xref: /freebsd/sys/kern/vfs_subr.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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  * These fences are intended for cases where some synchronization is
126  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
127  * and v_usecount) updates.  Access to v_iflags is generally synchronized
128  * by the interlock, but we have some internal assertions that check vnode
129  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
130  * for now.
131  */
132 #ifdef INVARIANTS
133 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
134 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
135 #else
136 #define	VNODE_REFCOUNT_FENCE_ACQ()
137 #define	VNODE_REFCOUNT_FENCE_REL()
138 #endif
139 
140 /*
141  * Number of vnodes in existence.  Increased whenever getnewvnode()
142  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
143  */
144 static u_long __exclusive_cache_line numvnodes;
145 
146 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
147     "Number of vnodes in existence");
148 
149 static counter_u64_t vnodes_created;
150 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
151     "Number of vnodes created by getnewvnode");
152 
153 /*
154  * Conversion tables for conversion from vnode types to inode formats
155  * and back.
156  */
157 enum vtype iftovt_tab[16] = {
158 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
159 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
160 };
161 int vttoif_tab[10] = {
162 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
163 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
164 };
165 
166 /*
167  * List of allocates vnodes in the system.
168  */
169 static TAILQ_HEAD(freelst, vnode) vnode_list;
170 static struct vnode *vnode_list_free_marker;
171 static struct vnode *vnode_list_reclaim_marker;
172 
173 /*
174  * "Free" vnode target.  Free vnodes are rarely completely free, but are
175  * just ones that are cheap to recycle.  Usually they are for files which
176  * have been stat'd but not read; these usually have inode and namecache
177  * data attached to them.  This target is the preferred minimum size of a
178  * sub-cache consisting mostly of such files. The system balances the size
179  * of this sub-cache with its complement to try to prevent either from
180  * thrashing while the other is relatively inactive.  The targets express
181  * a preference for the best balance.
182  *
183  * "Above" this target there are 2 further targets (watermarks) related
184  * to recyling of free vnodes.  In the best-operating case, the cache is
185  * exactly full, the free list has size between vlowat and vhiwat above the
186  * free target, and recycling from it and normal use maintains this state.
187  * Sometimes the free list is below vlowat or even empty, but this state
188  * is even better for immediate use provided the cache is not full.
189  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
190  * ones) to reach one of these states.  The watermarks are currently hard-
191  * coded as 4% and 9% of the available space higher.  These and the default
192  * of 25% for wantfreevnodes are too large if the memory size is large.
193  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
194  * whenever vnlru_proc() becomes active.
195  */
196 static long wantfreevnodes;
197 static long __exclusive_cache_line freevnodes;
198 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
199     &freevnodes, 0, "Number of \"free\" vnodes");
200 static long freevnodes_old;
201 
202 static counter_u64_t recycles_count;
203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
204     "Number of vnodes recycled to meet vnode cache targets");
205 
206 static counter_u64_t recycles_free_count;
207 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
208     "Number of free vnodes recycled to meet vnode cache targets");
209 
210 static counter_u64_t deferred_inact;
211 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
212     "Number of times inactive processing was deferred");
213 
214 /* To keep more than one thread at a time from running vfs_getnewfsid */
215 static struct mtx mntid_mtx;
216 
217 /*
218  * Lock for any access to the following:
219  *	vnode_list
220  *	numvnodes
221  *	freevnodes
222  */
223 static struct mtx __exclusive_cache_line vnode_list_mtx;
224 
225 /* Publicly exported FS */
226 struct nfs_public nfs_pub;
227 
228 static uma_zone_t buf_trie_zone;
229 static smr_t buf_trie_smr;
230 
231 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
232 static uma_zone_t vnode_zone;
233 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
234 
235 __read_frequently smr_t vfs_smr;
236 
237 /*
238  * The workitem queue.
239  *
240  * It is useful to delay writes of file data and filesystem metadata
241  * for tens of seconds so that quickly created and deleted files need
242  * not waste disk bandwidth being created and removed. To realize this,
243  * we append vnodes to a "workitem" queue. When running with a soft
244  * updates implementation, most pending metadata dependencies should
245  * not wait for more than a few seconds. Thus, mounted on block devices
246  * are delayed only about a half the time that file data is delayed.
247  * Similarly, directory updates are more critical, so are only delayed
248  * about a third the time that file data is delayed. Thus, there are
249  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
250  * one each second (driven off the filesystem syncer process). The
251  * syncer_delayno variable indicates the next queue that is to be processed.
252  * Items that need to be processed soon are placed in this queue:
253  *
254  *	syncer_workitem_pending[syncer_delayno]
255  *
256  * A delay of fifteen seconds is done by placing the request fifteen
257  * entries later in the queue:
258  *
259  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
260  *
261  */
262 static int syncer_delayno;
263 static long syncer_mask;
264 LIST_HEAD(synclist, bufobj);
265 static struct synclist *syncer_workitem_pending;
266 /*
267  * The sync_mtx protects:
268  *	bo->bo_synclist
269  *	sync_vnode_count
270  *	syncer_delayno
271  *	syncer_state
272  *	syncer_workitem_pending
273  *	syncer_worklist_len
274  *	rushjob
275  */
276 static struct mtx sync_mtx;
277 static struct cv sync_wakeup;
278 
279 #define SYNCER_MAXDELAY		32
280 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
281 static int syncdelay = 30;		/* max time to delay syncing data */
282 static int filedelay = 30;		/* time to delay syncing files */
283 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
284     "Time to delay syncing files (in seconds)");
285 static int dirdelay = 29;		/* time to delay syncing directories */
286 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
287     "Time to delay syncing directories (in seconds)");
288 static int metadelay = 28;		/* time to delay syncing metadata */
289 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
290     "Time to delay syncing metadata (in seconds)");
291 static int rushjob;		/* number of slots to run ASAP */
292 static int stat_rush_requests;	/* number of times I/O speeded up */
293 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
294     "Number of times I/O speeded up (rush requests)");
295 
296 #define	VDBATCH_SIZE 8
297 struct vdbatch {
298 	u_int index;
299 	long freevnodes;
300 	struct mtx lock;
301 	struct vnode *tab[VDBATCH_SIZE];
302 };
303 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
304 
305 static void	vdbatch_dequeue(struct vnode *vp);
306 
307 /*
308  * When shutting down the syncer, run it at four times normal speed.
309  */
310 #define SYNCER_SHUTDOWN_SPEEDUP		4
311 static int sync_vnode_count;
312 static int syncer_worklist_len;
313 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
314     syncer_state;
315 
316 /* Target for maximum number of vnodes. */
317 u_long desiredvnodes;
318 static u_long gapvnodes;		/* gap between wanted and desired */
319 static u_long vhiwat;		/* enough extras after expansion */
320 static u_long vlowat;		/* minimal extras before expansion */
321 static u_long vstir;		/* nonzero to stir non-free vnodes */
322 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
323 
324 static u_long vnlru_read_freevnodes(void);
325 
326 /*
327  * Note that no attempt is made to sanitize these parameters.
328  */
329 static int
330 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
331 {
332 	u_long val;
333 	int error;
334 
335 	val = desiredvnodes;
336 	error = sysctl_handle_long(oidp, &val, 0, req);
337 	if (error != 0 || req->newptr == NULL)
338 		return (error);
339 
340 	if (val == desiredvnodes)
341 		return (0);
342 	mtx_lock(&vnode_list_mtx);
343 	desiredvnodes = val;
344 	wantfreevnodes = desiredvnodes / 4;
345 	vnlru_recalc();
346 	mtx_unlock(&vnode_list_mtx);
347 	/*
348 	 * XXX There is no protection against multiple threads changing
349 	 * desiredvnodes at the same time. Locking above only helps vnlru and
350 	 * getnewvnode.
351 	 */
352 	vfs_hash_changesize(desiredvnodes);
353 	cache_changesize(desiredvnodes);
354 	return (0);
355 }
356 
357 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
358     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
359     "LU", "Target for maximum number of vnodes");
360 
361 static int
362 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
363 {
364 	u_long val;
365 	int error;
366 
367 	val = wantfreevnodes;
368 	error = sysctl_handle_long(oidp, &val, 0, req);
369 	if (error != 0 || req->newptr == NULL)
370 		return (error);
371 
372 	if (val == wantfreevnodes)
373 		return (0);
374 	mtx_lock(&vnode_list_mtx);
375 	wantfreevnodes = val;
376 	vnlru_recalc();
377 	mtx_unlock(&vnode_list_mtx);
378 	return (0);
379 }
380 
381 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
382     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
383     "LU", "Target for minimum number of \"free\" vnodes");
384 
385 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
386     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
387 static int vnlru_nowhere;
388 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
389     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
390 
391 static int
392 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
393 {
394 	struct vnode *vp;
395 	struct nameidata nd;
396 	char *buf;
397 	unsigned long ndflags;
398 	int error;
399 
400 	if (req->newptr == NULL)
401 		return (EINVAL);
402 	if (req->newlen >= PATH_MAX)
403 		return (E2BIG);
404 
405 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
406 	error = SYSCTL_IN(req, buf, req->newlen);
407 	if (error != 0)
408 		goto out;
409 
410 	buf[req->newlen] = '\0';
411 
412 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
413 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
414 	if ((error = namei(&nd)) != 0)
415 		goto out;
416 	vp = nd.ni_vp;
417 
418 	if (VN_IS_DOOMED(vp)) {
419 		/*
420 		 * This vnode is being recycled.  Return != 0 to let the caller
421 		 * know that the sysctl had no effect.  Return EAGAIN because a
422 		 * subsequent call will likely succeed (since namei will create
423 		 * a new vnode if necessary)
424 		 */
425 		error = EAGAIN;
426 		goto putvnode;
427 	}
428 
429 	counter_u64_add(recycles_count, 1);
430 	vgone(vp);
431 putvnode:
432 	NDFREE(&nd, 0);
433 out:
434 	free(buf, M_TEMP);
435 	return (error);
436 }
437 
438 static int
439 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
440 {
441 	struct thread *td = curthread;
442 	struct vnode *vp;
443 	struct file *fp;
444 	int error;
445 	int fd;
446 
447 	if (req->newptr == NULL)
448 		return (EBADF);
449 
450         error = sysctl_handle_int(oidp, &fd, 0, req);
451         if (error != 0)
452                 return (error);
453 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
454 	if (error != 0)
455 		return (error);
456 	vp = fp->f_vnode;
457 
458 	error = vn_lock(vp, LK_EXCLUSIVE);
459 	if (error != 0)
460 		goto drop;
461 
462 	counter_u64_add(recycles_count, 1);
463 	vgone(vp);
464 	VOP_UNLOCK(vp);
465 drop:
466 	fdrop(fp, td);
467 	return (error);
468 }
469 
470 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
471     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
472     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
473 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
474     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
475     sysctl_ftry_reclaim_vnode, "I",
476     "Try to reclaim a vnode by its file descriptor");
477 
478 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
479 static int vnsz2log;
480 
481 /*
482  * Support for the bufobj clean & dirty pctrie.
483  */
484 static void *
485 buf_trie_alloc(struct pctrie *ptree)
486 {
487 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
488 }
489 
490 static void
491 buf_trie_free(struct pctrie *ptree, void *node)
492 {
493 	uma_zfree_smr(buf_trie_zone, node);
494 }
495 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
496     buf_trie_smr);
497 
498 /*
499  * Initialize the vnode management data structures.
500  *
501  * Reevaluate the following cap on the number of vnodes after the physical
502  * memory size exceeds 512GB.  In the limit, as the physical memory size
503  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
504  */
505 #ifndef	MAXVNODES_MAX
506 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
507 #endif
508 
509 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
510 
511 static struct vnode *
512 vn_alloc_marker(struct mount *mp)
513 {
514 	struct vnode *vp;
515 
516 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
517 	vp->v_type = VMARKER;
518 	vp->v_mount = mp;
519 
520 	return (vp);
521 }
522 
523 static void
524 vn_free_marker(struct vnode *vp)
525 {
526 
527 	MPASS(vp->v_type == VMARKER);
528 	free(vp, M_VNODE_MARKER);
529 }
530 
531 #ifdef KASAN
532 static int
533 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
534 {
535 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
536 	return (0);
537 }
538 
539 static void
540 vnode_dtor(void *mem, int size, void *arg __unused)
541 {
542 	size_t end1, end2, off1, off2;
543 
544 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
545 	    offsetof(struct vnode, v_dbatchcpu),
546 	    "KASAN marks require updating");
547 
548 	off1 = offsetof(struct vnode, v_vnodelist);
549 	off2 = offsetof(struct vnode, v_dbatchcpu);
550 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
551 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
552 
553 	/*
554 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
555 	 * after the vnode has been freed.  Try to get some KASAN coverage by
556 	 * marking everything except those two fields as invalid.  Because
557 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
558 	 * the same 8-byte aligned word must also be marked valid.
559 	 */
560 
561 	/* Handle the area from the start until v_vnodelist... */
562 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
563 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
564 
565 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
566 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
567 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
568 	if (off2 > off1)
569 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
570 		    off2 - off1, KASAN_UMA_FREED);
571 
572 	/* ... and finally the area from v_dbatchcpu to the end. */
573 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
574 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
575 	    KASAN_UMA_FREED);
576 }
577 #endif /* KASAN */
578 
579 /*
580  * Initialize a vnode as it first enters the zone.
581  */
582 static int
583 vnode_init(void *mem, int size, int flags)
584 {
585 	struct vnode *vp;
586 
587 	vp = mem;
588 	bzero(vp, size);
589 	/*
590 	 * Setup locks.
591 	 */
592 	vp->v_vnlock = &vp->v_lock;
593 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
594 	/*
595 	 * By default, don't allow shared locks unless filesystems opt-in.
596 	 */
597 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
598 	    LK_NOSHARE | LK_IS_VNODE);
599 	/*
600 	 * Initialize bufobj.
601 	 */
602 	bufobj_init(&vp->v_bufobj, vp);
603 	/*
604 	 * Initialize namecache.
605 	 */
606 	cache_vnode_init(vp);
607 	/*
608 	 * Initialize rangelocks.
609 	 */
610 	rangelock_init(&vp->v_rl);
611 
612 	vp->v_dbatchcpu = NOCPU;
613 
614 	/*
615 	 * Check vhold_recycle_free for an explanation.
616 	 */
617 	vp->v_holdcnt = VHOLD_NO_SMR;
618 	vp->v_type = VNON;
619 	mtx_lock(&vnode_list_mtx);
620 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
621 	mtx_unlock(&vnode_list_mtx);
622 	return (0);
623 }
624 
625 /*
626  * Free a vnode when it is cleared from the zone.
627  */
628 static void
629 vnode_fini(void *mem, int size)
630 {
631 	struct vnode *vp;
632 	struct bufobj *bo;
633 
634 	vp = mem;
635 	vdbatch_dequeue(vp);
636 	mtx_lock(&vnode_list_mtx);
637 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
638 	mtx_unlock(&vnode_list_mtx);
639 	rangelock_destroy(&vp->v_rl);
640 	lockdestroy(vp->v_vnlock);
641 	mtx_destroy(&vp->v_interlock);
642 	bo = &vp->v_bufobj;
643 	rw_destroy(BO_LOCKPTR(bo));
644 
645 	kasan_mark(mem, size, size, 0);
646 }
647 
648 /*
649  * Provide the size of NFS nclnode and NFS fh for calculation of the
650  * vnode memory consumption.  The size is specified directly to
651  * eliminate dependency on NFS-private header.
652  *
653  * Other filesystems may use bigger or smaller (like UFS and ZFS)
654  * private inode data, but the NFS-based estimation is ample enough.
655  * Still, we care about differences in the size between 64- and 32-bit
656  * platforms.
657  *
658  * Namecache structure size is heuristically
659  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
660  */
661 #ifdef _LP64
662 #define	NFS_NCLNODE_SZ	(528 + 64)
663 #define	NC_SZ		148
664 #else
665 #define	NFS_NCLNODE_SZ	(360 + 32)
666 #define	NC_SZ		92
667 #endif
668 
669 static void
670 vntblinit(void *dummy __unused)
671 {
672 	struct vdbatch *vd;
673 	uma_ctor ctor;
674 	uma_dtor dtor;
675 	int cpu, physvnodes, virtvnodes;
676 	u_int i;
677 
678 	/*
679 	 * Desiredvnodes is a function of the physical memory size and the
680 	 * kernel's heap size.  Generally speaking, it scales with the
681 	 * physical memory size.  The ratio of desiredvnodes to the physical
682 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
683 	 * Thereafter, the
684 	 * marginal ratio of desiredvnodes to the physical memory size is
685 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
686 	 * size.  The memory required by desiredvnodes vnodes and vm objects
687 	 * must not exceed 1/10th of the kernel's heap size.
688 	 */
689 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
690 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
691 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
692 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
693 	desiredvnodes = min(physvnodes, virtvnodes);
694 	if (desiredvnodes > MAXVNODES_MAX) {
695 		if (bootverbose)
696 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
697 			    desiredvnodes, MAXVNODES_MAX);
698 		desiredvnodes = MAXVNODES_MAX;
699 	}
700 	wantfreevnodes = desiredvnodes / 4;
701 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
702 	TAILQ_INIT(&vnode_list);
703 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
704 	/*
705 	 * The lock is taken to appease WITNESS.
706 	 */
707 	mtx_lock(&vnode_list_mtx);
708 	vnlru_recalc();
709 	mtx_unlock(&vnode_list_mtx);
710 	vnode_list_free_marker = vn_alloc_marker(NULL);
711 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
712 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
713 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
714 
715 #ifdef KASAN
716 	ctor = vnode_ctor;
717 	dtor = vnode_dtor;
718 #else
719 	ctor = NULL;
720 	dtor = NULL;
721 #endif
722 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
723 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
724 	uma_zone_set_smr(vnode_zone, vfs_smr);
725 
726 	/*
727 	 * Preallocate enough nodes to support one-per buf so that
728 	 * we can not fail an insert.  reassignbuf() callers can not
729 	 * tolerate the insertion failure.
730 	 */
731 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
732 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
733 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
734 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
735 	uma_prealloc(buf_trie_zone, nbuf);
736 
737 	vnodes_created = counter_u64_alloc(M_WAITOK);
738 	recycles_count = counter_u64_alloc(M_WAITOK);
739 	recycles_free_count = counter_u64_alloc(M_WAITOK);
740 	deferred_inact = counter_u64_alloc(M_WAITOK);
741 
742 	/*
743 	 * Initialize the filesystem syncer.
744 	 */
745 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
746 	    &syncer_mask);
747 	syncer_maxdelay = syncer_mask + 1;
748 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
749 	cv_init(&sync_wakeup, "syncer");
750 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
751 		vnsz2log++;
752 	vnsz2log--;
753 
754 	CPU_FOREACH(cpu) {
755 		vd = DPCPU_ID_PTR((cpu), vd);
756 		bzero(vd, sizeof(*vd));
757 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
758 	}
759 }
760 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
761 
762 /*
763  * Mark a mount point as busy. Used to synchronize access and to delay
764  * unmounting. Eventually, mountlist_mtx is not released on failure.
765  *
766  * vfs_busy() is a custom lock, it can block the caller.
767  * vfs_busy() only sleeps if the unmount is active on the mount point.
768  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
769  * vnode belonging to mp.
770  *
771  * Lookup uses vfs_busy() to traverse mount points.
772  * root fs			var fs
773  * / vnode lock		A	/ vnode lock (/var)		D
774  * /var vnode lock	B	/log vnode lock(/var/log)	E
775  * vfs_busy lock	C	vfs_busy lock			F
776  *
777  * Within each file system, the lock order is C->A->B and F->D->E.
778  *
779  * When traversing across mounts, the system follows that lock order:
780  *
781  *        C->A->B
782  *              |
783  *              +->F->D->E
784  *
785  * The lookup() process for namei("/var") illustrates the process:
786  *  VOP_LOOKUP() obtains B while A is held
787  *  vfs_busy() obtains a shared lock on F while A and B are held
788  *  vput() releases lock on B
789  *  vput() releases lock on A
790  *  VFS_ROOT() obtains lock on D while shared lock on F is held
791  *  vfs_unbusy() releases shared lock on F
792  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
793  *    Attempt to lock A (instead of vp_crossmp) while D is held would
794  *    violate the global order, causing deadlocks.
795  *
796  * dounmount() locks B while F is drained.
797  */
798 int
799 vfs_busy(struct mount *mp, int flags)
800 {
801 	struct mount_pcpu *mpcpu;
802 
803 	MPASS((flags & ~MBF_MASK) == 0);
804 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
805 
806 	if (vfs_op_thread_enter(mp, mpcpu)) {
807 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
808 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
809 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
810 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
811 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
812 		vfs_op_thread_exit(mp, mpcpu);
813 		if (flags & MBF_MNTLSTLOCK)
814 			mtx_unlock(&mountlist_mtx);
815 		return (0);
816 	}
817 
818 	MNT_ILOCK(mp);
819 	vfs_assert_mount_counters(mp);
820 	MNT_REF(mp);
821 	/*
822 	 * If mount point is currently being unmounted, sleep until the
823 	 * mount point fate is decided.  If thread doing the unmounting fails,
824 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
825 	 * that this mount point has survived the unmount attempt and vfs_busy
826 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
827 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
828 	 * about to be really destroyed.  vfs_busy needs to release its
829 	 * reference on the mount point in this case and return with ENOENT,
830 	 * telling the caller that mount mount it tried to busy is no longer
831 	 * valid.
832 	 */
833 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
834 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
835 			MNT_REL(mp);
836 			MNT_IUNLOCK(mp);
837 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
838 			    __func__);
839 			return (ENOENT);
840 		}
841 		if (flags & MBF_MNTLSTLOCK)
842 			mtx_unlock(&mountlist_mtx);
843 		mp->mnt_kern_flag |= MNTK_MWAIT;
844 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
845 		if (flags & MBF_MNTLSTLOCK)
846 			mtx_lock(&mountlist_mtx);
847 		MNT_ILOCK(mp);
848 	}
849 	if (flags & MBF_MNTLSTLOCK)
850 		mtx_unlock(&mountlist_mtx);
851 	mp->mnt_lockref++;
852 	MNT_IUNLOCK(mp);
853 	return (0);
854 }
855 
856 /*
857  * Free a busy filesystem.
858  */
859 void
860 vfs_unbusy(struct mount *mp)
861 {
862 	struct mount_pcpu *mpcpu;
863 	int c;
864 
865 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
866 
867 	if (vfs_op_thread_enter(mp, mpcpu)) {
868 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
869 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
870 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
871 		vfs_op_thread_exit(mp, mpcpu);
872 		return;
873 	}
874 
875 	MNT_ILOCK(mp);
876 	vfs_assert_mount_counters(mp);
877 	MNT_REL(mp);
878 	c = --mp->mnt_lockref;
879 	if (mp->mnt_vfs_ops == 0) {
880 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
881 		MNT_IUNLOCK(mp);
882 		return;
883 	}
884 	if (c < 0)
885 		vfs_dump_mount_counters(mp);
886 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
887 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
888 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
889 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
890 		wakeup(&mp->mnt_lockref);
891 	}
892 	MNT_IUNLOCK(mp);
893 }
894 
895 /*
896  * Lookup a mount point by filesystem identifier.
897  */
898 struct mount *
899 vfs_getvfs(fsid_t *fsid)
900 {
901 	struct mount *mp;
902 
903 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
904 	mtx_lock(&mountlist_mtx);
905 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
906 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
907 			vfs_ref(mp);
908 			mtx_unlock(&mountlist_mtx);
909 			return (mp);
910 		}
911 	}
912 	mtx_unlock(&mountlist_mtx);
913 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
914 	return ((struct mount *) 0);
915 }
916 
917 /*
918  * Lookup a mount point by filesystem identifier, busying it before
919  * returning.
920  *
921  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
922  * cache for popular filesystem identifiers.  The cache is lockess, using
923  * the fact that struct mount's are never freed.  In worst case we may
924  * get pointer to unmounted or even different filesystem, so we have to
925  * check what we got, and go slow way if so.
926  */
927 struct mount *
928 vfs_busyfs(fsid_t *fsid)
929 {
930 #define	FSID_CACHE_SIZE	256
931 	typedef struct mount * volatile vmp_t;
932 	static vmp_t cache[FSID_CACHE_SIZE];
933 	struct mount *mp;
934 	int error;
935 	uint32_t hash;
936 
937 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
938 	hash = fsid->val[0] ^ fsid->val[1];
939 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
940 	mp = cache[hash];
941 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
942 		goto slow;
943 	if (vfs_busy(mp, 0) != 0) {
944 		cache[hash] = NULL;
945 		goto slow;
946 	}
947 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
948 		return (mp);
949 	else
950 	    vfs_unbusy(mp);
951 
952 slow:
953 	mtx_lock(&mountlist_mtx);
954 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
955 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
956 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
957 			if (error) {
958 				cache[hash] = NULL;
959 				mtx_unlock(&mountlist_mtx);
960 				return (NULL);
961 			}
962 			cache[hash] = mp;
963 			return (mp);
964 		}
965 	}
966 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
967 	mtx_unlock(&mountlist_mtx);
968 	return ((struct mount *) 0);
969 }
970 
971 /*
972  * Check if a user can access privileged mount options.
973  */
974 int
975 vfs_suser(struct mount *mp, struct thread *td)
976 {
977 	int error;
978 
979 	if (jailed(td->td_ucred)) {
980 		/*
981 		 * If the jail of the calling thread lacks permission for
982 		 * this type of file system, deny immediately.
983 		 */
984 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
985 			return (EPERM);
986 
987 		/*
988 		 * If the file system was mounted outside the jail of the
989 		 * calling thread, deny immediately.
990 		 */
991 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
992 			return (EPERM);
993 	}
994 
995 	/*
996 	 * If file system supports delegated administration, we don't check
997 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
998 	 * by the file system itself.
999 	 * If this is not the user that did original mount, we check for
1000 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1001 	 */
1002 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1003 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1004 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1005 			return (error);
1006 	}
1007 	return (0);
1008 }
1009 
1010 /*
1011  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1012  * will be used to create fake device numbers for stat().  Also try (but
1013  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1014  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1015  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1016  *
1017  * Keep in mind that several mounts may be running in parallel.  Starting
1018  * the search one past where the previous search terminated is both a
1019  * micro-optimization and a defense against returning the same fsid to
1020  * different mounts.
1021  */
1022 void
1023 vfs_getnewfsid(struct mount *mp)
1024 {
1025 	static uint16_t mntid_base;
1026 	struct mount *nmp;
1027 	fsid_t tfsid;
1028 	int mtype;
1029 
1030 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1031 	mtx_lock(&mntid_mtx);
1032 	mtype = mp->mnt_vfc->vfc_typenum;
1033 	tfsid.val[1] = mtype;
1034 	mtype = (mtype & 0xFF) << 24;
1035 	for (;;) {
1036 		tfsid.val[0] = makedev(255,
1037 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1038 		mntid_base++;
1039 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1040 			break;
1041 		vfs_rel(nmp);
1042 	}
1043 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1044 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1045 	mtx_unlock(&mntid_mtx);
1046 }
1047 
1048 /*
1049  * Knob to control the precision of file timestamps:
1050  *
1051  *   0 = seconds only; nanoseconds zeroed.
1052  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1053  *   2 = seconds and nanoseconds, truncated to microseconds.
1054  * >=3 = seconds and nanoseconds, maximum precision.
1055  */
1056 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1057 
1058 static int timestamp_precision = TSP_USEC;
1059 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1060     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1061     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1062     "3+: sec + ns (max. precision))");
1063 
1064 /*
1065  * Get a current timestamp.
1066  */
1067 void
1068 vfs_timestamp(struct timespec *tsp)
1069 {
1070 	struct timeval tv;
1071 
1072 	switch (timestamp_precision) {
1073 	case TSP_SEC:
1074 		tsp->tv_sec = time_second;
1075 		tsp->tv_nsec = 0;
1076 		break;
1077 	case TSP_HZ:
1078 		getnanotime(tsp);
1079 		break;
1080 	case TSP_USEC:
1081 		microtime(&tv);
1082 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1083 		break;
1084 	case TSP_NSEC:
1085 	default:
1086 		nanotime(tsp);
1087 		break;
1088 	}
1089 }
1090 
1091 /*
1092  * Set vnode attributes to VNOVAL
1093  */
1094 void
1095 vattr_null(struct vattr *vap)
1096 {
1097 
1098 	vap->va_type = VNON;
1099 	vap->va_size = VNOVAL;
1100 	vap->va_bytes = VNOVAL;
1101 	vap->va_mode = VNOVAL;
1102 	vap->va_nlink = VNOVAL;
1103 	vap->va_uid = VNOVAL;
1104 	vap->va_gid = VNOVAL;
1105 	vap->va_fsid = VNOVAL;
1106 	vap->va_fileid = VNOVAL;
1107 	vap->va_blocksize = VNOVAL;
1108 	vap->va_rdev = VNOVAL;
1109 	vap->va_atime.tv_sec = VNOVAL;
1110 	vap->va_atime.tv_nsec = VNOVAL;
1111 	vap->va_mtime.tv_sec = VNOVAL;
1112 	vap->va_mtime.tv_nsec = VNOVAL;
1113 	vap->va_ctime.tv_sec = VNOVAL;
1114 	vap->va_ctime.tv_nsec = VNOVAL;
1115 	vap->va_birthtime.tv_sec = VNOVAL;
1116 	vap->va_birthtime.tv_nsec = VNOVAL;
1117 	vap->va_flags = VNOVAL;
1118 	vap->va_gen = VNOVAL;
1119 	vap->va_vaflags = 0;
1120 }
1121 
1122 /*
1123  * Try to reduce the total number of vnodes.
1124  *
1125  * This routine (and its user) are buggy in at least the following ways:
1126  * - all parameters were picked years ago when RAM sizes were significantly
1127  *   smaller
1128  * - it can pick vnodes based on pages used by the vm object, but filesystems
1129  *   like ZFS don't use it making the pick broken
1130  * - since ZFS has its own aging policy it gets partially combated by this one
1131  * - a dedicated method should be provided for filesystems to let them decide
1132  *   whether the vnode should be recycled
1133  *
1134  * This routine is called when we have too many vnodes.  It attempts
1135  * to free <count> vnodes and will potentially free vnodes that still
1136  * have VM backing store (VM backing store is typically the cause
1137  * of a vnode blowout so we want to do this).  Therefore, this operation
1138  * is not considered cheap.
1139  *
1140  * A number of conditions may prevent a vnode from being reclaimed.
1141  * the buffer cache may have references on the vnode, a directory
1142  * vnode may still have references due to the namei cache representing
1143  * underlying files, or the vnode may be in active use.   It is not
1144  * desirable to reuse such vnodes.  These conditions may cause the
1145  * number of vnodes to reach some minimum value regardless of what
1146  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1147  *
1148  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1149  * 			 entries if this argument is strue
1150  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1151  *			 pages.
1152  * @param target	 How many vnodes to reclaim.
1153  * @return		 The number of vnodes that were reclaimed.
1154  */
1155 static int
1156 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1157 {
1158 	struct vnode *vp, *mvp;
1159 	struct mount *mp;
1160 	struct vm_object *object;
1161 	u_long done;
1162 	bool retried;
1163 
1164 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1165 
1166 	retried = false;
1167 	done = 0;
1168 
1169 	mvp = vnode_list_reclaim_marker;
1170 restart:
1171 	vp = mvp;
1172 	while (done < target) {
1173 		vp = TAILQ_NEXT(vp, v_vnodelist);
1174 		if (__predict_false(vp == NULL))
1175 			break;
1176 
1177 		if (__predict_false(vp->v_type == VMARKER))
1178 			continue;
1179 
1180 		/*
1181 		 * If it's been deconstructed already, it's still
1182 		 * referenced, or it exceeds the trigger, skip it.
1183 		 * Also skip free vnodes.  We are trying to make space
1184 		 * to expand the free list, not reduce it.
1185 		 */
1186 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1187 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1188 			goto next_iter;
1189 
1190 		if (vp->v_type == VBAD || vp->v_type == VNON)
1191 			goto next_iter;
1192 
1193 		object = atomic_load_ptr(&vp->v_object);
1194 		if (object == NULL || object->resident_page_count > trigger) {
1195 			goto next_iter;
1196 		}
1197 
1198 		/*
1199 		 * Handle races against vnode allocation. Filesystems lock the
1200 		 * vnode some time after it gets returned from getnewvnode,
1201 		 * despite type and hold count being manipulated earlier.
1202 		 * Resorting to checking v_mount restores guarantees present
1203 		 * before the global list was reworked to contain all vnodes.
1204 		 */
1205 		if (!VI_TRYLOCK(vp))
1206 			goto next_iter;
1207 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1208 			VI_UNLOCK(vp);
1209 			goto next_iter;
1210 		}
1211 		if (vp->v_mount == NULL) {
1212 			VI_UNLOCK(vp);
1213 			goto next_iter;
1214 		}
1215 		vholdl(vp);
1216 		VI_UNLOCK(vp);
1217 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1218 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1219 		mtx_unlock(&vnode_list_mtx);
1220 
1221 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1222 			vdrop(vp);
1223 			goto next_iter_unlocked;
1224 		}
1225 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1226 			vdrop(vp);
1227 			vn_finished_write(mp);
1228 			goto next_iter_unlocked;
1229 		}
1230 
1231 		VI_LOCK(vp);
1232 		if (vp->v_usecount > 0 ||
1233 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1234 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1235 		    vp->v_object->resident_page_count > trigger)) {
1236 			VOP_UNLOCK(vp);
1237 			vdropl(vp);
1238 			vn_finished_write(mp);
1239 			goto next_iter_unlocked;
1240 		}
1241 		counter_u64_add(recycles_count, 1);
1242 		vgonel(vp);
1243 		VOP_UNLOCK(vp);
1244 		vdropl(vp);
1245 		vn_finished_write(mp);
1246 		done++;
1247 next_iter_unlocked:
1248 		if (should_yield())
1249 			kern_yield(PRI_USER);
1250 		mtx_lock(&vnode_list_mtx);
1251 		goto restart;
1252 next_iter:
1253 		MPASS(vp->v_type != VMARKER);
1254 		if (!should_yield())
1255 			continue;
1256 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1257 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1258 		mtx_unlock(&vnode_list_mtx);
1259 		kern_yield(PRI_USER);
1260 		mtx_lock(&vnode_list_mtx);
1261 		goto restart;
1262 	}
1263 	if (done == 0 && !retried) {
1264 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1265 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1266 		retried = true;
1267 		goto restart;
1268 	}
1269 	return (done);
1270 }
1271 
1272 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1273 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1274     0,
1275     "limit on vnode free requests per call to the vnlru_free routine");
1276 
1277 /*
1278  * Attempt to reduce the free list by the requested amount.
1279  */
1280 static int
1281 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1282 {
1283 	struct vnode *vp;
1284 	struct mount *mp;
1285 	int ocount;
1286 
1287 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1288 	if (count > max_vnlru_free)
1289 		count = max_vnlru_free;
1290 	ocount = count;
1291 	vp = mvp;
1292 	for (;;) {
1293 		if (count == 0) {
1294 			break;
1295 		}
1296 		vp = TAILQ_NEXT(vp, v_vnodelist);
1297 		if (__predict_false(vp == NULL)) {
1298 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1299 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1300 			break;
1301 		}
1302 		if (__predict_false(vp->v_type == VMARKER))
1303 			continue;
1304 		if (vp->v_holdcnt > 0)
1305 			continue;
1306 		/*
1307 		 * Don't recycle if our vnode is from different type
1308 		 * of mount point.  Note that mp is type-safe, the
1309 		 * check does not reach unmapped address even if
1310 		 * vnode is reclaimed.
1311 		 */
1312 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1313 		    mp->mnt_op != mnt_op) {
1314 			continue;
1315 		}
1316 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1317 			continue;
1318 		}
1319 		if (!vhold_recycle_free(vp))
1320 			continue;
1321 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1322 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1323 		mtx_unlock(&vnode_list_mtx);
1324 		if (vtryrecycle(vp) == 0)
1325 			count--;
1326 		mtx_lock(&vnode_list_mtx);
1327 		vp = mvp;
1328 	}
1329 	return (ocount - count);
1330 }
1331 
1332 static int
1333 vnlru_free_locked(int count)
1334 {
1335 
1336 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1337 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1338 }
1339 
1340 void
1341 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1342 {
1343 
1344 	MPASS(mnt_op != NULL);
1345 	MPASS(mvp != NULL);
1346 	VNPASS(mvp->v_type == VMARKER, mvp);
1347 	mtx_lock(&vnode_list_mtx);
1348 	vnlru_free_impl(count, mnt_op, mvp);
1349 	mtx_unlock(&vnode_list_mtx);
1350 }
1351 
1352 struct vnode *
1353 vnlru_alloc_marker(void)
1354 {
1355 	struct vnode *mvp;
1356 
1357 	mvp = vn_alloc_marker(NULL);
1358 	mtx_lock(&vnode_list_mtx);
1359 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1360 	mtx_unlock(&vnode_list_mtx);
1361 	return (mvp);
1362 }
1363 
1364 void
1365 vnlru_free_marker(struct vnode *mvp)
1366 {
1367 	mtx_lock(&vnode_list_mtx);
1368 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1369 	mtx_unlock(&vnode_list_mtx);
1370 	vn_free_marker(mvp);
1371 }
1372 
1373 static void
1374 vnlru_recalc(void)
1375 {
1376 
1377 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1378 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1379 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1380 	vlowat = vhiwat / 2;
1381 }
1382 
1383 /*
1384  * Attempt to recycle vnodes in a context that is always safe to block.
1385  * Calling vlrurecycle() from the bowels of filesystem code has some
1386  * interesting deadlock problems.
1387  */
1388 static struct proc *vnlruproc;
1389 static int vnlruproc_sig;
1390 
1391 /*
1392  * The main freevnodes counter is only updated when threads requeue their vnode
1393  * batches. CPUs are conditionally walked to compute a more accurate total.
1394  *
1395  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1396  * at any given moment can still exceed slop, but it should not be by significant
1397  * margin in practice.
1398  */
1399 #define VNLRU_FREEVNODES_SLOP 128
1400 
1401 static __inline void
1402 vfs_freevnodes_inc(void)
1403 {
1404 	struct vdbatch *vd;
1405 
1406 	critical_enter();
1407 	vd = DPCPU_PTR(vd);
1408 	vd->freevnodes++;
1409 	critical_exit();
1410 }
1411 
1412 static __inline void
1413 vfs_freevnodes_dec(void)
1414 {
1415 	struct vdbatch *vd;
1416 
1417 	critical_enter();
1418 	vd = DPCPU_PTR(vd);
1419 	vd->freevnodes--;
1420 	critical_exit();
1421 }
1422 
1423 static u_long
1424 vnlru_read_freevnodes(void)
1425 {
1426 	struct vdbatch *vd;
1427 	long slop;
1428 	int cpu;
1429 
1430 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1431 	if (freevnodes > freevnodes_old)
1432 		slop = freevnodes - freevnodes_old;
1433 	else
1434 		slop = freevnodes_old - freevnodes;
1435 	if (slop < VNLRU_FREEVNODES_SLOP)
1436 		return (freevnodes >= 0 ? freevnodes : 0);
1437 	freevnodes_old = freevnodes;
1438 	CPU_FOREACH(cpu) {
1439 		vd = DPCPU_ID_PTR((cpu), vd);
1440 		freevnodes_old += vd->freevnodes;
1441 	}
1442 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1443 }
1444 
1445 static bool
1446 vnlru_under(u_long rnumvnodes, u_long limit)
1447 {
1448 	u_long rfreevnodes, space;
1449 
1450 	if (__predict_false(rnumvnodes > desiredvnodes))
1451 		return (true);
1452 
1453 	space = desiredvnodes - rnumvnodes;
1454 	if (space < limit) {
1455 		rfreevnodes = vnlru_read_freevnodes();
1456 		if (rfreevnodes > wantfreevnodes)
1457 			space += rfreevnodes - wantfreevnodes;
1458 	}
1459 	return (space < limit);
1460 }
1461 
1462 static bool
1463 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1464 {
1465 	long rfreevnodes, space;
1466 
1467 	if (__predict_false(rnumvnodes > desiredvnodes))
1468 		return (true);
1469 
1470 	space = desiredvnodes - rnumvnodes;
1471 	if (space < limit) {
1472 		rfreevnodes = atomic_load_long(&freevnodes);
1473 		if (rfreevnodes > wantfreevnodes)
1474 			space += rfreevnodes - wantfreevnodes;
1475 	}
1476 	return (space < limit);
1477 }
1478 
1479 static void
1480 vnlru_kick(void)
1481 {
1482 
1483 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1484 	if (vnlruproc_sig == 0) {
1485 		vnlruproc_sig = 1;
1486 		wakeup(vnlruproc);
1487 	}
1488 }
1489 
1490 static void
1491 vnlru_proc(void)
1492 {
1493 	u_long rnumvnodes, rfreevnodes, target;
1494 	unsigned long onumvnodes;
1495 	int done, force, trigger, usevnodes;
1496 	bool reclaim_nc_src, want_reread;
1497 
1498 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1499 	    SHUTDOWN_PRI_FIRST);
1500 
1501 	force = 0;
1502 	want_reread = false;
1503 	for (;;) {
1504 		kproc_suspend_check(vnlruproc);
1505 		mtx_lock(&vnode_list_mtx);
1506 		rnumvnodes = atomic_load_long(&numvnodes);
1507 
1508 		if (want_reread) {
1509 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1510 			want_reread = false;
1511 		}
1512 
1513 		/*
1514 		 * If numvnodes is too large (due to desiredvnodes being
1515 		 * adjusted using its sysctl, or emergency growth), first
1516 		 * try to reduce it by discarding from the free list.
1517 		 */
1518 		if (rnumvnodes > desiredvnodes) {
1519 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1520 			rnumvnodes = atomic_load_long(&numvnodes);
1521 		}
1522 		/*
1523 		 * Sleep if the vnode cache is in a good state.  This is
1524 		 * when it is not over-full and has space for about a 4%
1525 		 * or 9% expansion (by growing its size or inexcessively
1526 		 * reducing its free list).  Otherwise, try to reclaim
1527 		 * space for a 10% expansion.
1528 		 */
1529 		if (vstir && force == 0) {
1530 			force = 1;
1531 			vstir = 0;
1532 		}
1533 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1534 			vnlruproc_sig = 0;
1535 			wakeup(&vnlruproc_sig);
1536 			msleep(vnlruproc, &vnode_list_mtx,
1537 			    PVFS|PDROP, "vlruwt", hz);
1538 			continue;
1539 		}
1540 		rfreevnodes = vnlru_read_freevnodes();
1541 
1542 		onumvnodes = rnumvnodes;
1543 		/*
1544 		 * Calculate parameters for recycling.  These are the same
1545 		 * throughout the loop to give some semblance of fairness.
1546 		 * The trigger point is to avoid recycling vnodes with lots
1547 		 * of resident pages.  We aren't trying to free memory; we
1548 		 * are trying to recycle or at least free vnodes.
1549 		 */
1550 		if (rnumvnodes <= desiredvnodes)
1551 			usevnodes = rnumvnodes - rfreevnodes;
1552 		else
1553 			usevnodes = rnumvnodes;
1554 		if (usevnodes <= 0)
1555 			usevnodes = 1;
1556 		/*
1557 		 * The trigger value is is chosen to give a conservatively
1558 		 * large value to ensure that it alone doesn't prevent
1559 		 * making progress.  The value can easily be so large that
1560 		 * it is effectively infinite in some congested and
1561 		 * misconfigured cases, and this is necessary.  Normally
1562 		 * it is about 8 to 100 (pages), which is quite large.
1563 		 */
1564 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1565 		if (force < 2)
1566 			trigger = vsmalltrigger;
1567 		reclaim_nc_src = force >= 3;
1568 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1569 		target = target / 10 + 1;
1570 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1571 		mtx_unlock(&vnode_list_mtx);
1572 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1573 			uma_reclaim(UMA_RECLAIM_DRAIN);
1574 		if (done == 0) {
1575 			if (force == 0 || force == 1) {
1576 				force = 2;
1577 				continue;
1578 			}
1579 			if (force == 2) {
1580 				force = 3;
1581 				continue;
1582 			}
1583 			want_reread = true;
1584 			force = 0;
1585 			vnlru_nowhere++;
1586 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1587 		} else {
1588 			want_reread = true;
1589 			kern_yield(PRI_USER);
1590 		}
1591 	}
1592 }
1593 
1594 static struct kproc_desc vnlru_kp = {
1595 	"vnlru",
1596 	vnlru_proc,
1597 	&vnlruproc
1598 };
1599 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1600     &vnlru_kp);
1601 
1602 /*
1603  * Routines having to do with the management of the vnode table.
1604  */
1605 
1606 /*
1607  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1608  * before we actually vgone().  This function must be called with the vnode
1609  * held to prevent the vnode from being returned to the free list midway
1610  * through vgone().
1611  */
1612 static int
1613 vtryrecycle(struct vnode *vp)
1614 {
1615 	struct mount *vnmp;
1616 
1617 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1618 	VNASSERT(vp->v_holdcnt, vp,
1619 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1620 	/*
1621 	 * This vnode may found and locked via some other list, if so we
1622 	 * can't recycle it yet.
1623 	 */
1624 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1625 		CTR2(KTR_VFS,
1626 		    "%s: impossible to recycle, vp %p lock is already held",
1627 		    __func__, vp);
1628 		vdrop(vp);
1629 		return (EWOULDBLOCK);
1630 	}
1631 	/*
1632 	 * Don't recycle if its filesystem is being suspended.
1633 	 */
1634 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1635 		VOP_UNLOCK(vp);
1636 		CTR2(KTR_VFS,
1637 		    "%s: impossible to recycle, cannot start the write for %p",
1638 		    __func__, vp);
1639 		vdrop(vp);
1640 		return (EBUSY);
1641 	}
1642 	/*
1643 	 * If we got this far, we need to acquire the interlock and see if
1644 	 * anyone picked up this vnode from another list.  If not, we will
1645 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1646 	 * will skip over it.
1647 	 */
1648 	VI_LOCK(vp);
1649 	if (vp->v_usecount) {
1650 		VOP_UNLOCK(vp);
1651 		vdropl(vp);
1652 		vn_finished_write(vnmp);
1653 		CTR2(KTR_VFS,
1654 		    "%s: impossible to recycle, %p is already referenced",
1655 		    __func__, vp);
1656 		return (EBUSY);
1657 	}
1658 	if (!VN_IS_DOOMED(vp)) {
1659 		counter_u64_add(recycles_free_count, 1);
1660 		vgonel(vp);
1661 	}
1662 	VOP_UNLOCK(vp);
1663 	vdropl(vp);
1664 	vn_finished_write(vnmp);
1665 	return (0);
1666 }
1667 
1668 /*
1669  * Allocate a new vnode.
1670  *
1671  * The operation never returns an error. Returning an error was disabled
1672  * in r145385 (dated 2005) with the following comment:
1673  *
1674  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1675  *
1676  * Given the age of this commit (almost 15 years at the time of writing this
1677  * comment) restoring the ability to fail requires a significant audit of
1678  * all codepaths.
1679  *
1680  * The routine can try to free a vnode or stall for up to 1 second waiting for
1681  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1682  */
1683 static u_long vn_alloc_cyclecount;
1684 
1685 static struct vnode * __noinline
1686 vn_alloc_hard(struct mount *mp)
1687 {
1688 	u_long rnumvnodes, rfreevnodes;
1689 
1690 	mtx_lock(&vnode_list_mtx);
1691 	rnumvnodes = atomic_load_long(&numvnodes);
1692 	if (rnumvnodes + 1 < desiredvnodes) {
1693 		vn_alloc_cyclecount = 0;
1694 		goto alloc;
1695 	}
1696 	rfreevnodes = vnlru_read_freevnodes();
1697 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1698 		vn_alloc_cyclecount = 0;
1699 		vstir = 1;
1700 	}
1701 	/*
1702 	 * Grow the vnode cache if it will not be above its target max
1703 	 * after growing.  Otherwise, if the free list is nonempty, try
1704 	 * to reclaim 1 item from it before growing the cache (possibly
1705 	 * above its target max if the reclamation failed or is delayed).
1706 	 * Otherwise, wait for some space.  In all cases, schedule
1707 	 * vnlru_proc() if we are getting short of space.  The watermarks
1708 	 * should be chosen so that we never wait or even reclaim from
1709 	 * the free list to below its target minimum.
1710 	 */
1711 	if (vnlru_free_locked(1) > 0)
1712 		goto alloc;
1713 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1714 		/*
1715 		 * Wait for space for a new vnode.
1716 		 */
1717 		vnlru_kick();
1718 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1719 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1720 		    vnlru_read_freevnodes() > 1)
1721 			vnlru_free_locked(1);
1722 	}
1723 alloc:
1724 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1725 	if (vnlru_under(rnumvnodes, vlowat))
1726 		vnlru_kick();
1727 	mtx_unlock(&vnode_list_mtx);
1728 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1729 }
1730 
1731 static struct vnode *
1732 vn_alloc(struct mount *mp)
1733 {
1734 	u_long rnumvnodes;
1735 
1736 	if (__predict_false(vn_alloc_cyclecount != 0))
1737 		return (vn_alloc_hard(mp));
1738 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1739 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1740 		atomic_subtract_long(&numvnodes, 1);
1741 		return (vn_alloc_hard(mp));
1742 	}
1743 
1744 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1745 }
1746 
1747 static void
1748 vn_free(struct vnode *vp)
1749 {
1750 
1751 	atomic_subtract_long(&numvnodes, 1);
1752 	uma_zfree_smr(vnode_zone, vp);
1753 }
1754 
1755 /*
1756  * Return the next vnode from the free list.
1757  */
1758 int
1759 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1760     struct vnode **vpp)
1761 {
1762 	struct vnode *vp;
1763 	struct thread *td;
1764 	struct lock_object *lo;
1765 
1766 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1767 
1768 	KASSERT(vops->registered,
1769 	    ("%s: not registered vector op %p\n", __func__, vops));
1770 
1771 	td = curthread;
1772 	if (td->td_vp_reserved != NULL) {
1773 		vp = td->td_vp_reserved;
1774 		td->td_vp_reserved = NULL;
1775 	} else {
1776 		vp = vn_alloc(mp);
1777 	}
1778 	counter_u64_add(vnodes_created, 1);
1779 	/*
1780 	 * Locks are given the generic name "vnode" when created.
1781 	 * Follow the historic practice of using the filesystem
1782 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1783 	 *
1784 	 * Locks live in a witness group keyed on their name. Thus,
1785 	 * when a lock is renamed, it must also move from the witness
1786 	 * group of its old name to the witness group of its new name.
1787 	 *
1788 	 * The change only needs to be made when the vnode moves
1789 	 * from one filesystem type to another. We ensure that each
1790 	 * filesystem use a single static name pointer for its tag so
1791 	 * that we can compare pointers rather than doing a strcmp().
1792 	 */
1793 	lo = &vp->v_vnlock->lock_object;
1794 #ifdef WITNESS
1795 	if (lo->lo_name != tag) {
1796 #endif
1797 		lo->lo_name = tag;
1798 #ifdef WITNESS
1799 		WITNESS_DESTROY(lo);
1800 		WITNESS_INIT(lo, tag);
1801 	}
1802 #endif
1803 	/*
1804 	 * By default, don't allow shared locks unless filesystems opt-in.
1805 	 */
1806 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1807 	/*
1808 	 * Finalize various vnode identity bits.
1809 	 */
1810 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1811 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1812 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1813 	vp->v_type = VNON;
1814 	vp->v_op = vops;
1815 	vp->v_irflag = 0;
1816 	v_init_counters(vp);
1817 	vn_seqc_init(vp);
1818 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1819 #ifdef DIAGNOSTIC
1820 	if (mp == NULL && vops != &dead_vnodeops)
1821 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1822 #endif
1823 #ifdef MAC
1824 	mac_vnode_init(vp);
1825 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1826 		mac_vnode_associate_singlelabel(mp, vp);
1827 #endif
1828 	if (mp != NULL) {
1829 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1830 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1831 			vp->v_vflag |= VV_NOKNOTE;
1832 	}
1833 
1834 	/*
1835 	 * For the filesystems which do not use vfs_hash_insert(),
1836 	 * still initialize v_hash to have vfs_hash_index() useful.
1837 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1838 	 * its own hashing.
1839 	 */
1840 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1841 
1842 	*vpp = vp;
1843 	return (0);
1844 }
1845 
1846 void
1847 getnewvnode_reserve(void)
1848 {
1849 	struct thread *td;
1850 
1851 	td = curthread;
1852 	MPASS(td->td_vp_reserved == NULL);
1853 	td->td_vp_reserved = vn_alloc(NULL);
1854 }
1855 
1856 void
1857 getnewvnode_drop_reserve(void)
1858 {
1859 	struct thread *td;
1860 
1861 	td = curthread;
1862 	if (td->td_vp_reserved != NULL) {
1863 		vn_free(td->td_vp_reserved);
1864 		td->td_vp_reserved = NULL;
1865 	}
1866 }
1867 
1868 static void __noinline
1869 freevnode(struct vnode *vp)
1870 {
1871 	struct bufobj *bo;
1872 
1873 	/*
1874 	 * The vnode has been marked for destruction, so free it.
1875 	 *
1876 	 * The vnode will be returned to the zone where it will
1877 	 * normally remain until it is needed for another vnode. We
1878 	 * need to cleanup (or verify that the cleanup has already
1879 	 * been done) any residual data left from its current use
1880 	 * so as not to contaminate the freshly allocated vnode.
1881 	 */
1882 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1883 	/*
1884 	 * Paired with vgone.
1885 	 */
1886 	vn_seqc_write_end_free(vp);
1887 
1888 	bo = &vp->v_bufobj;
1889 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1890 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1891 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1892 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1893 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1894 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1895 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1896 	    ("clean blk trie not empty"));
1897 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1898 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1899 	    ("dirty blk trie not empty"));
1900 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1901 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1902 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1903 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1904 	    ("Dangling rangelock waiters"));
1905 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1906 	    ("Leaked inactivation"));
1907 	VI_UNLOCK(vp);
1908 #ifdef MAC
1909 	mac_vnode_destroy(vp);
1910 #endif
1911 	if (vp->v_pollinfo != NULL) {
1912 		destroy_vpollinfo(vp->v_pollinfo);
1913 		vp->v_pollinfo = NULL;
1914 	}
1915 	vp->v_mountedhere = NULL;
1916 	vp->v_unpcb = NULL;
1917 	vp->v_rdev = NULL;
1918 	vp->v_fifoinfo = NULL;
1919 	vp->v_iflag = 0;
1920 	vp->v_vflag = 0;
1921 	bo->bo_flag = 0;
1922 	vn_free(vp);
1923 }
1924 
1925 /*
1926  * Delete from old mount point vnode list, if on one.
1927  */
1928 static void
1929 delmntque(struct vnode *vp)
1930 {
1931 	struct mount *mp;
1932 
1933 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1934 
1935 	mp = vp->v_mount;
1936 	if (mp == NULL)
1937 		return;
1938 	MNT_ILOCK(mp);
1939 	VI_LOCK(vp);
1940 	vp->v_mount = NULL;
1941 	VI_UNLOCK(vp);
1942 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1943 		("bad mount point vnode list size"));
1944 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1945 	mp->mnt_nvnodelistsize--;
1946 	MNT_REL(mp);
1947 	MNT_IUNLOCK(mp);
1948 }
1949 
1950 static void
1951 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1952 {
1953 
1954 	vp->v_data = NULL;
1955 	vp->v_op = &dead_vnodeops;
1956 	vgone(vp);
1957 	vput(vp);
1958 }
1959 
1960 /*
1961  * Insert into list of vnodes for the new mount point, if available.
1962  */
1963 int
1964 insmntque1(struct vnode *vp, struct mount *mp,
1965 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1966 {
1967 
1968 	KASSERT(vp->v_mount == NULL,
1969 		("insmntque: vnode already on per mount vnode list"));
1970 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1971 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1972 
1973 	/*
1974 	 * We acquire the vnode interlock early to ensure that the
1975 	 * vnode cannot be recycled by another process releasing a
1976 	 * holdcnt on it before we get it on both the vnode list
1977 	 * and the active vnode list. The mount mutex protects only
1978 	 * manipulation of the vnode list and the vnode freelist
1979 	 * mutex protects only manipulation of the active vnode list.
1980 	 * Hence the need to hold the vnode interlock throughout.
1981 	 */
1982 	MNT_ILOCK(mp);
1983 	VI_LOCK(vp);
1984 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1985 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1986 	    mp->mnt_nvnodelistsize == 0)) &&
1987 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1988 		VI_UNLOCK(vp);
1989 		MNT_IUNLOCK(mp);
1990 		if (dtr != NULL)
1991 			dtr(vp, dtr_arg);
1992 		return (EBUSY);
1993 	}
1994 	vp->v_mount = mp;
1995 	MNT_REF(mp);
1996 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1997 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1998 		("neg mount point vnode list size"));
1999 	mp->mnt_nvnodelistsize++;
2000 	VI_UNLOCK(vp);
2001 	MNT_IUNLOCK(mp);
2002 	return (0);
2003 }
2004 
2005 int
2006 insmntque(struct vnode *vp, struct mount *mp)
2007 {
2008 
2009 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
2010 }
2011 
2012 /*
2013  * Flush out and invalidate all buffers associated with a bufobj
2014  * Called with the underlying object locked.
2015  */
2016 int
2017 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2018 {
2019 	int error;
2020 
2021 	BO_LOCK(bo);
2022 	if (flags & V_SAVE) {
2023 		error = bufobj_wwait(bo, slpflag, slptimeo);
2024 		if (error) {
2025 			BO_UNLOCK(bo);
2026 			return (error);
2027 		}
2028 		if (bo->bo_dirty.bv_cnt > 0) {
2029 			BO_UNLOCK(bo);
2030 			do {
2031 				error = BO_SYNC(bo, MNT_WAIT);
2032 			} while (error == ERELOOKUP);
2033 			if (error != 0)
2034 				return (error);
2035 			BO_LOCK(bo);
2036 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2037 				BO_UNLOCK(bo);
2038 				return (EBUSY);
2039 			}
2040 		}
2041 	}
2042 	/*
2043 	 * If you alter this loop please notice that interlock is dropped and
2044 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2045 	 * no race conditions occur from this.
2046 	 */
2047 	do {
2048 		error = flushbuflist(&bo->bo_clean,
2049 		    flags, bo, slpflag, slptimeo);
2050 		if (error == 0 && !(flags & V_CLEANONLY))
2051 			error = flushbuflist(&bo->bo_dirty,
2052 			    flags, bo, slpflag, slptimeo);
2053 		if (error != 0 && error != EAGAIN) {
2054 			BO_UNLOCK(bo);
2055 			return (error);
2056 		}
2057 	} while (error != 0);
2058 
2059 	/*
2060 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2061 	 * have write I/O in-progress but if there is a VM object then the
2062 	 * VM object can also have read-I/O in-progress.
2063 	 */
2064 	do {
2065 		bufobj_wwait(bo, 0, 0);
2066 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2067 			BO_UNLOCK(bo);
2068 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2069 			BO_LOCK(bo);
2070 		}
2071 	} while (bo->bo_numoutput > 0);
2072 	BO_UNLOCK(bo);
2073 
2074 	/*
2075 	 * Destroy the copy in the VM cache, too.
2076 	 */
2077 	if (bo->bo_object != NULL &&
2078 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2079 		VM_OBJECT_WLOCK(bo->bo_object);
2080 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2081 		    OBJPR_CLEANONLY : 0);
2082 		VM_OBJECT_WUNLOCK(bo->bo_object);
2083 	}
2084 
2085 #ifdef INVARIANTS
2086 	BO_LOCK(bo);
2087 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2088 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2089 	    bo->bo_clean.bv_cnt > 0))
2090 		panic("vinvalbuf: flush failed");
2091 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2092 	    bo->bo_dirty.bv_cnt > 0)
2093 		panic("vinvalbuf: flush dirty failed");
2094 	BO_UNLOCK(bo);
2095 #endif
2096 	return (0);
2097 }
2098 
2099 /*
2100  * Flush out and invalidate all buffers associated with a vnode.
2101  * Called with the underlying object locked.
2102  */
2103 int
2104 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2105 {
2106 
2107 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2108 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2109 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2110 		return (0);
2111 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2112 }
2113 
2114 /*
2115  * Flush out buffers on the specified list.
2116  *
2117  */
2118 static int
2119 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2120     int slptimeo)
2121 {
2122 	struct buf *bp, *nbp;
2123 	int retval, error;
2124 	daddr_t lblkno;
2125 	b_xflags_t xflags;
2126 
2127 	ASSERT_BO_WLOCKED(bo);
2128 
2129 	retval = 0;
2130 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2131 		/*
2132 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2133 		 * do not skip any buffers. If we are flushing only V_NORMAL
2134 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2135 		 * flushing only V_ALT buffers then skip buffers not marked
2136 		 * as BX_ALTDATA.
2137 		 */
2138 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2139 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2140 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2141 			continue;
2142 		}
2143 		if (nbp != NULL) {
2144 			lblkno = nbp->b_lblkno;
2145 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2146 		}
2147 		retval = EAGAIN;
2148 		error = BUF_TIMELOCK(bp,
2149 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2150 		    "flushbuf", slpflag, slptimeo);
2151 		if (error) {
2152 			BO_LOCK(bo);
2153 			return (error != ENOLCK ? error : EAGAIN);
2154 		}
2155 		KASSERT(bp->b_bufobj == bo,
2156 		    ("bp %p wrong b_bufobj %p should be %p",
2157 		    bp, bp->b_bufobj, bo));
2158 		/*
2159 		 * XXX Since there are no node locks for NFS, I
2160 		 * believe there is a slight chance that a delayed
2161 		 * write will occur while sleeping just above, so
2162 		 * check for it.
2163 		 */
2164 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2165 		    (flags & V_SAVE)) {
2166 			bremfree(bp);
2167 			bp->b_flags |= B_ASYNC;
2168 			bwrite(bp);
2169 			BO_LOCK(bo);
2170 			return (EAGAIN);	/* XXX: why not loop ? */
2171 		}
2172 		bremfree(bp);
2173 		bp->b_flags |= (B_INVAL | B_RELBUF);
2174 		bp->b_flags &= ~B_ASYNC;
2175 		brelse(bp);
2176 		BO_LOCK(bo);
2177 		if (nbp == NULL)
2178 			break;
2179 		nbp = gbincore(bo, lblkno);
2180 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2181 		    != xflags)
2182 			break;			/* nbp invalid */
2183 	}
2184 	return (retval);
2185 }
2186 
2187 int
2188 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2189 {
2190 	struct buf *bp;
2191 	int error;
2192 	daddr_t lblkno;
2193 
2194 	ASSERT_BO_LOCKED(bo);
2195 
2196 	for (lblkno = startn;;) {
2197 again:
2198 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2199 		if (bp == NULL || bp->b_lblkno >= endn ||
2200 		    bp->b_lblkno < startn)
2201 			break;
2202 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2203 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2204 		if (error != 0) {
2205 			BO_RLOCK(bo);
2206 			if (error == ENOLCK)
2207 				goto again;
2208 			return (error);
2209 		}
2210 		KASSERT(bp->b_bufobj == bo,
2211 		    ("bp %p wrong b_bufobj %p should be %p",
2212 		    bp, bp->b_bufobj, bo));
2213 		lblkno = bp->b_lblkno + 1;
2214 		if ((bp->b_flags & B_MANAGED) == 0)
2215 			bremfree(bp);
2216 		bp->b_flags |= B_RELBUF;
2217 		/*
2218 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2219 		 * pages backing each buffer in the range are unlikely to be
2220 		 * reused.  Dirty buffers will have the hint applied once
2221 		 * they've been written.
2222 		 */
2223 		if ((bp->b_flags & B_VMIO) != 0)
2224 			bp->b_flags |= B_NOREUSE;
2225 		brelse(bp);
2226 		BO_RLOCK(bo);
2227 	}
2228 	return (0);
2229 }
2230 
2231 /*
2232  * Truncate a file's buffer and pages to a specified length.  This
2233  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2234  * sync activity.
2235  */
2236 int
2237 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2238 {
2239 	struct buf *bp, *nbp;
2240 	struct bufobj *bo;
2241 	daddr_t startlbn;
2242 
2243 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2244 	    vp, blksize, (uintmax_t)length);
2245 
2246 	/*
2247 	 * Round up to the *next* lbn.
2248 	 */
2249 	startlbn = howmany(length, blksize);
2250 
2251 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2252 
2253 	bo = &vp->v_bufobj;
2254 restart_unlocked:
2255 	BO_LOCK(bo);
2256 
2257 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2258 		;
2259 
2260 	if (length > 0) {
2261 restartsync:
2262 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2263 			if (bp->b_lblkno > 0)
2264 				continue;
2265 			/*
2266 			 * Since we hold the vnode lock this should only
2267 			 * fail if we're racing with the buf daemon.
2268 			 */
2269 			if (BUF_LOCK(bp,
2270 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2271 			    BO_LOCKPTR(bo)) == ENOLCK)
2272 				goto restart_unlocked;
2273 
2274 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2275 			    ("buf(%p) on dirty queue without DELWRI", bp));
2276 
2277 			bremfree(bp);
2278 			bawrite(bp);
2279 			BO_LOCK(bo);
2280 			goto restartsync;
2281 		}
2282 	}
2283 
2284 	bufobj_wwait(bo, 0, 0);
2285 	BO_UNLOCK(bo);
2286 	vnode_pager_setsize(vp, length);
2287 
2288 	return (0);
2289 }
2290 
2291 /*
2292  * Invalidate the cached pages of a file's buffer within the range of block
2293  * numbers [startlbn, endlbn).
2294  */
2295 void
2296 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2297     int blksize)
2298 {
2299 	struct bufobj *bo;
2300 	off_t start, end;
2301 
2302 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2303 
2304 	start = blksize * startlbn;
2305 	end = blksize * endlbn;
2306 
2307 	bo = &vp->v_bufobj;
2308 	BO_LOCK(bo);
2309 	MPASS(blksize == bo->bo_bsize);
2310 
2311 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2312 		;
2313 
2314 	BO_UNLOCK(bo);
2315 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2316 }
2317 
2318 static int
2319 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2320     daddr_t startlbn, daddr_t endlbn)
2321 {
2322 	struct buf *bp, *nbp;
2323 	bool anyfreed;
2324 
2325 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2326 	ASSERT_BO_LOCKED(bo);
2327 
2328 	do {
2329 		anyfreed = false;
2330 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2331 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2332 				continue;
2333 			if (BUF_LOCK(bp,
2334 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2335 			    BO_LOCKPTR(bo)) == ENOLCK) {
2336 				BO_LOCK(bo);
2337 				return (EAGAIN);
2338 			}
2339 
2340 			bremfree(bp);
2341 			bp->b_flags |= B_INVAL | B_RELBUF;
2342 			bp->b_flags &= ~B_ASYNC;
2343 			brelse(bp);
2344 			anyfreed = true;
2345 
2346 			BO_LOCK(bo);
2347 			if (nbp != NULL &&
2348 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2349 			    nbp->b_vp != vp ||
2350 			    (nbp->b_flags & B_DELWRI) != 0))
2351 				return (EAGAIN);
2352 		}
2353 
2354 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2355 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2356 				continue;
2357 			if (BUF_LOCK(bp,
2358 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2359 			    BO_LOCKPTR(bo)) == ENOLCK) {
2360 				BO_LOCK(bo);
2361 				return (EAGAIN);
2362 			}
2363 			bremfree(bp);
2364 			bp->b_flags |= B_INVAL | B_RELBUF;
2365 			bp->b_flags &= ~B_ASYNC;
2366 			brelse(bp);
2367 			anyfreed = true;
2368 
2369 			BO_LOCK(bo);
2370 			if (nbp != NULL &&
2371 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2372 			    (nbp->b_vp != vp) ||
2373 			    (nbp->b_flags & B_DELWRI) == 0))
2374 				return (EAGAIN);
2375 		}
2376 	} while (anyfreed);
2377 	return (0);
2378 }
2379 
2380 static void
2381 buf_vlist_remove(struct buf *bp)
2382 {
2383 	struct bufv *bv;
2384 	b_xflags_t flags;
2385 
2386 	flags = bp->b_xflags;
2387 
2388 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2389 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2390 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2391 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2392 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2393 
2394 	if ((flags & BX_VNDIRTY) != 0)
2395 		bv = &bp->b_bufobj->bo_dirty;
2396 	else
2397 		bv = &bp->b_bufobj->bo_clean;
2398 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2399 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2400 	bv->bv_cnt--;
2401 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2402 }
2403 
2404 /*
2405  * Add the buffer to the sorted clean or dirty block list.
2406  *
2407  * NOTE: xflags is passed as a constant, optimizing this inline function!
2408  */
2409 static void
2410 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2411 {
2412 	struct bufv *bv;
2413 	struct buf *n;
2414 	int error;
2415 
2416 	ASSERT_BO_WLOCKED(bo);
2417 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2418 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2419 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2420 	    ("dead bo %p", bo));
2421 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2422 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2423 	bp->b_xflags |= xflags;
2424 	if (xflags & BX_VNDIRTY)
2425 		bv = &bo->bo_dirty;
2426 	else
2427 		bv = &bo->bo_clean;
2428 
2429 	/*
2430 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2431 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2432 	 * than _ge.
2433 	 */
2434 	if (bv->bv_cnt == 0 ||
2435 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2436 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2437 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2438 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2439 	else
2440 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2441 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2442 	if (error)
2443 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2444 	bv->bv_cnt++;
2445 }
2446 
2447 /*
2448  * Look up a buffer using the buffer tries.
2449  */
2450 struct buf *
2451 gbincore(struct bufobj *bo, daddr_t lblkno)
2452 {
2453 	struct buf *bp;
2454 
2455 	ASSERT_BO_LOCKED(bo);
2456 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2457 	if (bp != NULL)
2458 		return (bp);
2459 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2460 }
2461 
2462 /*
2463  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2464  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2465  * stability of the result.  Like other lockless lookups, the found buf may
2466  * already be invalid by the time this function returns.
2467  */
2468 struct buf *
2469 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2470 {
2471 	struct buf *bp;
2472 
2473 	ASSERT_BO_UNLOCKED(bo);
2474 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2475 	if (bp != NULL)
2476 		return (bp);
2477 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2478 }
2479 
2480 /*
2481  * Associate a buffer with a vnode.
2482  */
2483 void
2484 bgetvp(struct vnode *vp, struct buf *bp)
2485 {
2486 	struct bufobj *bo;
2487 
2488 	bo = &vp->v_bufobj;
2489 	ASSERT_BO_WLOCKED(bo);
2490 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2491 
2492 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2493 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2494 	    ("bgetvp: bp already attached! %p", bp));
2495 
2496 	vhold(vp);
2497 	bp->b_vp = vp;
2498 	bp->b_bufobj = bo;
2499 	/*
2500 	 * Insert onto list for new vnode.
2501 	 */
2502 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2503 }
2504 
2505 /*
2506  * Disassociate a buffer from a vnode.
2507  */
2508 void
2509 brelvp(struct buf *bp)
2510 {
2511 	struct bufobj *bo;
2512 	struct vnode *vp;
2513 
2514 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2515 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2516 
2517 	/*
2518 	 * Delete from old vnode list, if on one.
2519 	 */
2520 	vp = bp->b_vp;		/* XXX */
2521 	bo = bp->b_bufobj;
2522 	BO_LOCK(bo);
2523 	buf_vlist_remove(bp);
2524 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2525 		bo->bo_flag &= ~BO_ONWORKLST;
2526 		mtx_lock(&sync_mtx);
2527 		LIST_REMOVE(bo, bo_synclist);
2528 		syncer_worklist_len--;
2529 		mtx_unlock(&sync_mtx);
2530 	}
2531 	bp->b_vp = NULL;
2532 	bp->b_bufobj = NULL;
2533 	BO_UNLOCK(bo);
2534 	vdrop(vp);
2535 }
2536 
2537 /*
2538  * Add an item to the syncer work queue.
2539  */
2540 static void
2541 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2542 {
2543 	int slot;
2544 
2545 	ASSERT_BO_WLOCKED(bo);
2546 
2547 	mtx_lock(&sync_mtx);
2548 	if (bo->bo_flag & BO_ONWORKLST)
2549 		LIST_REMOVE(bo, bo_synclist);
2550 	else {
2551 		bo->bo_flag |= BO_ONWORKLST;
2552 		syncer_worklist_len++;
2553 	}
2554 
2555 	if (delay > syncer_maxdelay - 2)
2556 		delay = syncer_maxdelay - 2;
2557 	slot = (syncer_delayno + delay) & syncer_mask;
2558 
2559 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2560 	mtx_unlock(&sync_mtx);
2561 }
2562 
2563 static int
2564 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2565 {
2566 	int error, len;
2567 
2568 	mtx_lock(&sync_mtx);
2569 	len = syncer_worklist_len - sync_vnode_count;
2570 	mtx_unlock(&sync_mtx);
2571 	error = SYSCTL_OUT(req, &len, sizeof(len));
2572 	return (error);
2573 }
2574 
2575 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2576     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2577     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2578 
2579 static struct proc *updateproc;
2580 static void sched_sync(void);
2581 static struct kproc_desc up_kp = {
2582 	"syncer",
2583 	sched_sync,
2584 	&updateproc
2585 };
2586 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2587 
2588 static int
2589 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2590 {
2591 	struct vnode *vp;
2592 	struct mount *mp;
2593 
2594 	*bo = LIST_FIRST(slp);
2595 	if (*bo == NULL)
2596 		return (0);
2597 	vp = bo2vnode(*bo);
2598 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2599 		return (1);
2600 	/*
2601 	 * We use vhold in case the vnode does not
2602 	 * successfully sync.  vhold prevents the vnode from
2603 	 * going away when we unlock the sync_mtx so that
2604 	 * we can acquire the vnode interlock.
2605 	 */
2606 	vholdl(vp);
2607 	mtx_unlock(&sync_mtx);
2608 	VI_UNLOCK(vp);
2609 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2610 		vdrop(vp);
2611 		mtx_lock(&sync_mtx);
2612 		return (*bo == LIST_FIRST(slp));
2613 	}
2614 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2615 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2616 	VOP_UNLOCK(vp);
2617 	vn_finished_write(mp);
2618 	BO_LOCK(*bo);
2619 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2620 		/*
2621 		 * Put us back on the worklist.  The worklist
2622 		 * routine will remove us from our current
2623 		 * position and then add us back in at a later
2624 		 * position.
2625 		 */
2626 		vn_syncer_add_to_worklist(*bo, syncdelay);
2627 	}
2628 	BO_UNLOCK(*bo);
2629 	vdrop(vp);
2630 	mtx_lock(&sync_mtx);
2631 	return (0);
2632 }
2633 
2634 static int first_printf = 1;
2635 
2636 /*
2637  * System filesystem synchronizer daemon.
2638  */
2639 static void
2640 sched_sync(void)
2641 {
2642 	struct synclist *next, *slp;
2643 	struct bufobj *bo;
2644 	long starttime;
2645 	struct thread *td = curthread;
2646 	int last_work_seen;
2647 	int net_worklist_len;
2648 	int syncer_final_iter;
2649 	int error;
2650 
2651 	last_work_seen = 0;
2652 	syncer_final_iter = 0;
2653 	syncer_state = SYNCER_RUNNING;
2654 	starttime = time_uptime;
2655 	td->td_pflags |= TDP_NORUNNINGBUF;
2656 
2657 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2658 	    SHUTDOWN_PRI_LAST);
2659 
2660 	mtx_lock(&sync_mtx);
2661 	for (;;) {
2662 		if (syncer_state == SYNCER_FINAL_DELAY &&
2663 		    syncer_final_iter == 0) {
2664 			mtx_unlock(&sync_mtx);
2665 			kproc_suspend_check(td->td_proc);
2666 			mtx_lock(&sync_mtx);
2667 		}
2668 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2669 		if (syncer_state != SYNCER_RUNNING &&
2670 		    starttime != time_uptime) {
2671 			if (first_printf) {
2672 				printf("\nSyncing disks, vnodes remaining... ");
2673 				first_printf = 0;
2674 			}
2675 			printf("%d ", net_worklist_len);
2676 		}
2677 		starttime = time_uptime;
2678 
2679 		/*
2680 		 * Push files whose dirty time has expired.  Be careful
2681 		 * of interrupt race on slp queue.
2682 		 *
2683 		 * Skip over empty worklist slots when shutting down.
2684 		 */
2685 		do {
2686 			slp = &syncer_workitem_pending[syncer_delayno];
2687 			syncer_delayno += 1;
2688 			if (syncer_delayno == syncer_maxdelay)
2689 				syncer_delayno = 0;
2690 			next = &syncer_workitem_pending[syncer_delayno];
2691 			/*
2692 			 * If the worklist has wrapped since the
2693 			 * it was emptied of all but syncer vnodes,
2694 			 * switch to the FINAL_DELAY state and run
2695 			 * for one more second.
2696 			 */
2697 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2698 			    net_worklist_len == 0 &&
2699 			    last_work_seen == syncer_delayno) {
2700 				syncer_state = SYNCER_FINAL_DELAY;
2701 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2702 			}
2703 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2704 		    syncer_worklist_len > 0);
2705 
2706 		/*
2707 		 * Keep track of the last time there was anything
2708 		 * on the worklist other than syncer vnodes.
2709 		 * Return to the SHUTTING_DOWN state if any
2710 		 * new work appears.
2711 		 */
2712 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2713 			last_work_seen = syncer_delayno;
2714 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2715 			syncer_state = SYNCER_SHUTTING_DOWN;
2716 		while (!LIST_EMPTY(slp)) {
2717 			error = sync_vnode(slp, &bo, td);
2718 			if (error == 1) {
2719 				LIST_REMOVE(bo, bo_synclist);
2720 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2721 				continue;
2722 			}
2723 
2724 			if (first_printf == 0) {
2725 				/*
2726 				 * Drop the sync mutex, because some watchdog
2727 				 * drivers need to sleep while patting
2728 				 */
2729 				mtx_unlock(&sync_mtx);
2730 				wdog_kern_pat(WD_LASTVAL);
2731 				mtx_lock(&sync_mtx);
2732 			}
2733 		}
2734 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2735 			syncer_final_iter--;
2736 		/*
2737 		 * The variable rushjob allows the kernel to speed up the
2738 		 * processing of the filesystem syncer process. A rushjob
2739 		 * value of N tells the filesystem syncer to process the next
2740 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2741 		 * is used by the soft update code to speed up the filesystem
2742 		 * syncer process when the incore state is getting so far
2743 		 * ahead of the disk that the kernel memory pool is being
2744 		 * threatened with exhaustion.
2745 		 */
2746 		if (rushjob > 0) {
2747 			rushjob -= 1;
2748 			continue;
2749 		}
2750 		/*
2751 		 * Just sleep for a short period of time between
2752 		 * iterations when shutting down to allow some I/O
2753 		 * to happen.
2754 		 *
2755 		 * If it has taken us less than a second to process the
2756 		 * current work, then wait. Otherwise start right over
2757 		 * again. We can still lose time if any single round
2758 		 * takes more than two seconds, but it does not really
2759 		 * matter as we are just trying to generally pace the
2760 		 * filesystem activity.
2761 		 */
2762 		if (syncer_state != SYNCER_RUNNING ||
2763 		    time_uptime == starttime) {
2764 			thread_lock(td);
2765 			sched_prio(td, PPAUSE);
2766 			thread_unlock(td);
2767 		}
2768 		if (syncer_state != SYNCER_RUNNING)
2769 			cv_timedwait(&sync_wakeup, &sync_mtx,
2770 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2771 		else if (time_uptime == starttime)
2772 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2773 	}
2774 }
2775 
2776 /*
2777  * Request the syncer daemon to speed up its work.
2778  * We never push it to speed up more than half of its
2779  * normal turn time, otherwise it could take over the cpu.
2780  */
2781 int
2782 speedup_syncer(void)
2783 {
2784 	int ret = 0;
2785 
2786 	mtx_lock(&sync_mtx);
2787 	if (rushjob < syncdelay / 2) {
2788 		rushjob += 1;
2789 		stat_rush_requests += 1;
2790 		ret = 1;
2791 	}
2792 	mtx_unlock(&sync_mtx);
2793 	cv_broadcast(&sync_wakeup);
2794 	return (ret);
2795 }
2796 
2797 /*
2798  * Tell the syncer to speed up its work and run though its work
2799  * list several times, then tell it to shut down.
2800  */
2801 static void
2802 syncer_shutdown(void *arg, int howto)
2803 {
2804 
2805 	if (howto & RB_NOSYNC)
2806 		return;
2807 	mtx_lock(&sync_mtx);
2808 	syncer_state = SYNCER_SHUTTING_DOWN;
2809 	rushjob = 0;
2810 	mtx_unlock(&sync_mtx);
2811 	cv_broadcast(&sync_wakeup);
2812 	kproc_shutdown(arg, howto);
2813 }
2814 
2815 void
2816 syncer_suspend(void)
2817 {
2818 
2819 	syncer_shutdown(updateproc, 0);
2820 }
2821 
2822 void
2823 syncer_resume(void)
2824 {
2825 
2826 	mtx_lock(&sync_mtx);
2827 	first_printf = 1;
2828 	syncer_state = SYNCER_RUNNING;
2829 	mtx_unlock(&sync_mtx);
2830 	cv_broadcast(&sync_wakeup);
2831 	kproc_resume(updateproc);
2832 }
2833 
2834 /*
2835  * Move the buffer between the clean and dirty lists of its vnode.
2836  */
2837 void
2838 reassignbuf(struct buf *bp)
2839 {
2840 	struct vnode *vp;
2841 	struct bufobj *bo;
2842 	int delay;
2843 #ifdef INVARIANTS
2844 	struct bufv *bv;
2845 #endif
2846 
2847 	vp = bp->b_vp;
2848 	bo = bp->b_bufobj;
2849 
2850 	KASSERT((bp->b_flags & B_PAGING) == 0,
2851 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2852 
2853 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2854 	    bp, bp->b_vp, bp->b_flags);
2855 
2856 	BO_LOCK(bo);
2857 	buf_vlist_remove(bp);
2858 
2859 	/*
2860 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2861 	 * of clean buffers.
2862 	 */
2863 	if (bp->b_flags & B_DELWRI) {
2864 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2865 			switch (vp->v_type) {
2866 			case VDIR:
2867 				delay = dirdelay;
2868 				break;
2869 			case VCHR:
2870 				delay = metadelay;
2871 				break;
2872 			default:
2873 				delay = filedelay;
2874 			}
2875 			vn_syncer_add_to_worklist(bo, delay);
2876 		}
2877 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2878 	} else {
2879 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2880 
2881 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2882 			mtx_lock(&sync_mtx);
2883 			LIST_REMOVE(bo, bo_synclist);
2884 			syncer_worklist_len--;
2885 			mtx_unlock(&sync_mtx);
2886 			bo->bo_flag &= ~BO_ONWORKLST;
2887 		}
2888 	}
2889 #ifdef INVARIANTS
2890 	bv = &bo->bo_clean;
2891 	bp = TAILQ_FIRST(&bv->bv_hd);
2892 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2893 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2894 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2895 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2896 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2897 	bv = &bo->bo_dirty;
2898 	bp = TAILQ_FIRST(&bv->bv_hd);
2899 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2900 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2901 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2902 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2903 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2904 #endif
2905 	BO_UNLOCK(bo);
2906 }
2907 
2908 static void
2909 v_init_counters(struct vnode *vp)
2910 {
2911 
2912 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2913 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2914 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2915 
2916 	refcount_init(&vp->v_holdcnt, 1);
2917 	refcount_init(&vp->v_usecount, 1);
2918 }
2919 
2920 /*
2921  * Grab a particular vnode from the free list, increment its
2922  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2923  * is being destroyed.  Only callers who specify LK_RETRY will
2924  * see doomed vnodes.  If inactive processing was delayed in
2925  * vput try to do it here.
2926  *
2927  * usecount is manipulated using atomics without holding any locks.
2928  *
2929  * holdcnt can be manipulated using atomics without holding any locks,
2930  * except when transitioning 1<->0, in which case the interlock is held.
2931  *
2932  * Consumers which don't guarantee liveness of the vnode can use SMR to
2933  * try to get a reference. Note this operation can fail since the vnode
2934  * may be awaiting getting freed by the time they get to it.
2935  */
2936 enum vgetstate
2937 vget_prep_smr(struct vnode *vp)
2938 {
2939 	enum vgetstate vs;
2940 
2941 	VFS_SMR_ASSERT_ENTERED();
2942 
2943 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2944 		vs = VGET_USECOUNT;
2945 	} else {
2946 		if (vhold_smr(vp))
2947 			vs = VGET_HOLDCNT;
2948 		else
2949 			vs = VGET_NONE;
2950 	}
2951 	return (vs);
2952 }
2953 
2954 enum vgetstate
2955 vget_prep(struct vnode *vp)
2956 {
2957 	enum vgetstate vs;
2958 
2959 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2960 		vs = VGET_USECOUNT;
2961 	} else {
2962 		vhold(vp);
2963 		vs = VGET_HOLDCNT;
2964 	}
2965 	return (vs);
2966 }
2967 
2968 void
2969 vget_abort(struct vnode *vp, enum vgetstate vs)
2970 {
2971 
2972 	switch (vs) {
2973 	case VGET_USECOUNT:
2974 		vrele(vp);
2975 		break;
2976 	case VGET_HOLDCNT:
2977 		vdrop(vp);
2978 		break;
2979 	default:
2980 		__assert_unreachable();
2981 	}
2982 }
2983 
2984 int
2985 vget(struct vnode *vp, int flags)
2986 {
2987 	enum vgetstate vs;
2988 
2989 	vs = vget_prep(vp);
2990 	return (vget_finish(vp, flags, vs));
2991 }
2992 
2993 int
2994 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2995 {
2996 	int error;
2997 
2998 	if ((flags & LK_INTERLOCK) != 0)
2999 		ASSERT_VI_LOCKED(vp, __func__);
3000 	else
3001 		ASSERT_VI_UNLOCKED(vp, __func__);
3002 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3003 	VNPASS(vp->v_holdcnt > 0, vp);
3004 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3005 
3006 	error = vn_lock(vp, flags);
3007 	if (__predict_false(error != 0)) {
3008 		vget_abort(vp, vs);
3009 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3010 		    vp);
3011 		return (error);
3012 	}
3013 
3014 	vget_finish_ref(vp, vs);
3015 	return (0);
3016 }
3017 
3018 void
3019 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3020 {
3021 	int old;
3022 
3023 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3024 	VNPASS(vp->v_holdcnt > 0, vp);
3025 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3026 
3027 	if (vs == VGET_USECOUNT)
3028 		return;
3029 
3030 	/*
3031 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3032 	 * the vnode around. Otherwise someone else lended their hold count and
3033 	 * we have to drop ours.
3034 	 */
3035 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3036 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3037 	if (old != 0) {
3038 #ifdef INVARIANTS
3039 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3040 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3041 #else
3042 		refcount_release(&vp->v_holdcnt);
3043 #endif
3044 	}
3045 }
3046 
3047 void
3048 vref(struct vnode *vp)
3049 {
3050 	enum vgetstate vs;
3051 
3052 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3053 	vs = vget_prep(vp);
3054 	vget_finish_ref(vp, vs);
3055 }
3056 
3057 void
3058 vrefact(struct vnode *vp)
3059 {
3060 
3061 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3062 #ifdef INVARIANTS
3063 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3064 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3065 #else
3066 	refcount_acquire(&vp->v_usecount);
3067 #endif
3068 }
3069 
3070 void
3071 vlazy(struct vnode *vp)
3072 {
3073 	struct mount *mp;
3074 
3075 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3076 
3077 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3078 		return;
3079 	/*
3080 	 * We may get here for inactive routines after the vnode got doomed.
3081 	 */
3082 	if (VN_IS_DOOMED(vp))
3083 		return;
3084 	mp = vp->v_mount;
3085 	mtx_lock(&mp->mnt_listmtx);
3086 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3087 		vp->v_mflag |= VMP_LAZYLIST;
3088 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3089 		mp->mnt_lazyvnodelistsize++;
3090 	}
3091 	mtx_unlock(&mp->mnt_listmtx);
3092 }
3093 
3094 static void
3095 vunlazy(struct vnode *vp)
3096 {
3097 	struct mount *mp;
3098 
3099 	ASSERT_VI_LOCKED(vp, __func__);
3100 	VNPASS(!VN_IS_DOOMED(vp), vp);
3101 
3102 	mp = vp->v_mount;
3103 	mtx_lock(&mp->mnt_listmtx);
3104 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3105 	/*
3106 	 * Don't remove the vnode from the lazy list if another thread
3107 	 * has increased the hold count. It may have re-enqueued the
3108 	 * vnode to the lazy list and is now responsible for its
3109 	 * removal.
3110 	 */
3111 	if (vp->v_holdcnt == 0) {
3112 		vp->v_mflag &= ~VMP_LAZYLIST;
3113 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3114 		mp->mnt_lazyvnodelistsize--;
3115 	}
3116 	mtx_unlock(&mp->mnt_listmtx);
3117 }
3118 
3119 /*
3120  * This routine is only meant to be called from vgonel prior to dooming
3121  * the vnode.
3122  */
3123 static void
3124 vunlazy_gone(struct vnode *vp)
3125 {
3126 	struct mount *mp;
3127 
3128 	ASSERT_VOP_ELOCKED(vp, __func__);
3129 	ASSERT_VI_LOCKED(vp, __func__);
3130 	VNPASS(!VN_IS_DOOMED(vp), vp);
3131 
3132 	if (vp->v_mflag & VMP_LAZYLIST) {
3133 		mp = vp->v_mount;
3134 		mtx_lock(&mp->mnt_listmtx);
3135 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3136 		vp->v_mflag &= ~VMP_LAZYLIST;
3137 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3138 		mp->mnt_lazyvnodelistsize--;
3139 		mtx_unlock(&mp->mnt_listmtx);
3140 	}
3141 }
3142 
3143 static void
3144 vdefer_inactive(struct vnode *vp)
3145 {
3146 
3147 	ASSERT_VI_LOCKED(vp, __func__);
3148 	VNASSERT(vp->v_holdcnt > 0, vp,
3149 	    ("%s: vnode without hold count", __func__));
3150 	if (VN_IS_DOOMED(vp)) {
3151 		vdropl(vp);
3152 		return;
3153 	}
3154 	if (vp->v_iflag & VI_DEFINACT) {
3155 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3156 		vdropl(vp);
3157 		return;
3158 	}
3159 	if (vp->v_usecount > 0) {
3160 		vp->v_iflag &= ~VI_OWEINACT;
3161 		vdropl(vp);
3162 		return;
3163 	}
3164 	vlazy(vp);
3165 	vp->v_iflag |= VI_DEFINACT;
3166 	VI_UNLOCK(vp);
3167 	counter_u64_add(deferred_inact, 1);
3168 }
3169 
3170 static void
3171 vdefer_inactive_unlocked(struct vnode *vp)
3172 {
3173 
3174 	VI_LOCK(vp);
3175 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3176 		vdropl(vp);
3177 		return;
3178 	}
3179 	vdefer_inactive(vp);
3180 }
3181 
3182 enum vput_op { VRELE, VPUT, VUNREF };
3183 
3184 /*
3185  * Handle ->v_usecount transitioning to 0.
3186  *
3187  * By releasing the last usecount we take ownership of the hold count which
3188  * provides liveness of the vnode, meaning we have to vdrop.
3189  *
3190  * For all vnodes we may need to perform inactive processing. It requires an
3191  * exclusive lock on the vnode, while it is legal to call here with only a
3192  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3193  * inactive processing gets deferred to the syncer.
3194  *
3195  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3196  * on the lock being held all the way until VOP_INACTIVE. This in particular
3197  * happens with UFS which adds half-constructed vnodes to the hash, where they
3198  * can be found by other code.
3199  */
3200 static void
3201 vput_final(struct vnode *vp, enum vput_op func)
3202 {
3203 	int error;
3204 	bool want_unlock;
3205 
3206 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3207 	VNPASS(vp->v_holdcnt > 0, vp);
3208 
3209 	VI_LOCK(vp);
3210 
3211 	/*
3212 	 * By the time we got here someone else might have transitioned
3213 	 * the count back to > 0.
3214 	 */
3215 	if (vp->v_usecount > 0)
3216 		goto out;
3217 
3218 	/*
3219 	 * If the vnode is doomed vgone already performed inactive processing
3220 	 * (if needed).
3221 	 */
3222 	if (VN_IS_DOOMED(vp))
3223 		goto out;
3224 
3225 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3226 		goto out;
3227 
3228 	if (vp->v_iflag & VI_DOINGINACT)
3229 		goto out;
3230 
3231 	/*
3232 	 * Locking operations here will drop the interlock and possibly the
3233 	 * vnode lock, opening a window where the vnode can get doomed all the
3234 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3235 	 * perform inactive.
3236 	 */
3237 	vp->v_iflag |= VI_OWEINACT;
3238 	want_unlock = false;
3239 	error = 0;
3240 	switch (func) {
3241 	case VRELE:
3242 		switch (VOP_ISLOCKED(vp)) {
3243 		case LK_EXCLUSIVE:
3244 			break;
3245 		case LK_EXCLOTHER:
3246 		case 0:
3247 			want_unlock = true;
3248 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3249 			VI_LOCK(vp);
3250 			break;
3251 		default:
3252 			/*
3253 			 * The lock has at least one sharer, but we have no way
3254 			 * to conclude whether this is us. Play it safe and
3255 			 * defer processing.
3256 			 */
3257 			error = EAGAIN;
3258 			break;
3259 		}
3260 		break;
3261 	case VPUT:
3262 		want_unlock = true;
3263 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3264 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3265 			    LK_NOWAIT);
3266 			VI_LOCK(vp);
3267 		}
3268 		break;
3269 	case VUNREF:
3270 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3271 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3272 			VI_LOCK(vp);
3273 		}
3274 		break;
3275 	}
3276 	if (error == 0) {
3277 		if (func == VUNREF) {
3278 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3279 			    ("recursive vunref"));
3280 			vp->v_vflag |= VV_UNREF;
3281 		}
3282 		for (;;) {
3283 			error = vinactive(vp);
3284 			if (want_unlock)
3285 				VOP_UNLOCK(vp);
3286 			if (error != ERELOOKUP || !want_unlock)
3287 				break;
3288 			VOP_LOCK(vp, LK_EXCLUSIVE);
3289 		}
3290 		if (func == VUNREF)
3291 			vp->v_vflag &= ~VV_UNREF;
3292 		vdropl(vp);
3293 	} else {
3294 		vdefer_inactive(vp);
3295 	}
3296 	return;
3297 out:
3298 	if (func == VPUT)
3299 		VOP_UNLOCK(vp);
3300 	vdropl(vp);
3301 }
3302 
3303 /*
3304  * Decrement ->v_usecount for a vnode.
3305  *
3306  * Releasing the last use count requires additional processing, see vput_final
3307  * above for details.
3308  *
3309  * Comment above each variant denotes lock state on entry and exit.
3310  */
3311 
3312 /*
3313  * in: any
3314  * out: same as passed in
3315  */
3316 void
3317 vrele(struct vnode *vp)
3318 {
3319 
3320 	ASSERT_VI_UNLOCKED(vp, __func__);
3321 	if (!refcount_release(&vp->v_usecount))
3322 		return;
3323 	vput_final(vp, VRELE);
3324 }
3325 
3326 /*
3327  * in: locked
3328  * out: unlocked
3329  */
3330 void
3331 vput(struct vnode *vp)
3332 {
3333 
3334 	ASSERT_VOP_LOCKED(vp, __func__);
3335 	ASSERT_VI_UNLOCKED(vp, __func__);
3336 	if (!refcount_release(&vp->v_usecount)) {
3337 		VOP_UNLOCK(vp);
3338 		return;
3339 	}
3340 	vput_final(vp, VPUT);
3341 }
3342 
3343 /*
3344  * in: locked
3345  * out: locked
3346  */
3347 void
3348 vunref(struct vnode *vp)
3349 {
3350 
3351 	ASSERT_VOP_LOCKED(vp, __func__);
3352 	ASSERT_VI_UNLOCKED(vp, __func__);
3353 	if (!refcount_release(&vp->v_usecount))
3354 		return;
3355 	vput_final(vp, VUNREF);
3356 }
3357 
3358 void
3359 vhold(struct vnode *vp)
3360 {
3361 	int old;
3362 
3363 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3364 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3365 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3366 	    ("%s: wrong hold count %d", __func__, old));
3367 	if (old == 0)
3368 		vfs_freevnodes_dec();
3369 }
3370 
3371 void
3372 vholdnz(struct vnode *vp)
3373 {
3374 
3375 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3376 #ifdef INVARIANTS
3377 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3378 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3379 	    ("%s: wrong hold count %d", __func__, old));
3380 #else
3381 	atomic_add_int(&vp->v_holdcnt, 1);
3382 #endif
3383 }
3384 
3385 /*
3386  * Grab a hold count unless the vnode is freed.
3387  *
3388  * Only use this routine if vfs smr is the only protection you have against
3389  * freeing the vnode.
3390  *
3391  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3392  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3393  * the thread which managed to set the flag.
3394  *
3395  * It may be tempting to replace the loop with:
3396  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3397  * if (count & VHOLD_NO_SMR) {
3398  *     backpedal and error out;
3399  * }
3400  *
3401  * However, while this is more performant, it hinders debugging by eliminating
3402  * the previously mentioned invariant.
3403  */
3404 bool
3405 vhold_smr(struct vnode *vp)
3406 {
3407 	int count;
3408 
3409 	VFS_SMR_ASSERT_ENTERED();
3410 
3411 	count = atomic_load_int(&vp->v_holdcnt);
3412 	for (;;) {
3413 		if (count & VHOLD_NO_SMR) {
3414 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3415 			    ("non-zero hold count with flags %d\n", count));
3416 			return (false);
3417 		}
3418 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3419 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3420 			if (count == 0)
3421 				vfs_freevnodes_dec();
3422 			return (true);
3423 		}
3424 	}
3425 }
3426 
3427 /*
3428  * Hold a free vnode for recycling.
3429  *
3430  * Note: vnode_init references this comment.
3431  *
3432  * Attempts to recycle only need the global vnode list lock and have no use for
3433  * SMR.
3434  *
3435  * However, vnodes get inserted into the global list before they get fully
3436  * initialized and stay there until UMA decides to free the memory. This in
3437  * particular means the target can be found before it becomes usable and after
3438  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3439  * VHOLD_NO_SMR.
3440  *
3441  * Note: the vnode may gain more references after we transition the count 0->1.
3442  */
3443 static bool
3444 vhold_recycle_free(struct vnode *vp)
3445 {
3446 	int count;
3447 
3448 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3449 
3450 	count = atomic_load_int(&vp->v_holdcnt);
3451 	for (;;) {
3452 		if (count & VHOLD_NO_SMR) {
3453 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3454 			    ("non-zero hold count with flags %d\n", count));
3455 			return (false);
3456 		}
3457 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3458 		if (count > 0) {
3459 			return (false);
3460 		}
3461 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3462 			vfs_freevnodes_dec();
3463 			return (true);
3464 		}
3465 	}
3466 }
3467 
3468 static void __noinline
3469 vdbatch_process(struct vdbatch *vd)
3470 {
3471 	struct vnode *vp;
3472 	int i;
3473 
3474 	mtx_assert(&vd->lock, MA_OWNED);
3475 	MPASS(curthread->td_pinned > 0);
3476 	MPASS(vd->index == VDBATCH_SIZE);
3477 
3478 	mtx_lock(&vnode_list_mtx);
3479 	critical_enter();
3480 	freevnodes += vd->freevnodes;
3481 	for (i = 0; i < VDBATCH_SIZE; i++) {
3482 		vp = vd->tab[i];
3483 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3484 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3485 		MPASS(vp->v_dbatchcpu != NOCPU);
3486 		vp->v_dbatchcpu = NOCPU;
3487 	}
3488 	mtx_unlock(&vnode_list_mtx);
3489 	vd->freevnodes = 0;
3490 	bzero(vd->tab, sizeof(vd->tab));
3491 	vd->index = 0;
3492 	critical_exit();
3493 }
3494 
3495 static void
3496 vdbatch_enqueue(struct vnode *vp)
3497 {
3498 	struct vdbatch *vd;
3499 
3500 	ASSERT_VI_LOCKED(vp, __func__);
3501 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3502 	    ("%s: deferring requeue of a doomed vnode", __func__));
3503 
3504 	if (vp->v_dbatchcpu != NOCPU) {
3505 		VI_UNLOCK(vp);
3506 		return;
3507 	}
3508 
3509 	sched_pin();
3510 	vd = DPCPU_PTR(vd);
3511 	mtx_lock(&vd->lock);
3512 	MPASS(vd->index < VDBATCH_SIZE);
3513 	MPASS(vd->tab[vd->index] == NULL);
3514 	/*
3515 	 * A hack: we depend on being pinned so that we know what to put in
3516 	 * ->v_dbatchcpu.
3517 	 */
3518 	vp->v_dbatchcpu = curcpu;
3519 	vd->tab[vd->index] = vp;
3520 	vd->index++;
3521 	VI_UNLOCK(vp);
3522 	if (vd->index == VDBATCH_SIZE)
3523 		vdbatch_process(vd);
3524 	mtx_unlock(&vd->lock);
3525 	sched_unpin();
3526 }
3527 
3528 /*
3529  * This routine must only be called for vnodes which are about to be
3530  * deallocated. Supporting dequeue for arbitrary vndoes would require
3531  * validating that the locked batch matches.
3532  */
3533 static void
3534 vdbatch_dequeue(struct vnode *vp)
3535 {
3536 	struct vdbatch *vd;
3537 	int i;
3538 	short cpu;
3539 
3540 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3541 	    ("%s: called for a used vnode\n", __func__));
3542 
3543 	cpu = vp->v_dbatchcpu;
3544 	if (cpu == NOCPU)
3545 		return;
3546 
3547 	vd = DPCPU_ID_PTR(cpu, vd);
3548 	mtx_lock(&vd->lock);
3549 	for (i = 0; i < vd->index; i++) {
3550 		if (vd->tab[i] != vp)
3551 			continue;
3552 		vp->v_dbatchcpu = NOCPU;
3553 		vd->index--;
3554 		vd->tab[i] = vd->tab[vd->index];
3555 		vd->tab[vd->index] = NULL;
3556 		break;
3557 	}
3558 	mtx_unlock(&vd->lock);
3559 	/*
3560 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3561 	 */
3562 	MPASS(vp->v_dbatchcpu == NOCPU);
3563 }
3564 
3565 /*
3566  * Drop the hold count of the vnode.  If this is the last reference to
3567  * the vnode we place it on the free list unless it has been vgone'd
3568  * (marked VIRF_DOOMED) in which case we will free it.
3569  *
3570  * Because the vnode vm object keeps a hold reference on the vnode if
3571  * there is at least one resident non-cached page, the vnode cannot
3572  * leave the active list without the page cleanup done.
3573  */
3574 static void __noinline
3575 vdropl_final(struct vnode *vp)
3576 {
3577 
3578 	ASSERT_VI_LOCKED(vp, __func__);
3579 	VNPASS(VN_IS_DOOMED(vp), vp);
3580 	/*
3581 	 * Set the VHOLD_NO_SMR flag.
3582 	 *
3583 	 * We may be racing against vhold_smr. If they win we can just pretend
3584 	 * we never got this far, they will vdrop later.
3585 	 */
3586 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3587 		vfs_freevnodes_inc();
3588 		VI_UNLOCK(vp);
3589 		/*
3590 		 * We lost the aforementioned race. Any subsequent access is
3591 		 * invalid as they might have managed to vdropl on their own.
3592 		 */
3593 		return;
3594 	}
3595 	/*
3596 	 * Don't bump freevnodes as this one is going away.
3597 	 */
3598 	freevnode(vp);
3599 }
3600 
3601 void
3602 vdrop(struct vnode *vp)
3603 {
3604 
3605 	ASSERT_VI_UNLOCKED(vp, __func__);
3606 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3607 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3608 		return;
3609 	VI_LOCK(vp);
3610 	vdropl(vp);
3611 }
3612 
3613 void
3614 vdropl(struct vnode *vp)
3615 {
3616 
3617 	ASSERT_VI_LOCKED(vp, __func__);
3618 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3619 	if (!refcount_release(&vp->v_holdcnt)) {
3620 		VI_UNLOCK(vp);
3621 		return;
3622 	}
3623 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3624 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3625 	if (VN_IS_DOOMED(vp)) {
3626 		vdropl_final(vp);
3627 		return;
3628 	}
3629 
3630 	vfs_freevnodes_inc();
3631 	if (vp->v_mflag & VMP_LAZYLIST) {
3632 		vunlazy(vp);
3633 	}
3634 	/*
3635 	 * Also unlocks the interlock. We can't assert on it as we
3636 	 * released our hold and by now the vnode might have been
3637 	 * freed.
3638 	 */
3639 	vdbatch_enqueue(vp);
3640 }
3641 
3642 /*
3643  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3644  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3645  */
3646 static int
3647 vinactivef(struct vnode *vp)
3648 {
3649 	struct vm_object *obj;
3650 	int error;
3651 
3652 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3653 	ASSERT_VI_LOCKED(vp, "vinactive");
3654 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3655 	    ("vinactive: recursed on VI_DOINGINACT"));
3656 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3657 	vp->v_iflag |= VI_DOINGINACT;
3658 	vp->v_iflag &= ~VI_OWEINACT;
3659 	VI_UNLOCK(vp);
3660 	/*
3661 	 * Before moving off the active list, we must be sure that any
3662 	 * modified pages are converted into the vnode's dirty
3663 	 * buffers, since these will no longer be checked once the
3664 	 * vnode is on the inactive list.
3665 	 *
3666 	 * The write-out of the dirty pages is asynchronous.  At the
3667 	 * point that VOP_INACTIVE() is called, there could still be
3668 	 * pending I/O and dirty pages in the object.
3669 	 */
3670 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3671 	    vm_object_mightbedirty(obj)) {
3672 		VM_OBJECT_WLOCK(obj);
3673 		vm_object_page_clean(obj, 0, 0, 0);
3674 		VM_OBJECT_WUNLOCK(obj);
3675 	}
3676 	error = VOP_INACTIVE(vp);
3677 	VI_LOCK(vp);
3678 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3679 	    ("vinactive: lost VI_DOINGINACT"));
3680 	vp->v_iflag &= ~VI_DOINGINACT;
3681 	return (error);
3682 }
3683 
3684 int
3685 vinactive(struct vnode *vp)
3686 {
3687 
3688 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3689 	ASSERT_VI_LOCKED(vp, "vinactive");
3690 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3691 
3692 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3693 		return (0);
3694 	if (vp->v_iflag & VI_DOINGINACT)
3695 		return (0);
3696 	if (vp->v_usecount > 0) {
3697 		vp->v_iflag &= ~VI_OWEINACT;
3698 		return (0);
3699 	}
3700 	return (vinactivef(vp));
3701 }
3702 
3703 /*
3704  * Remove any vnodes in the vnode table belonging to mount point mp.
3705  *
3706  * If FORCECLOSE is not specified, there should not be any active ones,
3707  * return error if any are found (nb: this is a user error, not a
3708  * system error). If FORCECLOSE is specified, detach any active vnodes
3709  * that are found.
3710  *
3711  * If WRITECLOSE is set, only flush out regular file vnodes open for
3712  * writing.
3713  *
3714  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3715  *
3716  * `rootrefs' specifies the base reference count for the root vnode
3717  * of this filesystem. The root vnode is considered busy if its
3718  * v_usecount exceeds this value. On a successful return, vflush(, td)
3719  * will call vrele() on the root vnode exactly rootrefs times.
3720  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3721  * be zero.
3722  */
3723 #ifdef DIAGNOSTIC
3724 static int busyprt = 0;		/* print out busy vnodes */
3725 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3726 #endif
3727 
3728 int
3729 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3730 {
3731 	struct vnode *vp, *mvp, *rootvp = NULL;
3732 	struct vattr vattr;
3733 	int busy = 0, error;
3734 
3735 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3736 	    rootrefs, flags);
3737 	if (rootrefs > 0) {
3738 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3739 		    ("vflush: bad args"));
3740 		/*
3741 		 * Get the filesystem root vnode. We can vput() it
3742 		 * immediately, since with rootrefs > 0, it won't go away.
3743 		 */
3744 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3745 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3746 			    __func__, error);
3747 			return (error);
3748 		}
3749 		vput(rootvp);
3750 	}
3751 loop:
3752 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3753 		vholdl(vp);
3754 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3755 		if (error) {
3756 			vdrop(vp);
3757 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3758 			goto loop;
3759 		}
3760 		/*
3761 		 * Skip over a vnodes marked VV_SYSTEM.
3762 		 */
3763 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3764 			VOP_UNLOCK(vp);
3765 			vdrop(vp);
3766 			continue;
3767 		}
3768 		/*
3769 		 * If WRITECLOSE is set, flush out unlinked but still open
3770 		 * files (even if open only for reading) and regular file
3771 		 * vnodes open for writing.
3772 		 */
3773 		if (flags & WRITECLOSE) {
3774 			if (vp->v_object != NULL) {
3775 				VM_OBJECT_WLOCK(vp->v_object);
3776 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3777 				VM_OBJECT_WUNLOCK(vp->v_object);
3778 			}
3779 			do {
3780 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3781 			} while (error == ERELOOKUP);
3782 			if (error != 0) {
3783 				VOP_UNLOCK(vp);
3784 				vdrop(vp);
3785 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3786 				return (error);
3787 			}
3788 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3789 			VI_LOCK(vp);
3790 
3791 			if ((vp->v_type == VNON ||
3792 			    (error == 0 && vattr.va_nlink > 0)) &&
3793 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3794 				VOP_UNLOCK(vp);
3795 				vdropl(vp);
3796 				continue;
3797 			}
3798 		} else
3799 			VI_LOCK(vp);
3800 		/*
3801 		 * With v_usecount == 0, all we need to do is clear out the
3802 		 * vnode data structures and we are done.
3803 		 *
3804 		 * If FORCECLOSE is set, forcibly close the vnode.
3805 		 */
3806 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3807 			vgonel(vp);
3808 		} else {
3809 			busy++;
3810 #ifdef DIAGNOSTIC
3811 			if (busyprt)
3812 				vn_printf(vp, "vflush: busy vnode ");
3813 #endif
3814 		}
3815 		VOP_UNLOCK(vp);
3816 		vdropl(vp);
3817 	}
3818 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3819 		/*
3820 		 * If just the root vnode is busy, and if its refcount
3821 		 * is equal to `rootrefs', then go ahead and kill it.
3822 		 */
3823 		VI_LOCK(rootvp);
3824 		KASSERT(busy > 0, ("vflush: not busy"));
3825 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3826 		    ("vflush: usecount %d < rootrefs %d",
3827 		     rootvp->v_usecount, rootrefs));
3828 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3829 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3830 			vgone(rootvp);
3831 			VOP_UNLOCK(rootvp);
3832 			busy = 0;
3833 		} else
3834 			VI_UNLOCK(rootvp);
3835 	}
3836 	if (busy) {
3837 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3838 		    busy);
3839 		return (EBUSY);
3840 	}
3841 	for (; rootrefs > 0; rootrefs--)
3842 		vrele(rootvp);
3843 	return (0);
3844 }
3845 
3846 /*
3847  * Recycle an unused vnode to the front of the free list.
3848  */
3849 int
3850 vrecycle(struct vnode *vp)
3851 {
3852 	int recycled;
3853 
3854 	VI_LOCK(vp);
3855 	recycled = vrecyclel(vp);
3856 	VI_UNLOCK(vp);
3857 	return (recycled);
3858 }
3859 
3860 /*
3861  * vrecycle, with the vp interlock held.
3862  */
3863 int
3864 vrecyclel(struct vnode *vp)
3865 {
3866 	int recycled;
3867 
3868 	ASSERT_VOP_ELOCKED(vp, __func__);
3869 	ASSERT_VI_LOCKED(vp, __func__);
3870 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3871 	recycled = 0;
3872 	if (vp->v_usecount == 0) {
3873 		recycled = 1;
3874 		vgonel(vp);
3875 	}
3876 	return (recycled);
3877 }
3878 
3879 /*
3880  * Eliminate all activity associated with a vnode
3881  * in preparation for reuse.
3882  */
3883 void
3884 vgone(struct vnode *vp)
3885 {
3886 	VI_LOCK(vp);
3887 	vgonel(vp);
3888 	VI_UNLOCK(vp);
3889 }
3890 
3891 static void
3892 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3893     struct vnode *lowervp __unused)
3894 {
3895 }
3896 
3897 /*
3898  * Notify upper mounts about reclaimed or unlinked vnode.
3899  */
3900 void
3901 vfs_notify_upper(struct vnode *vp, int event)
3902 {
3903 	static struct vfsops vgonel_vfsops = {
3904 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3905 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3906 	};
3907 	struct mount *mp, *ump, *mmp;
3908 
3909 	mp = vp->v_mount;
3910 	if (mp == NULL)
3911 		return;
3912 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3913 		return;
3914 
3915 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3916 	mmp->mnt_op = &vgonel_vfsops;
3917 	mmp->mnt_kern_flag |= MNTK_MARKER;
3918 	MNT_ILOCK(mp);
3919 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3920 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3921 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3922 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3923 			continue;
3924 		}
3925 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3926 		MNT_IUNLOCK(mp);
3927 		switch (event) {
3928 		case VFS_NOTIFY_UPPER_RECLAIM:
3929 			VFS_RECLAIM_LOWERVP(ump, vp);
3930 			break;
3931 		case VFS_NOTIFY_UPPER_UNLINK:
3932 			VFS_UNLINK_LOWERVP(ump, vp);
3933 			break;
3934 		default:
3935 			KASSERT(0, ("invalid event %d", event));
3936 			break;
3937 		}
3938 		MNT_ILOCK(mp);
3939 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3940 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3941 	}
3942 	free(mmp, M_TEMP);
3943 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3944 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3945 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3946 		wakeup(&mp->mnt_uppers);
3947 	}
3948 	MNT_IUNLOCK(mp);
3949 }
3950 
3951 /*
3952  * vgone, with the vp interlock held.
3953  */
3954 static void
3955 vgonel(struct vnode *vp)
3956 {
3957 	struct thread *td;
3958 	struct mount *mp;
3959 	vm_object_t object;
3960 	bool active, doinginact, oweinact;
3961 
3962 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3963 	ASSERT_VI_LOCKED(vp, "vgonel");
3964 	VNASSERT(vp->v_holdcnt, vp,
3965 	    ("vgonel: vp %p has no reference.", vp));
3966 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3967 	td = curthread;
3968 
3969 	/*
3970 	 * Don't vgonel if we're already doomed.
3971 	 */
3972 	if (VN_IS_DOOMED(vp))
3973 		return;
3974 	/*
3975 	 * Paired with freevnode.
3976 	 */
3977 	vn_seqc_write_begin_locked(vp);
3978 	vunlazy_gone(vp);
3979 	vn_irflag_set_locked(vp, VIRF_DOOMED);
3980 
3981 	/*
3982 	 * Check to see if the vnode is in use.  If so, we have to
3983 	 * call VOP_CLOSE() and VOP_INACTIVE().
3984 	 *
3985 	 * It could be that VOP_INACTIVE() requested reclamation, in
3986 	 * which case we should avoid recursion, so check
3987 	 * VI_DOINGINACT.  This is not precise but good enough.
3988 	 */
3989 	active = vp->v_usecount > 0;
3990 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3991 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3992 
3993 	/*
3994 	 * If we need to do inactive VI_OWEINACT will be set.
3995 	 */
3996 	if (vp->v_iflag & VI_DEFINACT) {
3997 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3998 		vp->v_iflag &= ~VI_DEFINACT;
3999 		vdropl(vp);
4000 	} else {
4001 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4002 		VI_UNLOCK(vp);
4003 	}
4004 	cache_purge_vgone(vp);
4005 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4006 
4007 	/*
4008 	 * If purging an active vnode, it must be closed and
4009 	 * deactivated before being reclaimed.
4010 	 */
4011 	if (active)
4012 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4013 	if (!doinginact) {
4014 		do {
4015 			if (oweinact || active) {
4016 				VI_LOCK(vp);
4017 				vinactivef(vp);
4018 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4019 				VI_UNLOCK(vp);
4020 			}
4021 		} while (oweinact);
4022 	}
4023 	if (vp->v_type == VSOCK)
4024 		vfs_unp_reclaim(vp);
4025 
4026 	/*
4027 	 * Clean out any buffers associated with the vnode.
4028 	 * If the flush fails, just toss the buffers.
4029 	 */
4030 	mp = NULL;
4031 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4032 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4033 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4034 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4035 			;
4036 	}
4037 
4038 	BO_LOCK(&vp->v_bufobj);
4039 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4040 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4041 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4042 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4043 	    ("vp %p bufobj not invalidated", vp));
4044 
4045 	/*
4046 	 * For VMIO bufobj, BO_DEAD is set later, or in
4047 	 * vm_object_terminate() after the object's page queue is
4048 	 * flushed.
4049 	 */
4050 	object = vp->v_bufobj.bo_object;
4051 	if (object == NULL)
4052 		vp->v_bufobj.bo_flag |= BO_DEAD;
4053 	BO_UNLOCK(&vp->v_bufobj);
4054 
4055 	/*
4056 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4057 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4058 	 * should not touch the object borrowed from the lower vnode
4059 	 * (the handle check).
4060 	 */
4061 	if (object != NULL && object->type == OBJT_VNODE &&
4062 	    object->handle == vp)
4063 		vnode_destroy_vobject(vp);
4064 
4065 	/*
4066 	 * Reclaim the vnode.
4067 	 */
4068 	if (VOP_RECLAIM(vp))
4069 		panic("vgone: cannot reclaim");
4070 	if (mp != NULL)
4071 		vn_finished_secondary_write(mp);
4072 	VNASSERT(vp->v_object == NULL, vp,
4073 	    ("vop_reclaim left v_object vp=%p", vp));
4074 	/*
4075 	 * Clear the advisory locks and wake up waiting threads.
4076 	 */
4077 	(void)VOP_ADVLOCKPURGE(vp);
4078 	vp->v_lockf = NULL;
4079 	/*
4080 	 * Delete from old mount point vnode list.
4081 	 */
4082 	delmntque(vp);
4083 	/*
4084 	 * Done with purge, reset to the standard lock and invalidate
4085 	 * the vnode.
4086 	 */
4087 	VI_LOCK(vp);
4088 	vp->v_vnlock = &vp->v_lock;
4089 	vp->v_op = &dead_vnodeops;
4090 	vp->v_type = VBAD;
4091 }
4092 
4093 /*
4094  * Print out a description of a vnode.
4095  */
4096 static const char * const typename[] =
4097 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4098  "VMARKER"};
4099 
4100 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4101     "new hold count flag not added to vn_printf");
4102 
4103 void
4104 vn_printf(struct vnode *vp, const char *fmt, ...)
4105 {
4106 	va_list ap;
4107 	char buf[256], buf2[16];
4108 	u_long flags;
4109 	u_int holdcnt;
4110 	short irflag;
4111 
4112 	va_start(ap, fmt);
4113 	vprintf(fmt, ap);
4114 	va_end(ap);
4115 	printf("%p: ", (void *)vp);
4116 	printf("type %s\n", typename[vp->v_type]);
4117 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4118 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4119 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4120 	    vp->v_seqc_users);
4121 	switch (vp->v_type) {
4122 	case VDIR:
4123 		printf(" mountedhere %p\n", vp->v_mountedhere);
4124 		break;
4125 	case VCHR:
4126 		printf(" rdev %p\n", vp->v_rdev);
4127 		break;
4128 	case VSOCK:
4129 		printf(" socket %p\n", vp->v_unpcb);
4130 		break;
4131 	case VFIFO:
4132 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4133 		break;
4134 	default:
4135 		printf("\n");
4136 		break;
4137 	}
4138 	buf[0] = '\0';
4139 	buf[1] = '\0';
4140 	if (holdcnt & VHOLD_NO_SMR)
4141 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4142 	printf("    hold count flags (%s)\n", buf + 1);
4143 
4144 	buf[0] = '\0';
4145 	buf[1] = '\0';
4146 	irflag = vn_irflag_read(vp);
4147 	if (irflag & VIRF_DOOMED)
4148 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4149 	if (irflag & VIRF_PGREAD)
4150 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4151 	if (irflag & VIRF_MOUNTPOINT)
4152 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4153 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT);
4154 	if (flags != 0) {
4155 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4156 		strlcat(buf, buf2, sizeof(buf));
4157 	}
4158 	if (vp->v_vflag & VV_ROOT)
4159 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4160 	if (vp->v_vflag & VV_ISTTY)
4161 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4162 	if (vp->v_vflag & VV_NOSYNC)
4163 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4164 	if (vp->v_vflag & VV_ETERNALDEV)
4165 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4166 	if (vp->v_vflag & VV_CACHEDLABEL)
4167 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4168 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4169 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4170 	if (vp->v_vflag & VV_COPYONWRITE)
4171 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4172 	if (vp->v_vflag & VV_SYSTEM)
4173 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4174 	if (vp->v_vflag & VV_PROCDEP)
4175 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4176 	if (vp->v_vflag & VV_NOKNOTE)
4177 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4178 	if (vp->v_vflag & VV_DELETED)
4179 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4180 	if (vp->v_vflag & VV_MD)
4181 		strlcat(buf, "|VV_MD", sizeof(buf));
4182 	if (vp->v_vflag & VV_FORCEINSMQ)
4183 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4184 	if (vp->v_vflag & VV_READLINK)
4185 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4186 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4187 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4188 	    VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ |
4189 	    VV_READLINK);
4190 	if (flags != 0) {
4191 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4192 		strlcat(buf, buf2, sizeof(buf));
4193 	}
4194 	if (vp->v_iflag & VI_TEXT_REF)
4195 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4196 	if (vp->v_iflag & VI_MOUNT)
4197 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4198 	if (vp->v_iflag & VI_DOINGINACT)
4199 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4200 	if (vp->v_iflag & VI_OWEINACT)
4201 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4202 	if (vp->v_iflag & VI_DEFINACT)
4203 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4204 	if (vp->v_iflag & VI_FOPENING)
4205 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4206 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4207 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4208 	if (flags != 0) {
4209 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4210 		strlcat(buf, buf2, sizeof(buf));
4211 	}
4212 	if (vp->v_mflag & VMP_LAZYLIST)
4213 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4214 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4215 	if (flags != 0) {
4216 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4217 		strlcat(buf, buf2, sizeof(buf));
4218 	}
4219 	printf("    flags (%s)", buf + 1);
4220 	if (mtx_owned(VI_MTX(vp)))
4221 		printf(" VI_LOCKed");
4222 	printf("\n");
4223 	if (vp->v_object != NULL)
4224 		printf("    v_object %p ref %d pages %d "
4225 		    "cleanbuf %d dirtybuf %d\n",
4226 		    vp->v_object, vp->v_object->ref_count,
4227 		    vp->v_object->resident_page_count,
4228 		    vp->v_bufobj.bo_clean.bv_cnt,
4229 		    vp->v_bufobj.bo_dirty.bv_cnt);
4230 	printf("    ");
4231 	lockmgr_printinfo(vp->v_vnlock);
4232 	if (vp->v_data != NULL)
4233 		VOP_PRINT(vp);
4234 }
4235 
4236 #ifdef DDB
4237 /*
4238  * List all of the locked vnodes in the system.
4239  * Called when debugging the kernel.
4240  */
4241 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4242 {
4243 	struct mount *mp;
4244 	struct vnode *vp;
4245 
4246 	/*
4247 	 * Note: because this is DDB, we can't obey the locking semantics
4248 	 * for these structures, which means we could catch an inconsistent
4249 	 * state and dereference a nasty pointer.  Not much to be done
4250 	 * about that.
4251 	 */
4252 	db_printf("Locked vnodes\n");
4253 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4254 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4255 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4256 				vn_printf(vp, "vnode ");
4257 		}
4258 	}
4259 }
4260 
4261 /*
4262  * Show details about the given vnode.
4263  */
4264 DB_SHOW_COMMAND(vnode, db_show_vnode)
4265 {
4266 	struct vnode *vp;
4267 
4268 	if (!have_addr)
4269 		return;
4270 	vp = (struct vnode *)addr;
4271 	vn_printf(vp, "vnode ");
4272 }
4273 
4274 /*
4275  * Show details about the given mount point.
4276  */
4277 DB_SHOW_COMMAND(mount, db_show_mount)
4278 {
4279 	struct mount *mp;
4280 	struct vfsopt *opt;
4281 	struct statfs *sp;
4282 	struct vnode *vp;
4283 	char buf[512];
4284 	uint64_t mflags;
4285 	u_int flags;
4286 
4287 	if (!have_addr) {
4288 		/* No address given, print short info about all mount points. */
4289 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4290 			db_printf("%p %s on %s (%s)\n", mp,
4291 			    mp->mnt_stat.f_mntfromname,
4292 			    mp->mnt_stat.f_mntonname,
4293 			    mp->mnt_stat.f_fstypename);
4294 			if (db_pager_quit)
4295 				break;
4296 		}
4297 		db_printf("\nMore info: show mount <addr>\n");
4298 		return;
4299 	}
4300 
4301 	mp = (struct mount *)addr;
4302 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4303 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4304 
4305 	buf[0] = '\0';
4306 	mflags = mp->mnt_flag;
4307 #define	MNT_FLAG(flag)	do {						\
4308 	if (mflags & (flag)) {						\
4309 		if (buf[0] != '\0')					\
4310 			strlcat(buf, ", ", sizeof(buf));		\
4311 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4312 		mflags &= ~(flag);					\
4313 	}								\
4314 } while (0)
4315 	MNT_FLAG(MNT_RDONLY);
4316 	MNT_FLAG(MNT_SYNCHRONOUS);
4317 	MNT_FLAG(MNT_NOEXEC);
4318 	MNT_FLAG(MNT_NOSUID);
4319 	MNT_FLAG(MNT_NFS4ACLS);
4320 	MNT_FLAG(MNT_UNION);
4321 	MNT_FLAG(MNT_ASYNC);
4322 	MNT_FLAG(MNT_SUIDDIR);
4323 	MNT_FLAG(MNT_SOFTDEP);
4324 	MNT_FLAG(MNT_NOSYMFOLLOW);
4325 	MNT_FLAG(MNT_GJOURNAL);
4326 	MNT_FLAG(MNT_MULTILABEL);
4327 	MNT_FLAG(MNT_ACLS);
4328 	MNT_FLAG(MNT_NOATIME);
4329 	MNT_FLAG(MNT_NOCLUSTERR);
4330 	MNT_FLAG(MNT_NOCLUSTERW);
4331 	MNT_FLAG(MNT_SUJ);
4332 	MNT_FLAG(MNT_EXRDONLY);
4333 	MNT_FLAG(MNT_EXPORTED);
4334 	MNT_FLAG(MNT_DEFEXPORTED);
4335 	MNT_FLAG(MNT_EXPORTANON);
4336 	MNT_FLAG(MNT_EXKERB);
4337 	MNT_FLAG(MNT_EXPUBLIC);
4338 	MNT_FLAG(MNT_LOCAL);
4339 	MNT_FLAG(MNT_QUOTA);
4340 	MNT_FLAG(MNT_ROOTFS);
4341 	MNT_FLAG(MNT_USER);
4342 	MNT_FLAG(MNT_IGNORE);
4343 	MNT_FLAG(MNT_UPDATE);
4344 	MNT_FLAG(MNT_DELEXPORT);
4345 	MNT_FLAG(MNT_RELOAD);
4346 	MNT_FLAG(MNT_FORCE);
4347 	MNT_FLAG(MNT_SNAPSHOT);
4348 	MNT_FLAG(MNT_BYFSID);
4349 #undef MNT_FLAG
4350 	if (mflags != 0) {
4351 		if (buf[0] != '\0')
4352 			strlcat(buf, ", ", sizeof(buf));
4353 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4354 		    "0x%016jx", mflags);
4355 	}
4356 	db_printf("    mnt_flag = %s\n", buf);
4357 
4358 	buf[0] = '\0';
4359 	flags = mp->mnt_kern_flag;
4360 #define	MNT_KERN_FLAG(flag)	do {					\
4361 	if (flags & (flag)) {						\
4362 		if (buf[0] != '\0')					\
4363 			strlcat(buf, ", ", sizeof(buf));		\
4364 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4365 		flags &= ~(flag);					\
4366 	}								\
4367 } while (0)
4368 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4369 	MNT_KERN_FLAG(MNTK_ASYNC);
4370 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4371 	MNT_KERN_FLAG(MNTK_DRAINING);
4372 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4373 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4374 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4375 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4376 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4377 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4378 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4379 	MNT_KERN_FLAG(MNTK_MARKER);
4380 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4381 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4382 	MNT_KERN_FLAG(MNTK_NOASYNC);
4383 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4384 	MNT_KERN_FLAG(MNTK_MWAIT);
4385 	MNT_KERN_FLAG(MNTK_SUSPEND);
4386 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4387 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4388 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4389 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4390 #undef MNT_KERN_FLAG
4391 	if (flags != 0) {
4392 		if (buf[0] != '\0')
4393 			strlcat(buf, ", ", sizeof(buf));
4394 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4395 		    "0x%08x", flags);
4396 	}
4397 	db_printf("    mnt_kern_flag = %s\n", buf);
4398 
4399 	db_printf("    mnt_opt = ");
4400 	opt = TAILQ_FIRST(mp->mnt_opt);
4401 	if (opt != NULL) {
4402 		db_printf("%s", opt->name);
4403 		opt = TAILQ_NEXT(opt, link);
4404 		while (opt != NULL) {
4405 			db_printf(", %s", opt->name);
4406 			opt = TAILQ_NEXT(opt, link);
4407 		}
4408 	}
4409 	db_printf("\n");
4410 
4411 	sp = &mp->mnt_stat;
4412 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4413 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4414 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4415 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4416 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4417 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4418 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4419 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4420 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4421 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4422 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4423 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4424 
4425 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4426 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4427 	if (jailed(mp->mnt_cred))
4428 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4429 	db_printf(" }\n");
4430 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4431 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4432 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4433 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4434 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4435 	    mp->mnt_lazyvnodelistsize);
4436 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4437 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4438 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4439 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4440 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4441 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4442 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4443 	db_printf("    mnt_secondary_accwrites = %d\n",
4444 	    mp->mnt_secondary_accwrites);
4445 	db_printf("    mnt_gjprovider = %s\n",
4446 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4447 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4448 
4449 	db_printf("\n\nList of active vnodes\n");
4450 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4451 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4452 			vn_printf(vp, "vnode ");
4453 			if (db_pager_quit)
4454 				break;
4455 		}
4456 	}
4457 	db_printf("\n\nList of inactive vnodes\n");
4458 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4459 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4460 			vn_printf(vp, "vnode ");
4461 			if (db_pager_quit)
4462 				break;
4463 		}
4464 	}
4465 }
4466 #endif	/* DDB */
4467 
4468 /*
4469  * Fill in a struct xvfsconf based on a struct vfsconf.
4470  */
4471 static int
4472 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4473 {
4474 	struct xvfsconf xvfsp;
4475 
4476 	bzero(&xvfsp, sizeof(xvfsp));
4477 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4478 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4479 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4480 	xvfsp.vfc_flags = vfsp->vfc_flags;
4481 	/*
4482 	 * These are unused in userland, we keep them
4483 	 * to not break binary compatibility.
4484 	 */
4485 	xvfsp.vfc_vfsops = NULL;
4486 	xvfsp.vfc_next = NULL;
4487 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4488 }
4489 
4490 #ifdef COMPAT_FREEBSD32
4491 struct xvfsconf32 {
4492 	uint32_t	vfc_vfsops;
4493 	char		vfc_name[MFSNAMELEN];
4494 	int32_t		vfc_typenum;
4495 	int32_t		vfc_refcount;
4496 	int32_t		vfc_flags;
4497 	uint32_t	vfc_next;
4498 };
4499 
4500 static int
4501 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4502 {
4503 	struct xvfsconf32 xvfsp;
4504 
4505 	bzero(&xvfsp, sizeof(xvfsp));
4506 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4507 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4508 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4509 	xvfsp.vfc_flags = vfsp->vfc_flags;
4510 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4511 }
4512 #endif
4513 
4514 /*
4515  * Top level filesystem related information gathering.
4516  */
4517 static int
4518 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4519 {
4520 	struct vfsconf *vfsp;
4521 	int error;
4522 
4523 	error = 0;
4524 	vfsconf_slock();
4525 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4526 #ifdef COMPAT_FREEBSD32
4527 		if (req->flags & SCTL_MASK32)
4528 			error = vfsconf2x32(req, vfsp);
4529 		else
4530 #endif
4531 			error = vfsconf2x(req, vfsp);
4532 		if (error)
4533 			break;
4534 	}
4535 	vfsconf_sunlock();
4536 	return (error);
4537 }
4538 
4539 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4540     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4541     "S,xvfsconf", "List of all configured filesystems");
4542 
4543 #ifndef BURN_BRIDGES
4544 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4545 
4546 static int
4547 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4548 {
4549 	int *name = (int *)arg1 - 1;	/* XXX */
4550 	u_int namelen = arg2 + 1;	/* XXX */
4551 	struct vfsconf *vfsp;
4552 
4553 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4554 	    "please rebuild world\n");
4555 
4556 #if 1 || defined(COMPAT_PRELITE2)
4557 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4558 	if (namelen == 1)
4559 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4560 #endif
4561 
4562 	switch (name[1]) {
4563 	case VFS_MAXTYPENUM:
4564 		if (namelen != 2)
4565 			return (ENOTDIR);
4566 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4567 	case VFS_CONF:
4568 		if (namelen != 3)
4569 			return (ENOTDIR);	/* overloaded */
4570 		vfsconf_slock();
4571 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4572 			if (vfsp->vfc_typenum == name[2])
4573 				break;
4574 		}
4575 		vfsconf_sunlock();
4576 		if (vfsp == NULL)
4577 			return (EOPNOTSUPP);
4578 #ifdef COMPAT_FREEBSD32
4579 		if (req->flags & SCTL_MASK32)
4580 			return (vfsconf2x32(req, vfsp));
4581 		else
4582 #endif
4583 			return (vfsconf2x(req, vfsp));
4584 	}
4585 	return (EOPNOTSUPP);
4586 }
4587 
4588 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4589     CTLFLAG_MPSAFE, vfs_sysctl,
4590     "Generic filesystem");
4591 
4592 #if 1 || defined(COMPAT_PRELITE2)
4593 
4594 static int
4595 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4596 {
4597 	int error;
4598 	struct vfsconf *vfsp;
4599 	struct ovfsconf ovfs;
4600 
4601 	vfsconf_slock();
4602 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4603 		bzero(&ovfs, sizeof(ovfs));
4604 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4605 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4606 		ovfs.vfc_index = vfsp->vfc_typenum;
4607 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4608 		ovfs.vfc_flags = vfsp->vfc_flags;
4609 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4610 		if (error != 0) {
4611 			vfsconf_sunlock();
4612 			return (error);
4613 		}
4614 	}
4615 	vfsconf_sunlock();
4616 	return (0);
4617 }
4618 
4619 #endif /* 1 || COMPAT_PRELITE2 */
4620 #endif /* !BURN_BRIDGES */
4621 
4622 #define KINFO_VNODESLOP		10
4623 #ifdef notyet
4624 /*
4625  * Dump vnode list (via sysctl).
4626  */
4627 /* ARGSUSED */
4628 static int
4629 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4630 {
4631 	struct xvnode *xvn;
4632 	struct mount *mp;
4633 	struct vnode *vp;
4634 	int error, len, n;
4635 
4636 	/*
4637 	 * Stale numvnodes access is not fatal here.
4638 	 */
4639 	req->lock = 0;
4640 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4641 	if (!req->oldptr)
4642 		/* Make an estimate */
4643 		return (SYSCTL_OUT(req, 0, len));
4644 
4645 	error = sysctl_wire_old_buffer(req, 0);
4646 	if (error != 0)
4647 		return (error);
4648 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4649 	n = 0;
4650 	mtx_lock(&mountlist_mtx);
4651 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4652 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4653 			continue;
4654 		MNT_ILOCK(mp);
4655 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4656 			if (n == len)
4657 				break;
4658 			vref(vp);
4659 			xvn[n].xv_size = sizeof *xvn;
4660 			xvn[n].xv_vnode = vp;
4661 			xvn[n].xv_id = 0;	/* XXX compat */
4662 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4663 			XV_COPY(usecount);
4664 			XV_COPY(writecount);
4665 			XV_COPY(holdcnt);
4666 			XV_COPY(mount);
4667 			XV_COPY(numoutput);
4668 			XV_COPY(type);
4669 #undef XV_COPY
4670 			xvn[n].xv_flag = vp->v_vflag;
4671 
4672 			switch (vp->v_type) {
4673 			case VREG:
4674 			case VDIR:
4675 			case VLNK:
4676 				break;
4677 			case VBLK:
4678 			case VCHR:
4679 				if (vp->v_rdev == NULL) {
4680 					vrele(vp);
4681 					continue;
4682 				}
4683 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4684 				break;
4685 			case VSOCK:
4686 				xvn[n].xv_socket = vp->v_socket;
4687 				break;
4688 			case VFIFO:
4689 				xvn[n].xv_fifo = vp->v_fifoinfo;
4690 				break;
4691 			case VNON:
4692 			case VBAD:
4693 			default:
4694 				/* shouldn't happen? */
4695 				vrele(vp);
4696 				continue;
4697 			}
4698 			vrele(vp);
4699 			++n;
4700 		}
4701 		MNT_IUNLOCK(mp);
4702 		mtx_lock(&mountlist_mtx);
4703 		vfs_unbusy(mp);
4704 		if (n == len)
4705 			break;
4706 	}
4707 	mtx_unlock(&mountlist_mtx);
4708 
4709 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4710 	free(xvn, M_TEMP);
4711 	return (error);
4712 }
4713 
4714 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4715     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4716     "");
4717 #endif
4718 
4719 static void
4720 unmount_or_warn(struct mount *mp)
4721 {
4722 	int error;
4723 
4724 	error = dounmount(mp, MNT_FORCE, curthread);
4725 	if (error != 0) {
4726 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4727 		if (error == EBUSY)
4728 			printf("BUSY)\n");
4729 		else
4730 			printf("%d)\n", error);
4731 	}
4732 }
4733 
4734 /*
4735  * Unmount all filesystems. The list is traversed in reverse order
4736  * of mounting to avoid dependencies.
4737  */
4738 void
4739 vfs_unmountall(void)
4740 {
4741 	struct mount *mp, *tmp;
4742 
4743 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4744 
4745 	/*
4746 	 * Since this only runs when rebooting, it is not interlocked.
4747 	 */
4748 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4749 		vfs_ref(mp);
4750 
4751 		/*
4752 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4753 		 * unmount of the latter.
4754 		 */
4755 		if (mp == rootdevmp)
4756 			continue;
4757 
4758 		unmount_or_warn(mp);
4759 	}
4760 
4761 	if (rootdevmp != NULL)
4762 		unmount_or_warn(rootdevmp);
4763 }
4764 
4765 static void
4766 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4767 {
4768 
4769 	ASSERT_VI_LOCKED(vp, __func__);
4770 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4771 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4772 		vdropl(vp);
4773 		return;
4774 	}
4775 	if (vn_lock(vp, lkflags) == 0) {
4776 		VI_LOCK(vp);
4777 		vinactive(vp);
4778 		VOP_UNLOCK(vp);
4779 		vdropl(vp);
4780 		return;
4781 	}
4782 	vdefer_inactive_unlocked(vp);
4783 }
4784 
4785 static int
4786 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4787 {
4788 
4789 	return (vp->v_iflag & VI_DEFINACT);
4790 }
4791 
4792 static void __noinline
4793 vfs_periodic_inactive(struct mount *mp, int flags)
4794 {
4795 	struct vnode *vp, *mvp;
4796 	int lkflags;
4797 
4798 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4799 	if (flags != MNT_WAIT)
4800 		lkflags |= LK_NOWAIT;
4801 
4802 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4803 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4804 			VI_UNLOCK(vp);
4805 			continue;
4806 		}
4807 		vp->v_iflag &= ~VI_DEFINACT;
4808 		vfs_deferred_inactive(vp, lkflags);
4809 	}
4810 }
4811 
4812 static inline bool
4813 vfs_want_msync(struct vnode *vp)
4814 {
4815 	struct vm_object *obj;
4816 
4817 	/*
4818 	 * This test may be performed without any locks held.
4819 	 * We rely on vm_object's type stability.
4820 	 */
4821 	if (vp->v_vflag & VV_NOSYNC)
4822 		return (false);
4823 	obj = vp->v_object;
4824 	return (obj != NULL && vm_object_mightbedirty(obj));
4825 }
4826 
4827 static int
4828 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4829 {
4830 
4831 	if (vp->v_vflag & VV_NOSYNC)
4832 		return (false);
4833 	if (vp->v_iflag & VI_DEFINACT)
4834 		return (true);
4835 	return (vfs_want_msync(vp));
4836 }
4837 
4838 static void __noinline
4839 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4840 {
4841 	struct vnode *vp, *mvp;
4842 	struct vm_object *obj;
4843 	int lkflags, objflags;
4844 	bool seen_defer;
4845 
4846 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4847 	if (flags != MNT_WAIT) {
4848 		lkflags |= LK_NOWAIT;
4849 		objflags = OBJPC_NOSYNC;
4850 	} else {
4851 		objflags = OBJPC_SYNC;
4852 	}
4853 
4854 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4855 		seen_defer = false;
4856 		if (vp->v_iflag & VI_DEFINACT) {
4857 			vp->v_iflag &= ~VI_DEFINACT;
4858 			seen_defer = true;
4859 		}
4860 		if (!vfs_want_msync(vp)) {
4861 			if (seen_defer)
4862 				vfs_deferred_inactive(vp, lkflags);
4863 			else
4864 				VI_UNLOCK(vp);
4865 			continue;
4866 		}
4867 		if (vget(vp, lkflags) == 0) {
4868 			obj = vp->v_object;
4869 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4870 				VM_OBJECT_WLOCK(obj);
4871 				vm_object_page_clean(obj, 0, 0, objflags);
4872 				VM_OBJECT_WUNLOCK(obj);
4873 			}
4874 			vput(vp);
4875 			if (seen_defer)
4876 				vdrop(vp);
4877 		} else {
4878 			if (seen_defer)
4879 				vdefer_inactive_unlocked(vp);
4880 		}
4881 	}
4882 }
4883 
4884 void
4885 vfs_periodic(struct mount *mp, int flags)
4886 {
4887 
4888 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4889 
4890 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4891 		vfs_periodic_inactive(mp, flags);
4892 	else
4893 		vfs_periodic_msync_inactive(mp, flags);
4894 }
4895 
4896 static void
4897 destroy_vpollinfo_free(struct vpollinfo *vi)
4898 {
4899 
4900 	knlist_destroy(&vi->vpi_selinfo.si_note);
4901 	mtx_destroy(&vi->vpi_lock);
4902 	free(vi, M_VNODEPOLL);
4903 }
4904 
4905 static void
4906 destroy_vpollinfo(struct vpollinfo *vi)
4907 {
4908 
4909 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4910 	seldrain(&vi->vpi_selinfo);
4911 	destroy_vpollinfo_free(vi);
4912 }
4913 
4914 /*
4915  * Initialize per-vnode helper structure to hold poll-related state.
4916  */
4917 void
4918 v_addpollinfo(struct vnode *vp)
4919 {
4920 	struct vpollinfo *vi;
4921 
4922 	if (vp->v_pollinfo != NULL)
4923 		return;
4924 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4925 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4926 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4927 	    vfs_knlunlock, vfs_knl_assert_lock);
4928 	VI_LOCK(vp);
4929 	if (vp->v_pollinfo != NULL) {
4930 		VI_UNLOCK(vp);
4931 		destroy_vpollinfo_free(vi);
4932 		return;
4933 	}
4934 	vp->v_pollinfo = vi;
4935 	VI_UNLOCK(vp);
4936 }
4937 
4938 /*
4939  * Record a process's interest in events which might happen to
4940  * a vnode.  Because poll uses the historic select-style interface
4941  * internally, this routine serves as both the ``check for any
4942  * pending events'' and the ``record my interest in future events''
4943  * functions.  (These are done together, while the lock is held,
4944  * to avoid race conditions.)
4945  */
4946 int
4947 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4948 {
4949 
4950 	v_addpollinfo(vp);
4951 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4952 	if (vp->v_pollinfo->vpi_revents & events) {
4953 		/*
4954 		 * This leaves events we are not interested
4955 		 * in available for the other process which
4956 		 * which presumably had requested them
4957 		 * (otherwise they would never have been
4958 		 * recorded).
4959 		 */
4960 		events &= vp->v_pollinfo->vpi_revents;
4961 		vp->v_pollinfo->vpi_revents &= ~events;
4962 
4963 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4964 		return (events);
4965 	}
4966 	vp->v_pollinfo->vpi_events |= events;
4967 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4968 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4969 	return (0);
4970 }
4971 
4972 /*
4973  * Routine to create and manage a filesystem syncer vnode.
4974  */
4975 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4976 static int	sync_fsync(struct  vop_fsync_args *);
4977 static int	sync_inactive(struct  vop_inactive_args *);
4978 static int	sync_reclaim(struct  vop_reclaim_args *);
4979 
4980 static struct vop_vector sync_vnodeops = {
4981 	.vop_bypass =	VOP_EOPNOTSUPP,
4982 	.vop_close =	sync_close,		/* close */
4983 	.vop_fsync =	sync_fsync,		/* fsync */
4984 	.vop_inactive =	sync_inactive,	/* inactive */
4985 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4986 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4987 	.vop_lock1 =	vop_stdlock,	/* lock */
4988 	.vop_unlock =	vop_stdunlock,	/* unlock */
4989 	.vop_islocked =	vop_stdislocked,	/* islocked */
4990 };
4991 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4992 
4993 /*
4994  * Create a new filesystem syncer vnode for the specified mount point.
4995  */
4996 void
4997 vfs_allocate_syncvnode(struct mount *mp)
4998 {
4999 	struct vnode *vp;
5000 	struct bufobj *bo;
5001 	static long start, incr, next;
5002 	int error;
5003 
5004 	/* Allocate a new vnode */
5005 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5006 	if (error != 0)
5007 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5008 	vp->v_type = VNON;
5009 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5010 	vp->v_vflag |= VV_FORCEINSMQ;
5011 	error = insmntque(vp, mp);
5012 	if (error != 0)
5013 		panic("vfs_allocate_syncvnode: insmntque() failed");
5014 	vp->v_vflag &= ~VV_FORCEINSMQ;
5015 	VOP_UNLOCK(vp);
5016 	/*
5017 	 * Place the vnode onto the syncer worklist. We attempt to
5018 	 * scatter them about on the list so that they will go off
5019 	 * at evenly distributed times even if all the filesystems
5020 	 * are mounted at once.
5021 	 */
5022 	next += incr;
5023 	if (next == 0 || next > syncer_maxdelay) {
5024 		start /= 2;
5025 		incr /= 2;
5026 		if (start == 0) {
5027 			start = syncer_maxdelay / 2;
5028 			incr = syncer_maxdelay;
5029 		}
5030 		next = start;
5031 	}
5032 	bo = &vp->v_bufobj;
5033 	BO_LOCK(bo);
5034 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5035 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5036 	mtx_lock(&sync_mtx);
5037 	sync_vnode_count++;
5038 	if (mp->mnt_syncer == NULL) {
5039 		mp->mnt_syncer = vp;
5040 		vp = NULL;
5041 	}
5042 	mtx_unlock(&sync_mtx);
5043 	BO_UNLOCK(bo);
5044 	if (vp != NULL) {
5045 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5046 		vgone(vp);
5047 		vput(vp);
5048 	}
5049 }
5050 
5051 void
5052 vfs_deallocate_syncvnode(struct mount *mp)
5053 {
5054 	struct vnode *vp;
5055 
5056 	mtx_lock(&sync_mtx);
5057 	vp = mp->mnt_syncer;
5058 	if (vp != NULL)
5059 		mp->mnt_syncer = NULL;
5060 	mtx_unlock(&sync_mtx);
5061 	if (vp != NULL)
5062 		vrele(vp);
5063 }
5064 
5065 /*
5066  * Do a lazy sync of the filesystem.
5067  */
5068 static int
5069 sync_fsync(struct vop_fsync_args *ap)
5070 {
5071 	struct vnode *syncvp = ap->a_vp;
5072 	struct mount *mp = syncvp->v_mount;
5073 	int error, save;
5074 	struct bufobj *bo;
5075 
5076 	/*
5077 	 * We only need to do something if this is a lazy evaluation.
5078 	 */
5079 	if (ap->a_waitfor != MNT_LAZY)
5080 		return (0);
5081 
5082 	/*
5083 	 * Move ourselves to the back of the sync list.
5084 	 */
5085 	bo = &syncvp->v_bufobj;
5086 	BO_LOCK(bo);
5087 	vn_syncer_add_to_worklist(bo, syncdelay);
5088 	BO_UNLOCK(bo);
5089 
5090 	/*
5091 	 * Walk the list of vnodes pushing all that are dirty and
5092 	 * not already on the sync list.
5093 	 */
5094 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5095 		return (0);
5096 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5097 		vfs_unbusy(mp);
5098 		return (0);
5099 	}
5100 	save = curthread_pflags_set(TDP_SYNCIO);
5101 	/*
5102 	 * The filesystem at hand may be idle with free vnodes stored in the
5103 	 * batch.  Return them instead of letting them stay there indefinitely.
5104 	 */
5105 	vfs_periodic(mp, MNT_NOWAIT);
5106 	error = VFS_SYNC(mp, MNT_LAZY);
5107 	curthread_pflags_restore(save);
5108 	vn_finished_write(mp);
5109 	vfs_unbusy(mp);
5110 	return (error);
5111 }
5112 
5113 /*
5114  * The syncer vnode is no referenced.
5115  */
5116 static int
5117 sync_inactive(struct vop_inactive_args *ap)
5118 {
5119 
5120 	vgone(ap->a_vp);
5121 	return (0);
5122 }
5123 
5124 /*
5125  * The syncer vnode is no longer needed and is being decommissioned.
5126  *
5127  * Modifications to the worklist must be protected by sync_mtx.
5128  */
5129 static int
5130 sync_reclaim(struct vop_reclaim_args *ap)
5131 {
5132 	struct vnode *vp = ap->a_vp;
5133 	struct bufobj *bo;
5134 
5135 	bo = &vp->v_bufobj;
5136 	BO_LOCK(bo);
5137 	mtx_lock(&sync_mtx);
5138 	if (vp->v_mount->mnt_syncer == vp)
5139 		vp->v_mount->mnt_syncer = NULL;
5140 	if (bo->bo_flag & BO_ONWORKLST) {
5141 		LIST_REMOVE(bo, bo_synclist);
5142 		syncer_worklist_len--;
5143 		sync_vnode_count--;
5144 		bo->bo_flag &= ~BO_ONWORKLST;
5145 	}
5146 	mtx_unlock(&sync_mtx);
5147 	BO_UNLOCK(bo);
5148 
5149 	return (0);
5150 }
5151 
5152 int
5153 vn_need_pageq_flush(struct vnode *vp)
5154 {
5155 	struct vm_object *obj;
5156 
5157 	obj = vp->v_object;
5158 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5159 	    vm_object_mightbedirty(obj));
5160 }
5161 
5162 /*
5163  * Check if vnode represents a disk device
5164  */
5165 bool
5166 vn_isdisk_error(struct vnode *vp, int *errp)
5167 {
5168 	int error;
5169 
5170 	if (vp->v_type != VCHR) {
5171 		error = ENOTBLK;
5172 		goto out;
5173 	}
5174 	error = 0;
5175 	dev_lock();
5176 	if (vp->v_rdev == NULL)
5177 		error = ENXIO;
5178 	else if (vp->v_rdev->si_devsw == NULL)
5179 		error = ENXIO;
5180 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5181 		error = ENOTBLK;
5182 	dev_unlock();
5183 out:
5184 	*errp = error;
5185 	return (error == 0);
5186 }
5187 
5188 bool
5189 vn_isdisk(struct vnode *vp)
5190 {
5191 	int error;
5192 
5193 	return (vn_isdisk_error(vp, &error));
5194 }
5195 
5196 /*
5197  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5198  * the comment above cache_fplookup for details.
5199  */
5200 int
5201 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5202 {
5203 	int error;
5204 
5205 	VFS_SMR_ASSERT_ENTERED();
5206 
5207 	/* Check the owner. */
5208 	if (cred->cr_uid == file_uid) {
5209 		if (file_mode & S_IXUSR)
5210 			return (0);
5211 		goto out_error;
5212 	}
5213 
5214 	/* Otherwise, check the groups (first match) */
5215 	if (groupmember(file_gid, cred)) {
5216 		if (file_mode & S_IXGRP)
5217 			return (0);
5218 		goto out_error;
5219 	}
5220 
5221 	/* Otherwise, check everyone else. */
5222 	if (file_mode & S_IXOTH)
5223 		return (0);
5224 out_error:
5225 	/*
5226 	 * Permission check failed, but it is possible denial will get overwritten
5227 	 * (e.g., when root is traversing through a 700 directory owned by someone
5228 	 * else).
5229 	 *
5230 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5231 	 * modules overriding this result. It's quite unclear what semantics
5232 	 * are allowed for them to operate, thus for safety we don't call them
5233 	 * from within the SMR section. This also means if any such modules
5234 	 * are present, we have to let the regular lookup decide.
5235 	 */
5236 	error = priv_check_cred_vfs_lookup_nomac(cred);
5237 	switch (error) {
5238 	case 0:
5239 		return (0);
5240 	case EAGAIN:
5241 		/*
5242 		 * MAC modules present.
5243 		 */
5244 		return (EAGAIN);
5245 	case EPERM:
5246 		return (EACCES);
5247 	default:
5248 		return (error);
5249 	}
5250 }
5251 
5252 /*
5253  * Common filesystem object access control check routine.  Accepts a
5254  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5255  * Returns 0 on success, or an errno on failure.
5256  */
5257 int
5258 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5259     accmode_t accmode, struct ucred *cred)
5260 {
5261 	accmode_t dac_granted;
5262 	accmode_t priv_granted;
5263 
5264 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5265 	    ("invalid bit in accmode"));
5266 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5267 	    ("VAPPEND without VWRITE"));
5268 
5269 	/*
5270 	 * Look for a normal, non-privileged way to access the file/directory
5271 	 * as requested.  If it exists, go with that.
5272 	 */
5273 
5274 	dac_granted = 0;
5275 
5276 	/* Check the owner. */
5277 	if (cred->cr_uid == file_uid) {
5278 		dac_granted |= VADMIN;
5279 		if (file_mode & S_IXUSR)
5280 			dac_granted |= VEXEC;
5281 		if (file_mode & S_IRUSR)
5282 			dac_granted |= VREAD;
5283 		if (file_mode & S_IWUSR)
5284 			dac_granted |= (VWRITE | VAPPEND);
5285 
5286 		if ((accmode & dac_granted) == accmode)
5287 			return (0);
5288 
5289 		goto privcheck;
5290 	}
5291 
5292 	/* Otherwise, check the groups (first match) */
5293 	if (groupmember(file_gid, cred)) {
5294 		if (file_mode & S_IXGRP)
5295 			dac_granted |= VEXEC;
5296 		if (file_mode & S_IRGRP)
5297 			dac_granted |= VREAD;
5298 		if (file_mode & S_IWGRP)
5299 			dac_granted |= (VWRITE | VAPPEND);
5300 
5301 		if ((accmode & dac_granted) == accmode)
5302 			return (0);
5303 
5304 		goto privcheck;
5305 	}
5306 
5307 	/* Otherwise, check everyone else. */
5308 	if (file_mode & S_IXOTH)
5309 		dac_granted |= VEXEC;
5310 	if (file_mode & S_IROTH)
5311 		dac_granted |= VREAD;
5312 	if (file_mode & S_IWOTH)
5313 		dac_granted |= (VWRITE | VAPPEND);
5314 	if ((accmode & dac_granted) == accmode)
5315 		return (0);
5316 
5317 privcheck:
5318 	/*
5319 	 * Build a privilege mask to determine if the set of privileges
5320 	 * satisfies the requirements when combined with the granted mask
5321 	 * from above.  For each privilege, if the privilege is required,
5322 	 * bitwise or the request type onto the priv_granted mask.
5323 	 */
5324 	priv_granted = 0;
5325 
5326 	if (type == VDIR) {
5327 		/*
5328 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5329 		 * requests, instead of PRIV_VFS_EXEC.
5330 		 */
5331 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5332 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5333 			priv_granted |= VEXEC;
5334 	} else {
5335 		/*
5336 		 * Ensure that at least one execute bit is on. Otherwise,
5337 		 * a privileged user will always succeed, and we don't want
5338 		 * this to happen unless the file really is executable.
5339 		 */
5340 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5341 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5342 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5343 			priv_granted |= VEXEC;
5344 	}
5345 
5346 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5347 	    !priv_check_cred(cred, PRIV_VFS_READ))
5348 		priv_granted |= VREAD;
5349 
5350 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5351 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5352 		priv_granted |= (VWRITE | VAPPEND);
5353 
5354 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5355 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5356 		priv_granted |= VADMIN;
5357 
5358 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5359 		return (0);
5360 	}
5361 
5362 	return ((accmode & VADMIN) ? EPERM : EACCES);
5363 }
5364 
5365 /*
5366  * Credential check based on process requesting service, and per-attribute
5367  * permissions.
5368  */
5369 int
5370 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5371     struct thread *td, accmode_t accmode)
5372 {
5373 
5374 	/*
5375 	 * Kernel-invoked always succeeds.
5376 	 */
5377 	if (cred == NOCRED)
5378 		return (0);
5379 
5380 	/*
5381 	 * Do not allow privileged processes in jail to directly manipulate
5382 	 * system attributes.
5383 	 */
5384 	switch (attrnamespace) {
5385 	case EXTATTR_NAMESPACE_SYSTEM:
5386 		/* Potentially should be: return (EPERM); */
5387 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5388 	case EXTATTR_NAMESPACE_USER:
5389 		return (VOP_ACCESS(vp, accmode, cred, td));
5390 	default:
5391 		return (EPERM);
5392 	}
5393 }
5394 
5395 #ifdef DEBUG_VFS_LOCKS
5396 /*
5397  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5398  * no longer ok to have an unlocked VFS.
5399  */
5400 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5401 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5402 
5403 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5404 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5405     "Drop into debugger on lock violation");
5406 
5407 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5408 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5409     0, "Check for interlock across VOPs");
5410 
5411 int vfs_badlock_print = 1;	/* Print lock violations. */
5412 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5413     0, "Print lock violations");
5414 
5415 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5416 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5417     0, "Print vnode details on lock violations");
5418 
5419 #ifdef KDB
5420 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5421 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5422     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5423 #endif
5424 
5425 static void
5426 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5427 {
5428 
5429 #ifdef KDB
5430 	if (vfs_badlock_backtrace)
5431 		kdb_backtrace();
5432 #endif
5433 	if (vfs_badlock_vnode)
5434 		vn_printf(vp, "vnode ");
5435 	if (vfs_badlock_print)
5436 		printf("%s: %p %s\n", str, (void *)vp, msg);
5437 	if (vfs_badlock_ddb)
5438 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5439 }
5440 
5441 void
5442 assert_vi_locked(struct vnode *vp, const char *str)
5443 {
5444 
5445 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5446 		vfs_badlock("interlock is not locked but should be", str, vp);
5447 }
5448 
5449 void
5450 assert_vi_unlocked(struct vnode *vp, const char *str)
5451 {
5452 
5453 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5454 		vfs_badlock("interlock is locked but should not be", str, vp);
5455 }
5456 
5457 void
5458 assert_vop_locked(struct vnode *vp, const char *str)
5459 {
5460 	int locked;
5461 
5462 	if (!IGNORE_LOCK(vp)) {
5463 		locked = VOP_ISLOCKED(vp);
5464 		if (locked == 0 || locked == LK_EXCLOTHER)
5465 			vfs_badlock("is not locked but should be", str, vp);
5466 	}
5467 }
5468 
5469 void
5470 assert_vop_unlocked(struct vnode *vp, const char *str)
5471 {
5472 
5473 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5474 		vfs_badlock("is locked but should not be", str, vp);
5475 }
5476 
5477 void
5478 assert_vop_elocked(struct vnode *vp, const char *str)
5479 {
5480 
5481 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5482 		vfs_badlock("is not exclusive locked but should be", str, vp);
5483 }
5484 #endif /* DEBUG_VFS_LOCKS */
5485 
5486 void
5487 vop_rename_fail(struct vop_rename_args *ap)
5488 {
5489 
5490 	if (ap->a_tvp != NULL)
5491 		vput(ap->a_tvp);
5492 	if (ap->a_tdvp == ap->a_tvp)
5493 		vrele(ap->a_tdvp);
5494 	else
5495 		vput(ap->a_tdvp);
5496 	vrele(ap->a_fdvp);
5497 	vrele(ap->a_fvp);
5498 }
5499 
5500 void
5501 vop_rename_pre(void *ap)
5502 {
5503 	struct vop_rename_args *a = ap;
5504 
5505 #ifdef DEBUG_VFS_LOCKS
5506 	if (a->a_tvp)
5507 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5508 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5509 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5510 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5511 
5512 	/* Check the source (from). */
5513 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5514 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5515 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5516 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5517 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5518 
5519 	/* Check the target. */
5520 	if (a->a_tvp)
5521 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5522 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5523 #endif
5524 	/*
5525 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5526 	 * in vop_rename_post but that's not going to work out since some
5527 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5528 	 *
5529 	 * For now filesystems are expected to do the relevant calls after they
5530 	 * decide what vnodes to operate on.
5531 	 */
5532 	if (a->a_tdvp != a->a_fdvp)
5533 		vhold(a->a_fdvp);
5534 	if (a->a_tvp != a->a_fvp)
5535 		vhold(a->a_fvp);
5536 	vhold(a->a_tdvp);
5537 	if (a->a_tvp)
5538 		vhold(a->a_tvp);
5539 }
5540 
5541 #ifdef DEBUG_VFS_LOCKS
5542 void
5543 vop_fplookup_vexec_debugpre(void *ap __unused)
5544 {
5545 
5546 	VFS_SMR_ASSERT_ENTERED();
5547 }
5548 
5549 void
5550 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5551 {
5552 
5553 	VFS_SMR_ASSERT_ENTERED();
5554 }
5555 
5556 void
5557 vop_fplookup_symlink_debugpre(void *ap __unused)
5558 {
5559 
5560 	VFS_SMR_ASSERT_ENTERED();
5561 }
5562 
5563 void
5564 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5565 {
5566 
5567 	VFS_SMR_ASSERT_ENTERED();
5568 }
5569 void
5570 vop_strategy_debugpre(void *ap)
5571 {
5572 	struct vop_strategy_args *a;
5573 	struct buf *bp;
5574 
5575 	a = ap;
5576 	bp = a->a_bp;
5577 
5578 	/*
5579 	 * Cluster ops lock their component buffers but not the IO container.
5580 	 */
5581 	if ((bp->b_flags & B_CLUSTER) != 0)
5582 		return;
5583 
5584 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5585 		if (vfs_badlock_print)
5586 			printf(
5587 			    "VOP_STRATEGY: bp is not locked but should be\n");
5588 		if (vfs_badlock_ddb)
5589 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5590 	}
5591 }
5592 
5593 void
5594 vop_lock_debugpre(void *ap)
5595 {
5596 	struct vop_lock1_args *a = ap;
5597 
5598 	if ((a->a_flags & LK_INTERLOCK) == 0)
5599 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5600 	else
5601 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5602 }
5603 
5604 void
5605 vop_lock_debugpost(void *ap, int rc)
5606 {
5607 	struct vop_lock1_args *a = ap;
5608 
5609 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5610 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5611 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5612 }
5613 
5614 void
5615 vop_unlock_debugpre(void *ap)
5616 {
5617 	struct vop_unlock_args *a = ap;
5618 
5619 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5620 }
5621 
5622 void
5623 vop_need_inactive_debugpre(void *ap)
5624 {
5625 	struct vop_need_inactive_args *a = ap;
5626 
5627 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5628 }
5629 
5630 void
5631 vop_need_inactive_debugpost(void *ap, int rc)
5632 {
5633 	struct vop_need_inactive_args *a = ap;
5634 
5635 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5636 }
5637 #endif
5638 
5639 void
5640 vop_create_pre(void *ap)
5641 {
5642 	struct vop_create_args *a;
5643 	struct vnode *dvp;
5644 
5645 	a = ap;
5646 	dvp = a->a_dvp;
5647 	vn_seqc_write_begin(dvp);
5648 }
5649 
5650 void
5651 vop_create_post(void *ap, int rc)
5652 {
5653 	struct vop_create_args *a;
5654 	struct vnode *dvp;
5655 
5656 	a = ap;
5657 	dvp = a->a_dvp;
5658 	vn_seqc_write_end(dvp);
5659 	if (!rc)
5660 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5661 }
5662 
5663 void
5664 vop_whiteout_pre(void *ap)
5665 {
5666 	struct vop_whiteout_args *a;
5667 	struct vnode *dvp;
5668 
5669 	a = ap;
5670 	dvp = a->a_dvp;
5671 	vn_seqc_write_begin(dvp);
5672 }
5673 
5674 void
5675 vop_whiteout_post(void *ap, int rc)
5676 {
5677 	struct vop_whiteout_args *a;
5678 	struct vnode *dvp;
5679 
5680 	a = ap;
5681 	dvp = a->a_dvp;
5682 	vn_seqc_write_end(dvp);
5683 }
5684 
5685 void
5686 vop_deleteextattr_pre(void *ap)
5687 {
5688 	struct vop_deleteextattr_args *a;
5689 	struct vnode *vp;
5690 
5691 	a = ap;
5692 	vp = a->a_vp;
5693 	vn_seqc_write_begin(vp);
5694 }
5695 
5696 void
5697 vop_deleteextattr_post(void *ap, int rc)
5698 {
5699 	struct vop_deleteextattr_args *a;
5700 	struct vnode *vp;
5701 
5702 	a = ap;
5703 	vp = a->a_vp;
5704 	vn_seqc_write_end(vp);
5705 	if (!rc)
5706 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5707 }
5708 
5709 void
5710 vop_link_pre(void *ap)
5711 {
5712 	struct vop_link_args *a;
5713 	struct vnode *vp, *tdvp;
5714 
5715 	a = ap;
5716 	vp = a->a_vp;
5717 	tdvp = a->a_tdvp;
5718 	vn_seqc_write_begin(vp);
5719 	vn_seqc_write_begin(tdvp);
5720 }
5721 
5722 void
5723 vop_link_post(void *ap, int rc)
5724 {
5725 	struct vop_link_args *a;
5726 	struct vnode *vp, *tdvp;
5727 
5728 	a = ap;
5729 	vp = a->a_vp;
5730 	tdvp = a->a_tdvp;
5731 	vn_seqc_write_end(vp);
5732 	vn_seqc_write_end(tdvp);
5733 	if (!rc) {
5734 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5735 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5736 	}
5737 }
5738 
5739 void
5740 vop_mkdir_pre(void *ap)
5741 {
5742 	struct vop_mkdir_args *a;
5743 	struct vnode *dvp;
5744 
5745 	a = ap;
5746 	dvp = a->a_dvp;
5747 	vn_seqc_write_begin(dvp);
5748 }
5749 
5750 void
5751 vop_mkdir_post(void *ap, int rc)
5752 {
5753 	struct vop_mkdir_args *a;
5754 	struct vnode *dvp;
5755 
5756 	a = ap;
5757 	dvp = a->a_dvp;
5758 	vn_seqc_write_end(dvp);
5759 	if (!rc)
5760 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5761 }
5762 
5763 #ifdef DEBUG_VFS_LOCKS
5764 void
5765 vop_mkdir_debugpost(void *ap, int rc)
5766 {
5767 	struct vop_mkdir_args *a;
5768 
5769 	a = ap;
5770 	if (!rc)
5771 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5772 }
5773 #endif
5774 
5775 void
5776 vop_mknod_pre(void *ap)
5777 {
5778 	struct vop_mknod_args *a;
5779 	struct vnode *dvp;
5780 
5781 	a = ap;
5782 	dvp = a->a_dvp;
5783 	vn_seqc_write_begin(dvp);
5784 }
5785 
5786 void
5787 vop_mknod_post(void *ap, int rc)
5788 {
5789 	struct vop_mknod_args *a;
5790 	struct vnode *dvp;
5791 
5792 	a = ap;
5793 	dvp = a->a_dvp;
5794 	vn_seqc_write_end(dvp);
5795 	if (!rc)
5796 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5797 }
5798 
5799 void
5800 vop_reclaim_post(void *ap, int rc)
5801 {
5802 	struct vop_reclaim_args *a;
5803 	struct vnode *vp;
5804 
5805 	a = ap;
5806 	vp = a->a_vp;
5807 	ASSERT_VOP_IN_SEQC(vp);
5808 	if (!rc)
5809 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5810 }
5811 
5812 void
5813 vop_remove_pre(void *ap)
5814 {
5815 	struct vop_remove_args *a;
5816 	struct vnode *dvp, *vp;
5817 
5818 	a = ap;
5819 	dvp = a->a_dvp;
5820 	vp = a->a_vp;
5821 	vn_seqc_write_begin(dvp);
5822 	vn_seqc_write_begin(vp);
5823 }
5824 
5825 void
5826 vop_remove_post(void *ap, int rc)
5827 {
5828 	struct vop_remove_args *a;
5829 	struct vnode *dvp, *vp;
5830 
5831 	a = ap;
5832 	dvp = a->a_dvp;
5833 	vp = a->a_vp;
5834 	vn_seqc_write_end(dvp);
5835 	vn_seqc_write_end(vp);
5836 	if (!rc) {
5837 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5838 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5839 	}
5840 }
5841 
5842 void
5843 vop_rename_post(void *ap, int rc)
5844 {
5845 	struct vop_rename_args *a = ap;
5846 	long hint;
5847 
5848 	if (!rc) {
5849 		hint = NOTE_WRITE;
5850 		if (a->a_fdvp == a->a_tdvp) {
5851 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5852 				hint |= NOTE_LINK;
5853 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5854 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5855 		} else {
5856 			hint |= NOTE_EXTEND;
5857 			if (a->a_fvp->v_type == VDIR)
5858 				hint |= NOTE_LINK;
5859 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5860 
5861 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5862 			    a->a_tvp->v_type == VDIR)
5863 				hint &= ~NOTE_LINK;
5864 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5865 		}
5866 
5867 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5868 		if (a->a_tvp)
5869 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5870 	}
5871 	if (a->a_tdvp != a->a_fdvp)
5872 		vdrop(a->a_fdvp);
5873 	if (a->a_tvp != a->a_fvp)
5874 		vdrop(a->a_fvp);
5875 	vdrop(a->a_tdvp);
5876 	if (a->a_tvp)
5877 		vdrop(a->a_tvp);
5878 }
5879 
5880 void
5881 vop_rmdir_pre(void *ap)
5882 {
5883 	struct vop_rmdir_args *a;
5884 	struct vnode *dvp, *vp;
5885 
5886 	a = ap;
5887 	dvp = a->a_dvp;
5888 	vp = a->a_vp;
5889 	vn_seqc_write_begin(dvp);
5890 	vn_seqc_write_begin(vp);
5891 }
5892 
5893 void
5894 vop_rmdir_post(void *ap, int rc)
5895 {
5896 	struct vop_rmdir_args *a;
5897 	struct vnode *dvp, *vp;
5898 
5899 	a = ap;
5900 	dvp = a->a_dvp;
5901 	vp = a->a_vp;
5902 	vn_seqc_write_end(dvp);
5903 	vn_seqc_write_end(vp);
5904 	if (!rc) {
5905 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5906 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5907 	}
5908 }
5909 
5910 void
5911 vop_setattr_pre(void *ap)
5912 {
5913 	struct vop_setattr_args *a;
5914 	struct vnode *vp;
5915 
5916 	a = ap;
5917 	vp = a->a_vp;
5918 	vn_seqc_write_begin(vp);
5919 }
5920 
5921 void
5922 vop_setattr_post(void *ap, int rc)
5923 {
5924 	struct vop_setattr_args *a;
5925 	struct vnode *vp;
5926 
5927 	a = ap;
5928 	vp = a->a_vp;
5929 	vn_seqc_write_end(vp);
5930 	if (!rc)
5931 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5932 }
5933 
5934 void
5935 vop_setacl_pre(void *ap)
5936 {
5937 	struct vop_setacl_args *a;
5938 	struct vnode *vp;
5939 
5940 	a = ap;
5941 	vp = a->a_vp;
5942 	vn_seqc_write_begin(vp);
5943 }
5944 
5945 void
5946 vop_setacl_post(void *ap, int rc __unused)
5947 {
5948 	struct vop_setacl_args *a;
5949 	struct vnode *vp;
5950 
5951 	a = ap;
5952 	vp = a->a_vp;
5953 	vn_seqc_write_end(vp);
5954 }
5955 
5956 void
5957 vop_setextattr_pre(void *ap)
5958 {
5959 	struct vop_setextattr_args *a;
5960 	struct vnode *vp;
5961 
5962 	a = ap;
5963 	vp = a->a_vp;
5964 	vn_seqc_write_begin(vp);
5965 }
5966 
5967 void
5968 vop_setextattr_post(void *ap, int rc)
5969 {
5970 	struct vop_setextattr_args *a;
5971 	struct vnode *vp;
5972 
5973 	a = ap;
5974 	vp = a->a_vp;
5975 	vn_seqc_write_end(vp);
5976 	if (!rc)
5977 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5978 }
5979 
5980 void
5981 vop_symlink_pre(void *ap)
5982 {
5983 	struct vop_symlink_args *a;
5984 	struct vnode *dvp;
5985 
5986 	a = ap;
5987 	dvp = a->a_dvp;
5988 	vn_seqc_write_begin(dvp);
5989 }
5990 
5991 void
5992 vop_symlink_post(void *ap, int rc)
5993 {
5994 	struct vop_symlink_args *a;
5995 	struct vnode *dvp;
5996 
5997 	a = ap;
5998 	dvp = a->a_dvp;
5999 	vn_seqc_write_end(dvp);
6000 	if (!rc)
6001 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6002 }
6003 
6004 void
6005 vop_open_post(void *ap, int rc)
6006 {
6007 	struct vop_open_args *a = ap;
6008 
6009 	if (!rc)
6010 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6011 }
6012 
6013 void
6014 vop_close_post(void *ap, int rc)
6015 {
6016 	struct vop_close_args *a = ap;
6017 
6018 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6019 	    !VN_IS_DOOMED(a->a_vp))) {
6020 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6021 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6022 	}
6023 }
6024 
6025 void
6026 vop_read_post(void *ap, int rc)
6027 {
6028 	struct vop_read_args *a = ap;
6029 
6030 	if (!rc)
6031 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6032 }
6033 
6034 void
6035 vop_read_pgcache_post(void *ap, int rc)
6036 {
6037 	struct vop_read_pgcache_args *a = ap;
6038 
6039 	if (!rc)
6040 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6041 }
6042 
6043 void
6044 vop_readdir_post(void *ap, int rc)
6045 {
6046 	struct vop_readdir_args *a = ap;
6047 
6048 	if (!rc)
6049 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6050 }
6051 
6052 static struct knlist fs_knlist;
6053 
6054 static void
6055 vfs_event_init(void *arg)
6056 {
6057 	knlist_init_mtx(&fs_knlist, NULL);
6058 }
6059 /* XXX - correct order? */
6060 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6061 
6062 void
6063 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6064 {
6065 
6066 	KNOTE_UNLOCKED(&fs_knlist, event);
6067 }
6068 
6069 static int	filt_fsattach(struct knote *kn);
6070 static void	filt_fsdetach(struct knote *kn);
6071 static int	filt_fsevent(struct knote *kn, long hint);
6072 
6073 struct filterops fs_filtops = {
6074 	.f_isfd = 0,
6075 	.f_attach = filt_fsattach,
6076 	.f_detach = filt_fsdetach,
6077 	.f_event = filt_fsevent
6078 };
6079 
6080 static int
6081 filt_fsattach(struct knote *kn)
6082 {
6083 
6084 	kn->kn_flags |= EV_CLEAR;
6085 	knlist_add(&fs_knlist, kn, 0);
6086 	return (0);
6087 }
6088 
6089 static void
6090 filt_fsdetach(struct knote *kn)
6091 {
6092 
6093 	knlist_remove(&fs_knlist, kn, 0);
6094 }
6095 
6096 static int
6097 filt_fsevent(struct knote *kn, long hint)
6098 {
6099 
6100 	kn->kn_fflags |= kn->kn_sfflags & hint;
6101 
6102 	return (kn->kn_fflags != 0);
6103 }
6104 
6105 static int
6106 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6107 {
6108 	struct vfsidctl vc;
6109 	int error;
6110 	struct mount *mp;
6111 
6112 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6113 	if (error)
6114 		return (error);
6115 	if (vc.vc_vers != VFS_CTL_VERS1)
6116 		return (EINVAL);
6117 	mp = vfs_getvfs(&vc.vc_fsid);
6118 	if (mp == NULL)
6119 		return (ENOENT);
6120 	/* ensure that a specific sysctl goes to the right filesystem. */
6121 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6122 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6123 		vfs_rel(mp);
6124 		return (EINVAL);
6125 	}
6126 	VCTLTOREQ(&vc, req);
6127 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6128 	vfs_rel(mp);
6129 	return (error);
6130 }
6131 
6132 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6133     NULL, 0, sysctl_vfs_ctl, "",
6134     "Sysctl by fsid");
6135 
6136 /*
6137  * Function to initialize a va_filerev field sensibly.
6138  * XXX: Wouldn't a random number make a lot more sense ??
6139  */
6140 u_quad_t
6141 init_va_filerev(void)
6142 {
6143 	struct bintime bt;
6144 
6145 	getbinuptime(&bt);
6146 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6147 }
6148 
6149 static int	filt_vfsread(struct knote *kn, long hint);
6150 static int	filt_vfswrite(struct knote *kn, long hint);
6151 static int	filt_vfsvnode(struct knote *kn, long hint);
6152 static void	filt_vfsdetach(struct knote *kn);
6153 static struct filterops vfsread_filtops = {
6154 	.f_isfd = 1,
6155 	.f_detach = filt_vfsdetach,
6156 	.f_event = filt_vfsread
6157 };
6158 static struct filterops vfswrite_filtops = {
6159 	.f_isfd = 1,
6160 	.f_detach = filt_vfsdetach,
6161 	.f_event = filt_vfswrite
6162 };
6163 static struct filterops vfsvnode_filtops = {
6164 	.f_isfd = 1,
6165 	.f_detach = filt_vfsdetach,
6166 	.f_event = filt_vfsvnode
6167 };
6168 
6169 static void
6170 vfs_knllock(void *arg)
6171 {
6172 	struct vnode *vp = arg;
6173 
6174 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6175 }
6176 
6177 static void
6178 vfs_knlunlock(void *arg)
6179 {
6180 	struct vnode *vp = arg;
6181 
6182 	VOP_UNLOCK(vp);
6183 }
6184 
6185 static void
6186 vfs_knl_assert_lock(void *arg, int what)
6187 {
6188 #ifdef DEBUG_VFS_LOCKS
6189 	struct vnode *vp = arg;
6190 
6191 	if (what == LA_LOCKED)
6192 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6193 	else
6194 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6195 #endif
6196 }
6197 
6198 int
6199 vfs_kqfilter(struct vop_kqfilter_args *ap)
6200 {
6201 	struct vnode *vp = ap->a_vp;
6202 	struct knote *kn = ap->a_kn;
6203 	struct knlist *knl;
6204 
6205 	switch (kn->kn_filter) {
6206 	case EVFILT_READ:
6207 		kn->kn_fop = &vfsread_filtops;
6208 		break;
6209 	case EVFILT_WRITE:
6210 		kn->kn_fop = &vfswrite_filtops;
6211 		break;
6212 	case EVFILT_VNODE:
6213 		kn->kn_fop = &vfsvnode_filtops;
6214 		break;
6215 	default:
6216 		return (EINVAL);
6217 	}
6218 
6219 	kn->kn_hook = (caddr_t)vp;
6220 
6221 	v_addpollinfo(vp);
6222 	if (vp->v_pollinfo == NULL)
6223 		return (ENOMEM);
6224 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6225 	vhold(vp);
6226 	knlist_add(knl, kn, 0);
6227 
6228 	return (0);
6229 }
6230 
6231 /*
6232  * Detach knote from vnode
6233  */
6234 static void
6235 filt_vfsdetach(struct knote *kn)
6236 {
6237 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6238 
6239 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6240 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6241 	vdrop(vp);
6242 }
6243 
6244 /*ARGSUSED*/
6245 static int
6246 filt_vfsread(struct knote *kn, long hint)
6247 {
6248 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6249 	struct vattr va;
6250 	int res;
6251 
6252 	/*
6253 	 * filesystem is gone, so set the EOF flag and schedule
6254 	 * the knote for deletion.
6255 	 */
6256 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6257 		VI_LOCK(vp);
6258 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6259 		VI_UNLOCK(vp);
6260 		return (1);
6261 	}
6262 
6263 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6264 		return (0);
6265 
6266 	VI_LOCK(vp);
6267 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6268 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6269 	VI_UNLOCK(vp);
6270 	return (res);
6271 }
6272 
6273 /*ARGSUSED*/
6274 static int
6275 filt_vfswrite(struct knote *kn, long hint)
6276 {
6277 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6278 
6279 	VI_LOCK(vp);
6280 
6281 	/*
6282 	 * filesystem is gone, so set the EOF flag and schedule
6283 	 * the knote for deletion.
6284 	 */
6285 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6286 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6287 
6288 	kn->kn_data = 0;
6289 	VI_UNLOCK(vp);
6290 	return (1);
6291 }
6292 
6293 static int
6294 filt_vfsvnode(struct knote *kn, long hint)
6295 {
6296 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6297 	int res;
6298 
6299 	VI_LOCK(vp);
6300 	if (kn->kn_sfflags & hint)
6301 		kn->kn_fflags |= hint;
6302 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6303 		kn->kn_flags |= EV_EOF;
6304 		VI_UNLOCK(vp);
6305 		return (1);
6306 	}
6307 	res = (kn->kn_fflags != 0);
6308 	VI_UNLOCK(vp);
6309 	return (res);
6310 }
6311 
6312 /*
6313  * Returns whether the directory is empty or not.
6314  * If it is empty, the return value is 0; otherwise
6315  * the return value is an error value (which may
6316  * be ENOTEMPTY).
6317  */
6318 int
6319 vfs_emptydir(struct vnode *vp)
6320 {
6321 	struct uio uio;
6322 	struct iovec iov;
6323 	struct dirent *dirent, *dp, *endp;
6324 	int error, eof;
6325 
6326 	error = 0;
6327 	eof = 0;
6328 
6329 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6330 
6331 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6332 	iov.iov_base = dirent;
6333 	iov.iov_len = sizeof(struct dirent);
6334 
6335 	uio.uio_iov = &iov;
6336 	uio.uio_iovcnt = 1;
6337 	uio.uio_offset = 0;
6338 	uio.uio_resid = sizeof(struct dirent);
6339 	uio.uio_segflg = UIO_SYSSPACE;
6340 	uio.uio_rw = UIO_READ;
6341 	uio.uio_td = curthread;
6342 
6343 	while (eof == 0 && error == 0) {
6344 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6345 		    NULL, NULL);
6346 		if (error != 0)
6347 			break;
6348 		endp = (void *)((uint8_t *)dirent +
6349 		    sizeof(struct dirent) - uio.uio_resid);
6350 		for (dp = dirent; dp < endp;
6351 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6352 			if (dp->d_type == DT_WHT)
6353 				continue;
6354 			if (dp->d_namlen == 0)
6355 				continue;
6356 			if (dp->d_type != DT_DIR &&
6357 			    dp->d_type != DT_UNKNOWN) {
6358 				error = ENOTEMPTY;
6359 				break;
6360 			}
6361 			if (dp->d_namlen > 2) {
6362 				error = ENOTEMPTY;
6363 				break;
6364 			}
6365 			if (dp->d_namlen == 1 &&
6366 			    dp->d_name[0] != '.') {
6367 				error = ENOTEMPTY;
6368 				break;
6369 			}
6370 			if (dp->d_namlen == 2 &&
6371 			    dp->d_name[1] != '.') {
6372 				error = ENOTEMPTY;
6373 				break;
6374 			}
6375 			uio.uio_resid = sizeof(struct dirent);
6376 		}
6377 	}
6378 	free(dirent, M_TEMP);
6379 	return (error);
6380 }
6381 
6382 int
6383 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6384 {
6385 	int error;
6386 
6387 	if (dp->d_reclen > ap->a_uio->uio_resid)
6388 		return (ENAMETOOLONG);
6389 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6390 	if (error) {
6391 		if (ap->a_ncookies != NULL) {
6392 			if (ap->a_cookies != NULL)
6393 				free(ap->a_cookies, M_TEMP);
6394 			ap->a_cookies = NULL;
6395 			*ap->a_ncookies = 0;
6396 		}
6397 		return (error);
6398 	}
6399 	if (ap->a_ncookies == NULL)
6400 		return (0);
6401 
6402 	KASSERT(ap->a_cookies,
6403 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6404 
6405 	*ap->a_cookies = realloc(*ap->a_cookies,
6406 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6407 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6408 	*ap->a_ncookies += 1;
6409 	return (0);
6410 }
6411 
6412 /*
6413  * The purpose of this routine is to remove granularity from accmode_t,
6414  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6415  * VADMIN and VAPPEND.
6416  *
6417  * If it returns 0, the caller is supposed to continue with the usual
6418  * access checks using 'accmode' as modified by this routine.  If it
6419  * returns nonzero value, the caller is supposed to return that value
6420  * as errno.
6421  *
6422  * Note that after this routine runs, accmode may be zero.
6423  */
6424 int
6425 vfs_unixify_accmode(accmode_t *accmode)
6426 {
6427 	/*
6428 	 * There is no way to specify explicit "deny" rule using
6429 	 * file mode or POSIX.1e ACLs.
6430 	 */
6431 	if (*accmode & VEXPLICIT_DENY) {
6432 		*accmode = 0;
6433 		return (0);
6434 	}
6435 
6436 	/*
6437 	 * None of these can be translated into usual access bits.
6438 	 * Also, the common case for NFSv4 ACLs is to not contain
6439 	 * either of these bits. Caller should check for VWRITE
6440 	 * on the containing directory instead.
6441 	 */
6442 	if (*accmode & (VDELETE_CHILD | VDELETE))
6443 		return (EPERM);
6444 
6445 	if (*accmode & VADMIN_PERMS) {
6446 		*accmode &= ~VADMIN_PERMS;
6447 		*accmode |= VADMIN;
6448 	}
6449 
6450 	/*
6451 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6452 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6453 	 */
6454 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6455 
6456 	return (0);
6457 }
6458 
6459 /*
6460  * Clear out a doomed vnode (if any) and replace it with a new one as long
6461  * as the fs is not being unmounted. Return the root vnode to the caller.
6462  */
6463 static int __noinline
6464 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6465 {
6466 	struct vnode *vp;
6467 	int error;
6468 
6469 restart:
6470 	if (mp->mnt_rootvnode != NULL) {
6471 		MNT_ILOCK(mp);
6472 		vp = mp->mnt_rootvnode;
6473 		if (vp != NULL) {
6474 			if (!VN_IS_DOOMED(vp)) {
6475 				vrefact(vp);
6476 				MNT_IUNLOCK(mp);
6477 				error = vn_lock(vp, flags);
6478 				if (error == 0) {
6479 					*vpp = vp;
6480 					return (0);
6481 				}
6482 				vrele(vp);
6483 				goto restart;
6484 			}
6485 			/*
6486 			 * Clear the old one.
6487 			 */
6488 			mp->mnt_rootvnode = NULL;
6489 		}
6490 		MNT_IUNLOCK(mp);
6491 		if (vp != NULL) {
6492 			vfs_op_barrier_wait(mp);
6493 			vrele(vp);
6494 		}
6495 	}
6496 	error = VFS_CACHEDROOT(mp, flags, vpp);
6497 	if (error != 0)
6498 		return (error);
6499 	if (mp->mnt_vfs_ops == 0) {
6500 		MNT_ILOCK(mp);
6501 		if (mp->mnt_vfs_ops != 0) {
6502 			MNT_IUNLOCK(mp);
6503 			return (0);
6504 		}
6505 		if (mp->mnt_rootvnode == NULL) {
6506 			vrefact(*vpp);
6507 			mp->mnt_rootvnode = *vpp;
6508 		} else {
6509 			if (mp->mnt_rootvnode != *vpp) {
6510 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6511 					panic("%s: mismatch between vnode returned "
6512 					    " by VFS_CACHEDROOT and the one cached "
6513 					    " (%p != %p)",
6514 					    __func__, *vpp, mp->mnt_rootvnode);
6515 				}
6516 			}
6517 		}
6518 		MNT_IUNLOCK(mp);
6519 	}
6520 	return (0);
6521 }
6522 
6523 int
6524 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6525 {
6526 	struct mount_pcpu *mpcpu;
6527 	struct vnode *vp;
6528 	int error;
6529 
6530 	if (!vfs_op_thread_enter(mp, mpcpu))
6531 		return (vfs_cache_root_fallback(mp, flags, vpp));
6532 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6533 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6534 		vfs_op_thread_exit(mp, mpcpu);
6535 		return (vfs_cache_root_fallback(mp, flags, vpp));
6536 	}
6537 	vrefact(vp);
6538 	vfs_op_thread_exit(mp, mpcpu);
6539 	error = vn_lock(vp, flags);
6540 	if (error != 0) {
6541 		vrele(vp);
6542 		return (vfs_cache_root_fallback(mp, flags, vpp));
6543 	}
6544 	*vpp = vp;
6545 	return (0);
6546 }
6547 
6548 struct vnode *
6549 vfs_cache_root_clear(struct mount *mp)
6550 {
6551 	struct vnode *vp;
6552 
6553 	/*
6554 	 * ops > 0 guarantees there is nobody who can see this vnode
6555 	 */
6556 	MPASS(mp->mnt_vfs_ops > 0);
6557 	vp = mp->mnt_rootvnode;
6558 	if (vp != NULL)
6559 		vn_seqc_write_begin(vp);
6560 	mp->mnt_rootvnode = NULL;
6561 	return (vp);
6562 }
6563 
6564 void
6565 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6566 {
6567 
6568 	MPASS(mp->mnt_vfs_ops > 0);
6569 	vrefact(vp);
6570 	mp->mnt_rootvnode = vp;
6571 }
6572 
6573 /*
6574  * These are helper functions for filesystems to traverse all
6575  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6576  *
6577  * This interface replaces MNT_VNODE_FOREACH.
6578  */
6579 
6580 struct vnode *
6581 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6582 {
6583 	struct vnode *vp;
6584 
6585 	if (should_yield())
6586 		kern_yield(PRI_USER);
6587 	MNT_ILOCK(mp);
6588 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6589 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6590 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6591 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6592 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6593 			continue;
6594 		VI_LOCK(vp);
6595 		if (VN_IS_DOOMED(vp)) {
6596 			VI_UNLOCK(vp);
6597 			continue;
6598 		}
6599 		break;
6600 	}
6601 	if (vp == NULL) {
6602 		__mnt_vnode_markerfree_all(mvp, mp);
6603 		/* MNT_IUNLOCK(mp); -- done in above function */
6604 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6605 		return (NULL);
6606 	}
6607 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6608 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6609 	MNT_IUNLOCK(mp);
6610 	return (vp);
6611 }
6612 
6613 struct vnode *
6614 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6615 {
6616 	struct vnode *vp;
6617 
6618 	*mvp = vn_alloc_marker(mp);
6619 	MNT_ILOCK(mp);
6620 	MNT_REF(mp);
6621 
6622 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6623 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6624 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6625 			continue;
6626 		VI_LOCK(vp);
6627 		if (VN_IS_DOOMED(vp)) {
6628 			VI_UNLOCK(vp);
6629 			continue;
6630 		}
6631 		break;
6632 	}
6633 	if (vp == NULL) {
6634 		MNT_REL(mp);
6635 		MNT_IUNLOCK(mp);
6636 		vn_free_marker(*mvp);
6637 		*mvp = NULL;
6638 		return (NULL);
6639 	}
6640 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6641 	MNT_IUNLOCK(mp);
6642 	return (vp);
6643 }
6644 
6645 void
6646 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6647 {
6648 
6649 	if (*mvp == NULL) {
6650 		MNT_IUNLOCK(mp);
6651 		return;
6652 	}
6653 
6654 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6655 
6656 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6657 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6658 	MNT_REL(mp);
6659 	MNT_IUNLOCK(mp);
6660 	vn_free_marker(*mvp);
6661 	*mvp = NULL;
6662 }
6663 
6664 /*
6665  * These are helper functions for filesystems to traverse their
6666  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6667  */
6668 static void
6669 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6670 {
6671 
6672 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6673 
6674 	MNT_ILOCK(mp);
6675 	MNT_REL(mp);
6676 	MNT_IUNLOCK(mp);
6677 	vn_free_marker(*mvp);
6678 	*mvp = NULL;
6679 }
6680 
6681 /*
6682  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6683  * conventional lock order during mnt_vnode_next_lazy iteration.
6684  *
6685  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6686  * The list lock is dropped and reacquired.  On success, both locks are held.
6687  * On failure, the mount vnode list lock is held but the vnode interlock is
6688  * not, and the procedure may have yielded.
6689  */
6690 static bool
6691 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6692     struct vnode *vp)
6693 {
6694 
6695 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6696 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6697 	    ("%s: bad marker", __func__));
6698 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6699 	    ("%s: inappropriate vnode", __func__));
6700 	ASSERT_VI_UNLOCKED(vp, __func__);
6701 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6702 
6703 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6704 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6705 
6706 	/*
6707 	 * Note we may be racing against vdrop which transitioned the hold
6708 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6709 	 * if we are the only user after we get the interlock we will just
6710 	 * vdrop.
6711 	 */
6712 	vhold(vp);
6713 	mtx_unlock(&mp->mnt_listmtx);
6714 	VI_LOCK(vp);
6715 	if (VN_IS_DOOMED(vp)) {
6716 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6717 		goto out_lost;
6718 	}
6719 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6720 	/*
6721 	 * There is nothing to do if we are the last user.
6722 	 */
6723 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6724 		goto out_lost;
6725 	mtx_lock(&mp->mnt_listmtx);
6726 	return (true);
6727 out_lost:
6728 	vdropl(vp);
6729 	maybe_yield();
6730 	mtx_lock(&mp->mnt_listmtx);
6731 	return (false);
6732 }
6733 
6734 static struct vnode *
6735 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6736     void *cbarg)
6737 {
6738 	struct vnode *vp;
6739 
6740 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6741 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6742 restart:
6743 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6744 	while (vp != NULL) {
6745 		if (vp->v_type == VMARKER) {
6746 			vp = TAILQ_NEXT(vp, v_lazylist);
6747 			continue;
6748 		}
6749 		/*
6750 		 * See if we want to process the vnode. Note we may encounter a
6751 		 * long string of vnodes we don't care about and hog the list
6752 		 * as a result. Check for it and requeue the marker.
6753 		 */
6754 		VNPASS(!VN_IS_DOOMED(vp), vp);
6755 		if (!cb(vp, cbarg)) {
6756 			if (!should_yield()) {
6757 				vp = TAILQ_NEXT(vp, v_lazylist);
6758 				continue;
6759 			}
6760 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6761 			    v_lazylist);
6762 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6763 			    v_lazylist);
6764 			mtx_unlock(&mp->mnt_listmtx);
6765 			kern_yield(PRI_USER);
6766 			mtx_lock(&mp->mnt_listmtx);
6767 			goto restart;
6768 		}
6769 		/*
6770 		 * Try-lock because this is the wrong lock order.
6771 		 */
6772 		if (!VI_TRYLOCK(vp) &&
6773 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6774 			goto restart;
6775 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6776 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6777 		    ("alien vnode on the lazy list %p %p", vp, mp));
6778 		VNPASS(vp->v_mount == mp, vp);
6779 		VNPASS(!VN_IS_DOOMED(vp), vp);
6780 		break;
6781 	}
6782 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6783 
6784 	/* Check if we are done */
6785 	if (vp == NULL) {
6786 		mtx_unlock(&mp->mnt_listmtx);
6787 		mnt_vnode_markerfree_lazy(mvp, mp);
6788 		return (NULL);
6789 	}
6790 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6791 	mtx_unlock(&mp->mnt_listmtx);
6792 	ASSERT_VI_LOCKED(vp, "lazy iter");
6793 	return (vp);
6794 }
6795 
6796 struct vnode *
6797 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6798     void *cbarg)
6799 {
6800 
6801 	if (should_yield())
6802 		kern_yield(PRI_USER);
6803 	mtx_lock(&mp->mnt_listmtx);
6804 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6805 }
6806 
6807 struct vnode *
6808 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6809     void *cbarg)
6810 {
6811 	struct vnode *vp;
6812 
6813 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6814 		return (NULL);
6815 
6816 	*mvp = vn_alloc_marker(mp);
6817 	MNT_ILOCK(mp);
6818 	MNT_REF(mp);
6819 	MNT_IUNLOCK(mp);
6820 
6821 	mtx_lock(&mp->mnt_listmtx);
6822 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6823 	if (vp == NULL) {
6824 		mtx_unlock(&mp->mnt_listmtx);
6825 		mnt_vnode_markerfree_lazy(mvp, mp);
6826 		return (NULL);
6827 	}
6828 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6829 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6830 }
6831 
6832 void
6833 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6834 {
6835 
6836 	if (*mvp == NULL)
6837 		return;
6838 
6839 	mtx_lock(&mp->mnt_listmtx);
6840 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6841 	mtx_unlock(&mp->mnt_listmtx);
6842 	mnt_vnode_markerfree_lazy(mvp, mp);
6843 }
6844 
6845 int
6846 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6847 {
6848 
6849 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6850 		cnp->cn_flags &= ~NOEXECCHECK;
6851 		return (0);
6852 	}
6853 
6854 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6855 }
6856 
6857 /*
6858  * Do not use this variant unless you have means other than the hold count
6859  * to prevent the vnode from getting freed.
6860  */
6861 void
6862 vn_seqc_write_begin_locked(struct vnode *vp)
6863 {
6864 
6865 	ASSERT_VI_LOCKED(vp, __func__);
6866 	VNPASS(vp->v_holdcnt > 0, vp);
6867 	VNPASS(vp->v_seqc_users >= 0, vp);
6868 	vp->v_seqc_users++;
6869 	if (vp->v_seqc_users == 1)
6870 		seqc_sleepable_write_begin(&vp->v_seqc);
6871 }
6872 
6873 void
6874 vn_seqc_write_begin(struct vnode *vp)
6875 {
6876 
6877 	VI_LOCK(vp);
6878 	vn_seqc_write_begin_locked(vp);
6879 	VI_UNLOCK(vp);
6880 }
6881 
6882 void
6883 vn_seqc_write_end_locked(struct vnode *vp)
6884 {
6885 
6886 	ASSERT_VI_LOCKED(vp, __func__);
6887 	VNPASS(vp->v_seqc_users > 0, vp);
6888 	vp->v_seqc_users--;
6889 	if (vp->v_seqc_users == 0)
6890 		seqc_sleepable_write_end(&vp->v_seqc);
6891 }
6892 
6893 void
6894 vn_seqc_write_end(struct vnode *vp)
6895 {
6896 
6897 	VI_LOCK(vp);
6898 	vn_seqc_write_end_locked(vp);
6899 	VI_UNLOCK(vp);
6900 }
6901 
6902 /*
6903  * Special case handling for allocating and freeing vnodes.
6904  *
6905  * The counter remains unchanged on free so that a doomed vnode will
6906  * keep testing as in modify as long as it is accessible with SMR.
6907  */
6908 static void
6909 vn_seqc_init(struct vnode *vp)
6910 {
6911 
6912 	vp->v_seqc = 0;
6913 	vp->v_seqc_users = 0;
6914 }
6915 
6916 static void
6917 vn_seqc_write_end_free(struct vnode *vp)
6918 {
6919 
6920 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6921 	VNPASS(vp->v_seqc_users == 1, vp);
6922 }
6923 
6924 void
6925 vn_irflag_set_locked(struct vnode *vp, short toset)
6926 {
6927 	short flags;
6928 
6929 	ASSERT_VI_LOCKED(vp, __func__);
6930 	flags = vn_irflag_read(vp);
6931 	VNASSERT((flags & toset) == 0, vp,
6932 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
6933 	    __func__, flags, toset));
6934 	atomic_store_short(&vp->v_irflag, flags | toset);
6935 }
6936 
6937 void
6938 vn_irflag_set(struct vnode *vp, short toset)
6939 {
6940 
6941 	VI_LOCK(vp);
6942 	vn_irflag_set_locked(vp, toset);
6943 	VI_UNLOCK(vp);
6944 }
6945 
6946 void
6947 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6948 {
6949 	short flags;
6950 
6951 	ASSERT_VI_LOCKED(vp, __func__);
6952 	flags = vn_irflag_read(vp);
6953 	atomic_store_short(&vp->v_irflag, flags | toset);
6954 }
6955 
6956 void
6957 vn_irflag_set_cond(struct vnode *vp, short toset)
6958 {
6959 
6960 	VI_LOCK(vp);
6961 	vn_irflag_set_cond_locked(vp, toset);
6962 	VI_UNLOCK(vp);
6963 }
6964 
6965 void
6966 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6967 {
6968 	short flags;
6969 
6970 	ASSERT_VI_LOCKED(vp, __func__);
6971 	flags = vn_irflag_read(vp);
6972 	VNASSERT((flags & tounset) == tounset, vp,
6973 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
6974 	    __func__, flags, tounset));
6975 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
6976 }
6977 
6978 void
6979 vn_irflag_unset(struct vnode *vp, short tounset)
6980 {
6981 
6982 	VI_LOCK(vp);
6983 	vn_irflag_unset_locked(vp, tounset);
6984 	VI_UNLOCK(vp);
6985 }
6986