 	return (0);
359 }
360 
361 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
362     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
363     "LU", "Target for maximum number of vnodes");
364 
365 static int
366 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
367 {
368 	u_long val;
369 	int error;
370 
371 	val = wantfreevnodes;
372 	error = sysctl_handle_long(oidp, &val, 0, req);
373 	if (error != 0 || req->newptr == NULL)
374 		return (error);
375 
376 	if (val == wantfreevnodes)
377 		return (0);
378 	mtx_lock(&vnode_list_mtx);
379 	wantfreevnodes = val;
380 	vnlru_recalc();
381 	mtx_unlock(&vnode_list_mtx);
382 	return (0);
383 }
384 
385 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
386     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
387     "LU", "Target for minimum number of \"free\" vnodes");
388 
389 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
390     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
391 static int vnlru_nowhere;
392 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
393     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
394 
395 static int
396 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
397 {
398 	struct vnode *vp;
399 	struct nameidata nd;
400 	char *buf;
401 	unsigned long ndflags;
402 	int error;
403 
404 	if (req->newptr == NULL)
405 		return (EINVAL);
406 	if (req->newlen >= PATH_MAX)
407 		return (E2BIG);
408 
409 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
410 	error = SYSCTL_IN(req, buf, req->newlen);
411 	if (error != 0)
412 		goto out;
413 
414 	buf[req->newlen] = '\0';
415 
416 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
417 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
418 	if ((error = namei(&nd)) != 0)
419 		goto out;
420 	vp = nd.ni_vp;
421 
422 	if (VN_IS_DOOMED(vp)) {
423 		/*
424 		 * This vnode is being recycled.  Return != 0 to let the caller
425 		 * know that the sysctl had no effect.  Return EAGAIN because a
426 		 * subsequent call will likely succeed (since namei will create
427 		 * a new vnode if necessary)
428 		 */
429 		error = EAGAIN;
430 		goto putvnode;
431 	}
432 
433 	counter_u64_add(recycles_count, 1);
434 	vgone(vp);
435 putvnode:
436 	NDFREE(&nd, 0);
437 out:
438 	free(buf, M_TEMP);
439 	return (error);
440 }
441 
442 static int
443 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
444 {
445 	struct thread *td = curthread;
446 	struct vnode *vp;
447 	struct file *fp;
448 	int error;
449 	int fd;
450 
451 	if (req->newptr == NULL)
452 		return (EBADF);
453 
454         error = sysctl_handle_int(oidp, &fd, 0, req);
455         if (error != 0)
456                 return (error);
457 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
458 	if (error != 0)
459 		return (error);
460 	vp = fp->f_vnode;
461 
462 	error = vn_lock(vp, LK_EXCLUSIVE);
463 	if (error != 0)
464 		goto drop;
465 
466 	counter_u64_add(recycles_count, 1);
467 	vgone(vp);
468 	VOP_UNLOCK(vp);
469 drop:
470 	fdrop(fp, td);
471 	return (error);
472 }
473 
474 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
475     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
476     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
477 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
478     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
479     sysctl_ftry_reclaim_vnode, "I",
480     "Try to reclaim a vnode by its file descriptor");
481 
482 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
483 static int vnsz2log;
484 
485 /*
486  * Support for the bufobj clean & dirty pctrie.
487  */
488 static void *
489 buf_trie_alloc(struct pctrie *ptree)
490 {
491 
492 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
493 }
494 
495 static void
496 buf_trie_free(struct pctrie *ptree, void *node)
497 {
498 
499 	uma_zfree(buf_trie_zone, node);
500 }
501 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
502 
503 /*
504  * Initialize the vnode management data structures.
505  *
506  * Reevaluate the following cap on the number of vnodes after the physical
507  * memory size exceeds 512GB.  In the limit, as the physical memory size
508  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
509  */
510 #ifndef	MAXVNODES_MAX
511 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
512 #endif
513 
514 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
515 
516 static struct vnode *
517 vn_alloc_marker(struct mount *mp)
518 {
519 	struct vnode *vp;
520 
521 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
522 	vp->v_type = VMARKER;
523 	vp->v_mount = mp;
524 
525 	return (vp);
526 }
527 
528 static void
529 vn_free_marker(struct vnode *vp)
530 {
531 
532 	MPASS(vp->v_type == VMARKER);
533 	free(vp, M_VNODE_MARKER);
534 }
535 
536 /*
537  * Initialize a vnode as it first enters the zone.
538  */
539 static int
540 vnode_init(void *mem, int size, int flags)
541 {
542 	struct vnode *vp;
543 
544 	vp = mem;
545 	bzero(vp, size);
546 	/*
547 	 * Setup locks.
548 	 */
549 	vp->v_vnlock = &vp->v_lock;
550 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
551 	/*
552 	 * By default, don't allow shared locks unless filesystems opt-in.
553 	 */
554 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
555 	    LK_NOSHARE | LK_IS_VNODE);
556 	/*
557 	 * Initialize bufobj.
558 	 */
559 	bufobj_init(&vp->v_bufobj, vp);
560 	/*
561 	 * Initialize namecache.
562 	 */
563 	LIST_INIT(&vp->v_cache_src);
564 	TAILQ_INIT(&vp->v_cache_dst);
565 	/*
566 	 * Initialize rangelocks.
567 	 */
568 	rangelock_init(&vp->v_rl);
569 
570 	vp->v_dbatchcpu = NOCPU;
571 
572 	mtx_lock(&vnode_list_mtx);
573 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
574 	mtx_unlock(&vnode_list_mtx);
575 	return (0);
576 }
577 
578 /*
579  * Free a vnode when it is cleared from the zone.
580  */
581 static void
582 vnode_fini(void *mem, int size)
583 {
584 	struct vnode *vp;
585 	struct bufobj *bo;
586 
587 	vp = mem;
588 	vdbatch_dequeue(vp);
589 	mtx_lock(&vnode_list_mtx);
590 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
591 	mtx_unlock(&vnode_list_mtx);
592 	rangelock_destroy(&vp->v_rl);
593 	lockdestroy(vp->v_vnlock);
594 	mtx_destroy(&vp->v_interlock);
595 	bo = &vp->v_bufobj;
596 	rw_destroy(BO_LOCKPTR(bo));
597 }
598 
599 /*
600  * Provide the size of NFS nclnode and NFS fh for calculation of the
601  * vnode memory consumption.  The size is specified directly to
602  * eliminate dependency on NFS-private header.
603  *
604  * Other filesystems may use bigger or smaller (like UFS and ZFS)
605  * private inode data, but the NFS-based estimation is ample enough.
606  * Still, we care about differences in the size between 64- and 32-bit
607  * platforms.
608  *
609  * Namecache structure size is heuristically
610  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
611  */
612 #ifdef _LP64
613 #define	NFS_NCLNODE_SZ	(528 + 64)
614 #define	NC_SZ		148
615 #else
616 #define	NFS_NCLNODE_SZ	(360 + 32)
617 #define	NC_SZ		92
618 #endif
619 
620 static void
621 vntblinit(void *dummy __unused)
622 {
623 	struct vdbatch *vd;
624 	int cpu, physvnodes, virtvnodes;
625 	u_int i;
626 
627 	/*
628 	 * Desiredvnodes is a function of the physical memory size and the
629 	 * kernel's heap size.  Generally speaking, it scales with the
630 	 * physical memory size.  The ratio of desiredvnodes to the physical
631 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
632 	 * Thereafter, the
633 	 * marginal ratio of desiredvnodes to the physical memory size is
634 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
635 	 * size.  The memory required by desiredvnodes vnodes and vm objects
636 	 * must not exceed 1/10th of the kernel's heap size.
637 	 */
638 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
639 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
640 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
641 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
642 	desiredvnodes = min(physvnodes, virtvnodes);
643 	if (desiredvnodes > MAXVNODES_MAX) {
644 		if (bootverbose)
645 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
646 			    desiredvnodes, MAXVNODES_MAX);
647 		desiredvnodes = MAXVNODES_MAX;
648 	}
649 	wantfreevnodes = desiredvnodes / 4;
650 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
651 	TAILQ_INIT(&vnode_list);
652 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
653 	/*
654 	 * The lock is taken to appease WITNESS.
655 	 */
656 	mtx_lock(&vnode_list_mtx);
657 	vnlru_recalc();
658 	mtx_unlock(&vnode_list_mtx);
659 	vnode_list_free_marker = vn_alloc_marker(NULL);
660 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
661 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
662 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
663 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
664 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
665 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
666 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
667 	/*
668 	 * Preallocate enough nodes to support one-per buf so that
669 	 * we can not fail an insert.  reassignbuf() callers can not
670 	 * tolerate the insertion failure.
671 	 */
672 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
673 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
674 	    UMA_ZONE_NOFREE);
675 	uma_prealloc(buf_trie_zone, nbuf);
676 
677 	vnodes_created = counter_u64_alloc(M_WAITOK);
678 	recycles_count = counter_u64_alloc(M_WAITOK);
679 	recycles_free_count = counter_u64_alloc(M_WAITOK);
680 	deferred_inact = counter_u64_alloc(M_WAITOK);
681 
682 	/*
683 	 * Initialize the filesystem syncer.
684 	 */
685 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
686 	    &syncer_mask);
687 	syncer_maxdelay = syncer_mask + 1;
688 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
689 	cv_init(&sync_wakeup, "syncer");
690 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
691 		vnsz2log++;
692 	vnsz2log--;
693 
694 	CPU_FOREACH(cpu) {
695 		vd = DPCPU_ID_PTR((cpu), vd);
696 		bzero(vd, sizeof(*vd));
697 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
698 	}
699 }
700 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
701 
702 /*
703  * Mark a mount point as busy. Used to synchronize access and to delay
704  * unmounting. Eventually, mountlist_mtx is not released on failure.
705  *
706  * vfs_busy() is a custom lock, it can block the caller.
707  * vfs_busy() only sleeps if the unmount is active on the mount point.
708  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
709  * vnode belonging to mp.
710  *
711  * Lookup uses vfs_busy() to traverse mount points.
712  * root fs			var fs
713  * / vnode lock		A	/ vnode lock (/var)		D
714  * /var vnode lock	B	/log vnode lock(/var/log)	E
715  * vfs_busy lock	C	vfs_busy lock			F
716  *
717  * Within each file system, the lock order is C->A->B and F->D->E.
718  *
719  * When traversing across mounts, the system follows that lock order:
720  *
721  *        C->A->B
722  *              |
723  *              +->F->D->E
724  *
725  * The lookup() process for namei("/var") illustrates the process:
726  *  VOP_LOOKUP() obtains B while A is held
727  *  vfs_busy() obtains a shared lock on F while A and B are held
728  *  vput() releases lock on B
729  *  vput() releases lock on A
730  *  VFS_ROOT() obtains lock on D while shared lock on F is held
731  *  vfs_unbusy() releases shared lock on F
732  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
733  *    Attempt to lock A (instead of vp_crossmp) while D is held would
734  *    violate the global order, causing deadlocks.
735  *
736  * dounmount() locks B while F is drained.
737  */
738 int
739 vfs_busy(struct mount *mp, int flags)
740 {
741 
742 	MPASS((flags & ~MBF_MASK) == 0);
743 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
744 
745 	if (vfs_op_thread_enter(mp)) {
746 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
747 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
748 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
749 		vfs_mp_count_add_pcpu(mp, ref, 1);
750 		vfs_mp_count_add_pcpu(mp, lockref, 1);
751 		vfs_op_thread_exit(mp);
752 		if (flags & MBF_MNTLSTLOCK)
753 			mtx_unlock(&mountlist_mtx);
754 		return (0);
755 	}
756 
757 	MNT_ILOCK(mp);
758 	vfs_assert_mount_counters(mp);
759 	MNT_REF(mp);
760 	/*
761 	 * If mount point is currently being unmounted, sleep until the
762 	 * mount point fate is decided.  If thread doing the unmounting fails,
763 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
764 	 * that this mount point has survived the unmount attempt and vfs_busy
765 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
766 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
767 	 * about to be really destroyed.  vfs_busy needs to release its
768 	 * reference on the mount point in this case and return with ENOENT,
769 	 * telling the caller that mount mount it tried to busy is no longer
770 	 * valid.
771 	 */
772 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
773 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
774 			MNT_REL(mp);
775 			MNT_IUNLOCK(mp);
776 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
777 			    __func__);
778 			return (ENOENT);
779 		}
780 		if (flags & MBF_MNTLSTLOCK)
781 			mtx_unlock(&mountlist_mtx);
782 		mp->mnt_kern_flag |= MNTK_MWAIT;
783 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
784 		if (flags & MBF_MNTLSTLOCK)
785 			mtx_lock(&mountlist_mtx);
786 		MNT_ILOCK(mp);
787 	}
788 	if (flags & MBF_MNTLSTLOCK)
789 		mtx_unlock(&mountlist_mtx);
790 	mp->mnt_lockref++;
791 	MNT_IUNLOCK(mp);
792 	return (0);
793 }
794 
795 /*
796  * Free a busy filesystem.
797  */
798 void
799 vfs_unbusy(struct mount *mp)
800 {
801 	int c;
802 
803 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
804 
805 	if (vfs_op_thread_enter(mp)) {
806 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
807 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
808 		vfs_mp_count_sub_pcpu(mp, ref, 1);
809 		vfs_op_thread_exit(mp);
810 		return;
811 	}
812 
813 	MNT_ILOCK(mp);
814 	vfs_assert_mount_counters(mp);
815 	MNT_REL(mp);
816 	c = --mp->mnt_lockref;
817 	if (mp->mnt_vfs_ops == 0) {
818 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
819 		MNT_IUNLOCK(mp);
820 		return;
821 	}
822 	if (c < 0)
823 		vfs_dump_mount_counters(mp);
824 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
825 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
826 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
827 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
828 		wakeup(&mp->mnt_lockref);
829 	}
830 	MNT_IUNLOCK(mp);
831 }
832 
833 /*
834  * Lookup a mount point by filesystem identifier.
835  */
836 struct mount *
837 vfs_getvfs(fsid_t *fsid)
838 {
839 	struct mount *mp;
840 
841 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
842 	mtx_lock(&mountlist_mtx);
843 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
844 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
845 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
846 			vfs_ref(mp);
847 			mtx_unlock(&mountlist_mtx);
848 			return (mp);
849 		}
850 	}
851 	mtx_unlock(&mountlist_mtx);
852 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
853 	return ((struct mount *) 0);
854 }
855 
856 /*
857  * Lookup a mount point by filesystem identifier, busying it before
858  * returning.
859  *
860  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
861  * cache for popular filesystem identifiers.  The cache is lockess, using
862  * the fact that struct mount's are never freed.  In worst case we may
863  * get pointer to unmounted or even different filesystem, so we have to
864  * check what we got, and go slow way if so.
865  */
866 struct mount *
867 vfs_busyfs(fsid_t *fsid)
868 {
869 #define	FSID_CACHE_SIZE	256
870 	typedef struct mount * volatile vmp_t;
871 	static vmp_t cache[FSID_CACHE_SIZE];
872 	struct mount *mp;
873 	int error;
874 	uint32_t hash;
875 
876 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
877 	hash = fsid->val[0] ^ fsid->val[1];
878 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
879 	mp = cache[hash];
880 	if (mp == NULL ||
881 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
882 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
883 		goto slow;
884 	if (vfs_busy(mp, 0) != 0) {
885 		cache[hash] = NULL;
886 		goto slow;
887 	}
888 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
889 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
890 		return (mp);
891 	else
892 	    vfs_unbusy(mp);
893 
894 slow:
895 	mtx_lock(&mountlist_mtx);
896 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
897 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
898 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
899 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
900 			if (error) {
901 				cache[hash] = NULL;
902 				mtx_unlock(&mountlist_mtx);
903 				return (NULL);
904 			}
905 			cache[hash] = mp;
906 			return (mp);
907 		}
908 	}
909 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
910 	mtx_unlock(&mountlist_mtx);
911 	return ((struct mount *) 0);
912 }
913 
914 /*
915  * Check if a user can access privileged mount options.
916  */
917 int
918 vfs_suser(struct mount *mp, struct thread *td)
919 {
920 	int error;
921 
922 	if (jailed(td->td_ucred)) {
923 		/*
924 		 * If the jail of the calling thread lacks permission for
925 		 * this type of file system, deny immediately.
926 		 */
927 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
928 			return (EPERM);
929 
930 		/*
931 		 * If the file system was mounted outside the jail of the
932 		 * calling thread, deny immediately.
933 		 */
934 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
935 			return (EPERM);
936 	}
937 
938 	/*
939 	 * If file system supports delegated administration, we don't check
940 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
941 	 * by the file system itself.
942 	 * If this is not the user that did original mount, we check for
943 	 * the PRIV_VFS_MOUNT_OWNER privilege.
944 	 */
945 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
946 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
947 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
948 			return (error);
949 	}
950 	return (0);
951 }
952 
953 /*
954  * Get a new unique fsid.  Try to make its val[0] unique, since this value
955  * will be used to create fake device numbers for stat().  Also try (but
956  * not so hard) make its val[0] unique mod 2^16, since some emulators only
957  * support 16-bit device numbers.  We end up with unique val[0]'s for the
958  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
959  *
960  * Keep in mind that several mounts may be running in parallel.  Starting
961  * the search one past where the previous search terminated is both a
962  * micro-optimization and a defense against returning the same fsid to
963  * different mounts.
964  */
965 void
966 vfs_getnewfsid(struct mount *mp)
967 {
968 	static uint16_t mntid_base;
969 	struct mount *nmp;
970 	fsid_t tfsid;
971 	int mtype;
972 
973 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
974 	mtx_lock(&mntid_mtx);
975 	mtype = mp->mnt_vfc->vfc_typenum;
976 	tfsid.val[1] = mtype;
977 	mtype = (mtype & 0xFF) << 24;
978 	for (;;) {
979 		tfsid.val[0] = makedev(255,
980 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
981 		mntid_base++;
982 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
983 			break;
984 		vfs_rel(nmp);
985 	}
986 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
987 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
988 	mtx_unlock(&mntid_mtx);
989 }
990 
991 /*
992  * Knob to control the precision of file timestamps:
993  *
994  *   0 = seconds only; nanoseconds zeroed.
995  *   1 = seconds and nanoseconds, accurate within 1/HZ.
996  *   2 = seconds and nanoseconds, truncated to microseconds.
997  * >=3 = seconds and nanoseconds, maximum precision.
998  */
999 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1000 
1001 static int timestamp_precision = TSP_USEC;
1002 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1003     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1004     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1005     "3+: sec + ns (max. precision))");
1006 
1007 /*
1008  * Get a current timestamp.
1009  */
1010 void
1011 vfs_timestamp(struct timespec *tsp)
1012 {
1013 	struct timeval tv;
1014 
1015 	switch (timestamp_precision) {
1016 	case TSP_SEC:
1017 		tsp->tv_sec = time_second;
1018 		tsp->tv_nsec = 0;
1019 		break;
1020 	case TSP_HZ:
1021 		getnanotime(tsp);
1022 		break;
1023 	case TSP_USEC:
1024 		microtime(&tv);
1025 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1026 		break;
1027 	case TSP_NSEC:
1028 	default:
1029 		nanotime(tsp);
1030 		break;
1031 	}
1032 }
1033 
1034 /*
1035  * Set vnode attributes to VNOVAL
1036  */
1037 void
1038 vattr_null(struct vattr *vap)
1039 {
1040 
1041 	vap->va_type = VNON;
1042 	vap->va_size = VNOVAL;
1043 	vap->va_bytes = VNOVAL;
1044 	vap->va_mode = VNOVAL;
1045 	vap->va_nlink = VNOVAL;
1046 	vap->va_uid = VNOVAL;
1047 	vap->va_gid = VNOVAL;
1048 	vap->va_fsid = VNOVAL;
1049 	vap->va_fileid = VNOVAL;
1050 	vap->va_blocksize = VNOVAL;
1051 	vap->va_rdev = VNOVAL;
1052 	vap->va_atime.tv_sec = VNOVAL;
1053 	vap->va_atime.tv_nsec = VNOVAL;
1054 	vap->va_mtime.tv_sec = VNOVAL;
1055 	vap->va_mtime.tv_nsec = VNOVAL;
1056 	vap->va_ctime.tv_sec = VNOVAL;
1057 	vap->va_ctime.tv_nsec = VNOVAL;
1058 	vap->va_birthtime.tv_sec = VNOVAL;
1059 	vap->va_birthtime.tv_nsec = VNOVAL;
1060 	vap->va_flags = VNOVAL;
1061 	vap->va_gen = VNOVAL;
1062 	vap->va_vaflags = 0;
1063 }
1064 
1065 /*
1066  * Try to reduce the total number of vnodes.
1067  *
1068  * This routine (and its user) are buggy in at least the following ways:
1069  * - all parameters were picked years ago when RAM sizes were significantly
1070  *   smaller
1071  * - it can pick vnodes based on pages used by the vm object, but filesystems
1072  *   like ZFS don't use it making the pick broken
1073  * - since ZFS has its own aging policy it gets partially combated by this one
1074  * - a dedicated method should be provided for filesystems to let them decide
1075  *   whether the vnode should be recycled
1076  *
1077  * This routine is called when we have too many vnodes.  It attempts
1078  * to free <count> vnodes and will potentially free vnodes that still
1079  * have VM backing store (VM backing store is typically the cause
1080  * of a vnode blowout so we want to do this).  Therefore, this operation
1081  * is not considered cheap.
1082  *
1083  * A number of conditions may prevent a vnode from being reclaimed.
1084  * the buffer cache may have references on the vnode, a directory
1085  * vnode may still have references due to the namei cache representing
1086  * underlying files, or the vnode may be in active use.   It is not
1087  * desirable to reuse such vnodes.  These conditions may cause the
1088  * number of vnodes to reach some minimum value regardless of what
1089  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1090  *
1091  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1092  * 			 entries if this argument is strue
1093  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1094  *			 pages.
1095  * @param target	 How many vnodes to reclaim.
1096  * @return		 The number of vnodes that were reclaimed.
1097  */
1098 static int
1099 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1100 {
1101 	struct vnode *vp, *mvp;
1102 	struct mount *mp;
1103 	struct vm_object *object;
1104 	u_long done;
1105 	bool retried;
1106 
1107 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1108 
1109 	retried = false;
1110 	done = 0;
1111 
1112 	mvp = vnode_list_reclaim_marker;
1113 restart:
1114 	vp = mvp;
1115 	while (done < target) {
1116 		vp = TAILQ_NEXT(vp, v_vnodelist);
1117 		if (__predict_false(vp == NULL))
1118 			break;
1119 
1120 		if (__predict_false(vp->v_type == VMARKER))
1121 			continue;
1122 
1123 		/*
1124 		 * If it's been deconstructed already, it's still
1125 		 * referenced, or it exceeds the trigger, skip it.
1126 		 * Also skip free vnodes.  We are trying to make space
1127 		 * to expand the free list, not reduce it.
1128 		 */
1129 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1130 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1131 			goto next_iter;
1132 
1133 		if (vp->v_type == VBAD || vp->v_type == VNON)
1134 			goto next_iter;
1135 
1136 		if (!VI_TRYLOCK(vp))
1137 			goto next_iter;
1138 
1139 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1140 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1141 		    VN_IS_DOOMED(vp) || vp->v_type == VNON) {
1142 			VI_UNLOCK(vp);
1143 			goto next_iter;
1144 		}
1145 
1146 		object = atomic_load_ptr(&vp->v_object);
1147 		if (object == NULL || object->resident_page_count > trigger) {
1148 			VI_UNLOCK(vp);
1149 			goto next_iter;
1150 		}
1151 
1152 		vholdl(vp);
1153 		VI_UNLOCK(vp);
1154 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1155 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1156 		mtx_unlock(&vnode_list_mtx);
1157 
1158 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1159 			vdrop(vp);
1160 			goto next_iter_unlocked;
1161 		}
1162 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1163 			vdrop(vp);
1164 			vn_finished_write(mp);
1165 			goto next_iter_unlocked;
1166 		}
1167 
1168 		VI_LOCK(vp);
1169 		if (vp->v_usecount > 0 ||
1170 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1171 		    (vp->v_object != NULL &&
1172 		    vp->v_object->resident_page_count > trigger)) {
1173 			VOP_UNLOCK(vp);
1174 			vdropl(vp);
1175 			vn_finished_write(mp);
1176 			goto next_iter_unlocked;
1177 		}
1178 		counter_u64_add(recycles_count, 1);
1179 		vgonel(vp);
1180 		VOP_UNLOCK(vp);
1181 		vdropl(vp);
1182 		vn_finished_write(mp);
1183 		done++;
1184 next_iter_unlocked:
1185 		if (should_yield())
1186 			kern_yield(PRI_USER);
1187 		mtx_lock(&vnode_list_mtx);
1188 		goto restart;
1189 next_iter:
1190 		MPASS(vp->v_type != VMARKER);
1191 		if (!should_yield())
1192 			continue;
1193 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1194 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1195 		mtx_unlock(&vnode_list_mtx);
1196 		kern_yield(PRI_USER);
1197 		mtx_lock(&vnode_list_mtx);
1198 		goto restart;
1199 	}
1200 	if (done == 0 && !retried) {
1201 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1202 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1203 		retried = true;
1204 		goto restart;
1205 	}
1206 	return (done);
1207 }
1208 
1209 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1210 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1211     0,
1212     "limit on vnode free requests per call to the vnlru_free routine");
1213 
1214 /*
1215  * Attempt to reduce the free list by the requested amount.
1216  */
1217 static int
1218 vnlru_free_locked(int count, struct vfsops *mnt_op)
1219 {
1220 	struct vnode *vp, *mvp;
1221 	struct mount *mp;
1222 	int ocount;
1223 
1224 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1225 	if (count > max_vnlru_free)
1226 		count = max_vnlru_free;
1227 	ocount = count;
1228 	mvp = vnode_list_free_marker;
1229 restart:
1230 	vp = mvp;
1231 	while (count > 0) {
1232 		vp = TAILQ_NEXT(vp, v_vnodelist);
1233 		if (__predict_false(vp == NULL)) {
1234 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1235 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1236 			break;
1237 		}
1238 		if (__predict_false(vp->v_type == VMARKER))
1239 			continue;
1240 
1241 		/*
1242 		 * Don't recycle if our vnode is from different type
1243 		 * of mount point.  Note that mp is type-safe, the
1244 		 * check does not reach unmapped address even if
1245 		 * vnode is reclaimed.
1246 		 * Don't recycle if we can't get the interlock without
1247 		 * blocking.
1248 		 */
1249 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1250 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1251 			continue;
1252 		}
1253 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1254 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1255 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1256 			VI_UNLOCK(vp);
1257 			continue;
1258 		}
1259 		vholdl(vp);
1260 		count--;
1261 		mtx_unlock(&vnode_list_mtx);
1262 		VI_UNLOCK(vp);
1263 		vtryrecycle(vp);
1264 		vdrop(vp);
1265 		mtx_lock(&vnode_list_mtx);
1266 		goto restart;
1267 	}
1268 	return (ocount - count);
1269 }
1270 
1271 void
1272 vnlru_free(int count, struct vfsops *mnt_op)
1273 {
1274 
1275 	mtx_lock(&vnode_list_mtx);
1276 	vnlru_free_locked(count, mnt_op);
1277 	mtx_unlock(&vnode_list_mtx);
1278 }
1279 
1280 static void
1281 vnlru_recalc(void)
1282 {
1283 
1284 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1285 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1286 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1287 	vlowat = vhiwat / 2;
1288 }
1289 
1290 /*
1291  * Attempt to recycle vnodes in a context that is always safe to block.
1292  * Calling vlrurecycle() from the bowels of filesystem code has some
1293  * interesting deadlock problems.
1294  */
1295 static struct proc *vnlruproc;
1296 static int vnlruproc_sig;
1297 
1298 /*
1299  * The main freevnodes counter is only updated when threads requeue their vnode
1300  * batches. CPUs are conditionally walked to compute a more accurate total.
1301  *
1302  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1303  * at any given moment can still exceed slop, but it should not be by significant
1304  * margin in practice.
1305  */
1306 #define VNLRU_FREEVNODES_SLOP 128
1307 
1308 static u_long
1309 vnlru_read_freevnodes(void)
1310 {
1311 	struct vdbatch *vd;
1312 	long slop;
1313 	int cpu;
1314 
1315 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1316 	if (freevnodes > freevnodes_old)
1317 		slop = freevnodes - freevnodes_old;
1318 	else
1319 		slop = freevnodes_old - freevnodes;
1320 	if (slop < VNLRU_FREEVNODES_SLOP)
1321 		return (freevnodes >= 0 ? freevnodes : 0);
1322 	freevnodes_old = freevnodes;
1323 	CPU_FOREACH(cpu) {
1324 		vd = DPCPU_ID_PTR((cpu), vd);
1325 		freevnodes_old += vd->freevnodes;
1326 	}
1327 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1328 }
1329 
1330 static bool
1331 vnlru_under(u_long rnumvnodes, u_long limit)
1332 {
1333 	u_long rfreevnodes, space;
1334 
1335 	if (__predict_false(rnumvnodes > desiredvnodes))
1336 		return (true);
1337 
1338 	space = desiredvnodes - rnumvnodes;
1339 	if (space < limit) {
1340 		rfreevnodes = vnlru_read_freevnodes();
1341 		if (rfreevnodes > wantfreevnodes)
1342 			space += rfreevnodes - wantfreevnodes;
1343 	}
1344 	return (space < limit);
1345 }
1346 
1347 static bool
1348 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1349 {
1350 	long rfreevnodes, space;
1351 
1352 	if (__predict_false(rnumvnodes > desiredvnodes))
1353 		return (true);
1354 
1355 	space = desiredvnodes - rnumvnodes;
1356 	if (space < limit) {
1357 		rfreevnodes = atomic_load_long(&freevnodes);
1358 		if (rfreevnodes > wantfreevnodes)
1359 			space += rfreevnodes - wantfreevnodes;
1360 	}
1361 	return (space < limit);
1362 }
1363 
1364 static void
1365 vnlru_kick(void)
1366 {
1367 
1368 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1369 	if (vnlruproc_sig == 0) {
1370 		vnlruproc_sig = 1;
1371 		wakeup(vnlruproc);
1372 	}
1373 }
1374 
1375 static void
1376 vnlru_proc(void)
1377 {
1378 	u_long rnumvnodes, rfreevnodes, target;
1379 	unsigned long onumvnodes;
1380 	int done, force, trigger, usevnodes;
1381 	bool reclaim_nc_src, want_reread;
1382 
1383 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1384 	    SHUTDOWN_PRI_FIRST);
1385 
1386 	force = 0;
1387 	want_reread = false;
1388 	for (;;) {
1389 		kproc_suspend_check(vnlruproc);
1390 		mtx_lock(&vnode_list_mtx);
1391 		rnumvnodes = atomic_load_long(&numvnodes);
1392 
1393 		if (want_reread) {
1394 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1395 			want_reread = false;
1396 		}
1397 
1398 		/*
1399 		 * If numvnodes is too large (due to desiredvnodes being
1400 		 * adjusted using its sysctl, or emergency growth), first
1401 		 * try to reduce it by discarding from the free list.
1402 		 */
1403 		if (rnumvnodes > desiredvnodes) {
1404 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1405 			rnumvnodes = atomic_load_long(&numvnodes);
1406 		}
1407 		/*
1408 		 * Sleep if the vnode cache is in a good state.  This is
1409 		 * when it is not over-full and has space for about a 4%
1410 		 * or 9% expansion (by growing its size or inexcessively
1411 		 * reducing its free list).  Otherwise, try to reclaim
1412 		 * space for a 10% expansion.
1413 		 */
1414 		if (vstir && force == 0) {
1415 			force = 1;
1416 			vstir = 0;
1417 		}
1418 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1419 			vnlruproc_sig = 0;
1420 			wakeup(&vnlruproc_sig);
1421 			msleep(vnlruproc, &vnode_list_mtx,
1422 			    PVFS|PDROP, "vlruwt", hz);
1423 			continue;
1424 		}
1425 		rfreevnodes = vnlru_read_freevnodes();
1426 
1427 		onumvnodes = rnumvnodes;
1428 		/*
1429 		 * Calculate parameters for recycling.  These are the same
1430 		 * throughout the loop to give some semblance of fairness.
1431 		 * The trigger point is to avoid recycling vnodes with lots
1432 		 * of resident pages.  We aren't trying to free memory; we
1433 		 * are trying to recycle or at least free vnodes.
1434 		 */
1435 		if (rnumvnodes <= desiredvnodes)
1436 			usevnodes = rnumvnodes - rfreevnodes;
1437 		else
1438 			usevnodes = rnumvnodes;
1439 		if (usevnodes <= 0)
1440 			usevnodes = 1;
1441 		/*
1442 		 * The trigger value is is chosen to give a conservatively
1443 		 * large value to ensure that it alone doesn't prevent
1444 		 * making progress.  The value can easily be so large that
1445 		 * it is effectively infinite in some congested and
1446 		 * misconfigured cases, and this is necessary.  Normally
1447 		 * it is about 8 to 100 (pages), which is quite large.
1448 		 */
1449 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1450 		if (force < 2)
1451 			trigger = vsmalltrigger;
1452 		reclaim_nc_src = force >= 3;
1453 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1454 		target = target / 10 + 1;
1455 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1456 		mtx_unlock(&vnode_list_mtx);
1457 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1458 			uma_reclaim(UMA_RECLAIM_DRAIN);
1459 		if (done == 0) {
1460 			if (force == 0 || force == 1) {
1461 				force = 2;
1462 				continue;
1463 			}
1464 			if (force == 2) {
1465 				force = 3;
1466 				continue;
1467 			}
1468 			want_reread = true;
1469 			force = 0;
1470 			vnlru_nowhere++;
1471 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1472 		} else {
1473 			want_reread = true;
1474 			kern_yield(PRI_USER);
1475 		}
1476 	}
1477 }
1478 
1479 static struct kproc_desc vnlru_kp = {
1480 	"vnlru",
1481 	vnlru_proc,
1482 	&vnlruproc
1483 };
1484 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1485     &vnlru_kp);
1486 
1487 /*
1488  * Routines having to do with the management of the vnode table.
1489  */
1490 
1491 /*
1492  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1493  * before we actually vgone().  This function must be called with the vnode
1494  * held to prevent the vnode from being returned to the free list midway
1495  * through vgone().
1496  */
1497 static int
1498 vtryrecycle(struct vnode *vp)
1499 {
1500 	struct mount *vnmp;
1501 
1502 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1503 	VNASSERT(vp->v_holdcnt, vp,
1504 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1505 	/*
1506 	 * This vnode may found and locked via some other list, if so we
1507 	 * can't recycle it yet.
1508 	 */
1509 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1510 		CTR2(KTR_VFS,
1511 		    "%s: impossible to recycle, vp %p lock is already held",
1512 		    __func__, vp);
1513 		return (EWOULDBLOCK);
1514 	}
1515 	/*
1516 	 * Don't recycle if its filesystem is being suspended.
1517 	 */
1518 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1519 		VOP_UNLOCK(vp);
1520 		CTR2(KTR_VFS,
1521 		    "%s: impossible to recycle, cannot start the write for %p",
1522 		    __func__, vp);
1523 		return (EBUSY);
1524 	}
1525 	/*
1526 	 * If we got this far, we need to acquire the interlock and see if
1527 	 * anyone picked up this vnode from another list.  If not, we will
1528 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1529 	 * will skip over it.
1530 	 */
1531 	VI_LOCK(vp);
1532 	if (vp->v_usecount) {
1533 		VOP_UNLOCK(vp);
1534 		VI_UNLOCK(vp);
1535 		vn_finished_write(vnmp);
1536 		CTR2(KTR_VFS,
1537 		    "%s: impossible to recycle, %p is already referenced",
1538 		    __func__, vp);
1539 		return (EBUSY);
1540 	}
1541 	if (!VN_IS_DOOMED(vp)) {
1542 		counter_u64_add(recycles_free_count, 1);
1543 		vgonel(vp);
1544 	}
1545 	VOP_UNLOCK(vp);
1546 	VI_UNLOCK(vp);
1547 	vn_finished_write(vnmp);
1548 	return (0);
1549 }
1550 
1551 /*
1552  * Allocate a new vnode.
1553  *
1554  * The operation never returns an error. Returning an error was disabled
1555  * in r145385 (dated 2005) with the following comment:
1556  *
1557  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1558  *
1559  * Given the age of this commit (almost 15 years at the time of writing this
1560  * comment) restoring the ability to fail requires a significant audit of
1561  * all codepaths.
1562  *
1563  * The routine can try to free a vnode or stall for up to 1 second waiting for
1564  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1565  */
1566 static u_long vn_alloc_cyclecount;
1567 
1568 static struct vnode * __noinline
1569 vn_alloc_hard(struct mount *mp)
1570 {
1571 	u_long rnumvnodes, rfreevnodes;
1572 
1573 	mtx_lock(&vnode_list_mtx);
1574 	rnumvnodes = atomic_load_long(&numvnodes);
1575 	if (rnumvnodes + 1 < desiredvnodes) {
1576 		vn_alloc_cyclecount = 0;
1577 		goto alloc;
1578 	}
1579 	rfreevnodes = vnlru_read_freevnodes();
1580 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1581 		vn_alloc_cyclecount = 0;
1582 		vstir = 1;
1583 	}
1584 	/*
1585 	 * Grow the vnode cache if it will not be above its target max
1586 	 * after growing.  Otherwise, if the free list is nonempty, try
1587 	 * to reclaim 1 item from it before growing the cache (possibly
1588 	 * above its target max if the reclamation failed or is delayed).
1589 	 * Otherwise, wait for some space.  In all cases, schedule
1590 	 * vnlru_proc() if we are getting short of space.  The watermarks
1591 	 * should be chosen so that we never wait or even reclaim from
1592 	 * the free list to below its target minimum.
1593 	 */
1594 	if (vnlru_free_locked(1, NULL) > 0)
1595 		goto alloc;
1596 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1597 		/*
1598 		 * Wait for space for a new vnode.
1599 		 */
1600 		vnlru_kick();
1601 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1602 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1603 		    vnlru_read_freevnodes() > 1)
1604 			vnlru_free_locked(1, NULL);
1605 	}
1606 alloc:
1607 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1608 	if (vnlru_under(rnumvnodes, vlowat))
1609 		vnlru_kick();
1610 	mtx_unlock(&vnode_list_mtx);
1611 	return (uma_zalloc(vnode_zone, M_WAITOK));
1612 }
1613 
1614 static struct vnode *
1615 vn_alloc(struct mount *mp)
1616 {
1617 	u_long rnumvnodes;
1618 
1619 	if (__predict_false(vn_alloc_cyclecount != 0))
1620 		return (vn_alloc_hard(mp));
1621 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1622 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1623 		atomic_subtract_long(&numvnodes, 1);
1624 		return (vn_alloc_hard(mp));
1625 	}
1626 
1627 	return (uma_zalloc(vnode_zone, M_WAITOK));
1628 }
1629 
1630 static void
1631 vn_free(struct vnode *vp)
1632 {
1633 
1634 	atomic_subtract_long(&numvnodes, 1);
1635 	uma_zfree(vnode_zone, vp);
1636 }
1637 
1638 /*
1639  * Return the next vnode from the free list.
1640  */
1641 int
1642 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1643     struct vnode **vpp)
1644 {
1645 	struct vnode *vp;
1646 	struct thread *td;
1647 	struct lock_object *lo;
1648 
1649 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1650 
1651 	KASSERT(vops->registered,
1652 	    ("%s: not registered vector op %p\n", __func__, vops));
1653 
1654 	td = curthread;
1655 	if (td->td_vp_reserved != NULL) {
1656 		vp = td->td_vp_reserved;
1657 		td->td_vp_reserved = NULL;
1658 	} else {
1659 		vp = vn_alloc(mp);
1660 	}
1661 	counter_u64_add(vnodes_created, 1);
1662 	/*
1663 	 * Locks are given the generic name "vnode" when created.
1664 	 * Follow the historic practice of using the filesystem
1665 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1666 	 *
1667 	 * Locks live in a witness group keyed on their name. Thus,
1668 	 * when a lock is renamed, it must also move from the witness
1669 	 * group of its old name to the witness group of its new name.
1670 	 *
1671 	 * The change only needs to be made when the vnode moves
1672 	 * from one filesystem type to another. We ensure that each
1673 	 * filesystem use a single static name pointer for its tag so
1674 	 * that we can compare pointers rather than doing a strcmp().
1675 	 */
1676 	lo = &vp->v_vnlock->lock_object;
1677 #ifdef WITNESS
1678 	if (lo->lo_name != tag) {
1679 #endif
1680 		lo->lo_name = tag;
1681 #ifdef WITNESS
1682 		WITNESS_DESTROY(lo);
1683 		WITNESS_INIT(lo, tag);
1684 	}
1685 #endif
1686 	/*
1687 	 * By default, don't allow shared locks unless filesystems opt-in.
1688 	 */
1689 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1690 	/*
1691 	 * Finalize various vnode identity bits.
1692 	 */
1693 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1694 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1695 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1696 	vp->v_type = VNON;
1697 	vp->v_op = vops;
1698 	v_init_counters(vp);
1699 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1700 #ifdef DIAGNOSTIC
1701 	if (mp == NULL && vops != &dead_vnodeops)
1702 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1703 #endif
1704 #ifdef MAC
1705 	mac_vnode_init(vp);
1706 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1707 		mac_vnode_associate_singlelabel(mp, vp);
1708 #endif
1709 	if (mp != NULL) {
1710 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1711 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1712 			vp->v_vflag |= VV_NOKNOTE;
1713 	}
1714 
1715 	/*
1716 	 * For the filesystems which do not use vfs_hash_insert(),
1717 	 * still initialize v_hash to have vfs_hash_index() useful.
1718 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1719 	 * its own hashing.
1720 	 */
1721 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1722 
1723 	*vpp = vp;
1724 	return (0);
1725 }
1726 
1727 void
1728 getnewvnode_reserve(void)
1729 {
1730 	struct thread *td;
1731 
1732 	td = curthread;
1733 	MPASS(td->td_vp_reserved == NULL);
1734 	td->td_vp_reserved = vn_alloc(NULL);
1735 }
1736 
1737 void
1738 getnewvnode_drop_reserve(void)
1739 {
1740 	struct thread *td;
1741 
1742 	td = curthread;
1743 	if (td->td_vp_reserved != NULL) {
1744 		vn_free(td->td_vp_reserved);
1745 		td->td_vp_reserved = NULL;
1746 	}
1747 }
1748 
1749 static void
1750 freevnode(struct vnode *vp)
1751 {
1752 	struct bufobj *bo;
1753 
1754 	/*
1755 	 * The vnode has been marked for destruction, so free it.
1756 	 *
1757 	 * The vnode will be returned to the zone where it will
1758 	 * normally remain until it is needed for another vnode. We
1759 	 * need to cleanup (or verify that the cleanup has already
1760 	 * been done) any residual data left from its current use
1761 	 * so as not to contaminate the freshly allocated vnode.
1762 	 */
1763 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1764 	bo = &vp->v_bufobj;
1765 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1766 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1767 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1768 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1769 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1770 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1771 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1772 	    ("clean blk trie not empty"));
1773 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1774 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1775 	    ("dirty blk trie not empty"));
1776 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1777 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1778 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1779 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1780 	    ("Dangling rangelock waiters"));
1781 	VI_UNLOCK(vp);
1782 #ifdef MAC
1783 	mac_vnode_destroy(vp);
1784 #endif
1785 	if (vp->v_pollinfo != NULL) {
1786 		destroy_vpollinfo(vp->v_pollinfo);
1787 		vp->v_pollinfo = NULL;
1788 	}
1789 #ifdef INVARIANTS
1790 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1791 	vp->v_op = NULL;
1792 #endif
1793 	vp->v_mountedhere = NULL;
1794 	vp->v_unpcb = NULL;
1795 	vp->v_rdev = NULL;
1796 	vp->v_fifoinfo = NULL;
1797 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1798 	vp->v_irflag = 0;
1799 	vp->v_iflag = 0;
1800 	vp->v_vflag = 0;
1801 	bo->bo_flag = 0;
1802 	vn_free(vp);
1803 }
1804 
1805 /*
1806  * Delete from old mount point vnode list, if on one.
1807  */
1808 static void
1809 delmntque(struct vnode *vp)
1810 {
1811 	struct mount *mp;
1812 
1813 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1814 
1815 	mp = vp->v_mount;
1816 	if (mp == NULL)
1817 		return;
1818 	MNT_ILOCK(mp);
1819 	VI_LOCK(vp);
1820 	vp->v_mount = NULL;
1821 	VI_UNLOCK(vp);
1822 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1823 		("bad mount point vnode list size"));
1824 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1825 	mp->mnt_nvnodelistsize--;
1826 	MNT_REL(mp);
1827 	MNT_IUNLOCK(mp);
1828 }
1829 
1830 static void
1831 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1832 {
1833 
1834 	vp->v_data = NULL;
1835 	vp->v_op = &dead_vnodeops;
1836 	vgone(vp);
1837 	vput(vp);
1838 }
1839 
1840 /*
1841  * Insert into list of vnodes for the new mount point, if available.
1842  */
1843 int
1844 insmntque1(struct vnode *vp, struct mount *mp,
1845 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1846 {
1847 
1848 	KASSERT(vp->v_mount == NULL,
1849 		("insmntque: vnode already on per mount vnode list"));
1850 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1851 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1852 
1853 	/*
1854 	 * We acquire the vnode interlock early to ensure that the
1855 	 * vnode cannot be recycled by another process releasing a
1856 	 * holdcnt on it before we get it on both the vnode list
1857 	 * and the active vnode list. The mount mutex protects only
1858 	 * manipulation of the vnode list and the vnode freelist
1859 	 * mutex protects only manipulation of the active vnode list.
1860 	 * Hence the need to hold the vnode interlock throughout.
1861 	 */
1862 	MNT_ILOCK(mp);
1863 	VI_LOCK(vp);
1864 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1865 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1866 	    mp->mnt_nvnodelistsize == 0)) &&
1867 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1868 		VI_UNLOCK(vp);
1869 		MNT_IUNLOCK(mp);
1870 		if (dtr != NULL)
1871 			dtr(vp, dtr_arg);
1872 		return (EBUSY);
1873 	}
1874 	vp->v_mount = mp;
1875 	MNT_REF(mp);
1876 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1877 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1878 		("neg mount point vnode list size"));
1879 	mp->mnt_nvnodelistsize++;
1880 	VI_UNLOCK(vp);
1881 	MNT_IUNLOCK(mp);
1882 	return (0);
1883 }
1884 
1885 int
1886 insmntque(struct vnode *vp, struct mount *mp)
1887 {
1888 
1889 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1890 }
1891 
1892 /*
1893  * Flush out and invalidate all buffers associated with a bufobj
1894  * Called with the underlying object locked.
1895  */
1896 int
1897 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1898 {
1899 	int error;
1900 
1901 	BO_LOCK(bo);
1902 	if (flags & V_SAVE) {
1903 		error = bufobj_wwait(bo, slpflag, slptimeo);
1904 		if (error) {
1905 			BO_UNLOCK(bo);
1906 			return (error);
1907 		}
1908 		if (bo->bo_dirty.bv_cnt > 0) {
1909 			BO_UNLOCK(bo);
1910 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1911 				return (error);
1912 			/*
1913 			 * XXX We could save a lock/unlock if this was only
1914 			 * enabled under INVARIANTS
1915 			 */
1916 			BO_LOCK(bo);
1917 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1918 				panic("vinvalbuf: dirty bufs");
1919 		}
1920 	}
1921 	/*
1922 	 * If you alter this loop please notice that interlock is dropped and
1923 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1924 	 * no race conditions occur from this.
1925 	 */
1926 	do {
1927 		error = flushbuflist(&bo->bo_clean,
1928 		    flags, bo, slpflag, slptimeo);
1929 		if (error == 0 && !(flags & V_CLEANONLY))
1930 			error = flushbuflist(&bo->bo_dirty,
1931 			    flags, bo, slpflag, slptimeo);
1932 		if (error != 0 && error != EAGAIN) {
1933 			BO_UNLOCK(bo);
1934 			return (error);
1935 		}
1936 	} while (error != 0);
1937 
1938 	/*
1939 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1940 	 * have write I/O in-progress but if there is a VM object then the
1941 	 * VM object can also have read-I/O in-progress.
1942 	 */
1943 	do {
1944 		bufobj_wwait(bo, 0, 0);
1945 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1946 			BO_UNLOCK(bo);
1947 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1948 			BO_LOCK(bo);
1949 		}
1950 	} while (bo->bo_numoutput > 0);
1951 	BO_UNLOCK(bo);
1952 
1953 	/*
1954 	 * Destroy the copy in the VM cache, too.
1955 	 */
1956 	if (bo->bo_object != NULL &&
1957 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1958 		VM_OBJECT_WLOCK(bo->bo_object);
1959 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1960 		    OBJPR_CLEANONLY : 0);
1961 		VM_OBJECT_WUNLOCK(bo->bo_object);
1962 	}
1963 
1964 #ifdef INVARIANTS
1965 	BO_LOCK(bo);
1966 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1967 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1968 	    bo->bo_clean.bv_cnt > 0))
1969 		panic("vinvalbuf: flush failed");
1970 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1971 	    bo->bo_dirty.bv_cnt > 0)
1972 		panic("vinvalbuf: flush dirty failed");
1973 	BO_UNLOCK(bo);
1974 #endif
1975 	return (0);
1976 }
1977 
1978 /*
1979  * Flush out and invalidate all buffers associated with a vnode.
1980  * Called with the underlying object locked.
1981  */
1982 int
1983 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1984 {
1985 
1986 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1987 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1988 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1989 		return (0);
1990 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1991 }
1992 
1993 /*
1994  * Flush out buffers on the specified list.
1995  *
1996  */
1997 static int
1998 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1999     int slptimeo)
2000 {
2001 	struct buf *bp, *nbp;
2002 	int retval, error;
2003 	daddr_t lblkno;
2004 	b_xflags_t xflags;
2005 
2006 	ASSERT_BO_WLOCKED(bo);
2007 
2008 	retval = 0;
2009 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2010 		/*
2011 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2012 		 * do not skip any buffers. If we are flushing only V_NORMAL
2013 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2014 		 * flushing only V_ALT buffers then skip buffers not marked
2015 		 * as BX_ALTDATA.
2016 		 */
2017 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2018 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2019 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2020 			continue;
2021 		}
2022 		if (nbp != NULL) {
2023 			lblkno = nbp->b_lblkno;
2024 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2025 		}
2026 		retval = EAGAIN;
2027 		error = BUF_TIMELOCK(bp,
2028 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2029 		    "flushbuf", slpflag, slptimeo);
2030 		if (error) {
2031 			BO_LOCK(bo);
2032 			return (error != ENOLCK ? error : EAGAIN);
2033 		}
2034 		KASSERT(bp->b_bufobj == bo,
2035 		    ("bp %p wrong b_bufobj %p should be %p",
2036 		    bp, bp->b_bufobj, bo));
2037 		/*
2038 		 * XXX Since there are no node locks for NFS, I
2039 		 * believe there is a slight chance that a delayed
2040 		 * write will occur while sleeping just above, so
2041 		 * check for it.
2042 		 */
2043 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2044 		    (flags & V_SAVE)) {
2045 			bremfree(bp);
2046 			bp->b_flags |= B_ASYNC;
2047 			bwrite(bp);
2048 			BO_LOCK(bo);
2049 			return (EAGAIN);	/* XXX: why not loop ? */
2050 		}
2051 		bremfree(bp);
2052 		bp->b_flags |= (B_INVAL | B_RELBUF);
2053 		bp->b_flags &= ~B_ASYNC;
2054 		brelse(bp);
2055 		BO_LOCK(bo);
2056 		if (nbp == NULL)
2057 			break;
2058 		nbp = gbincore(bo, lblkno);
2059 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2060 		    != xflags)
2061 			break;			/* nbp invalid */
2062 	}
2063 	return (retval);
2064 }
2065 
2066 int
2067 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2068 {
2069 	struct buf *bp;
2070 	int error;
2071 	daddr_t lblkno;
2072 
2073 	ASSERT_BO_LOCKED(bo);
2074 
2075 	for (lblkno = startn;;) {
2076 again:
2077 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2078 		if (bp == NULL || bp->b_lblkno >= endn ||
2079 		    bp->b_lblkno < startn)
2080 			break;
2081 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2082 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2083 		if (error != 0) {
2084 			BO_RLOCK(bo);
2085 			if (error == ENOLCK)
2086 				goto again;
2087 			return (error);
2088 		}
2089 		KASSERT(bp->b_bufobj == bo,
2090 		    ("bp %p wrong b_bufobj %p should be %p",
2091 		    bp, bp->b_bufobj, bo));
2092 		lblkno = bp->b_lblkno + 1;
2093 		if ((bp->b_flags & B_MANAGED) == 0)
2094 			bremfree(bp);
2095 		bp->b_flags |= B_RELBUF;
2096 		/*
2097 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2098 		 * pages backing each buffer in the range are unlikely to be
2099 		 * reused.  Dirty buffers will have the hint applied once
2100 		 * they've been written.
2101 		 */
2102 		if ((bp->b_flags & B_VMIO) != 0)
2103 			bp->b_flags |= B_NOREUSE;
2104 		brelse(bp);
2105 		BO_RLOCK(bo);
2106 	}
2107 	return (0);
2108 }
2109 
2110 /*
2111  * Truncate a file's buffer and pages to a specified length.  This
2112  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2113  * sync activity.
2114  */
2115 int
2116 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2117 {
2118 	struct buf *bp, *nbp;
2119 	struct bufobj *bo;
2120 	daddr_t startlbn;
2121 
2122 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2123 	    vp, blksize, (uintmax_t)length);
2124 
2125 	/*
2126 	 * Round up to the *next* lbn.
2127 	 */
2128 	startlbn = howmany(length, blksize);
2129 
2130 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2131 
2132 	bo = &vp->v_bufobj;
2133 restart_unlocked:
2134 	BO_LOCK(bo);
2135 
2136 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2137 		;
2138 
2139 	if (length > 0) {
2140 restartsync:
2141 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2142 			if (bp->b_lblkno > 0)
2143 				continue;
2144 			/*
2145 			 * Since we hold the vnode lock this should only
2146 			 * fail if we're racing with the buf daemon.
2147 			 */
2148 			if (BUF_LOCK(bp,
2149 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2150 			    BO_LOCKPTR(bo)) == ENOLCK)
2151 				goto restart_unlocked;
2152 
2153 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2154 			    ("buf(%p) on dirty queue without DELWRI", bp));
2155 
2156 			bremfree(bp);
2157 			bawrite(bp);
2158 			BO_LOCK(bo);
2159 			goto restartsync;
2160 		}
2161 	}
2162 
2163 	bufobj_wwait(bo, 0, 0);
2164 	BO_UNLOCK(bo);
2165 	vnode_pager_setsize(vp, length);
2166 
2167 	return (0);
2168 }
2169 
2170 /*
2171  * Invalidate the cached pages of a file's buffer within the range of block
2172  * numbers [startlbn, endlbn).
2173  */
2174 void
2175 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2176     int blksize)
2177 {
2178 	struct bufobj *bo;
2179 	off_t start, end;
2180 
2181 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2182 
2183 	start = blksize * startlbn;
2184 	end = blksize * endlbn;
2185 
2186 	bo = &vp->v_bufobj;
2187 	BO_LOCK(bo);
2188 	MPASS(blksize == bo->bo_bsize);
2189 
2190 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2191 		;
2192 
2193 	BO_UNLOCK(bo);
2194 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2195 }
2196 
2197 static int
2198 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2199     daddr_t startlbn, daddr_t endlbn)
2200 {
2201 	struct buf *bp, *nbp;
2202 	bool anyfreed;
2203 
2204 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2205 	ASSERT_BO_LOCKED(bo);
2206 
2207 	do {
2208 		anyfreed = false;
2209 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2210 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2211 				continue;
2212 			if (BUF_LOCK(bp,
2213 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2214 			    BO_LOCKPTR(bo)) == ENOLCK) {
2215 				BO_LOCK(bo);
2216 				return (EAGAIN);
2217 			}
2218 
2219 			bremfree(bp);
2220 			bp->b_flags |= B_INVAL | B_RELBUF;
2221 			bp->b_flags &= ~B_ASYNC;
2222 			brelse(bp);
2223 			anyfreed = true;
2224 
2225 			BO_LOCK(bo);
2226 			if (nbp != NULL &&
2227 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2228 			    nbp->b_vp != vp ||
2229 			    (nbp->b_flags & B_DELWRI) != 0))
2230 				return (EAGAIN);
2231 		}
2232 
2233 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2234 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2235 				continue;
2236 			if (BUF_LOCK(bp,
2237 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2238 			    BO_LOCKPTR(bo)) == ENOLCK) {
2239 				BO_LOCK(bo);
2240 				return (EAGAIN);
2241 			}
2242 			bremfree(bp);
2243 			bp->b_flags |= B_INVAL | B_RELBUF;
2244 			bp->b_flags &= ~B_ASYNC;
2245 			brelse(bp);
2246 			anyfreed = true;
2247 
2248 			BO_LOCK(bo);
2249 			if (nbp != NULL &&
2250 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2251 			    (nbp->b_vp != vp) ||
2252 			    (nbp->b_flags & B_DELWRI) == 0))
2253 				return (EAGAIN);
2254 		}
2255 	} while (anyfreed);
2256 	return (0);
2257 }
2258 
2259 static void
2260 buf_vlist_remove(struct buf *bp)
2261 {
2262 	struct bufv *bv;
2263 
2264 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2265 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2266 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2267 	    (BX_VNDIRTY|BX_VNCLEAN),
2268 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2269 	if (bp->b_xflags & BX_VNDIRTY)
2270 		bv = &bp->b_bufobj->bo_dirty;
2271 	else
2272 		bv = &bp->b_bufobj->bo_clean;
2273 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2274 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2275 	bv->bv_cnt--;
2276 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2277 }
2278 
2279 /*
2280  * Add the buffer to the sorted clean or dirty block list.
2281  *
2282  * NOTE: xflags is passed as a constant, optimizing this inline function!
2283  */
2284 static void
2285 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2286 {
2287 	struct bufv *bv;
2288 	struct buf *n;
2289 	int error;
2290 
2291 	ASSERT_BO_WLOCKED(bo);
2292 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2293 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2294 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2295 	    ("dead bo %p", bo));
2296 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2297 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2298 	bp->b_xflags |= xflags;
2299 	if (xflags & BX_VNDIRTY)
2300 		bv = &bo->bo_dirty;
2301 	else
2302 		bv = &bo->bo_clean;
2303 
2304 	/*
2305 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2306 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2307 	 * than _ge.
2308 	 */
2309 	if (bv->bv_cnt == 0 ||
2310 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2311 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2312 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2313 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2314 	else
2315 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2316 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2317 	if (error)
2318 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2319 	bv->bv_cnt++;
2320 }
2321 
2322 /*
2323  * Look up a buffer using the buffer tries.
2324  */
2325 struct buf *
2326 gbincore(struct bufobj *bo, daddr_t lblkno)
2327 {
2328 	struct buf *bp;
2329 
2330 	ASSERT_BO_LOCKED(bo);
2331 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2332 	if (bp != NULL)
2333 		return (bp);
2334 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2335 }
2336 
2337 /*
2338  * Associate a buffer with a vnode.
2339  */
2340 void
2341 bgetvp(struct vnode *vp, struct buf *bp)
2342 {
2343 	struct bufobj *bo;
2344 
2345 	bo = &vp->v_bufobj;
2346 	ASSERT_BO_WLOCKED(bo);
2347 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2348 
2349 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2350 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2351 	    ("bgetvp: bp already attached! %p", bp));
2352 
2353 	vhold(vp);
2354 	bp->b_vp = vp;
2355 	bp->b_bufobj = bo;
2356 	/*
2357 	 * Insert onto list for new vnode.
2358 	 */
2359 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2360 }
2361 
2362 /*
2363  * Disassociate a buffer from a vnode.
2364  */
2365 void
2366 brelvp(struct buf *bp)
2367 {
2368 	struct bufobj *bo;
2369 	struct vnode *vp;
2370 
2371 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2372 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2373 
2374 	/*
2375 	 * Delete from old vnode list, if on one.
2376 	 */
2377 	vp = bp->b_vp;		/* XXX */
2378 	bo = bp->b_bufobj;
2379 	BO_LOCK(bo);
2380 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2381 		buf_vlist_remove(bp);
2382 	else
2383 		panic("brelvp: Buffer %p not on queue.", bp);
2384 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2385 		bo->bo_flag &= ~BO_ONWORKLST;
2386 		mtx_lock(&sync_mtx);
2387 		LIST_REMOVE(bo, bo_synclist);
2388 		syncer_worklist_len--;
2389 		mtx_unlock(&sync_mtx);
2390 	}
2391 	bp->b_vp = NULL;
2392 	bp->b_bufobj = NULL;
2393 	BO_UNLOCK(bo);
2394 	vdrop(vp);
2395 }
2396 
2397 /*
2398  * Add an item to the syncer work queue.
2399  */
2400 static void
2401 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2402 {
2403 	int slot;
2404 
2405 	ASSERT_BO_WLOCKED(bo);
2406 
2407 	mtx_lock(&sync_mtx);
2408 	if (bo->bo_flag & BO_ONWORKLST)
2409 		LIST_REMOVE(bo, bo_synclist);
2410 	else {
2411 		bo->bo_flag |= BO_ONWORKLST;
2412 		syncer_worklist_len++;
2413 	}
2414 
2415 	if (delay > syncer_maxdelay - 2)
2416 		delay = syncer_maxdelay - 2;
2417 	slot = (syncer_delayno + delay) & syncer_mask;
2418 
2419 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2420 	mtx_unlock(&sync_mtx);
2421 }
2422 
2423 static int
2424 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2425 {
2426 	int error, len;
2427 
2428 	mtx_lock(&sync_mtx);
2429 	len = syncer_worklist_len - sync_vnode_count;
2430 	mtx_unlock(&sync_mtx);
2431 	error = SYSCTL_OUT(req, &len, sizeof(len));
2432 	return (error);
2433 }
2434 
2435 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2436     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2437     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2438 
2439 static struct proc *updateproc;
2440 static void sched_sync(void);
2441 static struct kproc_desc up_kp = {
2442 	"syncer",
2443 	sched_sync,
2444 	&updateproc
2445 };
2446 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2447 
2448 static int
2449 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2450 {
2451 	struct vnode *vp;
2452 	struct mount *mp;
2453 
2454 	*bo = LIST_FIRST(slp);
2455 	if (*bo == NULL)
2456 		return (0);
2457 	vp = bo2vnode(*bo);
2458 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2459 		return (1);
2460 	/*
2461 	 * We use vhold in case the vnode does not
2462 	 * successfully sync.  vhold prevents the vnode from
2463 	 * going away when we unlock the sync_mtx so that
2464 	 * we can acquire the vnode interlock.
2465 	 */
2466 	vholdl(vp);
2467 	mtx_unlock(&sync_mtx);
2468 	VI_UNLOCK(vp);
2469 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2470 		vdrop(vp);
2471 		mtx_lock(&sync_mtx);
2472 		return (*bo == LIST_FIRST(slp));
2473 	}
2474 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2475 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2476 	VOP_UNLOCK(vp);
2477 	vn_finished_write(mp);
2478 	BO_LOCK(*bo);
2479 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2480 		/*
2481 		 * Put us back on the worklist.  The worklist
2482 		 * routine will remove us from our current
2483 		 * position and then add us back in at a later
2484 		 * position.
2485 		 */
2486 		vn_syncer_add_to_worklist(*bo, syncdelay);
2487 	}
2488 	BO_UNLOCK(*bo);
2489 	vdrop(vp);
2490 	mtx_lock(&sync_mtx);
2491 	return (0);
2492 }
2493 
2494 static int first_printf = 1;
2495 
2496 /*
2497  * System filesystem synchronizer daemon.
2498  */
2499 static void
2500 sched_sync(void)
2501 {
2502 	struct synclist *next, *slp;
2503 	struct bufobj *bo;
2504 	long starttime;
2505 	struct thread *td = curthread;
2506 	int last_work_seen;
2507 	int net_worklist_len;
2508 	int syncer_final_iter;
2509 	int error;
2510 
2511 	last_work_seen = 0;
2512 	syncer_final_iter = 0;
2513 	syncer_state = SYNCER_RUNNING;
2514 	starttime = time_uptime;
2515 	td->td_pflags |= TDP_NORUNNINGBUF;
2516 
2517 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2518 	    SHUTDOWN_PRI_LAST);
2519 
2520 	mtx_lock(&sync_mtx);
2521 	for (;;) {
2522 		if (syncer_state == SYNCER_FINAL_DELAY &&
2523 		    syncer_final_iter == 0) {
2524 			mtx_unlock(&sync_mtx);
2525 			kproc_suspend_check(td->td_proc);
2526 			mtx_lock(&sync_mtx);
2527 		}
2528 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2529 		if (syncer_state != SYNCER_RUNNING &&
2530 		    starttime != time_uptime) {
2531 			if (first_printf) {
2532 				printf("\nSyncing disks, vnodes remaining... ");
2533 				first_printf = 0;
2534 			}
2535 			printf("%d ", net_worklist_len);
2536 		}
2537 		starttime = time_uptime;
2538 
2539 		/*
2540 		 * Push files whose dirty time has expired.  Be careful
2541 		 * of interrupt race on slp queue.
2542 		 *
2543 		 * Skip over empty worklist slots when shutting down.
2544 		 */
2545 		do {
2546 			slp = &syncer_workitem_pending[syncer_delayno];
2547 			syncer_delayno += 1;
2548 			if (syncer_delayno == syncer_maxdelay)
2549 				syncer_delayno = 0;
2550 			next = &syncer_workitem_pending[syncer_delayno];
2551 			/*
2552 			 * If the worklist has wrapped since the
2553 			 * it was emptied of all but syncer vnodes,
2554 			 * switch to the FINAL_DELAY state and run
2555 			 * for one more second.
2556 			 */
2557 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2558 			    net_worklist_len == 0 &&
2559 			    last_work_seen == syncer_delayno) {
2560 				syncer_state = SYNCER_FINAL_DELAY;
2561 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2562 			}
2563 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2564 		    syncer_worklist_len > 0);
2565 
2566 		/*
2567 		 * Keep track of the last time there was anything
2568 		 * on the worklist other than syncer vnodes.
2569 		 * Return to the SHUTTING_DOWN state if any
2570 		 * new work appears.
2571 		 */
2572 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2573 			last_work_seen = syncer_delayno;
2574 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2575 			syncer_state = SYNCER_SHUTTING_DOWN;
2576 		while (!LIST_EMPTY(slp)) {
2577 			error = sync_vnode(slp, &bo, td);
2578 			if (error == 1) {
2579 				LIST_REMOVE(bo, bo_synclist);
2580 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2581 				continue;
2582 			}
2583 
2584 			if (first_printf == 0) {
2585 				/*
2586 				 * Drop the sync mutex, because some watchdog
2587 				 * drivers need to sleep while patting
2588 				 */
2589 				mtx_unlock(&sync_mtx);
2590 				wdog_kern_pat(WD_LASTVAL);
2591 				mtx_lock(&sync_mtx);
2592 			}
2593 
2594 		}
2595 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2596 			syncer_final_iter--;
2597 		/*
2598 		 * The variable rushjob allows the kernel to speed up the
2599 		 * processing of the filesystem syncer process. A rushjob
2600 		 * value of N tells the filesystem syncer to process the next
2601 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2602 		 * is used by the soft update code to speed up the filesystem
2603 		 * syncer process when the incore state is getting so far
2604 		 * ahead of the disk that the kernel memory pool is being
2605 		 * threatened with exhaustion.
2606 		 */
2607 		if (rushjob > 0) {
2608 			rushjob -= 1;
2609 			continue;
2610 		}
2611 		/*
2612 		 * Just sleep for a short period of time between
2613 		 * iterations when shutting down to allow some I/O
2614 		 * to happen.
2615 		 *
2616 		 * If it has taken us less than a second to process the
2617 		 * current work, then wait. Otherwise start right over
2618 		 * again. We can still lose time if any single round
2619 		 * takes more than two seconds, but it does not really
2620 		 * matter as we are just trying to generally pace the
2621 		 * filesystem activity.
2622 		 */
2623 		if (syncer_state != SYNCER_RUNNING ||
2624 		    time_uptime == starttime) {
2625 			thread_lock(td);
2626 			sched_prio(td, PPAUSE);
2627 			thread_unlock(td);
2628 		}
2629 		if (syncer_state != SYNCER_RUNNING)
2630 			cv_timedwait(&sync_wakeup, &sync_mtx,
2631 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2632 		else if (time_uptime == starttime)
2633 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2634 	}
2635 }
2636 
2637 /*
2638  * Request the syncer daemon to speed up its work.
2639  * We never push it to speed up more than half of its
2640  * normal turn time, otherwise it could take over the cpu.
2641  */
2642 int
2643 speedup_syncer(void)
2644 {
2645 	int ret = 0;
2646 
2647 	mtx_lock(&sync_mtx);
2648 	if (rushjob < syncdelay / 2) {
2649 		rushjob += 1;
2650 		stat_rush_requests += 1;
2651 		ret = 1;
2652 	}
2653 	mtx_unlock(&sync_mtx);
2654 	cv_broadcast(&sync_wakeup);
2655 	return (ret);
2656 }
2657 
2658 /*
2659  * Tell the syncer to speed up its work and run though its work
2660  * list several times, then tell it to shut down.
2661  */
2662 static void
2663 syncer_shutdown(void *arg, int howto)
2664 {
2665 
2666 	if (howto & RB_NOSYNC)
2667 		return;
2668 	mtx_lock(&sync_mtx);
2669 	syncer_state = SYNCER_SHUTTING_DOWN;
2670 	rushjob = 0;
2671 	mtx_unlock(&sync_mtx);
2672 	cv_broadcast(&sync_wakeup);
2673 	kproc_shutdown(arg, howto);
2674 }
2675 
2676 void
2677 syncer_suspend(void)
2678 {
2679 
2680 	syncer_shutdown(updateproc, 0);
2681 }
2682 
2683 void
2684 syncer_resume(void)
2685 {
2686 
2687 	mtx_lock(&sync_mtx);
2688 	first_printf = 1;
2689 	syncer_state = SYNCER_RUNNING;
2690 	mtx_unlock(&sync_mtx);
2691 	cv_broadcast(&sync_wakeup);
2692 	kproc_resume(updateproc);
2693 }
2694 
2695 /*
2696  * Reassign a buffer from one vnode to another.
2697  * Used to assign file specific control information
2698  * (indirect blocks) to the vnode to which they belong.
2699  */
2700 void
2701 reassignbuf(struct buf *bp)
2702 {
2703 	struct vnode *vp;
2704 	struct bufobj *bo;
2705 	int delay;
2706 #ifdef INVARIANTS
2707 	struct bufv *bv;
2708 #endif
2709 
2710 	vp = bp->b_vp;
2711 	bo = bp->b_bufobj;
2712 	++reassignbufcalls;
2713 
2714 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2715 	    bp, bp->b_vp, bp->b_flags);
2716 	/*
2717 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2718 	 * is not fully linked in.
2719 	 */
2720 	if (bp->b_flags & B_PAGING)
2721 		panic("cannot reassign paging buffer");
2722 
2723 	/*
2724 	 * Delete from old vnode list, if on one.
2725 	 */
2726 	BO_LOCK(bo);
2727 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2728 		buf_vlist_remove(bp);
2729 	else
2730 		panic("reassignbuf: Buffer %p not on queue.", bp);
2731 	/*
2732 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2733 	 * of clean buffers.
2734 	 */
2735 	if (bp->b_flags & B_DELWRI) {
2736 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2737 			switch (vp->v_type) {
2738 			case VDIR:
2739 				delay = dirdelay;
2740 				break;
2741 			case VCHR:
2742 				delay = metadelay;
2743 				break;
2744 			default:
2745 				delay = filedelay;
2746 			}
2747 			vn_syncer_add_to_worklist(bo, delay);
2748 		}
2749 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2750 	} else {
2751 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2752 
2753 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2754 			mtx_lock(&sync_mtx);
2755 			LIST_REMOVE(bo, bo_synclist);
2756 			syncer_worklist_len--;
2757 			mtx_unlock(&sync_mtx);
2758 			bo->bo_flag &= ~BO_ONWORKLST;
2759 		}
2760 	}
2761 #ifdef INVARIANTS
2762 	bv = &bo->bo_clean;
2763 	bp = TAILQ_FIRST(&bv->bv_hd);
2764 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2765 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2766 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2767 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2768 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2769 	bv = &bo->bo_dirty;
2770 	bp = TAILQ_FIRST(&bv->bv_hd);
2771 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2772 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2773 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2774 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2775 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2776 #endif
2777 	BO_UNLOCK(bo);
2778 }
2779 
2780 static void
2781 v_init_counters(struct vnode *vp)
2782 {
2783 
2784 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2785 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2786 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2787 
2788 	refcount_init(&vp->v_holdcnt, 1);
2789 	refcount_init(&vp->v_usecount, 1);
2790 }
2791 
2792 /*
2793  * Increment si_usecount of the associated device, if any.
2794  */
2795 static void
2796 v_incr_devcount(struct vnode *vp)
2797 {
2798 
2799 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2800 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2801 		dev_lock();
2802 		vp->v_rdev->si_usecount++;
2803 		dev_unlock();
2804 	}
2805 }
2806 
2807 /*
2808  * Decrement si_usecount of the associated device, if any.
2809  *
2810  * The caller is required to hold the interlock when transitioning a VCHR use
2811  * count to zero. This prevents a race with devfs_reclaim_vchr() that would
2812  * leak a si_usecount reference. The vnode lock will also prevent this race
2813  * if it is held while dropping the last ref.
2814  *
2815  * The race is:
2816  *
2817  * CPU1					CPU2
2818  *				  	devfs_reclaim_vchr
2819  * make v_usecount == 0
2820  * 				    	  VI_LOCK
2821  * 				    	  sees v_usecount == 0, no updates
2822  * 				    	  vp->v_rdev = NULL;
2823  * 				    	  ...
2824  * 				    	  VI_UNLOCK
2825  * VI_LOCK
2826  * v_decr_devcount
2827  *   sees v_rdev == NULL, no updates
2828  *
2829  * In this scenario si_devcount decrement is not performed.
2830  */
2831 static void
2832 v_decr_devcount(struct vnode *vp)
2833 {
2834 
2835 	ASSERT_VOP_LOCKED(vp, __func__);
2836 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2837 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2838 		dev_lock();
2839 		VNPASS(vp->v_rdev->si_usecount > 0, vp);
2840 		vp->v_rdev->si_usecount--;
2841 		dev_unlock();
2842 	}
2843 }
2844 
2845 /*
2846  * Grab a particular vnode from the free list, increment its
2847  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2848  * is being destroyed.  Only callers who specify LK_RETRY will
2849  * see doomed vnodes.  If inactive processing was delayed in
2850  * vput try to do it here.
2851  *
2852  * usecount is manipulated using atomics without holding any locks.
2853  *
2854  * holdcnt can be manipulated using atomics without holding any locks,
2855  * except when transitioning 1<->0, in which case the interlock is held.
2856  */
2857 enum vgetstate
2858 vget_prep(struct vnode *vp)
2859 {
2860 	enum vgetstate vs;
2861 
2862 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2863 		vs = VGET_USECOUNT;
2864 	} else {
2865 		vhold(vp);
2866 		vs = VGET_HOLDCNT;
2867 	}
2868 	return (vs);
2869 }
2870 
2871 int
2872 vget(struct vnode *vp, int flags, struct thread *td)
2873 {
2874 	enum vgetstate vs;
2875 
2876 	MPASS(td == curthread);
2877 
2878 	vs = vget_prep(vp);
2879 	return (vget_finish(vp, flags, vs));
2880 }
2881 
2882 static int __noinline
2883 vget_finish_vchr(struct vnode *vp)
2884 {
2885 
2886 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2887 
2888 	/*
2889 	 * See the comment in vget_finish before usecount bump.
2890 	 */
2891 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2892 #ifdef INVARIANTS
2893 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2894 		VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2895 #else
2896 		refcount_release(&vp->v_holdcnt);
2897 #endif
2898 		return (0);
2899 	}
2900 
2901 	VI_LOCK(vp);
2902 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2903 #ifdef INVARIANTS
2904 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2905 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2906 #else
2907 		refcount_release(&vp->v_holdcnt);
2908 #endif
2909 		VI_UNLOCK(vp);
2910 		return (0);
2911 	}
2912 	v_incr_devcount(vp);
2913 	refcount_acquire(&vp->v_usecount);
2914 	VI_UNLOCK(vp);
2915 	return (0);
2916 }
2917 
2918 int
2919 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2920 {
2921 	int error, old;
2922 
2923 	if ((flags & LK_INTERLOCK) != 0)
2924 		ASSERT_VI_LOCKED(vp, __func__);
2925 	else
2926 		ASSERT_VI_UNLOCKED(vp, __func__);
2927 	VNPASS(vp->v_holdcnt > 0, vp);
2928 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2929 
2930 	error = vn_lock(vp, flags);
2931 	if (__predict_false(error != 0)) {
2932 		if (vs == VGET_USECOUNT)
2933 			vrele(vp);
2934 		else
2935 			vdrop(vp);
2936 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2937 		    vp);
2938 		return (error);
2939 	}
2940 
2941 	if (vs == VGET_USECOUNT)
2942 		return (0);
2943 
2944 	if (__predict_false(vp->v_type == VCHR))
2945 		return (vget_finish_vchr(vp));
2946 
2947 	/*
2948 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2949 	 * the vnode around. Otherwise someone else lended their hold count and
2950 	 * we have to drop ours.
2951 	 */
2952 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2953 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2954 	if (old != 0) {
2955 #ifdef INVARIANTS
2956 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2957 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2958 #else
2959 		refcount_release(&vp->v_holdcnt);
2960 #endif
2961 	}
2962 	return (0);
2963 }
2964 
2965 /*
2966  * Increase the reference (use) and hold count of a vnode.
2967  * This will also remove the vnode from the free list if it is presently free.
2968  */
2969 static void __noinline
2970 vref_vchr(struct vnode *vp, bool interlock)
2971 {
2972 
2973 	/*
2974 	 * See the comment in vget_finish before usecount bump.
2975 	 */
2976 	if (!interlock) {
2977 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2978 			VNODE_REFCOUNT_FENCE_ACQ();
2979 			VNASSERT(vp->v_holdcnt > 0, vp,
2980 			    ("%s: active vnode not held", __func__));
2981 			return;
2982 		}
2983 		VI_LOCK(vp);
2984 		/*
2985 		 * By the time we get here the vnode might have been doomed, at
2986 		 * which point the 0->1 use count transition is no longer
2987 		 * protected by the interlock. Since it can't bounce back to
2988 		 * VCHR and requires vref semantics, punt it back
2989 		 */
2990 		if (__predict_false(vp->v_type == VBAD)) {
2991 			VI_UNLOCK(vp);
2992 			vref(vp);
2993 			return;
2994 		}
2995 	}
2996 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2997 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2998 		VNODE_REFCOUNT_FENCE_ACQ();
2999 		VNASSERT(vp->v_holdcnt > 0, vp,
3000 		    ("%s: active vnode not held", __func__));
3001 		if (!interlock)
3002 			VI_UNLOCK(vp);
3003 		return;
3004 	}
3005 	vhold(vp);
3006 	v_incr_devcount(vp);
3007 	refcount_acquire(&vp->v_usecount);
3008 	if (!interlock)
3009 		VI_UNLOCK(vp);
3010 	return;
3011 }
3012 
3013 void
3014 vref(struct vnode *vp)
3015 {
3016 	int old;
3017 
3018 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3019 	if (__predict_false(vp->v_type == VCHR)) {
3020 		 vref_vchr(vp, false);
3021 		 return;
3022 	}
3023 
3024 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3025 		VNODE_REFCOUNT_FENCE_ACQ();
3026 		VNASSERT(vp->v_holdcnt > 0, vp,
3027 		    ("%s: active vnode not held", __func__));
3028 		return;
3029 	}
3030 	vhold(vp);
3031 	/*
3032 	 * See the comment in vget_finish.
3033 	 */
3034 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3035 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3036 	if (old != 0) {
3037 #ifdef INVARIANTS
3038 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3039 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3040 #else
3041 		refcount_release(&vp->v_holdcnt);
3042 #endif
3043 	}
3044 }
3045 
3046 void
3047 vrefl(struct vnode *vp)
3048 {
3049 
3050 	ASSERT_VI_LOCKED(vp, __func__);
3051 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3052 	if (__predict_false(vp->v_type == VCHR)) {
3053 		vref_vchr(vp, true);
3054 		return;
3055 	}
3056 	vref(vp);
3057 }
3058 
3059 void
3060 vrefact(struct vnode *vp)
3061 {
3062 
3063 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3064 #ifdef INVARIANTS
3065 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3066 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3067 #else
3068 	refcount_acquire(&vp->v_usecount);
3069 #endif
3070 }
3071 
3072 void
3073 vrefactn(struct vnode *vp, u_int n)
3074 {
3075 
3076 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3077 #ifdef INVARIANTS
3078 	int old = atomic_fetchadd_int(&vp->v_usecount, n);
3079 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3080 #else
3081 	atomic_add_int(&vp->v_usecount, n);
3082 #endif
3083 }
3084 
3085 /*
3086  * Return reference count of a vnode.
3087  *
3088  * The results of this call are only guaranteed when some mechanism is used to
3089  * stop other processes from gaining references to the vnode.  This may be the
3090  * case if the caller holds the only reference.  This is also useful when stale
3091  * data is acceptable as race conditions may be accounted for by some other
3092  * means.
3093  */
3094 int
3095 vrefcnt(struct vnode *vp)
3096 {
3097 
3098 	return (vp->v_usecount);
3099 }
3100 
3101 void
3102 vlazy(struct vnode *vp)
3103 {
3104 	struct mount *mp;
3105 
3106 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3107 
3108 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3109 		return;
3110 	/*
3111 	 * We may get here for inactive routines after the vnode got doomed.
3112 	 */
3113 	if (VN_IS_DOOMED(vp))
3114 		return;
3115 	mp = vp->v_mount;
3116 	mtx_lock(&mp->mnt_listmtx);
3117 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3118 		vp->v_mflag |= VMP_LAZYLIST;
3119 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3120 		mp->mnt_lazyvnodelistsize++;
3121 	}
3122 	mtx_unlock(&mp->mnt_listmtx);
3123 }
3124 
3125 /*
3126  * This routine is only meant to be called from vgonel prior to dooming
3127  * the vnode.
3128  */
3129 static void
3130 vunlazy_gone(struct vnode *vp)
3131 {
3132 	struct mount *mp;
3133 
3134 	ASSERT_VOP_ELOCKED(vp, __func__);
3135 	ASSERT_VI_LOCKED(vp, __func__);
3136 	VNPASS(!VN_IS_DOOMED(vp), vp);
3137 
3138 	if (vp->v_mflag & VMP_LAZYLIST) {
3139 		mp = vp->v_mount;
3140 		mtx_lock(&mp->mnt_listmtx);
3141 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3142 		vp->v_mflag &= ~VMP_LAZYLIST;
3143 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3144 		mp->mnt_lazyvnodelistsize--;
3145 		mtx_unlock(&mp->mnt_listmtx);
3146 	}
3147 }
3148 
3149 static void
3150 vdefer_inactive(struct vnode *vp)
3151 {
3152 
3153 	ASSERT_VI_LOCKED(vp, __func__);
3154 	VNASSERT(vp->v_holdcnt > 0, vp,
3155 	    ("%s: vnode without hold count", __func__));
3156 	if (VN_IS_DOOMED(vp)) {
3157 		vdropl(vp);
3158 		return;
3159 	}
3160 	if (vp->v_iflag & VI_DEFINACT) {
3161 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3162 		vdropl(vp);
3163 		return;
3164 	}
3165 	if (vp->v_usecount > 0) {
3166 		vp->v_iflag &= ~VI_OWEINACT;
3167 		vdropl(vp);
3168 		return;
3169 	}
3170 	vlazy(vp);
3171 	vp->v_iflag |= VI_DEFINACT;
3172 	VI_UNLOCK(vp);
3173 	counter_u64_add(deferred_inact, 1);
3174 }
3175 
3176 static void
3177 vdefer_inactive_unlocked(struct vnode *vp)
3178 {
3179 
3180 	VI_LOCK(vp);
3181 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3182 		vdropl(vp);
3183 		return;
3184 	}
3185 	vdefer_inactive(vp);
3186 }
3187 
3188 enum vput_op { VRELE, VPUT, VUNREF };
3189 
3190 /*
3191  * Handle ->v_usecount transitioning to 0.
3192  *
3193  * By releasing the last usecount we take ownership of the hold count which
3194  * provides liveness of the vnode, meaning we have to vdrop.
3195  *
3196  * If the vnode is of type VCHR we may need to decrement si_usecount, see
3197  * v_decr_devcount for details.
3198  *
3199  * For all vnodes we may need to perform inactive processing. It requires an
3200  * exclusive lock on the vnode, while it is legal to call here with only a
3201  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3202  * inactive processing gets deferred to the syncer.
3203  *
3204  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3205  * on the lock being held all the way until VOP_INACTIVE. This in particular
3206  * happens with UFS which adds half-constructed vnodes to the hash, where they
3207  * can be found by other code.
3208  */
3209 static void
3210 vput_final(struct vnode *vp, enum vput_op func)
3211 {
3212 	int error;
3213 	bool want_unlock;
3214 
3215 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3216 	VNPASS(vp->v_holdcnt > 0, vp);
3217 
3218 	VI_LOCK(vp);
3219 	if (__predict_false(vp->v_type == VCHR && func != VRELE))
3220 		v_decr_devcount(vp);
3221 
3222 	/*
3223 	 * By the time we got here someone else might have transitioned
3224 	 * the count back to > 0.
3225 	 */
3226 	if (vp->v_usecount > 0)
3227 		goto out;
3228 
3229 	/*
3230 	 * If the vnode is doomed vgone already performed inactive processing
3231 	 * (if needed).
3232 	 */
3233 	if (VN_IS_DOOMED(vp))
3234 		goto out;
3235 
3236 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3237 		goto out;
3238 
3239 	if (vp->v_iflag & VI_DOINGINACT)
3240 		goto out;
3241 
3242 	/*
3243 	 * Locking operations here will drop the interlock and possibly the
3244 	 * vnode lock, opening a window where the vnode can get doomed all the
3245 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3246 	 * perform inactive.
3247 	 */
3248 	vp->v_iflag |= VI_OWEINACT;
3249 	want_unlock = false;
3250 	error = 0;
3251 	switch (func) {
3252 	case VRELE:
3253 		switch (VOP_ISLOCKED(vp)) {
3254 		case LK_EXCLUSIVE:
3255 			break;
3256 		case LK_EXCLOTHER:
3257 		case 0:
3258 			want_unlock = true;
3259 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3260 			VI_LOCK(vp);
3261 			break;
3262 		default:
3263 			/*
3264 			 * The lock has at least one sharer, but we have no way
3265 			 * to conclude whether this is us. Play it safe and
3266 			 * defer processing.
3267 			 */
3268 			error = EAGAIN;
3269 			break;
3270 		}
3271 		break;
3272 	case VPUT:
3273 		want_unlock = true;
3274 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3275 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3276 			    LK_NOWAIT);
3277 			VI_LOCK(vp);
3278 		}
3279 		break;
3280 	case VUNREF:
3281 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3282 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3283 			VI_LOCK(vp);
3284 		}
3285 		break;
3286 	}
3287 	if (error == 0) {
3288 		vinactive(vp);
3289 		if (want_unlock)
3290 			VOP_UNLOCK(vp);
3291 		vdropl(vp);
3292 	} else {
3293 		vdefer_inactive(vp);
3294 	}
3295 	return;
3296 out:
3297 	if (func == VPUT)
3298 		VOP_UNLOCK(vp);
3299 	vdropl(vp);
3300 }
3301 
3302 /*
3303  * Decrement ->v_usecount for a vnode.
3304  *
3305  * Releasing the last use count requires additional processing, see vput_final
3306  * above for details.
3307  *
3308  * Note that releasing use count without the vnode lock requires special casing
3309  * for VCHR, see v_decr_devcount for details.
3310  *
3311  * Comment above each variant denotes lock state on entry and exit.
3312  */
3313 
3314 static void __noinline
3315 vrele_vchr(struct vnode *vp)
3316 {
3317 
3318 	if (refcount_release_if_not_last(&vp->v_usecount))
3319 		return;
3320 	VI_LOCK(vp);
3321 	if (!refcount_release(&vp->v_usecount)) {
3322 		VI_UNLOCK(vp);
3323 		return;
3324 	}
3325 	v_decr_devcount(vp);
3326 	VI_UNLOCK(vp);
3327 	vput_final(vp, VRELE);
3328 }
3329 
3330 /*
3331  * in: any
3332  * out: same as passed in
3333  */
3334 void
3335 vrele(struct vnode *vp)
3336 {
3337 
3338 	ASSERT_VI_UNLOCKED(vp, __func__);
3339 	if (__predict_false(vp->v_type == VCHR)) {
3340 		vrele_vchr(vp);
3341 		return;
3342 	}
3343 	if (!refcount_release(&vp->v_usecount))
3344 		return;
3345 	vput_final(vp, VRELE);
3346 }
3347 
3348 /*
3349  * in: locked
3350  * out: unlocked
3351  */
3352 void
3353 vput(struct vnode *vp)
3354 {
3355 
3356 	ASSERT_VOP_LOCKED(vp, __func__);
3357 	ASSERT_VI_UNLOCKED(vp, __func__);
3358 	if (!refcount_release(&vp->v_usecount)) {
3359 		VOP_UNLOCK(vp);
3360 		return;
3361 	}
3362 	vput_final(vp, VPUT);
3363 }
3364 
3365 /*
3366  * in: locked
3367  * out: locked
3368  */
3369 void
3370 vunref(struct vnode *vp)
3371 {
3372 
3373 	ASSERT_VOP_LOCKED(vp, __func__);
3374 	ASSERT_VI_UNLOCKED(vp, __func__);
3375 	if (!refcount_release(&vp->v_usecount))
3376 		return;
3377 	vput_final(vp, VUNREF);
3378 }
3379 
3380 void
3381 vhold(struct vnode *vp)
3382 {
3383 	struct vdbatch *vd;
3384 	int old;
3385 
3386 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3387 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3388 	VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3389 	if (old != 0)
3390 		return;
3391 	critical_enter();
3392 	vd = DPCPU_PTR(vd);
3393 	vd->freevnodes--;
3394 	critical_exit();
3395 }
3396 
3397 void
3398 vholdl(struct vnode *vp)
3399 {
3400 
3401 	ASSERT_VI_LOCKED(vp, __func__);
3402 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3403 	vhold(vp);
3404 }
3405 
3406 void
3407 vholdnz(struct vnode *vp)
3408 {
3409 
3410 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3411 #ifdef INVARIANTS
3412 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3413 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3414 #else
3415 	atomic_add_int(&vp->v_holdcnt, 1);
3416 #endif
3417 }
3418 
3419 static void __noinline
3420 vdbatch_process(struct vdbatch *vd)
3421 {
3422 	struct vnode *vp;
3423 	int i;
3424 
3425 	mtx_assert(&vd->lock, MA_OWNED);
3426 	MPASS(curthread->td_pinned > 0);
3427 	MPASS(vd->index == VDBATCH_SIZE);
3428 
3429 	mtx_lock(&vnode_list_mtx);
3430 	critical_enter();
3431 	freevnodes += vd->freevnodes;
3432 	for (i = 0; i < VDBATCH_SIZE; i++) {
3433 		vp = vd->tab[i];
3434 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3435 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3436 		MPASS(vp->v_dbatchcpu != NOCPU);
3437 		vp->v_dbatchcpu = NOCPU;
3438 	}
3439 	mtx_unlock(&vnode_list_mtx);
3440 	vd->freevnodes = 0;
3441 	bzero(vd->tab, sizeof(vd->tab));
3442 	vd->index = 0;
3443 	critical_exit();
3444 }
3445 
3446 static void
3447 vdbatch_enqueue(struct vnode *vp)
3448 {
3449 	struct vdbatch *vd;
3450 
3451 	ASSERT_VI_LOCKED(vp, __func__);
3452 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3453 	    ("%s: deferring requeue of a doomed vnode", __func__));
3454 
3455 	critical_enter();
3456 	vd = DPCPU_PTR(vd);
3457 	vd->freevnodes++;
3458 	if (vp->v_dbatchcpu != NOCPU) {
3459 		VI_UNLOCK(vp);
3460 		critical_exit();
3461 		return;
3462 	}
3463 
3464 	sched_pin();
3465 	critical_exit();
3466 	mtx_lock(&vd->lock);
3467 	MPASS(vd->index < VDBATCH_SIZE);
3468 	MPASS(vd->tab[vd->index] == NULL);
3469 	/*
3470 	 * A hack: we depend on being pinned so that we know what to put in
3471 	 * ->v_dbatchcpu.
3472 	 */
3473 	vp->v_dbatchcpu = curcpu;
3474 	vd->tab[vd->index] = vp;
3475 	vd->index++;
3476 	VI_UNLOCK(vp);
3477 	if (vd->index == VDBATCH_SIZE)
3478 		vdbatch_process(vd);
3479 	mtx_unlock(&vd->lock);
3480 	sched_unpin();
3481 }
3482 
3483 /*
3484  * This routine must only be called for vnodes which are about to be
3485  * deallocated. Supporting dequeue for arbitrary vndoes would require
3486  * validating that the locked batch matches.
3487  */
3488 static void
3489 vdbatch_dequeue(struct vnode *vp)
3490 {
3491 	struct vdbatch *vd;
3492 	int i;
3493 	short cpu;
3494 
3495 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3496 	    ("%s: called for a used vnode\n", __func__));
3497 
3498 	cpu = vp->v_dbatchcpu;
3499 	if (cpu == NOCPU)
3500 		return;
3501 
3502 	vd = DPCPU_ID_PTR(cpu, vd);
3503 	mtx_lock(&vd->lock);
3504 	for (i = 0; i < vd->index; i++) {
3505 		if (vd->tab[i] != vp)
3506 			continue;
3507 		vp->v_dbatchcpu = NOCPU;
3508 		vd->index--;
3509 		vd->tab[i] = vd->tab[vd->index];
3510 		vd->tab[vd->index] = NULL;
3511 		break;
3512 	}
3513 	mtx_unlock(&vd->lock);
3514 	/*
3515 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3516 	 */
3517 	MPASS(vp->v_dbatchcpu == NOCPU);
3518 }
3519 
3520 /*
3521  * Drop the hold count of the vnode.  If this is the last reference to
3522  * the vnode we place it on the free list unless it has been vgone'd
3523  * (marked VIRF_DOOMED) in which case we will free it.
3524  *
3525  * Because the vnode vm object keeps a hold reference on the vnode if
3526  * there is at least one resident non-cached page, the vnode cannot
3527  * leave the active list without the page cleanup done.
3528  */
3529 static void
3530 vdrop_deactivate(struct vnode *vp)
3531 {
3532 	struct mount *mp;
3533 
3534 	ASSERT_VI_LOCKED(vp, __func__);
3535 	/*
3536 	 * Mark a vnode as free: remove it from its active list
3537 	 * and put it up for recycling on the freelist.
3538 	 */
3539 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3540 	    ("vdrop: returning doomed vnode"));
3541 	VNASSERT(vp->v_op != NULL, vp,
3542 	    ("vdrop: vnode already reclaimed."));
3543 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3544 	    ("vnode with VI_OWEINACT set"));
3545 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3546 	    ("vnode with VI_DEFINACT set"));
3547 	if (vp->v_mflag & VMP_LAZYLIST) {
3548 		mp = vp->v_mount;
3549 		mtx_lock(&mp->mnt_listmtx);
3550 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3551 		/*
3552 		 * Don't remove the vnode from the lazy list if another thread
3553 		 * has increased the hold count. It may have re-enqueued the
3554 		 * vnode to the lazy list and is now responsible for its
3555 		 * removal.
3556 		 */
3557 		if (vp->v_holdcnt == 0) {
3558 			vp->v_mflag &= ~VMP_LAZYLIST;
3559 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3560 			mp->mnt_lazyvnodelistsize--;
3561 		}
3562 		mtx_unlock(&mp->mnt_listmtx);
3563 	}
3564 	vdbatch_enqueue(vp);
3565 }
3566 
3567 void
3568 vdrop(struct vnode *vp)
3569 {
3570 
3571 	ASSERT_VI_UNLOCKED(vp, __func__);
3572 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3573 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3574 		return;
3575 	VI_LOCK(vp);
3576 	vdropl(vp);
3577 }
3578 
3579 void
3580 vdropl(struct vnode *vp)
3581 {
3582 
3583 	ASSERT_VI_LOCKED(vp, __func__);
3584 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3585 	if (!refcount_release(&vp->v_holdcnt)) {
3586 		VI_UNLOCK(vp);
3587 		return;
3588 	}
3589 	if (VN_IS_DOOMED(vp)) {
3590 		freevnode(vp);
3591 		return;
3592 	}
3593 	vdrop_deactivate(vp);
3594 }
3595 
3596 /*
3597  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3598  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3599  */
3600 static void
3601 vinactivef(struct vnode *vp)
3602 {
3603 	struct vm_object *obj;
3604 
3605 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3606 	ASSERT_VI_LOCKED(vp, "vinactive");
3607 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3608 	    ("vinactive: recursed on VI_DOINGINACT"));
3609 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3610 	vp->v_iflag |= VI_DOINGINACT;
3611 	vp->v_iflag &= ~VI_OWEINACT;
3612 	VI_UNLOCK(vp);
3613 	/*
3614 	 * Before moving off the active list, we must be sure that any
3615 	 * modified pages are converted into the vnode's dirty
3616 	 * buffers, since these will no longer be checked once the
3617 	 * vnode is on the inactive list.
3618 	 *
3619 	 * The write-out of the dirty pages is asynchronous.  At the
3620 	 * point that VOP_INACTIVE() is called, there could still be
3621 	 * pending I/O and dirty pages in the object.
3622 	 */
3623 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3624 	    vm_object_mightbedirty(obj)) {
3625 		VM_OBJECT_WLOCK(obj);
3626 		vm_object_page_clean(obj, 0, 0, 0);
3627 		VM_OBJECT_WUNLOCK(obj);
3628 	}
3629 	VOP_INACTIVE(vp, curthread);
3630 	VI_LOCK(vp);
3631 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3632 	    ("vinactive: lost VI_DOINGINACT"));
3633 	vp->v_iflag &= ~VI_DOINGINACT;
3634 }
3635 
3636 void
3637 vinactive(struct vnode *vp)
3638 {
3639 
3640 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3641 	ASSERT_VI_LOCKED(vp, "vinactive");
3642 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3643 
3644 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3645 		return;
3646 	if (vp->v_iflag & VI_DOINGINACT)
3647 		return;
3648 	if (vp->v_usecount > 0) {
3649 		vp->v_iflag &= ~VI_OWEINACT;
3650 		return;
3651 	}
3652 	vinactivef(vp);
3653 }
3654 
3655 /*
3656  * Remove any vnodes in the vnode table belonging to mount point mp.
3657  *
3658  * If FORCECLOSE is not specified, there should not be any active ones,
3659  * return error if any are found (nb: this is a user error, not a
3660  * system error). If FORCECLOSE is specified, detach any active vnodes
3661  * that are found.
3662  *
3663  * If WRITECLOSE is set, only flush out regular file vnodes open for
3664  * writing.
3665  *
3666  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3667  *
3668  * `rootrefs' specifies the base reference count for the root vnode
3669  * of this filesystem. The root vnode is considered busy if its
3670  * v_usecount exceeds this value. On a successful return, vflush(, td)
3671  * will call vrele() on the root vnode exactly rootrefs times.
3672  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3673  * be zero.
3674  */
3675 #ifdef DIAGNOSTIC
3676 static int busyprt = 0;		/* print out busy vnodes */
3677 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3678 #endif
3679 
3680 int
3681 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3682 {
3683 	struct vnode *vp, *mvp, *rootvp = NULL;
3684 	struct vattr vattr;
3685 	int busy = 0, error;
3686 
3687 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3688 	    rootrefs, flags);
3689 	if (rootrefs > 0) {
3690 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3691 		    ("vflush: bad args"));
3692 		/*
3693 		 * Get the filesystem root vnode. We can vput() it
3694 		 * immediately, since with rootrefs > 0, it won't go away.
3695 		 */
3696 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3697 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3698 			    __func__, error);
3699 			return (error);
3700 		}
3701 		vput(rootvp);
3702 	}
3703 loop:
3704 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3705 		vholdl(vp);
3706 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3707 		if (error) {
3708 			vdrop(vp);
3709 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3710 			goto loop;
3711 		}
3712 		/*
3713 		 * Skip over a vnodes marked VV_SYSTEM.
3714 		 */
3715 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3716 			VOP_UNLOCK(vp);
3717 			vdrop(vp);
3718 			continue;
3719 		}
3720 		/*
3721 		 * If WRITECLOSE is set, flush out unlinked but still open
3722 		 * files (even if open only for reading) and regular file
3723 		 * vnodes open for writing.
3724 		 */
3725 		if (flags & WRITECLOSE) {
3726 			if (vp->v_object != NULL) {
3727 				VM_OBJECT_WLOCK(vp->v_object);
3728 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3729 				VM_OBJECT_WUNLOCK(vp->v_object);
3730 			}
3731 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3732 			if (error != 0) {
3733 				VOP_UNLOCK(vp);
3734 				vdrop(vp);
3735 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3736 				return (error);
3737 			}
3738 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3739 			VI_LOCK(vp);
3740 
3741 			if ((vp->v_type == VNON ||
3742 			    (error == 0 && vattr.va_nlink > 0)) &&
3743 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3744 				VOP_UNLOCK(vp);
3745 				vdropl(vp);
3746 				continue;
3747 			}
3748 		} else
3749 			VI_LOCK(vp);
3750 		/*
3751 		 * With v_usecount == 0, all we need to do is clear out the
3752 		 * vnode data structures and we are done.
3753 		 *
3754 		 * If FORCECLOSE is set, forcibly close the vnode.
3755 		 */
3756 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3757 			vgonel(vp);
3758 		} else {
3759 			busy++;
3760 #ifdef DIAGNOSTIC
3761 			if (busyprt)
3762 				vn_printf(vp, "vflush: busy vnode ");
3763 #endif
3764 		}
3765 		VOP_UNLOCK(vp);
3766 		vdropl(vp);
3767 	}
3768 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3769 		/*
3770 		 * If just the root vnode is busy, and if its refcount
3771 		 * is equal to `rootrefs', then go ahead and kill it.
3772 		 */
3773 		VI_LOCK(rootvp);
3774 		KASSERT(busy > 0, ("vflush: not busy"));
3775 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3776 		    ("vflush: usecount %d < rootrefs %d",
3777 		     rootvp->v_usecount, rootrefs));
3778 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3779 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3780 			vgone(rootvp);
3781 			VOP_UNLOCK(rootvp);
3782 			busy = 0;
3783 		} else
3784 			VI_UNLOCK(rootvp);
3785 	}
3786 	if (busy) {
3787 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3788 		    busy);
3789 		return (EBUSY);
3790 	}
3791 	for (; rootrefs > 0; rootrefs--)
3792 		vrele(rootvp);
3793 	return (0);
3794 }
3795 
3796 /*
3797  * Recycle an unused vnode to the front of the free list.
3798  */
3799 int
3800 vrecycle(struct vnode *vp)
3801 {
3802 	int recycled;
3803 
3804 	VI_LOCK(vp);
3805 	recycled = vrecyclel(vp);
3806 	VI_UNLOCK(vp);
3807 	return (recycled);
3808 }
3809 
3810 /*
3811  * vrecycle, with the vp interlock held.
3812  */
3813 int
3814 vrecyclel(struct vnode *vp)
3815 {
3816 	int recycled;
3817 
3818 	ASSERT_VOP_ELOCKED(vp, __func__);
3819 	ASSERT_VI_LOCKED(vp, __func__);
3820 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3821 	recycled = 0;
3822 	if (vp->v_usecount == 0) {
3823 		recycled = 1;
3824 		vgonel(vp);
3825 	}
3826 	return (recycled);
3827 }
3828 
3829 /*
3830  * Eliminate all activity associated with a vnode
3831  * in preparation for reuse.
3832  */
3833 void
3834 vgone(struct vnode *vp)
3835 {
3836 	VI_LOCK(vp);
3837 	vgonel(vp);
3838 	VI_UNLOCK(vp);
3839 }
3840 
3841 static void
3842 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3843     struct vnode *lowervp __unused)
3844 {
3845 }
3846 
3847 /*
3848  * Notify upper mounts about reclaimed or unlinked vnode.
3849  */
3850 void
3851 vfs_notify_upper(struct vnode *vp, int event)
3852 {
3853 	static struct vfsops vgonel_vfsops = {
3854 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3855 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3856 	};
3857 	struct mount *mp, *ump, *mmp;
3858 
3859 	mp = vp->v_mount;
3860 	if (mp == NULL)
3861 		return;
3862 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3863 		return;
3864 
3865 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3866 	mmp->mnt_op = &vgonel_vfsops;
3867 	mmp->mnt_kern_flag |= MNTK_MARKER;
3868 	MNT_ILOCK(mp);
3869 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3870 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3871 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3872 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3873 			continue;
3874 		}
3875 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3876 		MNT_IUNLOCK(mp);
3877 		switch (event) {
3878 		case VFS_NOTIFY_UPPER_RECLAIM:
3879 			VFS_RECLAIM_LOWERVP(ump, vp);
3880 			break;
3881 		case VFS_NOTIFY_UPPER_UNLINK:
3882 			VFS_UNLINK_LOWERVP(ump, vp);
3883 			break;
3884 		default:
3885 			KASSERT(0, ("invalid event %d", event));
3886 			break;
3887 		}
3888 		MNT_ILOCK(mp);
3889 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3890 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3891 	}
3892 	free(mmp, M_TEMP);
3893 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3894 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3895 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3896 		wakeup(&mp->mnt_uppers);
3897 	}
3898 	MNT_IUNLOCK(mp);
3899 }
3900 
3901 /*
3902  * vgone, with the vp interlock held.
3903  */
3904 static void
3905 vgonel(struct vnode *vp)
3906 {
3907 	struct thread *td;
3908 	struct mount *mp;
3909 	vm_object_t object;
3910 	bool active, oweinact;
3911 
3912 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3913 	ASSERT_VI_LOCKED(vp, "vgonel");
3914 	VNASSERT(vp->v_holdcnt, vp,
3915 	    ("vgonel: vp %p has no reference.", vp));
3916 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3917 	td = curthread;
3918 
3919 	/*
3920 	 * Don't vgonel if we're already doomed.
3921 	 */
3922 	if (vp->v_irflag & VIRF_DOOMED)
3923 		return;
3924 	vunlazy_gone(vp);
3925 	vp->v_irflag |= VIRF_DOOMED;
3926 
3927 	/*
3928 	 * Check to see if the vnode is in use.  If so, we have to call
3929 	 * VOP_CLOSE() and VOP_INACTIVE().
3930 	 */
3931 	active = vp->v_usecount > 0;
3932 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3933 	/*
3934 	 * If we need to do inactive VI_OWEINACT will be set.
3935 	 */
3936 	if (vp->v_iflag & VI_DEFINACT) {
3937 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3938 		vp->v_iflag &= ~VI_DEFINACT;
3939 		vdropl(vp);
3940 	} else {
3941 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3942 		VI_UNLOCK(vp);
3943 	}
3944 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3945 
3946 	/*
3947 	 * If purging an active vnode, it must be closed and
3948 	 * deactivated before being reclaimed.
3949 	 */
3950 	if (active)
3951 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3952 	if (oweinact || active) {
3953 		VI_LOCK(vp);
3954 		vinactivef(vp);
3955 		VI_UNLOCK(vp);
3956 	}
3957 	if (vp->v_type == VSOCK)
3958 		vfs_unp_reclaim(vp);
3959 
3960 	/*
3961 	 * Clean out any buffers associated with the vnode.
3962 	 * If the flush fails, just toss the buffers.
3963 	 */
3964 	mp = NULL;
3965 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3966 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3967 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3968 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3969 			;
3970 	}
3971 
3972 	BO_LOCK(&vp->v_bufobj);
3973 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3974 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3975 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3976 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3977 	    ("vp %p bufobj not invalidated", vp));
3978 
3979 	/*
3980 	 * For VMIO bufobj, BO_DEAD is set later, or in
3981 	 * vm_object_terminate() after the object's page queue is
3982 	 * flushed.
3983 	 */
3984 	object = vp->v_bufobj.bo_object;
3985 	if (object == NULL)
3986 		vp->v_bufobj.bo_flag |= BO_DEAD;
3987 	BO_UNLOCK(&vp->v_bufobj);
3988 
3989 	/*
3990 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3991 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3992 	 * should not touch the object borrowed from the lower vnode
3993 	 * (the handle check).
3994 	 */
3995 	if (object != NULL && object->type == OBJT_VNODE &&
3996 	    object->handle == vp)
3997 		vnode_destroy_vobject(vp);
3998 
3999 	/*
4000 	 * Reclaim the vnode.
4001 	 */
4002 	if (VOP_RECLAIM(vp, td))
4003 		panic("vgone: cannot reclaim");
4004 	if (mp != NULL)
4005 		vn_finished_secondary_write(mp);
4006 	VNASSERT(vp->v_object == NULL, vp,
4007 	    ("vop_reclaim left v_object vp=%p", vp));
4008 	/*
4009 	 * Clear the advisory locks and wake up waiting threads.
4010 	 */
4011 	(void)VOP_ADVLOCKPURGE(vp);
4012 	vp->v_lockf = NULL;
4013 	/*
4014 	 * Delete from old mount point vnode list.
4015 	 */
4016 	delmntque(vp);
4017 	cache_purge(vp);
4018 	/*
4019 	 * Done with purge, reset to the standard lock and invalidate
4020 	 * the vnode.
4021 	 */
4022 	VI_LOCK(vp);
4023 	vp->v_vnlock = &vp->v_lock;
4024 	vp->v_op = &dead_vnodeops;
4025 	vp->v_type = VBAD;
4026 }
4027 
4028 /*
4029  * Calculate the total number of references to a special device.
4030  */
4031 int
4032 vcount(struct vnode *vp)
4033 {
4034 	int count;
4035 
4036 	dev_lock();
4037 	count = vp->v_rdev->si_usecount;
4038 	dev_unlock();
4039 	return (count);
4040 }
4041 
4042 /*
4043  * Print out a description of a vnode.
4044  */
4045 static const char * const typename[] =
4046 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4047  "VMARKER"};
4048 
4049 void
4050 vn_printf(struct vnode *vp, const char *fmt, ...)
4051 {
4052 	va_list ap;
4053 	char buf[256], buf2[16];
4054 	u_long flags;
4055 
4056 	va_start(ap, fmt);
4057 	vprintf(fmt, ap);
4058 	va_end(ap);
4059 	printf("%p: ", (void *)vp);
4060 	printf("type %s\n", typename[vp->v_type]);
4061 	printf("    usecount %d, writecount %d, refcount %d",
4062 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
4063 	switch (vp->v_type) {
4064 	case VDIR:
4065 		printf(" mountedhere %p\n", vp->v_mountedhere);
4066 		break;
4067 	case VCHR:
4068 		printf(" rdev %p\n", vp->v_rdev);
4069 		break;
4070 	case VSOCK:
4071 		printf(" socket %p\n", vp->v_unpcb);
4072 		break;
4073 	case VFIFO:
4074 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4075 		break;
4076 	default:
4077 		printf("\n");
4078 		break;
4079 	}
4080 	buf[0] = '\0';
4081 	buf[1] = '\0';
4082 	if (vp->v_irflag & VIRF_DOOMED)
4083 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4084 	flags = vp->v_irflag & ~(VIRF_DOOMED);
4085 	if (flags != 0) {
4086 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4087 		strlcat(buf, buf2, sizeof(buf));
4088 	}
4089 	if (vp->v_vflag & VV_ROOT)
4090 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4091 	if (vp->v_vflag & VV_ISTTY)
4092 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4093 	if (vp->v_vflag & VV_NOSYNC)
4094 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4095 	if (vp->v_vflag & VV_ETERNALDEV)
4096 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4097 	if (vp->v_vflag & VV_CACHEDLABEL)
4098 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4099 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4100 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4101 	if (vp->v_vflag & VV_COPYONWRITE)
4102 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4103 	if (vp->v_vflag & VV_SYSTEM)
4104 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4105 	if (vp->v_vflag & VV_PROCDEP)
4106 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4107 	if (vp->v_vflag & VV_NOKNOTE)
4108 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4109 	if (vp->v_vflag & VV_DELETED)
4110 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4111 	if (vp->v_vflag & VV_MD)
4112 		strlcat(buf, "|VV_MD", sizeof(buf));
4113 	if (vp->v_vflag & VV_FORCEINSMQ)
4114 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4115 	if (vp->v_vflag & VV_READLINK)
4116 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4117 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4118 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4119 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4120 	if (flags != 0) {
4121 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4122 		strlcat(buf, buf2, sizeof(buf));
4123 	}
4124 	if (vp->v_iflag & VI_TEXT_REF)
4125 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4126 	if (vp->v_iflag & VI_MOUNT)
4127 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4128 	if (vp->v_iflag & VI_DOINGINACT)
4129 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4130 	if (vp->v_iflag & VI_OWEINACT)
4131 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4132 	if (vp->v_iflag & VI_DEFINACT)
4133 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4134 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4135 	    VI_OWEINACT | VI_DEFINACT);
4136 	if (flags != 0) {
4137 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4138 		strlcat(buf, buf2, sizeof(buf));
4139 	}
4140 	if (vp->v_mflag & VMP_LAZYLIST)
4141 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4142 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4143 	if (flags != 0) {
4144 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4145 		strlcat(buf, buf2, sizeof(buf));
4146 	}
4147 	printf("    flags (%s)\n", buf + 1);
4148 	if (mtx_owned(VI_MTX(vp)))
4149 		printf(" VI_LOCKed");
4150 	if (vp->v_object != NULL)
4151 		printf("    v_object %p ref %d pages %d "
4152 		    "cleanbuf %d dirtybuf %d\n",
4153 		    vp->v_object, vp->v_object->ref_count,
4154 		    vp->v_object->resident_page_count,
4155 		    vp->v_bufobj.bo_clean.bv_cnt,
4156 		    vp->v_bufobj.bo_dirty.bv_cnt);
4157 	printf("    ");
4158 	lockmgr_printinfo(vp->v_vnlock);
4159 	if (vp->v_data != NULL)
4160 		VOP_PRINT(vp);
4161 }
4162 
4163 #ifdef DDB
4164 /*
4165  * List all of the locked vnodes in the system.
4166  * Called when debugging the kernel.
4167  */
4168 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4169 {
4170 	struct mount *mp;
4171 	struct vnode *vp;
4172 
4173 	/*
4174 	 * Note: because this is DDB, we can't obey the locking semantics
4175 	 * for these structures, which means we could catch an inconsistent
4176 	 * state and dereference a nasty pointer.  Not much to be done
4177 	 * about that.
4178 	 */
4179 	db_printf("Locked vnodes\n");
4180 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4181 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4182 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4183 				vn_printf(vp, "vnode ");
4184 		}
4185 	}
4186 }
4187 
4188 /*
4189  * Show details about the given vnode.
4190  */
4191 DB_SHOW_COMMAND(vnode, db_show_vnode)
4192 {
4193 	struct vnode *vp;
4194 
4195 	if (!have_addr)
4196 		return;
4197 	vp = (struct vnode *)addr;
4198 	vn_printf(vp, "vnode ");
4199 }
4200 
4201 /*
4202  * Show details about the given mount point.
4203  */
4204 DB_SHOW_COMMAND(mount, db_show_mount)
4205 {
4206 	struct mount *mp;
4207 	struct vfsopt *opt;
4208 	struct statfs *sp;
4209 	struct vnode *vp;
4210 	char buf[512];
4211 	uint64_t mflags;
4212 	u_int flags;
4213 
4214 	if (!have_addr) {
4215 		/* No address given, print short info about all mount points. */
4216 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4217 			db_printf("%p %s on %s (%s)\n", mp,
4218 			    mp->mnt_stat.f_mntfromname,
4219 			    mp->mnt_stat.f_mntonname,
4220 			    mp->mnt_stat.f_fstypename);
4221 			if (db_pager_quit)
4222 				break;
4223 		}
4224 		db_printf("\nMore info: show mount <addr>\n");
4225 		return;
4226 	}
4227 
4228 	mp = (struct mount *)addr;
4229 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4230 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4231 
4232 	buf[0] = '\0';
4233 	mflags = mp->mnt_flag;
4234 #define	MNT_FLAG(flag)	do {						\
4235 	if (mflags & (flag)) {						\
4236 		if (buf[0] != '\0')					\
4237 			strlcat(buf, ", ", sizeof(buf));		\
4238 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4239 		mflags &= ~(flag);					\
4240 	}								\
4241 } while (0)
4242 	MNT_FLAG(MNT_RDONLY);
4243 	MNT_FLAG(MNT_SYNCHRONOUS);
4244 	MNT_FLAG(MNT_NOEXEC);
4245 	MNT_FLAG(MNT_NOSUID);
4246 	MNT_FLAG(MNT_NFS4ACLS);
4247 	MNT_FLAG(MNT_UNION);
4248 	MNT_FLAG(MNT_ASYNC);
4249 	MNT_FLAG(MNT_SUIDDIR);
4250 	MNT_FLAG(MNT_SOFTDEP);
4251 	MNT_FLAG(MNT_NOSYMFOLLOW);
4252 	MNT_FLAG(MNT_GJOURNAL);
4253 	MNT_FLAG(MNT_MULTILABEL);
4254 	MNT_FLAG(MNT_ACLS);
4255 	MNT_FLAG(MNT_NOATIME);
4256 	MNT_FLAG(MNT_NOCLUSTERR);
4257 	MNT_FLAG(MNT_NOCLUSTERW);
4258 	MNT_FLAG(MNT_SUJ);
4259 	MNT_FLAG(MNT_EXRDONLY);
4260 	MNT_FLAG(MNT_EXPORTED);
4261 	MNT_FLAG(MNT_DEFEXPORTED);
4262 	MNT_FLAG(MNT_EXPORTANON);
4263 	MNT_FLAG(MNT_EXKERB);
4264 	MNT_FLAG(MNT_EXPUBLIC);
4265 	MNT_FLAG(MNT_LOCAL);
4266 	MNT_FLAG(MNT_QUOTA);
4267 	MNT_FLAG(MNT_ROOTFS);
4268 	MNT_FLAG(MNT_USER);
4269 	MNT_FLAG(MNT_IGNORE);
4270 	MNT_FLAG(MNT_UPDATE);
4271 	MNT_FLAG(MNT_DELEXPORT);
4272 	MNT_FLAG(MNT_RELOAD);
4273 	MNT_FLAG(MNT_FORCE);
4274 	MNT_FLAG(MNT_SNAPSHOT);
4275 	MNT_FLAG(MNT_BYFSID);
4276 #undef MNT_FLAG
4277 	if (mflags != 0) {
4278 		if (buf[0] != '\0')
4279 			strlcat(buf, ", ", sizeof(buf));
4280 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4281 		    "0x%016jx", mflags);
4282 	}
4283 	db_printf("    mnt_flag = %s\n", buf);
4284 
4285 	buf[0] = '\0';
4286 	flags = mp->mnt_kern_flag;
4287 #define	MNT_KERN_FLAG(flag)	do {					\
4288 	if (flags & (flag)) {						\
4289 		if (buf[0] != '\0')					\
4290 			strlcat(buf, ", ", sizeof(buf));		\
4291 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4292 		flags &= ~(flag);					\
4293 	}								\
4294 } while (0)
4295 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4296 	MNT_KERN_FLAG(MNTK_ASYNC);
4297 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4298 	MNT_KERN_FLAG(MNTK_DRAINING);
4299 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4300 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4301 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4302 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4303 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4304 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4305 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4306 	MNT_KERN_FLAG(MNTK_MARKER);
4307 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4308 	MNT_KERN_FLAG(MNTK_NOASYNC);
4309 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4310 	MNT_KERN_FLAG(MNTK_MWAIT);
4311 	MNT_KERN_FLAG(MNTK_SUSPEND);
4312 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4313 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4314 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4315 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4316 #undef MNT_KERN_FLAG
4317 	if (flags != 0) {
4318 		if (buf[0] != '\0')
4319 			strlcat(buf, ", ", sizeof(buf));
4320 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4321 		    "0x%08x", flags);
4322 	}
4323 	db_printf("    mnt_kern_flag = %s\n", buf);
4324 
4325 	db_printf("    mnt_opt = ");
4326 	opt = TAILQ_FIRST(mp->mnt_opt);
4327 	if (opt != NULL) {
4328 		db_printf("%s", opt->name);
4329 		opt = TAILQ_NEXT(opt, link);
4330 		while (opt != NULL) {
4331 			db_printf(", %s", opt->name);
4332 			opt = TAILQ_NEXT(opt, link);
4333 		}
4334 	}
4335 	db_printf("\n");
4336 
4337 	sp = &mp->mnt_stat;
4338 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4339 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4340 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4341 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4342 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4343 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4344 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4345 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4346 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4347 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4348 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4349 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4350 
4351 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4352 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4353 	if (jailed(mp->mnt_cred))
4354 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4355 	db_printf(" }\n");
4356 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4357 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4358 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4359 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4360 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4361 	    mp->mnt_lazyvnodelistsize);
4362 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4363 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4364 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4365 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4366 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4367 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4368 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4369 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4370 	db_printf("    mnt_secondary_accwrites = %d\n",
4371 	    mp->mnt_secondary_accwrites);
4372 	db_printf("    mnt_gjprovider = %s\n",
4373 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4374 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4375 
4376 	db_printf("\n\nList of active vnodes\n");
4377 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4378 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4379 			vn_printf(vp, "vnode ");
4380 			if (db_pager_quit)
4381 				break;
4382 		}
4383 	}
4384 	db_printf("\n\nList of inactive vnodes\n");
4385 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4386 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4387 			vn_printf(vp, "vnode ");
4388 			if (db_pager_quit)
4389 				break;
4390 		}
4391 	}
4392 }
4393 #endif	/* DDB */
4394 
4395 /*
4396  * Fill in a struct xvfsconf based on a struct vfsconf.
4397  */
4398 static int
4399 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4400 {
4401 	struct xvfsconf xvfsp;
4402 
4403 	bzero(&xvfsp, sizeof(xvfsp));
4404 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4405 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4406 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4407 	xvfsp.vfc_flags = vfsp->vfc_flags;
4408 	/*
4409 	 * These are unused in userland, we keep them
4410 	 * to not break binary compatibility.
4411 	 */
4412 	xvfsp.vfc_vfsops = NULL;
4413 	xvfsp.vfc_next = NULL;
4414 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4415 }
4416 
4417 #ifdef COMPAT_FREEBSD32
4418 struct xvfsconf32 {
4419 	uint32_t	vfc_vfsops;
4420 	char		vfc_name[MFSNAMELEN];
4421 	int32_t		vfc_typenum;
4422 	int32_t		vfc_refcount;
4423 	int32_t		vfc_flags;
4424 	uint32_t	vfc_next;
4425 };
4426 
4427 static int
4428 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4429 {
4430 	struct xvfsconf32 xvfsp;
4431 
4432 	bzero(&xvfsp, sizeof(xvfsp));
4433 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4434 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4435 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4436 	xvfsp.vfc_flags = vfsp->vfc_flags;
4437 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4438 }
4439 #endif
4440 
4441 /*
4442  * Top level filesystem related information gathering.
4443  */
4444 static int
4445 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4446 {
4447 	struct vfsconf *vfsp;
4448 	int error;
4449 
4450 	error = 0;
4451 	vfsconf_slock();
4452 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4453 #ifdef COMPAT_FREEBSD32
4454 		if (req->flags & SCTL_MASK32)
4455 			error = vfsconf2x32(req, vfsp);
4456 		else
4457 #endif
4458 			error = vfsconf2x(req, vfsp);
4459 		if (error)
4460 			break;
4461 	}
4462 	vfsconf_sunlock();
4463 	return (error);
4464 }
4465 
4466 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4467     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4468     "S,xvfsconf", "List of all configured filesystems");
4469 
4470 #ifndef BURN_BRIDGES
4471 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4472 
4473 static int
4474 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4475 {
4476 	int *name = (int *)arg1 - 1;	/* XXX */
4477 	u_int namelen = arg2 + 1;	/* XXX */
4478 	struct vfsconf *vfsp;
4479 
4480 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4481 	    "please rebuild world\n");
4482 
4483 #if 1 || defined(COMPAT_PRELITE2)
4484 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4485 	if (namelen == 1)
4486 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4487 #endif
4488 
4489 	switch (name[1]) {
4490 	case VFS_MAXTYPENUM:
4491 		if (namelen != 2)
4492 			return (ENOTDIR);
4493 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4494 	case VFS_CONF:
4495 		if (namelen != 3)
4496 			return (ENOTDIR);	/* overloaded */
4497 		vfsconf_slock();
4498 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4499 			if (vfsp->vfc_typenum == name[2])
4500 				break;
4501 		}
4502 		vfsconf_sunlock();
4503 		if (vfsp == NULL)
4504 			return (EOPNOTSUPP);
4505 #ifdef COMPAT_FREEBSD32
4506 		if (req->flags & SCTL_MASK32)
4507 			return (vfsconf2x32(req, vfsp));
4508 		else
4509 #endif
4510 			return (vfsconf2x(req, vfsp));
4511 	}
4512 	return (EOPNOTSUPP);
4513 }
4514 
4515 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4516     CTLFLAG_MPSAFE, vfs_sysctl,
4517     "Generic filesystem");
4518 
4519 #if 1 || defined(COMPAT_PRELITE2)
4520 
4521 static int
4522 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4523 {
4524 	int error;
4525 	struct vfsconf *vfsp;
4526 	struct ovfsconf ovfs;
4527 
4528 	vfsconf_slock();
4529 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4530 		bzero(&ovfs, sizeof(ovfs));
4531 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4532 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4533 		ovfs.vfc_index = vfsp->vfc_typenum;
4534 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4535 		ovfs.vfc_flags = vfsp->vfc_flags;
4536 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4537 		if (error != 0) {
4538 			vfsconf_sunlock();
4539 			return (error);
4540 		}
4541 	}
4542 	vfsconf_sunlock();
4543 	return (0);
4544 }
4545 
4546 #endif /* 1 || COMPAT_PRELITE2 */
4547 #endif /* !BURN_BRIDGES */
4548 
4549 #define KINFO_VNODESLOP		10
4550 #ifdef notyet
4551 /*
4552  * Dump vnode list (via sysctl).
4553  */
4554 /* ARGSUSED */
4555 static int
4556 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4557 {
4558 	struct xvnode *xvn;
4559 	struct mount *mp;
4560 	struct vnode *vp;
4561 	int error, len, n;
4562 
4563 	/*
4564 	 * Stale numvnodes access is not fatal here.
4565 	 */
4566 	req->lock = 0;
4567 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4568 	if (!req->oldptr)
4569 		/* Make an estimate */
4570 		return (SYSCTL_OUT(req, 0, len));
4571 
4572 	error = sysctl_wire_old_buffer(req, 0);
4573 	if (error != 0)
4574 		return (error);
4575 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4576 	n = 0;
4577 	mtx_lock(&mountlist_mtx);
4578 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4579 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4580 			continue;
4581 		MNT_ILOCK(mp);
4582 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4583 			if (n == len)
4584 				break;
4585 			vref(vp);
4586 			xvn[n].xv_size = sizeof *xvn;
4587 			xvn[n].xv_vnode = vp;
4588 			xvn[n].xv_id = 0;	/* XXX compat */
4589 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4590 			XV_COPY(usecount);
4591 			XV_COPY(writecount);
4592 			XV_COPY(holdcnt);
4593 			XV_COPY(mount);
4594 			XV_COPY(numoutput);
4595 			XV_COPY(type);
4596 #undef XV_COPY
4597 			xvn[n].xv_flag = vp->v_vflag;
4598 
4599 			switch (vp->v_type) {
4600 			case VREG:
4601 			case VDIR:
4602 			case VLNK:
4603 				break;
4604 			case VBLK:
4605 			case VCHR:
4606 				if (vp->v_rdev == NULL) {
4607 					vrele(vp);
4608 					continue;
4609 				}
4610 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4611 				break;
4612 			case VSOCK:
4613 				xvn[n].xv_socket = vp->v_socket;
4614 				break;
4615 			case VFIFO:
4616 				xvn[n].xv_fifo = vp->v_fifoinfo;
4617 				break;
4618 			case VNON:
4619 			case VBAD:
4620 			default:
4621 				/* shouldn't happen? */
4622 				vrele(vp);
4623 				continue;
4624 			}
4625 			vrele(vp);
4626 			++n;
4627 		}
4628 		MNT_IUNLOCK(mp);
4629 		mtx_lock(&mountlist_mtx);
4630 		vfs_unbusy(mp);
4631 		if (n == len)
4632 			break;
4633 	}
4634 	mtx_unlock(&mountlist_mtx);
4635 
4636 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4637 	free(xvn, M_TEMP);
4638 	return (error);
4639 }
4640 
4641 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4642     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4643     "");
4644 #endif
4645 
4646 static void
4647 unmount_or_warn(struct mount *mp)
4648 {
4649 	int error;
4650 
4651 	error = dounmount(mp, MNT_FORCE, curthread);
4652 	if (error != 0) {
4653 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4654 		if (error == EBUSY)
4655 			printf("BUSY)\n");
4656 		else
4657 			printf("%d)\n", error);
4658 	}
4659 }
4660 
4661 /*
4662  * Unmount all filesystems. The list is traversed in reverse order
4663  * of mounting to avoid dependencies.
4664  */
4665 void
4666 vfs_unmountall(void)
4667 {
4668 	struct mount *mp, *tmp;
4669 
4670 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4671 
4672 	/*
4673 	 * Since this only runs when rebooting, it is not interlocked.
4674 	 */
4675 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4676 		vfs_ref(mp);
4677 
4678 		/*
4679 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4680 		 * unmount of the latter.
4681 		 */
4682 		if (mp == rootdevmp)
4683 			continue;
4684 
4685 		unmount_or_warn(mp);
4686 	}
4687 
4688 	if (rootdevmp != NULL)
4689 		unmount_or_warn(rootdevmp);
4690 }
4691 
4692 static void
4693 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4694 {
4695 
4696 	ASSERT_VI_LOCKED(vp, __func__);
4697 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4698 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4699 		vdropl(vp);
4700 		return;
4701 	}
4702 	if (vn_lock(vp, lkflags) == 0) {
4703 		VI_LOCK(vp);
4704 		vinactive(vp);
4705 		VOP_UNLOCK(vp);
4706 		vdropl(vp);
4707 		return;
4708 	}
4709 	vdefer_inactive_unlocked(vp);
4710 }
4711 
4712 static int
4713 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4714 {
4715 
4716 	return (vp->v_iflag & VI_DEFINACT);
4717 }
4718 
4719 static void __noinline
4720 vfs_periodic_inactive(struct mount *mp, int flags)
4721 {
4722 	struct vnode *vp, *mvp;
4723 	int lkflags;
4724 
4725 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4726 	if (flags != MNT_WAIT)
4727 		lkflags |= LK_NOWAIT;
4728 
4729 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4730 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4731 			VI_UNLOCK(vp);
4732 			continue;
4733 		}
4734 		vp->v_iflag &= ~VI_DEFINACT;
4735 		vfs_deferred_inactive(vp, lkflags);
4736 	}
4737 }
4738 
4739 static inline bool
4740 vfs_want_msync(struct vnode *vp)
4741 {
4742 	struct vm_object *obj;
4743 
4744 	/*
4745 	 * This test may be performed without any locks held.
4746 	 * We rely on vm_object's type stability.
4747 	 */
4748 	if (vp->v_vflag & VV_NOSYNC)
4749 		return (false);
4750 	obj = vp->v_object;
4751 	return (obj != NULL && vm_object_mightbedirty(obj));
4752 }
4753 
4754 static int
4755 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4756 {
4757 
4758 	if (vp->v_vflag & VV_NOSYNC)
4759 		return (false);
4760 	if (vp->v_iflag & VI_DEFINACT)
4761 		return (true);
4762 	return (vfs_want_msync(vp));
4763 }
4764 
4765 static void __noinline
4766 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4767 {
4768 	struct vnode *vp, *mvp;
4769 	struct vm_object *obj;
4770 	struct thread *td;
4771 	int lkflags, objflags;
4772 	bool seen_defer;
4773 
4774 	td = curthread;
4775 
4776 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4777 	if (flags != MNT_WAIT) {
4778 		lkflags |= LK_NOWAIT;
4779 		objflags = OBJPC_NOSYNC;
4780 	} else {
4781 		objflags = OBJPC_SYNC;
4782 	}
4783 
4784 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4785 		seen_defer = false;
4786 		if (vp->v_iflag & VI_DEFINACT) {
4787 			vp->v_iflag &= ~VI_DEFINACT;
4788 			seen_defer = true;
4789 		}
4790 		if (!vfs_want_msync(vp)) {
4791 			if (seen_defer)
4792 				vfs_deferred_inactive(vp, lkflags);
4793 			else
4794 				VI_UNLOCK(vp);
4795 			continue;
4796 		}
4797 		if (vget(vp, lkflags, td) == 0) {
4798 			obj = vp->v_object;
4799 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4800 				VM_OBJECT_WLOCK(obj);
4801 				vm_object_page_clean(obj, 0, 0, objflags);
4802 				VM_OBJECT_WUNLOCK(obj);
4803 			}
4804 			vput(vp);
4805 			if (seen_defer)
4806 				vdrop(vp);
4807 		} else {
4808 			if (seen_defer)
4809 				vdefer_inactive_unlocked(vp);
4810 		}
4811 	}
4812 }
4813 
4814 void
4815 vfs_periodic(struct mount *mp, int flags)
4816 {
4817 
4818 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4819 
4820 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4821 		vfs_periodic_inactive(mp, flags);
4822 	else
4823 		vfs_periodic_msync_inactive(mp, flags);
4824 }
4825 
4826 static void
4827 destroy_vpollinfo_free(struct vpollinfo *vi)
4828 {
4829 
4830 	knlist_destroy(&vi->vpi_selinfo.si_note);
4831 	mtx_destroy(&vi->vpi_lock);
4832 	uma_zfree(vnodepoll_zone, vi);
4833 }
4834 
4835 static void
4836 destroy_vpollinfo(struct vpollinfo *vi)
4837 {
4838 
4839 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4840 	seldrain(&vi->vpi_selinfo);
4841 	destroy_vpollinfo_free(vi);
4842 }
4843 
4844 /*
4845  * Initialize per-vnode helper structure to hold poll-related state.
4846  */
4847 void
4848 v_addpollinfo(struct vnode *vp)
4849 {
4850 	struct vpollinfo *vi;
4851 
4852 	if (vp->v_pollinfo != NULL)
4853 		return;
4854 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4855 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4856 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4857 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4858 	VI_LOCK(vp);
4859 	if (vp->v_pollinfo != NULL) {
4860 		VI_UNLOCK(vp);
4861 		destroy_vpollinfo_free(vi);
4862 		return;
4863 	}
4864 	vp->v_pollinfo = vi;
4865 	VI_UNLOCK(vp);
4866 }
4867 
4868 /*
4869  * Record a process's interest in events which might happen to
4870  * a vnode.  Because poll uses the historic select-style interface
4871  * internally, this routine serves as both the ``check for any
4872  * pending events'' and the ``record my interest in future events''
4873  * functions.  (These are done together, while the lock is held,
4874  * to avoid race conditions.)
4875  */
4876 int
4877 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4878 {
4879 
4880 	v_addpollinfo(vp);
4881 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4882 	if (vp->v_pollinfo->vpi_revents & events) {
4883 		/*
4884 		 * This leaves events we are not interested
4885 		 * in available for the other process which
4886 		 * which presumably had requested them
4887 		 * (otherwise they would never have been
4888 		 * recorded).
4889 		 */
4890 		events &= vp->v_pollinfo->vpi_revents;
4891 		vp->v_pollinfo->vpi_revents &= ~events;
4892 
4893 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4894 		return (events);
4895 	}
4896 	vp->v_pollinfo->vpi_events |= events;
4897 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4898 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4899 	return (0);
4900 }
4901 
4902 /*
4903  * Routine to create and manage a filesystem syncer vnode.
4904  */
4905 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4906 static int	sync_fsync(struct  vop_fsync_args *);
4907 static int	sync_inactive(struct  vop_inactive_args *);
4908 static int	sync_reclaim(struct  vop_reclaim_args *);
4909 
4910 static struct vop_vector sync_vnodeops = {
4911 	.vop_bypass =	VOP_EOPNOTSUPP,
4912 	.vop_close =	sync_close,		/* close */
4913 	.vop_fsync =	sync_fsync,		/* fsync */
4914 	.vop_inactive =	sync_inactive,	/* inactive */
4915 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4916 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4917 	.vop_lock1 =	vop_stdlock,	/* lock */
4918 	.vop_unlock =	vop_stdunlock,	/* unlock */
4919 	.vop_islocked =	vop_stdislocked,	/* islocked */
4920 };
4921 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4922 
4923 /*
4924  * Create a new filesystem syncer vnode for the specified mount point.
4925  */
4926 void
4927 vfs_allocate_syncvnode(struct mount *mp)
4928 {
4929 	struct vnode *vp;
4930 	struct bufobj *bo;
4931 	static long start, incr, next;
4932 	int error;
4933 
4934 	/* Allocate a new vnode */
4935 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4936 	if (error != 0)
4937 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4938 	vp->v_type = VNON;
4939 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4940 	vp->v_vflag |= VV_FORCEINSMQ;
4941 	error = insmntque(vp, mp);
4942 	if (error != 0)
4943 		panic("vfs_allocate_syncvnode: insmntque() failed");
4944 	vp->v_vflag &= ~VV_FORCEINSMQ;
4945 	VOP_UNLOCK(vp);
4946 	/*
4947 	 * Place the vnode onto the syncer worklist. We attempt to
4948 	 * scatter them about on the list so that they will go off
4949 	 * at evenly distributed times even if all the filesystems
4950 	 * are mounted at once.
4951 	 */
4952 	next += incr;
4953 	if (next == 0 || next > syncer_maxdelay) {
4954 		start /= 2;
4955 		incr /= 2;
4956 		if (start == 0) {
4957 			start = syncer_maxdelay / 2;
4958 			incr = syncer_maxdelay;
4959 		}
4960 		next = start;
4961 	}
4962 	bo = &vp->v_bufobj;
4963 	BO_LOCK(bo);
4964 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4965 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4966 	mtx_lock(&sync_mtx);
4967 	sync_vnode_count++;
4968 	if (mp->mnt_syncer == NULL) {
4969 		mp->mnt_syncer = vp;
4970 		vp = NULL;
4971 	}
4972 	mtx_unlock(&sync_mtx);
4973 	BO_UNLOCK(bo);
4974 	if (vp != NULL) {
4975 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4976 		vgone(vp);
4977 		vput(vp);
4978 	}
4979 }
4980 
4981 void
4982 vfs_deallocate_syncvnode(struct mount *mp)
4983 {
4984 	struct vnode *vp;
4985 
4986 	mtx_lock(&sync_mtx);
4987 	vp = mp->mnt_syncer;
4988 	if (vp != NULL)
4989 		mp->mnt_syncer = NULL;
4990 	mtx_unlock(&sync_mtx);
4991 	if (vp != NULL)
4992 		vrele(vp);
4993 }
4994 
4995 /*
4996  * Do a lazy sync of the filesystem.
4997  */
4998 static int
4999 sync_fsync(struct vop_fsync_args *ap)
5000 {
5001 	struct vnode *syncvp = ap->a_vp;
5002 	struct mount *mp = syncvp->v_mount;
5003 	int error, save;
5004 	struct bufobj *bo;
5005 
5006 	/*
5007 	 * We only need to do something if this is a lazy evaluation.
5008 	 */
5009 	if (ap->a_waitfor != MNT_LAZY)
5010 		return (0);
5011 
5012 	/*
5013 	 * Move ourselves to the back of the sync list.
5014 	 */
5015 	bo = &syncvp->v_bufobj;
5016 	BO_LOCK(bo);
5017 	vn_syncer_add_to_worklist(bo, syncdelay);
5018 	BO_UNLOCK(bo);
5019 
5020 	/*
5021 	 * Walk the list of vnodes pushing all that are dirty and
5022 	 * not already on the sync list.
5023 	 */
5024 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5025 		return (0);
5026 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5027 		vfs_unbusy(mp);
5028 		return (0);
5029 	}
5030 	save = curthread_pflags_set(TDP_SYNCIO);
5031 	/*
5032 	 * The filesystem at hand may be idle with free vnodes stored in the
5033 	 * batch.  Return them instead of letting them stay there indefinitely.
5034 	 */
5035 	vfs_periodic(mp, MNT_NOWAIT);
5036 	error = VFS_SYNC(mp, MNT_LAZY);
5037 	curthread_pflags_restore(save);
5038 	vn_finished_write(mp);
5039 	vfs_unbusy(mp);
5040 	return (error);
5041 }
5042 
5043 /*
5044  * The syncer vnode is no referenced.
5045  */
5046 static int
5047 sync_inactive(struct vop_inactive_args *ap)
5048 {
5049 
5050 	vgone(ap->a_vp);
5051 	return (0);
5052 }
5053 
5054 /*
5055  * The syncer vnode is no longer needed and is being decommissioned.
5056  *
5057  * Modifications to the worklist must be protected by sync_mtx.
5058  */
5059 static int
5060 sync_reclaim(struct vop_reclaim_args *ap)
5061 {
5062 	struct vnode *vp = ap->a_vp;
5063 	struct bufobj *bo;
5064 
5065 	bo = &vp->v_bufobj;
5066 	BO_LOCK(bo);
5067 	mtx_lock(&sync_mtx);
5068 	if (vp->v_mount->mnt_syncer == vp)
5069 		vp->v_mount->mnt_syncer = NULL;
5070 	if (bo->bo_flag & BO_ONWORKLST) {
5071 		LIST_REMOVE(bo, bo_synclist);
5072 		syncer_worklist_len--;
5073 		sync_vnode_count--;
5074 		bo->bo_flag &= ~BO_ONWORKLST;
5075 	}
5076 	mtx_unlock(&sync_mtx);
5077 	BO_UNLOCK(bo);
5078 
5079 	return (0);
5080 }
5081 
5082 int
5083 vn_need_pageq_flush(struct vnode *vp)
5084 {
5085 	struct vm_object *obj;
5086 	int need;
5087 
5088 	MPASS(mtx_owned(VI_MTX(vp)));
5089 	need = 0;
5090 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5091 	    vm_object_mightbedirty(obj))
5092 		need = 1;
5093 	return (need);
5094 }
5095 
5096 /*
5097  * Check if vnode represents a disk device
5098  */
5099 int
5100 vn_isdisk(struct vnode *vp, int *errp)
5101 {
5102 	int error;
5103 
5104 	if (vp->v_type != VCHR) {
5105 		error = ENOTBLK;
5106 		goto out;
5107 	}
5108 	error = 0;
5109 	dev_lock();
5110 	if (vp->v_rdev == NULL)
5111 		error = ENXIO;
5112 	else if (vp->v_rdev->si_devsw == NULL)
5113 		error = ENXIO;
5114 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5115 		error = ENOTBLK;
5116 	dev_unlock();
5117 out:
5118 	if (errp != NULL)
5119 		*errp = error;
5120 	return (error == 0);
5121 }
5122 
5123 /*
5124  * Common filesystem object access control check routine.  Accepts a
5125  * vnode's type, "mode", uid and gid, requested access mode, credentials,
5126  * and optional call-by-reference privused argument allowing vaccess()
5127  * to indicate to the caller whether privilege was used to satisfy the
5128  * request (obsoleted).  Returns 0 on success, or an errno on failure.
5129  */
5130 int
5131 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5132     accmode_t accmode, struct ucred *cred, int *privused)
5133 {
5134 	accmode_t dac_granted;
5135 	accmode_t priv_granted;
5136 
5137 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5138 	    ("invalid bit in accmode"));
5139 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5140 	    ("VAPPEND without VWRITE"));
5141 
5142 	/*
5143 	 * Look for a normal, non-privileged way to access the file/directory
5144 	 * as requested.  If it exists, go with that.
5145 	 */
5146 
5147 	if (privused != NULL)
5148 		*privused = 0;
5149 
5150 	dac_granted = 0;
5151 
5152 	/* Check the owner. */
5153 	if (cred->cr_uid == file_uid) {
5154 		dac_granted |= VADMIN;
5155 		if (file_mode & S_IXUSR)
5156 			dac_granted |= VEXEC;
5157 		if (file_mode & S_IRUSR)
5158 			dac_granted |= VREAD;
5159 		if (file_mode & S_IWUSR)
5160 			dac_granted |= (VWRITE | VAPPEND);
5161 
5162 		if ((accmode & dac_granted) == accmode)
5163 			return (0);
5164 
5165 		goto privcheck;
5166 	}
5167 
5168 	/* Otherwise, check the groups (first match) */
5169 	if (groupmember(file_gid, cred)) {
5170 		if (file_mode & S_IXGRP)
5171 			dac_granted |= VEXEC;
5172 		if (file_mode & S_IRGRP)
5173 			dac_granted |= VREAD;
5174 		if (file_mode & S_IWGRP)
5175 			dac_granted |= (VWRITE | VAPPEND);
5176 
5177 		if ((accmode & dac_granted) == accmode)
5178 			return (0);
5179 
5180 		goto privcheck;
5181 	}
5182 
5183 	/* Otherwise, check everyone else. */
5184 	if (file_mode & S_IXOTH)
5185 		dac_granted |= VEXEC;
5186 	if (file_mode & S_IROTH)
5187 		dac_granted |= VREAD;
5188 	if (file_mode & S_IWOTH)
5189 		dac_granted |= (VWRITE | VAPPEND);
5190 	if ((accmode & dac_granted) == accmode)
5191 		return (0);
5192 
5193 privcheck:
5194 	/*
5195 	 * Build a privilege mask to determine if the set of privileges
5196 	 * satisfies the requirements when combined with the granted mask
5197 	 * from above.  For each privilege, if the privilege is required,
5198 	 * bitwise or the request type onto the priv_granted mask.
5199 	 */
5200 	priv_granted = 0;
5201 
5202 	if (type == VDIR) {
5203 		/*
5204 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5205 		 * requests, instead of PRIV_VFS_EXEC.
5206 		 */
5207 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5208 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5209 			priv_granted |= VEXEC;
5210 	} else {
5211 		/*
5212 		 * Ensure that at least one execute bit is on. Otherwise,
5213 		 * a privileged user will always succeed, and we don't want
5214 		 * this to happen unless the file really is executable.
5215 		 */
5216 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5217 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5218 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5219 			priv_granted |= VEXEC;
5220 	}
5221 
5222 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5223 	    !priv_check_cred(cred, PRIV_VFS_READ))
5224 		priv_granted |= VREAD;
5225 
5226 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5227 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5228 		priv_granted |= (VWRITE | VAPPEND);
5229 
5230 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5231 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5232 		priv_granted |= VADMIN;
5233 
5234 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5235 		/* XXX audit: privilege used */
5236 		if (privused != NULL)
5237 			*privused = 1;
5238 		return (0);
5239 	}
5240 
5241 	return ((accmode & VADMIN) ? EPERM : EACCES);
5242 }
5243 
5244 /*
5245  * Credential check based on process requesting service, and per-attribute
5246  * permissions.
5247  */
5248 int
5249 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5250     struct thread *td, accmode_t accmode)
5251 {
5252 
5253 	/*
5254 	 * Kernel-invoked always succeeds.
5255 	 */
5256 	if (cred == NOCRED)
5257 		return (0);
5258 
5259 	/*
5260 	 * Do not allow privileged processes in jail to directly manipulate
5261 	 * system attributes.
5262 	 */
5263 	switch (attrnamespace) {
5264 	case EXTATTR_NAMESPACE_SYSTEM:
5265 		/* Potentially should be: return (EPERM); */
5266 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5267 	case EXTATTR_NAMESPACE_USER:
5268 		return (VOP_ACCESS(vp, accmode, cred, td));
5269 	default:
5270 		return (EPERM);
5271 	}
5272 }
5273 
5274 #ifdef DEBUG_VFS_LOCKS
5275 /*
5276  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5277  * no longer ok to have an unlocked VFS.
5278  */
5279 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5280 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5281 
5282 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5283 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5284     "Drop into debugger on lock violation");
5285 
5286 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5287 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5288     0, "Check for interlock across VOPs");
5289 
5290 int vfs_badlock_print = 1;	/* Print lock violations. */
5291 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5292     0, "Print lock violations");
5293 
5294 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5295 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5296     0, "Print vnode details on lock violations");
5297 
5298 #ifdef KDB
5299 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5300 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5301     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5302 #endif
5303 
5304 static void
5305 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5306 {
5307 
5308 #ifdef KDB
5309 	if (vfs_badlock_backtrace)
5310 		kdb_backtrace();
5311 #endif
5312 	if (vfs_badlock_vnode)
5313 		vn_printf(vp, "vnode ");
5314 	if (vfs_badlock_print)
5315 		printf("%s: %p %s\n", str, (void *)vp, msg);
5316 	if (vfs_badlock_ddb)
5317 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5318 }
5319 
5320 void
5321 assert_vi_locked(struct vnode *vp, const char *str)
5322 {
5323 
5324 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5325 		vfs_badlock("interlock is not locked but should be", str, vp);
5326 }
5327 
5328 void
5329 assert_vi_unlocked(struct vnode *vp, const char *str)
5330 {
5331 
5332 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5333 		vfs_badlock("interlock is locked but should not be", str, vp);
5334 }
5335 
5336 void
5337 assert_vop_locked(struct vnode *vp, const char *str)
5338 {
5339 	int locked;
5340 
5341 	if (!IGNORE_LOCK(vp)) {
5342 		locked = VOP_ISLOCKED(vp);
5343 		if (locked == 0 || locked == LK_EXCLOTHER)
5344 			vfs_badlock("is not locked but should be", str, vp);
5345 	}
5346 }
5347 
5348 void
5349 assert_vop_unlocked(struct vnode *vp, const char *str)
5350 {
5351 
5352 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5353 		vfs_badlock("is locked but should not be", str, vp);
5354 }
5355 
5356 void
5357 assert_vop_elocked(struct vnode *vp, const char *str)
5358 {
5359 
5360 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5361 		vfs_badlock("is not exclusive locked but should be", str, vp);
5362 }
5363 #endif /* DEBUG_VFS_LOCKS */
5364 
5365 void
5366 vop_rename_fail(struct vop_rename_args *ap)
5367 {
5368 
5369 	if (ap->a_tvp != NULL)
5370 		vput(ap->a_tvp);
5371 	if (ap->a_tdvp == ap->a_tvp)
5372 		vrele(ap->a_tdvp);
5373 	else
5374 		vput(ap->a_tdvp);
5375 	vrele(ap->a_fdvp);
5376 	vrele(ap->a_fvp);
5377 }
5378 
5379 void
5380 vop_rename_pre(void *ap)
5381 {
5382 	struct vop_rename_args *a = ap;
5383 
5384 #ifdef DEBUG_VFS_LOCKS
5385 	if (a->a_tvp)
5386 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5387 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5388 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5389 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5390 
5391 	/* Check the source (from). */
5392 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5393 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5394 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5395 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5396 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5397 
5398 	/* Check the target. */
5399 	if (a->a_tvp)
5400 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5401 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5402 #endif
5403 	if (a->a_tdvp != a->a_fdvp)
5404 		vhold(a->a_fdvp);
5405 	if (a->a_tvp != a->a_fvp)
5406 		vhold(a->a_fvp);
5407 	vhold(a->a_tdvp);
5408 	if (a->a_tvp)
5409 		vhold(a->a_tvp);
5410 }
5411 
5412 #ifdef DEBUG_VFS_LOCKS
5413 void
5414 vop_strategy_pre(void *ap)
5415 {
5416 	struct vop_strategy_args *a;
5417 	struct buf *bp;
5418 
5419 	a = ap;
5420 	bp = a->a_bp;
5421 
5422 	/*
5423 	 * Cluster ops lock their component buffers but not the IO container.
5424 	 */
5425 	if ((bp->b_flags & B_CLUSTER) != 0)
5426 		return;
5427 
5428 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5429 		if (vfs_badlock_print)
5430 			printf(
5431 			    "VOP_STRATEGY: bp is not locked but should be\n");
5432 		if (vfs_badlock_ddb)
5433 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5434 	}
5435 }
5436 
5437 void
5438 vop_lock_pre(void *ap)
5439 {
5440 	struct vop_lock1_args *a = ap;
5441 
5442 	if ((a->a_flags & LK_INTERLOCK) == 0)
5443 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5444 	else
5445 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5446 }
5447 
5448 void
5449 vop_lock_post(void *ap, int rc)
5450 {
5451 	struct vop_lock1_args *a = ap;
5452 
5453 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5454 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5455 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5456 }
5457 
5458 void
5459 vop_unlock_pre(void *ap)
5460 {
5461 	struct vop_unlock_args *a = ap;
5462 
5463 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5464 }
5465 
5466 void
5467 vop_need_inactive_pre(void *ap)
5468 {
5469 	struct vop_need_inactive_args *a = ap;
5470 
5471 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5472 }
5473 
5474 void
5475 vop_need_inactive_post(void *ap, int rc)
5476 {
5477 	struct vop_need_inactive_args *a = ap;
5478 
5479 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5480 }
5481 #endif
5482 
5483 void
5484 vop_create_post(void *ap, int rc)
5485 {
5486 	struct vop_create_args *a = ap;
5487 
5488 	if (!rc)
5489 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5490 }
5491 
5492 void
5493 vop_deleteextattr_post(void *ap, int rc)
5494 {
5495 	struct vop_deleteextattr_args *a = ap;
5496 
5497 	if (!rc)
5498 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5499 }
5500 
5501 void
5502 vop_link_post(void *ap, int rc)
5503 {
5504 	struct vop_link_args *a = ap;
5505 
5506 	if (!rc) {
5507 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5508 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5509 	}
5510 }
5511 
5512 void
5513 vop_mkdir_post(void *ap, int rc)
5514 {
5515 	struct vop_mkdir_args *a = ap;
5516 
5517 	if (!rc)
5518 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5519 }
5520 
5521 void
5522 vop_mknod_post(void *ap, int rc)
5523 {
5524 	struct vop_mknod_args *a = ap;
5525 
5526 	if (!rc)
5527 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5528 }
5529 
5530 void
5531 vop_reclaim_post(void *ap, int rc)
5532 {
5533 	struct vop_reclaim_args *a = ap;
5534 
5535 	if (!rc)
5536 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5537 }
5538 
5539 void
5540 vop_remove_post(void *ap, int rc)
5541 {
5542 	struct vop_remove_args *a = ap;
5543 
5544 	if (!rc) {
5545 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5546 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5547 	}
5548 }
5549 
5550 void
5551 vop_rename_post(void *ap, int rc)
5552 {
5553 	struct vop_rename_args *a = ap;
5554 	long hint;
5555 
5556 	if (!rc) {
5557 		hint = NOTE_WRITE;
5558 		if (a->a_fdvp == a->a_tdvp) {
5559 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5560 				hint |= NOTE_LINK;
5561 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5562 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5563 		} else {
5564 			hint |= NOTE_EXTEND;
5565 			if (a->a_fvp->v_type == VDIR)
5566 				hint |= NOTE_LINK;
5567 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5568 
5569 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5570 			    a->a_tvp->v_type == VDIR)
5571 				hint &= ~NOTE_LINK;
5572 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5573 		}
5574 
5575 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5576 		if (a->a_tvp)
5577 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5578 	}
5579 	if (a->a_tdvp != a->a_fdvp)
5580 		vdrop(a->a_fdvp);
5581 	if (a->a_tvp != a->a_fvp)
5582 		vdrop(a->a_fvp);
5583 	vdrop(a->a_tdvp);
5584 	if (a->a_tvp)
5585 		vdrop(a->a_tvp);
5586 }
5587 
5588 void
5589 vop_rmdir_post(void *ap, int rc)
5590 {
5591 	struct vop_rmdir_args *a = ap;
5592 
5593 	if (!rc) {
5594 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5595 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5596 	}
5597 }
5598 
5599 void
5600 vop_setattr_post(void *ap, int rc)
5601 {
5602 	struct vop_setattr_args *a = ap;
5603 
5604 	if (!rc)
5605 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5606 }
5607 
5608 void
5609 vop_setextattr_post(void *ap, int rc)
5610 {
5611 	struct vop_setextattr_args *a = ap;
5612 
5613 	if (!rc)
5614 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5615 }
5616 
5617 void
5618 vop_symlink_post(void *ap, int rc)
5619 {
5620 	struct vop_symlink_args *a = ap;
5621 
5622 	if (!rc)
5623 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5624 }
5625 
5626 void
5627 vop_open_post(void *ap, int rc)
5628 {
5629 	struct vop_open_args *a = ap;
5630 
5631 	if (!rc)
5632 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5633 }
5634 
5635 void
5636 vop_close_post(void *ap, int rc)
5637 {
5638 	struct vop_close_args *a = ap;
5639 
5640 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5641 	    !VN_IS_DOOMED(a->a_vp))) {
5642 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5643 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5644 	}
5645 }
5646 
5647 void
5648 vop_read_post(void *ap, int rc)
5649 {
5650 	struct vop_read_args *a = ap;
5651 
5652 	if (!rc)
5653 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5654 }
5655 
5656 void
5657 vop_readdir_post(void *ap, int rc)
5658 {
5659 	struct vop_readdir_args *a = ap;
5660 
5661 	if (!rc)
5662 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5663 }
5664 
5665 static struct knlist fs_knlist;
5666 
5667 static void
5668 vfs_event_init(void *arg)
5669 {
5670 	knlist_init_mtx(&fs_knlist, NULL);
5671 }
5672 /* XXX - correct order? */
5673 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5674 
5675 void
5676 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5677 {
5678 
5679 	KNOTE_UNLOCKED(&fs_knlist, event);
5680 }
5681 
5682 static int	filt_fsattach(struct knote *kn);
5683 static void	filt_fsdetach(struct knote *kn);
5684 static int	filt_fsevent(struct knote *kn, long hint);
5685 
5686 struct filterops fs_filtops = {
5687 	.f_isfd = 0,
5688 	.f_attach = filt_fsattach,
5689 	.f_detach = filt_fsdetach,
5690 	.f_event = filt_fsevent
5691 };
5692 
5693 static int
5694 filt_fsattach(struct knote *kn)
5695 {
5696 
5697 	kn->kn_flags |= EV_CLEAR;
5698 	knlist_add(&fs_knlist, kn, 0);
5699 	return (0);
5700 }
5701 
5702 static void
5703 filt_fsdetach(struct knote *kn)
5704 {
5705 
5706 	knlist_remove(&fs_knlist, kn, 0);
5707 }
5708 
5709 static int
5710 filt_fsevent(struct knote *kn, long hint)
5711 {
5712 
5713 	kn->kn_fflags |= hint;
5714 	return (kn->kn_fflags != 0);
5715 }
5716 
5717 static int
5718 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5719 {
5720 	struct vfsidctl vc;
5721 	int error;
5722 	struct mount *mp;
5723 
5724 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5725 	if (error)
5726 		return (error);
5727 	if (vc.vc_vers != VFS_CTL_VERS1)
5728 		return (EINVAL);
5729 	mp = vfs_getvfs(&vc.vc_fsid);
5730 	if (mp == NULL)
5731 		return (ENOENT);
5732 	/* ensure that a specific sysctl goes to the right filesystem. */
5733 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5734 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5735 		vfs_rel(mp);
5736 		return (EINVAL);
5737 	}
5738 	VCTLTOREQ(&vc, req);
5739 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5740 	vfs_rel(mp);
5741 	return (error);
5742 }
5743 
5744 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5745     NULL, 0, sysctl_vfs_ctl, "",
5746     "Sysctl by fsid");
5747 
5748 /*
5749  * Function to initialize a va_filerev field sensibly.
5750  * XXX: Wouldn't a random number make a lot more sense ??
5751  */
5752 u_quad_t
5753 init_va_filerev(void)
5754 {
5755 	struct bintime bt;
5756 
5757 	getbinuptime(&bt);
5758 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5759 }
5760 
5761 static int	filt_vfsread(struct knote *kn, long hint);
5762 static int	filt_vfswrite(struct knote *kn, long hint);
5763 static int	filt_vfsvnode(struct knote *kn, long hint);
5764 static void	filt_vfsdetach(struct knote *kn);
5765 static struct filterops vfsread_filtops = {
5766 	.f_isfd = 1,
5767 	.f_detach = filt_vfsdetach,
5768 	.f_event = filt_vfsread
5769 };
5770 static struct filterops vfswrite_filtops = {
5771 	.f_isfd = 1,
5772 	.f_detach = filt_vfsdetach,
5773 	.f_event = filt_vfswrite
5774 };
5775 static struct filterops vfsvnode_filtops = {
5776 	.f_isfd = 1,
5777 	.f_detach = filt_vfsdetach,
5778 	.f_event = filt_vfsvnode
5779 };
5780 
5781 static void
5782 vfs_knllock(void *arg)
5783 {
5784 	struct vnode *vp = arg;
5785 
5786 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5787 }
5788 
5789 static void
5790 vfs_knlunlock(void *arg)
5791 {
5792 	struct vnode *vp = arg;
5793 
5794 	VOP_UNLOCK(vp);
5795 }
5796 
5797 static void
5798 vfs_knl_assert_locked(void *arg)
5799 {
5800 #ifdef DEBUG_VFS_LOCKS
5801 	struct vnode *vp = arg;
5802 
5803 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5804 #endif
5805 }
5806 
5807 static void
5808 vfs_knl_assert_unlocked(void *arg)
5809 {
5810 #ifdef DEBUG_VFS_LOCKS
5811 	struct vnode *vp = arg;
5812 
5813 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5814 #endif
5815 }
5816 
5817 int
5818 vfs_kqfilter(struct vop_kqfilter_args *ap)
5819 {
5820 	struct vnode *vp = ap->a_vp;
5821 	struct knote *kn = ap->a_kn;
5822 	struct knlist *knl;
5823 
5824 	switch (kn->kn_filter) {
5825 	case EVFILT_READ:
5826 		kn->kn_fop = &vfsread_filtops;
5827 		break;
5828 	case EVFILT_WRITE:
5829 		kn->kn_fop = &vfswrite_filtops;
5830 		break;
5831 	case EVFILT_VNODE:
5832 		kn->kn_fop = &vfsvnode_filtops;
5833 		break;
5834 	default:
5835 		return (EINVAL);
5836 	}
5837 
5838 	kn->kn_hook = (caddr_t)vp;
5839 
5840 	v_addpollinfo(vp);
5841 	if (vp->v_pollinfo == NULL)
5842 		return (ENOMEM);
5843 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5844 	vhold(vp);
5845 	knlist_add(knl, kn, 0);
5846 
5847 	return (0);
5848 }
5849 
5850 /*
5851  * Detach knote from vnode
5852  */
5853 static void
5854 filt_vfsdetach(struct knote *kn)
5855 {
5856 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5857 
5858 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5859 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5860 	vdrop(vp);
5861 }
5862 
5863 /*ARGSUSED*/
5864 static int
5865 filt_vfsread(struct knote *kn, long hint)
5866 {
5867 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5868 	struct vattr va;
5869 	int res;
5870 
5871 	/*
5872 	 * filesystem is gone, so set the EOF flag and schedule
5873 	 * the knote for deletion.
5874 	 */
5875 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5876 		VI_LOCK(vp);
5877 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5878 		VI_UNLOCK(vp);
5879 		return (1);
5880 	}
5881 
5882 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5883 		return (0);
5884 
5885 	VI_LOCK(vp);
5886 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5887 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5888 	VI_UNLOCK(vp);
5889 	return (res);
5890 }
5891 
5892 /*ARGSUSED*/
5893 static int
5894 filt_vfswrite(struct knote *kn, long hint)
5895 {
5896 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5897 
5898 	VI_LOCK(vp);
5899 
5900 	/*
5901 	 * filesystem is gone, so set the EOF flag and schedule
5902 	 * the knote for deletion.
5903 	 */
5904 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5905 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5906 
5907 	kn->kn_data = 0;
5908 	VI_UNLOCK(vp);
5909 	return (1);
5910 }
5911 
5912 static int
5913 filt_vfsvnode(struct knote *kn, long hint)
5914 {
5915 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5916 	int res;
5917 
5918 	VI_LOCK(vp);
5919 	if (kn->kn_sfflags & hint)
5920 		kn->kn_fflags |= hint;
5921 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5922 		kn->kn_flags |= EV_EOF;
5923 		VI_UNLOCK(vp);
5924 		return (1);
5925 	}
5926 	res = (kn->kn_fflags != 0);
5927 	VI_UNLOCK(vp);
5928 	return (res);
5929 }
5930 
5931 /*
5932  * Returns whether the directory is empty or not.
5933  * If it is empty, the return value is 0; otherwise
5934  * the return value is an error value (which may
5935  * be ENOTEMPTY).
5936  */
5937 int
5938 vfs_emptydir(struct vnode *vp)
5939 {
5940 	struct uio uio;
5941 	struct iovec iov;
5942 	struct dirent *dirent, *dp, *endp;
5943 	int error, eof;
5944 
5945 	error = 0;
5946 	eof = 0;
5947 
5948 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5949 
5950 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5951 	iov.iov_base = dirent;
5952 	iov.iov_len = sizeof(struct dirent);
5953 
5954 	uio.uio_iov = &iov;
5955 	uio.uio_iovcnt = 1;
5956 	uio.uio_offset = 0;
5957 	uio.uio_resid = sizeof(struct dirent);
5958 	uio.uio_segflg = UIO_SYSSPACE;
5959 	uio.uio_rw = UIO_READ;
5960 	uio.uio_td = curthread;
5961 
5962 	while (eof == 0 && error == 0) {
5963 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5964 		    NULL, NULL);
5965 		if (error != 0)
5966 			break;
5967 		endp = (void *)((uint8_t *)dirent +
5968 		    sizeof(struct dirent) - uio.uio_resid);
5969 		for (dp = dirent; dp < endp;
5970 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5971 			if (dp->d_type == DT_WHT)
5972 				continue;
5973 			if (dp->d_namlen == 0)
5974 				continue;
5975 			if (dp->d_type != DT_DIR &&
5976 			    dp->d_type != DT_UNKNOWN) {
5977 				error = ENOTEMPTY;
5978 				break;
5979 			}
5980 			if (dp->d_namlen > 2) {
5981 				error = ENOTEMPTY;
5982 				break;
5983 			}
5984 			if (dp->d_namlen == 1 &&
5985 			    dp->d_name[0] != '.') {
5986 				error = ENOTEMPTY;
5987 				break;
5988 			}
5989 			if (dp->d_namlen == 2 &&
5990 			    dp->d_name[1] != '.') {
5991 				error = ENOTEMPTY;
5992 				break;
5993 			}
5994 			uio.uio_resid = sizeof(struct dirent);
5995 		}
5996 	}
5997 	free(dirent, M_TEMP);
5998 	return (error);
5999 }
6000 
6001 int
6002 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6003 {
6004 	int error;
6005 
6006 	if (dp->d_reclen > ap->a_uio->uio_resid)
6007 		return (ENAMETOOLONG);
6008 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6009 	if (error) {
6010 		if (ap->a_ncookies != NULL) {
6011 			if (ap->a_cookies != NULL)
6012 				free(ap->a_cookies, M_TEMP);
6013 			ap->a_cookies = NULL;
6014 			*ap->a_ncookies = 0;
6015 		}
6016 		return (error);
6017 	}
6018 	if (ap->a_ncookies == NULL)
6019 		return (0);
6020 
6021 	KASSERT(ap->a_cookies,
6022 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6023 
6024 	*ap->a_cookies = realloc(*ap->a_cookies,
6025 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6026 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6027 	*ap->a_ncookies += 1;
6028 	return (0);
6029 }
6030 
6031 /*
6032  * The purpose of this routine is to remove granularity from accmode_t,
6033  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6034  * VADMIN and VAPPEND.
6035  *
6036  * If it returns 0, the caller is supposed to continue with the usual
6037  * access checks using 'accmode' as modified by this routine.  If it
6038  * returns nonzero value, the caller is supposed to return that value
6039  * as errno.
6040  *
6041  * Note that after this routine runs, accmode may be zero.
6042  */
6043 int
6044 vfs_unixify_accmode(accmode_t *accmode)
6045 {
6046 	/*
6047 	 * There is no way to specify explicit "deny" rule using
6048 	 * file mode or POSIX.1e ACLs.
6049 	 */
6050 	if (*accmode & VEXPLICIT_DENY) {
6051 		*accmode = 0;
6052 		return (0);
6053 	}
6054 
6055 	/*
6056 	 * None of these can be translated into usual access bits.
6057 	 * Also, the common case for NFSv4 ACLs is to not contain
6058 	 * either of these bits. Caller should check for VWRITE
6059 	 * on the containing directory instead.
6060 	 */
6061 	if (*accmode & (VDELETE_CHILD | VDELETE))
6062 		return (EPERM);
6063 
6064 	if (*accmode & VADMIN_PERMS) {
6065 		*accmode &= ~VADMIN_PERMS;
6066 		*accmode |= VADMIN;
6067 	}
6068 
6069 	/*
6070 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6071 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6072 	 */
6073 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6074 
6075 	return (0);
6076 }
6077 
6078 /*
6079  * Clear out a doomed vnode (if any) and replace it with a new one as long
6080  * as the fs is not being unmounted. Return the root vnode to the caller.
6081  */
6082 static int __noinline
6083 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6084 {
6085 	struct vnode *vp;
6086 	int error;
6087 
6088 restart:
6089 	if (mp->mnt_rootvnode != NULL) {
6090 		MNT_ILOCK(mp);
6091 		vp = mp->mnt_rootvnode;
6092 		if (vp != NULL) {
6093 			if (!VN_IS_DOOMED(vp)) {
6094 				vrefact(vp);
6095 				MNT_IUNLOCK(mp);
6096 				error = vn_lock(vp, flags);
6097 				if (error == 0) {
6098 					*vpp = vp;
6099 					return (0);
6100 				}
6101 				vrele(vp);
6102 				goto restart;
6103 			}
6104 			/*
6105 			 * Clear the old one.
6106 			 */
6107 			mp->mnt_rootvnode = NULL;
6108 		}
6109 		MNT_IUNLOCK(mp);
6110 		if (vp != NULL) {
6111 			vfs_op_barrier_wait(mp);
6112 			vrele(vp);
6113 		}
6114 	}
6115 	error = VFS_CACHEDROOT(mp, flags, vpp);
6116 	if (error != 0)
6117 		return (error);
6118 	if (mp->mnt_vfs_ops == 0) {
6119 		MNT_ILOCK(mp);
6120 		if (mp->mnt_vfs_ops != 0) {
6121 			MNT_IUNLOCK(mp);
6122 			return (0);
6123 		}
6124 		if (mp->mnt_rootvnode == NULL) {
6125 			vrefact(*vpp);
6126 			mp->mnt_rootvnode = *vpp;
6127 		} else {
6128 			if (mp->mnt_rootvnode != *vpp) {
6129 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6130 					panic("%s: mismatch between vnode returned "
6131 					    " by VFS_CACHEDROOT and the one cached "
6132 					    " (%p != %p)",
6133 					    __func__, *vpp, mp->mnt_rootvnode);
6134 				}
6135 			}
6136 		}
6137 		MNT_IUNLOCK(mp);
6138 	}
6139 	return (0);
6140 }
6141 
6142 int
6143 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6144 {
6145 	struct vnode *vp;
6146 	int error;
6147 
6148 	if (!vfs_op_thread_enter(mp))
6149 		return (vfs_cache_root_fallback(mp, flags, vpp));
6150 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6151 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6152 		vfs_op_thread_exit(mp);
6153 		return (vfs_cache_root_fallback(mp, flags, vpp));
6154 	}
6155 	vrefact(vp);
6156 	vfs_op_thread_exit(mp);
6157 	error = vn_lock(vp, flags);
6158 	if (error != 0) {
6159 		vrele(vp);
6160 		return (vfs_cache_root_fallback(mp, flags, vpp));
6161 	}
6162 	*vpp = vp;
6163 	return (0);
6164 }
6165 
6166 struct vnode *
6167 vfs_cache_root_clear(struct mount *mp)
6168 {
6169 	struct vnode *vp;
6170 
6171 	/*
6172 	 * ops > 0 guarantees there is nobody who can see this vnode
6173 	 */
6174 	MPASS(mp->mnt_vfs_ops > 0);
6175 	vp = mp->mnt_rootvnode;
6176 	mp->mnt_rootvnode = NULL;
6177 	return (vp);
6178 }
6179 
6180 void
6181 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6182 {
6183 
6184 	MPASS(mp->mnt_vfs_ops > 0);
6185 	vrefact(vp);
6186 	mp->mnt_rootvnode = vp;
6187 }
6188 
6189 /*
6190  * These are helper functions for filesystems to traverse all
6191  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6192  *
6193  * This interface replaces MNT_VNODE_FOREACH.
6194  */
6195 
6196 struct vnode *
6197 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6198 {
6199 	struct vnode *vp;
6200 
6201 	if (should_yield())
6202 		kern_yield(PRI_USER);
6203 	MNT_ILOCK(mp);
6204 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6205 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6206 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6207 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6208 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6209 			continue;
6210 		VI_LOCK(vp);
6211 		if (VN_IS_DOOMED(vp)) {
6212 			VI_UNLOCK(vp);
6213 			continue;
6214 		}
6215 		break;
6216 	}
6217 	if (vp == NULL) {
6218 		__mnt_vnode_markerfree_all(mvp, mp);
6219 		/* MNT_IUNLOCK(mp); -- done in above function */
6220 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6221 		return (NULL);
6222 	}
6223 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6224 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6225 	MNT_IUNLOCK(mp);
6226 	return (vp);
6227 }
6228 
6229 struct vnode *
6230 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6231 {
6232 	struct vnode *vp;
6233 
6234 	*mvp = vn_alloc_marker(mp);
6235 	MNT_ILOCK(mp);
6236 	MNT_REF(mp);
6237 
6238 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6239 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6240 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6241 			continue;
6242 		VI_LOCK(vp);
6243 		if (VN_IS_DOOMED(vp)) {
6244 			VI_UNLOCK(vp);
6245 			continue;
6246 		}
6247 		break;
6248 	}
6249 	if (vp == NULL) {
6250 		MNT_REL(mp);
6251 		MNT_IUNLOCK(mp);
6252 		vn_free_marker(*mvp);
6253 		*mvp = NULL;
6254 		return (NULL);
6255 	}
6256 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6257 	MNT_IUNLOCK(mp);
6258 	return (vp);
6259 }
6260 
6261 void
6262 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6263 {
6264 
6265 	if (*mvp == NULL) {
6266 		MNT_IUNLOCK(mp);
6267 		return;
6268 	}
6269 
6270 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6271 
6272 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6273 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6274 	MNT_REL(mp);
6275 	MNT_IUNLOCK(mp);
6276 	vn_free_marker(*mvp);
6277 	*mvp = NULL;
6278 }
6279 
6280 /*
6281  * These are helper functions for filesystems to traverse their
6282  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6283  */
6284 static void
6285 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6286 {
6287 
6288 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6289 
6290 	MNT_ILOCK(mp);
6291 	MNT_REL(mp);
6292 	MNT_IUNLOCK(mp);
6293 	vn_free_marker(*mvp);
6294 	*mvp = NULL;
6295 }
6296 
6297 /*
6298  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6299  * conventional lock order during mnt_vnode_next_lazy iteration.
6300  *
6301  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6302  * The list lock is dropped and reacquired.  On success, both locks are held.
6303  * On failure, the mount vnode list lock is held but the vnode interlock is
6304  * not, and the procedure may have yielded.
6305  */
6306 static bool
6307 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6308     struct vnode *vp)
6309 {
6310 
6311 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6312 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6313 	    ("%s: bad marker", __func__));
6314 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6315 	    ("%s: inappropriate vnode", __func__));
6316 	ASSERT_VI_UNLOCKED(vp, __func__);
6317 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6318 
6319 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6320 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6321 
6322 	/*
6323 	 * Note we may be racing against vdrop which transitioned the hold
6324 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6325 	 * if we are the only user after we get the interlock we will just
6326 	 * vdrop.
6327 	 */
6328 	vhold(vp);
6329 	mtx_unlock(&mp->mnt_listmtx);
6330 	VI_LOCK(vp);
6331 	if (VN_IS_DOOMED(vp)) {
6332 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6333 		goto out_lost;
6334 	}
6335 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6336 	/*
6337 	 * There is nothing to do if we are the last user.
6338 	 */
6339 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6340 		goto out_lost;
6341 	mtx_lock(&mp->mnt_listmtx);
6342 	return (true);
6343 out_lost:
6344 	vdropl(vp);
6345 	maybe_yield();
6346 	mtx_lock(&mp->mnt_listmtx);
6347 	return (false);
6348 }
6349 
6350 static struct vnode *
6351 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6352     void *cbarg)
6353 {
6354 	struct vnode *vp;
6355 
6356 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6357 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6358 restart:
6359 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6360 	while (vp != NULL) {
6361 		if (vp->v_type == VMARKER) {
6362 			vp = TAILQ_NEXT(vp, v_lazylist);
6363 			continue;
6364 		}
6365 		/*
6366 		 * See if we want to process the vnode. Note we may encounter a
6367 		 * long string of vnodes we don't care about and hog the list
6368 		 * as a result. Check for it and requeue the marker.
6369 		 */
6370 		VNPASS(!VN_IS_DOOMED(vp), vp);
6371 		if (!cb(vp, cbarg)) {
6372 			if (!should_yield()) {
6373 				vp = TAILQ_NEXT(vp, v_lazylist);
6374 				continue;
6375 			}
6376 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6377 			    v_lazylist);
6378 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6379 			    v_lazylist);
6380 			mtx_unlock(&mp->mnt_listmtx);
6381 			kern_yield(PRI_USER);
6382 			mtx_lock(&mp->mnt_listmtx);
6383 			goto restart;
6384 		}
6385 		/*
6386 		 * Try-lock because this is the wrong lock order.
6387 		 */
6388 		if (!VI_TRYLOCK(vp) &&
6389 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6390 			goto restart;
6391 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6392 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6393 		    ("alien vnode on the lazy list %p %p", vp, mp));
6394 		VNPASS(vp->v_mount == mp, vp);
6395 		VNPASS(!VN_IS_DOOMED(vp), vp);
6396 		break;
6397 	}
6398 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6399 
6400 	/* Check if we are done */
6401 	if (vp == NULL) {
6402 		mtx_unlock(&mp->mnt_listmtx);
6403 		mnt_vnode_markerfree_lazy(mvp, mp);
6404 		return (NULL);
6405 	}
6406 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6407 	mtx_unlock(&mp->mnt_listmtx);
6408 	ASSERT_VI_LOCKED(vp, "lazy iter");
6409 	return (vp);
6410 }
6411 
6412 struct vnode *
6413 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6414     void *cbarg)
6415 {
6416 
6417 	if (should_yield())
6418 		kern_yield(PRI_USER);
6419 	mtx_lock(&mp->mnt_listmtx);
6420 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6421 }
6422 
6423 struct vnode *
6424 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6425     void *cbarg)
6426 {
6427 	struct vnode *vp;
6428 
6429 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6430 		return (NULL);
6431 
6432 	*mvp = vn_alloc_marker(mp);
6433 	MNT_ILOCK(mp);
6434 	MNT_REF(mp);
6435 	MNT_IUNLOCK(mp);
6436 
6437 	mtx_lock(&mp->mnt_listmtx);
6438 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6439 	if (vp == NULL) {
6440 		mtx_unlock(&mp->mnt_listmtx);
6441 		mnt_vnode_markerfree_lazy(mvp, mp);
6442 		return (NULL);
6443 	}
6444 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6445 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6446 }
6447 
6448 void
6449 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6450 {
6451 
6452 	if (*mvp == NULL)
6453 		return;
6454 
6455 	mtx_lock(&mp->mnt_listmtx);
6456 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6457 	mtx_unlock(&mp->mnt_listmtx);
6458 	mnt_vnode_markerfree_lazy(mvp, mp);
6459 }
6460 
6461 int
6462 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6463 {
6464 
6465 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6466 		cnp->cn_flags &= ~NOEXECCHECK;
6467 		return (0);
6468 	}
6469 
6470 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6471 }
6472