xref: /freebsd/sys/kern/vfs_subr.c (revision 5bb3134a8c21cb87b30e135ef168483f0333dabb)
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/asan.h>
52 #include <sys/bio.h>
53 #include <sys/buf.h>
54 #include <sys/capsicum.h>
55 #include <sys/condvar.h>
56 #include <sys/conf.h>
57 #include <sys/counter.h>
58 #include <sys/dirent.h>
59 #include <sys/event.h>
60 #include <sys/eventhandler.h>
61 #include <sys/extattr.h>
62 #include <sys/file.h>
63 #include <sys/fcntl.h>
64 #include <sys/jail.h>
65 #include <sys/kdb.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/ktr.h>
69 #include <sys/lockf.h>
70 #include <sys/malloc.h>
71 #include <sys/mount.h>
72 #include <sys/namei.h>
73 #include <sys/pctrie.h>
74 #include <sys/priv.h>
75 #include <sys/reboot.h>
76 #include <sys/refcount.h>
77 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/smr.h>
81 #include <sys/smp.h>
82 #include <sys/stat.h>
83 #include <sys/sysctl.h>
84 #include <sys/syslog.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/watchdog.h>
88 
89 #include <machine/stdarg.h>
90 
91 #include <security/mac/mac_framework.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_kern.h>
100 #include <vm/uma.h>
101 
102 #ifdef DDB
103 #include <ddb/ddb.h>
104 #endif
105 
106 static void	delmntque(struct vnode *vp);
107 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
108 		    int slpflag, int slptimeo);
109 static void	syncer_shutdown(void *arg, int howto);
110 static int	vtryrecycle(struct vnode *vp);
111 static void	v_init_counters(struct vnode *);
112 static void	vn_seqc_init(struct vnode *);
113 static void	vn_seqc_write_end_free(struct vnode *vp);
114 static void	vgonel(struct vnode *);
115 static bool	vhold_recycle_free(struct vnode *);
116 static void	vfs_knllock(void *arg);
117 static void	vfs_knlunlock(void *arg);
118 static void	vfs_knl_assert_lock(void *arg, int what);
119 static void	destroy_vpollinfo(struct vpollinfo *vi);
120 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
121 		    daddr_t startlbn, daddr_t endlbn);
122 static void	vnlru_recalc(void);
123 
124 /*
125  * Number of vnodes in existence.  Increased whenever getnewvnode()
126  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
127  */
128 static u_long __exclusive_cache_line numvnodes;
129 
130 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
131     "Number of vnodes in existence");
132 
133 static counter_u64_t vnodes_created;
134 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
135     "Number of vnodes created by getnewvnode");
136 
137 /*
138  * Conversion tables for conversion from vnode types to inode formats
139  * and back.
140  */
141 enum vtype iftovt_tab[16] = {
142 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
143 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
144 };
145 int vttoif_tab[10] = {
146 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
147 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
148 };
149 
150 /*
151  * List of allocates vnodes in the system.
152  */
153 static TAILQ_HEAD(freelst, vnode) vnode_list;
154 static struct vnode *vnode_list_free_marker;
155 static struct vnode *vnode_list_reclaim_marker;
156 
157 /*
158  * "Free" vnode target.  Free vnodes are rarely completely free, but are
159  * just ones that are cheap to recycle.  Usually they are for files which
160  * have been stat'd but not read; these usually have inode and namecache
161  * data attached to them.  This target is the preferred minimum size of a
162  * sub-cache consisting mostly of such files. The system balances the size
163  * of this sub-cache with its complement to try to prevent either from
164  * thrashing while the other is relatively inactive.  The targets express
165  * a preference for the best balance.
166  *
167  * "Above" this target there are 2 further targets (watermarks) related
168  * to recyling of free vnodes.  In the best-operating case, the cache is
169  * exactly full, the free list has size between vlowat and vhiwat above the
170  * free target, and recycling from it and normal use maintains this state.
171  * Sometimes the free list is below vlowat or even empty, but this state
172  * is even better for immediate use provided the cache is not full.
173  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
174  * ones) to reach one of these states.  The watermarks are currently hard-
175  * coded as 4% and 9% of the available space higher.  These and the default
176  * of 25% for wantfreevnodes are too large if the memory size is large.
177  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
178  * whenever vnlru_proc() becomes active.
179  */
180 static long wantfreevnodes;
181 static long __exclusive_cache_line freevnodes;
182 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
183     &freevnodes, 0, "Number of \"free\" vnodes");
184 static long freevnodes_old;
185 
186 static counter_u64_t recycles_count;
187 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
188     "Number of vnodes recycled to meet vnode cache targets");
189 
190 static counter_u64_t recycles_free_count;
191 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
192     "Number of free vnodes recycled to meet vnode cache targets");
193 
194 static counter_u64_t deferred_inact;
195 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
196     "Number of times inactive processing was deferred");
197 
198 /* To keep more than one thread at a time from running vfs_getnewfsid */
199 static struct mtx mntid_mtx;
200 
201 /*
202  * Lock for any access to the following:
203  *	vnode_list
204  *	numvnodes
205  *	freevnodes
206  */
207 static struct mtx __exclusive_cache_line vnode_list_mtx;
208 
209 /* Publicly exported FS */
210 struct nfs_public nfs_pub;
211 
212 static uma_zone_t buf_trie_zone;
213 static smr_t buf_trie_smr;
214 
215 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
216 static uma_zone_t vnode_zone;
217 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
218 
219 __read_frequently smr_t vfs_smr;
220 
221 /*
222  * The workitem queue.
223  *
224  * It is useful to delay writes of file data and filesystem metadata
225  * for tens of seconds so that quickly created and deleted files need
226  * not waste disk bandwidth being created and removed. To realize this,
227  * we append vnodes to a "workitem" queue. When running with a soft
228  * updates implementation, most pending metadata dependencies should
229  * not wait for more than a few seconds. Thus, mounted on block devices
230  * are delayed only about a half the time that file data is delayed.
231  * Similarly, directory updates are more critical, so are only delayed
232  * about a third the time that file data is delayed. Thus, there are
233  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
234  * one each second (driven off the filesystem syncer process). The
235  * syncer_delayno variable indicates the next queue that is to be processed.
236  * Items that need to be processed soon are placed in this queue:
237  *
238  *	syncer_workitem_pending[syncer_delayno]
239  *
240  * A delay of fifteen seconds is done by placing the request fifteen
241  * entries later in the queue:
242  *
243  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
244  *
245  */
246 static int syncer_delayno;
247 static long syncer_mask;
248 LIST_HEAD(synclist, bufobj);
249 static struct synclist *syncer_workitem_pending;
250 /*
251  * The sync_mtx protects:
252  *	bo->bo_synclist
253  *	sync_vnode_count
254  *	syncer_delayno
255  *	syncer_state
256  *	syncer_workitem_pending
257  *	syncer_worklist_len
258  *	rushjob
259  */
260 static struct mtx sync_mtx;
261 static struct cv sync_wakeup;
262 
263 #define SYNCER_MAXDELAY		32
264 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
265 static int syncdelay = 30;		/* max time to delay syncing data */
266 static int filedelay = 30;		/* time to delay syncing files */
267 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
268     "Time to delay syncing files (in seconds)");
269 static int dirdelay = 29;		/* time to delay syncing directories */
270 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
271     "Time to delay syncing directories (in seconds)");
272 static int metadelay = 28;		/* time to delay syncing metadata */
273 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
274     "Time to delay syncing metadata (in seconds)");
275 static int rushjob;		/* number of slots to run ASAP */
276 static int stat_rush_requests;	/* number of times I/O speeded up */
277 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
278     "Number of times I/O speeded up (rush requests)");
279 
280 #define	VDBATCH_SIZE 8
281 struct vdbatch {
282 	u_int index;
283 	long freevnodes;
284 	struct mtx lock;
285 	struct vnode *tab[VDBATCH_SIZE];
286 };
287 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
288 
289 static void	vdbatch_dequeue(struct vnode *vp);
290 
291 /*
292  * When shutting down the syncer, run it at four times normal speed.
293  */
294 #define SYNCER_SHUTDOWN_SPEEDUP		4
295 static int sync_vnode_count;
296 static int syncer_worklist_len;
297 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
298     syncer_state;
299 
300 /* Target for maximum number of vnodes. */
301 u_long desiredvnodes;
302 static u_long gapvnodes;		/* gap between wanted and desired */
303 static u_long vhiwat;		/* enough extras after expansion */
304 static u_long vlowat;		/* minimal extras before expansion */
305 static u_long vstir;		/* nonzero to stir non-free vnodes */
306 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
307 
308 static u_long vnlru_read_freevnodes(void);
309 
310 /*
311  * Note that no attempt is made to sanitize these parameters.
312  */
313 static int
314 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
315 {
316 	u_long val;
317 	int error;
318 
319 	val = desiredvnodes;
320 	error = sysctl_handle_long(oidp, &val, 0, req);
321 	if (error != 0 || req->newptr == NULL)
322 		return (error);
323 
324 	if (val == desiredvnodes)
325 		return (0);
326 	mtx_lock(&vnode_list_mtx);
327 	desiredvnodes = val;
328 	wantfreevnodes = desiredvnodes / 4;
329 	vnlru_recalc();
330 	mtx_unlock(&vnode_list_mtx);
331 	/*
332 	 * XXX There is no protection against multiple threads changing
333 	 * desiredvnodes at the same time. Locking above only helps vnlru and
334 	 * getnewvnode.
335 	 */
336 	vfs_hash_changesize(desiredvnodes);
337 	cache_changesize(desiredvnodes);
338 	return (0);
339 }
340 
341 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
342     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
343     "LU", "Target for maximum number of vnodes");
344 
345 static int
346 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
347 {
348 	u_long val;
349 	int error;
350 
351 	val = wantfreevnodes;
352 	error = sysctl_handle_long(oidp, &val, 0, req);
353 	if (error != 0 || req->newptr == NULL)
354 		return (error);
355 
356 	if (val == wantfreevnodes)
357 		return (0);
358 	mtx_lock(&vnode_list_mtx);
359 	wantfreevnodes = val;
360 	vnlru_recalc();
361 	mtx_unlock(&vnode_list_mtx);
362 	return (0);
363 }
364 
365 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
366     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
367     "LU", "Target for minimum number of \"free\" vnodes");
368 
369 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
370     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
371 static int vnlru_nowhere;
372 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
373     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
374 
375 static int
376 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
377 {
378 	struct vnode *vp;
379 	struct nameidata nd;
380 	char *buf;
381 	unsigned long ndflags;
382 	int error;
383 
384 	if (req->newptr == NULL)
385 		return (EINVAL);
386 	if (req->newlen >= PATH_MAX)
387 		return (E2BIG);
388 
389 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
390 	error = SYSCTL_IN(req, buf, req->newlen);
391 	if (error != 0)
392 		goto out;
393 
394 	buf[req->newlen] = '\0';
395 
396 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
397 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
398 	if ((error = namei(&nd)) != 0)
399 		goto out;
400 	vp = nd.ni_vp;
401 
402 	if (VN_IS_DOOMED(vp)) {
403 		/*
404 		 * This vnode is being recycled.  Return != 0 to let the caller
405 		 * know that the sysctl had no effect.  Return EAGAIN because a
406 		 * subsequent call will likely succeed (since namei will create
407 		 * a new vnode if necessary)
408 		 */
409 		error = EAGAIN;
410 		goto putvnode;
411 	}
412 
413 	counter_u64_add(recycles_count, 1);
414 	vgone(vp);
415 putvnode:
416 	NDFREE(&nd, 0);
417 out:
418 	free(buf, M_TEMP);
419 	return (error);
420 }
421 
422 static int
423 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
424 {
425 	struct thread *td = curthread;
426 	struct vnode *vp;
427 	struct file *fp;
428 	int error;
429 	int fd;
430 
431 	if (req->newptr == NULL)
432 		return (EBADF);
433 
434         error = sysctl_handle_int(oidp, &fd, 0, req);
435         if (error != 0)
436                 return (error);
437 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
438 	if (error != 0)
439 		return (error);
440 	vp = fp->f_vnode;
441 
442 	error = vn_lock(vp, LK_EXCLUSIVE);
443 	if (error != 0)
444 		goto drop;
445 
446 	counter_u64_add(recycles_count, 1);
447 	vgone(vp);
448 	VOP_UNLOCK(vp);
449 drop:
450 	fdrop(fp, td);
451 	return (error);
452 }
453 
454 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
455     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
456     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
457 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
458     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
459     sysctl_ftry_reclaim_vnode, "I",
460     "Try to reclaim a vnode by its file descriptor");
461 
462 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
463 static int vnsz2log;
464 
465 /*
466  * Support for the bufobj clean & dirty pctrie.
467  */
468 static void *
469 buf_trie_alloc(struct pctrie *ptree)
470 {
471 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
472 }
473 
474 static void
475 buf_trie_free(struct pctrie *ptree, void *node)
476 {
477 	uma_zfree_smr(buf_trie_zone, node);
478 }
479 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
480     buf_trie_smr);
481 
482 /*
483  * Initialize the vnode management data structures.
484  *
485  * Reevaluate the following cap on the number of vnodes after the physical
486  * memory size exceeds 512GB.  In the limit, as the physical memory size
487  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
488  */
489 #ifndef	MAXVNODES_MAX
490 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
491 #endif
492 
493 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
494 
495 static struct vnode *
496 vn_alloc_marker(struct mount *mp)
497 {
498 	struct vnode *vp;
499 
500 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
501 	vp->v_type = VMARKER;
502 	vp->v_mount = mp;
503 
504 	return (vp);
505 }
506 
507 static void
508 vn_free_marker(struct vnode *vp)
509 {
510 
511 	MPASS(vp->v_type == VMARKER);
512 	free(vp, M_VNODE_MARKER);
513 }
514 
515 #ifdef KASAN
516 static int
517 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
518 {
519 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
520 	return (0);
521 }
522 
523 static void
524 vnode_dtor(void *mem, int size, void *arg __unused)
525 {
526 	size_t end1, end2, off1, off2;
527 
528 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
529 	    offsetof(struct vnode, v_dbatchcpu),
530 	    "KASAN marks require updating");
531 
532 	off1 = offsetof(struct vnode, v_vnodelist);
533 	off2 = offsetof(struct vnode, v_dbatchcpu);
534 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
535 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
536 
537 	/*
538 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
539 	 * after the vnode has been freed.  Try to get some KASAN coverage by
540 	 * marking everything except those two fields as invalid.  Because
541 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
542 	 * the same 8-byte aligned word must also be marked valid.
543 	 */
544 
545 	/* Handle the area from the start until v_vnodelist... */
546 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
547 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
548 
549 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
550 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
551 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
552 	if (off2 > off1)
553 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
554 		    off2 - off1, KASAN_UMA_FREED);
555 
556 	/* ... and finally the area from v_dbatchcpu to the end. */
557 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
558 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
559 	    KASAN_UMA_FREED);
560 }
561 #endif /* KASAN */
562 
563 /*
564  * Initialize a vnode as it first enters the zone.
565  */
566 static int
567 vnode_init(void *mem, int size, int flags)
568 {
569 	struct vnode *vp;
570 
571 	vp = mem;
572 	bzero(vp, size);
573 	/*
574 	 * Setup locks.
575 	 */
576 	vp->v_vnlock = &vp->v_lock;
577 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
578 	/*
579 	 * By default, don't allow shared locks unless filesystems opt-in.
580 	 */
581 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
582 	    LK_NOSHARE | LK_IS_VNODE);
583 	/*
584 	 * Initialize bufobj.
585 	 */
586 	bufobj_init(&vp->v_bufobj, vp);
587 	/*
588 	 * Initialize namecache.
589 	 */
590 	cache_vnode_init(vp);
591 	/*
592 	 * Initialize rangelocks.
593 	 */
594 	rangelock_init(&vp->v_rl);
595 
596 	vp->v_dbatchcpu = NOCPU;
597 
598 	/*
599 	 * Check vhold_recycle_free for an explanation.
600 	 */
601 	vp->v_holdcnt = VHOLD_NO_SMR;
602 	vp->v_type = VNON;
603 	mtx_lock(&vnode_list_mtx);
604 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
605 	mtx_unlock(&vnode_list_mtx);
606 	return (0);
607 }
608 
609 /*
610  * Free a vnode when it is cleared from the zone.
611  */
612 static void
613 vnode_fini(void *mem, int size)
614 {
615 	struct vnode *vp;
616 	struct bufobj *bo;
617 
618 	vp = mem;
619 	vdbatch_dequeue(vp);
620 	mtx_lock(&vnode_list_mtx);
621 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
622 	mtx_unlock(&vnode_list_mtx);
623 	rangelock_destroy(&vp->v_rl);
624 	lockdestroy(vp->v_vnlock);
625 	mtx_destroy(&vp->v_interlock);
626 	bo = &vp->v_bufobj;
627 	rw_destroy(BO_LOCKPTR(bo));
628 
629 	kasan_mark(mem, size, size, 0);
630 }
631 
632 /*
633  * Provide the size of NFS nclnode and NFS fh for calculation of the
634  * vnode memory consumption.  The size is specified directly to
635  * eliminate dependency on NFS-private header.
636  *
637  * Other filesystems may use bigger or smaller (like UFS and ZFS)
638  * private inode data, but the NFS-based estimation is ample enough.
639  * Still, we care about differences in the size between 64- and 32-bit
640  * platforms.
641  *
642  * Namecache structure size is heuristically
643  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
644  */
645 #ifdef _LP64
646 #define	NFS_NCLNODE_SZ	(528 + 64)
647 #define	NC_SZ		148
648 #else
649 #define	NFS_NCLNODE_SZ	(360 + 32)
650 #define	NC_SZ		92
651 #endif
652 
653 static void
654 vntblinit(void *dummy __unused)
655 {
656 	struct vdbatch *vd;
657 	uma_ctor ctor;
658 	uma_dtor dtor;
659 	int cpu, physvnodes, virtvnodes;
660 	u_int i;
661 
662 	/*
663 	 * Desiredvnodes is a function of the physical memory size and the
664 	 * kernel's heap size.  Generally speaking, it scales with the
665 	 * physical memory size.  The ratio of desiredvnodes to the physical
666 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
667 	 * Thereafter, the
668 	 * marginal ratio of desiredvnodes to the physical memory size is
669 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
670 	 * size.  The memory required by desiredvnodes vnodes and vm objects
671 	 * must not exceed 1/10th of the kernel's heap size.
672 	 */
673 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
674 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
675 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
676 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
677 	desiredvnodes = min(physvnodes, virtvnodes);
678 	if (desiredvnodes > MAXVNODES_MAX) {
679 		if (bootverbose)
680 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
681 			    desiredvnodes, MAXVNODES_MAX);
682 		desiredvnodes = MAXVNODES_MAX;
683 	}
684 	wantfreevnodes = desiredvnodes / 4;
685 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
686 	TAILQ_INIT(&vnode_list);
687 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
688 	/*
689 	 * The lock is taken to appease WITNESS.
690 	 */
691 	mtx_lock(&vnode_list_mtx);
692 	vnlru_recalc();
693 	mtx_unlock(&vnode_list_mtx);
694 	vnode_list_free_marker = vn_alloc_marker(NULL);
695 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
696 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
697 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
698 
699 #ifdef KASAN
700 	ctor = vnode_ctor;
701 	dtor = vnode_dtor;
702 #else
703 	ctor = NULL;
704 	dtor = NULL;
705 #endif
706 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
707 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
708 	uma_zone_set_smr(vnode_zone, vfs_smr);
709 
710 	/*
711 	 * Preallocate enough nodes to support one-per buf so that
712 	 * we can not fail an insert.  reassignbuf() callers can not
713 	 * tolerate the insertion failure.
714 	 */
715 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
716 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
717 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
718 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
719 	uma_prealloc(buf_trie_zone, nbuf);
720 
721 	vnodes_created = counter_u64_alloc(M_WAITOK);
722 	recycles_count = counter_u64_alloc(M_WAITOK);
723 	recycles_free_count = counter_u64_alloc(M_WAITOK);
724 	deferred_inact = counter_u64_alloc(M_WAITOK);
725 
726 	/*
727 	 * Initialize the filesystem syncer.
728 	 */
729 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
730 	    &syncer_mask);
731 	syncer_maxdelay = syncer_mask + 1;
732 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
733 	cv_init(&sync_wakeup, "syncer");
734 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
735 		vnsz2log++;
736 	vnsz2log--;
737 
738 	CPU_FOREACH(cpu) {
739 		vd = DPCPU_ID_PTR((cpu), vd);
740 		bzero(vd, sizeof(*vd));
741 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
742 	}
743 }
744 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
745 
746 /*
747  * Mark a mount point as busy. Used to synchronize access and to delay
748  * unmounting. Eventually, mountlist_mtx is not released on failure.
749  *
750  * vfs_busy() is a custom lock, it can block the caller.
751  * vfs_busy() only sleeps if the unmount is active on the mount point.
752  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
753  * vnode belonging to mp.
754  *
755  * Lookup uses vfs_busy() to traverse mount points.
756  * root fs			var fs
757  * / vnode lock		A	/ vnode lock (/var)		D
758  * /var vnode lock	B	/log vnode lock(/var/log)	E
759  * vfs_busy lock	C	vfs_busy lock			F
760  *
761  * Within each file system, the lock order is C->A->B and F->D->E.
762  *
763  * When traversing across mounts, the system follows that lock order:
764  *
765  *        C->A->B
766  *              |
767  *              +->F->D->E
768  *
769  * The lookup() process for namei("/var") illustrates the process:
770  *  VOP_LOOKUP() obtains B while A is held
771  *  vfs_busy() obtains a shared lock on F while A and B are held
772  *  vput() releases lock on B
773  *  vput() releases lock on A
774  *  VFS_ROOT() obtains lock on D while shared lock on F is held
775  *  vfs_unbusy() releases shared lock on F
776  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
777  *    Attempt to lock A (instead of vp_crossmp) while D is held would
778  *    violate the global order, causing deadlocks.
779  *
780  * dounmount() locks B while F is drained.
781  */
782 int
783 vfs_busy(struct mount *mp, int flags)
784 {
785 	struct mount_pcpu *mpcpu;
786 
787 	MPASS((flags & ~MBF_MASK) == 0);
788 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
789 
790 	if (vfs_op_thread_enter(mp, mpcpu)) {
791 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
792 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
793 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
794 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
795 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
796 		vfs_op_thread_exit(mp, mpcpu);
797 		if (flags & MBF_MNTLSTLOCK)
798 			mtx_unlock(&mountlist_mtx);
799 		return (0);
800 	}
801 
802 	MNT_ILOCK(mp);
803 	vfs_assert_mount_counters(mp);
804 	MNT_REF(mp);
805 	/*
806 	 * If mount point is currently being unmounted, sleep until the
807 	 * mount point fate is decided.  If thread doing the unmounting fails,
808 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
809 	 * that this mount point has survived the unmount attempt and vfs_busy
810 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
811 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
812 	 * about to be really destroyed.  vfs_busy needs to release its
813 	 * reference on the mount point in this case and return with ENOENT,
814 	 * telling the caller that mount mount it tried to busy is no longer
815 	 * valid.
816 	 */
817 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
818 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
819 		    ("%s: non-empty upper mount list with pending unmount",
820 		    __func__));
821 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
822 			MNT_REL(mp);
823 			MNT_IUNLOCK(mp);
824 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
825 			    __func__);
826 			return (ENOENT);
827 		}
828 		if (flags & MBF_MNTLSTLOCK)
829 			mtx_unlock(&mountlist_mtx);
830 		mp->mnt_kern_flag |= MNTK_MWAIT;
831 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
832 		if (flags & MBF_MNTLSTLOCK)
833 			mtx_lock(&mountlist_mtx);
834 		MNT_ILOCK(mp);
835 	}
836 	if (flags & MBF_MNTLSTLOCK)
837 		mtx_unlock(&mountlist_mtx);
838 	mp->mnt_lockref++;
839 	MNT_IUNLOCK(mp);
840 	return (0);
841 }
842 
843 /*
844  * Free a busy filesystem.
845  */
846 void
847 vfs_unbusy(struct mount *mp)
848 {
849 	struct mount_pcpu *mpcpu;
850 	int c;
851 
852 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
853 
854 	if (vfs_op_thread_enter(mp, mpcpu)) {
855 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
856 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
857 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
858 		vfs_op_thread_exit(mp, mpcpu);
859 		return;
860 	}
861 
862 	MNT_ILOCK(mp);
863 	vfs_assert_mount_counters(mp);
864 	MNT_REL(mp);
865 	c = --mp->mnt_lockref;
866 	if (mp->mnt_vfs_ops == 0) {
867 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
868 		MNT_IUNLOCK(mp);
869 		return;
870 	}
871 	if (c < 0)
872 		vfs_dump_mount_counters(mp);
873 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
874 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
875 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
876 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
877 		wakeup(&mp->mnt_lockref);
878 	}
879 	MNT_IUNLOCK(mp);
880 }
881 
882 /*
883  * Lookup a mount point by filesystem identifier.
884  */
885 struct mount *
886 vfs_getvfs(fsid_t *fsid)
887 {
888 	struct mount *mp;
889 
890 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
891 	mtx_lock(&mountlist_mtx);
892 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
893 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
894 			vfs_ref(mp);
895 			mtx_unlock(&mountlist_mtx);
896 			return (mp);
897 		}
898 	}
899 	mtx_unlock(&mountlist_mtx);
900 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
901 	return ((struct mount *) 0);
902 }
903 
904 /*
905  * Lookup a mount point by filesystem identifier, busying it before
906  * returning.
907  *
908  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
909  * cache for popular filesystem identifiers.  The cache is lockess, using
910  * the fact that struct mount's are never freed.  In worst case we may
911  * get pointer to unmounted or even different filesystem, so we have to
912  * check what we got, and go slow way if so.
913  */
914 struct mount *
915 vfs_busyfs(fsid_t *fsid)
916 {
917 #define	FSID_CACHE_SIZE	256
918 	typedef struct mount * volatile vmp_t;
919 	static vmp_t cache[FSID_CACHE_SIZE];
920 	struct mount *mp;
921 	int error;
922 	uint32_t hash;
923 
924 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
925 	hash = fsid->val[0] ^ fsid->val[1];
926 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
927 	mp = cache[hash];
928 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
929 		goto slow;
930 	if (vfs_busy(mp, 0) != 0) {
931 		cache[hash] = NULL;
932 		goto slow;
933 	}
934 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
935 		return (mp);
936 	else
937 	    vfs_unbusy(mp);
938 
939 slow:
940 	mtx_lock(&mountlist_mtx);
941 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
942 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
943 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
944 			if (error) {
945 				cache[hash] = NULL;
946 				mtx_unlock(&mountlist_mtx);
947 				return (NULL);
948 			}
949 			cache[hash] = mp;
950 			return (mp);
951 		}
952 	}
953 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
954 	mtx_unlock(&mountlist_mtx);
955 	return ((struct mount *) 0);
956 }
957 
958 /*
959  * Check if a user can access privileged mount options.
960  */
961 int
962 vfs_suser(struct mount *mp, struct thread *td)
963 {
964 	int error;
965 
966 	if (jailed(td->td_ucred)) {
967 		/*
968 		 * If the jail of the calling thread lacks permission for
969 		 * this type of file system, deny immediately.
970 		 */
971 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
972 			return (EPERM);
973 
974 		/*
975 		 * If the file system was mounted outside the jail of the
976 		 * calling thread, deny immediately.
977 		 */
978 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
979 			return (EPERM);
980 	}
981 
982 	/*
983 	 * If file system supports delegated administration, we don't check
984 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
985 	 * by the file system itself.
986 	 * If this is not the user that did original mount, we check for
987 	 * the PRIV_VFS_MOUNT_OWNER privilege.
988 	 */
989 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
990 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
991 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
992 			return (error);
993 	}
994 	return (0);
995 }
996 
997 /*
998  * Get a new unique fsid.  Try to make its val[0] unique, since this value
999  * will be used to create fake device numbers for stat().  Also try (but
1000  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1001  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1002  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1003  *
1004  * Keep in mind that several mounts may be running in parallel.  Starting
1005  * the search one past where the previous search terminated is both a
1006  * micro-optimization and a defense against returning the same fsid to
1007  * different mounts.
1008  */
1009 void
1010 vfs_getnewfsid(struct mount *mp)
1011 {
1012 	static uint16_t mntid_base;
1013 	struct mount *nmp;
1014 	fsid_t tfsid;
1015 	int mtype;
1016 
1017 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1018 	mtx_lock(&mntid_mtx);
1019 	mtype = mp->mnt_vfc->vfc_typenum;
1020 	tfsid.val[1] = mtype;
1021 	mtype = (mtype & 0xFF) << 24;
1022 	for (;;) {
1023 		tfsid.val[0] = makedev(255,
1024 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1025 		mntid_base++;
1026 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1027 			break;
1028 		vfs_rel(nmp);
1029 	}
1030 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1031 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1032 	mtx_unlock(&mntid_mtx);
1033 }
1034 
1035 /*
1036  * Knob to control the precision of file timestamps:
1037  *
1038  *   0 = seconds only; nanoseconds zeroed.
1039  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1040  *   2 = seconds and nanoseconds, truncated to microseconds.
1041  * >=3 = seconds and nanoseconds, maximum precision.
1042  */
1043 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1044 
1045 static int timestamp_precision = TSP_USEC;
1046 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1047     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1048     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1049     "3+: sec + ns (max. precision))");
1050 
1051 /*
1052  * Get a current timestamp.
1053  */
1054 void
1055 vfs_timestamp(struct timespec *tsp)
1056 {
1057 	struct timeval tv;
1058 
1059 	switch (timestamp_precision) {
1060 	case TSP_SEC:
1061 		tsp->tv_sec = time_second;
1062 		tsp->tv_nsec = 0;
1063 		break;
1064 	case TSP_HZ:
1065 		getnanotime(tsp);
1066 		break;
1067 	case TSP_USEC:
1068 		microtime(&tv);
1069 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1070 		break;
1071 	case TSP_NSEC:
1072 	default:
1073 		nanotime(tsp);
1074 		break;
1075 	}
1076 }
1077 
1078 /*
1079  * Set vnode attributes to VNOVAL
1080  */
1081 void
1082 vattr_null(struct vattr *vap)
1083 {
1084 
1085 	vap->va_type = VNON;
1086 	vap->va_size = VNOVAL;
1087 	vap->va_bytes = VNOVAL;
1088 	vap->va_mode = VNOVAL;
1089 	vap->va_nlink = VNOVAL;
1090 	vap->va_uid = VNOVAL;
1091 	vap->va_gid = VNOVAL;
1092 	vap->va_fsid = VNOVAL;
1093 	vap->va_fileid = VNOVAL;
1094 	vap->va_blocksize = VNOVAL;
1095 	vap->va_rdev = VNOVAL;
1096 	vap->va_atime.tv_sec = VNOVAL;
1097 	vap->va_atime.tv_nsec = VNOVAL;
1098 	vap->va_mtime.tv_sec = VNOVAL;
1099 	vap->va_mtime.tv_nsec = VNOVAL;
1100 	vap->va_ctime.tv_sec = VNOVAL;
1101 	vap->va_ctime.tv_nsec = VNOVAL;
1102 	vap->va_birthtime.tv_sec = VNOVAL;
1103 	vap->va_birthtime.tv_nsec = VNOVAL;
1104 	vap->va_flags = VNOVAL;
1105 	vap->va_gen = VNOVAL;
1106 	vap->va_vaflags = 0;
1107 }
1108 
1109 /*
1110  * Try to reduce the total number of vnodes.
1111  *
1112  * This routine (and its user) are buggy in at least the following ways:
1113  * - all parameters were picked years ago when RAM sizes were significantly
1114  *   smaller
1115  * - it can pick vnodes based on pages used by the vm object, but filesystems
1116  *   like ZFS don't use it making the pick broken
1117  * - since ZFS has its own aging policy it gets partially combated by this one
1118  * - a dedicated method should be provided for filesystems to let them decide
1119  *   whether the vnode should be recycled
1120  *
1121  * This routine is called when we have too many vnodes.  It attempts
1122  * to free <count> vnodes and will potentially free vnodes that still
1123  * have VM backing store (VM backing store is typically the cause
1124  * of a vnode blowout so we want to do this).  Therefore, this operation
1125  * is not considered cheap.
1126  *
1127  * A number of conditions may prevent a vnode from being reclaimed.
1128  * the buffer cache may have references on the vnode, a directory
1129  * vnode may still have references due to the namei cache representing
1130  * underlying files, or the vnode may be in active use.   It is not
1131  * desirable to reuse such vnodes.  These conditions may cause the
1132  * number of vnodes to reach some minimum value regardless of what
1133  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1134  *
1135  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1136  * 			 entries if this argument is strue
1137  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1138  *			 pages.
1139  * @param target	 How many vnodes to reclaim.
1140  * @return		 The number of vnodes that were reclaimed.
1141  */
1142 static int
1143 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1144 {
1145 	struct vnode *vp, *mvp;
1146 	struct mount *mp;
1147 	struct vm_object *object;
1148 	u_long done;
1149 	bool retried;
1150 
1151 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1152 
1153 	retried = false;
1154 	done = 0;
1155 
1156 	mvp = vnode_list_reclaim_marker;
1157 restart:
1158 	vp = mvp;
1159 	while (done < target) {
1160 		vp = TAILQ_NEXT(vp, v_vnodelist);
1161 		if (__predict_false(vp == NULL))
1162 			break;
1163 
1164 		if (__predict_false(vp->v_type == VMARKER))
1165 			continue;
1166 
1167 		/*
1168 		 * If it's been deconstructed already, it's still
1169 		 * referenced, or it exceeds the trigger, skip it.
1170 		 * Also skip free vnodes.  We are trying to make space
1171 		 * to expand the free list, not reduce it.
1172 		 */
1173 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1174 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1175 			goto next_iter;
1176 
1177 		if (vp->v_type == VBAD || vp->v_type == VNON)
1178 			goto next_iter;
1179 
1180 		object = atomic_load_ptr(&vp->v_object);
1181 		if (object == NULL || object->resident_page_count > trigger) {
1182 			goto next_iter;
1183 		}
1184 
1185 		/*
1186 		 * Handle races against vnode allocation. Filesystems lock the
1187 		 * vnode some time after it gets returned from getnewvnode,
1188 		 * despite type and hold count being manipulated earlier.
1189 		 * Resorting to checking v_mount restores guarantees present
1190 		 * before the global list was reworked to contain all vnodes.
1191 		 */
1192 		if (!VI_TRYLOCK(vp))
1193 			goto next_iter;
1194 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1195 			VI_UNLOCK(vp);
1196 			goto next_iter;
1197 		}
1198 		if (vp->v_mount == NULL) {
1199 			VI_UNLOCK(vp);
1200 			goto next_iter;
1201 		}
1202 		vholdl(vp);
1203 		VI_UNLOCK(vp);
1204 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1205 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1206 		mtx_unlock(&vnode_list_mtx);
1207 
1208 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1209 			vdrop(vp);
1210 			goto next_iter_unlocked;
1211 		}
1212 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1213 			vdrop(vp);
1214 			vn_finished_write(mp);
1215 			goto next_iter_unlocked;
1216 		}
1217 
1218 		VI_LOCK(vp);
1219 		if (vp->v_usecount > 0 ||
1220 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1221 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1222 		    vp->v_object->resident_page_count > trigger)) {
1223 			VOP_UNLOCK(vp);
1224 			vdropl(vp);
1225 			vn_finished_write(mp);
1226 			goto next_iter_unlocked;
1227 		}
1228 		counter_u64_add(recycles_count, 1);
1229 		vgonel(vp);
1230 		VOP_UNLOCK(vp);
1231 		vdropl(vp);
1232 		vn_finished_write(mp);
1233 		done++;
1234 next_iter_unlocked:
1235 		if (should_yield())
1236 			kern_yield(PRI_USER);
1237 		mtx_lock(&vnode_list_mtx);
1238 		goto restart;
1239 next_iter:
1240 		MPASS(vp->v_type != VMARKER);
1241 		if (!should_yield())
1242 			continue;
1243 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1244 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1245 		mtx_unlock(&vnode_list_mtx);
1246 		kern_yield(PRI_USER);
1247 		mtx_lock(&vnode_list_mtx);
1248 		goto restart;
1249 	}
1250 	if (done == 0 && !retried) {
1251 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1252 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1253 		retried = true;
1254 		goto restart;
1255 	}
1256 	return (done);
1257 }
1258 
1259 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1260 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1261     0,
1262     "limit on vnode free requests per call to the vnlru_free routine");
1263 
1264 /*
1265  * Attempt to reduce the free list by the requested amount.
1266  */
1267 static int
1268 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1269 {
1270 	struct vnode *vp;
1271 	struct mount *mp;
1272 	int ocount;
1273 
1274 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1275 	if (count > max_vnlru_free)
1276 		count = max_vnlru_free;
1277 	ocount = count;
1278 	vp = mvp;
1279 	for (;;) {
1280 		if (count == 0) {
1281 			break;
1282 		}
1283 		vp = TAILQ_NEXT(vp, v_vnodelist);
1284 		if (__predict_false(vp == NULL)) {
1285 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1286 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1287 			break;
1288 		}
1289 		if (__predict_false(vp->v_type == VMARKER))
1290 			continue;
1291 		if (vp->v_holdcnt > 0)
1292 			continue;
1293 		/*
1294 		 * Don't recycle if our vnode is from different type
1295 		 * of mount point.  Note that mp is type-safe, the
1296 		 * check does not reach unmapped address even if
1297 		 * vnode is reclaimed.
1298 		 */
1299 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1300 		    mp->mnt_op != mnt_op) {
1301 			continue;
1302 		}
1303 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1304 			continue;
1305 		}
1306 		if (!vhold_recycle_free(vp))
1307 			continue;
1308 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1309 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1310 		mtx_unlock(&vnode_list_mtx);
1311 		if (vtryrecycle(vp) == 0)
1312 			count--;
1313 		mtx_lock(&vnode_list_mtx);
1314 		vp = mvp;
1315 	}
1316 	return (ocount - count);
1317 }
1318 
1319 static int
1320 vnlru_free_locked(int count)
1321 {
1322 
1323 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1324 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1325 }
1326 
1327 void
1328 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1329 {
1330 
1331 	MPASS(mnt_op != NULL);
1332 	MPASS(mvp != NULL);
1333 	VNPASS(mvp->v_type == VMARKER, mvp);
1334 	mtx_lock(&vnode_list_mtx);
1335 	vnlru_free_impl(count, mnt_op, mvp);
1336 	mtx_unlock(&vnode_list_mtx);
1337 }
1338 
1339 struct vnode *
1340 vnlru_alloc_marker(void)
1341 {
1342 	struct vnode *mvp;
1343 
1344 	mvp = vn_alloc_marker(NULL);
1345 	mtx_lock(&vnode_list_mtx);
1346 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1347 	mtx_unlock(&vnode_list_mtx);
1348 	return (mvp);
1349 }
1350 
1351 void
1352 vnlru_free_marker(struct vnode *mvp)
1353 {
1354 	mtx_lock(&vnode_list_mtx);
1355 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1356 	mtx_unlock(&vnode_list_mtx);
1357 	vn_free_marker(mvp);
1358 }
1359 
1360 static void
1361 vnlru_recalc(void)
1362 {
1363 
1364 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1365 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1366 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1367 	vlowat = vhiwat / 2;
1368 }
1369 
1370 /*
1371  * Attempt to recycle vnodes in a context that is always safe to block.
1372  * Calling vlrurecycle() from the bowels of filesystem code has some
1373  * interesting deadlock problems.
1374  */
1375 static struct proc *vnlruproc;
1376 static int vnlruproc_sig;
1377 
1378 /*
1379  * The main freevnodes counter is only updated when threads requeue their vnode
1380  * batches. CPUs are conditionally walked to compute a more accurate total.
1381  *
1382  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1383  * at any given moment can still exceed slop, but it should not be by significant
1384  * margin in practice.
1385  */
1386 #define VNLRU_FREEVNODES_SLOP 128
1387 
1388 static __inline void
1389 vfs_freevnodes_inc(void)
1390 {
1391 	struct vdbatch *vd;
1392 
1393 	critical_enter();
1394 	vd = DPCPU_PTR(vd);
1395 	vd->freevnodes++;
1396 	critical_exit();
1397 }
1398 
1399 static __inline void
1400 vfs_freevnodes_dec(void)
1401 {
1402 	struct vdbatch *vd;
1403 
1404 	critical_enter();
1405 	vd = DPCPU_PTR(vd);
1406 	vd->freevnodes--;
1407 	critical_exit();
1408 }
1409 
1410 static u_long
1411 vnlru_read_freevnodes(void)
1412 {
1413 	struct vdbatch *vd;
1414 	long slop;
1415 	int cpu;
1416 
1417 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1418 	if (freevnodes > freevnodes_old)
1419 		slop = freevnodes - freevnodes_old;
1420 	else
1421 		slop = freevnodes_old - freevnodes;
1422 	if (slop < VNLRU_FREEVNODES_SLOP)
1423 		return (freevnodes >= 0 ? freevnodes : 0);
1424 	freevnodes_old = freevnodes;
1425 	CPU_FOREACH(cpu) {
1426 		vd = DPCPU_ID_PTR((cpu), vd);
1427 		freevnodes_old += vd->freevnodes;
1428 	}
1429 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1430 }
1431 
1432 static bool
1433 vnlru_under(u_long rnumvnodes, u_long limit)
1434 {
1435 	u_long rfreevnodes, space;
1436 
1437 	if (__predict_false(rnumvnodes > desiredvnodes))
1438 		return (true);
1439 
1440 	space = desiredvnodes - rnumvnodes;
1441 	if (space < limit) {
1442 		rfreevnodes = vnlru_read_freevnodes();
1443 		if (rfreevnodes > wantfreevnodes)
1444 			space += rfreevnodes - wantfreevnodes;
1445 	}
1446 	return (space < limit);
1447 }
1448 
1449 static bool
1450 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1451 {
1452 	long rfreevnodes, space;
1453 
1454 	if (__predict_false(rnumvnodes > desiredvnodes))
1455 		return (true);
1456 
1457 	space = desiredvnodes - rnumvnodes;
1458 	if (space < limit) {
1459 		rfreevnodes = atomic_load_long(&freevnodes);
1460 		if (rfreevnodes > wantfreevnodes)
1461 			space += rfreevnodes - wantfreevnodes;
1462 	}
1463 	return (space < limit);
1464 }
1465 
1466 static void
1467 vnlru_kick(void)
1468 {
1469 
1470 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1471 	if (vnlruproc_sig == 0) {
1472 		vnlruproc_sig = 1;
1473 		wakeup(vnlruproc);
1474 	}
1475 }
1476 
1477 static void
1478 vnlru_proc(void)
1479 {
1480 	u_long rnumvnodes, rfreevnodes, target;
1481 	unsigned long onumvnodes;
1482 	int done, force, trigger, usevnodes;
1483 	bool reclaim_nc_src, want_reread;
1484 
1485 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1486 	    SHUTDOWN_PRI_FIRST);
1487 
1488 	force = 0;
1489 	want_reread = false;
1490 	for (;;) {
1491 		kproc_suspend_check(vnlruproc);
1492 		mtx_lock(&vnode_list_mtx);
1493 		rnumvnodes = atomic_load_long(&numvnodes);
1494 
1495 		if (want_reread) {
1496 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1497 			want_reread = false;
1498 		}
1499 
1500 		/*
1501 		 * If numvnodes is too large (due to desiredvnodes being
1502 		 * adjusted using its sysctl, or emergency growth), first
1503 		 * try to reduce it by discarding from the free list.
1504 		 */
1505 		if (rnumvnodes > desiredvnodes) {
1506 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1507 			rnumvnodes = atomic_load_long(&numvnodes);
1508 		}
1509 		/*
1510 		 * Sleep if the vnode cache is in a good state.  This is
1511 		 * when it is not over-full and has space for about a 4%
1512 		 * or 9% expansion (by growing its size or inexcessively
1513 		 * reducing its free list).  Otherwise, try to reclaim
1514 		 * space for a 10% expansion.
1515 		 */
1516 		if (vstir && force == 0) {
1517 			force = 1;
1518 			vstir = 0;
1519 		}
1520 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1521 			vnlruproc_sig = 0;
1522 			wakeup(&vnlruproc_sig);
1523 			msleep(vnlruproc, &vnode_list_mtx,
1524 			    PVFS|PDROP, "vlruwt", hz);
1525 			continue;
1526 		}
1527 		rfreevnodes = vnlru_read_freevnodes();
1528 
1529 		onumvnodes = rnumvnodes;
1530 		/*
1531 		 * Calculate parameters for recycling.  These are the same
1532 		 * throughout the loop to give some semblance of fairness.
1533 		 * The trigger point is to avoid recycling vnodes with lots
1534 		 * of resident pages.  We aren't trying to free memory; we
1535 		 * are trying to recycle or at least free vnodes.
1536 		 */
1537 		if (rnumvnodes <= desiredvnodes)
1538 			usevnodes = rnumvnodes - rfreevnodes;
1539 		else
1540 			usevnodes = rnumvnodes;
1541 		if (usevnodes <= 0)
1542 			usevnodes = 1;
1543 		/*
1544 		 * The trigger value is is chosen to give a conservatively
1545 		 * large value to ensure that it alone doesn't prevent
1546 		 * making progress.  The value can easily be so large that
1547 		 * it is effectively infinite in some congested and
1548 		 * misconfigured cases, and this is necessary.  Normally
1549 		 * it is about 8 to 100 (pages), which is quite large.
1550 		 */
1551 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1552 		if (force < 2)
1553 			trigger = vsmalltrigger;
1554 		reclaim_nc_src = force >= 3;
1555 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1556 		target = target / 10 + 1;
1557 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1558 		mtx_unlock(&vnode_list_mtx);
1559 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1560 			uma_reclaim(UMA_RECLAIM_DRAIN);
1561 		if (done == 0) {
1562 			if (force == 0 || force == 1) {
1563 				force = 2;
1564 				continue;
1565 			}
1566 			if (force == 2) {
1567 				force = 3;
1568 				continue;
1569 			}
1570 			want_reread = true;
1571 			force = 0;
1572 			vnlru_nowhere++;
1573 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1574 		} else {
1575 			want_reread = true;
1576 			kern_yield(PRI_USER);
1577 		}
1578 	}
1579 }
1580 
1581 static struct kproc_desc vnlru_kp = {
1582 	"vnlru",
1583 	vnlru_proc,
1584 	&vnlruproc
1585 };
1586 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1587     &vnlru_kp);
1588 
1589 /*
1590  * Routines having to do with the management of the vnode table.
1591  */
1592 
1593 /*
1594  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1595  * before we actually vgone().  This function must be called with the vnode
1596  * held to prevent the vnode from being returned to the free list midway
1597  * through vgone().
1598  */
1599 static int
1600 vtryrecycle(struct vnode *vp)
1601 {
1602 	struct mount *vnmp;
1603 
1604 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1605 	VNASSERT(vp->v_holdcnt, vp,
1606 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1607 	/*
1608 	 * This vnode may found and locked via some other list, if so we
1609 	 * can't recycle it yet.
1610 	 */
1611 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1612 		CTR2(KTR_VFS,
1613 		    "%s: impossible to recycle, vp %p lock is already held",
1614 		    __func__, vp);
1615 		vdrop(vp);
1616 		return (EWOULDBLOCK);
1617 	}
1618 	/*
1619 	 * Don't recycle if its filesystem is being suspended.
1620 	 */
1621 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1622 		VOP_UNLOCK(vp);
1623 		CTR2(KTR_VFS,
1624 		    "%s: impossible to recycle, cannot start the write for %p",
1625 		    __func__, vp);
1626 		vdrop(vp);
1627 		return (EBUSY);
1628 	}
1629 	/*
1630 	 * If we got this far, we need to acquire the interlock and see if
1631 	 * anyone picked up this vnode from another list.  If not, we will
1632 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1633 	 * will skip over it.
1634 	 */
1635 	VI_LOCK(vp);
1636 	if (vp->v_usecount) {
1637 		VOP_UNLOCK(vp);
1638 		vdropl(vp);
1639 		vn_finished_write(vnmp);
1640 		CTR2(KTR_VFS,
1641 		    "%s: impossible to recycle, %p is already referenced",
1642 		    __func__, vp);
1643 		return (EBUSY);
1644 	}
1645 	if (!VN_IS_DOOMED(vp)) {
1646 		counter_u64_add(recycles_free_count, 1);
1647 		vgonel(vp);
1648 	}
1649 	VOP_UNLOCK(vp);
1650 	vdropl(vp);
1651 	vn_finished_write(vnmp);
1652 	return (0);
1653 }
1654 
1655 /*
1656  * Allocate a new vnode.
1657  *
1658  * The operation never returns an error. Returning an error was disabled
1659  * in r145385 (dated 2005) with the following comment:
1660  *
1661  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1662  *
1663  * Given the age of this commit (almost 15 years at the time of writing this
1664  * comment) restoring the ability to fail requires a significant audit of
1665  * all codepaths.
1666  *
1667  * The routine can try to free a vnode or stall for up to 1 second waiting for
1668  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1669  */
1670 static u_long vn_alloc_cyclecount;
1671 
1672 static struct vnode * __noinline
1673 vn_alloc_hard(struct mount *mp)
1674 {
1675 	u_long rnumvnodes, rfreevnodes;
1676 
1677 	mtx_lock(&vnode_list_mtx);
1678 	rnumvnodes = atomic_load_long(&numvnodes);
1679 	if (rnumvnodes + 1 < desiredvnodes) {
1680 		vn_alloc_cyclecount = 0;
1681 		goto alloc;
1682 	}
1683 	rfreevnodes = vnlru_read_freevnodes();
1684 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1685 		vn_alloc_cyclecount = 0;
1686 		vstir = 1;
1687 	}
1688 	/*
1689 	 * Grow the vnode cache if it will not be above its target max
1690 	 * after growing.  Otherwise, if the free list is nonempty, try
1691 	 * to reclaim 1 item from it before growing the cache (possibly
1692 	 * above its target max if the reclamation failed or is delayed).
1693 	 * Otherwise, wait for some space.  In all cases, schedule
1694 	 * vnlru_proc() if we are getting short of space.  The watermarks
1695 	 * should be chosen so that we never wait or even reclaim from
1696 	 * the free list to below its target minimum.
1697 	 */
1698 	if (vnlru_free_locked(1) > 0)
1699 		goto alloc;
1700 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1701 		/*
1702 		 * Wait for space for a new vnode.
1703 		 */
1704 		vnlru_kick();
1705 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1706 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1707 		    vnlru_read_freevnodes() > 1)
1708 			vnlru_free_locked(1);
1709 	}
1710 alloc:
1711 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1712 	if (vnlru_under(rnumvnodes, vlowat))
1713 		vnlru_kick();
1714 	mtx_unlock(&vnode_list_mtx);
1715 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1716 }
1717 
1718 static struct vnode *
1719 vn_alloc(struct mount *mp)
1720 {
1721 	u_long rnumvnodes;
1722 
1723 	if (__predict_false(vn_alloc_cyclecount != 0))
1724 		return (vn_alloc_hard(mp));
1725 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1726 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1727 		atomic_subtract_long(&numvnodes, 1);
1728 		return (vn_alloc_hard(mp));
1729 	}
1730 
1731 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1732 }
1733 
1734 static void
1735 vn_free(struct vnode *vp)
1736 {
1737 
1738 	atomic_subtract_long(&numvnodes, 1);
1739 	uma_zfree_smr(vnode_zone, vp);
1740 }
1741 
1742 /*
1743  * Return the next vnode from the free list.
1744  */
1745 int
1746 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1747     struct vnode **vpp)
1748 {
1749 	struct vnode *vp;
1750 	struct thread *td;
1751 	struct lock_object *lo;
1752 
1753 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1754 
1755 	KASSERT(vops->registered,
1756 	    ("%s: not registered vector op %p\n", __func__, vops));
1757 
1758 	td = curthread;
1759 	if (td->td_vp_reserved != NULL) {
1760 		vp = td->td_vp_reserved;
1761 		td->td_vp_reserved = NULL;
1762 	} else {
1763 		vp = vn_alloc(mp);
1764 	}
1765 	counter_u64_add(vnodes_created, 1);
1766 	/*
1767 	 * Locks are given the generic name "vnode" when created.
1768 	 * Follow the historic practice of using the filesystem
1769 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1770 	 *
1771 	 * Locks live in a witness group keyed on their name. Thus,
1772 	 * when a lock is renamed, it must also move from the witness
1773 	 * group of its old name to the witness group of its new name.
1774 	 *
1775 	 * The change only needs to be made when the vnode moves
1776 	 * from one filesystem type to another. We ensure that each
1777 	 * filesystem use a single static name pointer for its tag so
1778 	 * that we can compare pointers rather than doing a strcmp().
1779 	 */
1780 	lo = &vp->v_vnlock->lock_object;
1781 #ifdef WITNESS
1782 	if (lo->lo_name != tag) {
1783 #endif
1784 		lo->lo_name = tag;
1785 #ifdef WITNESS
1786 		WITNESS_DESTROY(lo);
1787 		WITNESS_INIT(lo, tag);
1788 	}
1789 #endif
1790 	/*
1791 	 * By default, don't allow shared locks unless filesystems opt-in.
1792 	 */
1793 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1794 	/*
1795 	 * Finalize various vnode identity bits.
1796 	 */
1797 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1798 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1799 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1800 	vp->v_type = VNON;
1801 	vp->v_op = vops;
1802 	vp->v_irflag = 0;
1803 	v_init_counters(vp);
1804 	vn_seqc_init(vp);
1805 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1806 #ifdef DIAGNOSTIC
1807 	if (mp == NULL && vops != &dead_vnodeops)
1808 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1809 #endif
1810 #ifdef MAC
1811 	mac_vnode_init(vp);
1812 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1813 		mac_vnode_associate_singlelabel(mp, vp);
1814 #endif
1815 	if (mp != NULL) {
1816 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1817 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1818 			vp->v_vflag |= VV_NOKNOTE;
1819 	}
1820 
1821 	/*
1822 	 * For the filesystems which do not use vfs_hash_insert(),
1823 	 * still initialize v_hash to have vfs_hash_index() useful.
1824 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1825 	 * its own hashing.
1826 	 */
1827 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1828 
1829 	*vpp = vp;
1830 	return (0);
1831 }
1832 
1833 void
1834 getnewvnode_reserve(void)
1835 {
1836 	struct thread *td;
1837 
1838 	td = curthread;
1839 	MPASS(td->td_vp_reserved == NULL);
1840 	td->td_vp_reserved = vn_alloc(NULL);
1841 }
1842 
1843 void
1844 getnewvnode_drop_reserve(void)
1845 {
1846 	struct thread *td;
1847 
1848 	td = curthread;
1849 	if (td->td_vp_reserved != NULL) {
1850 		vn_free(td->td_vp_reserved);
1851 		td->td_vp_reserved = NULL;
1852 	}
1853 }
1854 
1855 static void __noinline
1856 freevnode(struct vnode *vp)
1857 {
1858 	struct bufobj *bo;
1859 
1860 	/*
1861 	 * The vnode has been marked for destruction, so free it.
1862 	 *
1863 	 * The vnode will be returned to the zone where it will
1864 	 * normally remain until it is needed for another vnode. We
1865 	 * need to cleanup (or verify that the cleanup has already
1866 	 * been done) any residual data left from its current use
1867 	 * so as not to contaminate the freshly allocated vnode.
1868 	 */
1869 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1870 	/*
1871 	 * Paired with vgone.
1872 	 */
1873 	vn_seqc_write_end_free(vp);
1874 
1875 	bo = &vp->v_bufobj;
1876 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1877 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1878 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1879 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1880 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1881 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1882 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1883 	    ("clean blk trie not empty"));
1884 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1885 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1886 	    ("dirty blk trie not empty"));
1887 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1888 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1889 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1890 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1891 	    ("Dangling rangelock waiters"));
1892 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1893 	    ("Leaked inactivation"));
1894 	VI_UNLOCK(vp);
1895 #ifdef MAC
1896 	mac_vnode_destroy(vp);
1897 #endif
1898 	if (vp->v_pollinfo != NULL) {
1899 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1900 		destroy_vpollinfo(vp->v_pollinfo);
1901 		VOP_UNLOCK(vp);
1902 		vp->v_pollinfo = NULL;
1903 	}
1904 	vp->v_mountedhere = NULL;
1905 	vp->v_unpcb = NULL;
1906 	vp->v_rdev = NULL;
1907 	vp->v_fifoinfo = NULL;
1908 	vp->v_iflag = 0;
1909 	vp->v_vflag = 0;
1910 	bo->bo_flag = 0;
1911 	vn_free(vp);
1912 }
1913 
1914 /*
1915  * Delete from old mount point vnode list, if on one.
1916  */
1917 static void
1918 delmntque(struct vnode *vp)
1919 {
1920 	struct mount *mp;
1921 
1922 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1923 
1924 	mp = vp->v_mount;
1925 	if (mp == NULL)
1926 		return;
1927 	MNT_ILOCK(mp);
1928 	VI_LOCK(vp);
1929 	vp->v_mount = NULL;
1930 	VI_UNLOCK(vp);
1931 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1932 		("bad mount point vnode list size"));
1933 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1934 	mp->mnt_nvnodelistsize--;
1935 	MNT_REL(mp);
1936 	MNT_IUNLOCK(mp);
1937 }
1938 
1939 static void
1940 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1941 {
1942 
1943 	vp->v_data = NULL;
1944 	vp->v_op = &dead_vnodeops;
1945 	vgone(vp);
1946 	vput(vp);
1947 }
1948 
1949 /*
1950  * Insert into list of vnodes for the new mount point, if available.
1951  */
1952 int
1953 insmntque1(struct vnode *vp, struct mount *mp,
1954 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1955 {
1956 
1957 	KASSERT(vp->v_mount == NULL,
1958 		("insmntque: vnode already on per mount vnode list"));
1959 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1960 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1961 
1962 	/*
1963 	 * We acquire the vnode interlock early to ensure that the
1964 	 * vnode cannot be recycled by another process releasing a
1965 	 * holdcnt on it before we get it on both the vnode list
1966 	 * and the active vnode list. The mount mutex protects only
1967 	 * manipulation of the vnode list and the vnode freelist
1968 	 * mutex protects only manipulation of the active vnode list.
1969 	 * Hence the need to hold the vnode interlock throughout.
1970 	 */
1971 	MNT_ILOCK(mp);
1972 	VI_LOCK(vp);
1973 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1974 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1975 	    mp->mnt_nvnodelistsize == 0)) &&
1976 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1977 		VI_UNLOCK(vp);
1978 		MNT_IUNLOCK(mp);
1979 		if (dtr != NULL)
1980 			dtr(vp, dtr_arg);
1981 		return (EBUSY);
1982 	}
1983 	vp->v_mount = mp;
1984 	MNT_REF(mp);
1985 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1986 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1987 		("neg mount point vnode list size"));
1988 	mp->mnt_nvnodelistsize++;
1989 	VI_UNLOCK(vp);
1990 	MNT_IUNLOCK(mp);
1991 	return (0);
1992 }
1993 
1994 int
1995 insmntque(struct vnode *vp, struct mount *mp)
1996 {
1997 
1998 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1999 }
2000 
2001 /*
2002  * Flush out and invalidate all buffers associated with a bufobj
2003  * Called with the underlying object locked.
2004  */
2005 int
2006 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2007 {
2008 	int error;
2009 
2010 	BO_LOCK(bo);
2011 	if (flags & V_SAVE) {
2012 		error = bufobj_wwait(bo, slpflag, slptimeo);
2013 		if (error) {
2014 			BO_UNLOCK(bo);
2015 			return (error);
2016 		}
2017 		if (bo->bo_dirty.bv_cnt > 0) {
2018 			BO_UNLOCK(bo);
2019 			do {
2020 				error = BO_SYNC(bo, MNT_WAIT);
2021 			} while (error == ERELOOKUP);
2022 			if (error != 0)
2023 				return (error);
2024 			BO_LOCK(bo);
2025 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2026 				BO_UNLOCK(bo);
2027 				return (EBUSY);
2028 			}
2029 		}
2030 	}
2031 	/*
2032 	 * If you alter this loop please notice that interlock is dropped and
2033 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2034 	 * no race conditions occur from this.
2035 	 */
2036 	do {
2037 		error = flushbuflist(&bo->bo_clean,
2038 		    flags, bo, slpflag, slptimeo);
2039 		if (error == 0 && !(flags & V_CLEANONLY))
2040 			error = flushbuflist(&bo->bo_dirty,
2041 			    flags, bo, slpflag, slptimeo);
2042 		if (error != 0 && error != EAGAIN) {
2043 			BO_UNLOCK(bo);
2044 			return (error);
2045 		}
2046 	} while (error != 0);
2047 
2048 	/*
2049 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2050 	 * have write I/O in-progress but if there is a VM object then the
2051 	 * VM object can also have read-I/O in-progress.
2052 	 */
2053 	do {
2054 		bufobj_wwait(bo, 0, 0);
2055 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2056 			BO_UNLOCK(bo);
2057 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2058 			BO_LOCK(bo);
2059 		}
2060 	} while (bo->bo_numoutput > 0);
2061 	BO_UNLOCK(bo);
2062 
2063 	/*
2064 	 * Destroy the copy in the VM cache, too.
2065 	 */
2066 	if (bo->bo_object != NULL &&
2067 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2068 		VM_OBJECT_WLOCK(bo->bo_object);
2069 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2070 		    OBJPR_CLEANONLY : 0);
2071 		VM_OBJECT_WUNLOCK(bo->bo_object);
2072 	}
2073 
2074 #ifdef INVARIANTS
2075 	BO_LOCK(bo);
2076 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2077 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2078 	    bo->bo_clean.bv_cnt > 0))
2079 		panic("vinvalbuf: flush failed");
2080 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2081 	    bo->bo_dirty.bv_cnt > 0)
2082 		panic("vinvalbuf: flush dirty failed");
2083 	BO_UNLOCK(bo);
2084 #endif
2085 	return (0);
2086 }
2087 
2088 /*
2089  * Flush out and invalidate all buffers associated with a vnode.
2090  * Called with the underlying object locked.
2091  */
2092 int
2093 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2094 {
2095 
2096 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2097 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2098 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2099 		return (0);
2100 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2101 }
2102 
2103 /*
2104  * Flush out buffers on the specified list.
2105  *
2106  */
2107 static int
2108 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2109     int slptimeo)
2110 {
2111 	struct buf *bp, *nbp;
2112 	int retval, error;
2113 	daddr_t lblkno;
2114 	b_xflags_t xflags;
2115 
2116 	ASSERT_BO_WLOCKED(bo);
2117 
2118 	retval = 0;
2119 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2120 		/*
2121 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2122 		 * do not skip any buffers. If we are flushing only V_NORMAL
2123 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2124 		 * flushing only V_ALT buffers then skip buffers not marked
2125 		 * as BX_ALTDATA.
2126 		 */
2127 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2128 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2129 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2130 			continue;
2131 		}
2132 		if (nbp != NULL) {
2133 			lblkno = nbp->b_lblkno;
2134 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2135 		}
2136 		retval = EAGAIN;
2137 		error = BUF_TIMELOCK(bp,
2138 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2139 		    "flushbuf", slpflag, slptimeo);
2140 		if (error) {
2141 			BO_LOCK(bo);
2142 			return (error != ENOLCK ? error : EAGAIN);
2143 		}
2144 		KASSERT(bp->b_bufobj == bo,
2145 		    ("bp %p wrong b_bufobj %p should be %p",
2146 		    bp, bp->b_bufobj, bo));
2147 		/*
2148 		 * XXX Since there are no node locks for NFS, I
2149 		 * believe there is a slight chance that a delayed
2150 		 * write will occur while sleeping just above, so
2151 		 * check for it.
2152 		 */
2153 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2154 		    (flags & V_SAVE)) {
2155 			bremfree(bp);
2156 			bp->b_flags |= B_ASYNC;
2157 			bwrite(bp);
2158 			BO_LOCK(bo);
2159 			return (EAGAIN);	/* XXX: why not loop ? */
2160 		}
2161 		bremfree(bp);
2162 		bp->b_flags |= (B_INVAL | B_RELBUF);
2163 		bp->b_flags &= ~B_ASYNC;
2164 		brelse(bp);
2165 		BO_LOCK(bo);
2166 		if (nbp == NULL)
2167 			break;
2168 		nbp = gbincore(bo, lblkno);
2169 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2170 		    != xflags)
2171 			break;			/* nbp invalid */
2172 	}
2173 	return (retval);
2174 }
2175 
2176 int
2177 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2178 {
2179 	struct buf *bp;
2180 	int error;
2181 	daddr_t lblkno;
2182 
2183 	ASSERT_BO_LOCKED(bo);
2184 
2185 	for (lblkno = startn;;) {
2186 again:
2187 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2188 		if (bp == NULL || bp->b_lblkno >= endn ||
2189 		    bp->b_lblkno < startn)
2190 			break;
2191 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2192 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2193 		if (error != 0) {
2194 			BO_RLOCK(bo);
2195 			if (error == ENOLCK)
2196 				goto again;
2197 			return (error);
2198 		}
2199 		KASSERT(bp->b_bufobj == bo,
2200 		    ("bp %p wrong b_bufobj %p should be %p",
2201 		    bp, bp->b_bufobj, bo));
2202 		lblkno = bp->b_lblkno + 1;
2203 		if ((bp->b_flags & B_MANAGED) == 0)
2204 			bremfree(bp);
2205 		bp->b_flags |= B_RELBUF;
2206 		/*
2207 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2208 		 * pages backing each buffer in the range are unlikely to be
2209 		 * reused.  Dirty buffers will have the hint applied once
2210 		 * they've been written.
2211 		 */
2212 		if ((bp->b_flags & B_VMIO) != 0)
2213 			bp->b_flags |= B_NOREUSE;
2214 		brelse(bp);
2215 		BO_RLOCK(bo);
2216 	}
2217 	return (0);
2218 }
2219 
2220 /*
2221  * Truncate a file's buffer and pages to a specified length.  This
2222  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2223  * sync activity.
2224  */
2225 int
2226 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2227 {
2228 	struct buf *bp, *nbp;
2229 	struct bufobj *bo;
2230 	daddr_t startlbn;
2231 
2232 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2233 	    vp, blksize, (uintmax_t)length);
2234 
2235 	/*
2236 	 * Round up to the *next* lbn.
2237 	 */
2238 	startlbn = howmany(length, blksize);
2239 
2240 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2241 
2242 	bo = &vp->v_bufobj;
2243 restart_unlocked:
2244 	BO_LOCK(bo);
2245 
2246 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2247 		;
2248 
2249 	if (length > 0) {
2250 restartsync:
2251 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2252 			if (bp->b_lblkno > 0)
2253 				continue;
2254 			/*
2255 			 * Since we hold the vnode lock this should only
2256 			 * fail if we're racing with the buf daemon.
2257 			 */
2258 			if (BUF_LOCK(bp,
2259 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2260 			    BO_LOCKPTR(bo)) == ENOLCK)
2261 				goto restart_unlocked;
2262 
2263 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2264 			    ("buf(%p) on dirty queue without DELWRI", bp));
2265 
2266 			bremfree(bp);
2267 			bawrite(bp);
2268 			BO_LOCK(bo);
2269 			goto restartsync;
2270 		}
2271 	}
2272 
2273 	bufobj_wwait(bo, 0, 0);
2274 	BO_UNLOCK(bo);
2275 	vnode_pager_setsize(vp, length);
2276 
2277 	return (0);
2278 }
2279 
2280 /*
2281  * Invalidate the cached pages of a file's buffer within the range of block
2282  * numbers [startlbn, endlbn).
2283  */
2284 void
2285 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2286     int blksize)
2287 {
2288 	struct bufobj *bo;
2289 	off_t start, end;
2290 
2291 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2292 
2293 	start = blksize * startlbn;
2294 	end = blksize * endlbn;
2295 
2296 	bo = &vp->v_bufobj;
2297 	BO_LOCK(bo);
2298 	MPASS(blksize == bo->bo_bsize);
2299 
2300 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2301 		;
2302 
2303 	BO_UNLOCK(bo);
2304 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2305 }
2306 
2307 static int
2308 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2309     daddr_t startlbn, daddr_t endlbn)
2310 {
2311 	struct buf *bp, *nbp;
2312 	bool anyfreed;
2313 
2314 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2315 	ASSERT_BO_LOCKED(bo);
2316 
2317 	do {
2318 		anyfreed = false;
2319 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2320 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2321 				continue;
2322 			if (BUF_LOCK(bp,
2323 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2324 			    BO_LOCKPTR(bo)) == ENOLCK) {
2325 				BO_LOCK(bo);
2326 				return (EAGAIN);
2327 			}
2328 
2329 			bremfree(bp);
2330 			bp->b_flags |= B_INVAL | B_RELBUF;
2331 			bp->b_flags &= ~B_ASYNC;
2332 			brelse(bp);
2333 			anyfreed = true;
2334 
2335 			BO_LOCK(bo);
2336 			if (nbp != NULL &&
2337 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2338 			    nbp->b_vp != vp ||
2339 			    (nbp->b_flags & B_DELWRI) != 0))
2340 				return (EAGAIN);
2341 		}
2342 
2343 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2344 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2345 				continue;
2346 			if (BUF_LOCK(bp,
2347 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2348 			    BO_LOCKPTR(bo)) == ENOLCK) {
2349 				BO_LOCK(bo);
2350 				return (EAGAIN);
2351 			}
2352 			bremfree(bp);
2353 			bp->b_flags |= B_INVAL | B_RELBUF;
2354 			bp->b_flags &= ~B_ASYNC;
2355 			brelse(bp);
2356 			anyfreed = true;
2357 
2358 			BO_LOCK(bo);
2359 			if (nbp != NULL &&
2360 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2361 			    (nbp->b_vp != vp) ||
2362 			    (nbp->b_flags & B_DELWRI) == 0))
2363 				return (EAGAIN);
2364 		}
2365 	} while (anyfreed);
2366 	return (0);
2367 }
2368 
2369 static void
2370 buf_vlist_remove(struct buf *bp)
2371 {
2372 	struct bufv *bv;
2373 	b_xflags_t flags;
2374 
2375 	flags = bp->b_xflags;
2376 
2377 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2378 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2379 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2380 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2381 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2382 
2383 	if ((flags & BX_VNDIRTY) != 0)
2384 		bv = &bp->b_bufobj->bo_dirty;
2385 	else
2386 		bv = &bp->b_bufobj->bo_clean;
2387 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2388 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2389 	bv->bv_cnt--;
2390 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2391 }
2392 
2393 /*
2394  * Add the buffer to the sorted clean or dirty block list.
2395  *
2396  * NOTE: xflags is passed as a constant, optimizing this inline function!
2397  */
2398 static void
2399 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2400 {
2401 	struct bufv *bv;
2402 	struct buf *n;
2403 	int error;
2404 
2405 	ASSERT_BO_WLOCKED(bo);
2406 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2407 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2408 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2409 	    ("dead bo %p", bo));
2410 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2411 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2412 	bp->b_xflags |= xflags;
2413 	if (xflags & BX_VNDIRTY)
2414 		bv = &bo->bo_dirty;
2415 	else
2416 		bv = &bo->bo_clean;
2417 
2418 	/*
2419 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2420 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2421 	 * than _ge.
2422 	 */
2423 	if (bv->bv_cnt == 0 ||
2424 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2425 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2426 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2427 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2428 	else
2429 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2430 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2431 	if (error)
2432 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2433 	bv->bv_cnt++;
2434 }
2435 
2436 /*
2437  * Look up a buffer using the buffer tries.
2438  */
2439 struct buf *
2440 gbincore(struct bufobj *bo, daddr_t lblkno)
2441 {
2442 	struct buf *bp;
2443 
2444 	ASSERT_BO_LOCKED(bo);
2445 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2446 	if (bp != NULL)
2447 		return (bp);
2448 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2449 }
2450 
2451 /*
2452  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2453  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2454  * stability of the result.  Like other lockless lookups, the found buf may
2455  * already be invalid by the time this function returns.
2456  */
2457 struct buf *
2458 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2459 {
2460 	struct buf *bp;
2461 
2462 	ASSERT_BO_UNLOCKED(bo);
2463 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2464 	if (bp != NULL)
2465 		return (bp);
2466 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2467 }
2468 
2469 /*
2470  * Associate a buffer with a vnode.
2471  */
2472 void
2473 bgetvp(struct vnode *vp, struct buf *bp)
2474 {
2475 	struct bufobj *bo;
2476 
2477 	bo = &vp->v_bufobj;
2478 	ASSERT_BO_WLOCKED(bo);
2479 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2480 
2481 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2482 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2483 	    ("bgetvp: bp already attached! %p", bp));
2484 
2485 	vhold(vp);
2486 	bp->b_vp = vp;
2487 	bp->b_bufobj = bo;
2488 	/*
2489 	 * Insert onto list for new vnode.
2490 	 */
2491 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2492 }
2493 
2494 /*
2495  * Disassociate a buffer from a vnode.
2496  */
2497 void
2498 brelvp(struct buf *bp)
2499 {
2500 	struct bufobj *bo;
2501 	struct vnode *vp;
2502 
2503 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2504 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2505 
2506 	/*
2507 	 * Delete from old vnode list, if on one.
2508 	 */
2509 	vp = bp->b_vp;		/* XXX */
2510 	bo = bp->b_bufobj;
2511 	BO_LOCK(bo);
2512 	buf_vlist_remove(bp);
2513 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2514 		bo->bo_flag &= ~BO_ONWORKLST;
2515 		mtx_lock(&sync_mtx);
2516 		LIST_REMOVE(bo, bo_synclist);
2517 		syncer_worklist_len--;
2518 		mtx_unlock(&sync_mtx);
2519 	}
2520 	bp->b_vp = NULL;
2521 	bp->b_bufobj = NULL;
2522 	BO_UNLOCK(bo);
2523 	vdrop(vp);
2524 }
2525 
2526 /*
2527  * Add an item to the syncer work queue.
2528  */
2529 static void
2530 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2531 {
2532 	int slot;
2533 
2534 	ASSERT_BO_WLOCKED(bo);
2535 
2536 	mtx_lock(&sync_mtx);
2537 	if (bo->bo_flag & BO_ONWORKLST)
2538 		LIST_REMOVE(bo, bo_synclist);
2539 	else {
2540 		bo->bo_flag |= BO_ONWORKLST;
2541 		syncer_worklist_len++;
2542 	}
2543 
2544 	if (delay > syncer_maxdelay - 2)
2545 		delay = syncer_maxdelay - 2;
2546 	slot = (syncer_delayno + delay) & syncer_mask;
2547 
2548 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2549 	mtx_unlock(&sync_mtx);
2550 }
2551 
2552 static int
2553 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2554 {
2555 	int error, len;
2556 
2557 	mtx_lock(&sync_mtx);
2558 	len = syncer_worklist_len - sync_vnode_count;
2559 	mtx_unlock(&sync_mtx);
2560 	error = SYSCTL_OUT(req, &len, sizeof(len));
2561 	return (error);
2562 }
2563 
2564 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2565     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2566     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2567 
2568 static struct proc *updateproc;
2569 static void sched_sync(void);
2570 static struct kproc_desc up_kp = {
2571 	"syncer",
2572 	sched_sync,
2573 	&updateproc
2574 };
2575 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2576 
2577 static int
2578 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2579 {
2580 	struct vnode *vp;
2581 	struct mount *mp;
2582 
2583 	*bo = LIST_FIRST(slp);
2584 	if (*bo == NULL)
2585 		return (0);
2586 	vp = bo2vnode(*bo);
2587 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2588 		return (1);
2589 	/*
2590 	 * We use vhold in case the vnode does not
2591 	 * successfully sync.  vhold prevents the vnode from
2592 	 * going away when we unlock the sync_mtx so that
2593 	 * we can acquire the vnode interlock.
2594 	 */
2595 	vholdl(vp);
2596 	mtx_unlock(&sync_mtx);
2597 	VI_UNLOCK(vp);
2598 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2599 		vdrop(vp);
2600 		mtx_lock(&sync_mtx);
2601 		return (*bo == LIST_FIRST(slp));
2602 	}
2603 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2604 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2605 	VOP_UNLOCK(vp);
2606 	vn_finished_write(mp);
2607 	BO_LOCK(*bo);
2608 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2609 		/*
2610 		 * Put us back on the worklist.  The worklist
2611 		 * routine will remove us from our current
2612 		 * position and then add us back in at a later
2613 		 * position.
2614 		 */
2615 		vn_syncer_add_to_worklist(*bo, syncdelay);
2616 	}
2617 	BO_UNLOCK(*bo);
2618 	vdrop(vp);
2619 	mtx_lock(&sync_mtx);
2620 	return (0);
2621 }
2622 
2623 static int first_printf = 1;
2624 
2625 /*
2626  * System filesystem synchronizer daemon.
2627  */
2628 static void
2629 sched_sync(void)
2630 {
2631 	struct synclist *next, *slp;
2632 	struct bufobj *bo;
2633 	long starttime;
2634 	struct thread *td = curthread;
2635 	int last_work_seen;
2636 	int net_worklist_len;
2637 	int syncer_final_iter;
2638 	int error;
2639 
2640 	last_work_seen = 0;
2641 	syncer_final_iter = 0;
2642 	syncer_state = SYNCER_RUNNING;
2643 	starttime = time_uptime;
2644 	td->td_pflags |= TDP_NORUNNINGBUF;
2645 
2646 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2647 	    SHUTDOWN_PRI_LAST);
2648 
2649 	mtx_lock(&sync_mtx);
2650 	for (;;) {
2651 		if (syncer_state == SYNCER_FINAL_DELAY &&
2652 		    syncer_final_iter == 0) {
2653 			mtx_unlock(&sync_mtx);
2654 			kproc_suspend_check(td->td_proc);
2655 			mtx_lock(&sync_mtx);
2656 		}
2657 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2658 		if (syncer_state != SYNCER_RUNNING &&
2659 		    starttime != time_uptime) {
2660 			if (first_printf) {
2661 				printf("\nSyncing disks, vnodes remaining... ");
2662 				first_printf = 0;
2663 			}
2664 			printf("%d ", net_worklist_len);
2665 		}
2666 		starttime = time_uptime;
2667 
2668 		/*
2669 		 * Push files whose dirty time has expired.  Be careful
2670 		 * of interrupt race on slp queue.
2671 		 *
2672 		 * Skip over empty worklist slots when shutting down.
2673 		 */
2674 		do {
2675 			slp = &syncer_workitem_pending[syncer_delayno];
2676 			syncer_delayno += 1;
2677 			if (syncer_delayno == syncer_maxdelay)
2678 				syncer_delayno = 0;
2679 			next = &syncer_workitem_pending[syncer_delayno];
2680 			/*
2681 			 * If the worklist has wrapped since the
2682 			 * it was emptied of all but syncer vnodes,
2683 			 * switch to the FINAL_DELAY state and run
2684 			 * for one more second.
2685 			 */
2686 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2687 			    net_worklist_len == 0 &&
2688 			    last_work_seen == syncer_delayno) {
2689 				syncer_state = SYNCER_FINAL_DELAY;
2690 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2691 			}
2692 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2693 		    syncer_worklist_len > 0);
2694 
2695 		/*
2696 		 * Keep track of the last time there was anything
2697 		 * on the worklist other than syncer vnodes.
2698 		 * Return to the SHUTTING_DOWN state if any
2699 		 * new work appears.
2700 		 */
2701 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2702 			last_work_seen = syncer_delayno;
2703 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2704 			syncer_state = SYNCER_SHUTTING_DOWN;
2705 		while (!LIST_EMPTY(slp)) {
2706 			error = sync_vnode(slp, &bo, td);
2707 			if (error == 1) {
2708 				LIST_REMOVE(bo, bo_synclist);
2709 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2710 				continue;
2711 			}
2712 
2713 			if (first_printf == 0) {
2714 				/*
2715 				 * Drop the sync mutex, because some watchdog
2716 				 * drivers need to sleep while patting
2717 				 */
2718 				mtx_unlock(&sync_mtx);
2719 				wdog_kern_pat(WD_LASTVAL);
2720 				mtx_lock(&sync_mtx);
2721 			}
2722 		}
2723 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2724 			syncer_final_iter--;
2725 		/*
2726 		 * The variable rushjob allows the kernel to speed up the
2727 		 * processing of the filesystem syncer process. A rushjob
2728 		 * value of N tells the filesystem syncer to process the next
2729 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2730 		 * is used by the soft update code to speed up the filesystem
2731 		 * syncer process when the incore state is getting so far
2732 		 * ahead of the disk that the kernel memory pool is being
2733 		 * threatened with exhaustion.
2734 		 */
2735 		if (rushjob > 0) {
2736 			rushjob -= 1;
2737 			continue;
2738 		}
2739 		/*
2740 		 * Just sleep for a short period of time between
2741 		 * iterations when shutting down to allow some I/O
2742 		 * to happen.
2743 		 *
2744 		 * If it has taken us less than a second to process the
2745 		 * current work, then wait. Otherwise start right over
2746 		 * again. We can still lose time if any single round
2747 		 * takes more than two seconds, but it does not really
2748 		 * matter as we are just trying to generally pace the
2749 		 * filesystem activity.
2750 		 */
2751 		if (syncer_state != SYNCER_RUNNING ||
2752 		    time_uptime == starttime) {
2753 			thread_lock(td);
2754 			sched_prio(td, PPAUSE);
2755 			thread_unlock(td);
2756 		}
2757 		if (syncer_state != SYNCER_RUNNING)
2758 			cv_timedwait(&sync_wakeup, &sync_mtx,
2759 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2760 		else if (time_uptime == starttime)
2761 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2762 	}
2763 }
2764 
2765 /*
2766  * Request the syncer daemon to speed up its work.
2767  * We never push it to speed up more than half of its
2768  * normal turn time, otherwise it could take over the cpu.
2769  */
2770 int
2771 speedup_syncer(void)
2772 {
2773 	int ret = 0;
2774 
2775 	mtx_lock(&sync_mtx);
2776 	if (rushjob < syncdelay / 2) {
2777 		rushjob += 1;
2778 		stat_rush_requests += 1;
2779 		ret = 1;
2780 	}
2781 	mtx_unlock(&sync_mtx);
2782 	cv_broadcast(&sync_wakeup);
2783 	return (ret);
2784 }
2785 
2786 /*
2787  * Tell the syncer to speed up its work and run though its work
2788  * list several times, then tell it to shut down.
2789  */
2790 static void
2791 syncer_shutdown(void *arg, int howto)
2792 {
2793 
2794 	if (howto & RB_NOSYNC)
2795 		return;
2796 	mtx_lock(&sync_mtx);
2797 	syncer_state = SYNCER_SHUTTING_DOWN;
2798 	rushjob = 0;
2799 	mtx_unlock(&sync_mtx);
2800 	cv_broadcast(&sync_wakeup);
2801 	kproc_shutdown(arg, howto);
2802 }
2803 
2804 void
2805 syncer_suspend(void)
2806 {
2807 
2808 	syncer_shutdown(updateproc, 0);
2809 }
2810 
2811 void
2812 syncer_resume(void)
2813 {
2814 
2815 	mtx_lock(&sync_mtx);
2816 	first_printf = 1;
2817 	syncer_state = SYNCER_RUNNING;
2818 	mtx_unlock(&sync_mtx);
2819 	cv_broadcast(&sync_wakeup);
2820 	kproc_resume(updateproc);
2821 }
2822 
2823 /*
2824  * Move the buffer between the clean and dirty lists of its vnode.
2825  */
2826 void
2827 reassignbuf(struct buf *bp)
2828 {
2829 	struct vnode *vp;
2830 	struct bufobj *bo;
2831 	int delay;
2832 #ifdef INVARIANTS
2833 	struct bufv *bv;
2834 #endif
2835 
2836 	vp = bp->b_vp;
2837 	bo = bp->b_bufobj;
2838 
2839 	KASSERT((bp->b_flags & B_PAGING) == 0,
2840 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2841 
2842 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2843 	    bp, bp->b_vp, bp->b_flags);
2844 
2845 	BO_LOCK(bo);
2846 	buf_vlist_remove(bp);
2847 
2848 	/*
2849 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2850 	 * of clean buffers.
2851 	 */
2852 	if (bp->b_flags & B_DELWRI) {
2853 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2854 			switch (vp->v_type) {
2855 			case VDIR:
2856 				delay = dirdelay;
2857 				break;
2858 			case VCHR:
2859 				delay = metadelay;
2860 				break;
2861 			default:
2862 				delay = filedelay;
2863 			}
2864 			vn_syncer_add_to_worklist(bo, delay);
2865 		}
2866 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2867 	} else {
2868 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2869 
2870 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2871 			mtx_lock(&sync_mtx);
2872 			LIST_REMOVE(bo, bo_synclist);
2873 			syncer_worklist_len--;
2874 			mtx_unlock(&sync_mtx);
2875 			bo->bo_flag &= ~BO_ONWORKLST;
2876 		}
2877 	}
2878 #ifdef INVARIANTS
2879 	bv = &bo->bo_clean;
2880 	bp = TAILQ_FIRST(&bv->bv_hd);
2881 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2882 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2883 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2884 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2885 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2886 	bv = &bo->bo_dirty;
2887 	bp = TAILQ_FIRST(&bv->bv_hd);
2888 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2889 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2890 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2891 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2892 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2893 #endif
2894 	BO_UNLOCK(bo);
2895 }
2896 
2897 static void
2898 v_init_counters(struct vnode *vp)
2899 {
2900 
2901 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2902 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2903 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2904 
2905 	refcount_init(&vp->v_holdcnt, 1);
2906 	refcount_init(&vp->v_usecount, 1);
2907 }
2908 
2909 /*
2910  * Grab a particular vnode from the free list, increment its
2911  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2912  * is being destroyed.  Only callers who specify LK_RETRY will
2913  * see doomed vnodes.  If inactive processing was delayed in
2914  * vput try to do it here.
2915  *
2916  * usecount is manipulated using atomics without holding any locks.
2917  *
2918  * holdcnt can be manipulated using atomics without holding any locks,
2919  * except when transitioning 1<->0, in which case the interlock is held.
2920  *
2921  * Consumers which don't guarantee liveness of the vnode can use SMR to
2922  * try to get a reference. Note this operation can fail since the vnode
2923  * may be awaiting getting freed by the time they get to it.
2924  */
2925 enum vgetstate
2926 vget_prep_smr(struct vnode *vp)
2927 {
2928 	enum vgetstate vs;
2929 
2930 	VFS_SMR_ASSERT_ENTERED();
2931 
2932 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2933 		vs = VGET_USECOUNT;
2934 	} else {
2935 		if (vhold_smr(vp))
2936 			vs = VGET_HOLDCNT;
2937 		else
2938 			vs = VGET_NONE;
2939 	}
2940 	return (vs);
2941 }
2942 
2943 enum vgetstate
2944 vget_prep(struct vnode *vp)
2945 {
2946 	enum vgetstate vs;
2947 
2948 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2949 		vs = VGET_USECOUNT;
2950 	} else {
2951 		vhold(vp);
2952 		vs = VGET_HOLDCNT;
2953 	}
2954 	return (vs);
2955 }
2956 
2957 void
2958 vget_abort(struct vnode *vp, enum vgetstate vs)
2959 {
2960 
2961 	switch (vs) {
2962 	case VGET_USECOUNT:
2963 		vrele(vp);
2964 		break;
2965 	case VGET_HOLDCNT:
2966 		vdrop(vp);
2967 		break;
2968 	default:
2969 		__assert_unreachable();
2970 	}
2971 }
2972 
2973 int
2974 vget(struct vnode *vp, int flags)
2975 {
2976 	enum vgetstate vs;
2977 
2978 	vs = vget_prep(vp);
2979 	return (vget_finish(vp, flags, vs));
2980 }
2981 
2982 int
2983 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2984 {
2985 	int error;
2986 
2987 	if ((flags & LK_INTERLOCK) != 0)
2988 		ASSERT_VI_LOCKED(vp, __func__);
2989 	else
2990 		ASSERT_VI_UNLOCKED(vp, __func__);
2991 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2992 	VNPASS(vp->v_holdcnt > 0, vp);
2993 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2994 
2995 	error = vn_lock(vp, flags);
2996 	if (__predict_false(error != 0)) {
2997 		vget_abort(vp, vs);
2998 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2999 		    vp);
3000 		return (error);
3001 	}
3002 
3003 	vget_finish_ref(vp, vs);
3004 	return (0);
3005 }
3006 
3007 void
3008 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3009 {
3010 	int old;
3011 
3012 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3013 	VNPASS(vp->v_holdcnt > 0, vp);
3014 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3015 
3016 	if (vs == VGET_USECOUNT)
3017 		return;
3018 
3019 	/*
3020 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3021 	 * the vnode around. Otherwise someone else lended their hold count and
3022 	 * we have to drop ours.
3023 	 */
3024 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3025 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3026 	if (old != 0) {
3027 #ifdef INVARIANTS
3028 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3029 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3030 #else
3031 		refcount_release(&vp->v_holdcnt);
3032 #endif
3033 	}
3034 }
3035 
3036 void
3037 vref(struct vnode *vp)
3038 {
3039 	enum vgetstate vs;
3040 
3041 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3042 	vs = vget_prep(vp);
3043 	vget_finish_ref(vp, vs);
3044 }
3045 
3046 void
3047 vrefact(struct vnode *vp)
3048 {
3049 
3050 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3051 #ifdef INVARIANTS
3052 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3053 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3054 #else
3055 	refcount_acquire(&vp->v_usecount);
3056 #endif
3057 }
3058 
3059 void
3060 vlazy(struct vnode *vp)
3061 {
3062 	struct mount *mp;
3063 
3064 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3065 
3066 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3067 		return;
3068 	/*
3069 	 * We may get here for inactive routines after the vnode got doomed.
3070 	 */
3071 	if (VN_IS_DOOMED(vp))
3072 		return;
3073 	mp = vp->v_mount;
3074 	mtx_lock(&mp->mnt_listmtx);
3075 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3076 		vp->v_mflag |= VMP_LAZYLIST;
3077 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3078 		mp->mnt_lazyvnodelistsize++;
3079 	}
3080 	mtx_unlock(&mp->mnt_listmtx);
3081 }
3082 
3083 static void
3084 vunlazy(struct vnode *vp)
3085 {
3086 	struct mount *mp;
3087 
3088 	ASSERT_VI_LOCKED(vp, __func__);
3089 	VNPASS(!VN_IS_DOOMED(vp), vp);
3090 
3091 	mp = vp->v_mount;
3092 	mtx_lock(&mp->mnt_listmtx);
3093 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3094 	/*
3095 	 * Don't remove the vnode from the lazy list if another thread
3096 	 * has increased the hold count. It may have re-enqueued the
3097 	 * vnode to the lazy list and is now responsible for its
3098 	 * removal.
3099 	 */
3100 	if (vp->v_holdcnt == 0) {
3101 		vp->v_mflag &= ~VMP_LAZYLIST;
3102 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3103 		mp->mnt_lazyvnodelistsize--;
3104 	}
3105 	mtx_unlock(&mp->mnt_listmtx);
3106 }
3107 
3108 /*
3109  * This routine is only meant to be called from vgonel prior to dooming
3110  * the vnode.
3111  */
3112 static void
3113 vunlazy_gone(struct vnode *vp)
3114 {
3115 	struct mount *mp;
3116 
3117 	ASSERT_VOP_ELOCKED(vp, __func__);
3118 	ASSERT_VI_LOCKED(vp, __func__);
3119 	VNPASS(!VN_IS_DOOMED(vp), vp);
3120 
3121 	if (vp->v_mflag & VMP_LAZYLIST) {
3122 		mp = vp->v_mount;
3123 		mtx_lock(&mp->mnt_listmtx);
3124 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3125 		vp->v_mflag &= ~VMP_LAZYLIST;
3126 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3127 		mp->mnt_lazyvnodelistsize--;
3128 		mtx_unlock(&mp->mnt_listmtx);
3129 	}
3130 }
3131 
3132 static void
3133 vdefer_inactive(struct vnode *vp)
3134 {
3135 
3136 	ASSERT_VI_LOCKED(vp, __func__);
3137 	VNASSERT(vp->v_holdcnt > 0, vp,
3138 	    ("%s: vnode without hold count", __func__));
3139 	if (VN_IS_DOOMED(vp)) {
3140 		vdropl(vp);
3141 		return;
3142 	}
3143 	if (vp->v_iflag & VI_DEFINACT) {
3144 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3145 		vdropl(vp);
3146 		return;
3147 	}
3148 	if (vp->v_usecount > 0) {
3149 		vp->v_iflag &= ~VI_OWEINACT;
3150 		vdropl(vp);
3151 		return;
3152 	}
3153 	vlazy(vp);
3154 	vp->v_iflag |= VI_DEFINACT;
3155 	VI_UNLOCK(vp);
3156 	counter_u64_add(deferred_inact, 1);
3157 }
3158 
3159 static void
3160 vdefer_inactive_unlocked(struct vnode *vp)
3161 {
3162 
3163 	VI_LOCK(vp);
3164 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3165 		vdropl(vp);
3166 		return;
3167 	}
3168 	vdefer_inactive(vp);
3169 }
3170 
3171 enum vput_op { VRELE, VPUT, VUNREF };
3172 
3173 /*
3174  * Handle ->v_usecount transitioning to 0.
3175  *
3176  * By releasing the last usecount we take ownership of the hold count which
3177  * provides liveness of the vnode, meaning we have to vdrop.
3178  *
3179  * For all vnodes we may need to perform inactive processing. It requires an
3180  * exclusive lock on the vnode, while it is legal to call here with only a
3181  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3182  * inactive processing gets deferred to the syncer.
3183  *
3184  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3185  * on the lock being held all the way until VOP_INACTIVE. This in particular
3186  * happens with UFS which adds half-constructed vnodes to the hash, where they
3187  * can be found by other code.
3188  */
3189 static void
3190 vput_final(struct vnode *vp, enum vput_op func)
3191 {
3192 	int error;
3193 	bool want_unlock;
3194 
3195 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3196 	VNPASS(vp->v_holdcnt > 0, vp);
3197 
3198 	VI_LOCK(vp);
3199 
3200 	/*
3201 	 * By the time we got here someone else might have transitioned
3202 	 * the count back to > 0.
3203 	 */
3204 	if (vp->v_usecount > 0)
3205 		goto out;
3206 
3207 	/*
3208 	 * If the vnode is doomed vgone already performed inactive processing
3209 	 * (if needed).
3210 	 */
3211 	if (VN_IS_DOOMED(vp))
3212 		goto out;
3213 
3214 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3215 		goto out;
3216 
3217 	if (vp->v_iflag & VI_DOINGINACT)
3218 		goto out;
3219 
3220 	/*
3221 	 * Locking operations here will drop the interlock and possibly the
3222 	 * vnode lock, opening a window where the vnode can get doomed all the
3223 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3224 	 * perform inactive.
3225 	 */
3226 	vp->v_iflag |= VI_OWEINACT;
3227 	want_unlock = false;
3228 	error = 0;
3229 	switch (func) {
3230 	case VRELE:
3231 		switch (VOP_ISLOCKED(vp)) {
3232 		case LK_EXCLUSIVE:
3233 			break;
3234 		case LK_EXCLOTHER:
3235 		case 0:
3236 			want_unlock = true;
3237 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3238 			VI_LOCK(vp);
3239 			break;
3240 		default:
3241 			/*
3242 			 * The lock has at least one sharer, but we have no way
3243 			 * to conclude whether this is us. Play it safe and
3244 			 * defer processing.
3245 			 */
3246 			error = EAGAIN;
3247 			break;
3248 		}
3249 		break;
3250 	case VPUT:
3251 		want_unlock = true;
3252 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3253 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3254 			    LK_NOWAIT);
3255 			VI_LOCK(vp);
3256 		}
3257 		break;
3258 	case VUNREF:
3259 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3260 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3261 			VI_LOCK(vp);
3262 		}
3263 		break;
3264 	}
3265 	if (error == 0) {
3266 		if (func == VUNREF) {
3267 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3268 			    ("recursive vunref"));
3269 			vp->v_vflag |= VV_UNREF;
3270 		}
3271 		for (;;) {
3272 			error = vinactive(vp);
3273 			if (want_unlock)
3274 				VOP_UNLOCK(vp);
3275 			if (error != ERELOOKUP || !want_unlock)
3276 				break;
3277 			VOP_LOCK(vp, LK_EXCLUSIVE);
3278 		}
3279 		if (func == VUNREF)
3280 			vp->v_vflag &= ~VV_UNREF;
3281 		vdropl(vp);
3282 	} else {
3283 		vdefer_inactive(vp);
3284 	}
3285 	return;
3286 out:
3287 	if (func == VPUT)
3288 		VOP_UNLOCK(vp);
3289 	vdropl(vp);
3290 }
3291 
3292 /*
3293  * Decrement ->v_usecount for a vnode.
3294  *
3295  * Releasing the last use count requires additional processing, see vput_final
3296  * above for details.
3297  *
3298  * Comment above each variant denotes lock state on entry and exit.
3299  */
3300 
3301 /*
3302  * in: any
3303  * out: same as passed in
3304  */
3305 void
3306 vrele(struct vnode *vp)
3307 {
3308 
3309 	ASSERT_VI_UNLOCKED(vp, __func__);
3310 	if (!refcount_release(&vp->v_usecount))
3311 		return;
3312 	vput_final(vp, VRELE);
3313 }
3314 
3315 /*
3316  * in: locked
3317  * out: unlocked
3318  */
3319 void
3320 vput(struct vnode *vp)
3321 {
3322 
3323 	ASSERT_VOP_LOCKED(vp, __func__);
3324 	ASSERT_VI_UNLOCKED(vp, __func__);
3325 	if (!refcount_release(&vp->v_usecount)) {
3326 		VOP_UNLOCK(vp);
3327 		return;
3328 	}
3329 	vput_final(vp, VPUT);
3330 }
3331 
3332 /*
3333  * in: locked
3334  * out: locked
3335  */
3336 void
3337 vunref(struct vnode *vp)
3338 {
3339 
3340 	ASSERT_VOP_LOCKED(vp, __func__);
3341 	ASSERT_VI_UNLOCKED(vp, __func__);
3342 	if (!refcount_release(&vp->v_usecount))
3343 		return;
3344 	vput_final(vp, VUNREF);
3345 }
3346 
3347 void
3348 vhold(struct vnode *vp)
3349 {
3350 	int old;
3351 
3352 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3353 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3354 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3355 	    ("%s: wrong hold count %d", __func__, old));
3356 	if (old == 0)
3357 		vfs_freevnodes_dec();
3358 }
3359 
3360 void
3361 vholdnz(struct vnode *vp)
3362 {
3363 
3364 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3365 #ifdef INVARIANTS
3366 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3367 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3368 	    ("%s: wrong hold count %d", __func__, old));
3369 #else
3370 	atomic_add_int(&vp->v_holdcnt, 1);
3371 #endif
3372 }
3373 
3374 /*
3375  * Grab a hold count unless the vnode is freed.
3376  *
3377  * Only use this routine if vfs smr is the only protection you have against
3378  * freeing the vnode.
3379  *
3380  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3381  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3382  * the thread which managed to set the flag.
3383  *
3384  * It may be tempting to replace the loop with:
3385  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3386  * if (count & VHOLD_NO_SMR) {
3387  *     backpedal and error out;
3388  * }
3389  *
3390  * However, while this is more performant, it hinders debugging by eliminating
3391  * the previously mentioned invariant.
3392  */
3393 bool
3394 vhold_smr(struct vnode *vp)
3395 {
3396 	int count;
3397 
3398 	VFS_SMR_ASSERT_ENTERED();
3399 
3400 	count = atomic_load_int(&vp->v_holdcnt);
3401 	for (;;) {
3402 		if (count & VHOLD_NO_SMR) {
3403 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3404 			    ("non-zero hold count with flags %d\n", count));
3405 			return (false);
3406 		}
3407 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3408 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3409 			if (count == 0)
3410 				vfs_freevnodes_dec();
3411 			return (true);
3412 		}
3413 	}
3414 }
3415 
3416 /*
3417  * Hold a free vnode for recycling.
3418  *
3419  * Note: vnode_init references this comment.
3420  *
3421  * Attempts to recycle only need the global vnode list lock and have no use for
3422  * SMR.
3423  *
3424  * However, vnodes get inserted into the global list before they get fully
3425  * initialized and stay there until UMA decides to free the memory. This in
3426  * particular means the target can be found before it becomes usable and after
3427  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3428  * VHOLD_NO_SMR.
3429  *
3430  * Note: the vnode may gain more references after we transition the count 0->1.
3431  */
3432 static bool
3433 vhold_recycle_free(struct vnode *vp)
3434 {
3435 	int count;
3436 
3437 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3438 
3439 	count = atomic_load_int(&vp->v_holdcnt);
3440 	for (;;) {
3441 		if (count & VHOLD_NO_SMR) {
3442 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3443 			    ("non-zero hold count with flags %d\n", count));
3444 			return (false);
3445 		}
3446 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3447 		if (count > 0) {
3448 			return (false);
3449 		}
3450 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3451 			vfs_freevnodes_dec();
3452 			return (true);
3453 		}
3454 	}
3455 }
3456 
3457 static void __noinline
3458 vdbatch_process(struct vdbatch *vd)
3459 {
3460 	struct vnode *vp;
3461 	int i;
3462 
3463 	mtx_assert(&vd->lock, MA_OWNED);
3464 	MPASS(curthread->td_pinned > 0);
3465 	MPASS(vd->index == VDBATCH_SIZE);
3466 
3467 	mtx_lock(&vnode_list_mtx);
3468 	critical_enter();
3469 	freevnodes += vd->freevnodes;
3470 	for (i = 0; i < VDBATCH_SIZE; i++) {
3471 		vp = vd->tab[i];
3472 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3473 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3474 		MPASS(vp->v_dbatchcpu != NOCPU);
3475 		vp->v_dbatchcpu = NOCPU;
3476 	}
3477 	mtx_unlock(&vnode_list_mtx);
3478 	vd->freevnodes = 0;
3479 	bzero(vd->tab, sizeof(vd->tab));
3480 	vd->index = 0;
3481 	critical_exit();
3482 }
3483 
3484 static void
3485 vdbatch_enqueue(struct vnode *vp)
3486 {
3487 	struct vdbatch *vd;
3488 
3489 	ASSERT_VI_LOCKED(vp, __func__);
3490 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3491 	    ("%s: deferring requeue of a doomed vnode", __func__));
3492 
3493 	if (vp->v_dbatchcpu != NOCPU) {
3494 		VI_UNLOCK(vp);
3495 		return;
3496 	}
3497 
3498 	sched_pin();
3499 	vd = DPCPU_PTR(vd);
3500 	mtx_lock(&vd->lock);
3501 	MPASS(vd->index < VDBATCH_SIZE);
3502 	MPASS(vd->tab[vd->index] == NULL);
3503 	/*
3504 	 * A hack: we depend on being pinned so that we know what to put in
3505 	 * ->v_dbatchcpu.
3506 	 */
3507 	vp->v_dbatchcpu = curcpu;
3508 	vd->tab[vd->index] = vp;
3509 	vd->index++;
3510 	VI_UNLOCK(vp);
3511 	if (vd->index == VDBATCH_SIZE)
3512 		vdbatch_process(vd);
3513 	mtx_unlock(&vd->lock);
3514 	sched_unpin();
3515 }
3516 
3517 /*
3518  * This routine must only be called for vnodes which are about to be
3519  * deallocated. Supporting dequeue for arbitrary vndoes would require
3520  * validating that the locked batch matches.
3521  */
3522 static void
3523 vdbatch_dequeue(struct vnode *vp)
3524 {
3525 	struct vdbatch *vd;
3526 	int i;
3527 	short cpu;
3528 
3529 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3530 	    ("%s: called for a used vnode\n", __func__));
3531 
3532 	cpu = vp->v_dbatchcpu;
3533 	if (cpu == NOCPU)
3534 		return;
3535 
3536 	vd = DPCPU_ID_PTR(cpu, vd);
3537 	mtx_lock(&vd->lock);
3538 	for (i = 0; i < vd->index; i++) {
3539 		if (vd->tab[i] != vp)
3540 			continue;
3541 		vp->v_dbatchcpu = NOCPU;
3542 		vd->index--;
3543 		vd->tab[i] = vd->tab[vd->index];
3544 		vd->tab[vd->index] = NULL;
3545 		break;
3546 	}
3547 	mtx_unlock(&vd->lock);
3548 	/*
3549 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3550 	 */
3551 	MPASS(vp->v_dbatchcpu == NOCPU);
3552 }
3553 
3554 /*
3555  * Drop the hold count of the vnode.  If this is the last reference to
3556  * the vnode we place it on the free list unless it has been vgone'd
3557  * (marked VIRF_DOOMED) in which case we will free it.
3558  *
3559  * Because the vnode vm object keeps a hold reference on the vnode if
3560  * there is at least one resident non-cached page, the vnode cannot
3561  * leave the active list without the page cleanup done.
3562  */
3563 static void __noinline
3564 vdropl_final(struct vnode *vp)
3565 {
3566 
3567 	ASSERT_VI_LOCKED(vp, __func__);
3568 	VNPASS(VN_IS_DOOMED(vp), vp);
3569 	/*
3570 	 * Set the VHOLD_NO_SMR flag.
3571 	 *
3572 	 * We may be racing against vhold_smr. If they win we can just pretend
3573 	 * we never got this far, they will vdrop later.
3574 	 */
3575 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3576 		vfs_freevnodes_inc();
3577 		VI_UNLOCK(vp);
3578 		/*
3579 		 * We lost the aforementioned race. Any subsequent access is
3580 		 * invalid as they might have managed to vdropl on their own.
3581 		 */
3582 		return;
3583 	}
3584 	/*
3585 	 * Don't bump freevnodes as this one is going away.
3586 	 */
3587 	freevnode(vp);
3588 }
3589 
3590 void
3591 vdrop(struct vnode *vp)
3592 {
3593 
3594 	ASSERT_VI_UNLOCKED(vp, __func__);
3595 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3596 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3597 		return;
3598 	VI_LOCK(vp);
3599 	vdropl(vp);
3600 }
3601 
3602 void
3603 vdropl(struct vnode *vp)
3604 {
3605 
3606 	ASSERT_VI_LOCKED(vp, __func__);
3607 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3608 	if (!refcount_release(&vp->v_holdcnt)) {
3609 		VI_UNLOCK(vp);
3610 		return;
3611 	}
3612 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3613 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3614 	if (VN_IS_DOOMED(vp)) {
3615 		vdropl_final(vp);
3616 		return;
3617 	}
3618 
3619 	vfs_freevnodes_inc();
3620 	if (vp->v_mflag & VMP_LAZYLIST) {
3621 		vunlazy(vp);
3622 	}
3623 	/*
3624 	 * Also unlocks the interlock. We can't assert on it as we
3625 	 * released our hold and by now the vnode might have been
3626 	 * freed.
3627 	 */
3628 	vdbatch_enqueue(vp);
3629 }
3630 
3631 /*
3632  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3633  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3634  */
3635 static int
3636 vinactivef(struct vnode *vp)
3637 {
3638 	struct vm_object *obj;
3639 	int error;
3640 
3641 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3642 	ASSERT_VI_LOCKED(vp, "vinactive");
3643 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3644 	    ("vinactive: recursed on VI_DOINGINACT"));
3645 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3646 	vp->v_iflag |= VI_DOINGINACT;
3647 	vp->v_iflag &= ~VI_OWEINACT;
3648 	VI_UNLOCK(vp);
3649 	/*
3650 	 * Before moving off the active list, we must be sure that any
3651 	 * modified pages are converted into the vnode's dirty
3652 	 * buffers, since these will no longer be checked once the
3653 	 * vnode is on the inactive list.
3654 	 *
3655 	 * The write-out of the dirty pages is asynchronous.  At the
3656 	 * point that VOP_INACTIVE() is called, there could still be
3657 	 * pending I/O and dirty pages in the object.
3658 	 */
3659 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3660 	    vm_object_mightbedirty(obj)) {
3661 		VM_OBJECT_WLOCK(obj);
3662 		vm_object_page_clean(obj, 0, 0, 0);
3663 		VM_OBJECT_WUNLOCK(obj);
3664 	}
3665 	error = VOP_INACTIVE(vp);
3666 	VI_LOCK(vp);
3667 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3668 	    ("vinactive: lost VI_DOINGINACT"));
3669 	vp->v_iflag &= ~VI_DOINGINACT;
3670 	return (error);
3671 }
3672 
3673 int
3674 vinactive(struct vnode *vp)
3675 {
3676 
3677 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3678 	ASSERT_VI_LOCKED(vp, "vinactive");
3679 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3680 
3681 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3682 		return (0);
3683 	if (vp->v_iflag & VI_DOINGINACT)
3684 		return (0);
3685 	if (vp->v_usecount > 0) {
3686 		vp->v_iflag &= ~VI_OWEINACT;
3687 		return (0);
3688 	}
3689 	return (vinactivef(vp));
3690 }
3691 
3692 /*
3693  * Remove any vnodes in the vnode table belonging to mount point mp.
3694  *
3695  * If FORCECLOSE is not specified, there should not be any active ones,
3696  * return error if any are found (nb: this is a user error, not a
3697  * system error). If FORCECLOSE is specified, detach any active vnodes
3698  * that are found.
3699  *
3700  * If WRITECLOSE is set, only flush out regular file vnodes open for
3701  * writing.
3702  *
3703  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3704  *
3705  * `rootrefs' specifies the base reference count for the root vnode
3706  * of this filesystem. The root vnode is considered busy if its
3707  * v_usecount exceeds this value. On a successful return, vflush(, td)
3708  * will call vrele() on the root vnode exactly rootrefs times.
3709  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3710  * be zero.
3711  */
3712 #ifdef DIAGNOSTIC
3713 static int busyprt = 0;		/* print out busy vnodes */
3714 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3715 #endif
3716 
3717 int
3718 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3719 {
3720 	struct vnode *vp, *mvp, *rootvp = NULL;
3721 	struct vattr vattr;
3722 	int busy = 0, error;
3723 
3724 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3725 	    rootrefs, flags);
3726 	if (rootrefs > 0) {
3727 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3728 		    ("vflush: bad args"));
3729 		/*
3730 		 * Get the filesystem root vnode. We can vput() it
3731 		 * immediately, since with rootrefs > 0, it won't go away.
3732 		 */
3733 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3734 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3735 			    __func__, error);
3736 			return (error);
3737 		}
3738 		vput(rootvp);
3739 	}
3740 loop:
3741 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3742 		vholdl(vp);
3743 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3744 		if (error) {
3745 			vdrop(vp);
3746 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3747 			goto loop;
3748 		}
3749 		/*
3750 		 * Skip over a vnodes marked VV_SYSTEM.
3751 		 */
3752 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3753 			VOP_UNLOCK(vp);
3754 			vdrop(vp);
3755 			continue;
3756 		}
3757 		/*
3758 		 * If WRITECLOSE is set, flush out unlinked but still open
3759 		 * files (even if open only for reading) and regular file
3760 		 * vnodes open for writing.
3761 		 */
3762 		if (flags & WRITECLOSE) {
3763 			if (vp->v_object != NULL) {
3764 				VM_OBJECT_WLOCK(vp->v_object);
3765 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3766 				VM_OBJECT_WUNLOCK(vp->v_object);
3767 			}
3768 			do {
3769 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3770 			} while (error == ERELOOKUP);
3771 			if (error != 0) {
3772 				VOP_UNLOCK(vp);
3773 				vdrop(vp);
3774 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3775 				return (error);
3776 			}
3777 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3778 			VI_LOCK(vp);
3779 
3780 			if ((vp->v_type == VNON ||
3781 			    (error == 0 && vattr.va_nlink > 0)) &&
3782 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3783 				VOP_UNLOCK(vp);
3784 				vdropl(vp);
3785 				continue;
3786 			}
3787 		} else
3788 			VI_LOCK(vp);
3789 		/*
3790 		 * With v_usecount == 0, all we need to do is clear out the
3791 		 * vnode data structures and we are done.
3792 		 *
3793 		 * If FORCECLOSE is set, forcibly close the vnode.
3794 		 */
3795 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3796 			vgonel(vp);
3797 		} else {
3798 			busy++;
3799 #ifdef DIAGNOSTIC
3800 			if (busyprt)
3801 				vn_printf(vp, "vflush: busy vnode ");
3802 #endif
3803 		}
3804 		VOP_UNLOCK(vp);
3805 		vdropl(vp);
3806 	}
3807 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3808 		/*
3809 		 * If just the root vnode is busy, and if its refcount
3810 		 * is equal to `rootrefs', then go ahead and kill it.
3811 		 */
3812 		VI_LOCK(rootvp);
3813 		KASSERT(busy > 0, ("vflush: not busy"));
3814 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3815 		    ("vflush: usecount %d < rootrefs %d",
3816 		     rootvp->v_usecount, rootrefs));
3817 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3818 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3819 			vgone(rootvp);
3820 			VOP_UNLOCK(rootvp);
3821 			busy = 0;
3822 		} else
3823 			VI_UNLOCK(rootvp);
3824 	}
3825 	if (busy) {
3826 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3827 		    busy);
3828 		return (EBUSY);
3829 	}
3830 	for (; rootrefs > 0; rootrefs--)
3831 		vrele(rootvp);
3832 	return (0);
3833 }
3834 
3835 /*
3836  * Recycle an unused vnode to the front of the free list.
3837  */
3838 int
3839 vrecycle(struct vnode *vp)
3840 {
3841 	int recycled;
3842 
3843 	VI_LOCK(vp);
3844 	recycled = vrecyclel(vp);
3845 	VI_UNLOCK(vp);
3846 	return (recycled);
3847 }
3848 
3849 /*
3850  * vrecycle, with the vp interlock held.
3851  */
3852 int
3853 vrecyclel(struct vnode *vp)
3854 {
3855 	int recycled;
3856 
3857 	ASSERT_VOP_ELOCKED(vp, __func__);
3858 	ASSERT_VI_LOCKED(vp, __func__);
3859 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3860 	recycled = 0;
3861 	if (vp->v_usecount == 0) {
3862 		recycled = 1;
3863 		vgonel(vp);
3864 	}
3865 	return (recycled);
3866 }
3867 
3868 /*
3869  * Eliminate all activity associated with a vnode
3870  * in preparation for reuse.
3871  */
3872 void
3873 vgone(struct vnode *vp)
3874 {
3875 	VI_LOCK(vp);
3876 	vgonel(vp);
3877 	VI_UNLOCK(vp);
3878 }
3879 
3880 /*
3881  * Notify upper mounts about reclaimed or unlinked vnode.
3882  */
3883 void
3884 vfs_notify_upper(struct vnode *vp, int event)
3885 {
3886 	struct mount *mp;
3887 	struct mount_upper_node *ump;
3888 
3889 	mp = atomic_load_ptr(&vp->v_mount);
3890 	if (mp == NULL)
3891 		return;
3892 	if (TAILQ_EMPTY(&mp->mnt_notify))
3893 		return;
3894 
3895 	MNT_ILOCK(mp);
3896 	mp->mnt_upper_pending++;
3897 	KASSERT(mp->mnt_upper_pending > 0,
3898 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3899 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
3900 		MNT_IUNLOCK(mp);
3901 		switch (event) {
3902 		case VFS_NOTIFY_UPPER_RECLAIM:
3903 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
3904 			break;
3905 		case VFS_NOTIFY_UPPER_UNLINK:
3906 			VFS_UNLINK_LOWERVP(ump->mp, vp);
3907 			break;
3908 		default:
3909 			KASSERT(0, ("invalid event %d", event));
3910 			break;
3911 		}
3912 		MNT_ILOCK(mp);
3913 	}
3914 	mp->mnt_upper_pending--;
3915 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
3916 	    mp->mnt_upper_pending == 0) {
3917 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
3918 		wakeup(&mp->mnt_uppers);
3919 	}
3920 	MNT_IUNLOCK(mp);
3921 }
3922 
3923 /*
3924  * vgone, with the vp interlock held.
3925  */
3926 static void
3927 vgonel(struct vnode *vp)
3928 {
3929 	struct thread *td;
3930 	struct mount *mp;
3931 	vm_object_t object;
3932 	bool active, doinginact, oweinact;
3933 
3934 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3935 	ASSERT_VI_LOCKED(vp, "vgonel");
3936 	VNASSERT(vp->v_holdcnt, vp,
3937 	    ("vgonel: vp %p has no reference.", vp));
3938 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3939 	td = curthread;
3940 
3941 	/*
3942 	 * Don't vgonel if we're already doomed.
3943 	 */
3944 	if (VN_IS_DOOMED(vp))
3945 		return;
3946 	/*
3947 	 * Paired with freevnode.
3948 	 */
3949 	vn_seqc_write_begin_locked(vp);
3950 	vunlazy_gone(vp);
3951 	vn_irflag_set_locked(vp, VIRF_DOOMED);
3952 
3953 	/*
3954 	 * Check to see if the vnode is in use.  If so, we have to
3955 	 * call VOP_CLOSE() and VOP_INACTIVE().
3956 	 *
3957 	 * It could be that VOP_INACTIVE() requested reclamation, in
3958 	 * which case we should avoid recursion, so check
3959 	 * VI_DOINGINACT.  This is not precise but good enough.
3960 	 */
3961 	active = vp->v_usecount > 0;
3962 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3963 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3964 
3965 	/*
3966 	 * If we need to do inactive VI_OWEINACT will be set.
3967 	 */
3968 	if (vp->v_iflag & VI_DEFINACT) {
3969 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3970 		vp->v_iflag &= ~VI_DEFINACT;
3971 		vdropl(vp);
3972 	} else {
3973 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3974 		VI_UNLOCK(vp);
3975 	}
3976 	cache_purge_vgone(vp);
3977 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3978 
3979 	/*
3980 	 * If purging an active vnode, it must be closed and
3981 	 * deactivated before being reclaimed.
3982 	 */
3983 	if (active)
3984 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3985 	if (!doinginact) {
3986 		do {
3987 			if (oweinact || active) {
3988 				VI_LOCK(vp);
3989 				vinactivef(vp);
3990 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3991 				VI_UNLOCK(vp);
3992 			}
3993 		} while (oweinact);
3994 	}
3995 	if (vp->v_type == VSOCK)
3996 		vfs_unp_reclaim(vp);
3997 
3998 	/*
3999 	 * Clean out any buffers associated with the vnode.
4000 	 * If the flush fails, just toss the buffers.
4001 	 */
4002 	mp = NULL;
4003 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4004 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4005 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4006 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4007 			;
4008 	}
4009 
4010 	BO_LOCK(&vp->v_bufobj);
4011 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4012 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4013 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4014 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4015 	    ("vp %p bufobj not invalidated", vp));
4016 
4017 	/*
4018 	 * For VMIO bufobj, BO_DEAD is set later, or in
4019 	 * vm_object_terminate() after the object's page queue is
4020 	 * flushed.
4021 	 */
4022 	object = vp->v_bufobj.bo_object;
4023 	if (object == NULL)
4024 		vp->v_bufobj.bo_flag |= BO_DEAD;
4025 	BO_UNLOCK(&vp->v_bufobj);
4026 
4027 	/*
4028 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4029 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4030 	 * should not touch the object borrowed from the lower vnode
4031 	 * (the handle check).
4032 	 */
4033 	if (object != NULL && object->type == OBJT_VNODE &&
4034 	    object->handle == vp)
4035 		vnode_destroy_vobject(vp);
4036 
4037 	/*
4038 	 * Reclaim the vnode.
4039 	 */
4040 	if (VOP_RECLAIM(vp))
4041 		panic("vgone: cannot reclaim");
4042 	if (mp != NULL)
4043 		vn_finished_secondary_write(mp);
4044 	VNASSERT(vp->v_object == NULL, vp,
4045 	    ("vop_reclaim left v_object vp=%p", vp));
4046 	/*
4047 	 * Clear the advisory locks and wake up waiting threads.
4048 	 */
4049 	(void)VOP_ADVLOCKPURGE(vp);
4050 	vp->v_lockf = NULL;
4051 	/*
4052 	 * Delete from old mount point vnode list.
4053 	 */
4054 	delmntque(vp);
4055 	/*
4056 	 * Done with purge, reset to the standard lock and invalidate
4057 	 * the vnode.
4058 	 */
4059 	VI_LOCK(vp);
4060 	vp->v_vnlock = &vp->v_lock;
4061 	vp->v_op = &dead_vnodeops;
4062 	vp->v_type = VBAD;
4063 }
4064 
4065 /*
4066  * Print out a description of a vnode.
4067  */
4068 static const char * const typename[] =
4069 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4070  "VMARKER"};
4071 
4072 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4073     "new hold count flag not added to vn_printf");
4074 
4075 void
4076 vn_printf(struct vnode *vp, const char *fmt, ...)
4077 {
4078 	va_list ap;
4079 	char buf[256], buf2[16];
4080 	u_long flags;
4081 	u_int holdcnt;
4082 	short irflag;
4083 
4084 	va_start(ap, fmt);
4085 	vprintf(fmt, ap);
4086 	va_end(ap);
4087 	printf("%p: ", (void *)vp);
4088 	printf("type %s\n", typename[vp->v_type]);
4089 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4090 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4091 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4092 	    vp->v_seqc_users);
4093 	switch (vp->v_type) {
4094 	case VDIR:
4095 		printf(" mountedhere %p\n", vp->v_mountedhere);
4096 		break;
4097 	case VCHR:
4098 		printf(" rdev %p\n", vp->v_rdev);
4099 		break;
4100 	case VSOCK:
4101 		printf(" socket %p\n", vp->v_unpcb);
4102 		break;
4103 	case VFIFO:
4104 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4105 		break;
4106 	default:
4107 		printf("\n");
4108 		break;
4109 	}
4110 	buf[0] = '\0';
4111 	buf[1] = '\0';
4112 	if (holdcnt & VHOLD_NO_SMR)
4113 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4114 	printf("    hold count flags (%s)\n", buf + 1);
4115 
4116 	buf[0] = '\0';
4117 	buf[1] = '\0';
4118 	irflag = vn_irflag_read(vp);
4119 	if (irflag & VIRF_DOOMED)
4120 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4121 	if (irflag & VIRF_PGREAD)
4122 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4123 	if (irflag & VIRF_MOUNTPOINT)
4124 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4125 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT);
4126 	if (flags != 0) {
4127 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4128 		strlcat(buf, buf2, sizeof(buf));
4129 	}
4130 	if (vp->v_vflag & VV_ROOT)
4131 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4132 	if (vp->v_vflag & VV_ISTTY)
4133 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4134 	if (vp->v_vflag & VV_NOSYNC)
4135 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4136 	if (vp->v_vflag & VV_ETERNALDEV)
4137 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4138 	if (vp->v_vflag & VV_CACHEDLABEL)
4139 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4140 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4141 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4142 	if (vp->v_vflag & VV_COPYONWRITE)
4143 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4144 	if (vp->v_vflag & VV_SYSTEM)
4145 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4146 	if (vp->v_vflag & VV_PROCDEP)
4147 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4148 	if (vp->v_vflag & VV_NOKNOTE)
4149 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4150 	if (vp->v_vflag & VV_DELETED)
4151 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4152 	if (vp->v_vflag & VV_MD)
4153 		strlcat(buf, "|VV_MD", sizeof(buf));
4154 	if (vp->v_vflag & VV_FORCEINSMQ)
4155 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4156 	if (vp->v_vflag & VV_READLINK)
4157 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4158 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4159 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4160 	    VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ |
4161 	    VV_READLINK);
4162 	if (flags != 0) {
4163 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4164 		strlcat(buf, buf2, sizeof(buf));
4165 	}
4166 	if (vp->v_iflag & VI_TEXT_REF)
4167 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4168 	if (vp->v_iflag & VI_MOUNT)
4169 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4170 	if (vp->v_iflag & VI_DOINGINACT)
4171 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4172 	if (vp->v_iflag & VI_OWEINACT)
4173 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4174 	if (vp->v_iflag & VI_DEFINACT)
4175 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4176 	if (vp->v_iflag & VI_FOPENING)
4177 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4178 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4179 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4180 	if (flags != 0) {
4181 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4182 		strlcat(buf, buf2, sizeof(buf));
4183 	}
4184 	if (vp->v_mflag & VMP_LAZYLIST)
4185 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4186 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4187 	if (flags != 0) {
4188 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4189 		strlcat(buf, buf2, sizeof(buf));
4190 	}
4191 	printf("    flags (%s)", buf + 1);
4192 	if (mtx_owned(VI_MTX(vp)))
4193 		printf(" VI_LOCKed");
4194 	printf("\n");
4195 	if (vp->v_object != NULL)
4196 		printf("    v_object %p ref %d pages %d "
4197 		    "cleanbuf %d dirtybuf %d\n",
4198 		    vp->v_object, vp->v_object->ref_count,
4199 		    vp->v_object->resident_page_count,
4200 		    vp->v_bufobj.bo_clean.bv_cnt,
4201 		    vp->v_bufobj.bo_dirty.bv_cnt);
4202 	printf("    ");
4203 	lockmgr_printinfo(vp->v_vnlock);
4204 	if (vp->v_data != NULL)
4205 		VOP_PRINT(vp);
4206 }
4207 
4208 #ifdef DDB
4209 /*
4210  * List all of the locked vnodes in the system.
4211  * Called when debugging the kernel.
4212  */
4213 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4214 {
4215 	struct mount *mp;
4216 	struct vnode *vp;
4217 
4218 	/*
4219 	 * Note: because this is DDB, we can't obey the locking semantics
4220 	 * for these structures, which means we could catch an inconsistent
4221 	 * state and dereference a nasty pointer.  Not much to be done
4222 	 * about that.
4223 	 */
4224 	db_printf("Locked vnodes\n");
4225 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4226 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4227 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4228 				vn_printf(vp, "vnode ");
4229 		}
4230 	}
4231 }
4232 
4233 /*
4234  * Show details about the given vnode.
4235  */
4236 DB_SHOW_COMMAND(vnode, db_show_vnode)
4237 {
4238 	struct vnode *vp;
4239 
4240 	if (!have_addr)
4241 		return;
4242 	vp = (struct vnode *)addr;
4243 	vn_printf(vp, "vnode ");
4244 }
4245 
4246 /*
4247  * Show details about the given mount point.
4248  */
4249 DB_SHOW_COMMAND(mount, db_show_mount)
4250 {
4251 	struct mount *mp;
4252 	struct vfsopt *opt;
4253 	struct statfs *sp;
4254 	struct vnode *vp;
4255 	char buf[512];
4256 	uint64_t mflags;
4257 	u_int flags;
4258 
4259 	if (!have_addr) {
4260 		/* No address given, print short info about all mount points. */
4261 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4262 			db_printf("%p %s on %s (%s)\n", mp,
4263 			    mp->mnt_stat.f_mntfromname,
4264 			    mp->mnt_stat.f_mntonname,
4265 			    mp->mnt_stat.f_fstypename);
4266 			if (db_pager_quit)
4267 				break;
4268 		}
4269 		db_printf("\nMore info: show mount <addr>\n");
4270 		return;
4271 	}
4272 
4273 	mp = (struct mount *)addr;
4274 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4275 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4276 
4277 	buf[0] = '\0';
4278 	mflags = mp->mnt_flag;
4279 #define	MNT_FLAG(flag)	do {						\
4280 	if (mflags & (flag)) {						\
4281 		if (buf[0] != '\0')					\
4282 			strlcat(buf, ", ", sizeof(buf));		\
4283 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4284 		mflags &= ~(flag);					\
4285 	}								\
4286 } while (0)
4287 	MNT_FLAG(MNT_RDONLY);
4288 	MNT_FLAG(MNT_SYNCHRONOUS);
4289 	MNT_FLAG(MNT_NOEXEC);
4290 	MNT_FLAG(MNT_NOSUID);
4291 	MNT_FLAG(MNT_NFS4ACLS);
4292 	MNT_FLAG(MNT_UNION);
4293 	MNT_FLAG(MNT_ASYNC);
4294 	MNT_FLAG(MNT_SUIDDIR);
4295 	MNT_FLAG(MNT_SOFTDEP);
4296 	MNT_FLAG(MNT_NOSYMFOLLOW);
4297 	MNT_FLAG(MNT_GJOURNAL);
4298 	MNT_FLAG(MNT_MULTILABEL);
4299 	MNT_FLAG(MNT_ACLS);
4300 	MNT_FLAG(MNT_NOATIME);
4301 	MNT_FLAG(MNT_NOCLUSTERR);
4302 	MNT_FLAG(MNT_NOCLUSTERW);
4303 	MNT_FLAG(MNT_SUJ);
4304 	MNT_FLAG(MNT_EXRDONLY);
4305 	MNT_FLAG(MNT_EXPORTED);
4306 	MNT_FLAG(MNT_DEFEXPORTED);
4307 	MNT_FLAG(MNT_EXPORTANON);
4308 	MNT_FLAG(MNT_EXKERB);
4309 	MNT_FLAG(MNT_EXPUBLIC);
4310 	MNT_FLAG(MNT_LOCAL);
4311 	MNT_FLAG(MNT_QUOTA);
4312 	MNT_FLAG(MNT_ROOTFS);
4313 	MNT_FLAG(MNT_USER);
4314 	MNT_FLAG(MNT_IGNORE);
4315 	MNT_FLAG(MNT_UPDATE);
4316 	MNT_FLAG(MNT_DELEXPORT);
4317 	MNT_FLAG(MNT_RELOAD);
4318 	MNT_FLAG(MNT_FORCE);
4319 	MNT_FLAG(MNT_SNAPSHOT);
4320 	MNT_FLAG(MNT_BYFSID);
4321 #undef MNT_FLAG
4322 	if (mflags != 0) {
4323 		if (buf[0] != '\0')
4324 			strlcat(buf, ", ", sizeof(buf));
4325 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4326 		    "0x%016jx", mflags);
4327 	}
4328 	db_printf("    mnt_flag = %s\n", buf);
4329 
4330 	buf[0] = '\0';
4331 	flags = mp->mnt_kern_flag;
4332 #define	MNT_KERN_FLAG(flag)	do {					\
4333 	if (flags & (flag)) {						\
4334 		if (buf[0] != '\0')					\
4335 			strlcat(buf, ", ", sizeof(buf));		\
4336 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4337 		flags &= ~(flag);					\
4338 	}								\
4339 } while (0)
4340 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4341 	MNT_KERN_FLAG(MNTK_ASYNC);
4342 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4343 	MNT_KERN_FLAG(MNTK_DRAINING);
4344 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4345 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4346 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4347 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4348 	MNT_KERN_FLAG(MNTK_RECURSE);
4349 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4350 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4351 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4352 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4353 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4354 	MNT_KERN_FLAG(MNTK_NOASYNC);
4355 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4356 	MNT_KERN_FLAG(MNTK_MWAIT);
4357 	MNT_KERN_FLAG(MNTK_SUSPEND);
4358 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4359 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4360 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4361 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4362 #undef MNT_KERN_FLAG
4363 	if (flags != 0) {
4364 		if (buf[0] != '\0')
4365 			strlcat(buf, ", ", sizeof(buf));
4366 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4367 		    "0x%08x", flags);
4368 	}
4369 	db_printf("    mnt_kern_flag = %s\n", buf);
4370 
4371 	db_printf("    mnt_opt = ");
4372 	opt = TAILQ_FIRST(mp->mnt_opt);
4373 	if (opt != NULL) {
4374 		db_printf("%s", opt->name);
4375 		opt = TAILQ_NEXT(opt, link);
4376 		while (opt != NULL) {
4377 			db_printf(", %s", opt->name);
4378 			opt = TAILQ_NEXT(opt, link);
4379 		}
4380 	}
4381 	db_printf("\n");
4382 
4383 	sp = &mp->mnt_stat;
4384 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4385 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4386 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4387 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4388 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4389 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4390 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4391 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4392 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4393 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4394 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4395 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4396 
4397 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4398 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4399 	if (jailed(mp->mnt_cred))
4400 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4401 	db_printf(" }\n");
4402 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4403 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4404 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4405 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4406 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4407 	    mp->mnt_lazyvnodelistsize);
4408 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4409 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4410 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4411 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4412 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4413 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4414 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4415 	db_printf("    mnt_secondary_accwrites = %d\n",
4416 	    mp->mnt_secondary_accwrites);
4417 	db_printf("    mnt_gjprovider = %s\n",
4418 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4419 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4420 
4421 	db_printf("\n\nList of active vnodes\n");
4422 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4423 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4424 			vn_printf(vp, "vnode ");
4425 			if (db_pager_quit)
4426 				break;
4427 		}
4428 	}
4429 	db_printf("\n\nList of inactive vnodes\n");
4430 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4431 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4432 			vn_printf(vp, "vnode ");
4433 			if (db_pager_quit)
4434 				break;
4435 		}
4436 	}
4437 }
4438 #endif	/* DDB */
4439 
4440 /*
4441  * Fill in a struct xvfsconf based on a struct vfsconf.
4442  */
4443 static int
4444 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4445 {
4446 	struct xvfsconf xvfsp;
4447 
4448 	bzero(&xvfsp, sizeof(xvfsp));
4449 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4450 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4451 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4452 	xvfsp.vfc_flags = vfsp->vfc_flags;
4453 	/*
4454 	 * These are unused in userland, we keep them
4455 	 * to not break binary compatibility.
4456 	 */
4457 	xvfsp.vfc_vfsops = NULL;
4458 	xvfsp.vfc_next = NULL;
4459 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4460 }
4461 
4462 #ifdef COMPAT_FREEBSD32
4463 struct xvfsconf32 {
4464 	uint32_t	vfc_vfsops;
4465 	char		vfc_name[MFSNAMELEN];
4466 	int32_t		vfc_typenum;
4467 	int32_t		vfc_refcount;
4468 	int32_t		vfc_flags;
4469 	uint32_t	vfc_next;
4470 };
4471 
4472 static int
4473 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4474 {
4475 	struct xvfsconf32 xvfsp;
4476 
4477 	bzero(&xvfsp, sizeof(xvfsp));
4478 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4479 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4480 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4481 	xvfsp.vfc_flags = vfsp->vfc_flags;
4482 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4483 }
4484 #endif
4485 
4486 /*
4487  * Top level filesystem related information gathering.
4488  */
4489 static int
4490 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4491 {
4492 	struct vfsconf *vfsp;
4493 	int error;
4494 
4495 	error = 0;
4496 	vfsconf_slock();
4497 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4498 #ifdef COMPAT_FREEBSD32
4499 		if (req->flags & SCTL_MASK32)
4500 			error = vfsconf2x32(req, vfsp);
4501 		else
4502 #endif
4503 			error = vfsconf2x(req, vfsp);
4504 		if (error)
4505 			break;
4506 	}
4507 	vfsconf_sunlock();
4508 	return (error);
4509 }
4510 
4511 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4512     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4513     "S,xvfsconf", "List of all configured filesystems");
4514 
4515 #ifndef BURN_BRIDGES
4516 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4517 
4518 static int
4519 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4520 {
4521 	int *name = (int *)arg1 - 1;	/* XXX */
4522 	u_int namelen = arg2 + 1;	/* XXX */
4523 	struct vfsconf *vfsp;
4524 
4525 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4526 	    "please rebuild world\n");
4527 
4528 #if 1 || defined(COMPAT_PRELITE2)
4529 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4530 	if (namelen == 1)
4531 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4532 #endif
4533 
4534 	switch (name[1]) {
4535 	case VFS_MAXTYPENUM:
4536 		if (namelen != 2)
4537 			return (ENOTDIR);
4538 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4539 	case VFS_CONF:
4540 		if (namelen != 3)
4541 			return (ENOTDIR);	/* overloaded */
4542 		vfsconf_slock();
4543 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4544 			if (vfsp->vfc_typenum == name[2])
4545 				break;
4546 		}
4547 		vfsconf_sunlock();
4548 		if (vfsp == NULL)
4549 			return (EOPNOTSUPP);
4550 #ifdef COMPAT_FREEBSD32
4551 		if (req->flags & SCTL_MASK32)
4552 			return (vfsconf2x32(req, vfsp));
4553 		else
4554 #endif
4555 			return (vfsconf2x(req, vfsp));
4556 	}
4557 	return (EOPNOTSUPP);
4558 }
4559 
4560 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4561     CTLFLAG_MPSAFE, vfs_sysctl,
4562     "Generic filesystem");
4563 
4564 #if 1 || defined(COMPAT_PRELITE2)
4565 
4566 static int
4567 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4568 {
4569 	int error;
4570 	struct vfsconf *vfsp;
4571 	struct ovfsconf ovfs;
4572 
4573 	vfsconf_slock();
4574 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4575 		bzero(&ovfs, sizeof(ovfs));
4576 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4577 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4578 		ovfs.vfc_index = vfsp->vfc_typenum;
4579 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4580 		ovfs.vfc_flags = vfsp->vfc_flags;
4581 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4582 		if (error != 0) {
4583 			vfsconf_sunlock();
4584 			return (error);
4585 		}
4586 	}
4587 	vfsconf_sunlock();
4588 	return (0);
4589 }
4590 
4591 #endif /* 1 || COMPAT_PRELITE2 */
4592 #endif /* !BURN_BRIDGES */
4593 
4594 #define KINFO_VNODESLOP		10
4595 #ifdef notyet
4596 /*
4597  * Dump vnode list (via sysctl).
4598  */
4599 /* ARGSUSED */
4600 static int
4601 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4602 {
4603 	struct xvnode *xvn;
4604 	struct mount *mp;
4605 	struct vnode *vp;
4606 	int error, len, n;
4607 
4608 	/*
4609 	 * Stale numvnodes access is not fatal here.
4610 	 */
4611 	req->lock = 0;
4612 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4613 	if (!req->oldptr)
4614 		/* Make an estimate */
4615 		return (SYSCTL_OUT(req, 0, len));
4616 
4617 	error = sysctl_wire_old_buffer(req, 0);
4618 	if (error != 0)
4619 		return (error);
4620 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4621 	n = 0;
4622 	mtx_lock(&mountlist_mtx);
4623 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4624 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4625 			continue;
4626 		MNT_ILOCK(mp);
4627 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4628 			if (n == len)
4629 				break;
4630 			vref(vp);
4631 			xvn[n].xv_size = sizeof *xvn;
4632 			xvn[n].xv_vnode = vp;
4633 			xvn[n].xv_id = 0;	/* XXX compat */
4634 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4635 			XV_COPY(usecount);
4636 			XV_COPY(writecount);
4637 			XV_COPY(holdcnt);
4638 			XV_COPY(mount);
4639 			XV_COPY(numoutput);
4640 			XV_COPY(type);
4641 #undef XV_COPY
4642 			xvn[n].xv_flag = vp->v_vflag;
4643 
4644 			switch (vp->v_type) {
4645 			case VREG:
4646 			case VDIR:
4647 			case VLNK:
4648 				break;
4649 			case VBLK:
4650 			case VCHR:
4651 				if (vp->v_rdev == NULL) {
4652 					vrele(vp);
4653 					continue;
4654 				}
4655 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4656 				break;
4657 			case VSOCK:
4658 				xvn[n].xv_socket = vp->v_socket;
4659 				break;
4660 			case VFIFO:
4661 				xvn[n].xv_fifo = vp->v_fifoinfo;
4662 				break;
4663 			case VNON:
4664 			case VBAD:
4665 			default:
4666 				/* shouldn't happen? */
4667 				vrele(vp);
4668 				continue;
4669 			}
4670 			vrele(vp);
4671 			++n;
4672 		}
4673 		MNT_IUNLOCK(mp);
4674 		mtx_lock(&mountlist_mtx);
4675 		vfs_unbusy(mp);
4676 		if (n == len)
4677 			break;
4678 	}
4679 	mtx_unlock(&mountlist_mtx);
4680 
4681 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4682 	free(xvn, M_TEMP);
4683 	return (error);
4684 }
4685 
4686 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4687     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4688     "");
4689 #endif
4690 
4691 static void
4692 unmount_or_warn(struct mount *mp)
4693 {
4694 	int error;
4695 
4696 	error = dounmount(mp, MNT_FORCE, curthread);
4697 	if (error != 0) {
4698 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4699 		if (error == EBUSY)
4700 			printf("BUSY)\n");
4701 		else
4702 			printf("%d)\n", error);
4703 	}
4704 }
4705 
4706 /*
4707  * Unmount all filesystems. The list is traversed in reverse order
4708  * of mounting to avoid dependencies.
4709  */
4710 void
4711 vfs_unmountall(void)
4712 {
4713 	struct mount *mp, *tmp;
4714 
4715 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4716 
4717 	/*
4718 	 * Since this only runs when rebooting, it is not interlocked.
4719 	 */
4720 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4721 		vfs_ref(mp);
4722 
4723 		/*
4724 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4725 		 * unmount of the latter.
4726 		 */
4727 		if (mp == rootdevmp)
4728 			continue;
4729 
4730 		unmount_or_warn(mp);
4731 	}
4732 
4733 	if (rootdevmp != NULL)
4734 		unmount_or_warn(rootdevmp);
4735 }
4736 
4737 static void
4738 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4739 {
4740 
4741 	ASSERT_VI_LOCKED(vp, __func__);
4742 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4743 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4744 		vdropl(vp);
4745 		return;
4746 	}
4747 	if (vn_lock(vp, lkflags) == 0) {
4748 		VI_LOCK(vp);
4749 		vinactive(vp);
4750 		VOP_UNLOCK(vp);
4751 		vdropl(vp);
4752 		return;
4753 	}
4754 	vdefer_inactive_unlocked(vp);
4755 }
4756 
4757 static int
4758 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4759 {
4760 
4761 	return (vp->v_iflag & VI_DEFINACT);
4762 }
4763 
4764 static void __noinline
4765 vfs_periodic_inactive(struct mount *mp, int flags)
4766 {
4767 	struct vnode *vp, *mvp;
4768 	int lkflags;
4769 
4770 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4771 	if (flags != MNT_WAIT)
4772 		lkflags |= LK_NOWAIT;
4773 
4774 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4775 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4776 			VI_UNLOCK(vp);
4777 			continue;
4778 		}
4779 		vp->v_iflag &= ~VI_DEFINACT;
4780 		vfs_deferred_inactive(vp, lkflags);
4781 	}
4782 }
4783 
4784 static inline bool
4785 vfs_want_msync(struct vnode *vp)
4786 {
4787 	struct vm_object *obj;
4788 
4789 	/*
4790 	 * This test may be performed without any locks held.
4791 	 * We rely on vm_object's type stability.
4792 	 */
4793 	if (vp->v_vflag & VV_NOSYNC)
4794 		return (false);
4795 	obj = vp->v_object;
4796 	return (obj != NULL && vm_object_mightbedirty(obj));
4797 }
4798 
4799 static int
4800 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4801 {
4802 
4803 	if (vp->v_vflag & VV_NOSYNC)
4804 		return (false);
4805 	if (vp->v_iflag & VI_DEFINACT)
4806 		return (true);
4807 	return (vfs_want_msync(vp));
4808 }
4809 
4810 static void __noinline
4811 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4812 {
4813 	struct vnode *vp, *mvp;
4814 	struct vm_object *obj;
4815 	int lkflags, objflags;
4816 	bool seen_defer;
4817 
4818 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4819 	if (flags != MNT_WAIT) {
4820 		lkflags |= LK_NOWAIT;
4821 		objflags = OBJPC_NOSYNC;
4822 	} else {
4823 		objflags = OBJPC_SYNC;
4824 	}
4825 
4826 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4827 		seen_defer = false;
4828 		if (vp->v_iflag & VI_DEFINACT) {
4829 			vp->v_iflag &= ~VI_DEFINACT;
4830 			seen_defer = true;
4831 		}
4832 		if (!vfs_want_msync(vp)) {
4833 			if (seen_defer)
4834 				vfs_deferred_inactive(vp, lkflags);
4835 			else
4836 				VI_UNLOCK(vp);
4837 			continue;
4838 		}
4839 		if (vget(vp, lkflags) == 0) {
4840 			obj = vp->v_object;
4841 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4842 				VM_OBJECT_WLOCK(obj);
4843 				vm_object_page_clean(obj, 0, 0, objflags);
4844 				VM_OBJECT_WUNLOCK(obj);
4845 			}
4846 			vput(vp);
4847 			if (seen_defer)
4848 				vdrop(vp);
4849 		} else {
4850 			if (seen_defer)
4851 				vdefer_inactive_unlocked(vp);
4852 		}
4853 	}
4854 }
4855 
4856 void
4857 vfs_periodic(struct mount *mp, int flags)
4858 {
4859 
4860 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4861 
4862 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4863 		vfs_periodic_inactive(mp, flags);
4864 	else
4865 		vfs_periodic_msync_inactive(mp, flags);
4866 }
4867 
4868 static void
4869 destroy_vpollinfo_free(struct vpollinfo *vi)
4870 {
4871 
4872 	knlist_destroy(&vi->vpi_selinfo.si_note);
4873 	mtx_destroy(&vi->vpi_lock);
4874 	free(vi, M_VNODEPOLL);
4875 }
4876 
4877 static void
4878 destroy_vpollinfo(struct vpollinfo *vi)
4879 {
4880 
4881 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4882 	seldrain(&vi->vpi_selinfo);
4883 	destroy_vpollinfo_free(vi);
4884 }
4885 
4886 /*
4887  * Initialize per-vnode helper structure to hold poll-related state.
4888  */
4889 void
4890 v_addpollinfo(struct vnode *vp)
4891 {
4892 	struct vpollinfo *vi;
4893 
4894 	if (vp->v_pollinfo != NULL)
4895 		return;
4896 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4897 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4898 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4899 	    vfs_knlunlock, vfs_knl_assert_lock);
4900 	VI_LOCK(vp);
4901 	if (vp->v_pollinfo != NULL) {
4902 		VI_UNLOCK(vp);
4903 		destroy_vpollinfo_free(vi);
4904 		return;
4905 	}
4906 	vp->v_pollinfo = vi;
4907 	VI_UNLOCK(vp);
4908 }
4909 
4910 /*
4911  * Record a process's interest in events which might happen to
4912  * a vnode.  Because poll uses the historic select-style interface
4913  * internally, this routine serves as both the ``check for any
4914  * pending events'' and the ``record my interest in future events''
4915  * functions.  (These are done together, while the lock is held,
4916  * to avoid race conditions.)
4917  */
4918 int
4919 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4920 {
4921 
4922 	v_addpollinfo(vp);
4923 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4924 	if (vp->v_pollinfo->vpi_revents & events) {
4925 		/*
4926 		 * This leaves events we are not interested
4927 		 * in available for the other process which
4928 		 * which presumably had requested them
4929 		 * (otherwise they would never have been
4930 		 * recorded).
4931 		 */
4932 		events &= vp->v_pollinfo->vpi_revents;
4933 		vp->v_pollinfo->vpi_revents &= ~events;
4934 
4935 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4936 		return (events);
4937 	}
4938 	vp->v_pollinfo->vpi_events |= events;
4939 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4940 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4941 	return (0);
4942 }
4943 
4944 /*
4945  * Routine to create and manage a filesystem syncer vnode.
4946  */
4947 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4948 static int	sync_fsync(struct  vop_fsync_args *);
4949 static int	sync_inactive(struct  vop_inactive_args *);
4950 static int	sync_reclaim(struct  vop_reclaim_args *);
4951 
4952 static struct vop_vector sync_vnodeops = {
4953 	.vop_bypass =	VOP_EOPNOTSUPP,
4954 	.vop_close =	sync_close,		/* close */
4955 	.vop_fsync =	sync_fsync,		/* fsync */
4956 	.vop_inactive =	sync_inactive,	/* inactive */
4957 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4958 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4959 	.vop_lock1 =	vop_stdlock,	/* lock */
4960 	.vop_unlock =	vop_stdunlock,	/* unlock */
4961 	.vop_islocked =	vop_stdislocked,	/* islocked */
4962 };
4963 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4964 
4965 /*
4966  * Create a new filesystem syncer vnode for the specified mount point.
4967  */
4968 void
4969 vfs_allocate_syncvnode(struct mount *mp)
4970 {
4971 	struct vnode *vp;
4972 	struct bufobj *bo;
4973 	static long start, incr, next;
4974 	int error;
4975 
4976 	/* Allocate a new vnode */
4977 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4978 	if (error != 0)
4979 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4980 	vp->v_type = VNON;
4981 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4982 	vp->v_vflag |= VV_FORCEINSMQ;
4983 	error = insmntque(vp, mp);
4984 	if (error != 0)
4985 		panic("vfs_allocate_syncvnode: insmntque() failed");
4986 	vp->v_vflag &= ~VV_FORCEINSMQ;
4987 	VOP_UNLOCK(vp);
4988 	/*
4989 	 * Place the vnode onto the syncer worklist. We attempt to
4990 	 * scatter them about on the list so that they will go off
4991 	 * at evenly distributed times even if all the filesystems
4992 	 * are mounted at once.
4993 	 */
4994 	next += incr;
4995 	if (next == 0 || next > syncer_maxdelay) {
4996 		start /= 2;
4997 		incr /= 2;
4998 		if (start == 0) {
4999 			start = syncer_maxdelay / 2;
5000 			incr = syncer_maxdelay;
5001 		}
5002 		next = start;
5003 	}
5004 	bo = &vp->v_bufobj;
5005 	BO_LOCK(bo);
5006 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5007 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5008 	mtx_lock(&sync_mtx);
5009 	sync_vnode_count++;
5010 	if (mp->mnt_syncer == NULL) {
5011 		mp->mnt_syncer = vp;
5012 		vp = NULL;
5013 	}
5014 	mtx_unlock(&sync_mtx);
5015 	BO_UNLOCK(bo);
5016 	if (vp != NULL) {
5017 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5018 		vgone(vp);
5019 		vput(vp);
5020 	}
5021 }
5022 
5023 void
5024 vfs_deallocate_syncvnode(struct mount *mp)
5025 {
5026 	struct vnode *vp;
5027 
5028 	mtx_lock(&sync_mtx);
5029 	vp = mp->mnt_syncer;
5030 	if (vp != NULL)
5031 		mp->mnt_syncer = NULL;
5032 	mtx_unlock(&sync_mtx);
5033 	if (vp != NULL)
5034 		vrele(vp);
5035 }
5036 
5037 /*
5038  * Do a lazy sync of the filesystem.
5039  */
5040 static int
5041 sync_fsync(struct vop_fsync_args *ap)
5042 {
5043 	struct vnode *syncvp = ap->a_vp;
5044 	struct mount *mp = syncvp->v_mount;
5045 	int error, save;
5046 	struct bufobj *bo;
5047 
5048 	/*
5049 	 * We only need to do something if this is a lazy evaluation.
5050 	 */
5051 	if (ap->a_waitfor != MNT_LAZY)
5052 		return (0);
5053 
5054 	/*
5055 	 * Move ourselves to the back of the sync list.
5056 	 */
5057 	bo = &syncvp->v_bufobj;
5058 	BO_LOCK(bo);
5059 	vn_syncer_add_to_worklist(bo, syncdelay);
5060 	BO_UNLOCK(bo);
5061 
5062 	/*
5063 	 * Walk the list of vnodes pushing all that are dirty and
5064 	 * not already on the sync list.
5065 	 */
5066 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5067 		return (0);
5068 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5069 		vfs_unbusy(mp);
5070 		return (0);
5071 	}
5072 	save = curthread_pflags_set(TDP_SYNCIO);
5073 	/*
5074 	 * The filesystem at hand may be idle with free vnodes stored in the
5075 	 * batch.  Return them instead of letting them stay there indefinitely.
5076 	 */
5077 	vfs_periodic(mp, MNT_NOWAIT);
5078 	error = VFS_SYNC(mp, MNT_LAZY);
5079 	curthread_pflags_restore(save);
5080 	vn_finished_write(mp);
5081 	vfs_unbusy(mp);
5082 	return (error);
5083 }
5084 
5085 /*
5086  * The syncer vnode is no referenced.
5087  */
5088 static int
5089 sync_inactive(struct vop_inactive_args *ap)
5090 {
5091 
5092 	vgone(ap->a_vp);
5093 	return (0);
5094 }
5095 
5096 /*
5097  * The syncer vnode is no longer needed and is being decommissioned.
5098  *
5099  * Modifications to the worklist must be protected by sync_mtx.
5100  */
5101 static int
5102 sync_reclaim(struct vop_reclaim_args *ap)
5103 {
5104 	struct vnode *vp = ap->a_vp;
5105 	struct bufobj *bo;
5106 
5107 	bo = &vp->v_bufobj;
5108 	BO_LOCK(bo);
5109 	mtx_lock(&sync_mtx);
5110 	if (vp->v_mount->mnt_syncer == vp)
5111 		vp->v_mount->mnt_syncer = NULL;
5112 	if (bo->bo_flag & BO_ONWORKLST) {
5113 		LIST_REMOVE(bo, bo_synclist);
5114 		syncer_worklist_len--;
5115 		sync_vnode_count--;
5116 		bo->bo_flag &= ~BO_ONWORKLST;
5117 	}
5118 	mtx_unlock(&sync_mtx);
5119 	BO_UNLOCK(bo);
5120 
5121 	return (0);
5122 }
5123 
5124 int
5125 vn_need_pageq_flush(struct vnode *vp)
5126 {
5127 	struct vm_object *obj;
5128 
5129 	obj = vp->v_object;
5130 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5131 	    vm_object_mightbedirty(obj));
5132 }
5133 
5134 /*
5135  * Check if vnode represents a disk device
5136  */
5137 bool
5138 vn_isdisk_error(struct vnode *vp, int *errp)
5139 {
5140 	int error;
5141 
5142 	if (vp->v_type != VCHR) {
5143 		error = ENOTBLK;
5144 		goto out;
5145 	}
5146 	error = 0;
5147 	dev_lock();
5148 	if (vp->v_rdev == NULL)
5149 		error = ENXIO;
5150 	else if (vp->v_rdev->si_devsw == NULL)
5151 		error = ENXIO;
5152 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5153 		error = ENOTBLK;
5154 	dev_unlock();
5155 out:
5156 	*errp = error;
5157 	return (error == 0);
5158 }
5159 
5160 bool
5161 vn_isdisk(struct vnode *vp)
5162 {
5163 	int error;
5164 
5165 	return (vn_isdisk_error(vp, &error));
5166 }
5167 
5168 /*
5169  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5170  * the comment above cache_fplookup for details.
5171  */
5172 int
5173 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5174 {
5175 	int error;
5176 
5177 	VFS_SMR_ASSERT_ENTERED();
5178 
5179 	/* Check the owner. */
5180 	if (cred->cr_uid == file_uid) {
5181 		if (file_mode & S_IXUSR)
5182 			return (0);
5183 		goto out_error;
5184 	}
5185 
5186 	/* Otherwise, check the groups (first match) */
5187 	if (groupmember(file_gid, cred)) {
5188 		if (file_mode & S_IXGRP)
5189 			return (0);
5190 		goto out_error;
5191 	}
5192 
5193 	/* Otherwise, check everyone else. */
5194 	if (file_mode & S_IXOTH)
5195 		return (0);
5196 out_error:
5197 	/*
5198 	 * Permission check failed, but it is possible denial will get overwritten
5199 	 * (e.g., when root is traversing through a 700 directory owned by someone
5200 	 * else).
5201 	 *
5202 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5203 	 * modules overriding this result. It's quite unclear what semantics
5204 	 * are allowed for them to operate, thus for safety we don't call them
5205 	 * from within the SMR section. This also means if any such modules
5206 	 * are present, we have to let the regular lookup decide.
5207 	 */
5208 	error = priv_check_cred_vfs_lookup_nomac(cred);
5209 	switch (error) {
5210 	case 0:
5211 		return (0);
5212 	case EAGAIN:
5213 		/*
5214 		 * MAC modules present.
5215 		 */
5216 		return (EAGAIN);
5217 	case EPERM:
5218 		return (EACCES);
5219 	default:
5220 		return (error);
5221 	}
5222 }
5223 
5224 /*
5225  * Common filesystem object access control check routine.  Accepts a
5226  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5227  * Returns 0 on success, or an errno on failure.
5228  */
5229 int
5230 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5231     accmode_t accmode, struct ucred *cred)
5232 {
5233 	accmode_t dac_granted;
5234 	accmode_t priv_granted;
5235 
5236 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5237 	    ("invalid bit in accmode"));
5238 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5239 	    ("VAPPEND without VWRITE"));
5240 
5241 	/*
5242 	 * Look for a normal, non-privileged way to access the file/directory
5243 	 * as requested.  If it exists, go with that.
5244 	 */
5245 
5246 	dac_granted = 0;
5247 
5248 	/* Check the owner. */
5249 	if (cred->cr_uid == file_uid) {
5250 		dac_granted |= VADMIN;
5251 		if (file_mode & S_IXUSR)
5252 			dac_granted |= VEXEC;
5253 		if (file_mode & S_IRUSR)
5254 			dac_granted |= VREAD;
5255 		if (file_mode & S_IWUSR)
5256 			dac_granted |= (VWRITE | VAPPEND);
5257 
5258 		if ((accmode & dac_granted) == accmode)
5259 			return (0);
5260 
5261 		goto privcheck;
5262 	}
5263 
5264 	/* Otherwise, check the groups (first match) */
5265 	if (groupmember(file_gid, cred)) {
5266 		if (file_mode & S_IXGRP)
5267 			dac_granted |= VEXEC;
5268 		if (file_mode & S_IRGRP)
5269 			dac_granted |= VREAD;
5270 		if (file_mode & S_IWGRP)
5271 			dac_granted |= (VWRITE | VAPPEND);
5272 
5273 		if ((accmode & dac_granted) == accmode)
5274 			return (0);
5275 
5276 		goto privcheck;
5277 	}
5278 
5279 	/* Otherwise, check everyone else. */
5280 	if (file_mode & S_IXOTH)
5281 		dac_granted |= VEXEC;
5282 	if (file_mode & S_IROTH)
5283 		dac_granted |= VREAD;
5284 	if (file_mode & S_IWOTH)
5285 		dac_granted |= (VWRITE | VAPPEND);
5286 	if ((accmode & dac_granted) == accmode)
5287 		return (0);
5288 
5289 privcheck:
5290 	/*
5291 	 * Build a privilege mask to determine if the set of privileges
5292 	 * satisfies the requirements when combined with the granted mask
5293 	 * from above.  For each privilege, if the privilege is required,
5294 	 * bitwise or the request type onto the priv_granted mask.
5295 	 */
5296 	priv_granted = 0;
5297 
5298 	if (type == VDIR) {
5299 		/*
5300 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5301 		 * requests, instead of PRIV_VFS_EXEC.
5302 		 */
5303 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5304 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5305 			priv_granted |= VEXEC;
5306 	} else {
5307 		/*
5308 		 * Ensure that at least one execute bit is on. Otherwise,
5309 		 * a privileged user will always succeed, and we don't want
5310 		 * this to happen unless the file really is executable.
5311 		 */
5312 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5313 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5314 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5315 			priv_granted |= VEXEC;
5316 	}
5317 
5318 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5319 	    !priv_check_cred(cred, PRIV_VFS_READ))
5320 		priv_granted |= VREAD;
5321 
5322 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5323 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5324 		priv_granted |= (VWRITE | VAPPEND);
5325 
5326 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5327 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5328 		priv_granted |= VADMIN;
5329 
5330 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5331 		return (0);
5332 	}
5333 
5334 	return ((accmode & VADMIN) ? EPERM : EACCES);
5335 }
5336 
5337 /*
5338  * Credential check based on process requesting service, and per-attribute
5339  * permissions.
5340  */
5341 int
5342 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5343     struct thread *td, accmode_t accmode)
5344 {
5345 
5346 	/*
5347 	 * Kernel-invoked always succeeds.
5348 	 */
5349 	if (cred == NOCRED)
5350 		return (0);
5351 
5352 	/*
5353 	 * Do not allow privileged processes in jail to directly manipulate
5354 	 * system attributes.
5355 	 */
5356 	switch (attrnamespace) {
5357 	case EXTATTR_NAMESPACE_SYSTEM:
5358 		/* Potentially should be: return (EPERM); */
5359 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5360 	case EXTATTR_NAMESPACE_USER:
5361 		return (VOP_ACCESS(vp, accmode, cred, td));
5362 	default:
5363 		return (EPERM);
5364 	}
5365 }
5366 
5367 #ifdef DEBUG_VFS_LOCKS
5368 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5369 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5370     "Drop into debugger on lock violation");
5371 
5372 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5373 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5374     0, "Check for interlock across VOPs");
5375 
5376 int vfs_badlock_print = 1;	/* Print lock violations. */
5377 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5378     0, "Print lock violations");
5379 
5380 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5381 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5382     0, "Print vnode details on lock violations");
5383 
5384 #ifdef KDB
5385 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5386 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5387     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5388 #endif
5389 
5390 static void
5391 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5392 {
5393 
5394 #ifdef KDB
5395 	if (vfs_badlock_backtrace)
5396 		kdb_backtrace();
5397 #endif
5398 	if (vfs_badlock_vnode)
5399 		vn_printf(vp, "vnode ");
5400 	if (vfs_badlock_print)
5401 		printf("%s: %p %s\n", str, (void *)vp, msg);
5402 	if (vfs_badlock_ddb)
5403 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5404 }
5405 
5406 void
5407 assert_vi_locked(struct vnode *vp, const char *str)
5408 {
5409 
5410 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5411 		vfs_badlock("interlock is not locked but should be", str, vp);
5412 }
5413 
5414 void
5415 assert_vi_unlocked(struct vnode *vp, const char *str)
5416 {
5417 
5418 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5419 		vfs_badlock("interlock is locked but should not be", str, vp);
5420 }
5421 
5422 void
5423 assert_vop_locked(struct vnode *vp, const char *str)
5424 {
5425 	int locked;
5426 
5427 	if (KERNEL_PANICKED() || vp == NULL)
5428 		return;
5429 
5430 	locked = VOP_ISLOCKED(vp);
5431 	if (locked == 0 || locked == LK_EXCLOTHER)
5432 		vfs_badlock("is not locked but should be", str, vp);
5433 }
5434 
5435 void
5436 assert_vop_unlocked(struct vnode *vp, const char *str)
5437 {
5438 	if (KERNEL_PANICKED() || vp == NULL)
5439 		return;
5440 
5441 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5442 		vfs_badlock("is locked but should not be", str, vp);
5443 }
5444 
5445 void
5446 assert_vop_elocked(struct vnode *vp, const char *str)
5447 {
5448 	if (KERNEL_PANICKED() || vp == NULL)
5449 		return;
5450 
5451 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5452 		vfs_badlock("is not exclusive locked but should be", str, vp);
5453 }
5454 #endif /* DEBUG_VFS_LOCKS */
5455 
5456 void
5457 vop_rename_fail(struct vop_rename_args *ap)
5458 {
5459 
5460 	if (ap->a_tvp != NULL)
5461 		vput(ap->a_tvp);
5462 	if (ap->a_tdvp == ap->a_tvp)
5463 		vrele(ap->a_tdvp);
5464 	else
5465 		vput(ap->a_tdvp);
5466 	vrele(ap->a_fdvp);
5467 	vrele(ap->a_fvp);
5468 }
5469 
5470 void
5471 vop_rename_pre(void *ap)
5472 {
5473 	struct vop_rename_args *a = ap;
5474 
5475 #ifdef DEBUG_VFS_LOCKS
5476 	if (a->a_tvp)
5477 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5478 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5479 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5480 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5481 
5482 	/* Check the source (from). */
5483 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5484 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5485 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5486 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5487 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5488 
5489 	/* Check the target. */
5490 	if (a->a_tvp)
5491 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5492 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5493 #endif
5494 	/*
5495 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5496 	 * in vop_rename_post but that's not going to work out since some
5497 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5498 	 *
5499 	 * For now filesystems are expected to do the relevant calls after they
5500 	 * decide what vnodes to operate on.
5501 	 */
5502 	if (a->a_tdvp != a->a_fdvp)
5503 		vhold(a->a_fdvp);
5504 	if (a->a_tvp != a->a_fvp)
5505 		vhold(a->a_fvp);
5506 	vhold(a->a_tdvp);
5507 	if (a->a_tvp)
5508 		vhold(a->a_tvp);
5509 }
5510 
5511 #ifdef DEBUG_VFS_LOCKS
5512 void
5513 vop_fplookup_vexec_debugpre(void *ap __unused)
5514 {
5515 
5516 	VFS_SMR_ASSERT_ENTERED();
5517 }
5518 
5519 void
5520 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5521 {
5522 
5523 	VFS_SMR_ASSERT_ENTERED();
5524 }
5525 
5526 void
5527 vop_fplookup_symlink_debugpre(void *ap __unused)
5528 {
5529 
5530 	VFS_SMR_ASSERT_ENTERED();
5531 }
5532 
5533 void
5534 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5535 {
5536 
5537 	VFS_SMR_ASSERT_ENTERED();
5538 }
5539 
5540 static void
5541 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5542 {
5543 	if (vp->v_type == VCHR)
5544 		;
5545 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5546 		ASSERT_VOP_LOCKED(vp, name);
5547 	else
5548 		ASSERT_VOP_ELOCKED(vp, name);
5549 }
5550 
5551 void
5552 vop_fsync_debugpre(void *a)
5553 {
5554 	struct vop_fsync_args *ap;
5555 
5556 	ap = a;
5557 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5558 }
5559 
5560 void
5561 vop_fsync_debugpost(void *a, int rc __unused)
5562 {
5563 	struct vop_fsync_args *ap;
5564 
5565 	ap = a;
5566 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5567 }
5568 
5569 void
5570 vop_fdatasync_debugpre(void *a)
5571 {
5572 	struct vop_fdatasync_args *ap;
5573 
5574 	ap = a;
5575 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5576 }
5577 
5578 void
5579 vop_fdatasync_debugpost(void *a, int rc __unused)
5580 {
5581 	struct vop_fdatasync_args *ap;
5582 
5583 	ap = a;
5584 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5585 }
5586 
5587 void
5588 vop_strategy_debugpre(void *ap)
5589 {
5590 	struct vop_strategy_args *a;
5591 	struct buf *bp;
5592 
5593 	a = ap;
5594 	bp = a->a_bp;
5595 
5596 	/*
5597 	 * Cluster ops lock their component buffers but not the IO container.
5598 	 */
5599 	if ((bp->b_flags & B_CLUSTER) != 0)
5600 		return;
5601 
5602 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5603 		if (vfs_badlock_print)
5604 			printf(
5605 			    "VOP_STRATEGY: bp is not locked but should be\n");
5606 		if (vfs_badlock_ddb)
5607 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5608 	}
5609 }
5610 
5611 void
5612 vop_lock_debugpre(void *ap)
5613 {
5614 	struct vop_lock1_args *a = ap;
5615 
5616 	if ((a->a_flags & LK_INTERLOCK) == 0)
5617 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5618 	else
5619 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5620 }
5621 
5622 void
5623 vop_lock_debugpost(void *ap, int rc)
5624 {
5625 	struct vop_lock1_args *a = ap;
5626 
5627 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5628 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5629 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5630 }
5631 
5632 void
5633 vop_unlock_debugpre(void *ap)
5634 {
5635 	struct vop_unlock_args *a = ap;
5636 
5637 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5638 }
5639 
5640 void
5641 vop_need_inactive_debugpre(void *ap)
5642 {
5643 	struct vop_need_inactive_args *a = ap;
5644 
5645 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5646 }
5647 
5648 void
5649 vop_need_inactive_debugpost(void *ap, int rc)
5650 {
5651 	struct vop_need_inactive_args *a = ap;
5652 
5653 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5654 }
5655 #endif
5656 
5657 void
5658 vop_create_pre(void *ap)
5659 {
5660 	struct vop_create_args *a;
5661 	struct vnode *dvp;
5662 
5663 	a = ap;
5664 	dvp = a->a_dvp;
5665 	vn_seqc_write_begin(dvp);
5666 }
5667 
5668 void
5669 vop_create_post(void *ap, int rc)
5670 {
5671 	struct vop_create_args *a;
5672 	struct vnode *dvp;
5673 
5674 	a = ap;
5675 	dvp = a->a_dvp;
5676 	vn_seqc_write_end(dvp);
5677 	if (!rc)
5678 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5679 }
5680 
5681 void
5682 vop_whiteout_pre(void *ap)
5683 {
5684 	struct vop_whiteout_args *a;
5685 	struct vnode *dvp;
5686 
5687 	a = ap;
5688 	dvp = a->a_dvp;
5689 	vn_seqc_write_begin(dvp);
5690 }
5691 
5692 void
5693 vop_whiteout_post(void *ap, int rc)
5694 {
5695 	struct vop_whiteout_args *a;
5696 	struct vnode *dvp;
5697 
5698 	a = ap;
5699 	dvp = a->a_dvp;
5700 	vn_seqc_write_end(dvp);
5701 }
5702 
5703 void
5704 vop_deleteextattr_pre(void *ap)
5705 {
5706 	struct vop_deleteextattr_args *a;
5707 	struct vnode *vp;
5708 
5709 	a = ap;
5710 	vp = a->a_vp;
5711 	vn_seqc_write_begin(vp);
5712 }
5713 
5714 void
5715 vop_deleteextattr_post(void *ap, int rc)
5716 {
5717 	struct vop_deleteextattr_args *a;
5718 	struct vnode *vp;
5719 
5720 	a = ap;
5721 	vp = a->a_vp;
5722 	vn_seqc_write_end(vp);
5723 	if (!rc)
5724 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5725 }
5726 
5727 void
5728 vop_link_pre(void *ap)
5729 {
5730 	struct vop_link_args *a;
5731 	struct vnode *vp, *tdvp;
5732 
5733 	a = ap;
5734 	vp = a->a_vp;
5735 	tdvp = a->a_tdvp;
5736 	vn_seqc_write_begin(vp);
5737 	vn_seqc_write_begin(tdvp);
5738 }
5739 
5740 void
5741 vop_link_post(void *ap, int rc)
5742 {
5743 	struct vop_link_args *a;
5744 	struct vnode *vp, *tdvp;
5745 
5746 	a = ap;
5747 	vp = a->a_vp;
5748 	tdvp = a->a_tdvp;
5749 	vn_seqc_write_end(vp);
5750 	vn_seqc_write_end(tdvp);
5751 	if (!rc) {
5752 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5753 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5754 	}
5755 }
5756 
5757 void
5758 vop_mkdir_pre(void *ap)
5759 {
5760 	struct vop_mkdir_args *a;
5761 	struct vnode *dvp;
5762 
5763 	a = ap;
5764 	dvp = a->a_dvp;
5765 	vn_seqc_write_begin(dvp);
5766 }
5767 
5768 void
5769 vop_mkdir_post(void *ap, int rc)
5770 {
5771 	struct vop_mkdir_args *a;
5772 	struct vnode *dvp;
5773 
5774 	a = ap;
5775 	dvp = a->a_dvp;
5776 	vn_seqc_write_end(dvp);
5777 	if (!rc)
5778 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5779 }
5780 
5781 #ifdef DEBUG_VFS_LOCKS
5782 void
5783 vop_mkdir_debugpost(void *ap, int rc)
5784 {
5785 	struct vop_mkdir_args *a;
5786 
5787 	a = ap;
5788 	if (!rc)
5789 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5790 }
5791 #endif
5792 
5793 void
5794 vop_mknod_pre(void *ap)
5795 {
5796 	struct vop_mknod_args *a;
5797 	struct vnode *dvp;
5798 
5799 	a = ap;
5800 	dvp = a->a_dvp;
5801 	vn_seqc_write_begin(dvp);
5802 }
5803 
5804 void
5805 vop_mknod_post(void *ap, int rc)
5806 {
5807 	struct vop_mknod_args *a;
5808 	struct vnode *dvp;
5809 
5810 	a = ap;
5811 	dvp = a->a_dvp;
5812 	vn_seqc_write_end(dvp);
5813 	if (!rc)
5814 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5815 }
5816 
5817 void
5818 vop_reclaim_post(void *ap, int rc)
5819 {
5820 	struct vop_reclaim_args *a;
5821 	struct vnode *vp;
5822 
5823 	a = ap;
5824 	vp = a->a_vp;
5825 	ASSERT_VOP_IN_SEQC(vp);
5826 	if (!rc)
5827 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5828 }
5829 
5830 void
5831 vop_remove_pre(void *ap)
5832 {
5833 	struct vop_remove_args *a;
5834 	struct vnode *dvp, *vp;
5835 
5836 	a = ap;
5837 	dvp = a->a_dvp;
5838 	vp = a->a_vp;
5839 	vn_seqc_write_begin(dvp);
5840 	vn_seqc_write_begin(vp);
5841 }
5842 
5843 void
5844 vop_remove_post(void *ap, int rc)
5845 {
5846 	struct vop_remove_args *a;
5847 	struct vnode *dvp, *vp;
5848 
5849 	a = ap;
5850 	dvp = a->a_dvp;
5851 	vp = a->a_vp;
5852 	vn_seqc_write_end(dvp);
5853 	vn_seqc_write_end(vp);
5854 	if (!rc) {
5855 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5856 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5857 	}
5858 }
5859 
5860 void
5861 vop_rename_post(void *ap, int rc)
5862 {
5863 	struct vop_rename_args *a = ap;
5864 	long hint;
5865 
5866 	if (!rc) {
5867 		hint = NOTE_WRITE;
5868 		if (a->a_fdvp == a->a_tdvp) {
5869 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5870 				hint |= NOTE_LINK;
5871 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5872 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5873 		} else {
5874 			hint |= NOTE_EXTEND;
5875 			if (a->a_fvp->v_type == VDIR)
5876 				hint |= NOTE_LINK;
5877 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5878 
5879 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5880 			    a->a_tvp->v_type == VDIR)
5881 				hint &= ~NOTE_LINK;
5882 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5883 		}
5884 
5885 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5886 		if (a->a_tvp)
5887 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5888 	}
5889 	if (a->a_tdvp != a->a_fdvp)
5890 		vdrop(a->a_fdvp);
5891 	if (a->a_tvp != a->a_fvp)
5892 		vdrop(a->a_fvp);
5893 	vdrop(a->a_tdvp);
5894 	if (a->a_tvp)
5895 		vdrop(a->a_tvp);
5896 }
5897 
5898 void
5899 vop_rmdir_pre(void *ap)
5900 {
5901 	struct vop_rmdir_args *a;
5902 	struct vnode *dvp, *vp;
5903 
5904 	a = ap;
5905 	dvp = a->a_dvp;
5906 	vp = a->a_vp;
5907 	vn_seqc_write_begin(dvp);
5908 	vn_seqc_write_begin(vp);
5909 }
5910 
5911 void
5912 vop_rmdir_post(void *ap, int rc)
5913 {
5914 	struct vop_rmdir_args *a;
5915 	struct vnode *dvp, *vp;
5916 
5917 	a = ap;
5918 	dvp = a->a_dvp;
5919 	vp = a->a_vp;
5920 	vn_seqc_write_end(dvp);
5921 	vn_seqc_write_end(vp);
5922 	if (!rc) {
5923 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5924 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5925 	}
5926 }
5927 
5928 void
5929 vop_setattr_pre(void *ap)
5930 {
5931 	struct vop_setattr_args *a;
5932 	struct vnode *vp;
5933 
5934 	a = ap;
5935 	vp = a->a_vp;
5936 	vn_seqc_write_begin(vp);
5937 }
5938 
5939 void
5940 vop_setattr_post(void *ap, int rc)
5941 {
5942 	struct vop_setattr_args *a;
5943 	struct vnode *vp;
5944 
5945 	a = ap;
5946 	vp = a->a_vp;
5947 	vn_seqc_write_end(vp);
5948 	if (!rc)
5949 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5950 }
5951 
5952 void
5953 vop_setacl_pre(void *ap)
5954 {
5955 	struct vop_setacl_args *a;
5956 	struct vnode *vp;
5957 
5958 	a = ap;
5959 	vp = a->a_vp;
5960 	vn_seqc_write_begin(vp);
5961 }
5962 
5963 void
5964 vop_setacl_post(void *ap, int rc __unused)
5965 {
5966 	struct vop_setacl_args *a;
5967 	struct vnode *vp;
5968 
5969 	a = ap;
5970 	vp = a->a_vp;
5971 	vn_seqc_write_end(vp);
5972 }
5973 
5974 void
5975 vop_setextattr_pre(void *ap)
5976 {
5977 	struct vop_setextattr_args *a;
5978 	struct vnode *vp;
5979 
5980 	a = ap;
5981 	vp = a->a_vp;
5982 	vn_seqc_write_begin(vp);
5983 }
5984 
5985 void
5986 vop_setextattr_post(void *ap, int rc)
5987 {
5988 	struct vop_setextattr_args *a;
5989 	struct vnode *vp;
5990 
5991 	a = ap;
5992 	vp = a->a_vp;
5993 	vn_seqc_write_end(vp);
5994 	if (!rc)
5995 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5996 }
5997 
5998 void
5999 vop_symlink_pre(void *ap)
6000 {
6001 	struct vop_symlink_args *a;
6002 	struct vnode *dvp;
6003 
6004 	a = ap;
6005 	dvp = a->a_dvp;
6006 	vn_seqc_write_begin(dvp);
6007 }
6008 
6009 void
6010 vop_symlink_post(void *ap, int rc)
6011 {
6012 	struct vop_symlink_args *a;
6013 	struct vnode *dvp;
6014 
6015 	a = ap;
6016 	dvp = a->a_dvp;
6017 	vn_seqc_write_end(dvp);
6018 	if (!rc)
6019 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6020 }
6021 
6022 void
6023 vop_open_post(void *ap, int rc)
6024 {
6025 	struct vop_open_args *a = ap;
6026 
6027 	if (!rc)
6028 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6029 }
6030 
6031 void
6032 vop_close_post(void *ap, int rc)
6033 {
6034 	struct vop_close_args *a = ap;
6035 
6036 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6037 	    !VN_IS_DOOMED(a->a_vp))) {
6038 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6039 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6040 	}
6041 }
6042 
6043 void
6044 vop_read_post(void *ap, int rc)
6045 {
6046 	struct vop_read_args *a = ap;
6047 
6048 	if (!rc)
6049 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6050 }
6051 
6052 void
6053 vop_read_pgcache_post(void *ap, int rc)
6054 {
6055 	struct vop_read_pgcache_args *a = ap;
6056 
6057 	if (!rc)
6058 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6059 }
6060 
6061 void
6062 vop_readdir_post(void *ap, int rc)
6063 {
6064 	struct vop_readdir_args *a = ap;
6065 
6066 	if (!rc)
6067 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6068 }
6069 
6070 static struct knlist fs_knlist;
6071 
6072 static void
6073 vfs_event_init(void *arg)
6074 {
6075 	knlist_init_mtx(&fs_knlist, NULL);
6076 }
6077 /* XXX - correct order? */
6078 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6079 
6080 void
6081 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6082 {
6083 
6084 	KNOTE_UNLOCKED(&fs_knlist, event);
6085 }
6086 
6087 static int	filt_fsattach(struct knote *kn);
6088 static void	filt_fsdetach(struct knote *kn);
6089 static int	filt_fsevent(struct knote *kn, long hint);
6090 
6091 struct filterops fs_filtops = {
6092 	.f_isfd = 0,
6093 	.f_attach = filt_fsattach,
6094 	.f_detach = filt_fsdetach,
6095 	.f_event = filt_fsevent
6096 };
6097 
6098 static int
6099 filt_fsattach(struct knote *kn)
6100 {
6101 
6102 	kn->kn_flags |= EV_CLEAR;
6103 	knlist_add(&fs_knlist, kn, 0);
6104 	return (0);
6105 }
6106 
6107 static void
6108 filt_fsdetach(struct knote *kn)
6109 {
6110 
6111 	knlist_remove(&fs_knlist, kn, 0);
6112 }
6113 
6114 static int
6115 filt_fsevent(struct knote *kn, long hint)
6116 {
6117 
6118 	kn->kn_fflags |= kn->kn_sfflags & hint;
6119 
6120 	return (kn->kn_fflags != 0);
6121 }
6122 
6123 static int
6124 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6125 {
6126 	struct vfsidctl vc;
6127 	int error;
6128 	struct mount *mp;
6129 
6130 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6131 	if (error)
6132 		return (error);
6133 	if (vc.vc_vers != VFS_CTL_VERS1)
6134 		return (EINVAL);
6135 	mp = vfs_getvfs(&vc.vc_fsid);
6136 	if (mp == NULL)
6137 		return (ENOENT);
6138 	/* ensure that a specific sysctl goes to the right filesystem. */
6139 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6140 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6141 		vfs_rel(mp);
6142 		return (EINVAL);
6143 	}
6144 	VCTLTOREQ(&vc, req);
6145 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6146 	vfs_rel(mp);
6147 	return (error);
6148 }
6149 
6150 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6151     NULL, 0, sysctl_vfs_ctl, "",
6152     "Sysctl by fsid");
6153 
6154 /*
6155  * Function to initialize a va_filerev field sensibly.
6156  * XXX: Wouldn't a random number make a lot more sense ??
6157  */
6158 u_quad_t
6159 init_va_filerev(void)
6160 {
6161 	struct bintime bt;
6162 
6163 	getbinuptime(&bt);
6164 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6165 }
6166 
6167 static int	filt_vfsread(struct knote *kn, long hint);
6168 static int	filt_vfswrite(struct knote *kn, long hint);
6169 static int	filt_vfsvnode(struct knote *kn, long hint);
6170 static void	filt_vfsdetach(struct knote *kn);
6171 static struct filterops vfsread_filtops = {
6172 	.f_isfd = 1,
6173 	.f_detach = filt_vfsdetach,
6174 	.f_event = filt_vfsread
6175 };
6176 static struct filterops vfswrite_filtops = {
6177 	.f_isfd = 1,
6178 	.f_detach = filt_vfsdetach,
6179 	.f_event = filt_vfswrite
6180 };
6181 static struct filterops vfsvnode_filtops = {
6182 	.f_isfd = 1,
6183 	.f_detach = filt_vfsdetach,
6184 	.f_event = filt_vfsvnode
6185 };
6186 
6187 static void
6188 vfs_knllock(void *arg)
6189 {
6190 	struct vnode *vp = arg;
6191 
6192 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6193 }
6194 
6195 static void
6196 vfs_knlunlock(void *arg)
6197 {
6198 	struct vnode *vp = arg;
6199 
6200 	VOP_UNLOCK(vp);
6201 }
6202 
6203 static void
6204 vfs_knl_assert_lock(void *arg, int what)
6205 {
6206 #ifdef DEBUG_VFS_LOCKS
6207 	struct vnode *vp = arg;
6208 
6209 	if (what == LA_LOCKED)
6210 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6211 	else
6212 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6213 #endif
6214 }
6215 
6216 int
6217 vfs_kqfilter(struct vop_kqfilter_args *ap)
6218 {
6219 	struct vnode *vp = ap->a_vp;
6220 	struct knote *kn = ap->a_kn;
6221 	struct knlist *knl;
6222 
6223 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6224 	    kn->kn_filter != EVFILT_WRITE),
6225 	    ("READ/WRITE filter on a FIFO leaked through"));
6226 	switch (kn->kn_filter) {
6227 	case EVFILT_READ:
6228 		kn->kn_fop = &vfsread_filtops;
6229 		break;
6230 	case EVFILT_WRITE:
6231 		kn->kn_fop = &vfswrite_filtops;
6232 		break;
6233 	case EVFILT_VNODE:
6234 		kn->kn_fop = &vfsvnode_filtops;
6235 		break;
6236 	default:
6237 		return (EINVAL);
6238 	}
6239 
6240 	kn->kn_hook = (caddr_t)vp;
6241 
6242 	v_addpollinfo(vp);
6243 	if (vp->v_pollinfo == NULL)
6244 		return (ENOMEM);
6245 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6246 	vhold(vp);
6247 	knlist_add(knl, kn, 0);
6248 
6249 	return (0);
6250 }
6251 
6252 /*
6253  * Detach knote from vnode
6254  */
6255 static void
6256 filt_vfsdetach(struct knote *kn)
6257 {
6258 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6259 
6260 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6261 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6262 	vdrop(vp);
6263 }
6264 
6265 /*ARGSUSED*/
6266 static int
6267 filt_vfsread(struct knote *kn, long hint)
6268 {
6269 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6270 	struct vattr va;
6271 	int res;
6272 
6273 	/*
6274 	 * filesystem is gone, so set the EOF flag and schedule
6275 	 * the knote for deletion.
6276 	 */
6277 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6278 		VI_LOCK(vp);
6279 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6280 		VI_UNLOCK(vp);
6281 		return (1);
6282 	}
6283 
6284 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6285 		return (0);
6286 
6287 	VI_LOCK(vp);
6288 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6289 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6290 	VI_UNLOCK(vp);
6291 	return (res);
6292 }
6293 
6294 /*ARGSUSED*/
6295 static int
6296 filt_vfswrite(struct knote *kn, long hint)
6297 {
6298 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6299 
6300 	VI_LOCK(vp);
6301 
6302 	/*
6303 	 * filesystem is gone, so set the EOF flag and schedule
6304 	 * the knote for deletion.
6305 	 */
6306 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6307 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6308 
6309 	kn->kn_data = 0;
6310 	VI_UNLOCK(vp);
6311 	return (1);
6312 }
6313 
6314 static int
6315 filt_vfsvnode(struct knote *kn, long hint)
6316 {
6317 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6318 	int res;
6319 
6320 	VI_LOCK(vp);
6321 	if (kn->kn_sfflags & hint)
6322 		kn->kn_fflags |= hint;
6323 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6324 		kn->kn_flags |= EV_EOF;
6325 		VI_UNLOCK(vp);
6326 		return (1);
6327 	}
6328 	res = (kn->kn_fflags != 0);
6329 	VI_UNLOCK(vp);
6330 	return (res);
6331 }
6332 
6333 /*
6334  * Returns whether the directory is empty or not.
6335  * If it is empty, the return value is 0; otherwise
6336  * the return value is an error value (which may
6337  * be ENOTEMPTY).
6338  */
6339 int
6340 vfs_emptydir(struct vnode *vp)
6341 {
6342 	struct uio uio;
6343 	struct iovec iov;
6344 	struct dirent *dirent, *dp, *endp;
6345 	int error, eof;
6346 
6347 	error = 0;
6348 	eof = 0;
6349 
6350 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6351 	VNASSERT(vp->v_type == VDIR, vp, ("vp is not a directory"));
6352 
6353 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6354 	iov.iov_base = dirent;
6355 	iov.iov_len = sizeof(struct dirent);
6356 
6357 	uio.uio_iov = &iov;
6358 	uio.uio_iovcnt = 1;
6359 	uio.uio_offset = 0;
6360 	uio.uio_resid = sizeof(struct dirent);
6361 	uio.uio_segflg = UIO_SYSSPACE;
6362 	uio.uio_rw = UIO_READ;
6363 	uio.uio_td = curthread;
6364 
6365 	while (eof == 0 && error == 0) {
6366 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6367 		    NULL, NULL);
6368 		if (error != 0)
6369 			break;
6370 		endp = (void *)((uint8_t *)dirent +
6371 		    sizeof(struct dirent) - uio.uio_resid);
6372 		for (dp = dirent; dp < endp;
6373 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6374 			if (dp->d_type == DT_WHT)
6375 				continue;
6376 			if (dp->d_namlen == 0)
6377 				continue;
6378 			if (dp->d_type != DT_DIR &&
6379 			    dp->d_type != DT_UNKNOWN) {
6380 				error = ENOTEMPTY;
6381 				break;
6382 			}
6383 			if (dp->d_namlen > 2) {
6384 				error = ENOTEMPTY;
6385 				break;
6386 			}
6387 			if (dp->d_namlen == 1 &&
6388 			    dp->d_name[0] != '.') {
6389 				error = ENOTEMPTY;
6390 				break;
6391 			}
6392 			if (dp->d_namlen == 2 &&
6393 			    dp->d_name[1] != '.') {
6394 				error = ENOTEMPTY;
6395 				break;
6396 			}
6397 			uio.uio_resid = sizeof(struct dirent);
6398 		}
6399 	}
6400 	free(dirent, M_TEMP);
6401 	return (error);
6402 }
6403 
6404 int
6405 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6406 {
6407 	int error;
6408 
6409 	if (dp->d_reclen > ap->a_uio->uio_resid)
6410 		return (ENAMETOOLONG);
6411 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6412 	if (error) {
6413 		if (ap->a_ncookies != NULL) {
6414 			if (ap->a_cookies != NULL)
6415 				free(ap->a_cookies, M_TEMP);
6416 			ap->a_cookies = NULL;
6417 			*ap->a_ncookies = 0;
6418 		}
6419 		return (error);
6420 	}
6421 	if (ap->a_ncookies == NULL)
6422 		return (0);
6423 
6424 	KASSERT(ap->a_cookies,
6425 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6426 
6427 	*ap->a_cookies = realloc(*ap->a_cookies,
6428 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6429 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6430 	*ap->a_ncookies += 1;
6431 	return (0);
6432 }
6433 
6434 /*
6435  * The purpose of this routine is to remove granularity from accmode_t,
6436  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6437  * VADMIN and VAPPEND.
6438  *
6439  * If it returns 0, the caller is supposed to continue with the usual
6440  * access checks using 'accmode' as modified by this routine.  If it
6441  * returns nonzero value, the caller is supposed to return that value
6442  * as errno.
6443  *
6444  * Note that after this routine runs, accmode may be zero.
6445  */
6446 int
6447 vfs_unixify_accmode(accmode_t *accmode)
6448 {
6449 	/*
6450 	 * There is no way to specify explicit "deny" rule using
6451 	 * file mode or POSIX.1e ACLs.
6452 	 */
6453 	if (*accmode & VEXPLICIT_DENY) {
6454 		*accmode = 0;
6455 		return (0);
6456 	}
6457 
6458 	/*
6459 	 * None of these can be translated into usual access bits.
6460 	 * Also, the common case for NFSv4 ACLs is to not contain
6461 	 * either of these bits. Caller should check for VWRITE
6462 	 * on the containing directory instead.
6463 	 */
6464 	if (*accmode & (VDELETE_CHILD | VDELETE))
6465 		return (EPERM);
6466 
6467 	if (*accmode & VADMIN_PERMS) {
6468 		*accmode &= ~VADMIN_PERMS;
6469 		*accmode |= VADMIN;
6470 	}
6471 
6472 	/*
6473 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6474 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6475 	 */
6476 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6477 
6478 	return (0);
6479 }
6480 
6481 /*
6482  * Clear out a doomed vnode (if any) and replace it with a new one as long
6483  * as the fs is not being unmounted. Return the root vnode to the caller.
6484  */
6485 static int __noinline
6486 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6487 {
6488 	struct vnode *vp;
6489 	int error;
6490 
6491 restart:
6492 	if (mp->mnt_rootvnode != NULL) {
6493 		MNT_ILOCK(mp);
6494 		vp = mp->mnt_rootvnode;
6495 		if (vp != NULL) {
6496 			if (!VN_IS_DOOMED(vp)) {
6497 				vrefact(vp);
6498 				MNT_IUNLOCK(mp);
6499 				error = vn_lock(vp, flags);
6500 				if (error == 0) {
6501 					*vpp = vp;
6502 					return (0);
6503 				}
6504 				vrele(vp);
6505 				goto restart;
6506 			}
6507 			/*
6508 			 * Clear the old one.
6509 			 */
6510 			mp->mnt_rootvnode = NULL;
6511 		}
6512 		MNT_IUNLOCK(mp);
6513 		if (vp != NULL) {
6514 			vfs_op_barrier_wait(mp);
6515 			vrele(vp);
6516 		}
6517 	}
6518 	error = VFS_CACHEDROOT(mp, flags, vpp);
6519 	if (error != 0)
6520 		return (error);
6521 	if (mp->mnt_vfs_ops == 0) {
6522 		MNT_ILOCK(mp);
6523 		if (mp->mnt_vfs_ops != 0) {
6524 			MNT_IUNLOCK(mp);
6525 			return (0);
6526 		}
6527 		if (mp->mnt_rootvnode == NULL) {
6528 			vrefact(*vpp);
6529 			mp->mnt_rootvnode = *vpp;
6530 		} else {
6531 			if (mp->mnt_rootvnode != *vpp) {
6532 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6533 					panic("%s: mismatch between vnode returned "
6534 					    " by VFS_CACHEDROOT and the one cached "
6535 					    " (%p != %p)",
6536 					    __func__, *vpp, mp->mnt_rootvnode);
6537 				}
6538 			}
6539 		}
6540 		MNT_IUNLOCK(mp);
6541 	}
6542 	return (0);
6543 }
6544 
6545 int
6546 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6547 {
6548 	struct mount_pcpu *mpcpu;
6549 	struct vnode *vp;
6550 	int error;
6551 
6552 	if (!vfs_op_thread_enter(mp, mpcpu))
6553 		return (vfs_cache_root_fallback(mp, flags, vpp));
6554 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6555 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6556 		vfs_op_thread_exit(mp, mpcpu);
6557 		return (vfs_cache_root_fallback(mp, flags, vpp));
6558 	}
6559 	vrefact(vp);
6560 	vfs_op_thread_exit(mp, mpcpu);
6561 	error = vn_lock(vp, flags);
6562 	if (error != 0) {
6563 		vrele(vp);
6564 		return (vfs_cache_root_fallback(mp, flags, vpp));
6565 	}
6566 	*vpp = vp;
6567 	return (0);
6568 }
6569 
6570 struct vnode *
6571 vfs_cache_root_clear(struct mount *mp)
6572 {
6573 	struct vnode *vp;
6574 
6575 	/*
6576 	 * ops > 0 guarantees there is nobody who can see this vnode
6577 	 */
6578 	MPASS(mp->mnt_vfs_ops > 0);
6579 	vp = mp->mnt_rootvnode;
6580 	if (vp != NULL)
6581 		vn_seqc_write_begin(vp);
6582 	mp->mnt_rootvnode = NULL;
6583 	return (vp);
6584 }
6585 
6586 void
6587 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6588 {
6589 
6590 	MPASS(mp->mnt_vfs_ops > 0);
6591 	vrefact(vp);
6592 	mp->mnt_rootvnode = vp;
6593 }
6594 
6595 /*
6596  * These are helper functions for filesystems to traverse all
6597  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6598  *
6599  * This interface replaces MNT_VNODE_FOREACH.
6600  */
6601 
6602 struct vnode *
6603 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6604 {
6605 	struct vnode *vp;
6606 
6607 	if (should_yield())
6608 		kern_yield(PRI_USER);
6609 	MNT_ILOCK(mp);
6610 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6611 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6612 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6613 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6614 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6615 			continue;
6616 		VI_LOCK(vp);
6617 		if (VN_IS_DOOMED(vp)) {
6618 			VI_UNLOCK(vp);
6619 			continue;
6620 		}
6621 		break;
6622 	}
6623 	if (vp == NULL) {
6624 		__mnt_vnode_markerfree_all(mvp, mp);
6625 		/* MNT_IUNLOCK(mp); -- done in above function */
6626 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6627 		return (NULL);
6628 	}
6629 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6630 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6631 	MNT_IUNLOCK(mp);
6632 	return (vp);
6633 }
6634 
6635 struct vnode *
6636 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6637 {
6638 	struct vnode *vp;
6639 
6640 	*mvp = vn_alloc_marker(mp);
6641 	MNT_ILOCK(mp);
6642 	MNT_REF(mp);
6643 
6644 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6645 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6646 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6647 			continue;
6648 		VI_LOCK(vp);
6649 		if (VN_IS_DOOMED(vp)) {
6650 			VI_UNLOCK(vp);
6651 			continue;
6652 		}
6653 		break;
6654 	}
6655 	if (vp == NULL) {
6656 		MNT_REL(mp);
6657 		MNT_IUNLOCK(mp);
6658 		vn_free_marker(*mvp);
6659 		*mvp = NULL;
6660 		return (NULL);
6661 	}
6662 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6663 	MNT_IUNLOCK(mp);
6664 	return (vp);
6665 }
6666 
6667 void
6668 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6669 {
6670 
6671 	if (*mvp == NULL) {
6672 		MNT_IUNLOCK(mp);
6673 		return;
6674 	}
6675 
6676 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6677 
6678 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6679 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6680 	MNT_REL(mp);
6681 	MNT_IUNLOCK(mp);
6682 	vn_free_marker(*mvp);
6683 	*mvp = NULL;
6684 }
6685 
6686 /*
6687  * These are helper functions for filesystems to traverse their
6688  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6689  */
6690 static void
6691 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6692 {
6693 
6694 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6695 
6696 	MNT_ILOCK(mp);
6697 	MNT_REL(mp);
6698 	MNT_IUNLOCK(mp);
6699 	vn_free_marker(*mvp);
6700 	*mvp = NULL;
6701 }
6702 
6703 /*
6704  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6705  * conventional lock order during mnt_vnode_next_lazy iteration.
6706  *
6707  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6708  * The list lock is dropped and reacquired.  On success, both locks are held.
6709  * On failure, the mount vnode list lock is held but the vnode interlock is
6710  * not, and the procedure may have yielded.
6711  */
6712 static bool
6713 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6714     struct vnode *vp)
6715 {
6716 
6717 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6718 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6719 	    ("%s: bad marker", __func__));
6720 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6721 	    ("%s: inappropriate vnode", __func__));
6722 	ASSERT_VI_UNLOCKED(vp, __func__);
6723 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6724 
6725 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6726 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6727 
6728 	/*
6729 	 * Note we may be racing against vdrop which transitioned the hold
6730 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6731 	 * if we are the only user after we get the interlock we will just
6732 	 * vdrop.
6733 	 */
6734 	vhold(vp);
6735 	mtx_unlock(&mp->mnt_listmtx);
6736 	VI_LOCK(vp);
6737 	if (VN_IS_DOOMED(vp)) {
6738 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6739 		goto out_lost;
6740 	}
6741 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6742 	/*
6743 	 * There is nothing to do if we are the last user.
6744 	 */
6745 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6746 		goto out_lost;
6747 	mtx_lock(&mp->mnt_listmtx);
6748 	return (true);
6749 out_lost:
6750 	vdropl(vp);
6751 	maybe_yield();
6752 	mtx_lock(&mp->mnt_listmtx);
6753 	return (false);
6754 }
6755 
6756 static struct vnode *
6757 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6758     void *cbarg)
6759 {
6760 	struct vnode *vp;
6761 
6762 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6763 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6764 restart:
6765 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6766 	while (vp != NULL) {
6767 		if (vp->v_type == VMARKER) {
6768 			vp = TAILQ_NEXT(vp, v_lazylist);
6769 			continue;
6770 		}
6771 		/*
6772 		 * See if we want to process the vnode. Note we may encounter a
6773 		 * long string of vnodes we don't care about and hog the list
6774 		 * as a result. Check for it and requeue the marker.
6775 		 */
6776 		VNPASS(!VN_IS_DOOMED(vp), vp);
6777 		if (!cb(vp, cbarg)) {
6778 			if (!should_yield()) {
6779 				vp = TAILQ_NEXT(vp, v_lazylist);
6780 				continue;
6781 			}
6782 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6783 			    v_lazylist);
6784 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6785 			    v_lazylist);
6786 			mtx_unlock(&mp->mnt_listmtx);
6787 			kern_yield(PRI_USER);
6788 			mtx_lock(&mp->mnt_listmtx);
6789 			goto restart;
6790 		}
6791 		/*
6792 		 * Try-lock because this is the wrong lock order.
6793 		 */
6794 		if (!VI_TRYLOCK(vp) &&
6795 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6796 			goto restart;
6797 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6798 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6799 		    ("alien vnode on the lazy list %p %p", vp, mp));
6800 		VNPASS(vp->v_mount == mp, vp);
6801 		VNPASS(!VN_IS_DOOMED(vp), vp);
6802 		break;
6803 	}
6804 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6805 
6806 	/* Check if we are done */
6807 	if (vp == NULL) {
6808 		mtx_unlock(&mp->mnt_listmtx);
6809 		mnt_vnode_markerfree_lazy(mvp, mp);
6810 		return (NULL);
6811 	}
6812 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6813 	mtx_unlock(&mp->mnt_listmtx);
6814 	ASSERT_VI_LOCKED(vp, "lazy iter");
6815 	return (vp);
6816 }
6817 
6818 struct vnode *
6819 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6820     void *cbarg)
6821 {
6822 
6823 	if (should_yield())
6824 		kern_yield(PRI_USER);
6825 	mtx_lock(&mp->mnt_listmtx);
6826 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6827 }
6828 
6829 struct vnode *
6830 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6831     void *cbarg)
6832 {
6833 	struct vnode *vp;
6834 
6835 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6836 		return (NULL);
6837 
6838 	*mvp = vn_alloc_marker(mp);
6839 	MNT_ILOCK(mp);
6840 	MNT_REF(mp);
6841 	MNT_IUNLOCK(mp);
6842 
6843 	mtx_lock(&mp->mnt_listmtx);
6844 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6845 	if (vp == NULL) {
6846 		mtx_unlock(&mp->mnt_listmtx);
6847 		mnt_vnode_markerfree_lazy(mvp, mp);
6848 		return (NULL);
6849 	}
6850 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6851 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6852 }
6853 
6854 void
6855 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6856 {
6857 
6858 	if (*mvp == NULL)
6859 		return;
6860 
6861 	mtx_lock(&mp->mnt_listmtx);
6862 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6863 	mtx_unlock(&mp->mnt_listmtx);
6864 	mnt_vnode_markerfree_lazy(mvp, mp);
6865 }
6866 
6867 int
6868 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6869 {
6870 
6871 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6872 		cnp->cn_flags &= ~NOEXECCHECK;
6873 		return (0);
6874 	}
6875 
6876 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6877 }
6878 
6879 /*
6880  * Do not use this variant unless you have means other than the hold count
6881  * to prevent the vnode from getting freed.
6882  */
6883 void
6884 vn_seqc_write_begin_locked(struct vnode *vp)
6885 {
6886 
6887 	ASSERT_VI_LOCKED(vp, __func__);
6888 	VNPASS(vp->v_holdcnt > 0, vp);
6889 	VNPASS(vp->v_seqc_users >= 0, vp);
6890 	vp->v_seqc_users++;
6891 	if (vp->v_seqc_users == 1)
6892 		seqc_sleepable_write_begin(&vp->v_seqc);
6893 }
6894 
6895 void
6896 vn_seqc_write_begin(struct vnode *vp)
6897 {
6898 
6899 	VI_LOCK(vp);
6900 	vn_seqc_write_begin_locked(vp);
6901 	VI_UNLOCK(vp);
6902 }
6903 
6904 void
6905 vn_seqc_write_end_locked(struct vnode *vp)
6906 {
6907 
6908 	ASSERT_VI_LOCKED(vp, __func__);
6909 	VNPASS(vp->v_seqc_users > 0, vp);
6910 	vp->v_seqc_users--;
6911 	if (vp->v_seqc_users == 0)
6912 		seqc_sleepable_write_end(&vp->v_seqc);
6913 }
6914 
6915 void
6916 vn_seqc_write_end(struct vnode *vp)
6917 {
6918 
6919 	VI_LOCK(vp);
6920 	vn_seqc_write_end_locked(vp);
6921 	VI_UNLOCK(vp);
6922 }
6923 
6924 /*
6925  * Special case handling for allocating and freeing vnodes.
6926  *
6927  * The counter remains unchanged on free so that a doomed vnode will
6928  * keep testing as in modify as long as it is accessible with SMR.
6929  */
6930 static void
6931 vn_seqc_init(struct vnode *vp)
6932 {
6933 
6934 	vp->v_seqc = 0;
6935 	vp->v_seqc_users = 0;
6936 }
6937 
6938 static void
6939 vn_seqc_write_end_free(struct vnode *vp)
6940 {
6941 
6942 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6943 	VNPASS(vp->v_seqc_users == 1, vp);
6944 }
6945 
6946 void
6947 vn_irflag_set_locked(struct vnode *vp, short toset)
6948 {
6949 	short flags;
6950 
6951 	ASSERT_VI_LOCKED(vp, __func__);
6952 	flags = vn_irflag_read(vp);
6953 	VNASSERT((flags & toset) == 0, vp,
6954 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
6955 	    __func__, flags, toset));
6956 	atomic_store_short(&vp->v_irflag, flags | toset);
6957 }
6958 
6959 void
6960 vn_irflag_set(struct vnode *vp, short toset)
6961 {
6962 
6963 	VI_LOCK(vp);
6964 	vn_irflag_set_locked(vp, toset);
6965 	VI_UNLOCK(vp);
6966 }
6967 
6968 void
6969 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6970 {
6971 	short flags;
6972 
6973 	ASSERT_VI_LOCKED(vp, __func__);
6974 	flags = vn_irflag_read(vp);
6975 	atomic_store_short(&vp->v_irflag, flags | toset);
6976 }
6977 
6978 void
6979 vn_irflag_set_cond(struct vnode *vp, short toset)
6980 {
6981 
6982 	VI_LOCK(vp);
6983 	vn_irflag_set_cond_locked(vp, toset);
6984 	VI_UNLOCK(vp);
6985 }
6986 
6987 void
6988 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6989 {
6990 	short flags;
6991 
6992 	ASSERT_VI_LOCKED(vp, __func__);
6993 	flags = vn_irflag_read(vp);
6994 	VNASSERT((flags & tounset) == tounset, vp,
6995 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
6996 	    __func__, flags, tounset));
6997 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
6998 }
6999 
7000 void
7001 vn_irflag_unset(struct vnode *vp, short tounset)
7002 {
7003 
7004 	VI_LOCK(vp);
7005 	vn_irflag_unset_locked(vp, tounset);
7006 	VI_UNLOCK(vp);
7007 }
7008