xref: /freebsd/sys/kern/vfs_subr.c (revision 41ce62251c1ed0003fc13b8735de5f9eff4c5c03)
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 #ifdef INVARIANTS
1820 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1821 	vp->v_op = NULL;
1822 #endif
1823 	vp->v_mountedhere = NULL;
1824 	vp->v_unpcb = NULL;
1825 	vp->v_rdev = NULL;
1826 	vp->v_fifoinfo = NULL;
1827 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1828 	vp->v_irflag = 0;
1829 	vp->v_iflag = 0;
1830 	vp->v_vflag = 0;
1831 	bo->bo_flag = 0;
1832 	vn_free(vp);
1833 }
1834 
1835 /*
1836  * Delete from old mount point vnode list, if on one.
1837  */
1838 static void
1839 delmntque(struct vnode *vp)
1840 {
1841 	struct mount *mp;
1842 
1843 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1844 
1845 	mp = vp->v_mount;
1846 	if (mp == NULL)
1847 		return;
1848 	MNT_ILOCK(mp);
1849 	VI_LOCK(vp);
1850 	vp->v_mount = NULL;
1851 	VI_UNLOCK(vp);
1852 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1853 		("bad mount point vnode list size"));
1854 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1855 	mp->mnt_nvnodelistsize--;
1856 	MNT_REL(mp);
1857 	MNT_IUNLOCK(mp);
1858 }
1859 
1860 static void
1861 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1862 {
1863 
1864 	vp->v_data = NULL;
1865 	vp->v_op = &dead_vnodeops;
1866 	vgone(vp);
1867 	vput(vp);
1868 }
1869 
1870 /*
1871  * Insert into list of vnodes for the new mount point, if available.
1872  */
1873 int
1874 insmntque1(struct vnode *vp, struct mount *mp,
1875 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1876 {
1877 
1878 	KASSERT(vp->v_mount == NULL,
1879 		("insmntque: vnode already on per mount vnode list"));
1880 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1881 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1882 
1883 	/*
1884 	 * We acquire the vnode interlock early to ensure that the
1885 	 * vnode cannot be recycled by another process releasing a
1886 	 * holdcnt on it before we get it on both the vnode list
1887 	 * and the active vnode list. The mount mutex protects only
1888 	 * manipulation of the vnode list and the vnode freelist
1889 	 * mutex protects only manipulation of the active vnode list.
1890 	 * Hence the need to hold the vnode interlock throughout.
1891 	 */
1892 	MNT_ILOCK(mp);
1893 	VI_LOCK(vp);
1894 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1895 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1896 	    mp->mnt_nvnodelistsize == 0)) &&
1897 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1898 		VI_UNLOCK(vp);
1899 		MNT_IUNLOCK(mp);
1900 		if (dtr != NULL)
1901 			dtr(vp, dtr_arg);
1902 		return (EBUSY);
1903 	}
1904 	vp->v_mount = mp;
1905 	MNT_REF(mp);
1906 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1907 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1908 		("neg mount point vnode list size"));
1909 	mp->mnt_nvnodelistsize++;
1910 	VI_UNLOCK(vp);
1911 	MNT_IUNLOCK(mp);
1912 	return (0);
1913 }
1914 
1915 int
1916 insmntque(struct vnode *vp, struct mount *mp)
1917 {
1918 
1919 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1920 }
1921 
1922 /*
1923  * Flush out and invalidate all buffers associated with a bufobj
1924  * Called with the underlying object locked.
1925  */
1926 int
1927 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1928 {
1929 	int error;
1930 
1931 	BO_LOCK(bo);
1932 	if (flags & V_SAVE) {
1933 		error = bufobj_wwait(bo, slpflag, slptimeo);
1934 		if (error) {
1935 			BO_UNLOCK(bo);
1936 			return (error);
1937 		}
1938 		if (bo->bo_dirty.bv_cnt > 0) {
1939 			BO_UNLOCK(bo);
1940 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1941 				return (error);
1942 			/*
1943 			 * XXX We could save a lock/unlock if this was only
1944 			 * enabled under INVARIANTS
1945 			 */
1946 			BO_LOCK(bo);
1947 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1948 				panic("vinvalbuf: dirty bufs");
1949 		}
1950 	}
1951 	/*
1952 	 * If you alter this loop please notice that interlock is dropped and
1953 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1954 	 * no race conditions occur from this.
1955 	 */
1956 	do {
1957 		error = flushbuflist(&bo->bo_clean,
1958 		    flags, bo, slpflag, slptimeo);
1959 		if (error == 0 && !(flags & V_CLEANONLY))
1960 			error = flushbuflist(&bo->bo_dirty,
1961 			    flags, bo, slpflag, slptimeo);
1962 		if (error != 0 && error != EAGAIN) {
1963 			BO_UNLOCK(bo);
1964 			return (error);
1965 		}
1966 	} while (error != 0);
1967 
1968 	/*
1969 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1970 	 * have write I/O in-progress but if there is a VM object then the
1971 	 * VM object can also have read-I/O in-progress.
1972 	 */
1973 	do {
1974 		bufobj_wwait(bo, 0, 0);
1975 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1976 			BO_UNLOCK(bo);
1977 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1978 			BO_LOCK(bo);
1979 		}
1980 	} while (bo->bo_numoutput > 0);
1981 	BO_UNLOCK(bo);
1982 
1983 	/*
1984 	 * Destroy the copy in the VM cache, too.
1985 	 */
1986 	if (bo->bo_object != NULL &&
1987 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1988 		VM_OBJECT_WLOCK(bo->bo_object);
1989 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1990 		    OBJPR_CLEANONLY : 0);
1991 		VM_OBJECT_WUNLOCK(bo->bo_object);
1992 	}
1993 
1994 #ifdef INVARIANTS
1995 	BO_LOCK(bo);
1996 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1997 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1998 	    bo->bo_clean.bv_cnt > 0))
1999 		panic("vinvalbuf: flush failed");
2000 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2001 	    bo->bo_dirty.bv_cnt > 0)
2002 		panic("vinvalbuf: flush dirty failed");
2003 	BO_UNLOCK(bo);
2004 #endif
2005 	return (0);
2006 }
2007 
2008 /*
2009  * Flush out and invalidate all buffers associated with a vnode.
2010  * Called with the underlying object locked.
2011  */
2012 int
2013 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2014 {
2015 
2016 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2017 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2018 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2019 		return (0);
2020 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2021 }
2022 
2023 /*
2024  * Flush out buffers on the specified list.
2025  *
2026  */
2027 static int
2028 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2029     int slptimeo)
2030 {
2031 	struct buf *bp, *nbp;
2032 	int retval, error;
2033 	daddr_t lblkno;
2034 	b_xflags_t xflags;
2035 
2036 	ASSERT_BO_WLOCKED(bo);
2037 
2038 	retval = 0;
2039 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2040 		/*
2041 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2042 		 * do not skip any buffers. If we are flushing only V_NORMAL
2043 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2044 		 * flushing only V_ALT buffers then skip buffers not marked
2045 		 * as BX_ALTDATA.
2046 		 */
2047 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2048 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2049 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2050 			continue;
2051 		}
2052 		if (nbp != NULL) {
2053 			lblkno = nbp->b_lblkno;
2054 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2055 		}
2056 		retval = EAGAIN;
2057 		error = BUF_TIMELOCK(bp,
2058 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2059 		    "flushbuf", slpflag, slptimeo);
2060 		if (error) {
2061 			BO_LOCK(bo);
2062 			return (error != ENOLCK ? error : EAGAIN);
2063 		}
2064 		KASSERT(bp->b_bufobj == bo,
2065 		    ("bp %p wrong b_bufobj %p should be %p",
2066 		    bp, bp->b_bufobj, bo));
2067 		/*
2068 		 * XXX Since there are no node locks for NFS, I
2069 		 * believe there is a slight chance that a delayed
2070 		 * write will occur while sleeping just above, so
2071 		 * check for it.
2072 		 */
2073 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2074 		    (flags & V_SAVE)) {
2075 			bremfree(bp);
2076 			bp->b_flags |= B_ASYNC;
2077 			bwrite(bp);
2078 			BO_LOCK(bo);
2079 			return (EAGAIN);	/* XXX: why not loop ? */
2080 		}
2081 		bremfree(bp);
2082 		bp->b_flags |= (B_INVAL | B_RELBUF);
2083 		bp->b_flags &= ~B_ASYNC;
2084 		brelse(bp);
2085 		BO_LOCK(bo);
2086 		if (nbp == NULL)
2087 			break;
2088 		nbp = gbincore(bo, lblkno);
2089 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2090 		    != xflags)
2091 			break;			/* nbp invalid */
2092 	}
2093 	return (retval);
2094 }
2095 
2096 int
2097 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2098 {
2099 	struct buf *bp;
2100 	int error;
2101 	daddr_t lblkno;
2102 
2103 	ASSERT_BO_LOCKED(bo);
2104 
2105 	for (lblkno = startn;;) {
2106 again:
2107 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2108 		if (bp == NULL || bp->b_lblkno >= endn ||
2109 		    bp->b_lblkno < startn)
2110 			break;
2111 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2112 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2113 		if (error != 0) {
2114 			BO_RLOCK(bo);
2115 			if (error == ENOLCK)
2116 				goto again;
2117 			return (error);
2118 		}
2119 		KASSERT(bp->b_bufobj == bo,
2120 		    ("bp %p wrong b_bufobj %p should be %p",
2121 		    bp, bp->b_bufobj, bo));
2122 		lblkno = bp->b_lblkno + 1;
2123 		if ((bp->b_flags & B_MANAGED) == 0)
2124 			bremfree(bp);
2125 		bp->b_flags |= B_RELBUF;
2126 		/*
2127 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2128 		 * pages backing each buffer in the range are unlikely to be
2129 		 * reused.  Dirty buffers will have the hint applied once
2130 		 * they've been written.
2131 		 */
2132 		if ((bp->b_flags & B_VMIO) != 0)
2133 			bp->b_flags |= B_NOREUSE;
2134 		brelse(bp);
2135 		BO_RLOCK(bo);
2136 	}
2137 	return (0);
2138 }
2139 
2140 /*
2141  * Truncate a file's buffer and pages to a specified length.  This
2142  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2143  * sync activity.
2144  */
2145 int
2146 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2147 {
2148 	struct buf *bp, *nbp;
2149 	struct bufobj *bo;
2150 	daddr_t startlbn;
2151 
2152 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2153 	    vp, blksize, (uintmax_t)length);
2154 
2155 	/*
2156 	 * Round up to the *next* lbn.
2157 	 */
2158 	startlbn = howmany(length, blksize);
2159 
2160 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2161 
2162 	bo = &vp->v_bufobj;
2163 restart_unlocked:
2164 	BO_LOCK(bo);
2165 
2166 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2167 		;
2168 
2169 	if (length > 0) {
2170 restartsync:
2171 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2172 			if (bp->b_lblkno > 0)
2173 				continue;
2174 			/*
2175 			 * Since we hold the vnode lock this should only
2176 			 * fail if we're racing with the buf daemon.
2177 			 */
2178 			if (BUF_LOCK(bp,
2179 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2180 			    BO_LOCKPTR(bo)) == ENOLCK)
2181 				goto restart_unlocked;
2182 
2183 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2184 			    ("buf(%p) on dirty queue without DELWRI", bp));
2185 
2186 			bremfree(bp);
2187 			bawrite(bp);
2188 			BO_LOCK(bo);
2189 			goto restartsync;
2190 		}
2191 	}
2192 
2193 	bufobj_wwait(bo, 0, 0);
2194 	BO_UNLOCK(bo);
2195 	vnode_pager_setsize(vp, length);
2196 
2197 	return (0);
2198 }
2199 
2200 /*
2201  * Invalidate the cached pages of a file's buffer within the range of block
2202  * numbers [startlbn, endlbn).
2203  */
2204 void
2205 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2206     int blksize)
2207 {
2208 	struct bufobj *bo;
2209 	off_t start, end;
2210 
2211 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2212 
2213 	start = blksize * startlbn;
2214 	end = blksize * endlbn;
2215 
2216 	bo = &vp->v_bufobj;
2217 	BO_LOCK(bo);
2218 	MPASS(blksize == bo->bo_bsize);
2219 
2220 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2221 		;
2222 
2223 	BO_UNLOCK(bo);
2224 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2225 }
2226 
2227 static int
2228 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2229     daddr_t startlbn, daddr_t endlbn)
2230 {
2231 	struct buf *bp, *nbp;
2232 	bool anyfreed;
2233 
2234 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2235 	ASSERT_BO_LOCKED(bo);
2236 
2237 	do {
2238 		anyfreed = false;
2239 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2240 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2241 				continue;
2242 			if (BUF_LOCK(bp,
2243 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2244 			    BO_LOCKPTR(bo)) == ENOLCK) {
2245 				BO_LOCK(bo);
2246 				return (EAGAIN);
2247 			}
2248 
2249 			bremfree(bp);
2250 			bp->b_flags |= B_INVAL | B_RELBUF;
2251 			bp->b_flags &= ~B_ASYNC;
2252 			brelse(bp);
2253 			anyfreed = true;
2254 
2255 			BO_LOCK(bo);
2256 			if (nbp != NULL &&
2257 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2258 			    nbp->b_vp != vp ||
2259 			    (nbp->b_flags & B_DELWRI) != 0))
2260 				return (EAGAIN);
2261 		}
2262 
2263 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2264 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2265 				continue;
2266 			if (BUF_LOCK(bp,
2267 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2268 			    BO_LOCKPTR(bo)) == ENOLCK) {
2269 				BO_LOCK(bo);
2270 				return (EAGAIN);
2271 			}
2272 			bremfree(bp);
2273 			bp->b_flags |= B_INVAL | B_RELBUF;
2274 			bp->b_flags &= ~B_ASYNC;
2275 			brelse(bp);
2276 			anyfreed = true;
2277 
2278 			BO_LOCK(bo);
2279 			if (nbp != NULL &&
2280 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2281 			    (nbp->b_vp != vp) ||
2282 			    (nbp->b_flags & B_DELWRI) == 0))
2283 				return (EAGAIN);
2284 		}
2285 	} while (anyfreed);
2286 	return (0);
2287 }
2288 
2289 static void
2290 buf_vlist_remove(struct buf *bp)
2291 {
2292 	struct bufv *bv;
2293 	b_xflags_t flags;
2294 
2295 	flags = bp->b_xflags;
2296 
2297 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2298 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2299 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2300 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2301 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2302 
2303 	if ((flags & BX_VNDIRTY) != 0)
2304 		bv = &bp->b_bufobj->bo_dirty;
2305 	else
2306 		bv = &bp->b_bufobj->bo_clean;
2307 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2308 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2309 	bv->bv_cnt--;
2310 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2311 }
2312 
2313 /*
2314  * Add the buffer to the sorted clean or dirty block list.
2315  *
2316  * NOTE: xflags is passed as a constant, optimizing this inline function!
2317  */
2318 static void
2319 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2320 {
2321 	struct bufv *bv;
2322 	struct buf *n;
2323 	int error;
2324 
2325 	ASSERT_BO_WLOCKED(bo);
2326 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2327 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2328 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2329 	    ("dead bo %p", bo));
2330 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2331 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2332 	bp->b_xflags |= xflags;
2333 	if (xflags & BX_VNDIRTY)
2334 		bv = &bo->bo_dirty;
2335 	else
2336 		bv = &bo->bo_clean;
2337 
2338 	/*
2339 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2340 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2341 	 * than _ge.
2342 	 */
2343 	if (bv->bv_cnt == 0 ||
2344 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2345 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2346 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2347 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2348 	else
2349 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2350 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2351 	if (error)
2352 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2353 	bv->bv_cnt++;
2354 }
2355 
2356 /*
2357  * Look up a buffer using the buffer tries.
2358  */
2359 struct buf *
2360 gbincore(struct bufobj *bo, daddr_t lblkno)
2361 {
2362 	struct buf *bp;
2363 
2364 	ASSERT_BO_LOCKED(bo);
2365 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2366 	if (bp != NULL)
2367 		return (bp);
2368 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2369 }
2370 
2371 /*
2372  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2373  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2374  * stability of the result.  Like other lockless lookups, the found buf may
2375  * already be invalid by the time this function returns.
2376  */
2377 struct buf *
2378 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2379 {
2380 	struct buf *bp;
2381 
2382 	ASSERT_BO_UNLOCKED(bo);
2383 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2384 	if (bp != NULL)
2385 		return (bp);
2386 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2387 }
2388 
2389 /*
2390  * Associate a buffer with a vnode.
2391  */
2392 void
2393 bgetvp(struct vnode *vp, struct buf *bp)
2394 {
2395 	struct bufobj *bo;
2396 
2397 	bo = &vp->v_bufobj;
2398 	ASSERT_BO_WLOCKED(bo);
2399 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2400 
2401 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2402 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2403 	    ("bgetvp: bp already attached! %p", bp));
2404 
2405 	vhold(vp);
2406 	bp->b_vp = vp;
2407 	bp->b_bufobj = bo;
2408 	/*
2409 	 * Insert onto list for new vnode.
2410 	 */
2411 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2412 }
2413 
2414 /*
2415  * Disassociate a buffer from a vnode.
2416  */
2417 void
2418 brelvp(struct buf *bp)
2419 {
2420 	struct bufobj *bo;
2421 	struct vnode *vp;
2422 
2423 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2424 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2425 
2426 	/*
2427 	 * Delete from old vnode list, if on one.
2428 	 */
2429 	vp = bp->b_vp;		/* XXX */
2430 	bo = bp->b_bufobj;
2431 	BO_LOCK(bo);
2432 	buf_vlist_remove(bp);
2433 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2434 		bo->bo_flag &= ~BO_ONWORKLST;
2435 		mtx_lock(&sync_mtx);
2436 		LIST_REMOVE(bo, bo_synclist);
2437 		syncer_worklist_len--;
2438 		mtx_unlock(&sync_mtx);
2439 	}
2440 	bp->b_vp = NULL;
2441 	bp->b_bufobj = NULL;
2442 	BO_UNLOCK(bo);
2443 	vdrop(vp);
2444 }
2445 
2446 /*
2447  * Add an item to the syncer work queue.
2448  */
2449 static void
2450 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2451 {
2452 	int slot;
2453 
2454 	ASSERT_BO_WLOCKED(bo);
2455 
2456 	mtx_lock(&sync_mtx);
2457 	if (bo->bo_flag & BO_ONWORKLST)
2458 		LIST_REMOVE(bo, bo_synclist);
2459 	else {
2460 		bo->bo_flag |= BO_ONWORKLST;
2461 		syncer_worklist_len++;
2462 	}
2463 
2464 	if (delay > syncer_maxdelay - 2)
2465 		delay = syncer_maxdelay - 2;
2466 	slot = (syncer_delayno + delay) & syncer_mask;
2467 
2468 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2469 	mtx_unlock(&sync_mtx);
2470 }
2471 
2472 static int
2473 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2474 {
2475 	int error, len;
2476 
2477 	mtx_lock(&sync_mtx);
2478 	len = syncer_worklist_len - sync_vnode_count;
2479 	mtx_unlock(&sync_mtx);
2480 	error = SYSCTL_OUT(req, &len, sizeof(len));
2481 	return (error);
2482 }
2483 
2484 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2485     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2486     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2487 
2488 static struct proc *updateproc;
2489 static void sched_sync(void);
2490 static struct kproc_desc up_kp = {
2491 	"syncer",
2492 	sched_sync,
2493 	&updateproc
2494 };
2495 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2496 
2497 static int
2498 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2499 {
2500 	struct vnode *vp;
2501 	struct mount *mp;
2502 
2503 	*bo = LIST_FIRST(slp);
2504 	if (*bo == NULL)
2505 		return (0);
2506 	vp = bo2vnode(*bo);
2507 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2508 		return (1);
2509 	/*
2510 	 * We use vhold in case the vnode does not
2511 	 * successfully sync.  vhold prevents the vnode from
2512 	 * going away when we unlock the sync_mtx so that
2513 	 * we can acquire the vnode interlock.
2514 	 */
2515 	vholdl(vp);
2516 	mtx_unlock(&sync_mtx);
2517 	VI_UNLOCK(vp);
2518 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2519 		vdrop(vp);
2520 		mtx_lock(&sync_mtx);
2521 		return (*bo == LIST_FIRST(slp));
2522 	}
2523 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2524 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2525 	VOP_UNLOCK(vp);
2526 	vn_finished_write(mp);
2527 	BO_LOCK(*bo);
2528 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2529 		/*
2530 		 * Put us back on the worklist.  The worklist
2531 		 * routine will remove us from our current
2532 		 * position and then add us back in at a later
2533 		 * position.
2534 		 */
2535 		vn_syncer_add_to_worklist(*bo, syncdelay);
2536 	}
2537 	BO_UNLOCK(*bo);
2538 	vdrop(vp);
2539 	mtx_lock(&sync_mtx);
2540 	return (0);
2541 }
2542 
2543 static int first_printf = 1;
2544 
2545 /*
2546  * System filesystem synchronizer daemon.
2547  */
2548 static void
2549 sched_sync(void)
2550 {
2551 	struct synclist *next, *slp;
2552 	struct bufobj *bo;
2553 	long starttime;
2554 	struct thread *td = curthread;
2555 	int last_work_seen;
2556 	int net_worklist_len;
2557 	int syncer_final_iter;
2558 	int error;
2559 
2560 	last_work_seen = 0;
2561 	syncer_final_iter = 0;
2562 	syncer_state = SYNCER_RUNNING;
2563 	starttime = time_uptime;
2564 	td->td_pflags |= TDP_NORUNNINGBUF;
2565 
2566 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2567 	    SHUTDOWN_PRI_LAST);
2568 
2569 	mtx_lock(&sync_mtx);
2570 	for (;;) {
2571 		if (syncer_state == SYNCER_FINAL_DELAY &&
2572 		    syncer_final_iter == 0) {
2573 			mtx_unlock(&sync_mtx);
2574 			kproc_suspend_check(td->td_proc);
2575 			mtx_lock(&sync_mtx);
2576 		}
2577 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2578 		if (syncer_state != SYNCER_RUNNING &&
2579 		    starttime != time_uptime) {
2580 			if (first_printf) {
2581 				printf("\nSyncing disks, vnodes remaining... ");
2582 				first_printf = 0;
2583 			}
2584 			printf("%d ", net_worklist_len);
2585 		}
2586 		starttime = time_uptime;
2587 
2588 		/*
2589 		 * Push files whose dirty time has expired.  Be careful
2590 		 * of interrupt race on slp queue.
2591 		 *
2592 		 * Skip over empty worklist slots when shutting down.
2593 		 */
2594 		do {
2595 			slp = &syncer_workitem_pending[syncer_delayno];
2596 			syncer_delayno += 1;
2597 			if (syncer_delayno == syncer_maxdelay)
2598 				syncer_delayno = 0;
2599 			next = &syncer_workitem_pending[syncer_delayno];
2600 			/*
2601 			 * If the worklist has wrapped since the
2602 			 * it was emptied of all but syncer vnodes,
2603 			 * switch to the FINAL_DELAY state and run
2604 			 * for one more second.
2605 			 */
2606 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2607 			    net_worklist_len == 0 &&
2608 			    last_work_seen == syncer_delayno) {
2609 				syncer_state = SYNCER_FINAL_DELAY;
2610 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2611 			}
2612 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2613 		    syncer_worklist_len > 0);
2614 
2615 		/*
2616 		 * Keep track of the last time there was anything
2617 		 * on the worklist other than syncer vnodes.
2618 		 * Return to the SHUTTING_DOWN state if any
2619 		 * new work appears.
2620 		 */
2621 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2622 			last_work_seen = syncer_delayno;
2623 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2624 			syncer_state = SYNCER_SHUTTING_DOWN;
2625 		while (!LIST_EMPTY(slp)) {
2626 			error = sync_vnode(slp, &bo, td);
2627 			if (error == 1) {
2628 				LIST_REMOVE(bo, bo_synclist);
2629 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2630 				continue;
2631 			}
2632 
2633 			if (first_printf == 0) {
2634 				/*
2635 				 * Drop the sync mutex, because some watchdog
2636 				 * drivers need to sleep while patting
2637 				 */
2638 				mtx_unlock(&sync_mtx);
2639 				wdog_kern_pat(WD_LASTVAL);
2640 				mtx_lock(&sync_mtx);
2641 			}
2642 		}
2643 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2644 			syncer_final_iter--;
2645 		/*
2646 		 * The variable rushjob allows the kernel to speed up the
2647 		 * processing of the filesystem syncer process. A rushjob
2648 		 * value of N tells the filesystem syncer to process the next
2649 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2650 		 * is used by the soft update code to speed up the filesystem
2651 		 * syncer process when the incore state is getting so far
2652 		 * ahead of the disk that the kernel memory pool is being
2653 		 * threatened with exhaustion.
2654 		 */
2655 		if (rushjob > 0) {
2656 			rushjob -= 1;
2657 			continue;
2658 		}
2659 		/*
2660 		 * Just sleep for a short period of time between
2661 		 * iterations when shutting down to allow some I/O
2662 		 * to happen.
2663 		 *
2664 		 * If it has taken us less than a second to process the
2665 		 * current work, then wait. Otherwise start right over
2666 		 * again. We can still lose time if any single round
2667 		 * takes more than two seconds, but it does not really
2668 		 * matter as we are just trying to generally pace the
2669 		 * filesystem activity.
2670 		 */
2671 		if (syncer_state != SYNCER_RUNNING ||
2672 		    time_uptime == starttime) {
2673 			thread_lock(td);
2674 			sched_prio(td, PPAUSE);
2675 			thread_unlock(td);
2676 		}
2677 		if (syncer_state != SYNCER_RUNNING)
2678 			cv_timedwait(&sync_wakeup, &sync_mtx,
2679 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2680 		else if (time_uptime == starttime)
2681 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2682 	}
2683 }
2684 
2685 /*
2686  * Request the syncer daemon to speed up its work.
2687  * We never push it to speed up more than half of its
2688  * normal turn time, otherwise it could take over the cpu.
2689  */
2690 int
2691 speedup_syncer(void)
2692 {
2693 	int ret = 0;
2694 
2695 	mtx_lock(&sync_mtx);
2696 	if (rushjob < syncdelay / 2) {
2697 		rushjob += 1;
2698 		stat_rush_requests += 1;
2699 		ret = 1;
2700 	}
2701 	mtx_unlock(&sync_mtx);
2702 	cv_broadcast(&sync_wakeup);
2703 	return (ret);
2704 }
2705 
2706 /*
2707  * Tell the syncer to speed up its work and run though its work
2708  * list several times, then tell it to shut down.
2709  */
2710 static void
2711 syncer_shutdown(void *arg, int howto)
2712 {
2713 
2714 	if (howto & RB_NOSYNC)
2715 		return;
2716 	mtx_lock(&sync_mtx);
2717 	syncer_state = SYNCER_SHUTTING_DOWN;
2718 	rushjob = 0;
2719 	mtx_unlock(&sync_mtx);
2720 	cv_broadcast(&sync_wakeup);
2721 	kproc_shutdown(arg, howto);
2722 }
2723 
2724 void
2725 syncer_suspend(void)
2726 {
2727 
2728 	syncer_shutdown(updateproc, 0);
2729 }
2730 
2731 void
2732 syncer_resume(void)
2733 {
2734 
2735 	mtx_lock(&sync_mtx);
2736 	first_printf = 1;
2737 	syncer_state = SYNCER_RUNNING;
2738 	mtx_unlock(&sync_mtx);
2739 	cv_broadcast(&sync_wakeup);
2740 	kproc_resume(updateproc);
2741 }
2742 
2743 /*
2744  * Move the buffer between the clean and dirty lists of its vnode.
2745  */
2746 void
2747 reassignbuf(struct buf *bp)
2748 {
2749 	struct vnode *vp;
2750 	struct bufobj *bo;
2751 	int delay;
2752 #ifdef INVARIANTS
2753 	struct bufv *bv;
2754 #endif
2755 
2756 	vp = bp->b_vp;
2757 	bo = bp->b_bufobj;
2758 
2759 	KASSERT((bp->b_flags & B_PAGING) == 0,
2760 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2761 
2762 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2763 	    bp, bp->b_vp, bp->b_flags);
2764 
2765 	BO_LOCK(bo);
2766 	buf_vlist_remove(bp);
2767 
2768 	/*
2769 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2770 	 * of clean buffers.
2771 	 */
2772 	if (bp->b_flags & B_DELWRI) {
2773 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2774 			switch (vp->v_type) {
2775 			case VDIR:
2776 				delay = dirdelay;
2777 				break;
2778 			case VCHR:
2779 				delay = metadelay;
2780 				break;
2781 			default:
2782 				delay = filedelay;
2783 			}
2784 			vn_syncer_add_to_worklist(bo, delay);
2785 		}
2786 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2787 	} else {
2788 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2789 
2790 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2791 			mtx_lock(&sync_mtx);
2792 			LIST_REMOVE(bo, bo_synclist);
2793 			syncer_worklist_len--;
2794 			mtx_unlock(&sync_mtx);
2795 			bo->bo_flag &= ~BO_ONWORKLST;
2796 		}
2797 	}
2798 #ifdef INVARIANTS
2799 	bv = &bo->bo_clean;
2800 	bp = TAILQ_FIRST(&bv->bv_hd);
2801 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2802 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2803 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2804 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2805 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2806 	bv = &bo->bo_dirty;
2807 	bp = TAILQ_FIRST(&bv->bv_hd);
2808 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2809 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2810 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2811 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2812 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2813 #endif
2814 	BO_UNLOCK(bo);
2815 }
2816 
2817 static void
2818 v_init_counters(struct vnode *vp)
2819 {
2820 
2821 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2822 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2823 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2824 
2825 	refcount_init(&vp->v_holdcnt, 1);
2826 	refcount_init(&vp->v_usecount, 1);
2827 }
2828 
2829 /*
2830  * Grab a particular vnode from the free list, increment its
2831  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2832  * is being destroyed.  Only callers who specify LK_RETRY will
2833  * see doomed vnodes.  If inactive processing was delayed in
2834  * vput try to do it here.
2835  *
2836  * usecount is manipulated using atomics without holding any locks.
2837  *
2838  * holdcnt can be manipulated using atomics without holding any locks,
2839  * except when transitioning 1<->0, in which case the interlock is held.
2840  *
2841  * Consumers which don't guarantee liveness of the vnode can use SMR to
2842  * try to get a reference. Note this operation can fail since the vnode
2843  * may be awaiting getting freed by the time they get to it.
2844  */
2845 enum vgetstate
2846 vget_prep_smr(struct vnode *vp)
2847 {
2848 	enum vgetstate vs;
2849 
2850 	VFS_SMR_ASSERT_ENTERED();
2851 
2852 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2853 		vs = VGET_USECOUNT;
2854 	} else {
2855 		if (vhold_smr(vp))
2856 			vs = VGET_HOLDCNT;
2857 		else
2858 			vs = VGET_NONE;
2859 	}
2860 	return (vs);
2861 }
2862 
2863 enum vgetstate
2864 vget_prep(struct vnode *vp)
2865 {
2866 	enum vgetstate vs;
2867 
2868 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2869 		vs = VGET_USECOUNT;
2870 	} else {
2871 		vhold(vp);
2872 		vs = VGET_HOLDCNT;
2873 	}
2874 	return (vs);
2875 }
2876 
2877 void
2878 vget_abort(struct vnode *vp, enum vgetstate vs)
2879 {
2880 
2881 	switch (vs) {
2882 	case VGET_USECOUNT:
2883 		vrele(vp);
2884 		break;
2885 	case VGET_HOLDCNT:
2886 		vdrop(vp);
2887 		break;
2888 	default:
2889 		__assert_unreachable();
2890 	}
2891 }
2892 
2893 int
2894 vget(struct vnode *vp, int flags)
2895 {
2896 	enum vgetstate vs;
2897 
2898 	vs = vget_prep(vp);
2899 	return (vget_finish(vp, flags, vs));
2900 }
2901 
2902 int
2903 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2904 {
2905 	int error;
2906 
2907 	if ((flags & LK_INTERLOCK) != 0)
2908 		ASSERT_VI_LOCKED(vp, __func__);
2909 	else
2910 		ASSERT_VI_UNLOCKED(vp, __func__);
2911 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2912 	VNPASS(vp->v_holdcnt > 0, vp);
2913 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2914 
2915 	error = vn_lock(vp, flags);
2916 	if (__predict_false(error != 0)) {
2917 		vget_abort(vp, vs);
2918 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2919 		    vp);
2920 		return (error);
2921 	}
2922 
2923 	vget_finish_ref(vp, vs);
2924 	return (0);
2925 }
2926 
2927 void
2928 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
2929 {
2930 	int old;
2931 
2932 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2933 	VNPASS(vp->v_holdcnt > 0, vp);
2934 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2935 
2936 	if (vs == VGET_USECOUNT)
2937 		return;
2938 
2939 	/*
2940 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2941 	 * the vnode around. Otherwise someone else lended their hold count and
2942 	 * we have to drop ours.
2943 	 */
2944 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2945 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2946 	if (old != 0) {
2947 #ifdef INVARIANTS
2948 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2949 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2950 #else
2951 		refcount_release(&vp->v_holdcnt);
2952 #endif
2953 	}
2954 }
2955 
2956 void
2957 vref(struct vnode *vp)
2958 {
2959 	enum vgetstate vs;
2960 
2961 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2962 	vs = vget_prep(vp);
2963 	vget_finish_ref(vp, vs);
2964 }
2965 
2966 void
2967 vrefact(struct vnode *vp)
2968 {
2969 
2970 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2971 #ifdef INVARIANTS
2972 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2973 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
2974 #else
2975 	refcount_acquire(&vp->v_usecount);
2976 #endif
2977 }
2978 
2979 void
2980 vlazy(struct vnode *vp)
2981 {
2982 	struct mount *mp;
2983 
2984 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2985 
2986 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
2987 		return;
2988 	/*
2989 	 * We may get here for inactive routines after the vnode got doomed.
2990 	 */
2991 	if (VN_IS_DOOMED(vp))
2992 		return;
2993 	mp = vp->v_mount;
2994 	mtx_lock(&mp->mnt_listmtx);
2995 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
2996 		vp->v_mflag |= VMP_LAZYLIST;
2997 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
2998 		mp->mnt_lazyvnodelistsize++;
2999 	}
3000 	mtx_unlock(&mp->mnt_listmtx);
3001 }
3002 
3003 /*
3004  * This routine is only meant to be called from vgonel prior to dooming
3005  * the vnode.
3006  */
3007 static void
3008 vunlazy_gone(struct vnode *vp)
3009 {
3010 	struct mount *mp;
3011 
3012 	ASSERT_VOP_ELOCKED(vp, __func__);
3013 	ASSERT_VI_LOCKED(vp, __func__);
3014 	VNPASS(!VN_IS_DOOMED(vp), vp);
3015 
3016 	if (vp->v_mflag & VMP_LAZYLIST) {
3017 		mp = vp->v_mount;
3018 		mtx_lock(&mp->mnt_listmtx);
3019 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3020 		vp->v_mflag &= ~VMP_LAZYLIST;
3021 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3022 		mp->mnt_lazyvnodelistsize--;
3023 		mtx_unlock(&mp->mnt_listmtx);
3024 	}
3025 }
3026 
3027 static void
3028 vdefer_inactive(struct vnode *vp)
3029 {
3030 
3031 	ASSERT_VI_LOCKED(vp, __func__);
3032 	VNASSERT(vp->v_holdcnt > 0, vp,
3033 	    ("%s: vnode without hold count", __func__));
3034 	if (VN_IS_DOOMED(vp)) {
3035 		vdropl(vp);
3036 		return;
3037 	}
3038 	if (vp->v_iflag & VI_DEFINACT) {
3039 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3040 		vdropl(vp);
3041 		return;
3042 	}
3043 	if (vp->v_usecount > 0) {
3044 		vp->v_iflag &= ~VI_OWEINACT;
3045 		vdropl(vp);
3046 		return;
3047 	}
3048 	vlazy(vp);
3049 	vp->v_iflag |= VI_DEFINACT;
3050 	VI_UNLOCK(vp);
3051 	counter_u64_add(deferred_inact, 1);
3052 }
3053 
3054 static void
3055 vdefer_inactive_unlocked(struct vnode *vp)
3056 {
3057 
3058 	VI_LOCK(vp);
3059 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3060 		vdropl(vp);
3061 		return;
3062 	}
3063 	vdefer_inactive(vp);
3064 }
3065 
3066 enum vput_op { VRELE, VPUT, VUNREF };
3067 
3068 /*
3069  * Handle ->v_usecount transitioning to 0.
3070  *
3071  * By releasing the last usecount we take ownership of the hold count which
3072  * provides liveness of the vnode, meaning we have to vdrop.
3073  *
3074  * For all vnodes we may need to perform inactive processing. It requires an
3075  * exclusive lock on the vnode, while it is legal to call here with only a
3076  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3077  * inactive processing gets deferred to the syncer.
3078  *
3079  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3080  * on the lock being held all the way until VOP_INACTIVE. This in particular
3081  * happens with UFS which adds half-constructed vnodes to the hash, where they
3082  * can be found by other code.
3083  */
3084 static void
3085 vput_final(struct vnode *vp, enum vput_op func)
3086 {
3087 	int error;
3088 	bool want_unlock;
3089 
3090 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3091 	VNPASS(vp->v_holdcnt > 0, vp);
3092 
3093 	VI_LOCK(vp);
3094 
3095 	/*
3096 	 * By the time we got here someone else might have transitioned
3097 	 * the count back to > 0.
3098 	 */
3099 	if (vp->v_usecount > 0)
3100 		goto out;
3101 
3102 	/*
3103 	 * If the vnode is doomed vgone already performed inactive processing
3104 	 * (if needed).
3105 	 */
3106 	if (VN_IS_DOOMED(vp))
3107 		goto out;
3108 
3109 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3110 		goto out;
3111 
3112 	if (vp->v_iflag & VI_DOINGINACT)
3113 		goto out;
3114 
3115 	/*
3116 	 * Locking operations here will drop the interlock and possibly the
3117 	 * vnode lock, opening a window where the vnode can get doomed all the
3118 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3119 	 * perform inactive.
3120 	 */
3121 	vp->v_iflag |= VI_OWEINACT;
3122 	want_unlock = false;
3123 	error = 0;
3124 	switch (func) {
3125 	case VRELE:
3126 		switch (VOP_ISLOCKED(vp)) {
3127 		case LK_EXCLUSIVE:
3128 			break;
3129 		case LK_EXCLOTHER:
3130 		case 0:
3131 			want_unlock = true;
3132 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3133 			VI_LOCK(vp);
3134 			break;
3135 		default:
3136 			/*
3137 			 * The lock has at least one sharer, but we have no way
3138 			 * to conclude whether this is us. Play it safe and
3139 			 * defer processing.
3140 			 */
3141 			error = EAGAIN;
3142 			break;
3143 		}
3144 		break;
3145 	case VPUT:
3146 		want_unlock = true;
3147 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3148 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3149 			    LK_NOWAIT);
3150 			VI_LOCK(vp);
3151 		}
3152 		break;
3153 	case VUNREF:
3154 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3155 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3156 			VI_LOCK(vp);
3157 		}
3158 		break;
3159 	}
3160 	if (error == 0) {
3161 		vinactive(vp);
3162 		if (want_unlock)
3163 			VOP_UNLOCK(vp);
3164 		vdropl(vp);
3165 	} else {
3166 		vdefer_inactive(vp);
3167 	}
3168 	return;
3169 out:
3170 	if (func == VPUT)
3171 		VOP_UNLOCK(vp);
3172 	vdropl(vp);
3173 }
3174 
3175 /*
3176  * Decrement ->v_usecount for a vnode.
3177  *
3178  * Releasing the last use count requires additional processing, see vput_final
3179  * above for details.
3180  *
3181  * Comment above each variant denotes lock state on entry and exit.
3182  */
3183 
3184 /*
3185  * in: any
3186  * out: same as passed in
3187  */
3188 void
3189 vrele(struct vnode *vp)
3190 {
3191 
3192 	ASSERT_VI_UNLOCKED(vp, __func__);
3193 	if (!refcount_release(&vp->v_usecount))
3194 		return;
3195 	vput_final(vp, VRELE);
3196 }
3197 
3198 /*
3199  * in: locked
3200  * out: unlocked
3201  */
3202 void
3203 vput(struct vnode *vp)
3204 {
3205 
3206 	ASSERT_VOP_LOCKED(vp, __func__);
3207 	ASSERT_VI_UNLOCKED(vp, __func__);
3208 	if (!refcount_release(&vp->v_usecount)) {
3209 		VOP_UNLOCK(vp);
3210 		return;
3211 	}
3212 	vput_final(vp, VPUT);
3213 }
3214 
3215 /*
3216  * in: locked
3217  * out: locked
3218  */
3219 void
3220 vunref(struct vnode *vp)
3221 {
3222 
3223 	ASSERT_VOP_LOCKED(vp, __func__);
3224 	ASSERT_VI_UNLOCKED(vp, __func__);
3225 	if (!refcount_release(&vp->v_usecount))
3226 		return;
3227 	vput_final(vp, VUNREF);
3228 }
3229 
3230 void
3231 vhold(struct vnode *vp)
3232 {
3233 	int old;
3234 
3235 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3236 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3237 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3238 	    ("%s: wrong hold count %d", __func__, old));
3239 	if (old == 0)
3240 		vn_freevnodes_dec();
3241 }
3242 
3243 void
3244 vholdnz(struct vnode *vp)
3245 {
3246 
3247 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3248 #ifdef INVARIANTS
3249 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3250 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3251 	    ("%s: wrong hold count %d", __func__, old));
3252 #else
3253 	atomic_add_int(&vp->v_holdcnt, 1);
3254 #endif
3255 }
3256 
3257 /*
3258  * Grab a hold count unless the vnode is freed.
3259  *
3260  * Only use this routine if vfs smr is the only protection you have against
3261  * freeing the vnode.
3262  *
3263  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3264  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3265  * the thread which managed to set the flag.
3266  *
3267  * It may be tempting to replace the loop with:
3268  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3269  * if (count & VHOLD_NO_SMR) {
3270  *     backpedal and error out;
3271  * }
3272  *
3273  * However, while this is more performant, it hinders debugging by eliminating
3274  * the previously mentioned invariant.
3275  */
3276 bool
3277 vhold_smr(struct vnode *vp)
3278 {
3279 	int count;
3280 
3281 	VFS_SMR_ASSERT_ENTERED();
3282 
3283 	count = atomic_load_int(&vp->v_holdcnt);
3284 	for (;;) {
3285 		if (count & VHOLD_NO_SMR) {
3286 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3287 			    ("non-zero hold count with flags %d\n", count));
3288 			return (false);
3289 		}
3290 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3291 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3292 			if (count == 0)
3293 				vn_freevnodes_dec();
3294 			return (true);
3295 		}
3296 	}
3297 }
3298 
3299 /*
3300  * Hold a free vnode for recycling.
3301  *
3302  * Note: vnode_init references this comment.
3303  *
3304  * Attempts to recycle only need the global vnode list lock and have no use for
3305  * SMR.
3306  *
3307  * However, vnodes get inserted into the global list before they get fully
3308  * initialized and stay there until UMA decides to free the memory. This in
3309  * particular means the target can be found before it becomes usable and after
3310  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3311  * VHOLD_NO_SMR.
3312  *
3313  * Note: the vnode may gain more references after we transition the count 0->1.
3314  */
3315 static bool
3316 vhold_recycle_free(struct vnode *vp)
3317 {
3318 	int count;
3319 
3320 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3321 
3322 	count = atomic_load_int(&vp->v_holdcnt);
3323 	for (;;) {
3324 		if (count & VHOLD_NO_SMR) {
3325 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3326 			    ("non-zero hold count with flags %d\n", count));
3327 			return (false);
3328 		}
3329 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3330 		if (count > 0) {
3331 			return (false);
3332 		}
3333 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3334 			vn_freevnodes_dec();
3335 			return (true);
3336 		}
3337 	}
3338 }
3339 
3340 static void __noinline
3341 vdbatch_process(struct vdbatch *vd)
3342 {
3343 	struct vnode *vp;
3344 	int i;
3345 
3346 	mtx_assert(&vd->lock, MA_OWNED);
3347 	MPASS(curthread->td_pinned > 0);
3348 	MPASS(vd->index == VDBATCH_SIZE);
3349 
3350 	mtx_lock(&vnode_list_mtx);
3351 	critical_enter();
3352 	freevnodes += vd->freevnodes;
3353 	for (i = 0; i < VDBATCH_SIZE; i++) {
3354 		vp = vd->tab[i];
3355 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3356 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3357 		MPASS(vp->v_dbatchcpu != NOCPU);
3358 		vp->v_dbatchcpu = NOCPU;
3359 	}
3360 	mtx_unlock(&vnode_list_mtx);
3361 	vd->freevnodes = 0;
3362 	bzero(vd->tab, sizeof(vd->tab));
3363 	vd->index = 0;
3364 	critical_exit();
3365 }
3366 
3367 static void
3368 vdbatch_enqueue(struct vnode *vp)
3369 {
3370 	struct vdbatch *vd;
3371 
3372 	ASSERT_VI_LOCKED(vp, __func__);
3373 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3374 	    ("%s: deferring requeue of a doomed vnode", __func__));
3375 
3376 	if (vp->v_dbatchcpu != NOCPU) {
3377 		VI_UNLOCK(vp);
3378 		return;
3379 	}
3380 
3381 	sched_pin();
3382 	vd = DPCPU_PTR(vd);
3383 	mtx_lock(&vd->lock);
3384 	MPASS(vd->index < VDBATCH_SIZE);
3385 	MPASS(vd->tab[vd->index] == NULL);
3386 	/*
3387 	 * A hack: we depend on being pinned so that we know what to put in
3388 	 * ->v_dbatchcpu.
3389 	 */
3390 	vp->v_dbatchcpu = curcpu;
3391 	vd->tab[vd->index] = vp;
3392 	vd->index++;
3393 	VI_UNLOCK(vp);
3394 	if (vd->index == VDBATCH_SIZE)
3395 		vdbatch_process(vd);
3396 	mtx_unlock(&vd->lock);
3397 	sched_unpin();
3398 }
3399 
3400 /*
3401  * This routine must only be called for vnodes which are about to be
3402  * deallocated. Supporting dequeue for arbitrary vndoes would require
3403  * validating that the locked batch matches.
3404  */
3405 static void
3406 vdbatch_dequeue(struct vnode *vp)
3407 {
3408 	struct vdbatch *vd;
3409 	int i;
3410 	short cpu;
3411 
3412 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3413 	    ("%s: called for a used vnode\n", __func__));
3414 
3415 	cpu = vp->v_dbatchcpu;
3416 	if (cpu == NOCPU)
3417 		return;
3418 
3419 	vd = DPCPU_ID_PTR(cpu, vd);
3420 	mtx_lock(&vd->lock);
3421 	for (i = 0; i < vd->index; i++) {
3422 		if (vd->tab[i] != vp)
3423 			continue;
3424 		vp->v_dbatchcpu = NOCPU;
3425 		vd->index--;
3426 		vd->tab[i] = vd->tab[vd->index];
3427 		vd->tab[vd->index] = NULL;
3428 		break;
3429 	}
3430 	mtx_unlock(&vd->lock);
3431 	/*
3432 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3433 	 */
3434 	MPASS(vp->v_dbatchcpu == NOCPU);
3435 }
3436 
3437 /*
3438  * Drop the hold count of the vnode.  If this is the last reference to
3439  * the vnode we place it on the free list unless it has been vgone'd
3440  * (marked VIRF_DOOMED) in which case we will free it.
3441  *
3442  * Because the vnode vm object keeps a hold reference on the vnode if
3443  * there is at least one resident non-cached page, the vnode cannot
3444  * leave the active list without the page cleanup done.
3445  */
3446 static void
3447 vdrop_deactivate(struct vnode *vp)
3448 {
3449 	struct mount *mp;
3450 
3451 	ASSERT_VI_LOCKED(vp, __func__);
3452 	/*
3453 	 * Mark a vnode as free: remove it from its active list
3454 	 * and put it up for recycling on the freelist.
3455 	 */
3456 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3457 	    ("vdrop: returning doomed vnode"));
3458 	VNASSERT(vp->v_op != NULL, vp,
3459 	    ("vdrop: vnode already reclaimed."));
3460 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3461 	    ("vnode with VI_OWEINACT set"));
3462 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3463 	    ("vnode with VI_DEFINACT set"));
3464 	if (vp->v_mflag & VMP_LAZYLIST) {
3465 		mp = vp->v_mount;
3466 		mtx_lock(&mp->mnt_listmtx);
3467 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3468 		/*
3469 		 * Don't remove the vnode from the lazy list if another thread
3470 		 * has increased the hold count. It may have re-enqueued the
3471 		 * vnode to the lazy list and is now responsible for its
3472 		 * removal.
3473 		 */
3474 		if (vp->v_holdcnt == 0) {
3475 			vp->v_mflag &= ~VMP_LAZYLIST;
3476 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3477 			mp->mnt_lazyvnodelistsize--;
3478 		}
3479 		mtx_unlock(&mp->mnt_listmtx);
3480 	}
3481 	vdbatch_enqueue(vp);
3482 }
3483 
3484 static void __noinline
3485 vdropl_final(struct vnode *vp)
3486 {
3487 
3488 	ASSERT_VI_LOCKED(vp, __func__);
3489 	VNPASS(VN_IS_DOOMED(vp), vp);
3490 	/*
3491 	 * Set the VHOLD_NO_SMR flag.
3492 	 *
3493 	 * We may be racing against vhold_smr. If they win we can just pretend
3494 	 * we never got this far, they will vdrop later.
3495 	 */
3496 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3497 		vn_freevnodes_inc();
3498 		VI_UNLOCK(vp);
3499 		/*
3500 		 * We lost the aforementioned race. Any subsequent access is
3501 		 * invalid as they might have managed to vdropl on their own.
3502 		 */
3503 		return;
3504 	}
3505 	/*
3506 	 * Don't bump freevnodes as this one is going away.
3507 	 */
3508 	freevnode(vp);
3509 }
3510 
3511 void
3512 vdrop(struct vnode *vp)
3513 {
3514 
3515 	ASSERT_VI_UNLOCKED(vp, __func__);
3516 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3517 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3518 		return;
3519 	VI_LOCK(vp);
3520 	vdropl(vp);
3521 }
3522 
3523 void
3524 vdropl(struct vnode *vp)
3525 {
3526 
3527 	ASSERT_VI_LOCKED(vp, __func__);
3528 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3529 	if (!refcount_release(&vp->v_holdcnt)) {
3530 		VI_UNLOCK(vp);
3531 		return;
3532 	}
3533 	if (!VN_IS_DOOMED(vp)) {
3534 		vn_freevnodes_inc();
3535 		vdrop_deactivate(vp);
3536 		/*
3537 		 * Also unlocks the interlock. We can't assert on it as we
3538 		 * released our hold and by now the vnode might have been
3539 		 * freed.
3540 		 */
3541 		return;
3542 	}
3543 	vdropl_final(vp);
3544 }
3545 
3546 /*
3547  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3548  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3549  */
3550 static void
3551 vinactivef(struct vnode *vp)
3552 {
3553 	struct vm_object *obj;
3554 
3555 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3556 	ASSERT_VI_LOCKED(vp, "vinactive");
3557 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3558 	    ("vinactive: recursed on VI_DOINGINACT"));
3559 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3560 	vp->v_iflag |= VI_DOINGINACT;
3561 	vp->v_iflag &= ~VI_OWEINACT;
3562 	VI_UNLOCK(vp);
3563 	/*
3564 	 * Before moving off the active list, we must be sure that any
3565 	 * modified pages are converted into the vnode's dirty
3566 	 * buffers, since these will no longer be checked once the
3567 	 * vnode is on the inactive list.
3568 	 *
3569 	 * The write-out of the dirty pages is asynchronous.  At the
3570 	 * point that VOP_INACTIVE() is called, there could still be
3571 	 * pending I/O and dirty pages in the object.
3572 	 */
3573 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3574 	    vm_object_mightbedirty(obj)) {
3575 		VM_OBJECT_WLOCK(obj);
3576 		vm_object_page_clean(obj, 0, 0, 0);
3577 		VM_OBJECT_WUNLOCK(obj);
3578 	}
3579 	VOP_INACTIVE(vp);
3580 	VI_LOCK(vp);
3581 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3582 	    ("vinactive: lost VI_DOINGINACT"));
3583 	vp->v_iflag &= ~VI_DOINGINACT;
3584 }
3585 
3586 void
3587 vinactive(struct vnode *vp)
3588 {
3589 
3590 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3591 	ASSERT_VI_LOCKED(vp, "vinactive");
3592 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3593 
3594 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3595 		return;
3596 	if (vp->v_iflag & VI_DOINGINACT)
3597 		return;
3598 	if (vp->v_usecount > 0) {
3599 		vp->v_iflag &= ~VI_OWEINACT;
3600 		return;
3601 	}
3602 	vinactivef(vp);
3603 }
3604 
3605 /*
3606  * Remove any vnodes in the vnode table belonging to mount point mp.
3607  *
3608  * If FORCECLOSE is not specified, there should not be any active ones,
3609  * return error if any are found (nb: this is a user error, not a
3610  * system error). If FORCECLOSE is specified, detach any active vnodes
3611  * that are found.
3612  *
3613  * If WRITECLOSE is set, only flush out regular file vnodes open for
3614  * writing.
3615  *
3616  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3617  *
3618  * `rootrefs' specifies the base reference count for the root vnode
3619  * of this filesystem. The root vnode is considered busy if its
3620  * v_usecount exceeds this value. On a successful return, vflush(, td)
3621  * will call vrele() on the root vnode exactly rootrefs times.
3622  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3623  * be zero.
3624  */
3625 #ifdef DIAGNOSTIC
3626 static int busyprt = 0;		/* print out busy vnodes */
3627 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3628 #endif
3629 
3630 int
3631 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3632 {
3633 	struct vnode *vp, *mvp, *rootvp = NULL;
3634 	struct vattr vattr;
3635 	int busy = 0, error;
3636 
3637 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3638 	    rootrefs, flags);
3639 	if (rootrefs > 0) {
3640 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3641 		    ("vflush: bad args"));
3642 		/*
3643 		 * Get the filesystem root vnode. We can vput() it
3644 		 * immediately, since with rootrefs > 0, it won't go away.
3645 		 */
3646 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3647 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3648 			    __func__, error);
3649 			return (error);
3650 		}
3651 		vput(rootvp);
3652 	}
3653 loop:
3654 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3655 		vholdl(vp);
3656 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3657 		if (error) {
3658 			vdrop(vp);
3659 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3660 			goto loop;
3661 		}
3662 		/*
3663 		 * Skip over a vnodes marked VV_SYSTEM.
3664 		 */
3665 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3666 			VOP_UNLOCK(vp);
3667 			vdrop(vp);
3668 			continue;
3669 		}
3670 		/*
3671 		 * If WRITECLOSE is set, flush out unlinked but still open
3672 		 * files (even if open only for reading) and regular file
3673 		 * vnodes open for writing.
3674 		 */
3675 		if (flags & WRITECLOSE) {
3676 			if (vp->v_object != NULL) {
3677 				VM_OBJECT_WLOCK(vp->v_object);
3678 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3679 				VM_OBJECT_WUNLOCK(vp->v_object);
3680 			}
3681 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3682 			if (error != 0) {
3683 				VOP_UNLOCK(vp);
3684 				vdrop(vp);
3685 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3686 				return (error);
3687 			}
3688 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3689 			VI_LOCK(vp);
3690 
3691 			if ((vp->v_type == VNON ||
3692 			    (error == 0 && vattr.va_nlink > 0)) &&
3693 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3694 				VOP_UNLOCK(vp);
3695 				vdropl(vp);
3696 				continue;
3697 			}
3698 		} else
3699 			VI_LOCK(vp);
3700 		/*
3701 		 * With v_usecount == 0, all we need to do is clear out the
3702 		 * vnode data structures and we are done.
3703 		 *
3704 		 * If FORCECLOSE is set, forcibly close the vnode.
3705 		 */
3706 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3707 			vgonel(vp);
3708 		} else {
3709 			busy++;
3710 #ifdef DIAGNOSTIC
3711 			if (busyprt)
3712 				vn_printf(vp, "vflush: busy vnode ");
3713 #endif
3714 		}
3715 		VOP_UNLOCK(vp);
3716 		vdropl(vp);
3717 	}
3718 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3719 		/*
3720 		 * If just the root vnode is busy, and if its refcount
3721 		 * is equal to `rootrefs', then go ahead and kill it.
3722 		 */
3723 		VI_LOCK(rootvp);
3724 		KASSERT(busy > 0, ("vflush: not busy"));
3725 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3726 		    ("vflush: usecount %d < rootrefs %d",
3727 		     rootvp->v_usecount, rootrefs));
3728 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3729 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3730 			vgone(rootvp);
3731 			VOP_UNLOCK(rootvp);
3732 			busy = 0;
3733 		} else
3734 			VI_UNLOCK(rootvp);
3735 	}
3736 	if (busy) {
3737 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3738 		    busy);
3739 		return (EBUSY);
3740 	}
3741 	for (; rootrefs > 0; rootrefs--)
3742 		vrele(rootvp);
3743 	return (0);
3744 }
3745 
3746 /*
3747  * Recycle an unused vnode to the front of the free list.
3748  */
3749 int
3750 vrecycle(struct vnode *vp)
3751 {
3752 	int recycled;
3753 
3754 	VI_LOCK(vp);
3755 	recycled = vrecyclel(vp);
3756 	VI_UNLOCK(vp);
3757 	return (recycled);
3758 }
3759 
3760 /*
3761  * vrecycle, with the vp interlock held.
3762  */
3763 int
3764 vrecyclel(struct vnode *vp)
3765 {
3766 	int recycled;
3767 
3768 	ASSERT_VOP_ELOCKED(vp, __func__);
3769 	ASSERT_VI_LOCKED(vp, __func__);
3770 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3771 	recycled = 0;
3772 	if (vp->v_usecount == 0) {
3773 		recycled = 1;
3774 		vgonel(vp);
3775 	}
3776 	return (recycled);
3777 }
3778 
3779 /*
3780  * Eliminate all activity associated with a vnode
3781  * in preparation for reuse.
3782  */
3783 void
3784 vgone(struct vnode *vp)
3785 {
3786 	VI_LOCK(vp);
3787 	vgonel(vp);
3788 	VI_UNLOCK(vp);
3789 }
3790 
3791 static void
3792 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3793     struct vnode *lowervp __unused)
3794 {
3795 }
3796 
3797 /*
3798  * Notify upper mounts about reclaimed or unlinked vnode.
3799  */
3800 void
3801 vfs_notify_upper(struct vnode *vp, int event)
3802 {
3803 	static struct vfsops vgonel_vfsops = {
3804 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3805 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3806 	};
3807 	struct mount *mp, *ump, *mmp;
3808 
3809 	mp = vp->v_mount;
3810 	if (mp == NULL)
3811 		return;
3812 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3813 		return;
3814 
3815 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3816 	mmp->mnt_op = &vgonel_vfsops;
3817 	mmp->mnt_kern_flag |= MNTK_MARKER;
3818 	MNT_ILOCK(mp);
3819 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3820 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3821 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3822 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3823 			continue;
3824 		}
3825 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3826 		MNT_IUNLOCK(mp);
3827 		switch (event) {
3828 		case VFS_NOTIFY_UPPER_RECLAIM:
3829 			VFS_RECLAIM_LOWERVP(ump, vp);
3830 			break;
3831 		case VFS_NOTIFY_UPPER_UNLINK:
3832 			VFS_UNLINK_LOWERVP(ump, vp);
3833 			break;
3834 		default:
3835 			KASSERT(0, ("invalid event %d", event));
3836 			break;
3837 		}
3838 		MNT_ILOCK(mp);
3839 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3840 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3841 	}
3842 	free(mmp, M_TEMP);
3843 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3844 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3845 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3846 		wakeup(&mp->mnt_uppers);
3847 	}
3848 	MNT_IUNLOCK(mp);
3849 }
3850 
3851 /*
3852  * vgone, with the vp interlock held.
3853  */
3854 static void
3855 vgonel(struct vnode *vp)
3856 {
3857 	struct thread *td;
3858 	struct mount *mp;
3859 	vm_object_t object;
3860 	bool active, doinginact, oweinact;
3861 
3862 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3863 	ASSERT_VI_LOCKED(vp, "vgonel");
3864 	VNASSERT(vp->v_holdcnt, vp,
3865 	    ("vgonel: vp %p has no reference.", vp));
3866 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3867 	td = curthread;
3868 
3869 	/*
3870 	 * Don't vgonel if we're already doomed.
3871 	 */
3872 	if (vp->v_irflag & VIRF_DOOMED)
3873 		return;
3874 	/*
3875 	 * Paired with freevnode.
3876 	 */
3877 	vn_seqc_write_begin_locked(vp);
3878 	vunlazy_gone(vp);
3879 	vp->v_irflag |= VIRF_DOOMED;
3880 
3881 	/*
3882 	 * Check to see if the vnode is in use.  If so, we have to
3883 	 * call VOP_CLOSE() and VOP_INACTIVE().
3884 	 *
3885 	 * It could be that VOP_INACTIVE() requested reclamation, in
3886 	 * which case we should avoid recursion, so check
3887 	 * VI_DOINGINACT.  This is not precise but good enough.
3888 	 */
3889 	active = vp->v_usecount > 0;
3890 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3891 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3892 
3893 	/*
3894 	 * If we need to do inactive VI_OWEINACT will be set.
3895 	 */
3896 	if (vp->v_iflag & VI_DEFINACT) {
3897 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3898 		vp->v_iflag &= ~VI_DEFINACT;
3899 		vdropl(vp);
3900 	} else {
3901 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3902 		VI_UNLOCK(vp);
3903 	}
3904 	cache_purge_vgone(vp);
3905 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3906 
3907 	/*
3908 	 * If purging an active vnode, it must be closed and
3909 	 * deactivated before being reclaimed.
3910 	 */
3911 	if (active)
3912 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3913 	if ((oweinact || active) && !doinginact) {
3914 		VI_LOCK(vp);
3915 		vinactivef(vp);
3916 		VI_UNLOCK(vp);
3917 	}
3918 	if (vp->v_type == VSOCK)
3919 		vfs_unp_reclaim(vp);
3920 
3921 	/*
3922 	 * Clean out any buffers associated with the vnode.
3923 	 * If the flush fails, just toss the buffers.
3924 	 */
3925 	mp = NULL;
3926 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3927 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3928 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3929 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3930 			;
3931 	}
3932 
3933 	BO_LOCK(&vp->v_bufobj);
3934 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3935 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3936 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3937 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3938 	    ("vp %p bufobj not invalidated", vp));
3939 
3940 	/*
3941 	 * For VMIO bufobj, BO_DEAD is set later, or in
3942 	 * vm_object_terminate() after the object's page queue is
3943 	 * flushed.
3944 	 */
3945 	object = vp->v_bufobj.bo_object;
3946 	if (object == NULL)
3947 		vp->v_bufobj.bo_flag |= BO_DEAD;
3948 	BO_UNLOCK(&vp->v_bufobj);
3949 
3950 	/*
3951 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3952 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3953 	 * should not touch the object borrowed from the lower vnode
3954 	 * (the handle check).
3955 	 */
3956 	if (object != NULL && object->type == OBJT_VNODE &&
3957 	    object->handle == vp)
3958 		vnode_destroy_vobject(vp);
3959 
3960 	/*
3961 	 * Reclaim the vnode.
3962 	 */
3963 	if (VOP_RECLAIM(vp))
3964 		panic("vgone: cannot reclaim");
3965 	if (mp != NULL)
3966 		vn_finished_secondary_write(mp);
3967 	VNASSERT(vp->v_object == NULL, vp,
3968 	    ("vop_reclaim left v_object vp=%p", vp));
3969 	/*
3970 	 * Clear the advisory locks and wake up waiting threads.
3971 	 */
3972 	(void)VOP_ADVLOCKPURGE(vp);
3973 	vp->v_lockf = NULL;
3974 	/*
3975 	 * Delete from old mount point vnode list.
3976 	 */
3977 	delmntque(vp);
3978 	/*
3979 	 * Done with purge, reset to the standard lock and invalidate
3980 	 * the vnode.
3981 	 */
3982 	VI_LOCK(vp);
3983 	vp->v_vnlock = &vp->v_lock;
3984 	vp->v_op = &dead_vnodeops;
3985 	vp->v_type = VBAD;
3986 }
3987 
3988 /*
3989  * Print out a description of a vnode.
3990  */
3991 static const char * const typename[] =
3992 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3993  "VMARKER"};
3994 
3995 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
3996     "new hold count flag not added to vn_printf");
3997 
3998 void
3999 vn_printf(struct vnode *vp, const char *fmt, ...)
4000 {
4001 	va_list ap;
4002 	char buf[256], buf2[16];
4003 	u_long flags;
4004 	u_int holdcnt;
4005 
4006 	va_start(ap, fmt);
4007 	vprintf(fmt, ap);
4008 	va_end(ap);
4009 	printf("%p: ", (void *)vp);
4010 	printf("type %s\n", typename[vp->v_type]);
4011 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4012 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4013 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4014 	    vp->v_seqc_users);
4015 	switch (vp->v_type) {
4016 	case VDIR:
4017 		printf(" mountedhere %p\n", vp->v_mountedhere);
4018 		break;
4019 	case VCHR:
4020 		printf(" rdev %p\n", vp->v_rdev);
4021 		break;
4022 	case VSOCK:
4023 		printf(" socket %p\n", vp->v_unpcb);
4024 		break;
4025 	case VFIFO:
4026 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4027 		break;
4028 	default:
4029 		printf("\n");
4030 		break;
4031 	}
4032 	buf[0] = '\0';
4033 	buf[1] = '\0';
4034 	if (holdcnt & VHOLD_NO_SMR)
4035 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4036 	printf("    hold count flags (%s)\n", buf + 1);
4037 
4038 	buf[0] = '\0';
4039 	buf[1] = '\0';
4040 	if (vp->v_irflag & VIRF_DOOMED)
4041 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4042 	if (vp->v_irflag & VIRF_PGREAD)
4043 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4044 	flags = vp->v_irflag & ~(VIRF_DOOMED | VIRF_PGREAD);
4045 	if (flags != 0) {
4046 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4047 		strlcat(buf, buf2, sizeof(buf));
4048 	}
4049 	if (vp->v_vflag & VV_ROOT)
4050 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4051 	if (vp->v_vflag & VV_ISTTY)
4052 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4053 	if (vp->v_vflag & VV_NOSYNC)
4054 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4055 	if (vp->v_vflag & VV_ETERNALDEV)
4056 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4057 	if (vp->v_vflag & VV_CACHEDLABEL)
4058 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4059 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4060 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4061 	if (vp->v_vflag & VV_COPYONWRITE)
4062 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4063 	if (vp->v_vflag & VV_SYSTEM)
4064 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4065 	if (vp->v_vflag & VV_PROCDEP)
4066 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4067 	if (vp->v_vflag & VV_NOKNOTE)
4068 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4069 	if (vp->v_vflag & VV_DELETED)
4070 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4071 	if (vp->v_vflag & VV_MD)
4072 		strlcat(buf, "|VV_MD", sizeof(buf));
4073 	if (vp->v_vflag & VV_FORCEINSMQ)
4074 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4075 	if (vp->v_vflag & VV_READLINK)
4076 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4077 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4078 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4079 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
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)\n", buf + 1);
4108 	if (mtx_owned(VI_MTX(vp)))
4109 		printf(" VI_LOCKed");
4110 	if (vp->v_object != NULL)
4111 		printf("    v_object %p ref %d pages %d "
4112 		    "cleanbuf %d dirtybuf %d\n",
4113 		    vp->v_object, vp->v_object->ref_count,
4114 		    vp->v_object->resident_page_count,
4115 		    vp->v_bufobj.bo_clean.bv_cnt,
4116 		    vp->v_bufobj.bo_dirty.bv_cnt);
4117 	printf("    ");
4118 	lockmgr_printinfo(vp->v_vnlock);
4119 	if (vp->v_data != NULL)
4120 		VOP_PRINT(vp);
4121 }
4122 
4123 #ifdef DDB
4124 /*
4125  * List all of the locked vnodes in the system.
4126  * Called when debugging the kernel.
4127  */
4128 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4129 {
4130 	struct mount *mp;
4131 	struct vnode *vp;
4132 
4133 	/*
4134 	 * Note: because this is DDB, we can't obey the locking semantics
4135 	 * for these structures, which means we could catch an inconsistent
4136 	 * state and dereference a nasty pointer.  Not much to be done
4137 	 * about that.
4138 	 */
4139 	db_printf("Locked vnodes\n");
4140 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4141 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4142 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4143 				vn_printf(vp, "vnode ");
4144 		}
4145 	}
4146 }
4147 
4148 /*
4149  * Show details about the given vnode.
4150  */
4151 DB_SHOW_COMMAND(vnode, db_show_vnode)
4152 {
4153 	struct vnode *vp;
4154 
4155 	if (!have_addr)
4156 		return;
4157 	vp = (struct vnode *)addr;
4158 	vn_printf(vp, "vnode ");
4159 }
4160 
4161 /*
4162  * Show details about the given mount point.
4163  */
4164 DB_SHOW_COMMAND(mount, db_show_mount)
4165 {
4166 	struct mount *mp;
4167 	struct vfsopt *opt;
4168 	struct statfs *sp;
4169 	struct vnode *vp;
4170 	char buf[512];
4171 	uint64_t mflags;
4172 	u_int flags;
4173 
4174 	if (!have_addr) {
4175 		/* No address given, print short info about all mount points. */
4176 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4177 			db_printf("%p %s on %s (%s)\n", mp,
4178 			    mp->mnt_stat.f_mntfromname,
4179 			    mp->mnt_stat.f_mntonname,
4180 			    mp->mnt_stat.f_fstypename);
4181 			if (db_pager_quit)
4182 				break;
4183 		}
4184 		db_printf("\nMore info: show mount <addr>\n");
4185 		return;
4186 	}
4187 
4188 	mp = (struct mount *)addr;
4189 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4190 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4191 
4192 	buf[0] = '\0';
4193 	mflags = mp->mnt_flag;
4194 #define	MNT_FLAG(flag)	do {						\
4195 	if (mflags & (flag)) {						\
4196 		if (buf[0] != '\0')					\
4197 			strlcat(buf, ", ", sizeof(buf));		\
4198 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4199 		mflags &= ~(flag);					\
4200 	}								\
4201 } while (0)
4202 	MNT_FLAG(MNT_RDONLY);
4203 	MNT_FLAG(MNT_SYNCHRONOUS);
4204 	MNT_FLAG(MNT_NOEXEC);
4205 	MNT_FLAG(MNT_NOSUID);
4206 	MNT_FLAG(MNT_NFS4ACLS);
4207 	MNT_FLAG(MNT_UNION);
4208 	MNT_FLAG(MNT_ASYNC);
4209 	MNT_FLAG(MNT_SUIDDIR);
4210 	MNT_FLAG(MNT_SOFTDEP);
4211 	MNT_FLAG(MNT_NOSYMFOLLOW);
4212 	MNT_FLAG(MNT_GJOURNAL);
4213 	MNT_FLAG(MNT_MULTILABEL);
4214 	MNT_FLAG(MNT_ACLS);
4215 	MNT_FLAG(MNT_NOATIME);
4216 	MNT_FLAG(MNT_NOCLUSTERR);
4217 	MNT_FLAG(MNT_NOCLUSTERW);
4218 	MNT_FLAG(MNT_SUJ);
4219 	MNT_FLAG(MNT_EXRDONLY);
4220 	MNT_FLAG(MNT_EXPORTED);
4221 	MNT_FLAG(MNT_DEFEXPORTED);
4222 	MNT_FLAG(MNT_EXPORTANON);
4223 	MNT_FLAG(MNT_EXKERB);
4224 	MNT_FLAG(MNT_EXPUBLIC);
4225 	MNT_FLAG(MNT_LOCAL);
4226 	MNT_FLAG(MNT_QUOTA);
4227 	MNT_FLAG(MNT_ROOTFS);
4228 	MNT_FLAG(MNT_USER);
4229 	MNT_FLAG(MNT_IGNORE);
4230 	MNT_FLAG(MNT_UPDATE);
4231 	MNT_FLAG(MNT_DELEXPORT);
4232 	MNT_FLAG(MNT_RELOAD);
4233 	MNT_FLAG(MNT_FORCE);
4234 	MNT_FLAG(MNT_SNAPSHOT);
4235 	MNT_FLAG(MNT_BYFSID);
4236 #undef MNT_FLAG
4237 	if (mflags != 0) {
4238 		if (buf[0] != '\0')
4239 			strlcat(buf, ", ", sizeof(buf));
4240 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4241 		    "0x%016jx", mflags);
4242 	}
4243 	db_printf("    mnt_flag = %s\n", buf);
4244 
4245 	buf[0] = '\0';
4246 	flags = mp->mnt_kern_flag;
4247 #define	MNT_KERN_FLAG(flag)	do {					\
4248 	if (flags & (flag)) {						\
4249 		if (buf[0] != '\0')					\
4250 			strlcat(buf, ", ", sizeof(buf));		\
4251 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4252 		flags &= ~(flag);					\
4253 	}								\
4254 } while (0)
4255 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4256 	MNT_KERN_FLAG(MNTK_ASYNC);
4257 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4258 	MNT_KERN_FLAG(MNTK_DRAINING);
4259 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4260 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4261 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4262 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4263 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4264 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4265 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4266 	MNT_KERN_FLAG(MNTK_MARKER);
4267 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4268 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4269 	MNT_KERN_FLAG(MNTK_NOASYNC);
4270 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4271 	MNT_KERN_FLAG(MNTK_MWAIT);
4272 	MNT_KERN_FLAG(MNTK_SUSPEND);
4273 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4274 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4275 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4276 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4277 #undef MNT_KERN_FLAG
4278 	if (flags != 0) {
4279 		if (buf[0] != '\0')
4280 			strlcat(buf, ", ", sizeof(buf));
4281 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4282 		    "0x%08x", flags);
4283 	}
4284 	db_printf("    mnt_kern_flag = %s\n", buf);
4285 
4286 	db_printf("    mnt_opt = ");
4287 	opt = TAILQ_FIRST(mp->mnt_opt);
4288 	if (opt != NULL) {
4289 		db_printf("%s", opt->name);
4290 		opt = TAILQ_NEXT(opt, link);
4291 		while (opt != NULL) {
4292 			db_printf(", %s", opt->name);
4293 			opt = TAILQ_NEXT(opt, link);
4294 		}
4295 	}
4296 	db_printf("\n");
4297 
4298 	sp = &mp->mnt_stat;
4299 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4300 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4301 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4302 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4303 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4304 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4305 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4306 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4307 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4308 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4309 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4310 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4311 
4312 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4313 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4314 	if (jailed(mp->mnt_cred))
4315 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4316 	db_printf(" }\n");
4317 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4318 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4319 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4320 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4321 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4322 	    mp->mnt_lazyvnodelistsize);
4323 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4324 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4325 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4326 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4327 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4328 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4329 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4330 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4331 	db_printf("    mnt_secondary_accwrites = %d\n",
4332 	    mp->mnt_secondary_accwrites);
4333 	db_printf("    mnt_gjprovider = %s\n",
4334 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4335 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4336 
4337 	db_printf("\n\nList of active vnodes\n");
4338 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4339 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4340 			vn_printf(vp, "vnode ");
4341 			if (db_pager_quit)
4342 				break;
4343 		}
4344 	}
4345 	db_printf("\n\nList of inactive vnodes\n");
4346 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4347 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4348 			vn_printf(vp, "vnode ");
4349 			if (db_pager_quit)
4350 				break;
4351 		}
4352 	}
4353 }
4354 #endif	/* DDB */
4355 
4356 /*
4357  * Fill in a struct xvfsconf based on a struct vfsconf.
4358  */
4359 static int
4360 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4361 {
4362 	struct xvfsconf xvfsp;
4363 
4364 	bzero(&xvfsp, sizeof(xvfsp));
4365 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4366 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4367 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4368 	xvfsp.vfc_flags = vfsp->vfc_flags;
4369 	/*
4370 	 * These are unused in userland, we keep them
4371 	 * to not break binary compatibility.
4372 	 */
4373 	xvfsp.vfc_vfsops = NULL;
4374 	xvfsp.vfc_next = NULL;
4375 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4376 }
4377 
4378 #ifdef COMPAT_FREEBSD32
4379 struct xvfsconf32 {
4380 	uint32_t	vfc_vfsops;
4381 	char		vfc_name[MFSNAMELEN];
4382 	int32_t		vfc_typenum;
4383 	int32_t		vfc_refcount;
4384 	int32_t		vfc_flags;
4385 	uint32_t	vfc_next;
4386 };
4387 
4388 static int
4389 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4390 {
4391 	struct xvfsconf32 xvfsp;
4392 
4393 	bzero(&xvfsp, sizeof(xvfsp));
4394 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4395 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4396 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4397 	xvfsp.vfc_flags = vfsp->vfc_flags;
4398 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4399 }
4400 #endif
4401 
4402 /*
4403  * Top level filesystem related information gathering.
4404  */
4405 static int
4406 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4407 {
4408 	struct vfsconf *vfsp;
4409 	int error;
4410 
4411 	error = 0;
4412 	vfsconf_slock();
4413 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4414 #ifdef COMPAT_FREEBSD32
4415 		if (req->flags & SCTL_MASK32)
4416 			error = vfsconf2x32(req, vfsp);
4417 		else
4418 #endif
4419 			error = vfsconf2x(req, vfsp);
4420 		if (error)
4421 			break;
4422 	}
4423 	vfsconf_sunlock();
4424 	return (error);
4425 }
4426 
4427 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4428     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4429     "S,xvfsconf", "List of all configured filesystems");
4430 
4431 #ifndef BURN_BRIDGES
4432 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4433 
4434 static int
4435 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4436 {
4437 	int *name = (int *)arg1 - 1;	/* XXX */
4438 	u_int namelen = arg2 + 1;	/* XXX */
4439 	struct vfsconf *vfsp;
4440 
4441 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4442 	    "please rebuild world\n");
4443 
4444 #if 1 || defined(COMPAT_PRELITE2)
4445 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4446 	if (namelen == 1)
4447 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4448 #endif
4449 
4450 	switch (name[1]) {
4451 	case VFS_MAXTYPENUM:
4452 		if (namelen != 2)
4453 			return (ENOTDIR);
4454 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4455 	case VFS_CONF:
4456 		if (namelen != 3)
4457 			return (ENOTDIR);	/* overloaded */
4458 		vfsconf_slock();
4459 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4460 			if (vfsp->vfc_typenum == name[2])
4461 				break;
4462 		}
4463 		vfsconf_sunlock();
4464 		if (vfsp == NULL)
4465 			return (EOPNOTSUPP);
4466 #ifdef COMPAT_FREEBSD32
4467 		if (req->flags & SCTL_MASK32)
4468 			return (vfsconf2x32(req, vfsp));
4469 		else
4470 #endif
4471 			return (vfsconf2x(req, vfsp));
4472 	}
4473 	return (EOPNOTSUPP);
4474 }
4475 
4476 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4477     CTLFLAG_MPSAFE, vfs_sysctl,
4478     "Generic filesystem");
4479 
4480 #if 1 || defined(COMPAT_PRELITE2)
4481 
4482 static int
4483 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4484 {
4485 	int error;
4486 	struct vfsconf *vfsp;
4487 	struct ovfsconf ovfs;
4488 
4489 	vfsconf_slock();
4490 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4491 		bzero(&ovfs, sizeof(ovfs));
4492 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4493 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4494 		ovfs.vfc_index = vfsp->vfc_typenum;
4495 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4496 		ovfs.vfc_flags = vfsp->vfc_flags;
4497 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4498 		if (error != 0) {
4499 			vfsconf_sunlock();
4500 			return (error);
4501 		}
4502 	}
4503 	vfsconf_sunlock();
4504 	return (0);
4505 }
4506 
4507 #endif /* 1 || COMPAT_PRELITE2 */
4508 #endif /* !BURN_BRIDGES */
4509 
4510 #define KINFO_VNODESLOP		10
4511 #ifdef notyet
4512 /*
4513  * Dump vnode list (via sysctl).
4514  */
4515 /* ARGSUSED */
4516 static int
4517 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4518 {
4519 	struct xvnode *xvn;
4520 	struct mount *mp;
4521 	struct vnode *vp;
4522 	int error, len, n;
4523 
4524 	/*
4525 	 * Stale numvnodes access is not fatal here.
4526 	 */
4527 	req->lock = 0;
4528 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4529 	if (!req->oldptr)
4530 		/* Make an estimate */
4531 		return (SYSCTL_OUT(req, 0, len));
4532 
4533 	error = sysctl_wire_old_buffer(req, 0);
4534 	if (error != 0)
4535 		return (error);
4536 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4537 	n = 0;
4538 	mtx_lock(&mountlist_mtx);
4539 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4540 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4541 			continue;
4542 		MNT_ILOCK(mp);
4543 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4544 			if (n == len)
4545 				break;
4546 			vref(vp);
4547 			xvn[n].xv_size = sizeof *xvn;
4548 			xvn[n].xv_vnode = vp;
4549 			xvn[n].xv_id = 0;	/* XXX compat */
4550 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4551 			XV_COPY(usecount);
4552 			XV_COPY(writecount);
4553 			XV_COPY(holdcnt);
4554 			XV_COPY(mount);
4555 			XV_COPY(numoutput);
4556 			XV_COPY(type);
4557 #undef XV_COPY
4558 			xvn[n].xv_flag = vp->v_vflag;
4559 
4560 			switch (vp->v_type) {
4561 			case VREG:
4562 			case VDIR:
4563 			case VLNK:
4564 				break;
4565 			case VBLK:
4566 			case VCHR:
4567 				if (vp->v_rdev == NULL) {
4568 					vrele(vp);
4569 					continue;
4570 				}
4571 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4572 				break;
4573 			case VSOCK:
4574 				xvn[n].xv_socket = vp->v_socket;
4575 				break;
4576 			case VFIFO:
4577 				xvn[n].xv_fifo = vp->v_fifoinfo;
4578 				break;
4579 			case VNON:
4580 			case VBAD:
4581 			default:
4582 				/* shouldn't happen? */
4583 				vrele(vp);
4584 				continue;
4585 			}
4586 			vrele(vp);
4587 			++n;
4588 		}
4589 		MNT_IUNLOCK(mp);
4590 		mtx_lock(&mountlist_mtx);
4591 		vfs_unbusy(mp);
4592 		if (n == len)
4593 			break;
4594 	}
4595 	mtx_unlock(&mountlist_mtx);
4596 
4597 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4598 	free(xvn, M_TEMP);
4599 	return (error);
4600 }
4601 
4602 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4603     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4604     "");
4605 #endif
4606 
4607 static void
4608 unmount_or_warn(struct mount *mp)
4609 {
4610 	int error;
4611 
4612 	error = dounmount(mp, MNT_FORCE, curthread);
4613 	if (error != 0) {
4614 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4615 		if (error == EBUSY)
4616 			printf("BUSY)\n");
4617 		else
4618 			printf("%d)\n", error);
4619 	}
4620 }
4621 
4622 /*
4623  * Unmount all filesystems. The list is traversed in reverse order
4624  * of mounting to avoid dependencies.
4625  */
4626 void
4627 vfs_unmountall(void)
4628 {
4629 	struct mount *mp, *tmp;
4630 
4631 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4632 
4633 	/*
4634 	 * Since this only runs when rebooting, it is not interlocked.
4635 	 */
4636 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4637 		vfs_ref(mp);
4638 
4639 		/*
4640 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4641 		 * unmount of the latter.
4642 		 */
4643 		if (mp == rootdevmp)
4644 			continue;
4645 
4646 		unmount_or_warn(mp);
4647 	}
4648 
4649 	if (rootdevmp != NULL)
4650 		unmount_or_warn(rootdevmp);
4651 }
4652 
4653 static void
4654 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4655 {
4656 
4657 	ASSERT_VI_LOCKED(vp, __func__);
4658 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4659 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4660 		vdropl(vp);
4661 		return;
4662 	}
4663 	if (vn_lock(vp, lkflags) == 0) {
4664 		VI_LOCK(vp);
4665 		vinactive(vp);
4666 		VOP_UNLOCK(vp);
4667 		vdropl(vp);
4668 		return;
4669 	}
4670 	vdefer_inactive_unlocked(vp);
4671 }
4672 
4673 static int
4674 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4675 {
4676 
4677 	return (vp->v_iflag & VI_DEFINACT);
4678 }
4679 
4680 static void __noinline
4681 vfs_periodic_inactive(struct mount *mp, int flags)
4682 {
4683 	struct vnode *vp, *mvp;
4684 	int lkflags;
4685 
4686 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4687 	if (flags != MNT_WAIT)
4688 		lkflags |= LK_NOWAIT;
4689 
4690 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4691 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4692 			VI_UNLOCK(vp);
4693 			continue;
4694 		}
4695 		vp->v_iflag &= ~VI_DEFINACT;
4696 		vfs_deferred_inactive(vp, lkflags);
4697 	}
4698 }
4699 
4700 static inline bool
4701 vfs_want_msync(struct vnode *vp)
4702 {
4703 	struct vm_object *obj;
4704 
4705 	/*
4706 	 * This test may be performed without any locks held.
4707 	 * We rely on vm_object's type stability.
4708 	 */
4709 	if (vp->v_vflag & VV_NOSYNC)
4710 		return (false);
4711 	obj = vp->v_object;
4712 	return (obj != NULL && vm_object_mightbedirty(obj));
4713 }
4714 
4715 static int
4716 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4717 {
4718 
4719 	if (vp->v_vflag & VV_NOSYNC)
4720 		return (false);
4721 	if (vp->v_iflag & VI_DEFINACT)
4722 		return (true);
4723 	return (vfs_want_msync(vp));
4724 }
4725 
4726 static void __noinline
4727 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4728 {
4729 	struct vnode *vp, *mvp;
4730 	struct vm_object *obj;
4731 	int lkflags, objflags;
4732 	bool seen_defer;
4733 
4734 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4735 	if (flags != MNT_WAIT) {
4736 		lkflags |= LK_NOWAIT;
4737 		objflags = OBJPC_NOSYNC;
4738 	} else {
4739 		objflags = OBJPC_SYNC;
4740 	}
4741 
4742 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4743 		seen_defer = false;
4744 		if (vp->v_iflag & VI_DEFINACT) {
4745 			vp->v_iflag &= ~VI_DEFINACT;
4746 			seen_defer = true;
4747 		}
4748 		if (!vfs_want_msync(vp)) {
4749 			if (seen_defer)
4750 				vfs_deferred_inactive(vp, lkflags);
4751 			else
4752 				VI_UNLOCK(vp);
4753 			continue;
4754 		}
4755 		if (vget(vp, lkflags) == 0) {
4756 			obj = vp->v_object;
4757 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4758 				VM_OBJECT_WLOCK(obj);
4759 				vm_object_page_clean(obj, 0, 0, objflags);
4760 				VM_OBJECT_WUNLOCK(obj);
4761 			}
4762 			vput(vp);
4763 			if (seen_defer)
4764 				vdrop(vp);
4765 		} else {
4766 			if (seen_defer)
4767 				vdefer_inactive_unlocked(vp);
4768 		}
4769 	}
4770 }
4771 
4772 void
4773 vfs_periodic(struct mount *mp, int flags)
4774 {
4775 
4776 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4777 
4778 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4779 		vfs_periodic_inactive(mp, flags);
4780 	else
4781 		vfs_periodic_msync_inactive(mp, flags);
4782 }
4783 
4784 static void
4785 destroy_vpollinfo_free(struct vpollinfo *vi)
4786 {
4787 
4788 	knlist_destroy(&vi->vpi_selinfo.si_note);
4789 	mtx_destroy(&vi->vpi_lock);
4790 	free(vi, M_VNODEPOLL);
4791 }
4792 
4793 static void
4794 destroy_vpollinfo(struct vpollinfo *vi)
4795 {
4796 
4797 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4798 	seldrain(&vi->vpi_selinfo);
4799 	destroy_vpollinfo_free(vi);
4800 }
4801 
4802 /*
4803  * Initialize per-vnode helper structure to hold poll-related state.
4804  */
4805 void
4806 v_addpollinfo(struct vnode *vp)
4807 {
4808 	struct vpollinfo *vi;
4809 
4810 	if (vp->v_pollinfo != NULL)
4811 		return;
4812 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4813 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4814 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4815 	    vfs_knlunlock, vfs_knl_assert_lock);
4816 	VI_LOCK(vp);
4817 	if (vp->v_pollinfo != NULL) {
4818 		VI_UNLOCK(vp);
4819 		destroy_vpollinfo_free(vi);
4820 		return;
4821 	}
4822 	vp->v_pollinfo = vi;
4823 	VI_UNLOCK(vp);
4824 }
4825 
4826 /*
4827  * Record a process's interest in events which might happen to
4828  * a vnode.  Because poll uses the historic select-style interface
4829  * internally, this routine serves as both the ``check for any
4830  * pending events'' and the ``record my interest in future events''
4831  * functions.  (These are done together, while the lock is held,
4832  * to avoid race conditions.)
4833  */
4834 int
4835 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4836 {
4837 
4838 	v_addpollinfo(vp);
4839 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4840 	if (vp->v_pollinfo->vpi_revents & events) {
4841 		/*
4842 		 * This leaves events we are not interested
4843 		 * in available for the other process which
4844 		 * which presumably had requested them
4845 		 * (otherwise they would never have been
4846 		 * recorded).
4847 		 */
4848 		events &= vp->v_pollinfo->vpi_revents;
4849 		vp->v_pollinfo->vpi_revents &= ~events;
4850 
4851 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4852 		return (events);
4853 	}
4854 	vp->v_pollinfo->vpi_events |= events;
4855 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4856 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4857 	return (0);
4858 }
4859 
4860 /*
4861  * Routine to create and manage a filesystem syncer vnode.
4862  */
4863 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4864 static int	sync_fsync(struct  vop_fsync_args *);
4865 static int	sync_inactive(struct  vop_inactive_args *);
4866 static int	sync_reclaim(struct  vop_reclaim_args *);
4867 
4868 static struct vop_vector sync_vnodeops = {
4869 	.vop_bypass =	VOP_EOPNOTSUPP,
4870 	.vop_close =	sync_close,		/* close */
4871 	.vop_fsync =	sync_fsync,		/* fsync */
4872 	.vop_inactive =	sync_inactive,	/* inactive */
4873 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4874 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4875 	.vop_lock1 =	vop_stdlock,	/* lock */
4876 	.vop_unlock =	vop_stdunlock,	/* unlock */
4877 	.vop_islocked =	vop_stdislocked,	/* islocked */
4878 };
4879 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4880 
4881 /*
4882  * Create a new filesystem syncer vnode for the specified mount point.
4883  */
4884 void
4885 vfs_allocate_syncvnode(struct mount *mp)
4886 {
4887 	struct vnode *vp;
4888 	struct bufobj *bo;
4889 	static long start, incr, next;
4890 	int error;
4891 
4892 	/* Allocate a new vnode */
4893 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4894 	if (error != 0)
4895 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4896 	vp->v_type = VNON;
4897 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4898 	vp->v_vflag |= VV_FORCEINSMQ;
4899 	error = insmntque(vp, mp);
4900 	if (error != 0)
4901 		panic("vfs_allocate_syncvnode: insmntque() failed");
4902 	vp->v_vflag &= ~VV_FORCEINSMQ;
4903 	VOP_UNLOCK(vp);
4904 	/*
4905 	 * Place the vnode onto the syncer worklist. We attempt to
4906 	 * scatter them about on the list so that they will go off
4907 	 * at evenly distributed times even if all the filesystems
4908 	 * are mounted at once.
4909 	 */
4910 	next += incr;
4911 	if (next == 0 || next > syncer_maxdelay) {
4912 		start /= 2;
4913 		incr /= 2;
4914 		if (start == 0) {
4915 			start = syncer_maxdelay / 2;
4916 			incr = syncer_maxdelay;
4917 		}
4918 		next = start;
4919 	}
4920 	bo = &vp->v_bufobj;
4921 	BO_LOCK(bo);
4922 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4923 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4924 	mtx_lock(&sync_mtx);
4925 	sync_vnode_count++;
4926 	if (mp->mnt_syncer == NULL) {
4927 		mp->mnt_syncer = vp;
4928 		vp = NULL;
4929 	}
4930 	mtx_unlock(&sync_mtx);
4931 	BO_UNLOCK(bo);
4932 	if (vp != NULL) {
4933 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4934 		vgone(vp);
4935 		vput(vp);
4936 	}
4937 }
4938 
4939 void
4940 vfs_deallocate_syncvnode(struct mount *mp)
4941 {
4942 	struct vnode *vp;
4943 
4944 	mtx_lock(&sync_mtx);
4945 	vp = mp->mnt_syncer;
4946 	if (vp != NULL)
4947 		mp->mnt_syncer = NULL;
4948 	mtx_unlock(&sync_mtx);
4949 	if (vp != NULL)
4950 		vrele(vp);
4951 }
4952 
4953 /*
4954  * Do a lazy sync of the filesystem.
4955  */
4956 static int
4957 sync_fsync(struct vop_fsync_args *ap)
4958 {
4959 	struct vnode *syncvp = ap->a_vp;
4960 	struct mount *mp = syncvp->v_mount;
4961 	int error, save;
4962 	struct bufobj *bo;
4963 
4964 	/*
4965 	 * We only need to do something if this is a lazy evaluation.
4966 	 */
4967 	if (ap->a_waitfor != MNT_LAZY)
4968 		return (0);
4969 
4970 	/*
4971 	 * Move ourselves to the back of the sync list.
4972 	 */
4973 	bo = &syncvp->v_bufobj;
4974 	BO_LOCK(bo);
4975 	vn_syncer_add_to_worklist(bo, syncdelay);
4976 	BO_UNLOCK(bo);
4977 
4978 	/*
4979 	 * Walk the list of vnodes pushing all that are dirty and
4980 	 * not already on the sync list.
4981 	 */
4982 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4983 		return (0);
4984 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4985 		vfs_unbusy(mp);
4986 		return (0);
4987 	}
4988 	save = curthread_pflags_set(TDP_SYNCIO);
4989 	/*
4990 	 * The filesystem at hand may be idle with free vnodes stored in the
4991 	 * batch.  Return them instead of letting them stay there indefinitely.
4992 	 */
4993 	vfs_periodic(mp, MNT_NOWAIT);
4994 	error = VFS_SYNC(mp, MNT_LAZY);
4995 	curthread_pflags_restore(save);
4996 	vn_finished_write(mp);
4997 	vfs_unbusy(mp);
4998 	return (error);
4999 }
5000 
5001 /*
5002  * The syncer vnode is no referenced.
5003  */
5004 static int
5005 sync_inactive(struct vop_inactive_args *ap)
5006 {
5007 
5008 	vgone(ap->a_vp);
5009 	return (0);
5010 }
5011 
5012 /*
5013  * The syncer vnode is no longer needed and is being decommissioned.
5014  *
5015  * Modifications to the worklist must be protected by sync_mtx.
5016  */
5017 static int
5018 sync_reclaim(struct vop_reclaim_args *ap)
5019 {
5020 	struct vnode *vp = ap->a_vp;
5021 	struct bufobj *bo;
5022 
5023 	bo = &vp->v_bufobj;
5024 	BO_LOCK(bo);
5025 	mtx_lock(&sync_mtx);
5026 	if (vp->v_mount->mnt_syncer == vp)
5027 		vp->v_mount->mnt_syncer = NULL;
5028 	if (bo->bo_flag & BO_ONWORKLST) {
5029 		LIST_REMOVE(bo, bo_synclist);
5030 		syncer_worklist_len--;
5031 		sync_vnode_count--;
5032 		bo->bo_flag &= ~BO_ONWORKLST;
5033 	}
5034 	mtx_unlock(&sync_mtx);
5035 	BO_UNLOCK(bo);
5036 
5037 	return (0);
5038 }
5039 
5040 int
5041 vn_need_pageq_flush(struct vnode *vp)
5042 {
5043 	struct vm_object *obj;
5044 	int need;
5045 
5046 	MPASS(mtx_owned(VI_MTX(vp)));
5047 	need = 0;
5048 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5049 	    vm_object_mightbedirty(obj))
5050 		need = 1;
5051 	return (need);
5052 }
5053 
5054 /*
5055  * Check if vnode represents a disk device
5056  */
5057 bool
5058 vn_isdisk_error(struct vnode *vp, int *errp)
5059 {
5060 	int error;
5061 
5062 	if (vp->v_type != VCHR) {
5063 		error = ENOTBLK;
5064 		goto out;
5065 	}
5066 	error = 0;
5067 	dev_lock();
5068 	if (vp->v_rdev == NULL)
5069 		error = ENXIO;
5070 	else if (vp->v_rdev->si_devsw == NULL)
5071 		error = ENXIO;
5072 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5073 		error = ENOTBLK;
5074 	dev_unlock();
5075 out:
5076 	*errp = error;
5077 	return (error == 0);
5078 }
5079 
5080 bool
5081 vn_isdisk(struct vnode *vp)
5082 {
5083 	int error;
5084 
5085 	return (vn_isdisk_error(vp, &error));
5086 }
5087 
5088 /*
5089  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5090  * the comment above cache_fplookup for details.
5091  */
5092 int
5093 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5094 {
5095 	int error;
5096 
5097 	VFS_SMR_ASSERT_ENTERED();
5098 
5099 	/* Check the owner. */
5100 	if (cred->cr_uid == file_uid) {
5101 		if (file_mode & S_IXUSR)
5102 			return (0);
5103 		goto out_error;
5104 	}
5105 
5106 	/* Otherwise, check the groups (first match) */
5107 	if (groupmember(file_gid, cred)) {
5108 		if (file_mode & S_IXGRP)
5109 			return (0);
5110 		goto out_error;
5111 	}
5112 
5113 	/* Otherwise, check everyone else. */
5114 	if (file_mode & S_IXOTH)
5115 		return (0);
5116 out_error:
5117 	/*
5118 	 * Permission check failed, but it is possible denial will get overwritten
5119 	 * (e.g., when root is traversing through a 700 directory owned by someone
5120 	 * else).
5121 	 *
5122 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5123 	 * modules overriding this result. It's quite unclear what semantics
5124 	 * are allowed for them to operate, thus for safety we don't call them
5125 	 * from within the SMR section. This also means if any such modules
5126 	 * are present, we have to let the regular lookup decide.
5127 	 */
5128 	error = priv_check_cred_vfs_lookup_nomac(cred);
5129 	switch (error) {
5130 	case 0:
5131 		return (0);
5132 	case EAGAIN:
5133 		/*
5134 		 * MAC modules present.
5135 		 */
5136 		return (EAGAIN);
5137 	case EPERM:
5138 		return (EACCES);
5139 	default:
5140 		return (error);
5141 	}
5142 }
5143 
5144 /*
5145  * Common filesystem object access control check routine.  Accepts a
5146  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5147  * Returns 0 on success, or an errno on failure.
5148  */
5149 int
5150 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5151     accmode_t accmode, struct ucred *cred)
5152 {
5153 	accmode_t dac_granted;
5154 	accmode_t priv_granted;
5155 
5156 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5157 	    ("invalid bit in accmode"));
5158 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5159 	    ("VAPPEND without VWRITE"));
5160 
5161 	/*
5162 	 * Look for a normal, non-privileged way to access the file/directory
5163 	 * as requested.  If it exists, go with that.
5164 	 */
5165 
5166 	dac_granted = 0;
5167 
5168 	/* Check the owner. */
5169 	if (cred->cr_uid == file_uid) {
5170 		dac_granted |= VADMIN;
5171 		if (file_mode & S_IXUSR)
5172 			dac_granted |= VEXEC;
5173 		if (file_mode & S_IRUSR)
5174 			dac_granted |= VREAD;
5175 		if (file_mode & S_IWUSR)
5176 			dac_granted |= (VWRITE | VAPPEND);
5177 
5178 		if ((accmode & dac_granted) == accmode)
5179 			return (0);
5180 
5181 		goto privcheck;
5182 	}
5183 
5184 	/* Otherwise, check the groups (first match) */
5185 	if (groupmember(file_gid, cred)) {
5186 		if (file_mode & S_IXGRP)
5187 			dac_granted |= VEXEC;
5188 		if (file_mode & S_IRGRP)
5189 			dac_granted |= VREAD;
5190 		if (file_mode & S_IWGRP)
5191 			dac_granted |= (VWRITE | VAPPEND);
5192 
5193 		if ((accmode & dac_granted) == accmode)
5194 			return (0);
5195 
5196 		goto privcheck;
5197 	}
5198 
5199 	/* Otherwise, check everyone else. */
5200 	if (file_mode & S_IXOTH)
5201 		dac_granted |= VEXEC;
5202 	if (file_mode & S_IROTH)
5203 		dac_granted |= VREAD;
5204 	if (file_mode & S_IWOTH)
5205 		dac_granted |= (VWRITE | VAPPEND);
5206 	if ((accmode & dac_granted) == accmode)
5207 		return (0);
5208 
5209 privcheck:
5210 	/*
5211 	 * Build a privilege mask to determine if the set of privileges
5212 	 * satisfies the requirements when combined with the granted mask
5213 	 * from above.  For each privilege, if the privilege is required,
5214 	 * bitwise or the request type onto the priv_granted mask.
5215 	 */
5216 	priv_granted = 0;
5217 
5218 	if (type == VDIR) {
5219 		/*
5220 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5221 		 * requests, instead of PRIV_VFS_EXEC.
5222 		 */
5223 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5224 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5225 			priv_granted |= VEXEC;
5226 	} else {
5227 		/*
5228 		 * Ensure that at least one execute bit is on. Otherwise,
5229 		 * a privileged user will always succeed, and we don't want
5230 		 * this to happen unless the file really is executable.
5231 		 */
5232 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5233 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5234 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5235 			priv_granted |= VEXEC;
5236 	}
5237 
5238 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5239 	    !priv_check_cred(cred, PRIV_VFS_READ))
5240 		priv_granted |= VREAD;
5241 
5242 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5243 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5244 		priv_granted |= (VWRITE | VAPPEND);
5245 
5246 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5247 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5248 		priv_granted |= VADMIN;
5249 
5250 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5251 		return (0);
5252 	}
5253 
5254 	return ((accmode & VADMIN) ? EPERM : EACCES);
5255 }
5256 
5257 /*
5258  * Credential check based on process requesting service, and per-attribute
5259  * permissions.
5260  */
5261 int
5262 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5263     struct thread *td, accmode_t accmode)
5264 {
5265 
5266 	/*
5267 	 * Kernel-invoked always succeeds.
5268 	 */
5269 	if (cred == NOCRED)
5270 		return (0);
5271 
5272 	/*
5273 	 * Do not allow privileged processes in jail to directly manipulate
5274 	 * system attributes.
5275 	 */
5276 	switch (attrnamespace) {
5277 	case EXTATTR_NAMESPACE_SYSTEM:
5278 		/* Potentially should be: return (EPERM); */
5279 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5280 	case EXTATTR_NAMESPACE_USER:
5281 		return (VOP_ACCESS(vp, accmode, cred, td));
5282 	default:
5283 		return (EPERM);
5284 	}
5285 }
5286 
5287 #ifdef DEBUG_VFS_LOCKS
5288 /*
5289  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5290  * no longer ok to have an unlocked VFS.
5291  */
5292 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5293 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5294 
5295 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5296 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5297     "Drop into debugger on lock violation");
5298 
5299 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5300 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5301     0, "Check for interlock across VOPs");
5302 
5303 int vfs_badlock_print = 1;	/* Print lock violations. */
5304 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5305     0, "Print lock violations");
5306 
5307 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5308 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5309     0, "Print vnode details on lock violations");
5310 
5311 #ifdef KDB
5312 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5313 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5314     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5315 #endif
5316 
5317 static void
5318 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5319 {
5320 
5321 #ifdef KDB
5322 	if (vfs_badlock_backtrace)
5323 		kdb_backtrace();
5324 #endif
5325 	if (vfs_badlock_vnode)
5326 		vn_printf(vp, "vnode ");
5327 	if (vfs_badlock_print)
5328 		printf("%s: %p %s\n", str, (void *)vp, msg);
5329 	if (vfs_badlock_ddb)
5330 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5331 }
5332 
5333 void
5334 assert_vi_locked(struct vnode *vp, const char *str)
5335 {
5336 
5337 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5338 		vfs_badlock("interlock is not locked but should be", str, vp);
5339 }
5340 
5341 void
5342 assert_vi_unlocked(struct vnode *vp, const char *str)
5343 {
5344 
5345 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5346 		vfs_badlock("interlock is locked but should not be", str, vp);
5347 }
5348 
5349 void
5350 assert_vop_locked(struct vnode *vp, const char *str)
5351 {
5352 	int locked;
5353 
5354 	if (!IGNORE_LOCK(vp)) {
5355 		locked = VOP_ISLOCKED(vp);
5356 		if (locked == 0 || locked == LK_EXCLOTHER)
5357 			vfs_badlock("is not locked but should be", str, vp);
5358 	}
5359 }
5360 
5361 void
5362 assert_vop_unlocked(struct vnode *vp, const char *str)
5363 {
5364 
5365 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5366 		vfs_badlock("is locked but should not be", str, vp);
5367 }
5368 
5369 void
5370 assert_vop_elocked(struct vnode *vp, const char *str)
5371 {
5372 
5373 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5374 		vfs_badlock("is not exclusive locked but should be", str, vp);
5375 }
5376 #endif /* DEBUG_VFS_LOCKS */
5377 
5378 void
5379 vop_rename_fail(struct vop_rename_args *ap)
5380 {
5381 
5382 	if (ap->a_tvp != NULL)
5383 		vput(ap->a_tvp);
5384 	if (ap->a_tdvp == ap->a_tvp)
5385 		vrele(ap->a_tdvp);
5386 	else
5387 		vput(ap->a_tdvp);
5388 	vrele(ap->a_fdvp);
5389 	vrele(ap->a_fvp);
5390 }
5391 
5392 void
5393 vop_rename_pre(void *ap)
5394 {
5395 	struct vop_rename_args *a = ap;
5396 
5397 #ifdef DEBUG_VFS_LOCKS
5398 	if (a->a_tvp)
5399 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5400 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5401 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5402 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5403 
5404 	/* Check the source (from). */
5405 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5406 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5407 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5408 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5409 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5410 
5411 	/* Check the target. */
5412 	if (a->a_tvp)
5413 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5414 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5415 #endif
5416 	/*
5417 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5418 	 * in vop_rename_post but that's not going to work out since some
5419 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5420 	 *
5421 	 * For now filesystems are expected to do the relevant calls after they
5422 	 * decide what vnodes to operate on.
5423 	 */
5424 	if (a->a_tdvp != a->a_fdvp)
5425 		vhold(a->a_fdvp);
5426 	if (a->a_tvp != a->a_fvp)
5427 		vhold(a->a_fvp);
5428 	vhold(a->a_tdvp);
5429 	if (a->a_tvp)
5430 		vhold(a->a_tvp);
5431 }
5432 
5433 #ifdef DEBUG_VFS_LOCKS
5434 void
5435 vop_fplookup_vexec_debugpre(void *ap __unused)
5436 {
5437 
5438 	VFS_SMR_ASSERT_ENTERED();
5439 }
5440 
5441 void
5442 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5443 {
5444 
5445 	VFS_SMR_ASSERT_ENTERED();
5446 }
5447 
5448 void
5449 vop_strategy_debugpre(void *ap)
5450 {
5451 	struct vop_strategy_args *a;
5452 	struct buf *bp;
5453 
5454 	a = ap;
5455 	bp = a->a_bp;
5456 
5457 	/*
5458 	 * Cluster ops lock their component buffers but not the IO container.
5459 	 */
5460 	if ((bp->b_flags & B_CLUSTER) != 0)
5461 		return;
5462 
5463 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5464 		if (vfs_badlock_print)
5465 			printf(
5466 			    "VOP_STRATEGY: bp is not locked but should be\n");
5467 		if (vfs_badlock_ddb)
5468 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5469 	}
5470 }
5471 
5472 void
5473 vop_lock_debugpre(void *ap)
5474 {
5475 	struct vop_lock1_args *a = ap;
5476 
5477 	if ((a->a_flags & LK_INTERLOCK) == 0)
5478 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5479 	else
5480 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5481 }
5482 
5483 void
5484 vop_lock_debugpost(void *ap, int rc)
5485 {
5486 	struct vop_lock1_args *a = ap;
5487 
5488 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5489 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5490 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5491 }
5492 
5493 void
5494 vop_unlock_debugpre(void *ap)
5495 {
5496 	struct vop_unlock_args *a = ap;
5497 
5498 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5499 }
5500 
5501 void
5502 vop_need_inactive_debugpre(void *ap)
5503 {
5504 	struct vop_need_inactive_args *a = ap;
5505 
5506 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5507 }
5508 
5509 void
5510 vop_need_inactive_debugpost(void *ap, int rc)
5511 {
5512 	struct vop_need_inactive_args *a = ap;
5513 
5514 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5515 }
5516 #endif
5517 
5518 void
5519 vop_create_pre(void *ap)
5520 {
5521 	struct vop_create_args *a;
5522 	struct vnode *dvp;
5523 
5524 	a = ap;
5525 	dvp = a->a_dvp;
5526 	vn_seqc_write_begin(dvp);
5527 }
5528 
5529 void
5530 vop_create_post(void *ap, int rc)
5531 {
5532 	struct vop_create_args *a;
5533 	struct vnode *dvp;
5534 
5535 	a = ap;
5536 	dvp = a->a_dvp;
5537 	vn_seqc_write_end(dvp);
5538 	if (!rc)
5539 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5540 }
5541 
5542 void
5543 vop_whiteout_pre(void *ap)
5544 {
5545 	struct vop_whiteout_args *a;
5546 	struct vnode *dvp;
5547 
5548 	a = ap;
5549 	dvp = a->a_dvp;
5550 	vn_seqc_write_begin(dvp);
5551 }
5552 
5553 void
5554 vop_whiteout_post(void *ap, int rc)
5555 {
5556 	struct vop_whiteout_args *a;
5557 	struct vnode *dvp;
5558 
5559 	a = ap;
5560 	dvp = a->a_dvp;
5561 	vn_seqc_write_end(dvp);
5562 }
5563 
5564 void
5565 vop_deleteextattr_pre(void *ap)
5566 {
5567 	struct vop_deleteextattr_args *a;
5568 	struct vnode *vp;
5569 
5570 	a = ap;
5571 	vp = a->a_vp;
5572 	vn_seqc_write_begin(vp);
5573 }
5574 
5575 void
5576 vop_deleteextattr_post(void *ap, int rc)
5577 {
5578 	struct vop_deleteextattr_args *a;
5579 	struct vnode *vp;
5580 
5581 	a = ap;
5582 	vp = a->a_vp;
5583 	vn_seqc_write_end(vp);
5584 	if (!rc)
5585 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5586 }
5587 
5588 void
5589 vop_link_pre(void *ap)
5590 {
5591 	struct vop_link_args *a;
5592 	struct vnode *vp, *tdvp;
5593 
5594 	a = ap;
5595 	vp = a->a_vp;
5596 	tdvp = a->a_tdvp;
5597 	vn_seqc_write_begin(vp);
5598 	vn_seqc_write_begin(tdvp);
5599 }
5600 
5601 void
5602 vop_link_post(void *ap, int rc)
5603 {
5604 	struct vop_link_args *a;
5605 	struct vnode *vp, *tdvp;
5606 
5607 	a = ap;
5608 	vp = a->a_vp;
5609 	tdvp = a->a_tdvp;
5610 	vn_seqc_write_end(vp);
5611 	vn_seqc_write_end(tdvp);
5612 	if (!rc) {
5613 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5614 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5615 	}
5616 }
5617 
5618 void
5619 vop_mkdir_pre(void *ap)
5620 {
5621 	struct vop_mkdir_args *a;
5622 	struct vnode *dvp;
5623 
5624 	a = ap;
5625 	dvp = a->a_dvp;
5626 	vn_seqc_write_begin(dvp);
5627 }
5628 
5629 void
5630 vop_mkdir_post(void *ap, int rc)
5631 {
5632 	struct vop_mkdir_args *a;
5633 	struct vnode *dvp;
5634 
5635 	a = ap;
5636 	dvp = a->a_dvp;
5637 	vn_seqc_write_end(dvp);
5638 	if (!rc)
5639 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5640 }
5641 
5642 #ifdef DEBUG_VFS_LOCKS
5643 void
5644 vop_mkdir_debugpost(void *ap, int rc)
5645 {
5646 	struct vop_mkdir_args *a;
5647 
5648 	a = ap;
5649 	if (!rc)
5650 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5651 }
5652 #endif
5653 
5654 void
5655 vop_mknod_pre(void *ap)
5656 {
5657 	struct vop_mknod_args *a;
5658 	struct vnode *dvp;
5659 
5660 	a = ap;
5661 	dvp = a->a_dvp;
5662 	vn_seqc_write_begin(dvp);
5663 }
5664 
5665 void
5666 vop_mknod_post(void *ap, int rc)
5667 {
5668 	struct vop_mknod_args *a;
5669 	struct vnode *dvp;
5670 
5671 	a = ap;
5672 	dvp = a->a_dvp;
5673 	vn_seqc_write_end(dvp);
5674 	if (!rc)
5675 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5676 }
5677 
5678 void
5679 vop_reclaim_post(void *ap, int rc)
5680 {
5681 	struct vop_reclaim_args *a;
5682 	struct vnode *vp;
5683 
5684 	a = ap;
5685 	vp = a->a_vp;
5686 	ASSERT_VOP_IN_SEQC(vp);
5687 	if (!rc)
5688 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5689 }
5690 
5691 void
5692 vop_remove_pre(void *ap)
5693 {
5694 	struct vop_remove_args *a;
5695 	struct vnode *dvp, *vp;
5696 
5697 	a = ap;
5698 	dvp = a->a_dvp;
5699 	vp = a->a_vp;
5700 	vn_seqc_write_begin(dvp);
5701 	vn_seqc_write_begin(vp);
5702 }
5703 
5704 void
5705 vop_remove_post(void *ap, int rc)
5706 {
5707 	struct vop_remove_args *a;
5708 	struct vnode *dvp, *vp;
5709 
5710 	a = ap;
5711 	dvp = a->a_dvp;
5712 	vp = a->a_vp;
5713 	vn_seqc_write_end(dvp);
5714 	vn_seqc_write_end(vp);
5715 	if (!rc) {
5716 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5717 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5718 	}
5719 }
5720 
5721 void
5722 vop_rename_post(void *ap, int rc)
5723 {
5724 	struct vop_rename_args *a = ap;
5725 	long hint;
5726 
5727 	if (!rc) {
5728 		hint = NOTE_WRITE;
5729 		if (a->a_fdvp == a->a_tdvp) {
5730 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5731 				hint |= NOTE_LINK;
5732 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5733 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5734 		} else {
5735 			hint |= NOTE_EXTEND;
5736 			if (a->a_fvp->v_type == VDIR)
5737 				hint |= NOTE_LINK;
5738 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5739 
5740 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5741 			    a->a_tvp->v_type == VDIR)
5742 				hint &= ~NOTE_LINK;
5743 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5744 		}
5745 
5746 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5747 		if (a->a_tvp)
5748 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5749 	}
5750 	if (a->a_tdvp != a->a_fdvp)
5751 		vdrop(a->a_fdvp);
5752 	if (a->a_tvp != a->a_fvp)
5753 		vdrop(a->a_fvp);
5754 	vdrop(a->a_tdvp);
5755 	if (a->a_tvp)
5756 		vdrop(a->a_tvp);
5757 }
5758 
5759 void
5760 vop_rmdir_pre(void *ap)
5761 {
5762 	struct vop_rmdir_args *a;
5763 	struct vnode *dvp, *vp;
5764 
5765 	a = ap;
5766 	dvp = a->a_dvp;
5767 	vp = a->a_vp;
5768 	vn_seqc_write_begin(dvp);
5769 	vn_seqc_write_begin(vp);
5770 }
5771 
5772 void
5773 vop_rmdir_post(void *ap, int rc)
5774 {
5775 	struct vop_rmdir_args *a;
5776 	struct vnode *dvp, *vp;
5777 
5778 	a = ap;
5779 	dvp = a->a_dvp;
5780 	vp = a->a_vp;
5781 	vn_seqc_write_end(dvp);
5782 	vn_seqc_write_end(vp);
5783 	if (!rc) {
5784 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5785 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5786 	}
5787 }
5788 
5789 void
5790 vop_setattr_pre(void *ap)
5791 {
5792 	struct vop_setattr_args *a;
5793 	struct vnode *vp;
5794 
5795 	a = ap;
5796 	vp = a->a_vp;
5797 	vn_seqc_write_begin(vp);
5798 }
5799 
5800 void
5801 vop_setattr_post(void *ap, int rc)
5802 {
5803 	struct vop_setattr_args *a;
5804 	struct vnode *vp;
5805 
5806 	a = ap;
5807 	vp = a->a_vp;
5808 	vn_seqc_write_end(vp);
5809 	if (!rc)
5810 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5811 }
5812 
5813 void
5814 vop_setacl_pre(void *ap)
5815 {
5816 	struct vop_setacl_args *a;
5817 	struct vnode *vp;
5818 
5819 	a = ap;
5820 	vp = a->a_vp;
5821 	vn_seqc_write_begin(vp);
5822 }
5823 
5824 void
5825 vop_setacl_post(void *ap, int rc __unused)
5826 {
5827 	struct vop_setacl_args *a;
5828 	struct vnode *vp;
5829 
5830 	a = ap;
5831 	vp = a->a_vp;
5832 	vn_seqc_write_end(vp);
5833 }
5834 
5835 void
5836 vop_setextattr_pre(void *ap)
5837 {
5838 	struct vop_setextattr_args *a;
5839 	struct vnode *vp;
5840 
5841 	a = ap;
5842 	vp = a->a_vp;
5843 	vn_seqc_write_begin(vp);
5844 }
5845 
5846 void
5847 vop_setextattr_post(void *ap, int rc)
5848 {
5849 	struct vop_setextattr_args *a;
5850 	struct vnode *vp;
5851 
5852 	a = ap;
5853 	vp = a->a_vp;
5854 	vn_seqc_write_end(vp);
5855 	if (!rc)
5856 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5857 }
5858 
5859 void
5860 vop_symlink_pre(void *ap)
5861 {
5862 	struct vop_symlink_args *a;
5863 	struct vnode *dvp;
5864 
5865 	a = ap;
5866 	dvp = a->a_dvp;
5867 	vn_seqc_write_begin(dvp);
5868 }
5869 
5870 void
5871 vop_symlink_post(void *ap, int rc)
5872 {
5873 	struct vop_symlink_args *a;
5874 	struct vnode *dvp;
5875 
5876 	a = ap;
5877 	dvp = a->a_dvp;
5878 	vn_seqc_write_end(dvp);
5879 	if (!rc)
5880 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5881 }
5882 
5883 void
5884 vop_open_post(void *ap, int rc)
5885 {
5886 	struct vop_open_args *a = ap;
5887 
5888 	if (!rc)
5889 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5890 }
5891 
5892 void
5893 vop_close_post(void *ap, int rc)
5894 {
5895 	struct vop_close_args *a = ap;
5896 
5897 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5898 	    !VN_IS_DOOMED(a->a_vp))) {
5899 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5900 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5901 	}
5902 }
5903 
5904 void
5905 vop_read_post(void *ap, int rc)
5906 {
5907 	struct vop_read_args *a = ap;
5908 
5909 	if (!rc)
5910 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5911 }
5912 
5913 void
5914 vop_read_pgcache_post(void *ap, int rc)
5915 {
5916 	struct vop_read_pgcache_args *a = ap;
5917 
5918 	if (!rc)
5919 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
5920 }
5921 
5922 void
5923 vop_readdir_post(void *ap, int rc)
5924 {
5925 	struct vop_readdir_args *a = ap;
5926 
5927 	if (!rc)
5928 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5929 }
5930 
5931 static struct knlist fs_knlist;
5932 
5933 static void
5934 vfs_event_init(void *arg)
5935 {
5936 	knlist_init_mtx(&fs_knlist, NULL);
5937 }
5938 /* XXX - correct order? */
5939 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5940 
5941 void
5942 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5943 {
5944 
5945 	KNOTE_UNLOCKED(&fs_knlist, event);
5946 }
5947 
5948 static int	filt_fsattach(struct knote *kn);
5949 static void	filt_fsdetach(struct knote *kn);
5950 static int	filt_fsevent(struct knote *kn, long hint);
5951 
5952 struct filterops fs_filtops = {
5953 	.f_isfd = 0,
5954 	.f_attach = filt_fsattach,
5955 	.f_detach = filt_fsdetach,
5956 	.f_event = filt_fsevent
5957 };
5958 
5959 static int
5960 filt_fsattach(struct knote *kn)
5961 {
5962 
5963 	kn->kn_flags |= EV_CLEAR;
5964 	knlist_add(&fs_knlist, kn, 0);
5965 	return (0);
5966 }
5967 
5968 static void
5969 filt_fsdetach(struct knote *kn)
5970 {
5971 
5972 	knlist_remove(&fs_knlist, kn, 0);
5973 }
5974 
5975 static int
5976 filt_fsevent(struct knote *kn, long hint)
5977 {
5978 
5979 	kn->kn_fflags |= hint;
5980 	return (kn->kn_fflags != 0);
5981 }
5982 
5983 static int
5984 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5985 {
5986 	struct vfsidctl vc;
5987 	int error;
5988 	struct mount *mp;
5989 
5990 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5991 	if (error)
5992 		return (error);
5993 	if (vc.vc_vers != VFS_CTL_VERS1)
5994 		return (EINVAL);
5995 	mp = vfs_getvfs(&vc.vc_fsid);
5996 	if (mp == NULL)
5997 		return (ENOENT);
5998 	/* ensure that a specific sysctl goes to the right filesystem. */
5999 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6000 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6001 		vfs_rel(mp);
6002 		return (EINVAL);
6003 	}
6004 	VCTLTOREQ(&vc, req);
6005 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6006 	vfs_rel(mp);
6007 	return (error);
6008 }
6009 
6010 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6011     NULL, 0, sysctl_vfs_ctl, "",
6012     "Sysctl by fsid");
6013 
6014 /*
6015  * Function to initialize a va_filerev field sensibly.
6016  * XXX: Wouldn't a random number make a lot more sense ??
6017  */
6018 u_quad_t
6019 init_va_filerev(void)
6020 {
6021 	struct bintime bt;
6022 
6023 	getbinuptime(&bt);
6024 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6025 }
6026 
6027 static int	filt_vfsread(struct knote *kn, long hint);
6028 static int	filt_vfswrite(struct knote *kn, long hint);
6029 static int	filt_vfsvnode(struct knote *kn, long hint);
6030 static void	filt_vfsdetach(struct knote *kn);
6031 static struct filterops vfsread_filtops = {
6032 	.f_isfd = 1,
6033 	.f_detach = filt_vfsdetach,
6034 	.f_event = filt_vfsread
6035 };
6036 static struct filterops vfswrite_filtops = {
6037 	.f_isfd = 1,
6038 	.f_detach = filt_vfsdetach,
6039 	.f_event = filt_vfswrite
6040 };
6041 static struct filterops vfsvnode_filtops = {
6042 	.f_isfd = 1,
6043 	.f_detach = filt_vfsdetach,
6044 	.f_event = filt_vfsvnode
6045 };
6046 
6047 static void
6048 vfs_knllock(void *arg)
6049 {
6050 	struct vnode *vp = arg;
6051 
6052 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6053 }
6054 
6055 static void
6056 vfs_knlunlock(void *arg)
6057 {
6058 	struct vnode *vp = arg;
6059 
6060 	VOP_UNLOCK(vp);
6061 }
6062 
6063 static void
6064 vfs_knl_assert_lock(void *arg, int what)
6065 {
6066 #ifdef DEBUG_VFS_LOCKS
6067 	struct vnode *vp = arg;
6068 
6069 	if (what == LA_LOCKED)
6070 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6071 	else
6072 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6073 #endif
6074 }
6075 
6076 int
6077 vfs_kqfilter(struct vop_kqfilter_args *ap)
6078 {
6079 	struct vnode *vp = ap->a_vp;
6080 	struct knote *kn = ap->a_kn;
6081 	struct knlist *knl;
6082 
6083 	switch (kn->kn_filter) {
6084 	case EVFILT_READ:
6085 		kn->kn_fop = &vfsread_filtops;
6086 		break;
6087 	case EVFILT_WRITE:
6088 		kn->kn_fop = &vfswrite_filtops;
6089 		break;
6090 	case EVFILT_VNODE:
6091 		kn->kn_fop = &vfsvnode_filtops;
6092 		break;
6093 	default:
6094 		return (EINVAL);
6095 	}
6096 
6097 	kn->kn_hook = (caddr_t)vp;
6098 
6099 	v_addpollinfo(vp);
6100 	if (vp->v_pollinfo == NULL)
6101 		return (ENOMEM);
6102 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6103 	vhold(vp);
6104 	knlist_add(knl, kn, 0);
6105 
6106 	return (0);
6107 }
6108 
6109 /*
6110  * Detach knote from vnode
6111  */
6112 static void
6113 filt_vfsdetach(struct knote *kn)
6114 {
6115 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6116 
6117 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6118 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6119 	vdrop(vp);
6120 }
6121 
6122 /*ARGSUSED*/
6123 static int
6124 filt_vfsread(struct knote *kn, long hint)
6125 {
6126 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6127 	struct vattr va;
6128 	int res;
6129 
6130 	/*
6131 	 * filesystem is gone, so set the EOF flag and schedule
6132 	 * the knote for deletion.
6133 	 */
6134 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6135 		VI_LOCK(vp);
6136 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6137 		VI_UNLOCK(vp);
6138 		return (1);
6139 	}
6140 
6141 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6142 		return (0);
6143 
6144 	VI_LOCK(vp);
6145 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6146 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6147 	VI_UNLOCK(vp);
6148 	return (res);
6149 }
6150 
6151 /*ARGSUSED*/
6152 static int
6153 filt_vfswrite(struct knote *kn, long hint)
6154 {
6155 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6156 
6157 	VI_LOCK(vp);
6158 
6159 	/*
6160 	 * filesystem is gone, so set the EOF flag and schedule
6161 	 * the knote for deletion.
6162 	 */
6163 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6164 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6165 
6166 	kn->kn_data = 0;
6167 	VI_UNLOCK(vp);
6168 	return (1);
6169 }
6170 
6171 static int
6172 filt_vfsvnode(struct knote *kn, long hint)
6173 {
6174 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6175 	int res;
6176 
6177 	VI_LOCK(vp);
6178 	if (kn->kn_sfflags & hint)
6179 		kn->kn_fflags |= hint;
6180 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6181 		kn->kn_flags |= EV_EOF;
6182 		VI_UNLOCK(vp);
6183 		return (1);
6184 	}
6185 	res = (kn->kn_fflags != 0);
6186 	VI_UNLOCK(vp);
6187 	return (res);
6188 }
6189 
6190 /*
6191  * Returns whether the directory is empty or not.
6192  * If it is empty, the return value is 0; otherwise
6193  * the return value is an error value (which may
6194  * be ENOTEMPTY).
6195  */
6196 int
6197 vfs_emptydir(struct vnode *vp)
6198 {
6199 	struct uio uio;
6200 	struct iovec iov;
6201 	struct dirent *dirent, *dp, *endp;
6202 	int error, eof;
6203 
6204 	error = 0;
6205 	eof = 0;
6206 
6207 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6208 
6209 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6210 	iov.iov_base = dirent;
6211 	iov.iov_len = sizeof(struct dirent);
6212 
6213 	uio.uio_iov = &iov;
6214 	uio.uio_iovcnt = 1;
6215 	uio.uio_offset = 0;
6216 	uio.uio_resid = sizeof(struct dirent);
6217 	uio.uio_segflg = UIO_SYSSPACE;
6218 	uio.uio_rw = UIO_READ;
6219 	uio.uio_td = curthread;
6220 
6221 	while (eof == 0 && error == 0) {
6222 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6223 		    NULL, NULL);
6224 		if (error != 0)
6225 			break;
6226 		endp = (void *)((uint8_t *)dirent +
6227 		    sizeof(struct dirent) - uio.uio_resid);
6228 		for (dp = dirent; dp < endp;
6229 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6230 			if (dp->d_type == DT_WHT)
6231 				continue;
6232 			if (dp->d_namlen == 0)
6233 				continue;
6234 			if (dp->d_type != DT_DIR &&
6235 			    dp->d_type != DT_UNKNOWN) {
6236 				error = ENOTEMPTY;
6237 				break;
6238 			}
6239 			if (dp->d_namlen > 2) {
6240 				error = ENOTEMPTY;
6241 				break;
6242 			}
6243 			if (dp->d_namlen == 1 &&
6244 			    dp->d_name[0] != '.') {
6245 				error = ENOTEMPTY;
6246 				break;
6247 			}
6248 			if (dp->d_namlen == 2 &&
6249 			    dp->d_name[1] != '.') {
6250 				error = ENOTEMPTY;
6251 				break;
6252 			}
6253 			uio.uio_resid = sizeof(struct dirent);
6254 		}
6255 	}
6256 	free(dirent, M_TEMP);
6257 	return (error);
6258 }
6259 
6260 int
6261 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6262 {
6263 	int error;
6264 
6265 	if (dp->d_reclen > ap->a_uio->uio_resid)
6266 		return (ENAMETOOLONG);
6267 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6268 	if (error) {
6269 		if (ap->a_ncookies != NULL) {
6270 			if (ap->a_cookies != NULL)
6271 				free(ap->a_cookies, M_TEMP);
6272 			ap->a_cookies = NULL;
6273 			*ap->a_ncookies = 0;
6274 		}
6275 		return (error);
6276 	}
6277 	if (ap->a_ncookies == NULL)
6278 		return (0);
6279 
6280 	KASSERT(ap->a_cookies,
6281 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6282 
6283 	*ap->a_cookies = realloc(*ap->a_cookies,
6284 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6285 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6286 	*ap->a_ncookies += 1;
6287 	return (0);
6288 }
6289 
6290 /*
6291  * The purpose of this routine is to remove granularity from accmode_t,
6292  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6293  * VADMIN and VAPPEND.
6294  *
6295  * If it returns 0, the caller is supposed to continue with the usual
6296  * access checks using 'accmode' as modified by this routine.  If it
6297  * returns nonzero value, the caller is supposed to return that value
6298  * as errno.
6299  *
6300  * Note that after this routine runs, accmode may be zero.
6301  */
6302 int
6303 vfs_unixify_accmode(accmode_t *accmode)
6304 {
6305 	/*
6306 	 * There is no way to specify explicit "deny" rule using
6307 	 * file mode or POSIX.1e ACLs.
6308 	 */
6309 	if (*accmode & VEXPLICIT_DENY) {
6310 		*accmode = 0;
6311 		return (0);
6312 	}
6313 
6314 	/*
6315 	 * None of these can be translated into usual access bits.
6316 	 * Also, the common case for NFSv4 ACLs is to not contain
6317 	 * either of these bits. Caller should check for VWRITE
6318 	 * on the containing directory instead.
6319 	 */
6320 	if (*accmode & (VDELETE_CHILD | VDELETE))
6321 		return (EPERM);
6322 
6323 	if (*accmode & VADMIN_PERMS) {
6324 		*accmode &= ~VADMIN_PERMS;
6325 		*accmode |= VADMIN;
6326 	}
6327 
6328 	/*
6329 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6330 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6331 	 */
6332 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6333 
6334 	return (0);
6335 }
6336 
6337 /*
6338  * Clear out a doomed vnode (if any) and replace it with a new one as long
6339  * as the fs is not being unmounted. Return the root vnode to the caller.
6340  */
6341 static int __noinline
6342 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6343 {
6344 	struct vnode *vp;
6345 	int error;
6346 
6347 restart:
6348 	if (mp->mnt_rootvnode != NULL) {
6349 		MNT_ILOCK(mp);
6350 		vp = mp->mnt_rootvnode;
6351 		if (vp != NULL) {
6352 			if (!VN_IS_DOOMED(vp)) {
6353 				vrefact(vp);
6354 				MNT_IUNLOCK(mp);
6355 				error = vn_lock(vp, flags);
6356 				if (error == 0) {
6357 					*vpp = vp;
6358 					return (0);
6359 				}
6360 				vrele(vp);
6361 				goto restart;
6362 			}
6363 			/*
6364 			 * Clear the old one.
6365 			 */
6366 			mp->mnt_rootvnode = NULL;
6367 		}
6368 		MNT_IUNLOCK(mp);
6369 		if (vp != NULL) {
6370 			vfs_op_barrier_wait(mp);
6371 			vrele(vp);
6372 		}
6373 	}
6374 	error = VFS_CACHEDROOT(mp, flags, vpp);
6375 	if (error != 0)
6376 		return (error);
6377 	if (mp->mnt_vfs_ops == 0) {
6378 		MNT_ILOCK(mp);
6379 		if (mp->mnt_vfs_ops != 0) {
6380 			MNT_IUNLOCK(mp);
6381 			return (0);
6382 		}
6383 		if (mp->mnt_rootvnode == NULL) {
6384 			vrefact(*vpp);
6385 			mp->mnt_rootvnode = *vpp;
6386 		} else {
6387 			if (mp->mnt_rootvnode != *vpp) {
6388 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6389 					panic("%s: mismatch between vnode returned "
6390 					    " by VFS_CACHEDROOT and the one cached "
6391 					    " (%p != %p)",
6392 					    __func__, *vpp, mp->mnt_rootvnode);
6393 				}
6394 			}
6395 		}
6396 		MNT_IUNLOCK(mp);
6397 	}
6398 	return (0);
6399 }
6400 
6401 int
6402 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6403 {
6404 	struct mount_pcpu *mpcpu;
6405 	struct vnode *vp;
6406 	int error;
6407 
6408 	if (!vfs_op_thread_enter(mp, mpcpu))
6409 		return (vfs_cache_root_fallback(mp, flags, vpp));
6410 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6411 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6412 		vfs_op_thread_exit(mp, mpcpu);
6413 		return (vfs_cache_root_fallback(mp, flags, vpp));
6414 	}
6415 	vrefact(vp);
6416 	vfs_op_thread_exit(mp, mpcpu);
6417 	error = vn_lock(vp, flags);
6418 	if (error != 0) {
6419 		vrele(vp);
6420 		return (vfs_cache_root_fallback(mp, flags, vpp));
6421 	}
6422 	*vpp = vp;
6423 	return (0);
6424 }
6425 
6426 struct vnode *
6427 vfs_cache_root_clear(struct mount *mp)
6428 {
6429 	struct vnode *vp;
6430 
6431 	/*
6432 	 * ops > 0 guarantees there is nobody who can see this vnode
6433 	 */
6434 	MPASS(mp->mnt_vfs_ops > 0);
6435 	vp = mp->mnt_rootvnode;
6436 	if (vp != NULL)
6437 		vn_seqc_write_begin(vp);
6438 	mp->mnt_rootvnode = NULL;
6439 	return (vp);
6440 }
6441 
6442 void
6443 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6444 {
6445 
6446 	MPASS(mp->mnt_vfs_ops > 0);
6447 	vrefact(vp);
6448 	mp->mnt_rootvnode = vp;
6449 }
6450 
6451 /*
6452  * These are helper functions for filesystems to traverse all
6453  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6454  *
6455  * This interface replaces MNT_VNODE_FOREACH.
6456  */
6457 
6458 struct vnode *
6459 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6460 {
6461 	struct vnode *vp;
6462 
6463 	if (should_yield())
6464 		kern_yield(PRI_USER);
6465 	MNT_ILOCK(mp);
6466 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6467 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6468 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6469 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6470 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6471 			continue;
6472 		VI_LOCK(vp);
6473 		if (VN_IS_DOOMED(vp)) {
6474 			VI_UNLOCK(vp);
6475 			continue;
6476 		}
6477 		break;
6478 	}
6479 	if (vp == NULL) {
6480 		__mnt_vnode_markerfree_all(mvp, mp);
6481 		/* MNT_IUNLOCK(mp); -- done in above function */
6482 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6483 		return (NULL);
6484 	}
6485 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6486 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6487 	MNT_IUNLOCK(mp);
6488 	return (vp);
6489 }
6490 
6491 struct vnode *
6492 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6493 {
6494 	struct vnode *vp;
6495 
6496 	*mvp = vn_alloc_marker(mp);
6497 	MNT_ILOCK(mp);
6498 	MNT_REF(mp);
6499 
6500 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6501 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6502 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6503 			continue;
6504 		VI_LOCK(vp);
6505 		if (VN_IS_DOOMED(vp)) {
6506 			VI_UNLOCK(vp);
6507 			continue;
6508 		}
6509 		break;
6510 	}
6511 	if (vp == NULL) {
6512 		MNT_REL(mp);
6513 		MNT_IUNLOCK(mp);
6514 		vn_free_marker(*mvp);
6515 		*mvp = NULL;
6516 		return (NULL);
6517 	}
6518 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6519 	MNT_IUNLOCK(mp);
6520 	return (vp);
6521 }
6522 
6523 void
6524 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6525 {
6526 
6527 	if (*mvp == NULL) {
6528 		MNT_IUNLOCK(mp);
6529 		return;
6530 	}
6531 
6532 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6533 
6534 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6535 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6536 	MNT_REL(mp);
6537 	MNT_IUNLOCK(mp);
6538 	vn_free_marker(*mvp);
6539 	*mvp = NULL;
6540 }
6541 
6542 /*
6543  * These are helper functions for filesystems to traverse their
6544  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6545  */
6546 static void
6547 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6548 {
6549 
6550 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6551 
6552 	MNT_ILOCK(mp);
6553 	MNT_REL(mp);
6554 	MNT_IUNLOCK(mp);
6555 	vn_free_marker(*mvp);
6556 	*mvp = NULL;
6557 }
6558 
6559 /*
6560  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6561  * conventional lock order during mnt_vnode_next_lazy iteration.
6562  *
6563  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6564  * The list lock is dropped and reacquired.  On success, both locks are held.
6565  * On failure, the mount vnode list lock is held but the vnode interlock is
6566  * not, and the procedure may have yielded.
6567  */
6568 static bool
6569 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6570     struct vnode *vp)
6571 {
6572 
6573 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6574 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6575 	    ("%s: bad marker", __func__));
6576 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6577 	    ("%s: inappropriate vnode", __func__));
6578 	ASSERT_VI_UNLOCKED(vp, __func__);
6579 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6580 
6581 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6582 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6583 
6584 	/*
6585 	 * Note we may be racing against vdrop which transitioned the hold
6586 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6587 	 * if we are the only user after we get the interlock we will just
6588 	 * vdrop.
6589 	 */
6590 	vhold(vp);
6591 	mtx_unlock(&mp->mnt_listmtx);
6592 	VI_LOCK(vp);
6593 	if (VN_IS_DOOMED(vp)) {
6594 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6595 		goto out_lost;
6596 	}
6597 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6598 	/*
6599 	 * There is nothing to do if we are the last user.
6600 	 */
6601 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6602 		goto out_lost;
6603 	mtx_lock(&mp->mnt_listmtx);
6604 	return (true);
6605 out_lost:
6606 	vdropl(vp);
6607 	maybe_yield();
6608 	mtx_lock(&mp->mnt_listmtx);
6609 	return (false);
6610 }
6611 
6612 static struct vnode *
6613 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6614     void *cbarg)
6615 {
6616 	struct vnode *vp;
6617 
6618 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6619 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6620 restart:
6621 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6622 	while (vp != NULL) {
6623 		if (vp->v_type == VMARKER) {
6624 			vp = TAILQ_NEXT(vp, v_lazylist);
6625 			continue;
6626 		}
6627 		/*
6628 		 * See if we want to process the vnode. Note we may encounter a
6629 		 * long string of vnodes we don't care about and hog the list
6630 		 * as a result. Check for it and requeue the marker.
6631 		 */
6632 		VNPASS(!VN_IS_DOOMED(vp), vp);
6633 		if (!cb(vp, cbarg)) {
6634 			if (!should_yield()) {
6635 				vp = TAILQ_NEXT(vp, v_lazylist);
6636 				continue;
6637 			}
6638 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6639 			    v_lazylist);
6640 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6641 			    v_lazylist);
6642 			mtx_unlock(&mp->mnt_listmtx);
6643 			kern_yield(PRI_USER);
6644 			mtx_lock(&mp->mnt_listmtx);
6645 			goto restart;
6646 		}
6647 		/*
6648 		 * Try-lock because this is the wrong lock order.
6649 		 */
6650 		if (!VI_TRYLOCK(vp) &&
6651 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6652 			goto restart;
6653 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6654 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6655 		    ("alien vnode on the lazy list %p %p", vp, mp));
6656 		VNPASS(vp->v_mount == mp, vp);
6657 		VNPASS(!VN_IS_DOOMED(vp), vp);
6658 		break;
6659 	}
6660 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6661 
6662 	/* Check if we are done */
6663 	if (vp == NULL) {
6664 		mtx_unlock(&mp->mnt_listmtx);
6665 		mnt_vnode_markerfree_lazy(mvp, mp);
6666 		return (NULL);
6667 	}
6668 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6669 	mtx_unlock(&mp->mnt_listmtx);
6670 	ASSERT_VI_LOCKED(vp, "lazy iter");
6671 	return (vp);
6672 }
6673 
6674 struct vnode *
6675 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6676     void *cbarg)
6677 {
6678 
6679 	if (should_yield())
6680 		kern_yield(PRI_USER);
6681 	mtx_lock(&mp->mnt_listmtx);
6682 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6683 }
6684 
6685 struct vnode *
6686 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6687     void *cbarg)
6688 {
6689 	struct vnode *vp;
6690 
6691 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6692 		return (NULL);
6693 
6694 	*mvp = vn_alloc_marker(mp);
6695 	MNT_ILOCK(mp);
6696 	MNT_REF(mp);
6697 	MNT_IUNLOCK(mp);
6698 
6699 	mtx_lock(&mp->mnt_listmtx);
6700 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6701 	if (vp == NULL) {
6702 		mtx_unlock(&mp->mnt_listmtx);
6703 		mnt_vnode_markerfree_lazy(mvp, mp);
6704 		return (NULL);
6705 	}
6706 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6707 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6708 }
6709 
6710 void
6711 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6712 {
6713 
6714 	if (*mvp == NULL)
6715 		return;
6716 
6717 	mtx_lock(&mp->mnt_listmtx);
6718 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6719 	mtx_unlock(&mp->mnt_listmtx);
6720 	mnt_vnode_markerfree_lazy(mvp, mp);
6721 }
6722 
6723 int
6724 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6725 {
6726 
6727 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6728 		cnp->cn_flags &= ~NOEXECCHECK;
6729 		return (0);
6730 	}
6731 
6732 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6733 }
6734 
6735 /*
6736  * Do not use this variant unless you have means other than the hold count
6737  * to prevent the vnode from getting freed.
6738  */
6739 void
6740 vn_seqc_write_begin_unheld_locked(struct vnode *vp)
6741 {
6742 
6743 	ASSERT_VI_LOCKED(vp, __func__);
6744 	VNPASS(vp->v_seqc_users >= 0, vp);
6745 	vp->v_seqc_users++;
6746 	if (vp->v_seqc_users == 1)
6747 		seqc_sleepable_write_begin(&vp->v_seqc);
6748 }
6749 
6750 void
6751 vn_seqc_write_begin_locked(struct vnode *vp)
6752 {
6753 
6754 	ASSERT_VI_LOCKED(vp, __func__);
6755 	VNPASS(vp->v_holdcnt > 0, vp);
6756 	vn_seqc_write_begin_unheld_locked(vp);
6757 }
6758 
6759 void
6760 vn_seqc_write_begin(struct vnode *vp)
6761 {
6762 
6763 	VI_LOCK(vp);
6764 	vn_seqc_write_begin_locked(vp);
6765 	VI_UNLOCK(vp);
6766 }
6767 
6768 void
6769 vn_seqc_write_begin_unheld(struct vnode *vp)
6770 {
6771 
6772 	VI_LOCK(vp);
6773 	vn_seqc_write_begin_unheld_locked(vp);
6774 	VI_UNLOCK(vp);
6775 }
6776 
6777 void
6778 vn_seqc_write_end_locked(struct vnode *vp)
6779 {
6780 
6781 	ASSERT_VI_LOCKED(vp, __func__);
6782 	VNPASS(vp->v_seqc_users > 0, vp);
6783 	vp->v_seqc_users--;
6784 	if (vp->v_seqc_users == 0)
6785 		seqc_sleepable_write_end(&vp->v_seqc);
6786 }
6787 
6788 void
6789 vn_seqc_write_end(struct vnode *vp)
6790 {
6791 
6792 	VI_LOCK(vp);
6793 	vn_seqc_write_end_locked(vp);
6794 	VI_UNLOCK(vp);
6795 }
6796