xref: /freebsd/sys/kern/vfs_subr.c (revision f2d48b5e2c3b45850585e4d7aee324fe148afbf2)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smr.h>
80 #include <sys/smp.h>
81 #include <sys/stat.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 #include <sys/watchdog.h>
87 
88 #include <machine/stdarg.h>
89 
90 #include <security/mac/mac_framework.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_extern.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_kern.h>
99 #include <vm/uma.h>
100 
101 #ifdef DDB
102 #include <ddb/ddb.h>
103 #endif
104 
105 static void	delmntque(struct vnode *vp);
106 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
107 		    int slpflag, int slptimeo);
108 static void	syncer_shutdown(void *arg, int howto);
109 static int	vtryrecycle(struct vnode *vp);
110 static void	v_init_counters(struct vnode *);
111 static void	vgonel(struct vnode *);
112 static bool	vhold_recycle_free(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_lock(void *arg, int what);
116 static void	destroy_vpollinfo(struct vpollinfo *vi);
117 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
118 		    daddr_t startlbn, daddr_t endlbn);
119 static void	vnlru_recalc(void);
120 
121 /*
122  * These fences are intended for cases where some synchronization is
123  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
124  * and v_usecount) updates.  Access to v_iflags is generally synchronized
125  * by the interlock, but we have some internal assertions that check vnode
126  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
127  * for now.
128  */
129 #ifdef INVARIANTS
130 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
131 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
132 #else
133 #define	VNODE_REFCOUNT_FENCE_ACQ()
134 #define	VNODE_REFCOUNT_FENCE_REL()
135 #endif
136 
137 /*
138  * Number of vnodes in existence.  Increased whenever getnewvnode()
139  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
140  */
141 static u_long __exclusive_cache_line numvnodes;
142 
143 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
144     "Number of vnodes in existence");
145 
146 static counter_u64_t vnodes_created;
147 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
148     "Number of vnodes created by getnewvnode");
149 
150 /*
151  * Conversion tables for conversion from vnode types to inode formats
152  * and back.
153  */
154 enum vtype iftovt_tab[16] = {
155 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
156 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
157 };
158 int vttoif_tab[10] = {
159 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
160 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
161 };
162 
163 /*
164  * List of allocates vnodes in the system.
165  */
166 static TAILQ_HEAD(freelst, vnode) vnode_list;
167 static struct vnode *vnode_list_free_marker;
168 static struct vnode *vnode_list_reclaim_marker;
169 
170 /*
171  * "Free" vnode target.  Free vnodes are rarely completely free, but are
172  * just ones that are cheap to recycle.  Usually they are for files which
173  * have been stat'd but not read; these usually have inode and namecache
174  * data attached to them.  This target is the preferred minimum size of a
175  * sub-cache consisting mostly of such files. The system balances the size
176  * of this sub-cache with its complement to try to prevent either from
177  * thrashing while the other is relatively inactive.  The targets express
178  * a preference for the best balance.
179  *
180  * "Above" this target there are 2 further targets (watermarks) related
181  * to recyling of free vnodes.  In the best-operating case, the cache is
182  * exactly full, the free list has size between vlowat and vhiwat above the
183  * free target, and recycling from it and normal use maintains this state.
184  * Sometimes the free list is below vlowat or even empty, but this state
185  * is even better for immediate use provided the cache is not full.
186  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
187  * ones) to reach one of these states.  The watermarks are currently hard-
188  * coded as 4% and 9% of the available space higher.  These and the default
189  * of 25% for wantfreevnodes are too large if the memory size is large.
190  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
191  * whenever vnlru_proc() becomes active.
192  */
193 static long wantfreevnodes;
194 static long __exclusive_cache_line freevnodes;
195 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
196     &freevnodes, 0, "Number of \"free\" vnodes");
197 static long freevnodes_old;
198 
199 static counter_u64_t recycles_count;
200 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
201     "Number of vnodes recycled to meet vnode cache targets");
202 
203 static counter_u64_t recycles_free_count;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
205     "Number of free vnodes recycled to meet vnode cache targets");
206 
207 static counter_u64_t deferred_inact;
208 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
209     "Number of times inactive processing was deferred");
210 
211 /* To keep more than one thread at a time from running vfs_getnewfsid */
212 static struct mtx mntid_mtx;
213 
214 /*
215  * Lock for any access to the following:
216  *	vnode_list
217  *	numvnodes
218  *	freevnodes
219  */
220 static struct mtx __exclusive_cache_line vnode_list_mtx;
221 
222 /* Publicly exported FS */
223 struct nfs_public nfs_pub;
224 
225 static uma_zone_t buf_trie_zone;
226 static smr_t buf_trie_smr;
227 
228 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
229 static uma_zone_t vnode_zone;
230 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
231 
232 __read_frequently smr_t vfs_smr;
233 
234 /*
235  * The workitem queue.
236  *
237  * It is useful to delay writes of file data and filesystem metadata
238  * for tens of seconds so that quickly created and deleted files need
239  * not waste disk bandwidth being created and removed. To realize this,
240  * we append vnodes to a "workitem" queue. When running with a soft
241  * updates implementation, most pending metadata dependencies should
242  * not wait for more than a few seconds. Thus, mounted on block devices
243  * are delayed only about a half the time that file data is delayed.
244  * Similarly, directory updates are more critical, so are only delayed
245  * about a third the time that file data is delayed. Thus, there are
246  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
247  * one each second (driven off the filesystem syncer process). The
248  * syncer_delayno variable indicates the next queue that is to be processed.
249  * Items that need to be processed soon are placed in this queue:
250  *
251  *	syncer_workitem_pending[syncer_delayno]
252  *
253  * A delay of fifteen seconds is done by placing the request fifteen
254  * entries later in the queue:
255  *
256  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
257  *
258  */
259 static int syncer_delayno;
260 static long syncer_mask;
261 LIST_HEAD(synclist, bufobj);
262 static struct synclist *syncer_workitem_pending;
263 /*
264  * The sync_mtx protects:
265  *	bo->bo_synclist
266  *	sync_vnode_count
267  *	syncer_delayno
268  *	syncer_state
269  *	syncer_workitem_pending
270  *	syncer_worklist_len
271  *	rushjob
272  */
273 static struct mtx sync_mtx;
274 static struct cv sync_wakeup;
275 
276 #define SYNCER_MAXDELAY		32
277 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
278 static int syncdelay = 30;		/* max time to delay syncing data */
279 static int filedelay = 30;		/* time to delay syncing files */
280 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
281     "Time to delay syncing files (in seconds)");
282 static int dirdelay = 29;		/* time to delay syncing directories */
283 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
284     "Time to delay syncing directories (in seconds)");
285 static int metadelay = 28;		/* time to delay syncing metadata */
286 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
287     "Time to delay syncing metadata (in seconds)");
288 static int rushjob;		/* number of slots to run ASAP */
289 static int stat_rush_requests;	/* number of times I/O speeded up */
290 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
291     "Number of times I/O speeded up (rush requests)");
292 
293 #define	VDBATCH_SIZE 8
294 struct vdbatch {
295 	u_int index;
296 	long freevnodes;
297 	struct mtx lock;
298 	struct vnode *tab[VDBATCH_SIZE];
299 };
300 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
301 
302 static void	vdbatch_dequeue(struct vnode *vp);
303 
304 /*
305  * When shutting down the syncer, run it at four times normal speed.
306  */
307 #define SYNCER_SHUTDOWN_SPEEDUP		4
308 static int sync_vnode_count;
309 static int syncer_worklist_len;
310 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
311     syncer_state;
312 
313 /* Target for maximum number of vnodes. */
314 u_long desiredvnodes;
315 static u_long gapvnodes;		/* gap between wanted and desired */
316 static u_long vhiwat;		/* enough extras after expansion */
317 static u_long vlowat;		/* minimal extras before expansion */
318 static u_long vstir;		/* nonzero to stir non-free vnodes */
319 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
320 
321 static u_long vnlru_read_freevnodes(void);
322 
323 /*
324  * Note that no attempt is made to sanitize these parameters.
325  */
326 static int
327 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
328 {
329 	u_long val;
330 	int error;
331 
332 	val = desiredvnodes;
333 	error = sysctl_handle_long(oidp, &val, 0, req);
334 	if (error != 0 || req->newptr == NULL)
335 		return (error);
336 
337 	if (val == desiredvnodes)
338 		return (0);
339 	mtx_lock(&vnode_list_mtx);
340 	desiredvnodes = val;
341 	wantfreevnodes = desiredvnodes / 4;
342 	vnlru_recalc();
343 	mtx_unlock(&vnode_list_mtx);
344 	/*
345 	 * XXX There is no protection against multiple threads changing
346 	 * desiredvnodes at the same time. Locking above only helps vnlru and
347 	 * getnewvnode.
348 	 */
349 	vfs_hash_changesize(desiredvnodes);
350 	cache_changesize(desiredvnodes);
351 	return (0);
352 }
353 
354 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
355     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
356     "LU", "Target for maximum number of vnodes");
357 
358 static int
359 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
360 {
361 	u_long val;
362 	int error;
363 
364 	val = wantfreevnodes;
365 	error = sysctl_handle_long(oidp, &val, 0, req);
366 	if (error != 0 || req->newptr == NULL)
367 		return (error);
368 
369 	if (val == wantfreevnodes)
370 		return (0);
371 	mtx_lock(&vnode_list_mtx);
372 	wantfreevnodes = val;
373 	vnlru_recalc();
374 	mtx_unlock(&vnode_list_mtx);
375 	return (0);
376 }
377 
378 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
379     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
380     "LU", "Target for minimum number of \"free\" vnodes");
381 
382 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
383     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
384 static int vnlru_nowhere;
385 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
386     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
387 
388 static int
389 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
390 {
391 	struct vnode *vp;
392 	struct nameidata nd;
393 	char *buf;
394 	unsigned long ndflags;
395 	int error;
396 
397 	if (req->newptr == NULL)
398 		return (EINVAL);
399 	if (req->newlen >= PATH_MAX)
400 		return (E2BIG);
401 
402 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
403 	error = SYSCTL_IN(req, buf, req->newlen);
404 	if (error != 0)
405 		goto out;
406 
407 	buf[req->newlen] = '\0';
408 
409 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
410 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
411 	if ((error = namei(&nd)) != 0)
412 		goto out;
413 	vp = nd.ni_vp;
414 
415 	if (VN_IS_DOOMED(vp)) {
416 		/*
417 		 * This vnode is being recycled.  Return != 0 to let the caller
418 		 * know that the sysctl had no effect.  Return EAGAIN because a
419 		 * subsequent call will likely succeed (since namei will create
420 		 * a new vnode if necessary)
421 		 */
422 		error = EAGAIN;
423 		goto putvnode;
424 	}
425 
426 	counter_u64_add(recycles_count, 1);
427 	vgone(vp);
428 putvnode:
429 	NDFREE(&nd, 0);
430 out:
431 	free(buf, M_TEMP);
432 	return (error);
433 }
434 
435 static int
436 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
437 {
438 	struct thread *td = curthread;
439 	struct vnode *vp;
440 	struct file *fp;
441 	int error;
442 	int fd;
443 
444 	if (req->newptr == NULL)
445 		return (EBADF);
446 
447         error = sysctl_handle_int(oidp, &fd, 0, req);
448         if (error != 0)
449                 return (error);
450 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
451 	if (error != 0)
452 		return (error);
453 	vp = fp->f_vnode;
454 
455 	error = vn_lock(vp, LK_EXCLUSIVE);
456 	if (error != 0)
457 		goto drop;
458 
459 	counter_u64_add(recycles_count, 1);
460 	vgone(vp);
461 	VOP_UNLOCK(vp);
462 drop:
463 	fdrop(fp, td);
464 	return (error);
465 }
466 
467 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
468     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
469     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
470 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
471     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
472     sysctl_ftry_reclaim_vnode, "I",
473     "Try to reclaim a vnode by its file descriptor");
474 
475 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
476 static int vnsz2log;
477 
478 /*
479  * Support for the bufobj clean & dirty pctrie.
480  */
481 static void *
482 buf_trie_alloc(struct pctrie *ptree)
483 {
484 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
485 }
486 
487 static void
488 buf_trie_free(struct pctrie *ptree, void *node)
489 {
490 	uma_zfree_smr(buf_trie_zone, node);
491 }
492 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
493     buf_trie_smr);
494 
495 /*
496  * Initialize the vnode management data structures.
497  *
498  * Reevaluate the following cap on the number of vnodes after the physical
499  * memory size exceeds 512GB.  In the limit, as the physical memory size
500  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
501  */
502 #ifndef	MAXVNODES_MAX
503 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
504 #endif
505 
506 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
507 
508 static struct vnode *
509 vn_alloc_marker(struct mount *mp)
510 {
511 	struct vnode *vp;
512 
513 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
514 	vp->v_type = VMARKER;
515 	vp->v_mount = mp;
516 
517 	return (vp);
518 }
519 
520 static void
521 vn_free_marker(struct vnode *vp)
522 {
523 
524 	MPASS(vp->v_type == VMARKER);
525 	free(vp, M_VNODE_MARKER);
526 }
527 
528 /*
529  * Initialize a vnode as it first enters the zone.
530  */
531 static int
532 vnode_init(void *mem, int size, int flags)
533 {
534 	struct vnode *vp;
535 
536 	vp = mem;
537 	bzero(vp, size);
538 	/*
539 	 * Setup locks.
540 	 */
541 	vp->v_vnlock = &vp->v_lock;
542 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
543 	/*
544 	 * By default, don't allow shared locks unless filesystems opt-in.
545 	 */
546 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
547 	    LK_NOSHARE | LK_IS_VNODE);
548 	/*
549 	 * Initialize bufobj.
550 	 */
551 	bufobj_init(&vp->v_bufobj, vp);
552 	/*
553 	 * Initialize namecache.
554 	 */
555 	cache_vnode_init(vp);
556 	/*
557 	 * Initialize rangelocks.
558 	 */
559 	rangelock_init(&vp->v_rl);
560 
561 	vp->v_dbatchcpu = NOCPU;
562 
563 	/*
564 	 * Check vhold_recycle_free for an explanation.
565 	 */
566 	vp->v_holdcnt = VHOLD_NO_SMR;
567 	vp->v_type = VNON;
568 	mtx_lock(&vnode_list_mtx);
569 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
570 	mtx_unlock(&vnode_list_mtx);
571 	return (0);
572 }
573 
574 /*
575  * Free a vnode when it is cleared from the zone.
576  */
577 static void
578 vnode_fini(void *mem, int size)
579 {
580 	struct vnode *vp;
581 	struct bufobj *bo;
582 
583 	vp = mem;
584 	vdbatch_dequeue(vp);
585 	mtx_lock(&vnode_list_mtx);
586 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
587 	mtx_unlock(&vnode_list_mtx);
588 	rangelock_destroy(&vp->v_rl);
589 	lockdestroy(vp->v_vnlock);
590 	mtx_destroy(&vp->v_interlock);
591 	bo = &vp->v_bufobj;
592 	rw_destroy(BO_LOCKPTR(bo));
593 }
594 
595 /*
596  * Provide the size of NFS nclnode and NFS fh for calculation of the
597  * vnode memory consumption.  The size is specified directly to
598  * eliminate dependency on NFS-private header.
599  *
600  * Other filesystems may use bigger or smaller (like UFS and ZFS)
601  * private inode data, but the NFS-based estimation is ample enough.
602  * Still, we care about differences in the size between 64- and 32-bit
603  * platforms.
604  *
605  * Namecache structure size is heuristically
606  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
607  */
608 #ifdef _LP64
609 #define	NFS_NCLNODE_SZ	(528 + 64)
610 #define	NC_SZ		148
611 #else
612 #define	NFS_NCLNODE_SZ	(360 + 32)
613 #define	NC_SZ		92
614 #endif
615 
616 static void
617 vntblinit(void *dummy __unused)
618 {
619 	struct vdbatch *vd;
620 	int cpu, physvnodes, virtvnodes;
621 	u_int i;
622 
623 	/*
624 	 * Desiredvnodes is a function of the physical memory size and the
625 	 * kernel's heap size.  Generally speaking, it scales with the
626 	 * physical memory size.  The ratio of desiredvnodes to the physical
627 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
628 	 * Thereafter, the
629 	 * marginal ratio of desiredvnodes to the physical memory size is
630 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
631 	 * size.  The memory required by desiredvnodes vnodes and vm objects
632 	 * must not exceed 1/10th of the kernel's heap size.
633 	 */
634 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
635 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
636 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
637 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
638 	desiredvnodes = min(physvnodes, virtvnodes);
639 	if (desiredvnodes > MAXVNODES_MAX) {
640 		if (bootverbose)
641 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
642 			    desiredvnodes, MAXVNODES_MAX);
643 		desiredvnodes = MAXVNODES_MAX;
644 	}
645 	wantfreevnodes = desiredvnodes / 4;
646 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
647 	TAILQ_INIT(&vnode_list);
648 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
649 	/*
650 	 * The lock is taken to appease WITNESS.
651 	 */
652 	mtx_lock(&vnode_list_mtx);
653 	vnlru_recalc();
654 	mtx_unlock(&vnode_list_mtx);
655 	vnode_list_free_marker = vn_alloc_marker(NULL);
656 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
657 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
658 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
659 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
660 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
661 	uma_zone_set_smr(vnode_zone, vfs_smr);
662 	/*
663 	 * Preallocate enough nodes to support one-per buf so that
664 	 * we can not fail an insert.  reassignbuf() callers can not
665 	 * tolerate the insertion failure.
666 	 */
667 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
668 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
669 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
670 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
671 	uma_prealloc(buf_trie_zone, nbuf);
672 
673 	vnodes_created = counter_u64_alloc(M_WAITOK);
674 	recycles_count = counter_u64_alloc(M_WAITOK);
675 	recycles_free_count = counter_u64_alloc(M_WAITOK);
676 	deferred_inact = counter_u64_alloc(M_WAITOK);
677 
678 	/*
679 	 * Initialize the filesystem syncer.
680 	 */
681 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
682 	    &syncer_mask);
683 	syncer_maxdelay = syncer_mask + 1;
684 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
685 	cv_init(&sync_wakeup, "syncer");
686 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
687 		vnsz2log++;
688 	vnsz2log--;
689 
690 	CPU_FOREACH(cpu) {
691 		vd = DPCPU_ID_PTR((cpu), vd);
692 		bzero(vd, sizeof(*vd));
693 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
694 	}
695 }
696 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
697 
698 /*
699  * Mark a mount point as busy. Used to synchronize access and to delay
700  * unmounting. Eventually, mountlist_mtx is not released on failure.
701  *
702  * vfs_busy() is a custom lock, it can block the caller.
703  * vfs_busy() only sleeps if the unmount is active on the mount point.
704  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
705  * vnode belonging to mp.
706  *
707  * Lookup uses vfs_busy() to traverse mount points.
708  * root fs			var fs
709  * / vnode lock		A	/ vnode lock (/var)		D
710  * /var vnode lock	B	/log vnode lock(/var/log)	E
711  * vfs_busy lock	C	vfs_busy lock			F
712  *
713  * Within each file system, the lock order is C->A->B and F->D->E.
714  *
715  * When traversing across mounts, the system follows that lock order:
716  *
717  *        C->A->B
718  *              |
719  *              +->F->D->E
720  *
721  * The lookup() process for namei("/var") illustrates the process:
722  *  VOP_LOOKUP() obtains B while A is held
723  *  vfs_busy() obtains a shared lock on F while A and B are held
724  *  vput() releases lock on B
725  *  vput() releases lock on A
726  *  VFS_ROOT() obtains lock on D while shared lock on F is held
727  *  vfs_unbusy() releases shared lock on F
728  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
729  *    Attempt to lock A (instead of vp_crossmp) while D is held would
730  *    violate the global order, causing deadlocks.
731  *
732  * dounmount() locks B while F is drained.
733  */
734 int
735 vfs_busy(struct mount *mp, int flags)
736 {
737 	struct mount_pcpu *mpcpu;
738 
739 	MPASS((flags & ~MBF_MASK) == 0);
740 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
741 
742 	if (vfs_op_thread_enter(mp, mpcpu)) {
743 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
744 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
745 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
746 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
747 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
748 		vfs_op_thread_exit(mp, mpcpu);
749 		if (flags & MBF_MNTLSTLOCK)
750 			mtx_unlock(&mountlist_mtx);
751 		return (0);
752 	}
753 
754 	MNT_ILOCK(mp);
755 	vfs_assert_mount_counters(mp);
756 	MNT_REF(mp);
757 	/*
758 	 * If mount point is currently being unmounted, sleep until the
759 	 * mount point fate is decided.  If thread doing the unmounting fails,
760 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
761 	 * that this mount point has survived the unmount attempt and vfs_busy
762 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
763 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
764 	 * about to be really destroyed.  vfs_busy needs to release its
765 	 * reference on the mount point in this case and return with ENOENT,
766 	 * telling the caller that mount mount it tried to busy is no longer
767 	 * valid.
768 	 */
769 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
770 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
771 			MNT_REL(mp);
772 			MNT_IUNLOCK(mp);
773 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
774 			    __func__);
775 			return (ENOENT);
776 		}
777 		if (flags & MBF_MNTLSTLOCK)
778 			mtx_unlock(&mountlist_mtx);
779 		mp->mnt_kern_flag |= MNTK_MWAIT;
780 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
781 		if (flags & MBF_MNTLSTLOCK)
782 			mtx_lock(&mountlist_mtx);
783 		MNT_ILOCK(mp);
784 	}
785 	if (flags & MBF_MNTLSTLOCK)
786 		mtx_unlock(&mountlist_mtx);
787 	mp->mnt_lockref++;
788 	MNT_IUNLOCK(mp);
789 	return (0);
790 }
791 
792 /*
793  * Free a busy filesystem.
794  */
795 void
796 vfs_unbusy(struct mount *mp)
797 {
798 	struct mount_pcpu *mpcpu;
799 	int c;
800 
801 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
802 
803 	if (vfs_op_thread_enter(mp, mpcpu)) {
804 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
805 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
806 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
807 		vfs_op_thread_exit(mp, mpcpu);
808 		return;
809 	}
810 
811 	MNT_ILOCK(mp);
812 	vfs_assert_mount_counters(mp);
813 	MNT_REL(mp);
814 	c = --mp->mnt_lockref;
815 	if (mp->mnt_vfs_ops == 0) {
816 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
817 		MNT_IUNLOCK(mp);
818 		return;
819 	}
820 	if (c < 0)
821 		vfs_dump_mount_counters(mp);
822 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
823 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
824 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
825 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
826 		wakeup(&mp->mnt_lockref);
827 	}
828 	MNT_IUNLOCK(mp);
829 }
830 
831 /*
832  * Lookup a mount point by filesystem identifier.
833  */
834 struct mount *
835 vfs_getvfs(fsid_t *fsid)
836 {
837 	struct mount *mp;
838 
839 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
840 	mtx_lock(&mountlist_mtx);
841 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
842 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
843 			vfs_ref(mp);
844 			mtx_unlock(&mountlist_mtx);
845 			return (mp);
846 		}
847 	}
848 	mtx_unlock(&mountlist_mtx);
849 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
850 	return ((struct mount *) 0);
851 }
852 
853 /*
854  * Lookup a mount point by filesystem identifier, busying it before
855  * returning.
856  *
857  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
858  * cache for popular filesystem identifiers.  The cache is lockess, using
859  * the fact that struct mount's are never freed.  In worst case we may
860  * get pointer to unmounted or even different filesystem, so we have to
861  * check what we got, and go slow way if so.
862  */
863 struct mount *
864 vfs_busyfs(fsid_t *fsid)
865 {
866 #define	FSID_CACHE_SIZE	256
867 	typedef struct mount * volatile vmp_t;
868 	static vmp_t cache[FSID_CACHE_SIZE];
869 	struct mount *mp;
870 	int error;
871 	uint32_t hash;
872 
873 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
874 	hash = fsid->val[0] ^ fsid->val[1];
875 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
876 	mp = cache[hash];
877 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
878 		goto slow;
879 	if (vfs_busy(mp, 0) != 0) {
880 		cache[hash] = NULL;
881 		goto slow;
882 	}
883 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
884 		return (mp);
885 	else
886 	    vfs_unbusy(mp);
887 
888 slow:
889 	mtx_lock(&mountlist_mtx);
890 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
891 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
892 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
893 			if (error) {
894 				cache[hash] = NULL;
895 				mtx_unlock(&mountlist_mtx);
896 				return (NULL);
897 			}
898 			cache[hash] = mp;
899 			return (mp);
900 		}
901 	}
902 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
903 	mtx_unlock(&mountlist_mtx);
904 	return ((struct mount *) 0);
905 }
906 
907 /*
908  * Check if a user can access privileged mount options.
909  */
910 int
911 vfs_suser(struct mount *mp, struct thread *td)
912 {
913 	int error;
914 
915 	if (jailed(td->td_ucred)) {
916 		/*
917 		 * If the jail of the calling thread lacks permission for
918 		 * this type of file system, deny immediately.
919 		 */
920 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
921 			return (EPERM);
922 
923 		/*
924 		 * If the file system was mounted outside the jail of the
925 		 * calling thread, deny immediately.
926 		 */
927 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
928 			return (EPERM);
929 	}
930 
931 	/*
932 	 * If file system supports delegated administration, we don't check
933 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
934 	 * by the file system itself.
935 	 * If this is not the user that did original mount, we check for
936 	 * the PRIV_VFS_MOUNT_OWNER privilege.
937 	 */
938 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
939 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
940 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
941 			return (error);
942 	}
943 	return (0);
944 }
945 
946 /*
947  * Get a new unique fsid.  Try to make its val[0] unique, since this value
948  * will be used to create fake device numbers for stat().  Also try (but
949  * not so hard) make its val[0] unique mod 2^16, since some emulators only
950  * support 16-bit device numbers.  We end up with unique val[0]'s for the
951  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
952  *
953  * Keep in mind that several mounts may be running in parallel.  Starting
954  * the search one past where the previous search terminated is both a
955  * micro-optimization and a defense against returning the same fsid to
956  * different mounts.
957  */
958 void
959 vfs_getnewfsid(struct mount *mp)
960 {
961 	static uint16_t mntid_base;
962 	struct mount *nmp;
963 	fsid_t tfsid;
964 	int mtype;
965 
966 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
967 	mtx_lock(&mntid_mtx);
968 	mtype = mp->mnt_vfc->vfc_typenum;
969 	tfsid.val[1] = mtype;
970 	mtype = (mtype & 0xFF) << 24;
971 	for (;;) {
972 		tfsid.val[0] = makedev(255,
973 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
974 		mntid_base++;
975 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
976 			break;
977 		vfs_rel(nmp);
978 	}
979 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
980 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
981 	mtx_unlock(&mntid_mtx);
982 }
983 
984 /*
985  * Knob to control the precision of file timestamps:
986  *
987  *   0 = seconds only; nanoseconds zeroed.
988  *   1 = seconds and nanoseconds, accurate within 1/HZ.
989  *   2 = seconds and nanoseconds, truncated to microseconds.
990  * >=3 = seconds and nanoseconds, maximum precision.
991  */
992 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
993 
994 static int timestamp_precision = TSP_USEC;
995 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
996     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
997     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
998     "3+: sec + ns (max. precision))");
999 
1000 /*
1001  * Get a current timestamp.
1002  */
1003 void
1004 vfs_timestamp(struct timespec *tsp)
1005 {
1006 	struct timeval tv;
1007 
1008 	switch (timestamp_precision) {
1009 	case TSP_SEC:
1010 		tsp->tv_sec = time_second;
1011 		tsp->tv_nsec = 0;
1012 		break;
1013 	case TSP_HZ:
1014 		getnanotime(tsp);
1015 		break;
1016 	case TSP_USEC:
1017 		microtime(&tv);
1018 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1019 		break;
1020 	case TSP_NSEC:
1021 	default:
1022 		nanotime(tsp);
1023 		break;
1024 	}
1025 }
1026 
1027 /*
1028  * Set vnode attributes to VNOVAL
1029  */
1030 void
1031 vattr_null(struct vattr *vap)
1032 {
1033 
1034 	vap->va_type = VNON;
1035 	vap->va_size = VNOVAL;
1036 	vap->va_bytes = VNOVAL;
1037 	vap->va_mode = VNOVAL;
1038 	vap->va_nlink = VNOVAL;
1039 	vap->va_uid = VNOVAL;
1040 	vap->va_gid = VNOVAL;
1041 	vap->va_fsid = VNOVAL;
1042 	vap->va_fileid = VNOVAL;
1043 	vap->va_blocksize = VNOVAL;
1044 	vap->va_rdev = VNOVAL;
1045 	vap->va_atime.tv_sec = VNOVAL;
1046 	vap->va_atime.tv_nsec = VNOVAL;
1047 	vap->va_mtime.tv_sec = VNOVAL;
1048 	vap->va_mtime.tv_nsec = VNOVAL;
1049 	vap->va_ctime.tv_sec = VNOVAL;
1050 	vap->va_ctime.tv_nsec = VNOVAL;
1051 	vap->va_birthtime.tv_sec = VNOVAL;
1052 	vap->va_birthtime.tv_nsec = VNOVAL;
1053 	vap->va_flags = VNOVAL;
1054 	vap->va_gen = VNOVAL;
1055 	vap->va_vaflags = 0;
1056 }
1057 
1058 /*
1059  * Try to reduce the total number of vnodes.
1060  *
1061  * This routine (and its user) are buggy in at least the following ways:
1062  * - all parameters were picked years ago when RAM sizes were significantly
1063  *   smaller
1064  * - it can pick vnodes based on pages used by the vm object, but filesystems
1065  *   like ZFS don't use it making the pick broken
1066  * - since ZFS has its own aging policy it gets partially combated by this one
1067  * - a dedicated method should be provided for filesystems to let them decide
1068  *   whether the vnode should be recycled
1069  *
1070  * This routine is called when we have too many vnodes.  It attempts
1071  * to free <count> vnodes and will potentially free vnodes that still
1072  * have VM backing store (VM backing store is typically the cause
1073  * of a vnode blowout so we want to do this).  Therefore, this operation
1074  * is not considered cheap.
1075  *
1076  * A number of conditions may prevent a vnode from being reclaimed.
1077  * the buffer cache may have references on the vnode, a directory
1078  * vnode may still have references due to the namei cache representing
1079  * underlying files, or the vnode may be in active use.   It is not
1080  * desirable to reuse such vnodes.  These conditions may cause the
1081  * number of vnodes to reach some minimum value regardless of what
1082  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1083  *
1084  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1085  * 			 entries if this argument is strue
1086  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1087  *			 pages.
1088  * @param target	 How many vnodes to reclaim.
1089  * @return		 The number of vnodes that were reclaimed.
1090  */
1091 static int
1092 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1093 {
1094 	struct vnode *vp, *mvp;
1095 	struct mount *mp;
1096 	struct vm_object *object;
1097 	u_long done;
1098 	bool retried;
1099 
1100 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1101 
1102 	retried = false;
1103 	done = 0;
1104 
1105 	mvp = vnode_list_reclaim_marker;
1106 restart:
1107 	vp = mvp;
1108 	while (done < target) {
1109 		vp = TAILQ_NEXT(vp, v_vnodelist);
1110 		if (__predict_false(vp == NULL))
1111 			break;
1112 
1113 		if (__predict_false(vp->v_type == VMARKER))
1114 			continue;
1115 
1116 		/*
1117 		 * If it's been deconstructed already, it's still
1118 		 * referenced, or it exceeds the trigger, skip it.
1119 		 * Also skip free vnodes.  We are trying to make space
1120 		 * to expand the free list, not reduce it.
1121 		 */
1122 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1123 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1124 			goto next_iter;
1125 
1126 		if (vp->v_type == VBAD || vp->v_type == VNON)
1127 			goto next_iter;
1128 
1129 		object = atomic_load_ptr(&vp->v_object);
1130 		if (object == NULL || object->resident_page_count > trigger) {
1131 			goto next_iter;
1132 		}
1133 
1134 		/*
1135 		 * Handle races against vnode allocation. Filesystems lock the
1136 		 * vnode some time after it gets returned from getnewvnode,
1137 		 * despite type and hold count being manipulated earlier.
1138 		 * Resorting to checking v_mount restores guarantees present
1139 		 * before the global list was reworked to contain all vnodes.
1140 		 */
1141 		if (!VI_TRYLOCK(vp))
1142 			goto next_iter;
1143 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1144 			VI_UNLOCK(vp);
1145 			goto next_iter;
1146 		}
1147 		if (vp->v_mount == NULL) {
1148 			VI_UNLOCK(vp);
1149 			goto next_iter;
1150 		}
1151 		vholdl(vp);
1152 		VI_UNLOCK(vp);
1153 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1154 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1155 		mtx_unlock(&vnode_list_mtx);
1156 
1157 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1158 			vdrop(vp);
1159 			goto next_iter_unlocked;
1160 		}
1161 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1162 			vdrop(vp);
1163 			vn_finished_write(mp);
1164 			goto next_iter_unlocked;
1165 		}
1166 
1167 		VI_LOCK(vp);
1168 		if (vp->v_usecount > 0 ||
1169 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1170 		    (vp->v_object != NULL &&
1171 		    vp->v_object->resident_page_count > trigger)) {
1172 			VOP_UNLOCK(vp);
1173 			vdropl(vp);
1174 			vn_finished_write(mp);
1175 			goto next_iter_unlocked;
1176 		}
1177 		counter_u64_add(recycles_count, 1);
1178 		vgonel(vp);
1179 		VOP_UNLOCK(vp);
1180 		vdropl(vp);
1181 		vn_finished_write(mp);
1182 		done++;
1183 next_iter_unlocked:
1184 		if (should_yield())
1185 			kern_yield(PRI_USER);
1186 		mtx_lock(&vnode_list_mtx);
1187 		goto restart;
1188 next_iter:
1189 		MPASS(vp->v_type != VMARKER);
1190 		if (!should_yield())
1191 			continue;
1192 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1193 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1194 		mtx_unlock(&vnode_list_mtx);
1195 		kern_yield(PRI_USER);
1196 		mtx_lock(&vnode_list_mtx);
1197 		goto restart;
1198 	}
1199 	if (done == 0 && !retried) {
1200 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1201 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1202 		retried = true;
1203 		goto restart;
1204 	}
1205 	return (done);
1206 }
1207 
1208 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1209 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1210     0,
1211     "limit on vnode free requests per call to the vnlru_free routine");
1212 
1213 /*
1214  * Attempt to reduce the free list by the requested amount.
1215  */
1216 static int
1217 vnlru_free_locked(int count, struct vfsops *mnt_op)
1218 {
1219 	struct vnode *vp, *mvp;
1220 	struct mount *mp;
1221 	int ocount;
1222 
1223 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1224 	if (count > max_vnlru_free)
1225 		count = max_vnlru_free;
1226 	ocount = count;
1227 	mvp = vnode_list_free_marker;
1228 	vp = mvp;
1229 	for (;;) {
1230 		if (count == 0) {
1231 			break;
1232 		}
1233 		vp = TAILQ_NEXT(vp, v_vnodelist);
1234 		if (__predict_false(vp == NULL)) {
1235 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1236 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1237 			break;
1238 		}
1239 		if (__predict_false(vp->v_type == VMARKER))
1240 			continue;
1241 		if (vp->v_holdcnt > 0)
1242 			continue;
1243 		/*
1244 		 * Don't recycle if our vnode is from different type
1245 		 * of mount point.  Note that mp is type-safe, the
1246 		 * check does not reach unmapped address even if
1247 		 * vnode is reclaimed.
1248 		 */
1249 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1250 		    mp->mnt_op != mnt_op) {
1251 			continue;
1252 		}
1253 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1254 			continue;
1255 		}
1256 		if (!vhold_recycle_free(vp))
1257 			continue;
1258 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1259 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1260 		mtx_unlock(&vnode_list_mtx);
1261 		if (vtryrecycle(vp) == 0)
1262 			count--;
1263 		mtx_lock(&vnode_list_mtx);
1264 		vp = mvp;
1265 	}
1266 	return (ocount - count);
1267 }
1268 
1269 void
1270 vnlru_free(int count, struct vfsops *mnt_op)
1271 {
1272 
1273 	mtx_lock(&vnode_list_mtx);
1274 	vnlru_free_locked(count, mnt_op);
1275 	mtx_unlock(&vnode_list_mtx);
1276 }
1277 
1278 static void
1279 vnlru_recalc(void)
1280 {
1281 
1282 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1283 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1284 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1285 	vlowat = vhiwat / 2;
1286 }
1287 
1288 /*
1289  * Attempt to recycle vnodes in a context that is always safe to block.
1290  * Calling vlrurecycle() from the bowels of filesystem code has some
1291  * interesting deadlock problems.
1292  */
1293 static struct proc *vnlruproc;
1294 static int vnlruproc_sig;
1295 
1296 /*
1297  * The main freevnodes counter is only updated when threads requeue their vnode
1298  * batches. CPUs are conditionally walked to compute a more accurate total.
1299  *
1300  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1301  * at any given moment can still exceed slop, but it should not be by significant
1302  * margin in practice.
1303  */
1304 #define VNLRU_FREEVNODES_SLOP 128
1305 
1306 static __inline void
1307 vn_freevnodes_inc(void)
1308 {
1309 	struct vdbatch *vd;
1310 
1311 	critical_enter();
1312 	vd = DPCPU_PTR(vd);
1313 	vd->freevnodes++;
1314 	critical_exit();
1315 }
1316 
1317 static __inline void
1318 vn_freevnodes_dec(void)
1319 {
1320 	struct vdbatch *vd;
1321 
1322 	critical_enter();
1323 	vd = DPCPU_PTR(vd);
1324 	vd->freevnodes--;
1325 	critical_exit();
1326 }
1327 
1328 static u_long
1329 vnlru_read_freevnodes(void)
1330 {
1331 	struct vdbatch *vd;
1332 	long slop;
1333 	int cpu;
1334 
1335 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1336 	if (freevnodes > freevnodes_old)
1337 		slop = freevnodes - freevnodes_old;
1338 	else
1339 		slop = freevnodes_old - freevnodes;
1340 	if (slop < VNLRU_FREEVNODES_SLOP)
1341 		return (freevnodes >= 0 ? freevnodes : 0);
1342 	freevnodes_old = freevnodes;
1343 	CPU_FOREACH(cpu) {
1344 		vd = DPCPU_ID_PTR((cpu), vd);
1345 		freevnodes_old += vd->freevnodes;
1346 	}
1347 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1348 }
1349 
1350 static bool
1351 vnlru_under(u_long rnumvnodes, u_long limit)
1352 {
1353 	u_long rfreevnodes, space;
1354 
1355 	if (__predict_false(rnumvnodes > desiredvnodes))
1356 		return (true);
1357 
1358 	space = desiredvnodes - rnumvnodes;
1359 	if (space < limit) {
1360 		rfreevnodes = vnlru_read_freevnodes();
1361 		if (rfreevnodes > wantfreevnodes)
1362 			space += rfreevnodes - wantfreevnodes;
1363 	}
1364 	return (space < limit);
1365 }
1366 
1367 static bool
1368 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1369 {
1370 	long rfreevnodes, space;
1371 
1372 	if (__predict_false(rnumvnodes > desiredvnodes))
1373 		return (true);
1374 
1375 	space = desiredvnodes - rnumvnodes;
1376 	if (space < limit) {
1377 		rfreevnodes = atomic_load_long(&freevnodes);
1378 		if (rfreevnodes > wantfreevnodes)
1379 			space += rfreevnodes - wantfreevnodes;
1380 	}
1381 	return (space < limit);
1382 }
1383 
1384 static void
1385 vnlru_kick(void)
1386 {
1387 
1388 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1389 	if (vnlruproc_sig == 0) {
1390 		vnlruproc_sig = 1;
1391 		wakeup(vnlruproc);
1392 	}
1393 }
1394 
1395 static void
1396 vnlru_proc(void)
1397 {
1398 	u_long rnumvnodes, rfreevnodes, target;
1399 	unsigned long onumvnodes;
1400 	int done, force, trigger, usevnodes;
1401 	bool reclaim_nc_src, want_reread;
1402 
1403 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1404 	    SHUTDOWN_PRI_FIRST);
1405 
1406 	force = 0;
1407 	want_reread = false;
1408 	for (;;) {
1409 		kproc_suspend_check(vnlruproc);
1410 		mtx_lock(&vnode_list_mtx);
1411 		rnumvnodes = atomic_load_long(&numvnodes);
1412 
1413 		if (want_reread) {
1414 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1415 			want_reread = false;
1416 		}
1417 
1418 		/*
1419 		 * If numvnodes is too large (due to desiredvnodes being
1420 		 * adjusted using its sysctl, or emergency growth), first
1421 		 * try to reduce it by discarding from the free list.
1422 		 */
1423 		if (rnumvnodes > desiredvnodes) {
1424 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1425 			rnumvnodes = atomic_load_long(&numvnodes);
1426 		}
1427 		/*
1428 		 * Sleep if the vnode cache is in a good state.  This is
1429 		 * when it is not over-full and has space for about a 4%
1430 		 * or 9% expansion (by growing its size or inexcessively
1431 		 * reducing its free list).  Otherwise, try to reclaim
1432 		 * space for a 10% expansion.
1433 		 */
1434 		if (vstir && force == 0) {
1435 			force = 1;
1436 			vstir = 0;
1437 		}
1438 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1439 			vnlruproc_sig = 0;
1440 			wakeup(&vnlruproc_sig);
1441 			msleep(vnlruproc, &vnode_list_mtx,
1442 			    PVFS|PDROP, "vlruwt", hz);
1443 			continue;
1444 		}
1445 		rfreevnodes = vnlru_read_freevnodes();
1446 
1447 		onumvnodes = rnumvnodes;
1448 		/*
1449 		 * Calculate parameters for recycling.  These are the same
1450 		 * throughout the loop to give some semblance of fairness.
1451 		 * The trigger point is to avoid recycling vnodes with lots
1452 		 * of resident pages.  We aren't trying to free memory; we
1453 		 * are trying to recycle or at least free vnodes.
1454 		 */
1455 		if (rnumvnodes <= desiredvnodes)
1456 			usevnodes = rnumvnodes - rfreevnodes;
1457 		else
1458 			usevnodes = rnumvnodes;
1459 		if (usevnodes <= 0)
1460 			usevnodes = 1;
1461 		/*
1462 		 * The trigger value is is chosen to give a conservatively
1463 		 * large value to ensure that it alone doesn't prevent
1464 		 * making progress.  The value can easily be so large that
1465 		 * it is effectively infinite in some congested and
1466 		 * misconfigured cases, and this is necessary.  Normally
1467 		 * it is about 8 to 100 (pages), which is quite large.
1468 		 */
1469 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1470 		if (force < 2)
1471 			trigger = vsmalltrigger;
1472 		reclaim_nc_src = force >= 3;
1473 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1474 		target = target / 10 + 1;
1475 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1476 		mtx_unlock(&vnode_list_mtx);
1477 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1478 			uma_reclaim(UMA_RECLAIM_DRAIN);
1479 		if (done == 0) {
1480 			if (force == 0 || force == 1) {
1481 				force = 2;
1482 				continue;
1483 			}
1484 			if (force == 2) {
1485 				force = 3;
1486 				continue;
1487 			}
1488 			want_reread = true;
1489 			force = 0;
1490 			vnlru_nowhere++;
1491 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1492 		} else {
1493 			want_reread = true;
1494 			kern_yield(PRI_USER);
1495 		}
1496 	}
1497 }
1498 
1499 static struct kproc_desc vnlru_kp = {
1500 	"vnlru",
1501 	vnlru_proc,
1502 	&vnlruproc
1503 };
1504 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1505     &vnlru_kp);
1506 
1507 /*
1508  * Routines having to do with the management of the vnode table.
1509  */
1510 
1511 /*
1512  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1513  * before we actually vgone().  This function must be called with the vnode
1514  * held to prevent the vnode from being returned to the free list midway
1515  * through vgone().
1516  */
1517 static int
1518 vtryrecycle(struct vnode *vp)
1519 {
1520 	struct mount *vnmp;
1521 
1522 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1523 	VNASSERT(vp->v_holdcnt, vp,
1524 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1525 	/*
1526 	 * This vnode may found and locked via some other list, if so we
1527 	 * can't recycle it yet.
1528 	 */
1529 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1530 		CTR2(KTR_VFS,
1531 		    "%s: impossible to recycle, vp %p lock is already held",
1532 		    __func__, vp);
1533 		vdrop(vp);
1534 		return (EWOULDBLOCK);
1535 	}
1536 	/*
1537 	 * Don't recycle if its filesystem is being suspended.
1538 	 */
1539 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1540 		VOP_UNLOCK(vp);
1541 		CTR2(KTR_VFS,
1542 		    "%s: impossible to recycle, cannot start the write for %p",
1543 		    __func__, vp);
1544 		vdrop(vp);
1545 		return (EBUSY);
1546 	}
1547 	/*
1548 	 * If we got this far, we need to acquire the interlock and see if
1549 	 * anyone picked up this vnode from another list.  If not, we will
1550 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1551 	 * will skip over it.
1552 	 */
1553 	VI_LOCK(vp);
1554 	if (vp->v_usecount) {
1555 		VOP_UNLOCK(vp);
1556 		vdropl(vp);
1557 		vn_finished_write(vnmp);
1558 		CTR2(KTR_VFS,
1559 		    "%s: impossible to recycle, %p is already referenced",
1560 		    __func__, vp);
1561 		return (EBUSY);
1562 	}
1563 	if (!VN_IS_DOOMED(vp)) {
1564 		counter_u64_add(recycles_free_count, 1);
1565 		vgonel(vp);
1566 	}
1567 	VOP_UNLOCK(vp);
1568 	vdropl(vp);
1569 	vn_finished_write(vnmp);
1570 	return (0);
1571 }
1572 
1573 /*
1574  * Allocate a new vnode.
1575  *
1576  * The operation never returns an error. Returning an error was disabled
1577  * in r145385 (dated 2005) with the following comment:
1578  *
1579  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1580  *
1581  * Given the age of this commit (almost 15 years at the time of writing this
1582  * comment) restoring the ability to fail requires a significant audit of
1583  * all codepaths.
1584  *
1585  * The routine can try to free a vnode or stall for up to 1 second waiting for
1586  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1587  */
1588 static u_long vn_alloc_cyclecount;
1589 
1590 static struct vnode * __noinline
1591 vn_alloc_hard(struct mount *mp)
1592 {
1593 	u_long rnumvnodes, rfreevnodes;
1594 
1595 	mtx_lock(&vnode_list_mtx);
1596 	rnumvnodes = atomic_load_long(&numvnodes);
1597 	if (rnumvnodes + 1 < desiredvnodes) {
1598 		vn_alloc_cyclecount = 0;
1599 		goto alloc;
1600 	}
1601 	rfreevnodes = vnlru_read_freevnodes();
1602 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1603 		vn_alloc_cyclecount = 0;
1604 		vstir = 1;
1605 	}
1606 	/*
1607 	 * Grow the vnode cache if it will not be above its target max
1608 	 * after growing.  Otherwise, if the free list is nonempty, try
1609 	 * to reclaim 1 item from it before growing the cache (possibly
1610 	 * above its target max if the reclamation failed or is delayed).
1611 	 * Otherwise, wait for some space.  In all cases, schedule
1612 	 * vnlru_proc() if we are getting short of space.  The watermarks
1613 	 * should be chosen so that we never wait or even reclaim from
1614 	 * the free list to below its target minimum.
1615 	 */
1616 	if (vnlru_free_locked(1, NULL) > 0)
1617 		goto alloc;
1618 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1619 		/*
1620 		 * Wait for space for a new vnode.
1621 		 */
1622 		vnlru_kick();
1623 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1624 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1625 		    vnlru_read_freevnodes() > 1)
1626 			vnlru_free_locked(1, NULL);
1627 	}
1628 alloc:
1629 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1630 	if (vnlru_under(rnumvnodes, vlowat))
1631 		vnlru_kick();
1632 	mtx_unlock(&vnode_list_mtx);
1633 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1634 }
1635 
1636 static struct vnode *
1637 vn_alloc(struct mount *mp)
1638 {
1639 	u_long rnumvnodes;
1640 
1641 	if (__predict_false(vn_alloc_cyclecount != 0))
1642 		return (vn_alloc_hard(mp));
1643 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1644 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1645 		atomic_subtract_long(&numvnodes, 1);
1646 		return (vn_alloc_hard(mp));
1647 	}
1648 
1649 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1650 }
1651 
1652 static void
1653 vn_free(struct vnode *vp)
1654 {
1655 
1656 	atomic_subtract_long(&numvnodes, 1);
1657 	uma_zfree_smr(vnode_zone, vp);
1658 }
1659 
1660 /*
1661  * Return the next vnode from the free list.
1662  */
1663 int
1664 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1665     struct vnode **vpp)
1666 {
1667 	struct vnode *vp;
1668 	struct thread *td;
1669 	struct lock_object *lo;
1670 
1671 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1672 
1673 	KASSERT(vops->registered,
1674 	    ("%s: not registered vector op %p\n", __func__, vops));
1675 
1676 	td = curthread;
1677 	if (td->td_vp_reserved != NULL) {
1678 		vp = td->td_vp_reserved;
1679 		td->td_vp_reserved = NULL;
1680 	} else {
1681 		vp = vn_alloc(mp);
1682 	}
1683 	counter_u64_add(vnodes_created, 1);
1684 	/*
1685 	 * Locks are given the generic name "vnode" when created.
1686 	 * Follow the historic practice of using the filesystem
1687 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1688 	 *
1689 	 * Locks live in a witness group keyed on their name. Thus,
1690 	 * when a lock is renamed, it must also move from the witness
1691 	 * group of its old name to the witness group of its new name.
1692 	 *
1693 	 * The change only needs to be made when the vnode moves
1694 	 * from one filesystem type to another. We ensure that each
1695 	 * filesystem use a single static name pointer for its tag so
1696 	 * that we can compare pointers rather than doing a strcmp().
1697 	 */
1698 	lo = &vp->v_vnlock->lock_object;
1699 #ifdef WITNESS
1700 	if (lo->lo_name != tag) {
1701 #endif
1702 		lo->lo_name = tag;
1703 #ifdef WITNESS
1704 		WITNESS_DESTROY(lo);
1705 		WITNESS_INIT(lo, tag);
1706 	}
1707 #endif
1708 	/*
1709 	 * By default, don't allow shared locks unless filesystems opt-in.
1710 	 */
1711 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1712 	/*
1713 	 * Finalize various vnode identity bits.
1714 	 */
1715 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1716 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1717 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1718 	vp->v_type = VNON;
1719 	vp->v_op = vops;
1720 	v_init_counters(vp);
1721 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1722 #ifdef DIAGNOSTIC
1723 	if (mp == NULL && vops != &dead_vnodeops)
1724 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1725 #endif
1726 #ifdef MAC
1727 	mac_vnode_init(vp);
1728 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1729 		mac_vnode_associate_singlelabel(mp, vp);
1730 #endif
1731 	if (mp != NULL) {
1732 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1733 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1734 			vp->v_vflag |= VV_NOKNOTE;
1735 	}
1736 
1737 	/*
1738 	 * For the filesystems which do not use vfs_hash_insert(),
1739 	 * still initialize v_hash to have vfs_hash_index() useful.
1740 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1741 	 * its own hashing.
1742 	 */
1743 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1744 
1745 	*vpp = vp;
1746 	return (0);
1747 }
1748 
1749 void
1750 getnewvnode_reserve(void)
1751 {
1752 	struct thread *td;
1753 
1754 	td = curthread;
1755 	MPASS(td->td_vp_reserved == NULL);
1756 	td->td_vp_reserved = vn_alloc(NULL);
1757 }
1758 
1759 void
1760 getnewvnode_drop_reserve(void)
1761 {
1762 	struct thread *td;
1763 
1764 	td = curthread;
1765 	if (td->td_vp_reserved != NULL) {
1766 		vn_free(td->td_vp_reserved);
1767 		td->td_vp_reserved = NULL;
1768 	}
1769 }
1770 
1771 static void __noinline
1772 freevnode(struct vnode *vp)
1773 {
1774 	struct bufobj *bo;
1775 
1776 	/*
1777 	 * The vnode has been marked for destruction, so free it.
1778 	 *
1779 	 * The vnode will be returned to the zone where it will
1780 	 * normally remain until it is needed for another vnode. We
1781 	 * need to cleanup (or verify that the cleanup has already
1782 	 * been done) any residual data left from its current use
1783 	 * so as not to contaminate the freshly allocated vnode.
1784 	 */
1785 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1786 	/*
1787 	 * Paired with vgone.
1788 	 */
1789 	vn_seqc_write_end_locked(vp);
1790 	VNPASS(vp->v_seqc_users == 0, vp);
1791 
1792 	bo = &vp->v_bufobj;
1793 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1794 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1795 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1796 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1797 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1798 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1799 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1800 	    ("clean blk trie not empty"));
1801 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1802 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1803 	    ("dirty blk trie not empty"));
1804 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1805 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1806 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1807 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1808 	    ("Dangling rangelock waiters"));
1809 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1810 	    ("Leaked inactivation"));
1811 	VI_UNLOCK(vp);
1812 #ifdef MAC
1813 	mac_vnode_destroy(vp);
1814 #endif
1815 	if (vp->v_pollinfo != NULL) {
1816 		destroy_vpollinfo(vp->v_pollinfo);
1817 		vp->v_pollinfo = NULL;
1818 	}
1819 	vp->v_mountedhere = NULL;
1820 	vp->v_unpcb = NULL;
1821 	vp->v_rdev = NULL;
1822 	vp->v_fifoinfo = NULL;
1823 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1824 	vp->v_irflag = 0;
1825 	vp->v_iflag = 0;
1826 	vp->v_vflag = 0;
1827 	bo->bo_flag = 0;
1828 	vn_free(vp);
1829 }
1830 
1831 /*
1832  * Delete from old mount point vnode list, if on one.
1833  */
1834 static void
1835 delmntque(struct vnode *vp)
1836 {
1837 	struct mount *mp;
1838 
1839 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1840 
1841 	mp = vp->v_mount;
1842 	if (mp == NULL)
1843 		return;
1844 	MNT_ILOCK(mp);
1845 	VI_LOCK(vp);
1846 	vp->v_mount = NULL;
1847 	VI_UNLOCK(vp);
1848 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1849 		("bad mount point vnode list size"));
1850 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1851 	mp->mnt_nvnodelistsize--;
1852 	MNT_REL(mp);
1853 	MNT_IUNLOCK(mp);
1854 }
1855 
1856 static void
1857 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1858 {
1859 
1860 	vp->v_data = NULL;
1861 	vp->v_op = &dead_vnodeops;
1862 	vgone(vp);
1863 	vput(vp);
1864 }
1865 
1866 /*
1867  * Insert into list of vnodes for the new mount point, if available.
1868  */
1869 int
1870 insmntque1(struct vnode *vp, struct mount *mp,
1871 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1872 {
1873 
1874 	KASSERT(vp->v_mount == NULL,
1875 		("insmntque: vnode already on per mount vnode list"));
1876 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1877 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1878 
1879 	/*
1880 	 * We acquire the vnode interlock early to ensure that the
1881 	 * vnode cannot be recycled by another process releasing a
1882 	 * holdcnt on it before we get it on both the vnode list
1883 	 * and the active vnode list. The mount mutex protects only
1884 	 * manipulation of the vnode list and the vnode freelist
1885 	 * mutex protects only manipulation of the active vnode list.
1886 	 * Hence the need to hold the vnode interlock throughout.
1887 	 */
1888 	MNT_ILOCK(mp);
1889 	VI_LOCK(vp);
1890 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1891 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1892 	    mp->mnt_nvnodelistsize == 0)) &&
1893 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1894 		VI_UNLOCK(vp);
1895 		MNT_IUNLOCK(mp);
1896 		if (dtr != NULL)
1897 			dtr(vp, dtr_arg);
1898 		return (EBUSY);
1899 	}
1900 	vp->v_mount = mp;
1901 	MNT_REF(mp);
1902 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1903 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1904 		("neg mount point vnode list size"));
1905 	mp->mnt_nvnodelistsize++;
1906 	VI_UNLOCK(vp);
1907 	MNT_IUNLOCK(mp);
1908 	return (0);
1909 }
1910 
1911 int
1912 insmntque(struct vnode *vp, struct mount *mp)
1913 {
1914 
1915 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1916 }
1917 
1918 /*
1919  * Flush out and invalidate all buffers associated with a bufobj
1920  * Called with the underlying object locked.
1921  */
1922 int
1923 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1924 {
1925 	int error;
1926 
1927 	BO_LOCK(bo);
1928 	if (flags & V_SAVE) {
1929 		error = bufobj_wwait(bo, slpflag, slptimeo);
1930 		if (error) {
1931 			BO_UNLOCK(bo);
1932 			return (error);
1933 		}
1934 		if (bo->bo_dirty.bv_cnt > 0) {
1935 			BO_UNLOCK(bo);
1936 			do {
1937 				error = BO_SYNC(bo, MNT_WAIT);
1938 			} while (error == ERELOOKUP);
1939 			if (error != 0)
1940 				return (error);
1941 			/*
1942 			 * XXX We could save a lock/unlock if this was only
1943 			 * enabled under INVARIANTS
1944 			 */
1945 			BO_LOCK(bo);
1946 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1947 				panic("vinvalbuf: dirty bufs");
1948 		}
1949 	}
1950 	/*
1951 	 * If you alter this loop please notice that interlock is dropped and
1952 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1953 	 * no race conditions occur from this.
1954 	 */
1955 	do {
1956 		error = flushbuflist(&bo->bo_clean,
1957 		    flags, bo, slpflag, slptimeo);
1958 		if (error == 0 && !(flags & V_CLEANONLY))
1959 			error = flushbuflist(&bo->bo_dirty,
1960 			    flags, bo, slpflag, slptimeo);
1961 		if (error != 0 && error != EAGAIN) {
1962 			BO_UNLOCK(bo);
1963 			return (error);
1964 		}
1965 	} while (error != 0);
1966 
1967 	/*
1968 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1969 	 * have write I/O in-progress but if there is a VM object then the
1970 	 * VM object can also have read-I/O in-progress.
1971 	 */
1972 	do {
1973 		bufobj_wwait(bo, 0, 0);
1974 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1975 			BO_UNLOCK(bo);
1976 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1977 			BO_LOCK(bo);
1978 		}
1979 	} while (bo->bo_numoutput > 0);
1980 	BO_UNLOCK(bo);
1981 
1982 	/*
1983 	 * Destroy the copy in the VM cache, too.
1984 	 */
1985 	if (bo->bo_object != NULL &&
1986 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1987 		VM_OBJECT_WLOCK(bo->bo_object);
1988 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1989 		    OBJPR_CLEANONLY : 0);
1990 		VM_OBJECT_WUNLOCK(bo->bo_object);
1991 	}
1992 
1993 #ifdef INVARIANTS
1994 	BO_LOCK(bo);
1995 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1996 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1997 	    bo->bo_clean.bv_cnt > 0))
1998 		panic("vinvalbuf: flush failed");
1999 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2000 	    bo->bo_dirty.bv_cnt > 0)
2001 		panic("vinvalbuf: flush dirty failed");
2002 	BO_UNLOCK(bo);
2003 #endif
2004 	return (0);
2005 }
2006 
2007 /*
2008  * Flush out and invalidate all buffers associated with a vnode.
2009  * Called with the underlying object locked.
2010  */
2011 int
2012 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2013 {
2014 
2015 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2016 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2017 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2018 		return (0);
2019 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2020 }
2021 
2022 /*
2023  * Flush out buffers on the specified list.
2024  *
2025  */
2026 static int
2027 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2028     int slptimeo)
2029 {
2030 	struct buf *bp, *nbp;
2031 	int retval, error;
2032 	daddr_t lblkno;
2033 	b_xflags_t xflags;
2034 
2035 	ASSERT_BO_WLOCKED(bo);
2036 
2037 	retval = 0;
2038 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2039 		/*
2040 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2041 		 * do not skip any buffers. If we are flushing only V_NORMAL
2042 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2043 		 * flushing only V_ALT buffers then skip buffers not marked
2044 		 * as BX_ALTDATA.
2045 		 */
2046 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2047 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2048 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2049 			continue;
2050 		}
2051 		if (nbp != NULL) {
2052 			lblkno = nbp->b_lblkno;
2053 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2054 		}
2055 		retval = EAGAIN;
2056 		error = BUF_TIMELOCK(bp,
2057 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2058 		    "flushbuf", slpflag, slptimeo);
2059 		if (error) {
2060 			BO_LOCK(bo);
2061 			return (error != ENOLCK ? error : EAGAIN);
2062 		}
2063 		KASSERT(bp->b_bufobj == bo,
2064 		    ("bp %p wrong b_bufobj %p should be %p",
2065 		    bp, bp->b_bufobj, bo));
2066 		/*
2067 		 * XXX Since there are no node locks for NFS, I
2068 		 * believe there is a slight chance that a delayed
2069 		 * write will occur while sleeping just above, so
2070 		 * check for it.
2071 		 */
2072 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2073 		    (flags & V_SAVE)) {
2074 			bremfree(bp);
2075 			bp->b_flags |= B_ASYNC;
2076 			bwrite(bp);
2077 			BO_LOCK(bo);
2078 			return (EAGAIN);	/* XXX: why not loop ? */
2079 		}
2080 		bremfree(bp);
2081 		bp->b_flags |= (B_INVAL | B_RELBUF);
2082 		bp->b_flags &= ~B_ASYNC;
2083 		brelse(bp);
2084 		BO_LOCK(bo);
2085 		if (nbp == NULL)
2086 			break;
2087 		nbp = gbincore(bo, lblkno);
2088 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2089 		    != xflags)
2090 			break;			/* nbp invalid */
2091 	}
2092 	return (retval);
2093 }
2094 
2095 int
2096 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2097 {
2098 	struct buf *bp;
2099 	int error;
2100 	daddr_t lblkno;
2101 
2102 	ASSERT_BO_LOCKED(bo);
2103 
2104 	for (lblkno = startn;;) {
2105 again:
2106 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2107 		if (bp == NULL || bp->b_lblkno >= endn ||
2108 		    bp->b_lblkno < startn)
2109 			break;
2110 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2111 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2112 		if (error != 0) {
2113 			BO_RLOCK(bo);
2114 			if (error == ENOLCK)
2115 				goto again;
2116 			return (error);
2117 		}
2118 		KASSERT(bp->b_bufobj == bo,
2119 		    ("bp %p wrong b_bufobj %p should be %p",
2120 		    bp, bp->b_bufobj, bo));
2121 		lblkno = bp->b_lblkno + 1;
2122 		if ((bp->b_flags & B_MANAGED) == 0)
2123 			bremfree(bp);
2124 		bp->b_flags |= B_RELBUF;
2125 		/*
2126 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2127 		 * pages backing each buffer in the range are unlikely to be
2128 		 * reused.  Dirty buffers will have the hint applied once
2129 		 * they've been written.
2130 		 */
2131 		if ((bp->b_flags & B_VMIO) != 0)
2132 			bp->b_flags |= B_NOREUSE;
2133 		brelse(bp);
2134 		BO_RLOCK(bo);
2135 	}
2136 	return (0);
2137 }
2138 
2139 /*
2140  * Truncate a file's buffer and pages to a specified length.  This
2141  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2142  * sync activity.
2143  */
2144 int
2145 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2146 {
2147 	struct buf *bp, *nbp;
2148 	struct bufobj *bo;
2149 	daddr_t startlbn;
2150 
2151 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2152 	    vp, blksize, (uintmax_t)length);
2153 
2154 	/*
2155 	 * Round up to the *next* lbn.
2156 	 */
2157 	startlbn = howmany(length, blksize);
2158 
2159 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2160 
2161 	bo = &vp->v_bufobj;
2162 restart_unlocked:
2163 	BO_LOCK(bo);
2164 
2165 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2166 		;
2167 
2168 	if (length > 0) {
2169 restartsync:
2170 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2171 			if (bp->b_lblkno > 0)
2172 				continue;
2173 			/*
2174 			 * Since we hold the vnode lock this should only
2175 			 * fail if we're racing with the buf daemon.
2176 			 */
2177 			if (BUF_LOCK(bp,
2178 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2179 			    BO_LOCKPTR(bo)) == ENOLCK)
2180 				goto restart_unlocked;
2181 
2182 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2183 			    ("buf(%p) on dirty queue without DELWRI", bp));
2184 
2185 			bremfree(bp);
2186 			bawrite(bp);
2187 			BO_LOCK(bo);
2188 			goto restartsync;
2189 		}
2190 	}
2191 
2192 	bufobj_wwait(bo, 0, 0);
2193 	BO_UNLOCK(bo);
2194 	vnode_pager_setsize(vp, length);
2195 
2196 	return (0);
2197 }
2198 
2199 /*
2200  * Invalidate the cached pages of a file's buffer within the range of block
2201  * numbers [startlbn, endlbn).
2202  */
2203 void
2204 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2205     int blksize)
2206 {
2207 	struct bufobj *bo;
2208 	off_t start, end;
2209 
2210 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2211 
2212 	start = blksize * startlbn;
2213 	end = blksize * endlbn;
2214 
2215 	bo = &vp->v_bufobj;
2216 	BO_LOCK(bo);
2217 	MPASS(blksize == bo->bo_bsize);
2218 
2219 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2220 		;
2221 
2222 	BO_UNLOCK(bo);
2223 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2224 }
2225 
2226 static int
2227 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2228     daddr_t startlbn, daddr_t endlbn)
2229 {
2230 	struct buf *bp, *nbp;
2231 	bool anyfreed;
2232 
2233 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2234 	ASSERT_BO_LOCKED(bo);
2235 
2236 	do {
2237 		anyfreed = false;
2238 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2239 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2240 				continue;
2241 			if (BUF_LOCK(bp,
2242 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2243 			    BO_LOCKPTR(bo)) == ENOLCK) {
2244 				BO_LOCK(bo);
2245 				return (EAGAIN);
2246 			}
2247 
2248 			bremfree(bp);
2249 			bp->b_flags |= B_INVAL | B_RELBUF;
2250 			bp->b_flags &= ~B_ASYNC;
2251 			brelse(bp);
2252 			anyfreed = true;
2253 
2254 			BO_LOCK(bo);
2255 			if (nbp != NULL &&
2256 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2257 			    nbp->b_vp != vp ||
2258 			    (nbp->b_flags & B_DELWRI) != 0))
2259 				return (EAGAIN);
2260 		}
2261 
2262 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2263 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2264 				continue;
2265 			if (BUF_LOCK(bp,
2266 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2267 			    BO_LOCKPTR(bo)) == ENOLCK) {
2268 				BO_LOCK(bo);
2269 				return (EAGAIN);
2270 			}
2271 			bremfree(bp);
2272 			bp->b_flags |= B_INVAL | B_RELBUF;
2273 			bp->b_flags &= ~B_ASYNC;
2274 			brelse(bp);
2275 			anyfreed = true;
2276 
2277 			BO_LOCK(bo);
2278 			if (nbp != NULL &&
2279 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2280 			    (nbp->b_vp != vp) ||
2281 			    (nbp->b_flags & B_DELWRI) == 0))
2282 				return (EAGAIN);
2283 		}
2284 	} while (anyfreed);
2285 	return (0);
2286 }
2287 
2288 static void
2289 buf_vlist_remove(struct buf *bp)
2290 {
2291 	struct bufv *bv;
2292 	b_xflags_t flags;
2293 
2294 	flags = bp->b_xflags;
2295 
2296 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2297 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2298 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2299 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2300 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2301 
2302 	if ((flags & BX_VNDIRTY) != 0)
2303 		bv = &bp->b_bufobj->bo_dirty;
2304 	else
2305 		bv = &bp->b_bufobj->bo_clean;
2306 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2307 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2308 	bv->bv_cnt--;
2309 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2310 }
2311 
2312 /*
2313  * Add the buffer to the sorted clean or dirty block list.
2314  *
2315  * NOTE: xflags is passed as a constant, optimizing this inline function!
2316  */
2317 static void
2318 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2319 {
2320 	struct bufv *bv;
2321 	struct buf *n;
2322 	int error;
2323 
2324 	ASSERT_BO_WLOCKED(bo);
2325 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2326 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2327 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2328 	    ("dead bo %p", bo));
2329 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2330 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2331 	bp->b_xflags |= xflags;
2332 	if (xflags & BX_VNDIRTY)
2333 		bv = &bo->bo_dirty;
2334 	else
2335 		bv = &bo->bo_clean;
2336 
2337 	/*
2338 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2339 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2340 	 * than _ge.
2341 	 */
2342 	if (bv->bv_cnt == 0 ||
2343 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2344 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2345 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2346 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2347 	else
2348 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2349 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2350 	if (error)
2351 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2352 	bv->bv_cnt++;
2353 }
2354 
2355 /*
2356  * Look up a buffer using the buffer tries.
2357  */
2358 struct buf *
2359 gbincore(struct bufobj *bo, daddr_t lblkno)
2360 {
2361 	struct buf *bp;
2362 
2363 	ASSERT_BO_LOCKED(bo);
2364 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2365 	if (bp != NULL)
2366 		return (bp);
2367 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2368 }
2369 
2370 /*
2371  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2372  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2373  * stability of the result.  Like other lockless lookups, the found buf may
2374  * already be invalid by the time this function returns.
2375  */
2376 struct buf *
2377 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2378 {
2379 	struct buf *bp;
2380 
2381 	ASSERT_BO_UNLOCKED(bo);
2382 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2383 	if (bp != NULL)
2384 		return (bp);
2385 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2386 }
2387 
2388 /*
2389  * Associate a buffer with a vnode.
2390  */
2391 void
2392 bgetvp(struct vnode *vp, struct buf *bp)
2393 {
2394 	struct bufobj *bo;
2395 
2396 	bo = &vp->v_bufobj;
2397 	ASSERT_BO_WLOCKED(bo);
2398 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2399 
2400 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2401 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2402 	    ("bgetvp: bp already attached! %p", bp));
2403 
2404 	vhold(vp);
2405 	bp->b_vp = vp;
2406 	bp->b_bufobj = bo;
2407 	/*
2408 	 * Insert onto list for new vnode.
2409 	 */
2410 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2411 }
2412 
2413 /*
2414  * Disassociate a buffer from a vnode.
2415  */
2416 void
2417 brelvp(struct buf *bp)
2418 {
2419 	struct bufobj *bo;
2420 	struct vnode *vp;
2421 
2422 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2423 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2424 
2425 	/*
2426 	 * Delete from old vnode list, if on one.
2427 	 */
2428 	vp = bp->b_vp;		/* XXX */
2429 	bo = bp->b_bufobj;
2430 	BO_LOCK(bo);
2431 	buf_vlist_remove(bp);
2432 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2433 		bo->bo_flag &= ~BO_ONWORKLST;
2434 		mtx_lock(&sync_mtx);
2435 		LIST_REMOVE(bo, bo_synclist);
2436 		syncer_worklist_len--;
2437 		mtx_unlock(&sync_mtx);
2438 	}
2439 	bp->b_vp = NULL;
2440 	bp->b_bufobj = NULL;
2441 	BO_UNLOCK(bo);
2442 	vdrop(vp);
2443 }
2444 
2445 /*
2446  * Add an item to the syncer work queue.
2447  */
2448 static void
2449 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2450 {
2451 	int slot;
2452 
2453 	ASSERT_BO_WLOCKED(bo);
2454 
2455 	mtx_lock(&sync_mtx);
2456 	if (bo->bo_flag & BO_ONWORKLST)
2457 		LIST_REMOVE(bo, bo_synclist);
2458 	else {
2459 		bo->bo_flag |= BO_ONWORKLST;
2460 		syncer_worklist_len++;
2461 	}
2462 
2463 	if (delay > syncer_maxdelay - 2)
2464 		delay = syncer_maxdelay - 2;
2465 	slot = (syncer_delayno + delay) & syncer_mask;
2466 
2467 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2468 	mtx_unlock(&sync_mtx);
2469 }
2470 
2471 static int
2472 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2473 {
2474 	int error, len;
2475 
2476 	mtx_lock(&sync_mtx);
2477 	len = syncer_worklist_len - sync_vnode_count;
2478 	mtx_unlock(&sync_mtx);
2479 	error = SYSCTL_OUT(req, &len, sizeof(len));
2480 	return (error);
2481 }
2482 
2483 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2484     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2485     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2486 
2487 static struct proc *updateproc;
2488 static void sched_sync(void);
2489 static struct kproc_desc up_kp = {
2490 	"syncer",
2491 	sched_sync,
2492 	&updateproc
2493 };
2494 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2495 
2496 static int
2497 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2498 {
2499 	struct vnode *vp;
2500 	struct mount *mp;
2501 
2502 	*bo = LIST_FIRST(slp);
2503 	if (*bo == NULL)
2504 		return (0);
2505 	vp = bo2vnode(*bo);
2506 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2507 		return (1);
2508 	/*
2509 	 * We use vhold in case the vnode does not
2510 	 * successfully sync.  vhold prevents the vnode from
2511 	 * going away when we unlock the sync_mtx so that
2512 	 * we can acquire the vnode interlock.
2513 	 */
2514 	vholdl(vp);
2515 	mtx_unlock(&sync_mtx);
2516 	VI_UNLOCK(vp);
2517 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2518 		vdrop(vp);
2519 		mtx_lock(&sync_mtx);
2520 		return (*bo == LIST_FIRST(slp));
2521 	}
2522 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2523 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2524 	VOP_UNLOCK(vp);
2525 	vn_finished_write(mp);
2526 	BO_LOCK(*bo);
2527 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2528 		/*
2529 		 * Put us back on the worklist.  The worklist
2530 		 * routine will remove us from our current
2531 		 * position and then add us back in at a later
2532 		 * position.
2533 		 */
2534 		vn_syncer_add_to_worklist(*bo, syncdelay);
2535 	}
2536 	BO_UNLOCK(*bo);
2537 	vdrop(vp);
2538 	mtx_lock(&sync_mtx);
2539 	return (0);
2540 }
2541 
2542 static int first_printf = 1;
2543 
2544 /*
2545  * System filesystem synchronizer daemon.
2546  */
2547 static void
2548 sched_sync(void)
2549 {
2550 	struct synclist *next, *slp;
2551 	struct bufobj *bo;
2552 	long starttime;
2553 	struct thread *td = curthread;
2554 	int last_work_seen;
2555 	int net_worklist_len;
2556 	int syncer_final_iter;
2557 	int error;
2558 
2559 	last_work_seen = 0;
2560 	syncer_final_iter = 0;
2561 	syncer_state = SYNCER_RUNNING;
2562 	starttime = time_uptime;
2563 	td->td_pflags |= TDP_NORUNNINGBUF;
2564 
2565 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2566 	    SHUTDOWN_PRI_LAST);
2567 
2568 	mtx_lock(&sync_mtx);
2569 	for (;;) {
2570 		if (syncer_state == SYNCER_FINAL_DELAY &&
2571 		    syncer_final_iter == 0) {
2572 			mtx_unlock(&sync_mtx);
2573 			kproc_suspend_check(td->td_proc);
2574 			mtx_lock(&sync_mtx);
2575 		}
2576 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2577 		if (syncer_state != SYNCER_RUNNING &&
2578 		    starttime != time_uptime) {
2579 			if (first_printf) {
2580 				printf("\nSyncing disks, vnodes remaining... ");
2581 				first_printf = 0;
2582 			}
2583 			printf("%d ", net_worklist_len);
2584 		}
2585 		starttime = time_uptime;
2586 
2587 		/*
2588 		 * Push files whose dirty time has expired.  Be careful
2589 		 * of interrupt race on slp queue.
2590 		 *
2591 		 * Skip over empty worklist slots when shutting down.
2592 		 */
2593 		do {
2594 			slp = &syncer_workitem_pending[syncer_delayno];
2595 			syncer_delayno += 1;
2596 			if (syncer_delayno == syncer_maxdelay)
2597 				syncer_delayno = 0;
2598 			next = &syncer_workitem_pending[syncer_delayno];
2599 			/*
2600 			 * If the worklist has wrapped since the
2601 			 * it was emptied of all but syncer vnodes,
2602 			 * switch to the FINAL_DELAY state and run
2603 			 * for one more second.
2604 			 */
2605 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2606 			    net_worklist_len == 0 &&
2607 			    last_work_seen == syncer_delayno) {
2608 				syncer_state = SYNCER_FINAL_DELAY;
2609 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2610 			}
2611 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2612 		    syncer_worklist_len > 0);
2613 
2614 		/*
2615 		 * Keep track of the last time there was anything
2616 		 * on the worklist other than syncer vnodes.
2617 		 * Return to the SHUTTING_DOWN state if any
2618 		 * new work appears.
2619 		 */
2620 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2621 			last_work_seen = syncer_delayno;
2622 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2623 			syncer_state = SYNCER_SHUTTING_DOWN;
2624 		while (!LIST_EMPTY(slp)) {
2625 			error = sync_vnode(slp, &bo, td);
2626 			if (error == 1) {
2627 				LIST_REMOVE(bo, bo_synclist);
2628 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2629 				continue;
2630 			}
2631 
2632 			if (first_printf == 0) {
2633 				/*
2634 				 * Drop the sync mutex, because some watchdog
2635 				 * drivers need to sleep while patting
2636 				 */
2637 				mtx_unlock(&sync_mtx);
2638 				wdog_kern_pat(WD_LASTVAL);
2639 				mtx_lock(&sync_mtx);
2640 			}
2641 		}
2642 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2643 			syncer_final_iter--;
2644 		/*
2645 		 * The variable rushjob allows the kernel to speed up the
2646 		 * processing of the filesystem syncer process. A rushjob
2647 		 * value of N tells the filesystem syncer to process the next
2648 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2649 		 * is used by the soft update code to speed up the filesystem
2650 		 * syncer process when the incore state is getting so far
2651 		 * ahead of the disk that the kernel memory pool is being
2652 		 * threatened with exhaustion.
2653 		 */
2654 		if (rushjob > 0) {
2655 			rushjob -= 1;
2656 			continue;
2657 		}
2658 		/*
2659 		 * Just sleep for a short period of time between
2660 		 * iterations when shutting down to allow some I/O
2661 		 * to happen.
2662 		 *
2663 		 * If it has taken us less than a second to process the
2664 		 * current work, then wait. Otherwise start right over
2665 		 * again. We can still lose time if any single round
2666 		 * takes more than two seconds, but it does not really
2667 		 * matter as we are just trying to generally pace the
2668 		 * filesystem activity.
2669 		 */
2670 		if (syncer_state != SYNCER_RUNNING ||
2671 		    time_uptime == starttime) {
2672 			thread_lock(td);
2673 			sched_prio(td, PPAUSE);
2674 			thread_unlock(td);
2675 		}
2676 		if (syncer_state != SYNCER_RUNNING)
2677 			cv_timedwait(&sync_wakeup, &sync_mtx,
2678 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2679 		else if (time_uptime == starttime)
2680 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2681 	}
2682 }
2683 
2684 /*
2685  * Request the syncer daemon to speed up its work.
2686  * We never push it to speed up more than half of its
2687  * normal turn time, otherwise it could take over the cpu.
2688  */
2689 int
2690 speedup_syncer(void)
2691 {
2692 	int ret = 0;
2693 
2694 	mtx_lock(&sync_mtx);
2695 	if (rushjob < syncdelay / 2) {
2696 		rushjob += 1;
2697 		stat_rush_requests += 1;
2698 		ret = 1;
2699 	}
2700 	mtx_unlock(&sync_mtx);
2701 	cv_broadcast(&sync_wakeup);
2702 	return (ret);
2703 }
2704 
2705 /*
2706  * Tell the syncer to speed up its work and run though its work
2707  * list several times, then tell it to shut down.
2708  */
2709 static void
2710 syncer_shutdown(void *arg, int howto)
2711 {
2712 
2713 	if (howto & RB_NOSYNC)
2714 		return;
2715 	mtx_lock(&sync_mtx);
2716 	syncer_state = SYNCER_SHUTTING_DOWN;
2717 	rushjob = 0;
2718 	mtx_unlock(&sync_mtx);
2719 	cv_broadcast(&sync_wakeup);
2720 	kproc_shutdown(arg, howto);
2721 }
2722 
2723 void
2724 syncer_suspend(void)
2725 {
2726 
2727 	syncer_shutdown(updateproc, 0);
2728 }
2729 
2730 void
2731 syncer_resume(void)
2732 {
2733 
2734 	mtx_lock(&sync_mtx);
2735 	first_printf = 1;
2736 	syncer_state = SYNCER_RUNNING;
2737 	mtx_unlock(&sync_mtx);
2738 	cv_broadcast(&sync_wakeup);
2739 	kproc_resume(updateproc);
2740 }
2741 
2742 /*
2743  * Move the buffer between the clean and dirty lists of its vnode.
2744  */
2745 void
2746 reassignbuf(struct buf *bp)
2747 {
2748 	struct vnode *vp;
2749 	struct bufobj *bo;
2750 	int delay;
2751 #ifdef INVARIANTS
2752 	struct bufv *bv;
2753 #endif
2754 
2755 	vp = bp->b_vp;
2756 	bo = bp->b_bufobj;
2757 
2758 	KASSERT((bp->b_flags & B_PAGING) == 0,
2759 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2760 
2761 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2762 	    bp, bp->b_vp, bp->b_flags);
2763 
2764 	BO_LOCK(bo);
2765 	buf_vlist_remove(bp);
2766 
2767 	/*
2768 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2769 	 * of clean buffers.
2770 	 */
2771 	if (bp->b_flags & B_DELWRI) {
2772 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2773 			switch (vp->v_type) {
2774 			case VDIR:
2775 				delay = dirdelay;
2776 				break;
2777 			case VCHR:
2778 				delay = metadelay;
2779 				break;
2780 			default:
2781 				delay = filedelay;
2782 			}
2783 			vn_syncer_add_to_worklist(bo, delay);
2784 		}
2785 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2786 	} else {
2787 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2788 
2789 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2790 			mtx_lock(&sync_mtx);
2791 			LIST_REMOVE(bo, bo_synclist);
2792 			syncer_worklist_len--;
2793 			mtx_unlock(&sync_mtx);
2794 			bo->bo_flag &= ~BO_ONWORKLST;
2795 		}
2796 	}
2797 #ifdef INVARIANTS
2798 	bv = &bo->bo_clean;
2799 	bp = TAILQ_FIRST(&bv->bv_hd);
2800 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2801 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2802 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2803 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2804 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2805 	bv = &bo->bo_dirty;
2806 	bp = TAILQ_FIRST(&bv->bv_hd);
2807 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2808 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2809 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2810 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2811 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2812 #endif
2813 	BO_UNLOCK(bo);
2814 }
2815 
2816 static void
2817 v_init_counters(struct vnode *vp)
2818 {
2819 
2820 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2821 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2822 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2823 
2824 	refcount_init(&vp->v_holdcnt, 1);
2825 	refcount_init(&vp->v_usecount, 1);
2826 }
2827 
2828 /*
2829  * Grab a particular vnode from the free list, increment its
2830  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2831  * is being destroyed.  Only callers who specify LK_RETRY will
2832  * see doomed vnodes.  If inactive processing was delayed in
2833  * vput try to do it here.
2834  *
2835  * usecount is manipulated using atomics without holding any locks.
2836  *
2837  * holdcnt can be manipulated using atomics without holding any locks,
2838  * except when transitioning 1<->0, in which case the interlock is held.
2839  *
2840  * Consumers which don't guarantee liveness of the vnode can use SMR to
2841  * try to get a reference. Note this operation can fail since the vnode
2842  * may be awaiting getting freed by the time they get to it.
2843  */
2844 enum vgetstate
2845 vget_prep_smr(struct vnode *vp)
2846 {
2847 	enum vgetstate vs;
2848 
2849 	VFS_SMR_ASSERT_ENTERED();
2850 
2851 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2852 		vs = VGET_USECOUNT;
2853 	} else {
2854 		if (vhold_smr(vp))
2855 			vs = VGET_HOLDCNT;
2856 		else
2857 			vs = VGET_NONE;
2858 	}
2859 	return (vs);
2860 }
2861 
2862 enum vgetstate
2863 vget_prep(struct vnode *vp)
2864 {
2865 	enum vgetstate vs;
2866 
2867 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2868 		vs = VGET_USECOUNT;
2869 	} else {
2870 		vhold(vp);
2871 		vs = VGET_HOLDCNT;
2872 	}
2873 	return (vs);
2874 }
2875 
2876 void
2877 vget_abort(struct vnode *vp, enum vgetstate vs)
2878 {
2879 
2880 	switch (vs) {
2881 	case VGET_USECOUNT:
2882 		vrele(vp);
2883 		break;
2884 	case VGET_HOLDCNT:
2885 		vdrop(vp);
2886 		break;
2887 	default:
2888 		__assert_unreachable();
2889 	}
2890 }
2891 
2892 int
2893 vget(struct vnode *vp, int flags)
2894 {
2895 	enum vgetstate vs;
2896 
2897 	vs = vget_prep(vp);
2898 	return (vget_finish(vp, flags, vs));
2899 }
2900 
2901 int
2902 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2903 {
2904 	int error;
2905 
2906 	if ((flags & LK_INTERLOCK) != 0)
2907 		ASSERT_VI_LOCKED(vp, __func__);
2908 	else
2909 		ASSERT_VI_UNLOCKED(vp, __func__);
2910 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2911 	VNPASS(vp->v_holdcnt > 0, vp);
2912 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2913 
2914 	error = vn_lock(vp, flags);
2915 	if (__predict_false(error != 0)) {
2916 		vget_abort(vp, vs);
2917 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2918 		    vp);
2919 		return (error);
2920 	}
2921 
2922 	vget_finish_ref(vp, vs);
2923 	return (0);
2924 }
2925 
2926 void
2927 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
2928 {
2929 	int old;
2930 
2931 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2932 	VNPASS(vp->v_holdcnt > 0, vp);
2933 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2934 
2935 	if (vs == VGET_USECOUNT)
2936 		return;
2937 
2938 	/*
2939 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2940 	 * the vnode around. Otherwise someone else lended their hold count and
2941 	 * we have to drop ours.
2942 	 */
2943 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2944 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2945 	if (old != 0) {
2946 #ifdef INVARIANTS
2947 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2948 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2949 #else
2950 		refcount_release(&vp->v_holdcnt);
2951 #endif
2952 	}
2953 }
2954 
2955 void
2956 vref(struct vnode *vp)
2957 {
2958 	enum vgetstate vs;
2959 
2960 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2961 	vs = vget_prep(vp);
2962 	vget_finish_ref(vp, vs);
2963 }
2964 
2965 void
2966 vrefact(struct vnode *vp)
2967 {
2968 
2969 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2970 #ifdef INVARIANTS
2971 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2972 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
2973 #else
2974 	refcount_acquire(&vp->v_usecount);
2975 #endif
2976 }
2977 
2978 void
2979 vlazy(struct vnode *vp)
2980 {
2981 	struct mount *mp;
2982 
2983 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2984 
2985 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
2986 		return;
2987 	/*
2988 	 * We may get here for inactive routines after the vnode got doomed.
2989 	 */
2990 	if (VN_IS_DOOMED(vp))
2991 		return;
2992 	mp = vp->v_mount;
2993 	mtx_lock(&mp->mnt_listmtx);
2994 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
2995 		vp->v_mflag |= VMP_LAZYLIST;
2996 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
2997 		mp->mnt_lazyvnodelistsize++;
2998 	}
2999 	mtx_unlock(&mp->mnt_listmtx);
3000 }
3001 
3002 /*
3003  * This routine is only meant to be called from vgonel prior to dooming
3004  * the vnode.
3005  */
3006 static void
3007 vunlazy_gone(struct vnode *vp)
3008 {
3009 	struct mount *mp;
3010 
3011 	ASSERT_VOP_ELOCKED(vp, __func__);
3012 	ASSERT_VI_LOCKED(vp, __func__);
3013 	VNPASS(!VN_IS_DOOMED(vp), vp);
3014 
3015 	if (vp->v_mflag & VMP_LAZYLIST) {
3016 		mp = vp->v_mount;
3017 		mtx_lock(&mp->mnt_listmtx);
3018 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3019 		vp->v_mflag &= ~VMP_LAZYLIST;
3020 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3021 		mp->mnt_lazyvnodelistsize--;
3022 		mtx_unlock(&mp->mnt_listmtx);
3023 	}
3024 }
3025 
3026 static void
3027 vdefer_inactive(struct vnode *vp)
3028 {
3029 
3030 	ASSERT_VI_LOCKED(vp, __func__);
3031 	VNASSERT(vp->v_holdcnt > 0, vp,
3032 	    ("%s: vnode without hold count", __func__));
3033 	if (VN_IS_DOOMED(vp)) {
3034 		vdropl(vp);
3035 		return;
3036 	}
3037 	if (vp->v_iflag & VI_DEFINACT) {
3038 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3039 		vdropl(vp);
3040 		return;
3041 	}
3042 	if (vp->v_usecount > 0) {
3043 		vp->v_iflag &= ~VI_OWEINACT;
3044 		vdropl(vp);
3045 		return;
3046 	}
3047 	vlazy(vp);
3048 	vp->v_iflag |= VI_DEFINACT;
3049 	VI_UNLOCK(vp);
3050 	counter_u64_add(deferred_inact, 1);
3051 }
3052 
3053 static void
3054 vdefer_inactive_unlocked(struct vnode *vp)
3055 {
3056 
3057 	VI_LOCK(vp);
3058 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3059 		vdropl(vp);
3060 		return;
3061 	}
3062 	vdefer_inactive(vp);
3063 }
3064 
3065 enum vput_op { VRELE, VPUT, VUNREF };
3066 
3067 /*
3068  * Handle ->v_usecount transitioning to 0.
3069  *
3070  * By releasing the last usecount we take ownership of the hold count which
3071  * provides liveness of the vnode, meaning we have to vdrop.
3072  *
3073  * For all vnodes we may need to perform inactive processing. It requires an
3074  * exclusive lock on the vnode, while it is legal to call here with only a
3075  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3076  * inactive processing gets deferred to the syncer.
3077  *
3078  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3079  * on the lock being held all the way until VOP_INACTIVE. This in particular
3080  * happens with UFS which adds half-constructed vnodes to the hash, where they
3081  * can be found by other code.
3082  */
3083 static void
3084 vput_final(struct vnode *vp, enum vput_op func)
3085 {
3086 	int error;
3087 	bool want_unlock;
3088 
3089 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3090 	VNPASS(vp->v_holdcnt > 0, vp);
3091 
3092 	VI_LOCK(vp);
3093 
3094 	/*
3095 	 * By the time we got here someone else might have transitioned
3096 	 * the count back to > 0.
3097 	 */
3098 	if (vp->v_usecount > 0)
3099 		goto out;
3100 
3101 	/*
3102 	 * If the vnode is doomed vgone already performed inactive processing
3103 	 * (if needed).
3104 	 */
3105 	if (VN_IS_DOOMED(vp))
3106 		goto out;
3107 
3108 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3109 		goto out;
3110 
3111 	if (vp->v_iflag & VI_DOINGINACT)
3112 		goto out;
3113 
3114 	/*
3115 	 * Locking operations here will drop the interlock and possibly the
3116 	 * vnode lock, opening a window where the vnode can get doomed all the
3117 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3118 	 * perform inactive.
3119 	 */
3120 	vp->v_iflag |= VI_OWEINACT;
3121 	want_unlock = false;
3122 	error = 0;
3123 	switch (func) {
3124 	case VRELE:
3125 		switch (VOP_ISLOCKED(vp)) {
3126 		case LK_EXCLUSIVE:
3127 			break;
3128 		case LK_EXCLOTHER:
3129 		case 0:
3130 			want_unlock = true;
3131 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3132 			VI_LOCK(vp);
3133 			break;
3134 		default:
3135 			/*
3136 			 * The lock has at least one sharer, but we have no way
3137 			 * to conclude whether this is us. Play it safe and
3138 			 * defer processing.
3139 			 */
3140 			error = EAGAIN;
3141 			break;
3142 		}
3143 		break;
3144 	case VPUT:
3145 		want_unlock = true;
3146 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3147 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3148 			    LK_NOWAIT);
3149 			VI_LOCK(vp);
3150 		}
3151 		break;
3152 	case VUNREF:
3153 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3154 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3155 			VI_LOCK(vp);
3156 		}
3157 		break;
3158 	}
3159 	if (error == 0) {
3160 		vinactive(vp);
3161 		if (want_unlock)
3162 			VOP_UNLOCK(vp);
3163 		vdropl(vp);
3164 	} else {
3165 		vdefer_inactive(vp);
3166 	}
3167 	return;
3168 out:
3169 	if (func == VPUT)
3170 		VOP_UNLOCK(vp);
3171 	vdropl(vp);
3172 }
3173 
3174 /*
3175  * Decrement ->v_usecount for a vnode.
3176  *
3177  * Releasing the last use count requires additional processing, see vput_final
3178  * above for details.
3179  *
3180  * Comment above each variant denotes lock state on entry and exit.
3181  */
3182 
3183 /*
3184  * in: any
3185  * out: same as passed in
3186  */
3187 void
3188 vrele(struct vnode *vp)
3189 {
3190 
3191 	ASSERT_VI_UNLOCKED(vp, __func__);
3192 	if (!refcount_release(&vp->v_usecount))
3193 		return;
3194 	vput_final(vp, VRELE);
3195 }
3196 
3197 /*
3198  * in: locked
3199  * out: unlocked
3200  */
3201 void
3202 vput(struct vnode *vp)
3203 {
3204 
3205 	ASSERT_VOP_LOCKED(vp, __func__);
3206 	ASSERT_VI_UNLOCKED(vp, __func__);
3207 	if (!refcount_release(&vp->v_usecount)) {
3208 		VOP_UNLOCK(vp);
3209 		return;
3210 	}
3211 	vput_final(vp, VPUT);
3212 }
3213 
3214 /*
3215  * in: locked
3216  * out: locked
3217  */
3218 void
3219 vunref(struct vnode *vp)
3220 {
3221 
3222 	ASSERT_VOP_LOCKED(vp, __func__);
3223 	ASSERT_VI_UNLOCKED(vp, __func__);
3224 	if (!refcount_release(&vp->v_usecount))
3225 		return;
3226 	vput_final(vp, VUNREF);
3227 }
3228 
3229 void
3230 vhold(struct vnode *vp)
3231 {
3232 	int old;
3233 
3234 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3235 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3236 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3237 	    ("%s: wrong hold count %d", __func__, old));
3238 	if (old == 0)
3239 		vn_freevnodes_dec();
3240 }
3241 
3242 void
3243 vholdnz(struct vnode *vp)
3244 {
3245 
3246 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3247 #ifdef INVARIANTS
3248 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3249 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3250 	    ("%s: wrong hold count %d", __func__, old));
3251 #else
3252 	atomic_add_int(&vp->v_holdcnt, 1);
3253 #endif
3254 }
3255 
3256 /*
3257  * Grab a hold count unless the vnode is freed.
3258  *
3259  * Only use this routine if vfs smr is the only protection you have against
3260  * freeing the vnode.
3261  *
3262  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3263  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3264  * the thread which managed to set the flag.
3265  *
3266  * It may be tempting to replace the loop with:
3267  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3268  * if (count & VHOLD_NO_SMR) {
3269  *     backpedal and error out;
3270  * }
3271  *
3272  * However, while this is more performant, it hinders debugging by eliminating
3273  * the previously mentioned invariant.
3274  */
3275 bool
3276 vhold_smr(struct vnode *vp)
3277 {
3278 	int count;
3279 
3280 	VFS_SMR_ASSERT_ENTERED();
3281 
3282 	count = atomic_load_int(&vp->v_holdcnt);
3283 	for (;;) {
3284 		if (count & VHOLD_NO_SMR) {
3285 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3286 			    ("non-zero hold count with flags %d\n", count));
3287 			return (false);
3288 		}
3289 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3290 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3291 			if (count == 0)
3292 				vn_freevnodes_dec();
3293 			return (true);
3294 		}
3295 	}
3296 }
3297 
3298 /*
3299  * Hold a free vnode for recycling.
3300  *
3301  * Note: vnode_init references this comment.
3302  *
3303  * Attempts to recycle only need the global vnode list lock and have no use for
3304  * SMR.
3305  *
3306  * However, vnodes get inserted into the global list before they get fully
3307  * initialized and stay there until UMA decides to free the memory. This in
3308  * particular means the target can be found before it becomes usable and after
3309  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3310  * VHOLD_NO_SMR.
3311  *
3312  * Note: the vnode may gain more references after we transition the count 0->1.
3313  */
3314 static bool
3315 vhold_recycle_free(struct vnode *vp)
3316 {
3317 	int count;
3318 
3319 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3320 
3321 	count = atomic_load_int(&vp->v_holdcnt);
3322 	for (;;) {
3323 		if (count & VHOLD_NO_SMR) {
3324 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3325 			    ("non-zero hold count with flags %d\n", count));
3326 			return (false);
3327 		}
3328 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3329 		if (count > 0) {
3330 			return (false);
3331 		}
3332 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3333 			vn_freevnodes_dec();
3334 			return (true);
3335 		}
3336 	}
3337 }
3338 
3339 static void __noinline
3340 vdbatch_process(struct vdbatch *vd)
3341 {
3342 	struct vnode *vp;
3343 	int i;
3344 
3345 	mtx_assert(&vd->lock, MA_OWNED);
3346 	MPASS(curthread->td_pinned > 0);
3347 	MPASS(vd->index == VDBATCH_SIZE);
3348 
3349 	mtx_lock(&vnode_list_mtx);
3350 	critical_enter();
3351 	freevnodes += vd->freevnodes;
3352 	for (i = 0; i < VDBATCH_SIZE; i++) {
3353 		vp = vd->tab[i];
3354 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3355 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3356 		MPASS(vp->v_dbatchcpu != NOCPU);
3357 		vp->v_dbatchcpu = NOCPU;
3358 	}
3359 	mtx_unlock(&vnode_list_mtx);
3360 	vd->freevnodes = 0;
3361 	bzero(vd->tab, sizeof(vd->tab));
3362 	vd->index = 0;
3363 	critical_exit();
3364 }
3365 
3366 static void
3367 vdbatch_enqueue(struct vnode *vp)
3368 {
3369 	struct vdbatch *vd;
3370 
3371 	ASSERT_VI_LOCKED(vp, __func__);
3372 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3373 	    ("%s: deferring requeue of a doomed vnode", __func__));
3374 
3375 	if (vp->v_dbatchcpu != NOCPU) {
3376 		VI_UNLOCK(vp);
3377 		return;
3378 	}
3379 
3380 	sched_pin();
3381 	vd = DPCPU_PTR(vd);
3382 	mtx_lock(&vd->lock);
3383 	MPASS(vd->index < VDBATCH_SIZE);
3384 	MPASS(vd->tab[vd->index] == NULL);
3385 	/*
3386 	 * A hack: we depend on being pinned so that we know what to put in
3387 	 * ->v_dbatchcpu.
3388 	 */
3389 	vp->v_dbatchcpu = curcpu;
3390 	vd->tab[vd->index] = vp;
3391 	vd->index++;
3392 	VI_UNLOCK(vp);
3393 	if (vd->index == VDBATCH_SIZE)
3394 		vdbatch_process(vd);
3395 	mtx_unlock(&vd->lock);
3396 	sched_unpin();
3397 }
3398 
3399 /*
3400  * This routine must only be called for vnodes which are about to be
3401  * deallocated. Supporting dequeue for arbitrary vndoes would require
3402  * validating that the locked batch matches.
3403  */
3404 static void
3405 vdbatch_dequeue(struct vnode *vp)
3406 {
3407 	struct vdbatch *vd;
3408 	int i;
3409 	short cpu;
3410 
3411 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3412 	    ("%s: called for a used vnode\n", __func__));
3413 
3414 	cpu = vp->v_dbatchcpu;
3415 	if (cpu == NOCPU)
3416 		return;
3417 
3418 	vd = DPCPU_ID_PTR(cpu, vd);
3419 	mtx_lock(&vd->lock);
3420 	for (i = 0; i < vd->index; i++) {
3421 		if (vd->tab[i] != vp)
3422 			continue;
3423 		vp->v_dbatchcpu = NOCPU;
3424 		vd->index--;
3425 		vd->tab[i] = vd->tab[vd->index];
3426 		vd->tab[vd->index] = NULL;
3427 		break;
3428 	}
3429 	mtx_unlock(&vd->lock);
3430 	/*
3431 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3432 	 */
3433 	MPASS(vp->v_dbatchcpu == NOCPU);
3434 }
3435 
3436 /*
3437  * Drop the hold count of the vnode.  If this is the last reference to
3438  * the vnode we place it on the free list unless it has been vgone'd
3439  * (marked VIRF_DOOMED) in which case we will free it.
3440  *
3441  * Because the vnode vm object keeps a hold reference on the vnode if
3442  * there is at least one resident non-cached page, the vnode cannot
3443  * leave the active list without the page cleanup done.
3444  */
3445 static void
3446 vdrop_deactivate(struct vnode *vp)
3447 {
3448 	struct mount *mp;
3449 
3450 	ASSERT_VI_LOCKED(vp, __func__);
3451 	/*
3452 	 * Mark a vnode as free: remove it from its active list
3453 	 * and put it up for recycling on the freelist.
3454 	 */
3455 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3456 	    ("vdrop: returning doomed vnode"));
3457 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3458 	    ("vnode with VI_OWEINACT set"));
3459 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3460 	    ("vnode with VI_DEFINACT set"));
3461 	if (vp->v_mflag & VMP_LAZYLIST) {
3462 		mp = vp->v_mount;
3463 		mtx_lock(&mp->mnt_listmtx);
3464 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3465 		/*
3466 		 * Don't remove the vnode from the lazy list if another thread
3467 		 * has increased the hold count. It may have re-enqueued the
3468 		 * vnode to the lazy list and is now responsible for its
3469 		 * removal.
3470 		 */
3471 		if (vp->v_holdcnt == 0) {
3472 			vp->v_mflag &= ~VMP_LAZYLIST;
3473 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3474 			mp->mnt_lazyvnodelistsize--;
3475 		}
3476 		mtx_unlock(&mp->mnt_listmtx);
3477 	}
3478 	vdbatch_enqueue(vp);
3479 }
3480 
3481 static void __noinline
3482 vdropl_final(struct vnode *vp)
3483 {
3484 
3485 	ASSERT_VI_LOCKED(vp, __func__);
3486 	VNPASS(VN_IS_DOOMED(vp), vp);
3487 	/*
3488 	 * Set the VHOLD_NO_SMR flag.
3489 	 *
3490 	 * We may be racing against vhold_smr. If they win we can just pretend
3491 	 * we never got this far, they will vdrop later.
3492 	 */
3493 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3494 		vn_freevnodes_inc();
3495 		VI_UNLOCK(vp);
3496 		/*
3497 		 * We lost the aforementioned race. Any subsequent access is
3498 		 * invalid as they might have managed to vdropl on their own.
3499 		 */
3500 		return;
3501 	}
3502 	/*
3503 	 * Don't bump freevnodes as this one is going away.
3504 	 */
3505 	freevnode(vp);
3506 }
3507 
3508 void
3509 vdrop(struct vnode *vp)
3510 {
3511 
3512 	ASSERT_VI_UNLOCKED(vp, __func__);
3513 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3514 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3515 		return;
3516 	VI_LOCK(vp);
3517 	vdropl(vp);
3518 }
3519 
3520 void
3521 vdropl(struct vnode *vp)
3522 {
3523 
3524 	ASSERT_VI_LOCKED(vp, __func__);
3525 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3526 	if (!refcount_release(&vp->v_holdcnt)) {
3527 		VI_UNLOCK(vp);
3528 		return;
3529 	}
3530 	if (!VN_IS_DOOMED(vp)) {
3531 		vn_freevnodes_inc();
3532 		vdrop_deactivate(vp);
3533 		/*
3534 		 * Also unlocks the interlock. We can't assert on it as we
3535 		 * released our hold and by now the vnode might have been
3536 		 * freed.
3537 		 */
3538 		return;
3539 	}
3540 	vdropl_final(vp);
3541 }
3542 
3543 /*
3544  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3545  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3546  */
3547 static void
3548 vinactivef(struct vnode *vp)
3549 {
3550 	struct vm_object *obj;
3551 
3552 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3553 	ASSERT_VI_LOCKED(vp, "vinactive");
3554 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3555 	    ("vinactive: recursed on VI_DOINGINACT"));
3556 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3557 	vp->v_iflag |= VI_DOINGINACT;
3558 	vp->v_iflag &= ~VI_OWEINACT;
3559 	VI_UNLOCK(vp);
3560 	/*
3561 	 * Before moving off the active list, we must be sure that any
3562 	 * modified pages are converted into the vnode's dirty
3563 	 * buffers, since these will no longer be checked once the
3564 	 * vnode is on the inactive list.
3565 	 *
3566 	 * The write-out of the dirty pages is asynchronous.  At the
3567 	 * point that VOP_INACTIVE() is called, there could still be
3568 	 * pending I/O and dirty pages in the object.
3569 	 */
3570 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3571 	    vm_object_mightbedirty(obj)) {
3572 		VM_OBJECT_WLOCK(obj);
3573 		vm_object_page_clean(obj, 0, 0, 0);
3574 		VM_OBJECT_WUNLOCK(obj);
3575 	}
3576 	VOP_INACTIVE(vp);
3577 	VI_LOCK(vp);
3578 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3579 	    ("vinactive: lost VI_DOINGINACT"));
3580 	vp->v_iflag &= ~VI_DOINGINACT;
3581 }
3582 
3583 void
3584 vinactive(struct vnode *vp)
3585 {
3586 
3587 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3588 	ASSERT_VI_LOCKED(vp, "vinactive");
3589 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3590 
3591 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3592 		return;
3593 	if (vp->v_iflag & VI_DOINGINACT)
3594 		return;
3595 	if (vp->v_usecount > 0) {
3596 		vp->v_iflag &= ~VI_OWEINACT;
3597 		return;
3598 	}
3599 	vinactivef(vp);
3600 }
3601 
3602 /*
3603  * Remove any vnodes in the vnode table belonging to mount point mp.
3604  *
3605  * If FORCECLOSE is not specified, there should not be any active ones,
3606  * return error if any are found (nb: this is a user error, not a
3607  * system error). If FORCECLOSE is specified, detach any active vnodes
3608  * that are found.
3609  *
3610  * If WRITECLOSE is set, only flush out regular file vnodes open for
3611  * writing.
3612  *
3613  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3614  *
3615  * `rootrefs' specifies the base reference count for the root vnode
3616  * of this filesystem. The root vnode is considered busy if its
3617  * v_usecount exceeds this value. On a successful return, vflush(, td)
3618  * will call vrele() on the root vnode exactly rootrefs times.
3619  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3620  * be zero.
3621  */
3622 #ifdef DIAGNOSTIC
3623 static int busyprt = 0;		/* print out busy vnodes */
3624 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3625 #endif
3626 
3627 int
3628 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3629 {
3630 	struct vnode *vp, *mvp, *rootvp = NULL;
3631 	struct vattr vattr;
3632 	int busy = 0, error;
3633 
3634 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3635 	    rootrefs, flags);
3636 	if (rootrefs > 0) {
3637 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3638 		    ("vflush: bad args"));
3639 		/*
3640 		 * Get the filesystem root vnode. We can vput() it
3641 		 * immediately, since with rootrefs > 0, it won't go away.
3642 		 */
3643 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3644 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3645 			    __func__, error);
3646 			return (error);
3647 		}
3648 		vput(rootvp);
3649 	}
3650 loop:
3651 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3652 		vholdl(vp);
3653 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3654 		if (error) {
3655 			vdrop(vp);
3656 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3657 			goto loop;
3658 		}
3659 		/*
3660 		 * Skip over a vnodes marked VV_SYSTEM.
3661 		 */
3662 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3663 			VOP_UNLOCK(vp);
3664 			vdrop(vp);
3665 			continue;
3666 		}
3667 		/*
3668 		 * If WRITECLOSE is set, flush out unlinked but still open
3669 		 * files (even if open only for reading) and regular file
3670 		 * vnodes open for writing.
3671 		 */
3672 		if (flags & WRITECLOSE) {
3673 			if (vp->v_object != NULL) {
3674 				VM_OBJECT_WLOCK(vp->v_object);
3675 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3676 				VM_OBJECT_WUNLOCK(vp->v_object);
3677 			}
3678 			do {
3679 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3680 			} while (error == ERELOOKUP);
3681 			if (error != 0) {
3682 				VOP_UNLOCK(vp);
3683 				vdrop(vp);
3684 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3685 				return (error);
3686 			}
3687 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3688 			VI_LOCK(vp);
3689 
3690 			if ((vp->v_type == VNON ||
3691 			    (error == 0 && vattr.va_nlink > 0)) &&
3692 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3693 				VOP_UNLOCK(vp);
3694 				vdropl(vp);
3695 				continue;
3696 			}
3697 		} else
3698 			VI_LOCK(vp);
3699 		/*
3700 		 * With v_usecount == 0, all we need to do is clear out the
3701 		 * vnode data structures and we are done.
3702 		 *
3703 		 * If FORCECLOSE is set, forcibly close the vnode.
3704 		 */
3705 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3706 			vgonel(vp);
3707 		} else {
3708 			busy++;
3709 #ifdef DIAGNOSTIC
3710 			if (busyprt)
3711 				vn_printf(vp, "vflush: busy vnode ");
3712 #endif
3713 		}
3714 		VOP_UNLOCK(vp);
3715 		vdropl(vp);
3716 	}
3717 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3718 		/*
3719 		 * If just the root vnode is busy, and if its refcount
3720 		 * is equal to `rootrefs', then go ahead and kill it.
3721 		 */
3722 		VI_LOCK(rootvp);
3723 		KASSERT(busy > 0, ("vflush: not busy"));
3724 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3725 		    ("vflush: usecount %d < rootrefs %d",
3726 		     rootvp->v_usecount, rootrefs));
3727 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3728 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3729 			vgone(rootvp);
3730 			VOP_UNLOCK(rootvp);
3731 			busy = 0;
3732 		} else
3733 			VI_UNLOCK(rootvp);
3734 	}
3735 	if (busy) {
3736 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3737 		    busy);
3738 		return (EBUSY);
3739 	}
3740 	for (; rootrefs > 0; rootrefs--)
3741 		vrele(rootvp);
3742 	return (0);
3743 }
3744 
3745 /*
3746  * Recycle an unused vnode to the front of the free list.
3747  */
3748 int
3749 vrecycle(struct vnode *vp)
3750 {
3751 	int recycled;
3752 
3753 	VI_LOCK(vp);
3754 	recycled = vrecyclel(vp);
3755 	VI_UNLOCK(vp);
3756 	return (recycled);
3757 }
3758 
3759 /*
3760  * vrecycle, with the vp interlock held.
3761  */
3762 int
3763 vrecyclel(struct vnode *vp)
3764 {
3765 	int recycled;
3766 
3767 	ASSERT_VOP_ELOCKED(vp, __func__);
3768 	ASSERT_VI_LOCKED(vp, __func__);
3769 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3770 	recycled = 0;
3771 	if (vp->v_usecount == 0) {
3772 		recycled = 1;
3773 		vgonel(vp);
3774 	}
3775 	return (recycled);
3776 }
3777 
3778 /*
3779  * Eliminate all activity associated with a vnode
3780  * in preparation for reuse.
3781  */
3782 void
3783 vgone(struct vnode *vp)
3784 {
3785 	VI_LOCK(vp);
3786 	vgonel(vp);
3787 	VI_UNLOCK(vp);
3788 }
3789 
3790 static void
3791 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3792     struct vnode *lowervp __unused)
3793 {
3794 }
3795 
3796 /*
3797  * Notify upper mounts about reclaimed or unlinked vnode.
3798  */
3799 void
3800 vfs_notify_upper(struct vnode *vp, int event)
3801 {
3802 	static struct vfsops vgonel_vfsops = {
3803 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3804 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3805 	};
3806 	struct mount *mp, *ump, *mmp;
3807 
3808 	mp = vp->v_mount;
3809 	if (mp == NULL)
3810 		return;
3811 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3812 		return;
3813 
3814 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3815 	mmp->mnt_op = &vgonel_vfsops;
3816 	mmp->mnt_kern_flag |= MNTK_MARKER;
3817 	MNT_ILOCK(mp);
3818 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3819 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3820 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3821 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3822 			continue;
3823 		}
3824 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3825 		MNT_IUNLOCK(mp);
3826 		switch (event) {
3827 		case VFS_NOTIFY_UPPER_RECLAIM:
3828 			VFS_RECLAIM_LOWERVP(ump, vp);
3829 			break;
3830 		case VFS_NOTIFY_UPPER_UNLINK:
3831 			VFS_UNLINK_LOWERVP(ump, vp);
3832 			break;
3833 		default:
3834 			KASSERT(0, ("invalid event %d", event));
3835 			break;
3836 		}
3837 		MNT_ILOCK(mp);
3838 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3839 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3840 	}
3841 	free(mmp, M_TEMP);
3842 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3843 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3844 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3845 		wakeup(&mp->mnt_uppers);
3846 	}
3847 	MNT_IUNLOCK(mp);
3848 }
3849 
3850 /*
3851  * vgone, with the vp interlock held.
3852  */
3853 static void
3854 vgonel(struct vnode *vp)
3855 {
3856 	struct thread *td;
3857 	struct mount *mp;
3858 	vm_object_t object;
3859 	bool active, doinginact, oweinact;
3860 
3861 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3862 	ASSERT_VI_LOCKED(vp, "vgonel");
3863 	VNASSERT(vp->v_holdcnt, vp,
3864 	    ("vgonel: vp %p has no reference.", vp));
3865 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3866 	td = curthread;
3867 
3868 	/*
3869 	 * Don't vgonel if we're already doomed.
3870 	 */
3871 	if (vp->v_irflag & VIRF_DOOMED)
3872 		return;
3873 	/*
3874 	 * Paired with freevnode.
3875 	 */
3876 	vn_seqc_write_begin_locked(vp);
3877 	vunlazy_gone(vp);
3878 	vp->v_irflag |= VIRF_DOOMED;
3879 
3880 	/*
3881 	 * Check to see if the vnode is in use.  If so, we have to
3882 	 * call VOP_CLOSE() and VOP_INACTIVE().
3883 	 *
3884 	 * It could be that VOP_INACTIVE() requested reclamation, in
3885 	 * which case we should avoid recursion, so check
3886 	 * VI_DOINGINACT.  This is not precise but good enough.
3887 	 */
3888 	active = vp->v_usecount > 0;
3889 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3890 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3891 
3892 	/*
3893 	 * If we need to do inactive VI_OWEINACT will be set.
3894 	 */
3895 	if (vp->v_iflag & VI_DEFINACT) {
3896 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3897 		vp->v_iflag &= ~VI_DEFINACT;
3898 		vdropl(vp);
3899 	} else {
3900 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3901 		VI_UNLOCK(vp);
3902 	}
3903 	cache_purge_vgone(vp);
3904 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3905 
3906 	/*
3907 	 * If purging an active vnode, it must be closed and
3908 	 * deactivated before being reclaimed.
3909 	 */
3910 	if (active)
3911 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3912 	if ((oweinact || active) && !doinginact) {
3913 		VI_LOCK(vp);
3914 		vinactivef(vp);
3915 		VI_UNLOCK(vp);
3916 	}
3917 	if (vp->v_type == VSOCK)
3918 		vfs_unp_reclaim(vp);
3919 
3920 	/*
3921 	 * Clean out any buffers associated with the vnode.
3922 	 * If the flush fails, just toss the buffers.
3923 	 */
3924 	mp = NULL;
3925 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3926 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3927 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3928 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3929 			;
3930 	}
3931 
3932 	BO_LOCK(&vp->v_bufobj);
3933 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3934 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3935 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3936 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3937 	    ("vp %p bufobj not invalidated", vp));
3938 
3939 	/*
3940 	 * For VMIO bufobj, BO_DEAD is set later, or in
3941 	 * vm_object_terminate() after the object's page queue is
3942 	 * flushed.
3943 	 */
3944 	object = vp->v_bufobj.bo_object;
3945 	if (object == NULL)
3946 		vp->v_bufobj.bo_flag |= BO_DEAD;
3947 	BO_UNLOCK(&vp->v_bufobj);
3948 
3949 	/*
3950 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3951 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3952 	 * should not touch the object borrowed from the lower vnode
3953 	 * (the handle check).
3954 	 */
3955 	if (object != NULL && object->type == OBJT_VNODE &&
3956 	    object->handle == vp)
3957 		vnode_destroy_vobject(vp);
3958 
3959 	/*
3960 	 * Reclaim the vnode.
3961 	 */
3962 	if (VOP_RECLAIM(vp))
3963 		panic("vgone: cannot reclaim");
3964 	if (mp != NULL)
3965 		vn_finished_secondary_write(mp);
3966 	VNASSERT(vp->v_object == NULL, vp,
3967 	    ("vop_reclaim left v_object vp=%p", vp));
3968 	/*
3969 	 * Clear the advisory locks and wake up waiting threads.
3970 	 */
3971 	(void)VOP_ADVLOCKPURGE(vp);
3972 	vp->v_lockf = NULL;
3973 	/*
3974 	 * Delete from old mount point vnode list.
3975 	 */
3976 	delmntque(vp);
3977 	/*
3978 	 * Done with purge, reset to the standard lock and invalidate
3979 	 * the vnode.
3980 	 */
3981 	VI_LOCK(vp);
3982 	vp->v_vnlock = &vp->v_lock;
3983 	vp->v_op = &dead_vnodeops;
3984 	vp->v_type = VBAD;
3985 }
3986 
3987 /*
3988  * Print out a description of a vnode.
3989  */
3990 static const char * const typename[] =
3991 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3992  "VMARKER"};
3993 
3994 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
3995     "new hold count flag not added to vn_printf");
3996 
3997 void
3998 vn_printf(struct vnode *vp, const char *fmt, ...)
3999 {
4000 	va_list ap;
4001 	char buf[256], buf2[16];
4002 	u_long flags;
4003 	u_int holdcnt;
4004 
4005 	va_start(ap, fmt);
4006 	vprintf(fmt, ap);
4007 	va_end(ap);
4008 	printf("%p: ", (void *)vp);
4009 	printf("type %s\n", typename[vp->v_type]);
4010 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4011 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4012 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4013 	    vp->v_seqc_users);
4014 	switch (vp->v_type) {
4015 	case VDIR:
4016 		printf(" mountedhere %p\n", vp->v_mountedhere);
4017 		break;
4018 	case VCHR:
4019 		printf(" rdev %p\n", vp->v_rdev);
4020 		break;
4021 	case VSOCK:
4022 		printf(" socket %p\n", vp->v_unpcb);
4023 		break;
4024 	case VFIFO:
4025 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4026 		break;
4027 	default:
4028 		printf("\n");
4029 		break;
4030 	}
4031 	buf[0] = '\0';
4032 	buf[1] = '\0';
4033 	if (holdcnt & VHOLD_NO_SMR)
4034 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4035 	printf("    hold count flags (%s)\n", buf + 1);
4036 
4037 	buf[0] = '\0';
4038 	buf[1] = '\0';
4039 	if (vp->v_irflag & VIRF_DOOMED)
4040 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4041 	if (vp->v_irflag & VIRF_PGREAD)
4042 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4043 	flags = vp->v_irflag & ~(VIRF_DOOMED | VIRF_PGREAD);
4044 	if (flags != 0) {
4045 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4046 		strlcat(buf, buf2, sizeof(buf));
4047 	}
4048 	if (vp->v_vflag & VV_ROOT)
4049 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4050 	if (vp->v_vflag & VV_ISTTY)
4051 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4052 	if (vp->v_vflag & VV_NOSYNC)
4053 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4054 	if (vp->v_vflag & VV_ETERNALDEV)
4055 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4056 	if (vp->v_vflag & VV_CACHEDLABEL)
4057 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4058 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4059 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4060 	if (vp->v_vflag & VV_COPYONWRITE)
4061 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4062 	if (vp->v_vflag & VV_SYSTEM)
4063 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4064 	if (vp->v_vflag & VV_PROCDEP)
4065 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4066 	if (vp->v_vflag & VV_NOKNOTE)
4067 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4068 	if (vp->v_vflag & VV_DELETED)
4069 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4070 	if (vp->v_vflag & VV_MD)
4071 		strlcat(buf, "|VV_MD", sizeof(buf));
4072 	if (vp->v_vflag & VV_FORCEINSMQ)
4073 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4074 	if (vp->v_vflag & VV_READLINK)
4075 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4076 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4077 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4078 	    VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ |
4079 	    VV_READLINK);
4080 	if (flags != 0) {
4081 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4082 		strlcat(buf, buf2, sizeof(buf));
4083 	}
4084 	if (vp->v_iflag & VI_TEXT_REF)
4085 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4086 	if (vp->v_iflag & VI_MOUNT)
4087 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4088 	if (vp->v_iflag & VI_DOINGINACT)
4089 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4090 	if (vp->v_iflag & VI_OWEINACT)
4091 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4092 	if (vp->v_iflag & VI_DEFINACT)
4093 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4094 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4095 	    VI_OWEINACT | VI_DEFINACT);
4096 	if (flags != 0) {
4097 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4098 		strlcat(buf, buf2, sizeof(buf));
4099 	}
4100 	if (vp->v_mflag & VMP_LAZYLIST)
4101 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4102 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4103 	if (flags != 0) {
4104 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4105 		strlcat(buf, buf2, sizeof(buf));
4106 	}
4107 	printf("    flags (%s)", buf + 1);
4108 	if (mtx_owned(VI_MTX(vp)))
4109 		printf(" VI_LOCKed");
4110 	printf("\n");
4111 	if (vp->v_object != NULL)
4112 		printf("    v_object %p ref %d pages %d "
4113 		    "cleanbuf %d dirtybuf %d\n",
4114 		    vp->v_object, vp->v_object->ref_count,
4115 		    vp->v_object->resident_page_count,
4116 		    vp->v_bufobj.bo_clean.bv_cnt,
4117 		    vp->v_bufobj.bo_dirty.bv_cnt);
4118 	printf("    ");
4119 	lockmgr_printinfo(vp->v_vnlock);
4120 	if (vp->v_data != NULL)
4121 		VOP_PRINT(vp);
4122 }
4123 
4124 #ifdef DDB
4125 /*
4126  * List all of the locked vnodes in the system.
4127  * Called when debugging the kernel.
4128  */
4129 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4130 {
4131 	struct mount *mp;
4132 	struct vnode *vp;
4133 
4134 	/*
4135 	 * Note: because this is DDB, we can't obey the locking semantics
4136 	 * for these structures, which means we could catch an inconsistent
4137 	 * state and dereference a nasty pointer.  Not much to be done
4138 	 * about that.
4139 	 */
4140 	db_printf("Locked vnodes\n");
4141 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4142 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4143 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4144 				vn_printf(vp, "vnode ");
4145 		}
4146 	}
4147 }
4148 
4149 /*
4150  * Show details about the given vnode.
4151  */
4152 DB_SHOW_COMMAND(vnode, db_show_vnode)
4153 {
4154 	struct vnode *vp;
4155 
4156 	if (!have_addr)
4157 		return;
4158 	vp = (struct vnode *)addr;
4159 	vn_printf(vp, "vnode ");
4160 }
4161 
4162 /*
4163  * Show details about the given mount point.
4164  */
4165 DB_SHOW_COMMAND(mount, db_show_mount)
4166 {
4167 	struct mount *mp;
4168 	struct vfsopt *opt;
4169 	struct statfs *sp;
4170 	struct vnode *vp;
4171 	char buf[512];
4172 	uint64_t mflags;
4173 	u_int flags;
4174 
4175 	if (!have_addr) {
4176 		/* No address given, print short info about all mount points. */
4177 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4178 			db_printf("%p %s on %s (%s)\n", mp,
4179 			    mp->mnt_stat.f_mntfromname,
4180 			    mp->mnt_stat.f_mntonname,
4181 			    mp->mnt_stat.f_fstypename);
4182 			if (db_pager_quit)
4183 				break;
4184 		}
4185 		db_printf("\nMore info: show mount <addr>\n");
4186 		return;
4187 	}
4188 
4189 	mp = (struct mount *)addr;
4190 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4191 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4192 
4193 	buf[0] = '\0';
4194 	mflags = mp->mnt_flag;
4195 #define	MNT_FLAG(flag)	do {						\
4196 	if (mflags & (flag)) {						\
4197 		if (buf[0] != '\0')					\
4198 			strlcat(buf, ", ", sizeof(buf));		\
4199 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4200 		mflags &= ~(flag);					\
4201 	}								\
4202 } while (0)
4203 	MNT_FLAG(MNT_RDONLY);
4204 	MNT_FLAG(MNT_SYNCHRONOUS);
4205 	MNT_FLAG(MNT_NOEXEC);
4206 	MNT_FLAG(MNT_NOSUID);
4207 	MNT_FLAG(MNT_NFS4ACLS);
4208 	MNT_FLAG(MNT_UNION);
4209 	MNT_FLAG(MNT_ASYNC);
4210 	MNT_FLAG(MNT_SUIDDIR);
4211 	MNT_FLAG(MNT_SOFTDEP);
4212 	MNT_FLAG(MNT_NOSYMFOLLOW);
4213 	MNT_FLAG(MNT_GJOURNAL);
4214 	MNT_FLAG(MNT_MULTILABEL);
4215 	MNT_FLAG(MNT_ACLS);
4216 	MNT_FLAG(MNT_NOATIME);
4217 	MNT_FLAG(MNT_NOCLUSTERR);
4218 	MNT_FLAG(MNT_NOCLUSTERW);
4219 	MNT_FLAG(MNT_SUJ);
4220 	MNT_FLAG(MNT_EXRDONLY);
4221 	MNT_FLAG(MNT_EXPORTED);
4222 	MNT_FLAG(MNT_DEFEXPORTED);
4223 	MNT_FLAG(MNT_EXPORTANON);
4224 	MNT_FLAG(MNT_EXKERB);
4225 	MNT_FLAG(MNT_EXPUBLIC);
4226 	MNT_FLAG(MNT_LOCAL);
4227 	MNT_FLAG(MNT_QUOTA);
4228 	MNT_FLAG(MNT_ROOTFS);
4229 	MNT_FLAG(MNT_USER);
4230 	MNT_FLAG(MNT_IGNORE);
4231 	MNT_FLAG(MNT_UPDATE);
4232 	MNT_FLAG(MNT_DELEXPORT);
4233 	MNT_FLAG(MNT_RELOAD);
4234 	MNT_FLAG(MNT_FORCE);
4235 	MNT_FLAG(MNT_SNAPSHOT);
4236 	MNT_FLAG(MNT_BYFSID);
4237 #undef MNT_FLAG
4238 	if (mflags != 0) {
4239 		if (buf[0] != '\0')
4240 			strlcat(buf, ", ", sizeof(buf));
4241 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4242 		    "0x%016jx", mflags);
4243 	}
4244 	db_printf("    mnt_flag = %s\n", buf);
4245 
4246 	buf[0] = '\0';
4247 	flags = mp->mnt_kern_flag;
4248 #define	MNT_KERN_FLAG(flag)	do {					\
4249 	if (flags & (flag)) {						\
4250 		if (buf[0] != '\0')					\
4251 			strlcat(buf, ", ", sizeof(buf));		\
4252 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4253 		flags &= ~(flag);					\
4254 	}								\
4255 } while (0)
4256 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4257 	MNT_KERN_FLAG(MNTK_ASYNC);
4258 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4259 	MNT_KERN_FLAG(MNTK_DRAINING);
4260 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4261 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4262 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4263 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4264 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4265 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4266 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4267 	MNT_KERN_FLAG(MNTK_MARKER);
4268 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4269 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4270 	MNT_KERN_FLAG(MNTK_NOASYNC);
4271 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4272 	MNT_KERN_FLAG(MNTK_MWAIT);
4273 	MNT_KERN_FLAG(MNTK_SUSPEND);
4274 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4275 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4276 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4277 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4278 #undef MNT_KERN_FLAG
4279 	if (flags != 0) {
4280 		if (buf[0] != '\0')
4281 			strlcat(buf, ", ", sizeof(buf));
4282 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4283 		    "0x%08x", flags);
4284 	}
4285 	db_printf("    mnt_kern_flag = %s\n", buf);
4286 
4287 	db_printf("    mnt_opt = ");
4288 	opt = TAILQ_FIRST(mp->mnt_opt);
4289 	if (opt != NULL) {
4290 		db_printf("%s", opt->name);
4291 		opt = TAILQ_NEXT(opt, link);
4292 		while (opt != NULL) {
4293 			db_printf(", %s", opt->name);
4294 			opt = TAILQ_NEXT(opt, link);
4295 		}
4296 	}
4297 	db_printf("\n");
4298 
4299 	sp = &mp->mnt_stat;
4300 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4301 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4302 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4303 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4304 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4305 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4306 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4307 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4308 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4309 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4310 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4311 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4312 
4313 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4314 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4315 	if (jailed(mp->mnt_cred))
4316 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4317 	db_printf(" }\n");
4318 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4319 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4320 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4321 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4322 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4323 	    mp->mnt_lazyvnodelistsize);
4324 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4325 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4326 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4327 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4328 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4329 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4330 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4331 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4332 	db_printf("    mnt_secondary_accwrites = %d\n",
4333 	    mp->mnt_secondary_accwrites);
4334 	db_printf("    mnt_gjprovider = %s\n",
4335 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4336 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4337 
4338 	db_printf("\n\nList of active vnodes\n");
4339 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4340 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4341 			vn_printf(vp, "vnode ");
4342 			if (db_pager_quit)
4343 				break;
4344 		}
4345 	}
4346 	db_printf("\n\nList of inactive vnodes\n");
4347 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4348 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4349 			vn_printf(vp, "vnode ");
4350 			if (db_pager_quit)
4351 				break;
4352 		}
4353 	}
4354 }
4355 #endif	/* DDB */
4356 
4357 /*
4358  * Fill in a struct xvfsconf based on a struct vfsconf.
4359  */
4360 static int
4361 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4362 {
4363 	struct xvfsconf xvfsp;
4364 
4365 	bzero(&xvfsp, sizeof(xvfsp));
4366 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4367 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4368 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4369 	xvfsp.vfc_flags = vfsp->vfc_flags;
4370 	/*
4371 	 * These are unused in userland, we keep them
4372 	 * to not break binary compatibility.
4373 	 */
4374 	xvfsp.vfc_vfsops = NULL;
4375 	xvfsp.vfc_next = NULL;
4376 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4377 }
4378 
4379 #ifdef COMPAT_FREEBSD32
4380 struct xvfsconf32 {
4381 	uint32_t	vfc_vfsops;
4382 	char		vfc_name[MFSNAMELEN];
4383 	int32_t		vfc_typenum;
4384 	int32_t		vfc_refcount;
4385 	int32_t		vfc_flags;
4386 	uint32_t	vfc_next;
4387 };
4388 
4389 static int
4390 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4391 {
4392 	struct xvfsconf32 xvfsp;
4393 
4394 	bzero(&xvfsp, sizeof(xvfsp));
4395 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4396 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4397 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4398 	xvfsp.vfc_flags = vfsp->vfc_flags;
4399 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4400 }
4401 #endif
4402 
4403 /*
4404  * Top level filesystem related information gathering.
4405  */
4406 static int
4407 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4408 {
4409 	struct vfsconf *vfsp;
4410 	int error;
4411 
4412 	error = 0;
4413 	vfsconf_slock();
4414 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4415 #ifdef COMPAT_FREEBSD32
4416 		if (req->flags & SCTL_MASK32)
4417 			error = vfsconf2x32(req, vfsp);
4418 		else
4419 #endif
4420 			error = vfsconf2x(req, vfsp);
4421 		if (error)
4422 			break;
4423 	}
4424 	vfsconf_sunlock();
4425 	return (error);
4426 }
4427 
4428 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4429     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4430     "S,xvfsconf", "List of all configured filesystems");
4431 
4432 #ifndef BURN_BRIDGES
4433 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4434 
4435 static int
4436 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4437 {
4438 	int *name = (int *)arg1 - 1;	/* XXX */
4439 	u_int namelen = arg2 + 1;	/* XXX */
4440 	struct vfsconf *vfsp;
4441 
4442 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4443 	    "please rebuild world\n");
4444 
4445 #if 1 || defined(COMPAT_PRELITE2)
4446 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4447 	if (namelen == 1)
4448 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4449 #endif
4450 
4451 	switch (name[1]) {
4452 	case VFS_MAXTYPENUM:
4453 		if (namelen != 2)
4454 			return (ENOTDIR);
4455 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4456 	case VFS_CONF:
4457 		if (namelen != 3)
4458 			return (ENOTDIR);	/* overloaded */
4459 		vfsconf_slock();
4460 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4461 			if (vfsp->vfc_typenum == name[2])
4462 				break;
4463 		}
4464 		vfsconf_sunlock();
4465 		if (vfsp == NULL)
4466 			return (EOPNOTSUPP);
4467 #ifdef COMPAT_FREEBSD32
4468 		if (req->flags & SCTL_MASK32)
4469 			return (vfsconf2x32(req, vfsp));
4470 		else
4471 #endif
4472 			return (vfsconf2x(req, vfsp));
4473 	}
4474 	return (EOPNOTSUPP);
4475 }
4476 
4477 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4478     CTLFLAG_MPSAFE, vfs_sysctl,
4479     "Generic filesystem");
4480 
4481 #if 1 || defined(COMPAT_PRELITE2)
4482 
4483 static int
4484 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4485 {
4486 	int error;
4487 	struct vfsconf *vfsp;
4488 	struct ovfsconf ovfs;
4489 
4490 	vfsconf_slock();
4491 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4492 		bzero(&ovfs, sizeof(ovfs));
4493 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4494 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4495 		ovfs.vfc_index = vfsp->vfc_typenum;
4496 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4497 		ovfs.vfc_flags = vfsp->vfc_flags;
4498 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4499 		if (error != 0) {
4500 			vfsconf_sunlock();
4501 			return (error);
4502 		}
4503 	}
4504 	vfsconf_sunlock();
4505 	return (0);
4506 }
4507 
4508 #endif /* 1 || COMPAT_PRELITE2 */
4509 #endif /* !BURN_BRIDGES */
4510 
4511 #define KINFO_VNODESLOP		10
4512 #ifdef notyet
4513 /*
4514  * Dump vnode list (via sysctl).
4515  */
4516 /* ARGSUSED */
4517 static int
4518 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4519 {
4520 	struct xvnode *xvn;
4521 	struct mount *mp;
4522 	struct vnode *vp;
4523 	int error, len, n;
4524 
4525 	/*
4526 	 * Stale numvnodes access is not fatal here.
4527 	 */
4528 	req->lock = 0;
4529 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4530 	if (!req->oldptr)
4531 		/* Make an estimate */
4532 		return (SYSCTL_OUT(req, 0, len));
4533 
4534 	error = sysctl_wire_old_buffer(req, 0);
4535 	if (error != 0)
4536 		return (error);
4537 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4538 	n = 0;
4539 	mtx_lock(&mountlist_mtx);
4540 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4541 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4542 			continue;
4543 		MNT_ILOCK(mp);
4544 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4545 			if (n == len)
4546 				break;
4547 			vref(vp);
4548 			xvn[n].xv_size = sizeof *xvn;
4549 			xvn[n].xv_vnode = vp;
4550 			xvn[n].xv_id = 0;	/* XXX compat */
4551 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4552 			XV_COPY(usecount);
4553 			XV_COPY(writecount);
4554 			XV_COPY(holdcnt);
4555 			XV_COPY(mount);
4556 			XV_COPY(numoutput);
4557 			XV_COPY(type);
4558 #undef XV_COPY
4559 			xvn[n].xv_flag = vp->v_vflag;
4560 
4561 			switch (vp->v_type) {
4562 			case VREG:
4563 			case VDIR:
4564 			case VLNK:
4565 				break;
4566 			case VBLK:
4567 			case VCHR:
4568 				if (vp->v_rdev == NULL) {
4569 					vrele(vp);
4570 					continue;
4571 				}
4572 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4573 				break;
4574 			case VSOCK:
4575 				xvn[n].xv_socket = vp->v_socket;
4576 				break;
4577 			case VFIFO:
4578 				xvn[n].xv_fifo = vp->v_fifoinfo;
4579 				break;
4580 			case VNON:
4581 			case VBAD:
4582 			default:
4583 				/* shouldn't happen? */
4584 				vrele(vp);
4585 				continue;
4586 			}
4587 			vrele(vp);
4588 			++n;
4589 		}
4590 		MNT_IUNLOCK(mp);
4591 		mtx_lock(&mountlist_mtx);
4592 		vfs_unbusy(mp);
4593 		if (n == len)
4594 			break;
4595 	}
4596 	mtx_unlock(&mountlist_mtx);
4597 
4598 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4599 	free(xvn, M_TEMP);
4600 	return (error);
4601 }
4602 
4603 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4604     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4605     "");
4606 #endif
4607 
4608 static void
4609 unmount_or_warn(struct mount *mp)
4610 {
4611 	int error;
4612 
4613 	error = dounmount(mp, MNT_FORCE, curthread);
4614 	if (error != 0) {
4615 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4616 		if (error == EBUSY)
4617 			printf("BUSY)\n");
4618 		else
4619 			printf("%d)\n", error);
4620 	}
4621 }
4622 
4623 /*
4624  * Unmount all filesystems. The list is traversed in reverse order
4625  * of mounting to avoid dependencies.
4626  */
4627 void
4628 vfs_unmountall(void)
4629 {
4630 	struct mount *mp, *tmp;
4631 
4632 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4633 
4634 	/*
4635 	 * Since this only runs when rebooting, it is not interlocked.
4636 	 */
4637 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4638 		vfs_ref(mp);
4639 
4640 		/*
4641 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4642 		 * unmount of the latter.
4643 		 */
4644 		if (mp == rootdevmp)
4645 			continue;
4646 
4647 		unmount_or_warn(mp);
4648 	}
4649 
4650 	if (rootdevmp != NULL)
4651 		unmount_or_warn(rootdevmp);
4652 }
4653 
4654 static void
4655 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4656 {
4657 
4658 	ASSERT_VI_LOCKED(vp, __func__);
4659 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4660 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4661 		vdropl(vp);
4662 		return;
4663 	}
4664 	if (vn_lock(vp, lkflags) == 0) {
4665 		VI_LOCK(vp);
4666 		vinactive(vp);
4667 		VOP_UNLOCK(vp);
4668 		vdropl(vp);
4669 		return;
4670 	}
4671 	vdefer_inactive_unlocked(vp);
4672 }
4673 
4674 static int
4675 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4676 {
4677 
4678 	return (vp->v_iflag & VI_DEFINACT);
4679 }
4680 
4681 static void __noinline
4682 vfs_periodic_inactive(struct mount *mp, int flags)
4683 {
4684 	struct vnode *vp, *mvp;
4685 	int lkflags;
4686 
4687 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4688 	if (flags != MNT_WAIT)
4689 		lkflags |= LK_NOWAIT;
4690 
4691 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4692 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4693 			VI_UNLOCK(vp);
4694 			continue;
4695 		}
4696 		vp->v_iflag &= ~VI_DEFINACT;
4697 		vfs_deferred_inactive(vp, lkflags);
4698 	}
4699 }
4700 
4701 static inline bool
4702 vfs_want_msync(struct vnode *vp)
4703 {
4704 	struct vm_object *obj;
4705 
4706 	/*
4707 	 * This test may be performed without any locks held.
4708 	 * We rely on vm_object's type stability.
4709 	 */
4710 	if (vp->v_vflag & VV_NOSYNC)
4711 		return (false);
4712 	obj = vp->v_object;
4713 	return (obj != NULL && vm_object_mightbedirty(obj));
4714 }
4715 
4716 static int
4717 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4718 {
4719 
4720 	if (vp->v_vflag & VV_NOSYNC)
4721 		return (false);
4722 	if (vp->v_iflag & VI_DEFINACT)
4723 		return (true);
4724 	return (vfs_want_msync(vp));
4725 }
4726 
4727 static void __noinline
4728 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4729 {
4730 	struct vnode *vp, *mvp;
4731 	struct vm_object *obj;
4732 	int lkflags, objflags;
4733 	bool seen_defer;
4734 
4735 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4736 	if (flags != MNT_WAIT) {
4737 		lkflags |= LK_NOWAIT;
4738 		objflags = OBJPC_NOSYNC;
4739 	} else {
4740 		objflags = OBJPC_SYNC;
4741 	}
4742 
4743 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4744 		seen_defer = false;
4745 		if (vp->v_iflag & VI_DEFINACT) {
4746 			vp->v_iflag &= ~VI_DEFINACT;
4747 			seen_defer = true;
4748 		}
4749 		if (!vfs_want_msync(vp)) {
4750 			if (seen_defer)
4751 				vfs_deferred_inactive(vp, lkflags);
4752 			else
4753 				VI_UNLOCK(vp);
4754 			continue;
4755 		}
4756 		if (vget(vp, lkflags) == 0) {
4757 			obj = vp->v_object;
4758 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4759 				VM_OBJECT_WLOCK(obj);
4760 				vm_object_page_clean(obj, 0, 0, objflags);
4761 				VM_OBJECT_WUNLOCK(obj);
4762 			}
4763 			vput(vp);
4764 			if (seen_defer)
4765 				vdrop(vp);
4766 		} else {
4767 			if (seen_defer)
4768 				vdefer_inactive_unlocked(vp);
4769 		}
4770 	}
4771 }
4772 
4773 void
4774 vfs_periodic(struct mount *mp, int flags)
4775 {
4776 
4777 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4778 
4779 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4780 		vfs_periodic_inactive(mp, flags);
4781 	else
4782 		vfs_periodic_msync_inactive(mp, flags);
4783 }
4784 
4785 static void
4786 destroy_vpollinfo_free(struct vpollinfo *vi)
4787 {
4788 
4789 	knlist_destroy(&vi->vpi_selinfo.si_note);
4790 	mtx_destroy(&vi->vpi_lock);
4791 	free(vi, M_VNODEPOLL);
4792 }
4793 
4794 static void
4795 destroy_vpollinfo(struct vpollinfo *vi)
4796 {
4797 
4798 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4799 	seldrain(&vi->vpi_selinfo);
4800 	destroy_vpollinfo_free(vi);
4801 }
4802 
4803 /*
4804  * Initialize per-vnode helper structure to hold poll-related state.
4805  */
4806 void
4807 v_addpollinfo(struct vnode *vp)
4808 {
4809 	struct vpollinfo *vi;
4810 
4811 	if (vp->v_pollinfo != NULL)
4812 		return;
4813 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4814 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4815 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4816 	    vfs_knlunlock, vfs_knl_assert_lock);
4817 	VI_LOCK(vp);
4818 	if (vp->v_pollinfo != NULL) {
4819 		VI_UNLOCK(vp);
4820 		destroy_vpollinfo_free(vi);
4821 		return;
4822 	}
4823 	vp->v_pollinfo = vi;
4824 	VI_UNLOCK(vp);
4825 }
4826 
4827 /*
4828  * Record a process's interest in events which might happen to
4829  * a vnode.  Because poll uses the historic select-style interface
4830  * internally, this routine serves as both the ``check for any
4831  * pending events'' and the ``record my interest in future events''
4832  * functions.  (These are done together, while the lock is held,
4833  * to avoid race conditions.)
4834  */
4835 int
4836 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4837 {
4838 
4839 	v_addpollinfo(vp);
4840 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4841 	if (vp->v_pollinfo->vpi_revents & events) {
4842 		/*
4843 		 * This leaves events we are not interested
4844 		 * in available for the other process which
4845 		 * which presumably had requested them
4846 		 * (otherwise they would never have been
4847 		 * recorded).
4848 		 */
4849 		events &= vp->v_pollinfo->vpi_revents;
4850 		vp->v_pollinfo->vpi_revents &= ~events;
4851 
4852 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4853 		return (events);
4854 	}
4855 	vp->v_pollinfo->vpi_events |= events;
4856 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4857 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4858 	return (0);
4859 }
4860 
4861 /*
4862  * Routine to create and manage a filesystem syncer vnode.
4863  */
4864 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4865 static int	sync_fsync(struct  vop_fsync_args *);
4866 static int	sync_inactive(struct  vop_inactive_args *);
4867 static int	sync_reclaim(struct  vop_reclaim_args *);
4868 
4869 static struct vop_vector sync_vnodeops = {
4870 	.vop_bypass =	VOP_EOPNOTSUPP,
4871 	.vop_close =	sync_close,		/* close */
4872 	.vop_fsync =	sync_fsync,		/* fsync */
4873 	.vop_inactive =	sync_inactive,	/* inactive */
4874 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4875 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4876 	.vop_lock1 =	vop_stdlock,	/* lock */
4877 	.vop_unlock =	vop_stdunlock,	/* unlock */
4878 	.vop_islocked =	vop_stdislocked,	/* islocked */
4879 };
4880 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4881 
4882 /*
4883  * Create a new filesystem syncer vnode for the specified mount point.
4884  */
4885 void
4886 vfs_allocate_syncvnode(struct mount *mp)
4887 {
4888 	struct vnode *vp;
4889 	struct bufobj *bo;
4890 	static long start, incr, next;
4891 	int error;
4892 
4893 	/* Allocate a new vnode */
4894 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4895 	if (error != 0)
4896 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4897 	vp->v_type = VNON;
4898 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4899 	vp->v_vflag |= VV_FORCEINSMQ;
4900 	error = insmntque(vp, mp);
4901 	if (error != 0)
4902 		panic("vfs_allocate_syncvnode: insmntque() failed");
4903 	vp->v_vflag &= ~VV_FORCEINSMQ;
4904 	VOP_UNLOCK(vp);
4905 	/*
4906 	 * Place the vnode onto the syncer worklist. We attempt to
4907 	 * scatter them about on the list so that they will go off
4908 	 * at evenly distributed times even if all the filesystems
4909 	 * are mounted at once.
4910 	 */
4911 	next += incr;
4912 	if (next == 0 || next > syncer_maxdelay) {
4913 		start /= 2;
4914 		incr /= 2;
4915 		if (start == 0) {
4916 			start = syncer_maxdelay / 2;
4917 			incr = syncer_maxdelay;
4918 		}
4919 		next = start;
4920 	}
4921 	bo = &vp->v_bufobj;
4922 	BO_LOCK(bo);
4923 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4924 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4925 	mtx_lock(&sync_mtx);
4926 	sync_vnode_count++;
4927 	if (mp->mnt_syncer == NULL) {
4928 		mp->mnt_syncer = vp;
4929 		vp = NULL;
4930 	}
4931 	mtx_unlock(&sync_mtx);
4932 	BO_UNLOCK(bo);
4933 	if (vp != NULL) {
4934 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4935 		vgone(vp);
4936 		vput(vp);
4937 	}
4938 }
4939 
4940 void
4941 vfs_deallocate_syncvnode(struct mount *mp)
4942 {
4943 	struct vnode *vp;
4944 
4945 	mtx_lock(&sync_mtx);
4946 	vp = mp->mnt_syncer;
4947 	if (vp != NULL)
4948 		mp->mnt_syncer = NULL;
4949 	mtx_unlock(&sync_mtx);
4950 	if (vp != NULL)
4951 		vrele(vp);
4952 }
4953 
4954 /*
4955  * Do a lazy sync of the filesystem.
4956  */
4957 static int
4958 sync_fsync(struct vop_fsync_args *ap)
4959 {
4960 	struct vnode *syncvp = ap->a_vp;
4961 	struct mount *mp = syncvp->v_mount;
4962 	int error, save;
4963 	struct bufobj *bo;
4964 
4965 	/*
4966 	 * We only need to do something if this is a lazy evaluation.
4967 	 */
4968 	if (ap->a_waitfor != MNT_LAZY)
4969 		return (0);
4970 
4971 	/*
4972 	 * Move ourselves to the back of the sync list.
4973 	 */
4974 	bo = &syncvp->v_bufobj;
4975 	BO_LOCK(bo);
4976 	vn_syncer_add_to_worklist(bo, syncdelay);
4977 	BO_UNLOCK(bo);
4978 
4979 	/*
4980 	 * Walk the list of vnodes pushing all that are dirty and
4981 	 * not already on the sync list.
4982 	 */
4983 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4984 		return (0);
4985 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4986 		vfs_unbusy(mp);
4987 		return (0);
4988 	}
4989 	save = curthread_pflags_set(TDP_SYNCIO);
4990 	/*
4991 	 * The filesystem at hand may be idle with free vnodes stored in the
4992 	 * batch.  Return them instead of letting them stay there indefinitely.
4993 	 */
4994 	vfs_periodic(mp, MNT_NOWAIT);
4995 	error = VFS_SYNC(mp, MNT_LAZY);
4996 	curthread_pflags_restore(save);
4997 	vn_finished_write(mp);
4998 	vfs_unbusy(mp);
4999 	return (error);
5000 }
5001 
5002 /*
5003  * The syncer vnode is no referenced.
5004  */
5005 static int
5006 sync_inactive(struct vop_inactive_args *ap)
5007 {
5008 
5009 	vgone(ap->a_vp);
5010 	return (0);
5011 }
5012 
5013 /*
5014  * The syncer vnode is no longer needed and is being decommissioned.
5015  *
5016  * Modifications to the worklist must be protected by sync_mtx.
5017  */
5018 static int
5019 sync_reclaim(struct vop_reclaim_args *ap)
5020 {
5021 	struct vnode *vp = ap->a_vp;
5022 	struct bufobj *bo;
5023 
5024 	bo = &vp->v_bufobj;
5025 	BO_LOCK(bo);
5026 	mtx_lock(&sync_mtx);
5027 	if (vp->v_mount->mnt_syncer == vp)
5028 		vp->v_mount->mnt_syncer = NULL;
5029 	if (bo->bo_flag & BO_ONWORKLST) {
5030 		LIST_REMOVE(bo, bo_synclist);
5031 		syncer_worklist_len--;
5032 		sync_vnode_count--;
5033 		bo->bo_flag &= ~BO_ONWORKLST;
5034 	}
5035 	mtx_unlock(&sync_mtx);
5036 	BO_UNLOCK(bo);
5037 
5038 	return (0);
5039 }
5040 
5041 int
5042 vn_need_pageq_flush(struct vnode *vp)
5043 {
5044 	struct vm_object *obj;
5045 	int need;
5046 
5047 	MPASS(mtx_owned(VI_MTX(vp)));
5048 	need = 0;
5049 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5050 	    vm_object_mightbedirty(obj))
5051 		need = 1;
5052 	return (need);
5053 }
5054 
5055 /*
5056  * Check if vnode represents a disk device
5057  */
5058 bool
5059 vn_isdisk_error(struct vnode *vp, int *errp)
5060 {
5061 	int error;
5062 
5063 	if (vp->v_type != VCHR) {
5064 		error = ENOTBLK;
5065 		goto out;
5066 	}
5067 	error = 0;
5068 	dev_lock();
5069 	if (vp->v_rdev == NULL)
5070 		error = ENXIO;
5071 	else if (vp->v_rdev->si_devsw == NULL)
5072 		error = ENXIO;
5073 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5074 		error = ENOTBLK;
5075 	dev_unlock();
5076 out:
5077 	*errp = error;
5078 	return (error == 0);
5079 }
5080 
5081 bool
5082 vn_isdisk(struct vnode *vp)
5083 {
5084 	int error;
5085 
5086 	return (vn_isdisk_error(vp, &error));
5087 }
5088 
5089 /*
5090  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5091  * the comment above cache_fplookup for details.
5092  */
5093 int
5094 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5095 {
5096 	int error;
5097 
5098 	VFS_SMR_ASSERT_ENTERED();
5099 
5100 	/* Check the owner. */
5101 	if (cred->cr_uid == file_uid) {
5102 		if (file_mode & S_IXUSR)
5103 			return (0);
5104 		goto out_error;
5105 	}
5106 
5107 	/* Otherwise, check the groups (first match) */
5108 	if (groupmember(file_gid, cred)) {
5109 		if (file_mode & S_IXGRP)
5110 			return (0);
5111 		goto out_error;
5112 	}
5113 
5114 	/* Otherwise, check everyone else. */
5115 	if (file_mode & S_IXOTH)
5116 		return (0);
5117 out_error:
5118 	/*
5119 	 * Permission check failed, but it is possible denial will get overwritten
5120 	 * (e.g., when root is traversing through a 700 directory owned by someone
5121 	 * else).
5122 	 *
5123 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5124 	 * modules overriding this result. It's quite unclear what semantics
5125 	 * are allowed for them to operate, thus for safety we don't call them
5126 	 * from within the SMR section. This also means if any such modules
5127 	 * are present, we have to let the regular lookup decide.
5128 	 */
5129 	error = priv_check_cred_vfs_lookup_nomac(cred);
5130 	switch (error) {
5131 	case 0:
5132 		return (0);
5133 	case EAGAIN:
5134 		/*
5135 		 * MAC modules present.
5136 		 */
5137 		return (EAGAIN);
5138 	case EPERM:
5139 		return (EACCES);
5140 	default:
5141 		return (error);
5142 	}
5143 }
5144 
5145 /*
5146  * Common filesystem object access control check routine.  Accepts a
5147  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5148  * Returns 0 on success, or an errno on failure.
5149  */
5150 int
5151 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5152     accmode_t accmode, struct ucred *cred)
5153 {
5154 	accmode_t dac_granted;
5155 	accmode_t priv_granted;
5156 
5157 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5158 	    ("invalid bit in accmode"));
5159 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5160 	    ("VAPPEND without VWRITE"));
5161 
5162 	/*
5163 	 * Look for a normal, non-privileged way to access the file/directory
5164 	 * as requested.  If it exists, go with that.
5165 	 */
5166 
5167 	dac_granted = 0;
5168 
5169 	/* Check the owner. */
5170 	if (cred->cr_uid == file_uid) {
5171 		dac_granted |= VADMIN;
5172 		if (file_mode & S_IXUSR)
5173 			dac_granted |= VEXEC;
5174 		if (file_mode & S_IRUSR)
5175 			dac_granted |= VREAD;
5176 		if (file_mode & S_IWUSR)
5177 			dac_granted |= (VWRITE | VAPPEND);
5178 
5179 		if ((accmode & dac_granted) == accmode)
5180 			return (0);
5181 
5182 		goto privcheck;
5183 	}
5184 
5185 	/* Otherwise, check the groups (first match) */
5186 	if (groupmember(file_gid, cred)) {
5187 		if (file_mode & S_IXGRP)
5188 			dac_granted |= VEXEC;
5189 		if (file_mode & S_IRGRP)
5190 			dac_granted |= VREAD;
5191 		if (file_mode & S_IWGRP)
5192 			dac_granted |= (VWRITE | VAPPEND);
5193 
5194 		if ((accmode & dac_granted) == accmode)
5195 			return (0);
5196 
5197 		goto privcheck;
5198 	}
5199 
5200 	/* Otherwise, check everyone else. */
5201 	if (file_mode & S_IXOTH)
5202 		dac_granted |= VEXEC;
5203 	if (file_mode & S_IROTH)
5204 		dac_granted |= VREAD;
5205 	if (file_mode & S_IWOTH)
5206 		dac_granted |= (VWRITE | VAPPEND);
5207 	if ((accmode & dac_granted) == accmode)
5208 		return (0);
5209 
5210 privcheck:
5211 	/*
5212 	 * Build a privilege mask to determine if the set of privileges
5213 	 * satisfies the requirements when combined with the granted mask
5214 	 * from above.  For each privilege, if the privilege is required,
5215 	 * bitwise or the request type onto the priv_granted mask.
5216 	 */
5217 	priv_granted = 0;
5218 
5219 	if (type == VDIR) {
5220 		/*
5221 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5222 		 * requests, instead of PRIV_VFS_EXEC.
5223 		 */
5224 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5225 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5226 			priv_granted |= VEXEC;
5227 	} else {
5228 		/*
5229 		 * Ensure that at least one execute bit is on. Otherwise,
5230 		 * a privileged user will always succeed, and we don't want
5231 		 * this to happen unless the file really is executable.
5232 		 */
5233 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5234 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5235 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5236 			priv_granted |= VEXEC;
5237 	}
5238 
5239 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5240 	    !priv_check_cred(cred, PRIV_VFS_READ))
5241 		priv_granted |= VREAD;
5242 
5243 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5244 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5245 		priv_granted |= (VWRITE | VAPPEND);
5246 
5247 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5248 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5249 		priv_granted |= VADMIN;
5250 
5251 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5252 		return (0);
5253 	}
5254 
5255 	return ((accmode & VADMIN) ? EPERM : EACCES);
5256 }
5257 
5258 /*
5259  * Credential check based on process requesting service, and per-attribute
5260  * permissions.
5261  */
5262 int
5263 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5264     struct thread *td, accmode_t accmode)
5265 {
5266 
5267 	/*
5268 	 * Kernel-invoked always succeeds.
5269 	 */
5270 	if (cred == NOCRED)
5271 		return (0);
5272 
5273 	/*
5274 	 * Do not allow privileged processes in jail to directly manipulate
5275 	 * system attributes.
5276 	 */
5277 	switch (attrnamespace) {
5278 	case EXTATTR_NAMESPACE_SYSTEM:
5279 		/* Potentially should be: return (EPERM); */
5280 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5281 	case EXTATTR_NAMESPACE_USER:
5282 		return (VOP_ACCESS(vp, accmode, cred, td));
5283 	default:
5284 		return (EPERM);
5285 	}
5286 }
5287 
5288 #ifdef DEBUG_VFS_LOCKS
5289 /*
5290  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5291  * no longer ok to have an unlocked VFS.
5292  */
5293 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5294 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5295 
5296 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5297 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5298     "Drop into debugger on lock violation");
5299 
5300 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5301 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5302     0, "Check for interlock across VOPs");
5303 
5304 int vfs_badlock_print = 1;	/* Print lock violations. */
5305 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5306     0, "Print lock violations");
5307 
5308 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5309 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5310     0, "Print vnode details on lock violations");
5311 
5312 #ifdef KDB
5313 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5314 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5315     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5316 #endif
5317 
5318 static void
5319 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5320 {
5321 
5322 #ifdef KDB
5323 	if (vfs_badlock_backtrace)
5324 		kdb_backtrace();
5325 #endif
5326 	if (vfs_badlock_vnode)
5327 		vn_printf(vp, "vnode ");
5328 	if (vfs_badlock_print)
5329 		printf("%s: %p %s\n", str, (void *)vp, msg);
5330 	if (vfs_badlock_ddb)
5331 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5332 }
5333 
5334 void
5335 assert_vi_locked(struct vnode *vp, const char *str)
5336 {
5337 
5338 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5339 		vfs_badlock("interlock is not locked but should be", str, vp);
5340 }
5341 
5342 void
5343 assert_vi_unlocked(struct vnode *vp, const char *str)
5344 {
5345 
5346 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5347 		vfs_badlock("interlock is locked but should not be", str, vp);
5348 }
5349 
5350 void
5351 assert_vop_locked(struct vnode *vp, const char *str)
5352 {
5353 	int locked;
5354 
5355 	if (!IGNORE_LOCK(vp)) {
5356 		locked = VOP_ISLOCKED(vp);
5357 		if (locked == 0 || locked == LK_EXCLOTHER)
5358 			vfs_badlock("is not locked but should be", str, vp);
5359 	}
5360 }
5361 
5362 void
5363 assert_vop_unlocked(struct vnode *vp, const char *str)
5364 {
5365 
5366 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5367 		vfs_badlock("is locked but should not be", str, vp);
5368 }
5369 
5370 void
5371 assert_vop_elocked(struct vnode *vp, const char *str)
5372 {
5373 
5374 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5375 		vfs_badlock("is not exclusive locked but should be", str, vp);
5376 }
5377 #endif /* DEBUG_VFS_LOCKS */
5378 
5379 void
5380 vop_rename_fail(struct vop_rename_args *ap)
5381 {
5382 
5383 	if (ap->a_tvp != NULL)
5384 		vput(ap->a_tvp);
5385 	if (ap->a_tdvp == ap->a_tvp)
5386 		vrele(ap->a_tdvp);
5387 	else
5388 		vput(ap->a_tdvp);
5389 	vrele(ap->a_fdvp);
5390 	vrele(ap->a_fvp);
5391 }
5392 
5393 void
5394 vop_rename_pre(void *ap)
5395 {
5396 	struct vop_rename_args *a = ap;
5397 
5398 #ifdef DEBUG_VFS_LOCKS
5399 	if (a->a_tvp)
5400 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5401 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5402 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5403 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5404 
5405 	/* Check the source (from). */
5406 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5407 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5408 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5409 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5410 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5411 
5412 	/* Check the target. */
5413 	if (a->a_tvp)
5414 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5415 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5416 #endif
5417 	/*
5418 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5419 	 * in vop_rename_post but that's not going to work out since some
5420 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5421 	 *
5422 	 * For now filesystems are expected to do the relevant calls after they
5423 	 * decide what vnodes to operate on.
5424 	 */
5425 	if (a->a_tdvp != a->a_fdvp)
5426 		vhold(a->a_fdvp);
5427 	if (a->a_tvp != a->a_fvp)
5428 		vhold(a->a_fvp);
5429 	vhold(a->a_tdvp);
5430 	if (a->a_tvp)
5431 		vhold(a->a_tvp);
5432 }
5433 
5434 #ifdef DEBUG_VFS_LOCKS
5435 void
5436 vop_fplookup_vexec_debugpre(void *ap __unused)
5437 {
5438 
5439 	VFS_SMR_ASSERT_ENTERED();
5440 }
5441 
5442 void
5443 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5444 {
5445 
5446 	VFS_SMR_ASSERT_ENTERED();
5447 }
5448 
5449 void
5450 vop_strategy_debugpre(void *ap)
5451 {
5452 	struct vop_strategy_args *a;
5453 	struct buf *bp;
5454 
5455 	a = ap;
5456 	bp = a->a_bp;
5457 
5458 	/*
5459 	 * Cluster ops lock their component buffers but not the IO container.
5460 	 */
5461 	if ((bp->b_flags & B_CLUSTER) != 0)
5462 		return;
5463 
5464 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5465 		if (vfs_badlock_print)
5466 			printf(
5467 			    "VOP_STRATEGY: bp is not locked but should be\n");
5468 		if (vfs_badlock_ddb)
5469 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5470 	}
5471 }
5472 
5473 void
5474 vop_lock_debugpre(void *ap)
5475 {
5476 	struct vop_lock1_args *a = ap;
5477 
5478 	if ((a->a_flags & LK_INTERLOCK) == 0)
5479 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5480 	else
5481 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5482 }
5483 
5484 void
5485 vop_lock_debugpost(void *ap, int rc)
5486 {
5487 	struct vop_lock1_args *a = ap;
5488 
5489 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5490 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5491 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5492 }
5493 
5494 void
5495 vop_unlock_debugpre(void *ap)
5496 {
5497 	struct vop_unlock_args *a = ap;
5498 
5499 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5500 }
5501 
5502 void
5503 vop_need_inactive_debugpre(void *ap)
5504 {
5505 	struct vop_need_inactive_args *a = ap;
5506 
5507 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5508 }
5509 
5510 void
5511 vop_need_inactive_debugpost(void *ap, int rc)
5512 {
5513 	struct vop_need_inactive_args *a = ap;
5514 
5515 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5516 }
5517 #endif
5518 
5519 void
5520 vop_create_pre(void *ap)
5521 {
5522 	struct vop_create_args *a;
5523 	struct vnode *dvp;
5524 
5525 	a = ap;
5526 	dvp = a->a_dvp;
5527 	vn_seqc_write_begin(dvp);
5528 }
5529 
5530 void
5531 vop_create_post(void *ap, int rc)
5532 {
5533 	struct vop_create_args *a;
5534 	struct vnode *dvp;
5535 
5536 	a = ap;
5537 	dvp = a->a_dvp;
5538 	vn_seqc_write_end(dvp);
5539 	if (!rc)
5540 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5541 }
5542 
5543 void
5544 vop_whiteout_pre(void *ap)
5545 {
5546 	struct vop_whiteout_args *a;
5547 	struct vnode *dvp;
5548 
5549 	a = ap;
5550 	dvp = a->a_dvp;
5551 	vn_seqc_write_begin(dvp);
5552 }
5553 
5554 void
5555 vop_whiteout_post(void *ap, int rc)
5556 {
5557 	struct vop_whiteout_args *a;
5558 	struct vnode *dvp;
5559 
5560 	a = ap;
5561 	dvp = a->a_dvp;
5562 	vn_seqc_write_end(dvp);
5563 }
5564 
5565 void
5566 vop_deleteextattr_pre(void *ap)
5567 {
5568 	struct vop_deleteextattr_args *a;
5569 	struct vnode *vp;
5570 
5571 	a = ap;
5572 	vp = a->a_vp;
5573 	vn_seqc_write_begin(vp);
5574 }
5575 
5576 void
5577 vop_deleteextattr_post(void *ap, int rc)
5578 {
5579 	struct vop_deleteextattr_args *a;
5580 	struct vnode *vp;
5581 
5582 	a = ap;
5583 	vp = a->a_vp;
5584 	vn_seqc_write_end(vp);
5585 	if (!rc)
5586 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5587 }
5588 
5589 void
5590 vop_link_pre(void *ap)
5591 {
5592 	struct vop_link_args *a;
5593 	struct vnode *vp, *tdvp;
5594 
5595 	a = ap;
5596 	vp = a->a_vp;
5597 	tdvp = a->a_tdvp;
5598 	vn_seqc_write_begin(vp);
5599 	vn_seqc_write_begin(tdvp);
5600 }
5601 
5602 void
5603 vop_link_post(void *ap, int rc)
5604 {
5605 	struct vop_link_args *a;
5606 	struct vnode *vp, *tdvp;
5607 
5608 	a = ap;
5609 	vp = a->a_vp;
5610 	tdvp = a->a_tdvp;
5611 	vn_seqc_write_end(vp);
5612 	vn_seqc_write_end(tdvp);
5613 	if (!rc) {
5614 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5615 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5616 	}
5617 }
5618 
5619 void
5620 vop_mkdir_pre(void *ap)
5621 {
5622 	struct vop_mkdir_args *a;
5623 	struct vnode *dvp;
5624 
5625 	a = ap;
5626 	dvp = a->a_dvp;
5627 	vn_seqc_write_begin(dvp);
5628 }
5629 
5630 void
5631 vop_mkdir_post(void *ap, int rc)
5632 {
5633 	struct vop_mkdir_args *a;
5634 	struct vnode *dvp;
5635 
5636 	a = ap;
5637 	dvp = a->a_dvp;
5638 	vn_seqc_write_end(dvp);
5639 	if (!rc)
5640 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5641 }
5642 
5643 #ifdef DEBUG_VFS_LOCKS
5644 void
5645 vop_mkdir_debugpost(void *ap, int rc)
5646 {
5647 	struct vop_mkdir_args *a;
5648 
5649 	a = ap;
5650 	if (!rc)
5651 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5652 }
5653 #endif
5654 
5655 void
5656 vop_mknod_pre(void *ap)
5657 {
5658 	struct vop_mknod_args *a;
5659 	struct vnode *dvp;
5660 
5661 	a = ap;
5662 	dvp = a->a_dvp;
5663 	vn_seqc_write_begin(dvp);
5664 }
5665 
5666 void
5667 vop_mknod_post(void *ap, int rc)
5668 {
5669 	struct vop_mknod_args *a;
5670 	struct vnode *dvp;
5671 
5672 	a = ap;
5673 	dvp = a->a_dvp;
5674 	vn_seqc_write_end(dvp);
5675 	if (!rc)
5676 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5677 }
5678 
5679 void
5680 vop_reclaim_post(void *ap, int rc)
5681 {
5682 	struct vop_reclaim_args *a;
5683 	struct vnode *vp;
5684 
5685 	a = ap;
5686 	vp = a->a_vp;
5687 	ASSERT_VOP_IN_SEQC(vp);
5688 	if (!rc)
5689 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5690 }
5691 
5692 void
5693 vop_remove_pre(void *ap)
5694 {
5695 	struct vop_remove_args *a;
5696 	struct vnode *dvp, *vp;
5697 
5698 	a = ap;
5699 	dvp = a->a_dvp;
5700 	vp = a->a_vp;
5701 	vn_seqc_write_begin(dvp);
5702 	vn_seqc_write_begin(vp);
5703 }
5704 
5705 void
5706 vop_remove_post(void *ap, int rc)
5707 {
5708 	struct vop_remove_args *a;
5709 	struct vnode *dvp, *vp;
5710 
5711 	a = ap;
5712 	dvp = a->a_dvp;
5713 	vp = a->a_vp;
5714 	vn_seqc_write_end(dvp);
5715 	vn_seqc_write_end(vp);
5716 	if (!rc) {
5717 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5718 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5719 	}
5720 }
5721 
5722 void
5723 vop_rename_post(void *ap, int rc)
5724 {
5725 	struct vop_rename_args *a = ap;
5726 	long hint;
5727 
5728 	if (!rc) {
5729 		hint = NOTE_WRITE;
5730 		if (a->a_fdvp == a->a_tdvp) {
5731 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5732 				hint |= NOTE_LINK;
5733 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5734 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5735 		} else {
5736 			hint |= NOTE_EXTEND;
5737 			if (a->a_fvp->v_type == VDIR)
5738 				hint |= NOTE_LINK;
5739 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5740 
5741 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5742 			    a->a_tvp->v_type == VDIR)
5743 				hint &= ~NOTE_LINK;
5744 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5745 		}
5746 
5747 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5748 		if (a->a_tvp)
5749 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5750 	}
5751 	if (a->a_tdvp != a->a_fdvp)
5752 		vdrop(a->a_fdvp);
5753 	if (a->a_tvp != a->a_fvp)
5754 		vdrop(a->a_fvp);
5755 	vdrop(a->a_tdvp);
5756 	if (a->a_tvp)
5757 		vdrop(a->a_tvp);
5758 }
5759 
5760 void
5761 vop_rmdir_pre(void *ap)
5762 {
5763 	struct vop_rmdir_args *a;
5764 	struct vnode *dvp, *vp;
5765 
5766 	a = ap;
5767 	dvp = a->a_dvp;
5768 	vp = a->a_vp;
5769 	vn_seqc_write_begin(dvp);
5770 	vn_seqc_write_begin(vp);
5771 }
5772 
5773 void
5774 vop_rmdir_post(void *ap, int rc)
5775 {
5776 	struct vop_rmdir_args *a;
5777 	struct vnode *dvp, *vp;
5778 
5779 	a = ap;
5780 	dvp = a->a_dvp;
5781 	vp = a->a_vp;
5782 	vn_seqc_write_end(dvp);
5783 	vn_seqc_write_end(vp);
5784 	if (!rc) {
5785 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5786 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5787 	}
5788 }
5789 
5790 void
5791 vop_setattr_pre(void *ap)
5792 {
5793 	struct vop_setattr_args *a;
5794 	struct vnode *vp;
5795 
5796 	a = ap;
5797 	vp = a->a_vp;
5798 	vn_seqc_write_begin(vp);
5799 }
5800 
5801 void
5802 vop_setattr_post(void *ap, int rc)
5803 {
5804 	struct vop_setattr_args *a;
5805 	struct vnode *vp;
5806 
5807 	a = ap;
5808 	vp = a->a_vp;
5809 	vn_seqc_write_end(vp);
5810 	if (!rc)
5811 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5812 }
5813 
5814 void
5815 vop_setacl_pre(void *ap)
5816 {
5817 	struct vop_setacl_args *a;
5818 	struct vnode *vp;
5819 
5820 	a = ap;
5821 	vp = a->a_vp;
5822 	vn_seqc_write_begin(vp);
5823 }
5824 
5825 void
5826 vop_setacl_post(void *ap, int rc __unused)
5827 {
5828 	struct vop_setacl_args *a;
5829 	struct vnode *vp;
5830 
5831 	a = ap;
5832 	vp = a->a_vp;
5833 	vn_seqc_write_end(vp);
5834 }
5835 
5836 void
5837 vop_setextattr_pre(void *ap)
5838 {
5839 	struct vop_setextattr_args *a;
5840 	struct vnode *vp;
5841 
5842 	a = ap;
5843 	vp = a->a_vp;
5844 	vn_seqc_write_begin(vp);
5845 }
5846 
5847 void
5848 vop_setextattr_post(void *ap, int rc)
5849 {
5850 	struct vop_setextattr_args *a;
5851 	struct vnode *vp;
5852 
5853 	a = ap;
5854 	vp = a->a_vp;
5855 	vn_seqc_write_end(vp);
5856 	if (!rc)
5857 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5858 }
5859 
5860 void
5861 vop_symlink_pre(void *ap)
5862 {
5863 	struct vop_symlink_args *a;
5864 	struct vnode *dvp;
5865 
5866 	a = ap;
5867 	dvp = a->a_dvp;
5868 	vn_seqc_write_begin(dvp);
5869 }
5870 
5871 void
5872 vop_symlink_post(void *ap, int rc)
5873 {
5874 	struct vop_symlink_args *a;
5875 	struct vnode *dvp;
5876 
5877 	a = ap;
5878 	dvp = a->a_dvp;
5879 	vn_seqc_write_end(dvp);
5880 	if (!rc)
5881 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5882 }
5883 
5884 void
5885 vop_open_post(void *ap, int rc)
5886 {
5887 	struct vop_open_args *a = ap;
5888 
5889 	if (!rc)
5890 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5891 }
5892 
5893 void
5894 vop_close_post(void *ap, int rc)
5895 {
5896 	struct vop_close_args *a = ap;
5897 
5898 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5899 	    !VN_IS_DOOMED(a->a_vp))) {
5900 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5901 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5902 	}
5903 }
5904 
5905 void
5906 vop_read_post(void *ap, int rc)
5907 {
5908 	struct vop_read_args *a = ap;
5909 
5910 	if (!rc)
5911 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5912 }
5913 
5914 void
5915 vop_read_pgcache_post(void *ap, int rc)
5916 {
5917 	struct vop_read_pgcache_args *a = ap;
5918 
5919 	if (!rc)
5920 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
5921 }
5922 
5923 void
5924 vop_readdir_post(void *ap, int rc)
5925 {
5926 	struct vop_readdir_args *a = ap;
5927 
5928 	if (!rc)
5929 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5930 }
5931 
5932 static struct knlist fs_knlist;
5933 
5934 static void
5935 vfs_event_init(void *arg)
5936 {
5937 	knlist_init_mtx(&fs_knlist, NULL);
5938 }
5939 /* XXX - correct order? */
5940 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5941 
5942 void
5943 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5944 {
5945 
5946 	KNOTE_UNLOCKED(&fs_knlist, event);
5947 }
5948 
5949 static int	filt_fsattach(struct knote *kn);
5950 static void	filt_fsdetach(struct knote *kn);
5951 static int	filt_fsevent(struct knote *kn, long hint);
5952 
5953 struct filterops fs_filtops = {
5954 	.f_isfd = 0,
5955 	.f_attach = filt_fsattach,
5956 	.f_detach = filt_fsdetach,
5957 	.f_event = filt_fsevent
5958 };
5959 
5960 static int
5961 filt_fsattach(struct knote *kn)
5962 {
5963 
5964 	kn->kn_flags |= EV_CLEAR;
5965 	knlist_add(&fs_knlist, kn, 0);
5966 	return (0);
5967 }
5968 
5969 static void
5970 filt_fsdetach(struct knote *kn)
5971 {
5972 
5973 	knlist_remove(&fs_knlist, kn, 0);
5974 }
5975 
5976 static int
5977 filt_fsevent(struct knote *kn, long hint)
5978 {
5979 
5980 	kn->kn_fflags |= hint;
5981 	return (kn->kn_fflags != 0);
5982 }
5983 
5984 static int
5985 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5986 {
5987 	struct vfsidctl vc;
5988 	int error;
5989 	struct mount *mp;
5990 
5991 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5992 	if (error)
5993 		return (error);
5994 	if (vc.vc_vers != VFS_CTL_VERS1)
5995 		return (EINVAL);
5996 	mp = vfs_getvfs(&vc.vc_fsid);
5997 	if (mp == NULL)
5998 		return (ENOENT);
5999 	/* ensure that a specific sysctl goes to the right filesystem. */
6000 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6001 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6002 		vfs_rel(mp);
6003 		return (EINVAL);
6004 	}
6005 	VCTLTOREQ(&vc, req);
6006 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6007 	vfs_rel(mp);
6008 	return (error);
6009 }
6010 
6011 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6012     NULL, 0, sysctl_vfs_ctl, "",
6013     "Sysctl by fsid");
6014 
6015 /*
6016  * Function to initialize a va_filerev field sensibly.
6017  * XXX: Wouldn't a random number make a lot more sense ??
6018  */
6019 u_quad_t
6020 init_va_filerev(void)
6021 {
6022 	struct bintime bt;
6023 
6024 	getbinuptime(&bt);
6025 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6026 }
6027 
6028 static int	filt_vfsread(struct knote *kn, long hint);
6029 static int	filt_vfswrite(struct knote *kn, long hint);
6030 static int	filt_vfsvnode(struct knote *kn, long hint);
6031 static void	filt_vfsdetach(struct knote *kn);
6032 static struct filterops vfsread_filtops = {
6033 	.f_isfd = 1,
6034 	.f_detach = filt_vfsdetach,
6035 	.f_event = filt_vfsread
6036 };
6037 static struct filterops vfswrite_filtops = {
6038 	.f_isfd = 1,
6039 	.f_detach = filt_vfsdetach,
6040 	.f_event = filt_vfswrite
6041 };
6042 static struct filterops vfsvnode_filtops = {
6043 	.f_isfd = 1,
6044 	.f_detach = filt_vfsdetach,
6045 	.f_event = filt_vfsvnode
6046 };
6047 
6048 static void
6049 vfs_knllock(void *arg)
6050 {
6051 	struct vnode *vp = arg;
6052 
6053 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6054 }
6055 
6056 static void
6057 vfs_knlunlock(void *arg)
6058 {
6059 	struct vnode *vp = arg;
6060 
6061 	VOP_UNLOCK(vp);
6062 }
6063 
6064 static void
6065 vfs_knl_assert_lock(void *arg, int what)
6066 {
6067 #ifdef DEBUG_VFS_LOCKS
6068 	struct vnode *vp = arg;
6069 
6070 	if (what == LA_LOCKED)
6071 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6072 	else
6073 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6074 #endif
6075 }
6076 
6077 int
6078 vfs_kqfilter(struct vop_kqfilter_args *ap)
6079 {
6080 	struct vnode *vp = ap->a_vp;
6081 	struct knote *kn = ap->a_kn;
6082 	struct knlist *knl;
6083 
6084 	switch (kn->kn_filter) {
6085 	case EVFILT_READ:
6086 		kn->kn_fop = &vfsread_filtops;
6087 		break;
6088 	case EVFILT_WRITE:
6089 		kn->kn_fop = &vfswrite_filtops;
6090 		break;
6091 	case EVFILT_VNODE:
6092 		kn->kn_fop = &vfsvnode_filtops;
6093 		break;
6094 	default:
6095 		return (EINVAL);
6096 	}
6097 
6098 	kn->kn_hook = (caddr_t)vp;
6099 
6100 	v_addpollinfo(vp);
6101 	if (vp->v_pollinfo == NULL)
6102 		return (ENOMEM);
6103 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6104 	vhold(vp);
6105 	knlist_add(knl, kn, 0);
6106 
6107 	return (0);
6108 }
6109 
6110 /*
6111  * Detach knote from vnode
6112  */
6113 static void
6114 filt_vfsdetach(struct knote *kn)
6115 {
6116 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6117 
6118 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6119 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6120 	vdrop(vp);
6121 }
6122 
6123 /*ARGSUSED*/
6124 static int
6125 filt_vfsread(struct knote *kn, long hint)
6126 {
6127 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6128 	struct vattr va;
6129 	int res;
6130 
6131 	/*
6132 	 * filesystem is gone, so set the EOF flag and schedule
6133 	 * the knote for deletion.
6134 	 */
6135 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6136 		VI_LOCK(vp);
6137 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6138 		VI_UNLOCK(vp);
6139 		return (1);
6140 	}
6141 
6142 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6143 		return (0);
6144 
6145 	VI_LOCK(vp);
6146 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6147 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6148 	VI_UNLOCK(vp);
6149 	return (res);
6150 }
6151 
6152 /*ARGSUSED*/
6153 static int
6154 filt_vfswrite(struct knote *kn, long hint)
6155 {
6156 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6157 
6158 	VI_LOCK(vp);
6159 
6160 	/*
6161 	 * filesystem is gone, so set the EOF flag and schedule
6162 	 * the knote for deletion.
6163 	 */
6164 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6165 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6166 
6167 	kn->kn_data = 0;
6168 	VI_UNLOCK(vp);
6169 	return (1);
6170 }
6171 
6172 static int
6173 filt_vfsvnode(struct knote *kn, long hint)
6174 {
6175 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6176 	int res;
6177 
6178 	VI_LOCK(vp);
6179 	if (kn->kn_sfflags & hint)
6180 		kn->kn_fflags |= hint;
6181 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6182 		kn->kn_flags |= EV_EOF;
6183 		VI_UNLOCK(vp);
6184 		return (1);
6185 	}
6186 	res = (kn->kn_fflags != 0);
6187 	VI_UNLOCK(vp);
6188 	return (res);
6189 }
6190 
6191 /*
6192  * Returns whether the directory is empty or not.
6193  * If it is empty, the return value is 0; otherwise
6194  * the return value is an error value (which may
6195  * be ENOTEMPTY).
6196  */
6197 int
6198 vfs_emptydir(struct vnode *vp)
6199 {
6200 	struct uio uio;
6201 	struct iovec iov;
6202 	struct dirent *dirent, *dp, *endp;
6203 	int error, eof;
6204 
6205 	error = 0;
6206 	eof = 0;
6207 
6208 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6209 
6210 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6211 	iov.iov_base = dirent;
6212 	iov.iov_len = sizeof(struct dirent);
6213 
6214 	uio.uio_iov = &iov;
6215 	uio.uio_iovcnt = 1;
6216 	uio.uio_offset = 0;
6217 	uio.uio_resid = sizeof(struct dirent);
6218 	uio.uio_segflg = UIO_SYSSPACE;
6219 	uio.uio_rw = UIO_READ;
6220 	uio.uio_td = curthread;
6221 
6222 	while (eof == 0 && error == 0) {
6223 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6224 		    NULL, NULL);
6225 		if (error != 0)
6226 			break;
6227 		endp = (void *)((uint8_t *)dirent +
6228 		    sizeof(struct dirent) - uio.uio_resid);
6229 		for (dp = dirent; dp < endp;
6230 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6231 			if (dp->d_type == DT_WHT)
6232 				continue;
6233 			if (dp->d_namlen == 0)
6234 				continue;
6235 			if (dp->d_type != DT_DIR &&
6236 			    dp->d_type != DT_UNKNOWN) {
6237 				error = ENOTEMPTY;
6238 				break;
6239 			}
6240 			if (dp->d_namlen > 2) {
6241 				error = ENOTEMPTY;
6242 				break;
6243 			}
6244 			if (dp->d_namlen == 1 &&
6245 			    dp->d_name[0] != '.') {
6246 				error = ENOTEMPTY;
6247 				break;
6248 			}
6249 			if (dp->d_namlen == 2 &&
6250 			    dp->d_name[1] != '.') {
6251 				error = ENOTEMPTY;
6252 				break;
6253 			}
6254 			uio.uio_resid = sizeof(struct dirent);
6255 		}
6256 	}
6257 	free(dirent, M_TEMP);
6258 	return (error);
6259 }
6260 
6261 int
6262 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6263 {
6264 	int error;
6265 
6266 	if (dp->d_reclen > ap->a_uio->uio_resid)
6267 		return (ENAMETOOLONG);
6268 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6269 	if (error) {
6270 		if (ap->a_ncookies != NULL) {
6271 			if (ap->a_cookies != NULL)
6272 				free(ap->a_cookies, M_TEMP);
6273 			ap->a_cookies = NULL;
6274 			*ap->a_ncookies = 0;
6275 		}
6276 		return (error);
6277 	}
6278 	if (ap->a_ncookies == NULL)
6279 		return (0);
6280 
6281 	KASSERT(ap->a_cookies,
6282 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6283 
6284 	*ap->a_cookies = realloc(*ap->a_cookies,
6285 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6286 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6287 	*ap->a_ncookies += 1;
6288 	return (0);
6289 }
6290 
6291 /*
6292  * The purpose of this routine is to remove granularity from accmode_t,
6293  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6294  * VADMIN and VAPPEND.
6295  *
6296  * If it returns 0, the caller is supposed to continue with the usual
6297  * access checks using 'accmode' as modified by this routine.  If it
6298  * returns nonzero value, the caller is supposed to return that value
6299  * as errno.
6300  *
6301  * Note that after this routine runs, accmode may be zero.
6302  */
6303 int
6304 vfs_unixify_accmode(accmode_t *accmode)
6305 {
6306 	/*
6307 	 * There is no way to specify explicit "deny" rule using
6308 	 * file mode or POSIX.1e ACLs.
6309 	 */
6310 	if (*accmode & VEXPLICIT_DENY) {
6311 		*accmode = 0;
6312 		return (0);
6313 	}
6314 
6315 	/*
6316 	 * None of these can be translated into usual access bits.
6317 	 * Also, the common case for NFSv4 ACLs is to not contain
6318 	 * either of these bits. Caller should check for VWRITE
6319 	 * on the containing directory instead.
6320 	 */
6321 	if (*accmode & (VDELETE_CHILD | VDELETE))
6322 		return (EPERM);
6323 
6324 	if (*accmode & VADMIN_PERMS) {
6325 		*accmode &= ~VADMIN_PERMS;
6326 		*accmode |= VADMIN;
6327 	}
6328 
6329 	/*
6330 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6331 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6332 	 */
6333 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6334 
6335 	return (0);
6336 }
6337 
6338 /*
6339  * Clear out a doomed vnode (if any) and replace it with a new one as long
6340  * as the fs is not being unmounted. Return the root vnode to the caller.
6341  */
6342 static int __noinline
6343 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6344 {
6345 	struct vnode *vp;
6346 	int error;
6347 
6348 restart:
6349 	if (mp->mnt_rootvnode != NULL) {
6350 		MNT_ILOCK(mp);
6351 		vp = mp->mnt_rootvnode;
6352 		if (vp != NULL) {
6353 			if (!VN_IS_DOOMED(vp)) {
6354 				vrefact(vp);
6355 				MNT_IUNLOCK(mp);
6356 				error = vn_lock(vp, flags);
6357 				if (error == 0) {
6358 					*vpp = vp;
6359 					return (0);
6360 				}
6361 				vrele(vp);
6362 				goto restart;
6363 			}
6364 			/*
6365 			 * Clear the old one.
6366 			 */
6367 			mp->mnt_rootvnode = NULL;
6368 		}
6369 		MNT_IUNLOCK(mp);
6370 		if (vp != NULL) {
6371 			vfs_op_barrier_wait(mp);
6372 			vrele(vp);
6373 		}
6374 	}
6375 	error = VFS_CACHEDROOT(mp, flags, vpp);
6376 	if (error != 0)
6377 		return (error);
6378 	if (mp->mnt_vfs_ops == 0) {
6379 		MNT_ILOCK(mp);
6380 		if (mp->mnt_vfs_ops != 0) {
6381 			MNT_IUNLOCK(mp);
6382 			return (0);
6383 		}
6384 		if (mp->mnt_rootvnode == NULL) {
6385 			vrefact(*vpp);
6386 			mp->mnt_rootvnode = *vpp;
6387 		} else {
6388 			if (mp->mnt_rootvnode != *vpp) {
6389 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6390 					panic("%s: mismatch between vnode returned "
6391 					    " by VFS_CACHEDROOT and the one cached "
6392 					    " (%p != %p)",
6393 					    __func__, *vpp, mp->mnt_rootvnode);
6394 				}
6395 			}
6396 		}
6397 		MNT_IUNLOCK(mp);
6398 	}
6399 	return (0);
6400 }
6401 
6402 int
6403 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6404 {
6405 	struct mount_pcpu *mpcpu;
6406 	struct vnode *vp;
6407 	int error;
6408 
6409 	if (!vfs_op_thread_enter(mp, mpcpu))
6410 		return (vfs_cache_root_fallback(mp, flags, vpp));
6411 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6412 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6413 		vfs_op_thread_exit(mp, mpcpu);
6414 		return (vfs_cache_root_fallback(mp, flags, vpp));
6415 	}
6416 	vrefact(vp);
6417 	vfs_op_thread_exit(mp, mpcpu);
6418 	error = vn_lock(vp, flags);
6419 	if (error != 0) {
6420 		vrele(vp);
6421 		return (vfs_cache_root_fallback(mp, flags, vpp));
6422 	}
6423 	*vpp = vp;
6424 	return (0);
6425 }
6426 
6427 struct vnode *
6428 vfs_cache_root_clear(struct mount *mp)
6429 {
6430 	struct vnode *vp;
6431 
6432 	/*
6433 	 * ops > 0 guarantees there is nobody who can see this vnode
6434 	 */
6435 	MPASS(mp->mnt_vfs_ops > 0);
6436 	vp = mp->mnt_rootvnode;
6437 	if (vp != NULL)
6438 		vn_seqc_write_begin(vp);
6439 	mp->mnt_rootvnode = NULL;
6440 	return (vp);
6441 }
6442 
6443 void
6444 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6445 {
6446 
6447 	MPASS(mp->mnt_vfs_ops > 0);
6448 	vrefact(vp);
6449 	mp->mnt_rootvnode = vp;
6450 }
6451 
6452 /*
6453  * These are helper functions for filesystems to traverse all
6454  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6455  *
6456  * This interface replaces MNT_VNODE_FOREACH.
6457  */
6458 
6459 struct vnode *
6460 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6461 {
6462 	struct vnode *vp;
6463 
6464 	if (should_yield())
6465 		kern_yield(PRI_USER);
6466 	MNT_ILOCK(mp);
6467 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6468 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6469 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6470 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6471 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6472 			continue;
6473 		VI_LOCK(vp);
6474 		if (VN_IS_DOOMED(vp)) {
6475 			VI_UNLOCK(vp);
6476 			continue;
6477 		}
6478 		break;
6479 	}
6480 	if (vp == NULL) {
6481 		__mnt_vnode_markerfree_all(mvp, mp);
6482 		/* MNT_IUNLOCK(mp); -- done in above function */
6483 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6484 		return (NULL);
6485 	}
6486 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6487 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6488 	MNT_IUNLOCK(mp);
6489 	return (vp);
6490 }
6491 
6492 struct vnode *
6493 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6494 {
6495 	struct vnode *vp;
6496 
6497 	*mvp = vn_alloc_marker(mp);
6498 	MNT_ILOCK(mp);
6499 	MNT_REF(mp);
6500 
6501 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6502 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6503 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6504 			continue;
6505 		VI_LOCK(vp);
6506 		if (VN_IS_DOOMED(vp)) {
6507 			VI_UNLOCK(vp);
6508 			continue;
6509 		}
6510 		break;
6511 	}
6512 	if (vp == NULL) {
6513 		MNT_REL(mp);
6514 		MNT_IUNLOCK(mp);
6515 		vn_free_marker(*mvp);
6516 		*mvp = NULL;
6517 		return (NULL);
6518 	}
6519 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6520 	MNT_IUNLOCK(mp);
6521 	return (vp);
6522 }
6523 
6524 void
6525 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6526 {
6527 
6528 	if (*mvp == NULL) {
6529 		MNT_IUNLOCK(mp);
6530 		return;
6531 	}
6532 
6533 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6534 
6535 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6536 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6537 	MNT_REL(mp);
6538 	MNT_IUNLOCK(mp);
6539 	vn_free_marker(*mvp);
6540 	*mvp = NULL;
6541 }
6542 
6543 /*
6544  * These are helper functions for filesystems to traverse their
6545  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6546  */
6547 static void
6548 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6549 {
6550 
6551 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6552 
6553 	MNT_ILOCK(mp);
6554 	MNT_REL(mp);
6555 	MNT_IUNLOCK(mp);
6556 	vn_free_marker(*mvp);
6557 	*mvp = NULL;
6558 }
6559 
6560 /*
6561  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6562  * conventional lock order during mnt_vnode_next_lazy iteration.
6563  *
6564  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6565  * The list lock is dropped and reacquired.  On success, both locks are held.
6566  * On failure, the mount vnode list lock is held but the vnode interlock is
6567  * not, and the procedure may have yielded.
6568  */
6569 static bool
6570 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6571     struct vnode *vp)
6572 {
6573 
6574 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6575 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6576 	    ("%s: bad marker", __func__));
6577 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6578 	    ("%s: inappropriate vnode", __func__));
6579 	ASSERT_VI_UNLOCKED(vp, __func__);
6580 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6581 
6582 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6583 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6584 
6585 	/*
6586 	 * Note we may be racing against vdrop which transitioned the hold
6587 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6588 	 * if we are the only user after we get the interlock we will just
6589 	 * vdrop.
6590 	 */
6591 	vhold(vp);
6592 	mtx_unlock(&mp->mnt_listmtx);
6593 	VI_LOCK(vp);
6594 	if (VN_IS_DOOMED(vp)) {
6595 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6596 		goto out_lost;
6597 	}
6598 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6599 	/*
6600 	 * There is nothing to do if we are the last user.
6601 	 */
6602 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6603 		goto out_lost;
6604 	mtx_lock(&mp->mnt_listmtx);
6605 	return (true);
6606 out_lost:
6607 	vdropl(vp);
6608 	maybe_yield();
6609 	mtx_lock(&mp->mnt_listmtx);
6610 	return (false);
6611 }
6612 
6613 static struct vnode *
6614 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6615     void *cbarg)
6616 {
6617 	struct vnode *vp;
6618 
6619 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6620 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6621 restart:
6622 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6623 	while (vp != NULL) {
6624 		if (vp->v_type == VMARKER) {
6625 			vp = TAILQ_NEXT(vp, v_lazylist);
6626 			continue;
6627 		}
6628 		/*
6629 		 * See if we want to process the vnode. Note we may encounter a
6630 		 * long string of vnodes we don't care about and hog the list
6631 		 * as a result. Check for it and requeue the marker.
6632 		 */
6633 		VNPASS(!VN_IS_DOOMED(vp), vp);
6634 		if (!cb(vp, cbarg)) {
6635 			if (!should_yield()) {
6636 				vp = TAILQ_NEXT(vp, v_lazylist);
6637 				continue;
6638 			}
6639 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6640 			    v_lazylist);
6641 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6642 			    v_lazylist);
6643 			mtx_unlock(&mp->mnt_listmtx);
6644 			kern_yield(PRI_USER);
6645 			mtx_lock(&mp->mnt_listmtx);
6646 			goto restart;
6647 		}
6648 		/*
6649 		 * Try-lock because this is the wrong lock order.
6650 		 */
6651 		if (!VI_TRYLOCK(vp) &&
6652 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6653 			goto restart;
6654 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6655 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6656 		    ("alien vnode on the lazy list %p %p", vp, mp));
6657 		VNPASS(vp->v_mount == mp, vp);
6658 		VNPASS(!VN_IS_DOOMED(vp), vp);
6659 		break;
6660 	}
6661 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6662 
6663 	/* Check if we are done */
6664 	if (vp == NULL) {
6665 		mtx_unlock(&mp->mnt_listmtx);
6666 		mnt_vnode_markerfree_lazy(mvp, mp);
6667 		return (NULL);
6668 	}
6669 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6670 	mtx_unlock(&mp->mnt_listmtx);
6671 	ASSERT_VI_LOCKED(vp, "lazy iter");
6672 	return (vp);
6673 }
6674 
6675 struct vnode *
6676 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6677     void *cbarg)
6678 {
6679 
6680 	if (should_yield())
6681 		kern_yield(PRI_USER);
6682 	mtx_lock(&mp->mnt_listmtx);
6683 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6684 }
6685 
6686 struct vnode *
6687 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6688     void *cbarg)
6689 {
6690 	struct vnode *vp;
6691 
6692 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6693 		return (NULL);
6694 
6695 	*mvp = vn_alloc_marker(mp);
6696 	MNT_ILOCK(mp);
6697 	MNT_REF(mp);
6698 	MNT_IUNLOCK(mp);
6699 
6700 	mtx_lock(&mp->mnt_listmtx);
6701 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6702 	if (vp == NULL) {
6703 		mtx_unlock(&mp->mnt_listmtx);
6704 		mnt_vnode_markerfree_lazy(mvp, mp);
6705 		return (NULL);
6706 	}
6707 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6708 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6709 }
6710 
6711 void
6712 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6713 {
6714 
6715 	if (*mvp == NULL)
6716 		return;
6717 
6718 	mtx_lock(&mp->mnt_listmtx);
6719 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6720 	mtx_unlock(&mp->mnt_listmtx);
6721 	mnt_vnode_markerfree_lazy(mvp, mp);
6722 }
6723 
6724 int
6725 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6726 {
6727 
6728 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6729 		cnp->cn_flags &= ~NOEXECCHECK;
6730 		return (0);
6731 	}
6732 
6733 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6734 }
6735 
6736 /*
6737  * Do not use this variant unless you have means other than the hold count
6738  * to prevent the vnode from getting freed.
6739  */
6740 void
6741 vn_seqc_write_begin_unheld_locked(struct vnode *vp)
6742 {
6743 
6744 	ASSERT_VI_LOCKED(vp, __func__);
6745 	VNPASS(vp->v_seqc_users >= 0, vp);
6746 	vp->v_seqc_users++;
6747 	if (vp->v_seqc_users == 1)
6748 		seqc_sleepable_write_begin(&vp->v_seqc);
6749 }
6750 
6751 void
6752 vn_seqc_write_begin_locked(struct vnode *vp)
6753 {
6754 
6755 	ASSERT_VI_LOCKED(vp, __func__);
6756 	VNPASS(vp->v_holdcnt > 0, vp);
6757 	vn_seqc_write_begin_unheld_locked(vp);
6758 }
6759 
6760 void
6761 vn_seqc_write_begin(struct vnode *vp)
6762 {
6763 
6764 	VI_LOCK(vp);
6765 	vn_seqc_write_begin_locked(vp);
6766 	VI_UNLOCK(vp);
6767 }
6768 
6769 void
6770 vn_seqc_write_begin_unheld(struct vnode *vp)
6771 {
6772 
6773 	VI_LOCK(vp);
6774 	vn_seqc_write_begin_unheld_locked(vp);
6775 	VI_UNLOCK(vp);
6776 }
6777 
6778 void
6779 vn_seqc_write_end_locked(struct vnode *vp)
6780 {
6781 
6782 	ASSERT_VI_LOCKED(vp, __func__);
6783 	VNPASS(vp->v_seqc_users > 0, vp);
6784 	vp->v_seqc_users--;
6785 	if (vp->v_seqc_users == 0)
6786 		seqc_sleepable_write_end(&vp->v_seqc);
6787 }
6788 
6789 void
6790 vn_seqc_write_end(struct vnode *vp)
6791 {
6792 
6793 	VI_LOCK(vp);
6794 	vn_seqc_write_end_locked(vp);
6795 	VI_UNLOCK(vp);
6796 }
6797