xref: /freebsd/sys/kern/vfs_subr.c (revision fc7510aef78781b0068da1a6ba190a636a54d6e7)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/stat.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/uma.h>
99 
100 #ifdef DDB
101 #include <ddb/ddb.h>
102 #endif
103 
104 static void	delmntque(struct vnode *vp);
105 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 		    int slpflag, int slptimeo);
107 static void	syncer_shutdown(void *arg, int howto);
108 static int	vtryrecycle(struct vnode *vp);
109 static void	v_init_counters(struct vnode *);
110 static void	v_incr_devcount(struct vnode *);
111 static void	v_decr_devcount(struct vnode *);
112 static void	vgonel(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_locked(void *arg);
116 static void	vfs_knl_assert_unlocked(void *arg);
117 static void	destroy_vpollinfo(struct vpollinfo *vi);
118 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
119 		    daddr_t startlbn, daddr_t endlbn);
120 static void	vnlru_recalc(void);
121 
122 /*
123  * These fences are intended for cases where some synchronization is
124  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
125  * and v_usecount) updates.  Access to v_iflags is generally synchronized
126  * by the interlock, but we have some internal assertions that check vnode
127  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
128  * for now.
129  */
130 #ifdef INVARIANTS
131 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
132 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
133 #else
134 #define	VNODE_REFCOUNT_FENCE_ACQ()
135 #define	VNODE_REFCOUNT_FENCE_REL()
136 #endif
137 
138 /*
139  * Number of vnodes in existence.  Increased whenever getnewvnode()
140  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141  */
142 static u_long __exclusive_cache_line numvnodes;
143 
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145     "Number of vnodes in existence");
146 
147 static counter_u64_t vnodes_created;
148 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
149     "Number of vnodes created by getnewvnode");
150 
151 /*
152  * Conversion tables for conversion from vnode types to inode formats
153  * and back.
154  */
155 enum vtype iftovt_tab[16] = {
156 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
157 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
158 };
159 int vttoif_tab[10] = {
160 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
161 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
162 };
163 
164 /*
165  * List of allocates vnodes in the system.
166  */
167 static TAILQ_HEAD(freelst, vnode) vnode_list;
168 static struct vnode *vnode_list_free_marker;
169 static struct vnode *vnode_list_reclaim_marker;
170 
171 /*
172  * "Free" vnode target.  Free vnodes are rarely completely free, but are
173  * just ones that are cheap to recycle.  Usually they are for files which
174  * have been stat'd but not read; these usually have inode and namecache
175  * data attached to them.  This target is the preferred minimum size of a
176  * sub-cache consisting mostly of such files. The system balances the size
177  * of this sub-cache with its complement to try to prevent either from
178  * thrashing while the other is relatively inactive.  The targets express
179  * a preference for the best balance.
180  *
181  * "Above" this target there are 2 further targets (watermarks) related
182  * to recyling of free vnodes.  In the best-operating case, the cache is
183  * exactly full, the free list has size between vlowat and vhiwat above the
184  * free target, and recycling from it and normal use maintains this state.
185  * Sometimes the free list is below vlowat or even empty, but this state
186  * is even better for immediate use provided the cache is not full.
187  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
188  * ones) to reach one of these states.  The watermarks are currently hard-
189  * coded as 4% and 9% of the available space higher.  These and the default
190  * of 25% for wantfreevnodes are too large if the memory size is large.
191  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
192  * whenever vnlru_proc() becomes active.
193  */
194 static long wantfreevnodes;
195 static long __exclusive_cache_line freevnodes;
196 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
197     &freevnodes, 0, "Number of \"free\" vnodes");
198 static long freevnodes_old;
199 
200 static counter_u64_t recycles_count;
201 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
202     "Number of vnodes recycled to meet vnode cache targets");
203 
204 static counter_u64_t recycles_free_count;
205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
206     "Number of free vnodes recycled to meet vnode cache targets");
207 
208 /*
209  * Various variables used for debugging the new implementation of
210  * reassignbuf().
211  * XXX these are probably of (very) limited utility now.
212  */
213 static int reassignbufcalls;
214 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
215     &reassignbufcalls, 0, "Number of calls to reassignbuf");
216 
217 static counter_u64_t deferred_inact;
218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
219     "Number of times inactive processing was deferred");
220 
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
223 
224 /*
225  * Lock for any access to the following:
226  *	vnode_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx __exclusive_cache_line vnode_list_mtx;
231 
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
234 
235 static uma_zone_t buf_trie_zone;
236 
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
240 
241 /*
242  * The workitem queue.
243  *
244  * It is useful to delay writes of file data and filesystem metadata
245  * for tens of seconds so that quickly created and deleted files need
246  * not waste disk bandwidth being created and removed. To realize this,
247  * we append vnodes to a "workitem" queue. When running with a soft
248  * updates implementation, most pending metadata dependencies should
249  * not wait for more than a few seconds. Thus, mounted on block devices
250  * are delayed only about a half the time that file data is delayed.
251  * Similarly, directory updates are more critical, so are only delayed
252  * about a third the time that file data is delayed. Thus, there are
253  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254  * one each second (driven off the filesystem syncer process). The
255  * syncer_delayno variable indicates the next queue that is to be processed.
256  * Items that need to be processed soon are placed in this queue:
257  *
258  *	syncer_workitem_pending[syncer_delayno]
259  *
260  * A delay of fifteen seconds is done by placing the request fifteen
261  * entries later in the queue:
262  *
263  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264  *
265  */
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
270 /*
271  * The sync_mtx protects:
272  *	bo->bo_synclist
273  *	sync_vnode_count
274  *	syncer_delayno
275  *	syncer_state
276  *	syncer_workitem_pending
277  *	syncer_worklist_len
278  *	rushjob
279  */
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
282 
283 #define SYNCER_MAXDELAY		32
284 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285 static int syncdelay = 30;		/* max time to delay syncing data */
286 static int filedelay = 30;		/* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288     "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29;		/* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291     "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28;		/* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294     "Time to delay syncing metadata (in seconds)");
295 static int rushjob;		/* number of slots to run ASAP */
296 static int stat_rush_requests;	/* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298     "Number of times I/O speeded up (rush requests)");
299 
300 #define	VDBATCH_SIZE 8
301 struct vdbatch {
302 	u_int index;
303 	long freevnodes;
304 	struct mtx lock;
305 	struct vnode *tab[VDBATCH_SIZE];
306 };
307 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
308 
309 static void	vdbatch_dequeue(struct vnode *vp);
310 
311 /*
312  * When shutting down the syncer, run it at four times normal speed.
313  */
314 #define SYNCER_SHUTDOWN_SPEEDUP		4
315 static int sync_vnode_count;
316 static int syncer_worklist_len;
317 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
318     syncer_state;
319 
320 /* Target for maximum number of vnodes. */
321 u_long desiredvnodes;
322 static u_long gapvnodes;		/* gap between wanted and desired */
323 static u_long vhiwat;		/* enough extras after expansion */
324 static u_long vlowat;		/* minimal extras before expansion */
325 static u_long vstir;		/* nonzero to stir non-free vnodes */
326 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
327 
328 static u_long vnlru_read_freevnodes(void);
329 
330 /*
331  * Note that no attempt is made to sanitize these parameters.
332  */
333 static int
334 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
335 {
336 	u_long val;
337 	int error;
338 
339 	val = desiredvnodes;
340 	error = sysctl_handle_long(oidp, &val, 0, req);
341 	if (error != 0 || req->newptr == NULL)
342 		return (error);
343 
344 	if (val == desiredvnodes)
345 		return (0);
346 	mtx_lock(&vnode_list_mtx);
347 	desiredvnodes = val;
348 	wantfreevnodes = desiredvnodes / 4;
349 	vnlru_recalc();
350 	mtx_unlock(&vnode_list_mtx);
351 	/*
352 	 * XXX There is no protection against multiple threads changing
353 	 * desiredvnodes at the same time. Locking above only helps vnlru and
354 	 * getnewvnode.
355 	 */
356 	vfs_hash_changesize(desiredvnodes);
357 	cache_changesize(desiredvnodes);
358 	return (0);
359 }
360 
361 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
362     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
363     "LU", "Target for maximum number of vnodes");
364 
365 static int
366 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
367 {
368 	u_long val;
369 	int error;
370 
371 	val = wantfreevnodes;
372 	error = sysctl_handle_long(oidp, &val, 0, req);
373 	if (error != 0 || req->newptr == NULL)
374 		return (error);
375 
376 	if (val == wantfreevnodes)
377 		return (0);
378 	mtx_lock(&vnode_list_mtx);
379 	wantfreevnodes = val;
380 	vnlru_recalc();
381 	mtx_unlock(&vnode_list_mtx);
382 	return (0);
383 }
384 
385 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
386     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
387     "LU", "Target for minimum number of \"free\" vnodes");
388 
389 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
390     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
391 static int vnlru_nowhere;
392 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
393     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
394 
395 static int
396 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
397 {
398 	struct vnode *vp;
399 	struct nameidata nd;
400 	char *buf;
401 	unsigned long ndflags;
402 	int error;
403 
404 	if (req->newptr == NULL)
405 		return (EINVAL);
406 	if (req->newlen >= PATH_MAX)
407 		return (E2BIG);
408 
409 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
410 	error = SYSCTL_IN(req, buf, req->newlen);
411 	if (error != 0)
412 		goto out;
413 
414 	buf[req->newlen] = '\0';
415 
416 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
417 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
418 	if ((error = namei(&nd)) != 0)
419 		goto out;
420 	vp = nd.ni_vp;
421 
422 	if (VN_IS_DOOMED(vp)) {
423 		/*
424 		 * This vnode is being recycled.  Return != 0 to let the caller
425 		 * know that the sysctl had no effect.  Return EAGAIN because a
426 		 * subsequent call will likely succeed (since namei will create
427 		 * a new vnode if necessary)
428 		 */
429 		error = EAGAIN;
430 		goto putvnode;
431 	}
432 
433 	counter_u64_add(recycles_count, 1);
434 	vgone(vp);
435 putvnode:
436 	NDFREE(&nd, 0);
437 out:
438 	free(buf, M_TEMP);
439 	return (error);
440 }
441 
442 static int
443 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
444 {
445 	struct thread *td = curthread;
446 	struct vnode *vp;
447 	struct file *fp;
448 	int error;
449 	int fd;
450 
451 	if (req->newptr == NULL)
452 		return (EBADF);
453 
454         error = sysctl_handle_int(oidp, &fd, 0, req);
455         if (error != 0)
456                 return (error);
457 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
458 	if (error != 0)
459 		return (error);
460 	vp = fp->f_vnode;
461 
462 	error = vn_lock(vp, LK_EXCLUSIVE);
463 	if (error != 0)
464 		goto drop;
465 
466 	counter_u64_add(recycles_count, 1);
467 	vgone(vp);
468 	VOP_UNLOCK(vp);
469 drop:
470 	fdrop(fp, td);
471 	return (error);
472 }
473 
474 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
475     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
476     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
477 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
478     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
479     sysctl_ftry_reclaim_vnode, "I",
480     "Try to reclaim a vnode by its file descriptor");
481 
482 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
483 static int vnsz2log;
484 
485 /*
486  * Support for the bufobj clean & dirty pctrie.
487  */
488 static void *
489 buf_trie_alloc(struct pctrie *ptree)
490 {
491 
492 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
493 }
494 
495 static void
496 buf_trie_free(struct pctrie *ptree, void *node)
497 {
498 
499 	uma_zfree(buf_trie_zone, node);
500 }
501 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
502 
503 /*
504  * Initialize the vnode management data structures.
505  *
506  * Reevaluate the following cap on the number of vnodes after the physical
507  * memory size exceeds 512GB.  In the limit, as the physical memory size
508  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
509  */
510 #ifndef	MAXVNODES_MAX
511 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
512 #endif
513 
514 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
515 
516 static struct vnode *
517 vn_alloc_marker(struct mount *mp)
518 {
519 	struct vnode *vp;
520 
521 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
522 	vp->v_type = VMARKER;
523 	vp->v_mount = mp;
524 
525 	return (vp);
526 }
527 
528 static void
529 vn_free_marker(struct vnode *vp)
530 {
531 
532 	MPASS(vp->v_type == VMARKER);
533 	free(vp, M_VNODE_MARKER);
534 }
535 
536 /*
537  * Initialize a vnode as it first enters the zone.
538  */
539 static int
540 vnode_init(void *mem, int size, int flags)
541 {
542 	struct vnode *vp;
543 
544 	vp = mem;
545 	bzero(vp, size);
546 	/*
547 	 * Setup locks.
548 	 */
549 	vp->v_vnlock = &vp->v_lock;
550 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
551 	/*
552 	 * By default, don't allow shared locks unless filesystems opt-in.
553 	 */
554 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
555 	    LK_NOSHARE | LK_IS_VNODE);
556 	/*
557 	 * Initialize bufobj.
558 	 */
559 	bufobj_init(&vp->v_bufobj, vp);
560 	/*
561 	 * Initialize namecache.
562 	 */
563 	LIST_INIT(&vp->v_cache_src);
564 	TAILQ_INIT(&vp->v_cache_dst);
565 	/*
566 	 * Initialize rangelocks.
567 	 */
568 	rangelock_init(&vp->v_rl);
569 
570 	vp->v_dbatchcpu = NOCPU;
571 
572 	mtx_lock(&vnode_list_mtx);
573 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
574 	mtx_unlock(&vnode_list_mtx);
575 	return (0);
576 }
577 
578 /*
579  * Free a vnode when it is cleared from the zone.
580  */
581 static void
582 vnode_fini(void *mem, int size)
583 {
584 	struct vnode *vp;
585 	struct bufobj *bo;
586 
587 	vp = mem;
588 	vdbatch_dequeue(vp);
589 	mtx_lock(&vnode_list_mtx);
590 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
591 	mtx_unlock(&vnode_list_mtx);
592 	rangelock_destroy(&vp->v_rl);
593 	lockdestroy(vp->v_vnlock);
594 	mtx_destroy(&vp->v_interlock);
595 	bo = &vp->v_bufobj;
596 	rw_destroy(BO_LOCKPTR(bo));
597 }
598 
599 /*
600  * Provide the size of NFS nclnode and NFS fh for calculation of the
601  * vnode memory consumption.  The size is specified directly to
602  * eliminate dependency on NFS-private header.
603  *
604  * Other filesystems may use bigger or smaller (like UFS and ZFS)
605  * private inode data, but the NFS-based estimation is ample enough.
606  * Still, we care about differences in the size between 64- and 32-bit
607  * platforms.
608  *
609  * Namecache structure size is heuristically
610  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
611  */
612 #ifdef _LP64
613 #define	NFS_NCLNODE_SZ	(528 + 64)
614 #define	NC_SZ		148
615 #else
616 #define	NFS_NCLNODE_SZ	(360 + 32)
617 #define	NC_SZ		92
618 #endif
619 
620 static void
621 vntblinit(void *dummy __unused)
622 {
623 	struct vdbatch *vd;
624 	int cpu, physvnodes, virtvnodes;
625 	u_int i;
626 
627 	/*
628 	 * Desiredvnodes is a function of the physical memory size and the
629 	 * kernel's heap size.  Generally speaking, it scales with the
630 	 * physical memory size.  The ratio of desiredvnodes to the physical
631 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
632 	 * Thereafter, the
633 	 * marginal ratio of desiredvnodes to the physical memory size is
634 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
635 	 * size.  The memory required by desiredvnodes vnodes and vm objects
636 	 * must not exceed 1/10th of the kernel's heap size.
637 	 */
638 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
639 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
640 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
641 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
642 	desiredvnodes = min(physvnodes, virtvnodes);
643 	if (desiredvnodes > MAXVNODES_MAX) {
644 		if (bootverbose)
645 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
646 			    desiredvnodes, MAXVNODES_MAX);
647 		desiredvnodes = MAXVNODES_MAX;
648 	}
649 	wantfreevnodes = desiredvnodes / 4;
650 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
651 	TAILQ_INIT(&vnode_list);
652 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
653 	/*
654 	 * The lock is taken to appease WITNESS.
655 	 */
656 	mtx_lock(&vnode_list_mtx);
657 	vnlru_recalc();
658 	mtx_unlock(&vnode_list_mtx);
659 	vnode_list_free_marker = vn_alloc_marker(NULL);
660 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
661 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
662 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
663 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
664 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
665 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
666 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
667 	/*
668 	 * Preallocate enough nodes to support one-per buf so that
669 	 * we can not fail an insert.  reassignbuf() callers can not
670 	 * tolerate the insertion failure.
671 	 */
672 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
673 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
674 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
675 	uma_prealloc(buf_trie_zone, nbuf);
676 
677 	vnodes_created = counter_u64_alloc(M_WAITOK);
678 	recycles_count = counter_u64_alloc(M_WAITOK);
679 	recycles_free_count = counter_u64_alloc(M_WAITOK);
680 	deferred_inact = counter_u64_alloc(M_WAITOK);
681 
682 	/*
683 	 * Initialize the filesystem syncer.
684 	 */
685 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
686 	    &syncer_mask);
687 	syncer_maxdelay = syncer_mask + 1;
688 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
689 	cv_init(&sync_wakeup, "syncer");
690 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
691 		vnsz2log++;
692 	vnsz2log--;
693 
694 	CPU_FOREACH(cpu) {
695 		vd = DPCPU_ID_PTR((cpu), vd);
696 		bzero(vd, sizeof(*vd));
697 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
698 	}
699 }
700 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
701 
702 /*
703  * Mark a mount point as busy. Used to synchronize access and to delay
704  * unmounting. Eventually, mountlist_mtx is not released on failure.
705  *
706  * vfs_busy() is a custom lock, it can block the caller.
707  * vfs_busy() only sleeps if the unmount is active on the mount point.
708  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
709  * vnode belonging to mp.
710  *
711  * Lookup uses vfs_busy() to traverse mount points.
712  * root fs			var fs
713  * / vnode lock		A	/ vnode lock (/var)		D
714  * /var vnode lock	B	/log vnode lock(/var/log)	E
715  * vfs_busy lock	C	vfs_busy lock			F
716  *
717  * Within each file system, the lock order is C->A->B and F->D->E.
718  *
719  * When traversing across mounts, the system follows that lock order:
720  *
721  *        C->A->B
722  *              |
723  *              +->F->D->E
724  *
725  * The lookup() process for namei("/var") illustrates the process:
726  *  VOP_LOOKUP() obtains B while A is held
727  *  vfs_busy() obtains a shared lock on F while A and B are held
728  *  vput() releases lock on B
729  *  vput() releases lock on A
730  *  VFS_ROOT() obtains lock on D while shared lock on F is held
731  *  vfs_unbusy() releases shared lock on F
732  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
733  *    Attempt to lock A (instead of vp_crossmp) while D is held would
734  *    violate the global order, causing deadlocks.
735  *
736  * dounmount() locks B while F is drained.
737  */
738 int
739 vfs_busy(struct mount *mp, int flags)
740 {
741 
742 	MPASS((flags & ~MBF_MASK) == 0);
743 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
744 
745 	if (vfs_op_thread_enter(mp)) {
746 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
747 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
748 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
749 		vfs_mp_count_add_pcpu(mp, ref, 1);
750 		vfs_mp_count_add_pcpu(mp, lockref, 1);
751 		vfs_op_thread_exit(mp);
752 		if (flags & MBF_MNTLSTLOCK)
753 			mtx_unlock(&mountlist_mtx);
754 		return (0);
755 	}
756 
757 	MNT_ILOCK(mp);
758 	vfs_assert_mount_counters(mp);
759 	MNT_REF(mp);
760 	/*
761 	 * If mount point is currently being unmounted, sleep until the
762 	 * mount point fate is decided.  If thread doing the unmounting fails,
763 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
764 	 * that this mount point has survived the unmount attempt and vfs_busy
765 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
766 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
767 	 * about to be really destroyed.  vfs_busy needs to release its
768 	 * reference on the mount point in this case and return with ENOENT,
769 	 * telling the caller that mount mount it tried to busy is no longer
770 	 * valid.
771 	 */
772 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
773 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
774 			MNT_REL(mp);
775 			MNT_IUNLOCK(mp);
776 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
777 			    __func__);
778 			return (ENOENT);
779 		}
780 		if (flags & MBF_MNTLSTLOCK)
781 			mtx_unlock(&mountlist_mtx);
782 		mp->mnt_kern_flag |= MNTK_MWAIT;
783 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
784 		if (flags & MBF_MNTLSTLOCK)
785 			mtx_lock(&mountlist_mtx);
786 		MNT_ILOCK(mp);
787 	}
788 	if (flags & MBF_MNTLSTLOCK)
789 		mtx_unlock(&mountlist_mtx);
790 	mp->mnt_lockref++;
791 	MNT_IUNLOCK(mp);
792 	return (0);
793 }
794 
795 /*
796  * Free a busy filesystem.
797  */
798 void
799 vfs_unbusy(struct mount *mp)
800 {
801 	int c;
802 
803 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
804 
805 	if (vfs_op_thread_enter(mp)) {
806 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
807 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
808 		vfs_mp_count_sub_pcpu(mp, ref, 1);
809 		vfs_op_thread_exit(mp);
810 		return;
811 	}
812 
813 	MNT_ILOCK(mp);
814 	vfs_assert_mount_counters(mp);
815 	MNT_REL(mp);
816 	c = --mp->mnt_lockref;
817 	if (mp->mnt_vfs_ops == 0) {
818 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
819 		MNT_IUNLOCK(mp);
820 		return;
821 	}
822 	if (c < 0)
823 		vfs_dump_mount_counters(mp);
824 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
825 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
826 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
827 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
828 		wakeup(&mp->mnt_lockref);
829 	}
830 	MNT_IUNLOCK(mp);
831 }
832 
833 /*
834  * Lookup a mount point by filesystem identifier.
835  */
836 struct mount *
837 vfs_getvfs(fsid_t *fsid)
838 {
839 	struct mount *mp;
840 
841 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
842 	mtx_lock(&mountlist_mtx);
843 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
844 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
845 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
846 			vfs_ref(mp);
847 			mtx_unlock(&mountlist_mtx);
848 			return (mp);
849 		}
850 	}
851 	mtx_unlock(&mountlist_mtx);
852 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
853 	return ((struct mount *) 0);
854 }
855 
856 /*
857  * Lookup a mount point by filesystem identifier, busying it before
858  * returning.
859  *
860  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
861  * cache for popular filesystem identifiers.  The cache is lockess, using
862  * the fact that struct mount's are never freed.  In worst case we may
863  * get pointer to unmounted or even different filesystem, so we have to
864  * check what we got, and go slow way if so.
865  */
866 struct mount *
867 vfs_busyfs(fsid_t *fsid)
868 {
869 #define	FSID_CACHE_SIZE	256
870 	typedef struct mount * volatile vmp_t;
871 	static vmp_t cache[FSID_CACHE_SIZE];
872 	struct mount *mp;
873 	int error;
874 	uint32_t hash;
875 
876 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
877 	hash = fsid->val[0] ^ fsid->val[1];
878 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
879 	mp = cache[hash];
880 	if (mp == NULL ||
881 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
882 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
883 		goto slow;
884 	if (vfs_busy(mp, 0) != 0) {
885 		cache[hash] = NULL;
886 		goto slow;
887 	}
888 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
889 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
890 		return (mp);
891 	else
892 	    vfs_unbusy(mp);
893 
894 slow:
895 	mtx_lock(&mountlist_mtx);
896 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
897 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
898 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
899 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
900 			if (error) {
901 				cache[hash] = NULL;
902 				mtx_unlock(&mountlist_mtx);
903 				return (NULL);
904 			}
905 			cache[hash] = mp;
906 			return (mp);
907 		}
908 	}
909 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
910 	mtx_unlock(&mountlist_mtx);
911 	return ((struct mount *) 0);
912 }
913 
914 /*
915  * Check if a user can access privileged mount options.
916  */
917 int
918 vfs_suser(struct mount *mp, struct thread *td)
919 {
920 	int error;
921 
922 	if (jailed(td->td_ucred)) {
923 		/*
924 		 * If the jail of the calling thread lacks permission for
925 		 * this type of file system, deny immediately.
926 		 */
927 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
928 			return (EPERM);
929 
930 		/*
931 		 * If the file system was mounted outside the jail of the
932 		 * calling thread, deny immediately.
933 		 */
934 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
935 			return (EPERM);
936 	}
937 
938 	/*
939 	 * If file system supports delegated administration, we don't check
940 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
941 	 * by the file system itself.
942 	 * If this is not the user that did original mount, we check for
943 	 * the PRIV_VFS_MOUNT_OWNER privilege.
944 	 */
945 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
946 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
947 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
948 			return (error);
949 	}
950 	return (0);
951 }
952 
953 /*
954  * Get a new unique fsid.  Try to make its val[0] unique, since this value
955  * will be used to create fake device numbers for stat().  Also try (but
956  * not so hard) make its val[0] unique mod 2^16, since some emulators only
957  * support 16-bit device numbers.  We end up with unique val[0]'s for the
958  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
959  *
960  * Keep in mind that several mounts may be running in parallel.  Starting
961  * the search one past where the previous search terminated is both a
962  * micro-optimization and a defense against returning the same fsid to
963  * different mounts.
964  */
965 void
966 vfs_getnewfsid(struct mount *mp)
967 {
968 	static uint16_t mntid_base;
969 	struct mount *nmp;
970 	fsid_t tfsid;
971 	int mtype;
972 
973 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
974 	mtx_lock(&mntid_mtx);
975 	mtype = mp->mnt_vfc->vfc_typenum;
976 	tfsid.val[1] = mtype;
977 	mtype = (mtype & 0xFF) << 24;
978 	for (;;) {
979 		tfsid.val[0] = makedev(255,
980 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
981 		mntid_base++;
982 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
983 			break;
984 		vfs_rel(nmp);
985 	}
986 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
987 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
988 	mtx_unlock(&mntid_mtx);
989 }
990 
991 /*
992  * Knob to control the precision of file timestamps:
993  *
994  *   0 = seconds only; nanoseconds zeroed.
995  *   1 = seconds and nanoseconds, accurate within 1/HZ.
996  *   2 = seconds and nanoseconds, truncated to microseconds.
997  * >=3 = seconds and nanoseconds, maximum precision.
998  */
999 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1000 
1001 static int timestamp_precision = TSP_USEC;
1002 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1003     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1004     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1005     "3+: sec + ns (max. precision))");
1006 
1007 /*
1008  * Get a current timestamp.
1009  */
1010 void
1011 vfs_timestamp(struct timespec *tsp)
1012 {
1013 	struct timeval tv;
1014 
1015 	switch (timestamp_precision) {
1016 	case TSP_SEC:
1017 		tsp->tv_sec = time_second;
1018 		tsp->tv_nsec = 0;
1019 		break;
1020 	case TSP_HZ:
1021 		getnanotime(tsp);
1022 		break;
1023 	case TSP_USEC:
1024 		microtime(&tv);
1025 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1026 		break;
1027 	case TSP_NSEC:
1028 	default:
1029 		nanotime(tsp);
1030 		break;
1031 	}
1032 }
1033 
1034 /*
1035  * Set vnode attributes to VNOVAL
1036  */
1037 void
1038 vattr_null(struct vattr *vap)
1039 {
1040 
1041 	vap->va_type = VNON;
1042 	vap->va_size = VNOVAL;
1043 	vap->va_bytes = VNOVAL;
1044 	vap->va_mode = VNOVAL;
1045 	vap->va_nlink = VNOVAL;
1046 	vap->va_uid = VNOVAL;
1047 	vap->va_gid = VNOVAL;
1048 	vap->va_fsid = VNOVAL;
1049 	vap->va_fileid = VNOVAL;
1050 	vap->va_blocksize = VNOVAL;
1051 	vap->va_rdev = VNOVAL;
1052 	vap->va_atime.tv_sec = VNOVAL;
1053 	vap->va_atime.tv_nsec = VNOVAL;
1054 	vap->va_mtime.tv_sec = VNOVAL;
1055 	vap->va_mtime.tv_nsec = VNOVAL;
1056 	vap->va_ctime.tv_sec = VNOVAL;
1057 	vap->va_ctime.tv_nsec = VNOVAL;
1058 	vap->va_birthtime.tv_sec = VNOVAL;
1059 	vap->va_birthtime.tv_nsec = VNOVAL;
1060 	vap->va_flags = VNOVAL;
1061 	vap->va_gen = VNOVAL;
1062 	vap->va_vaflags = 0;
1063 }
1064 
1065 /*
1066  * Try to reduce the total number of vnodes.
1067  *
1068  * This routine (and its user) are buggy in at least the following ways:
1069  * - all parameters were picked years ago when RAM sizes were significantly
1070  *   smaller
1071  * - it can pick vnodes based on pages used by the vm object, but filesystems
1072  *   like ZFS don't use it making the pick broken
1073  * - since ZFS has its own aging policy it gets partially combated by this one
1074  * - a dedicated method should be provided for filesystems to let them decide
1075  *   whether the vnode should be recycled
1076  *
1077  * This routine is called when we have too many vnodes.  It attempts
1078  * to free <count> vnodes and will potentially free vnodes that still
1079  * have VM backing store (VM backing store is typically the cause
1080  * of a vnode blowout so we want to do this).  Therefore, this operation
1081  * is not considered cheap.
1082  *
1083  * A number of conditions may prevent a vnode from being reclaimed.
1084  * the buffer cache may have references on the vnode, a directory
1085  * vnode may still have references due to the namei cache representing
1086  * underlying files, or the vnode may be in active use.   It is not
1087  * desirable to reuse such vnodes.  These conditions may cause the
1088  * number of vnodes to reach some minimum value regardless of what
1089  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1090  *
1091  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1092  * 			 entries if this argument is strue
1093  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1094  *			 pages.
1095  * @param target	 How many vnodes to reclaim.
1096  * @return		 The number of vnodes that were reclaimed.
1097  */
1098 static int
1099 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1100 {
1101 	struct vnode *vp, *mvp;
1102 	struct mount *mp;
1103 	u_long done;
1104 	bool retried;
1105 
1106 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1107 
1108 	retried = false;
1109 	done = 0;
1110 
1111 	mvp = vnode_list_reclaim_marker;
1112 restart:
1113 	vp = mvp;
1114 	while (done < target) {
1115 		vp = TAILQ_NEXT(vp, v_vnodelist);
1116 		if (__predict_false(vp == NULL))
1117 			break;
1118 
1119 		if (__predict_false(vp->v_type == VMARKER))
1120 			continue;
1121 
1122 		/*
1123 		 * If it's been deconstructed already, it's still
1124 		 * referenced, or it exceeds the trigger, skip it.
1125 		 * Also skip free vnodes.  We are trying to make space
1126 		 * to expand the free list, not reduce it.
1127 		 */
1128 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1129 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1130 			goto next_iter;
1131 
1132 		if (vp->v_type == VBAD || vp->v_type == VNON)
1133 			goto next_iter;
1134 
1135 		if (!VI_TRYLOCK(vp))
1136 			goto next_iter;
1137 
1138 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1139 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1140 		    vp->v_type == VBAD || vp->v_type == VNON ||
1141 		    (vp->v_object != NULL &&
1142 		    vp->v_object->resident_page_count > trigger)) {
1143 			VI_UNLOCK(vp);
1144 			goto next_iter;
1145 		}
1146 		vholdl(vp);
1147 		VI_UNLOCK(vp);
1148 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1149 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1150 		mtx_unlock(&vnode_list_mtx);
1151 
1152 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1153 			vdrop(vp);
1154 			goto next_iter_unlocked;
1155 		}
1156 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1157 			vdrop(vp);
1158 			vn_finished_write(mp);
1159 			goto next_iter_unlocked;
1160 		}
1161 
1162 		VI_LOCK(vp);
1163 		if (vp->v_usecount > 0 ||
1164 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1165 		    (vp->v_object != NULL &&
1166 		    vp->v_object->resident_page_count > trigger)) {
1167 			VOP_UNLOCK(vp);
1168 			vdropl(vp);
1169 			vn_finished_write(mp);
1170 			goto next_iter_unlocked;
1171 		}
1172 		counter_u64_add(recycles_count, 1);
1173 		vgonel(vp);
1174 		VOP_UNLOCK(vp);
1175 		vdropl(vp);
1176 		vn_finished_write(mp);
1177 		done++;
1178 next_iter_unlocked:
1179 		if (should_yield())
1180 			kern_yield(PRI_USER);
1181 		mtx_lock(&vnode_list_mtx);
1182 		goto restart;
1183 next_iter:
1184 		MPASS(vp->v_type != VMARKER);
1185 		if (!should_yield())
1186 			continue;
1187 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1188 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1189 		mtx_unlock(&vnode_list_mtx);
1190 		kern_yield(PRI_USER);
1191 		mtx_lock(&vnode_list_mtx);
1192 		goto restart;
1193 	}
1194 	if (done == 0 && !retried) {
1195 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1196 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1197 		retried = true;
1198 		goto restart;
1199 	}
1200 	return (done);
1201 }
1202 
1203 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1204 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1205     0,
1206     "limit on vnode free requests per call to the vnlru_free routine");
1207 
1208 /*
1209  * Attempt to reduce the free list by the requested amount.
1210  */
1211 static int
1212 vnlru_free_locked(int count, struct vfsops *mnt_op)
1213 {
1214 	struct vnode *vp, *mvp;
1215 	struct mount *mp;
1216 	int ocount;
1217 
1218 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1219 	if (count > max_vnlru_free)
1220 		count = max_vnlru_free;
1221 	ocount = count;
1222 	mvp = vnode_list_free_marker;
1223 restart:
1224 	vp = mvp;
1225 	while (count > 0) {
1226 		vp = TAILQ_NEXT(vp, v_vnodelist);
1227 		if (__predict_false(vp == NULL)) {
1228 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1229 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1230 			break;
1231 		}
1232 		if (__predict_false(vp->v_type == VMARKER))
1233 			continue;
1234 
1235 		/*
1236 		 * Don't recycle if our vnode is from different type
1237 		 * of mount point.  Note that mp is type-safe, the
1238 		 * check does not reach unmapped address even if
1239 		 * vnode is reclaimed.
1240 		 * Don't recycle if we can't get the interlock without
1241 		 * blocking.
1242 		 */
1243 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1244 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1245 			continue;
1246 		}
1247 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1248 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1249 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1250 			VI_UNLOCK(vp);
1251 			continue;
1252 		}
1253 		vholdl(vp);
1254 		count--;
1255 		mtx_unlock(&vnode_list_mtx);
1256 		VI_UNLOCK(vp);
1257 		vtryrecycle(vp);
1258 		vdrop(vp);
1259 		mtx_lock(&vnode_list_mtx);
1260 		goto restart;
1261 	}
1262 	return (ocount - count);
1263 }
1264 
1265 void
1266 vnlru_free(int count, struct vfsops *mnt_op)
1267 {
1268 
1269 	mtx_lock(&vnode_list_mtx);
1270 	vnlru_free_locked(count, mnt_op);
1271 	mtx_unlock(&vnode_list_mtx);
1272 }
1273 
1274 static void
1275 vnlru_recalc(void)
1276 {
1277 
1278 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1279 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1280 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1281 	vlowat = vhiwat / 2;
1282 }
1283 
1284 /*
1285  * Attempt to recycle vnodes in a context that is always safe to block.
1286  * Calling vlrurecycle() from the bowels of filesystem code has some
1287  * interesting deadlock problems.
1288  */
1289 static struct proc *vnlruproc;
1290 static int vnlruproc_sig;
1291 
1292 /*
1293  * The main freevnodes counter is only updated when threads requeue their vnode
1294  * batches. CPUs are conditionally walked to compute a more accurate total.
1295  *
1296  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1297  * at any given moment can still exceed slop, but it should not be by significant
1298  * margin in practice.
1299  */
1300 #define VNLRU_FREEVNODES_SLOP 128
1301 
1302 static u_long
1303 vnlru_read_freevnodes(void)
1304 {
1305 	struct vdbatch *vd;
1306 	long slop;
1307 	int cpu;
1308 
1309 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1310 	if (freevnodes > freevnodes_old)
1311 		slop = freevnodes - freevnodes_old;
1312 	else
1313 		slop = freevnodes_old - freevnodes;
1314 	if (slop < VNLRU_FREEVNODES_SLOP)
1315 		return (freevnodes >= 0 ? freevnodes : 0);
1316 	freevnodes_old = freevnodes;
1317 	CPU_FOREACH(cpu) {
1318 		vd = DPCPU_ID_PTR((cpu), vd);
1319 		freevnodes_old += vd->freevnodes;
1320 	}
1321 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1322 }
1323 
1324 static bool
1325 vnlru_under(u_long rnumvnodes, u_long limit)
1326 {
1327 	u_long rfreevnodes, space;
1328 
1329 	if (__predict_false(rnumvnodes > desiredvnodes))
1330 		return (true);
1331 
1332 	space = desiredvnodes - rnumvnodes;
1333 	if (space < limit) {
1334 		rfreevnodes = vnlru_read_freevnodes();
1335 		if (rfreevnodes > wantfreevnodes)
1336 			space += rfreevnodes - wantfreevnodes;
1337 	}
1338 	return (space < limit);
1339 }
1340 
1341 static bool
1342 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1343 {
1344 	long rfreevnodes, space;
1345 
1346 	if (__predict_false(rnumvnodes > desiredvnodes))
1347 		return (true);
1348 
1349 	space = desiredvnodes - rnumvnodes;
1350 	if (space < limit) {
1351 		rfreevnodes = atomic_load_long(&freevnodes);
1352 		if (rfreevnodes > wantfreevnodes)
1353 			space += rfreevnodes - wantfreevnodes;
1354 	}
1355 	return (space < limit);
1356 }
1357 
1358 static void
1359 vnlru_kick(void)
1360 {
1361 
1362 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1363 	if (vnlruproc_sig == 0) {
1364 		vnlruproc_sig = 1;
1365 		wakeup(vnlruproc);
1366 	}
1367 }
1368 
1369 static void
1370 vnlru_proc(void)
1371 {
1372 	u_long rnumvnodes, rfreevnodes, target;
1373 	unsigned long onumvnodes;
1374 	int done, force, trigger, usevnodes;
1375 	bool reclaim_nc_src, want_reread;
1376 
1377 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1378 	    SHUTDOWN_PRI_FIRST);
1379 
1380 	force = 0;
1381 	want_reread = false;
1382 	for (;;) {
1383 		kproc_suspend_check(vnlruproc);
1384 		mtx_lock(&vnode_list_mtx);
1385 		rnumvnodes = atomic_load_long(&numvnodes);
1386 
1387 		if (want_reread) {
1388 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1389 			want_reread = false;
1390 		}
1391 
1392 		/*
1393 		 * If numvnodes is too large (due to desiredvnodes being
1394 		 * adjusted using its sysctl, or emergency growth), first
1395 		 * try to reduce it by discarding from the free list.
1396 		 */
1397 		if (rnumvnodes > desiredvnodes) {
1398 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1399 			rnumvnodes = atomic_load_long(&numvnodes);
1400 		}
1401 		/*
1402 		 * Sleep if the vnode cache is in a good state.  This is
1403 		 * when it is not over-full and has space for about a 4%
1404 		 * or 9% expansion (by growing its size or inexcessively
1405 		 * reducing its free list).  Otherwise, try to reclaim
1406 		 * space for a 10% expansion.
1407 		 */
1408 		if (vstir && force == 0) {
1409 			force = 1;
1410 			vstir = 0;
1411 		}
1412 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1413 			vnlruproc_sig = 0;
1414 			wakeup(&vnlruproc_sig);
1415 			msleep(vnlruproc, &vnode_list_mtx,
1416 			    PVFS|PDROP, "vlruwt", hz);
1417 			continue;
1418 		}
1419 		rfreevnodes = vnlru_read_freevnodes();
1420 
1421 		onumvnodes = rnumvnodes;
1422 		/*
1423 		 * Calculate parameters for recycling.  These are the same
1424 		 * throughout the loop to give some semblance of fairness.
1425 		 * The trigger point is to avoid recycling vnodes with lots
1426 		 * of resident pages.  We aren't trying to free memory; we
1427 		 * are trying to recycle or at least free vnodes.
1428 		 */
1429 		if (rnumvnodes <= desiredvnodes)
1430 			usevnodes = rnumvnodes - rfreevnodes;
1431 		else
1432 			usevnodes = rnumvnodes;
1433 		if (usevnodes <= 0)
1434 			usevnodes = 1;
1435 		/*
1436 		 * The trigger value is is chosen to give a conservatively
1437 		 * large value to ensure that it alone doesn't prevent
1438 		 * making progress.  The value can easily be so large that
1439 		 * it is effectively infinite in some congested and
1440 		 * misconfigured cases, and this is necessary.  Normally
1441 		 * it is about 8 to 100 (pages), which is quite large.
1442 		 */
1443 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1444 		if (force < 2)
1445 			trigger = vsmalltrigger;
1446 		reclaim_nc_src = force >= 3;
1447 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1448 		target = target / 10 + 1;
1449 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1450 		mtx_unlock(&vnode_list_mtx);
1451 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1452 			uma_reclaim(UMA_RECLAIM_DRAIN);
1453 		if (done == 0) {
1454 			if (force == 0 || force == 1) {
1455 				force = 2;
1456 				continue;
1457 			}
1458 			if (force == 2) {
1459 				force = 3;
1460 				continue;
1461 			}
1462 			want_reread = true;
1463 			force = 0;
1464 			vnlru_nowhere++;
1465 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1466 		} else {
1467 			want_reread = true;
1468 			kern_yield(PRI_USER);
1469 		}
1470 	}
1471 }
1472 
1473 static struct kproc_desc vnlru_kp = {
1474 	"vnlru",
1475 	vnlru_proc,
1476 	&vnlruproc
1477 };
1478 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1479     &vnlru_kp);
1480 
1481 /*
1482  * Routines having to do with the management of the vnode table.
1483  */
1484 
1485 /*
1486  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1487  * before we actually vgone().  This function must be called with the vnode
1488  * held to prevent the vnode from being returned to the free list midway
1489  * through vgone().
1490  */
1491 static int
1492 vtryrecycle(struct vnode *vp)
1493 {
1494 	struct mount *vnmp;
1495 
1496 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1497 	VNASSERT(vp->v_holdcnt, vp,
1498 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1499 	/*
1500 	 * This vnode may found and locked via some other list, if so we
1501 	 * can't recycle it yet.
1502 	 */
1503 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1504 		CTR2(KTR_VFS,
1505 		    "%s: impossible to recycle, vp %p lock is already held",
1506 		    __func__, vp);
1507 		return (EWOULDBLOCK);
1508 	}
1509 	/*
1510 	 * Don't recycle if its filesystem is being suspended.
1511 	 */
1512 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1513 		VOP_UNLOCK(vp);
1514 		CTR2(KTR_VFS,
1515 		    "%s: impossible to recycle, cannot start the write for %p",
1516 		    __func__, vp);
1517 		return (EBUSY);
1518 	}
1519 	/*
1520 	 * If we got this far, we need to acquire the interlock and see if
1521 	 * anyone picked up this vnode from another list.  If not, we will
1522 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1523 	 * will skip over it.
1524 	 */
1525 	VI_LOCK(vp);
1526 	if (vp->v_usecount) {
1527 		VOP_UNLOCK(vp);
1528 		VI_UNLOCK(vp);
1529 		vn_finished_write(vnmp);
1530 		CTR2(KTR_VFS,
1531 		    "%s: impossible to recycle, %p is already referenced",
1532 		    __func__, vp);
1533 		return (EBUSY);
1534 	}
1535 	if (!VN_IS_DOOMED(vp)) {
1536 		counter_u64_add(recycles_free_count, 1);
1537 		vgonel(vp);
1538 	}
1539 	VOP_UNLOCK(vp);
1540 	VI_UNLOCK(vp);
1541 	vn_finished_write(vnmp);
1542 	return (0);
1543 }
1544 
1545 /*
1546  * Allocate a new vnode.
1547  *
1548  * The operation never returns an error. Returning an error was disabled
1549  * in r145385 (dated 2005) with the following comment:
1550  *
1551  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1552  *
1553  * Given the age of this commit (almost 15 years at the time of writing this
1554  * comment) restoring the ability to fail requires a significant audit of
1555  * all codepaths.
1556  *
1557  * The routine can try to free a vnode or stall for up to 1 second waiting for
1558  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1559  */
1560 static u_long vn_alloc_cyclecount;
1561 
1562 static struct vnode * __noinline
1563 vn_alloc_hard(struct mount *mp)
1564 {
1565 	u_long rnumvnodes, rfreevnodes;
1566 
1567 	mtx_lock(&vnode_list_mtx);
1568 	rnumvnodes = atomic_load_long(&numvnodes);
1569 	if (rnumvnodes + 1 < desiredvnodes) {
1570 		vn_alloc_cyclecount = 0;
1571 		goto alloc;
1572 	}
1573 	rfreevnodes = vnlru_read_freevnodes();
1574 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1575 		vn_alloc_cyclecount = 0;
1576 		vstir = 1;
1577 	}
1578 	/*
1579 	 * Grow the vnode cache if it will not be above its target max
1580 	 * after growing.  Otherwise, if the free list is nonempty, try
1581 	 * to reclaim 1 item from it before growing the cache (possibly
1582 	 * above its target max if the reclamation failed or is delayed).
1583 	 * Otherwise, wait for some space.  In all cases, schedule
1584 	 * vnlru_proc() if we are getting short of space.  The watermarks
1585 	 * should be chosen so that we never wait or even reclaim from
1586 	 * the free list to below its target minimum.
1587 	 */
1588 	if (vnlru_free_locked(1, NULL) > 0)
1589 		goto alloc;
1590 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1591 		/*
1592 		 * Wait for space for a new vnode.
1593 		 */
1594 		vnlru_kick();
1595 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1596 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1597 		    vnlru_read_freevnodes() > 1)
1598 			vnlru_free_locked(1, NULL);
1599 	}
1600 alloc:
1601 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1602 	if (vnlru_under(rnumvnodes, vlowat))
1603 		vnlru_kick();
1604 	mtx_unlock(&vnode_list_mtx);
1605 	return (uma_zalloc(vnode_zone, M_WAITOK));
1606 }
1607 
1608 static struct vnode *
1609 vn_alloc(struct mount *mp)
1610 {
1611 	u_long rnumvnodes;
1612 
1613 	if (__predict_false(vn_alloc_cyclecount != 0))
1614 		return (vn_alloc_hard(mp));
1615 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1616 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1617 		atomic_subtract_long(&numvnodes, 1);
1618 		return (vn_alloc_hard(mp));
1619 	}
1620 
1621 	return (uma_zalloc(vnode_zone, M_WAITOK));
1622 }
1623 
1624 static void
1625 vn_free(struct vnode *vp)
1626 {
1627 
1628 	atomic_subtract_long(&numvnodes, 1);
1629 	uma_zfree(vnode_zone, vp);
1630 }
1631 
1632 /*
1633  * Return the next vnode from the free list.
1634  */
1635 int
1636 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1637     struct vnode **vpp)
1638 {
1639 	struct vnode *vp;
1640 	struct thread *td;
1641 	struct lock_object *lo;
1642 
1643 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1644 
1645 	KASSERT(vops->registered,
1646 	    ("%s: not registered vector op %p\n", __func__, vops));
1647 
1648 	td = curthread;
1649 	if (td->td_vp_reserved != NULL) {
1650 		vp = td->td_vp_reserved;
1651 		td->td_vp_reserved = NULL;
1652 	} else {
1653 		vp = vn_alloc(mp);
1654 	}
1655 	counter_u64_add(vnodes_created, 1);
1656 	/*
1657 	 * Locks are given the generic name "vnode" when created.
1658 	 * Follow the historic practice of using the filesystem
1659 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1660 	 *
1661 	 * Locks live in a witness group keyed on their name. Thus,
1662 	 * when a lock is renamed, it must also move from the witness
1663 	 * group of its old name to the witness group of its new name.
1664 	 *
1665 	 * The change only needs to be made when the vnode moves
1666 	 * from one filesystem type to another. We ensure that each
1667 	 * filesystem use a single static name pointer for its tag so
1668 	 * that we can compare pointers rather than doing a strcmp().
1669 	 */
1670 	lo = &vp->v_vnlock->lock_object;
1671 #ifdef WITNESS
1672 	if (lo->lo_name != tag) {
1673 #endif
1674 		lo->lo_name = tag;
1675 #ifdef WITNESS
1676 		WITNESS_DESTROY(lo);
1677 		WITNESS_INIT(lo, tag);
1678 	}
1679 #endif
1680 	/*
1681 	 * By default, don't allow shared locks unless filesystems opt-in.
1682 	 */
1683 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1684 	/*
1685 	 * Finalize various vnode identity bits.
1686 	 */
1687 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1688 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1689 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1690 	vp->v_type = VNON;
1691 	vp->v_op = vops;
1692 	v_init_counters(vp);
1693 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1694 #ifdef DIAGNOSTIC
1695 	if (mp == NULL && vops != &dead_vnodeops)
1696 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1697 #endif
1698 #ifdef MAC
1699 	mac_vnode_init(vp);
1700 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1701 		mac_vnode_associate_singlelabel(mp, vp);
1702 #endif
1703 	if (mp != NULL) {
1704 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1705 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1706 			vp->v_vflag |= VV_NOKNOTE;
1707 	}
1708 
1709 	/*
1710 	 * For the filesystems which do not use vfs_hash_insert(),
1711 	 * still initialize v_hash to have vfs_hash_index() useful.
1712 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1713 	 * its own hashing.
1714 	 */
1715 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1716 
1717 	*vpp = vp;
1718 	return (0);
1719 }
1720 
1721 void
1722 getnewvnode_reserve(void)
1723 {
1724 	struct thread *td;
1725 
1726 	td = curthread;
1727 	MPASS(td->td_vp_reserved == NULL);
1728 	td->td_vp_reserved = vn_alloc(NULL);
1729 }
1730 
1731 void
1732 getnewvnode_drop_reserve(void)
1733 {
1734 	struct thread *td;
1735 
1736 	td = curthread;
1737 	if (td->td_vp_reserved != NULL) {
1738 		vn_free(td->td_vp_reserved);
1739 		td->td_vp_reserved = NULL;
1740 	}
1741 }
1742 
1743 static void
1744 freevnode(struct vnode *vp)
1745 {
1746 	struct bufobj *bo;
1747 
1748 	/*
1749 	 * The vnode has been marked for destruction, so free it.
1750 	 *
1751 	 * The vnode will be returned to the zone where it will
1752 	 * normally remain until it is needed for another vnode. We
1753 	 * need to cleanup (or verify that the cleanup has already
1754 	 * been done) any residual data left from its current use
1755 	 * so as not to contaminate the freshly allocated vnode.
1756 	 */
1757 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1758 	bo = &vp->v_bufobj;
1759 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1760 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1761 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1762 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1763 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1764 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1765 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1766 	    ("clean blk trie not empty"));
1767 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1768 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1769 	    ("dirty blk trie not empty"));
1770 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1771 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1772 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1773 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1774 	    ("Dangling rangelock waiters"));
1775 	VI_UNLOCK(vp);
1776 #ifdef MAC
1777 	mac_vnode_destroy(vp);
1778 #endif
1779 	if (vp->v_pollinfo != NULL) {
1780 		destroy_vpollinfo(vp->v_pollinfo);
1781 		vp->v_pollinfo = NULL;
1782 	}
1783 #ifdef INVARIANTS
1784 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1785 	vp->v_op = NULL;
1786 #endif
1787 	vp->v_mountedhere = NULL;
1788 	vp->v_unpcb = NULL;
1789 	vp->v_rdev = NULL;
1790 	vp->v_fifoinfo = NULL;
1791 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1792 	vp->v_irflag = 0;
1793 	vp->v_iflag = 0;
1794 	vp->v_vflag = 0;
1795 	bo->bo_flag = 0;
1796 	vn_free(vp);
1797 }
1798 
1799 /*
1800  * Delete from old mount point vnode list, if on one.
1801  */
1802 static void
1803 delmntque(struct vnode *vp)
1804 {
1805 	struct mount *mp;
1806 
1807 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1808 
1809 	mp = vp->v_mount;
1810 	if (mp == NULL)
1811 		return;
1812 	MNT_ILOCK(mp);
1813 	VI_LOCK(vp);
1814 	vp->v_mount = NULL;
1815 	VI_UNLOCK(vp);
1816 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1817 		("bad mount point vnode list size"));
1818 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1819 	mp->mnt_nvnodelistsize--;
1820 	MNT_REL(mp);
1821 	MNT_IUNLOCK(mp);
1822 }
1823 
1824 static void
1825 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1826 {
1827 
1828 	vp->v_data = NULL;
1829 	vp->v_op = &dead_vnodeops;
1830 	vgone(vp);
1831 	vput(vp);
1832 }
1833 
1834 /*
1835  * Insert into list of vnodes for the new mount point, if available.
1836  */
1837 int
1838 insmntque1(struct vnode *vp, struct mount *mp,
1839 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1840 {
1841 
1842 	KASSERT(vp->v_mount == NULL,
1843 		("insmntque: vnode already on per mount vnode list"));
1844 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1845 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1846 
1847 	/*
1848 	 * We acquire the vnode interlock early to ensure that the
1849 	 * vnode cannot be recycled by another process releasing a
1850 	 * holdcnt on it before we get it on both the vnode list
1851 	 * and the active vnode list. The mount mutex protects only
1852 	 * manipulation of the vnode list and the vnode freelist
1853 	 * mutex protects only manipulation of the active vnode list.
1854 	 * Hence the need to hold the vnode interlock throughout.
1855 	 */
1856 	MNT_ILOCK(mp);
1857 	VI_LOCK(vp);
1858 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1859 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1860 	    mp->mnt_nvnodelistsize == 0)) &&
1861 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1862 		VI_UNLOCK(vp);
1863 		MNT_IUNLOCK(mp);
1864 		if (dtr != NULL)
1865 			dtr(vp, dtr_arg);
1866 		return (EBUSY);
1867 	}
1868 	vp->v_mount = mp;
1869 	MNT_REF(mp);
1870 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1871 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1872 		("neg mount point vnode list size"));
1873 	mp->mnt_nvnodelistsize++;
1874 	VI_UNLOCK(vp);
1875 	MNT_IUNLOCK(mp);
1876 	return (0);
1877 }
1878 
1879 int
1880 insmntque(struct vnode *vp, struct mount *mp)
1881 {
1882 
1883 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1884 }
1885 
1886 /*
1887  * Flush out and invalidate all buffers associated with a bufobj
1888  * Called with the underlying object locked.
1889  */
1890 int
1891 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1892 {
1893 	int error;
1894 
1895 	BO_LOCK(bo);
1896 	if (flags & V_SAVE) {
1897 		error = bufobj_wwait(bo, slpflag, slptimeo);
1898 		if (error) {
1899 			BO_UNLOCK(bo);
1900 			return (error);
1901 		}
1902 		if (bo->bo_dirty.bv_cnt > 0) {
1903 			BO_UNLOCK(bo);
1904 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1905 				return (error);
1906 			/*
1907 			 * XXX We could save a lock/unlock if this was only
1908 			 * enabled under INVARIANTS
1909 			 */
1910 			BO_LOCK(bo);
1911 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1912 				panic("vinvalbuf: dirty bufs");
1913 		}
1914 	}
1915 	/*
1916 	 * If you alter this loop please notice that interlock is dropped and
1917 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1918 	 * no race conditions occur from this.
1919 	 */
1920 	do {
1921 		error = flushbuflist(&bo->bo_clean,
1922 		    flags, bo, slpflag, slptimeo);
1923 		if (error == 0 && !(flags & V_CLEANONLY))
1924 			error = flushbuflist(&bo->bo_dirty,
1925 			    flags, bo, slpflag, slptimeo);
1926 		if (error != 0 && error != EAGAIN) {
1927 			BO_UNLOCK(bo);
1928 			return (error);
1929 		}
1930 	} while (error != 0);
1931 
1932 	/*
1933 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1934 	 * have write I/O in-progress but if there is a VM object then the
1935 	 * VM object can also have read-I/O in-progress.
1936 	 */
1937 	do {
1938 		bufobj_wwait(bo, 0, 0);
1939 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1940 			BO_UNLOCK(bo);
1941 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1942 			BO_LOCK(bo);
1943 		}
1944 	} while (bo->bo_numoutput > 0);
1945 	BO_UNLOCK(bo);
1946 
1947 	/*
1948 	 * Destroy the copy in the VM cache, too.
1949 	 */
1950 	if (bo->bo_object != NULL &&
1951 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1952 		VM_OBJECT_WLOCK(bo->bo_object);
1953 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1954 		    OBJPR_CLEANONLY : 0);
1955 		VM_OBJECT_WUNLOCK(bo->bo_object);
1956 	}
1957 
1958 #ifdef INVARIANTS
1959 	BO_LOCK(bo);
1960 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1961 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1962 	    bo->bo_clean.bv_cnt > 0))
1963 		panic("vinvalbuf: flush failed");
1964 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1965 	    bo->bo_dirty.bv_cnt > 0)
1966 		panic("vinvalbuf: flush dirty failed");
1967 	BO_UNLOCK(bo);
1968 #endif
1969 	return (0);
1970 }
1971 
1972 /*
1973  * Flush out and invalidate all buffers associated with a vnode.
1974  * Called with the underlying object locked.
1975  */
1976 int
1977 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1978 {
1979 
1980 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1981 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1982 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1983 		return (0);
1984 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1985 }
1986 
1987 /*
1988  * Flush out buffers on the specified list.
1989  *
1990  */
1991 static int
1992 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1993     int slptimeo)
1994 {
1995 	struct buf *bp, *nbp;
1996 	int retval, error;
1997 	daddr_t lblkno;
1998 	b_xflags_t xflags;
1999 
2000 	ASSERT_BO_WLOCKED(bo);
2001 
2002 	retval = 0;
2003 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2004 		/*
2005 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2006 		 * do not skip any buffers. If we are flushing only V_NORMAL
2007 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2008 		 * flushing only V_ALT buffers then skip buffers not marked
2009 		 * as BX_ALTDATA.
2010 		 */
2011 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2012 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2013 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2014 			continue;
2015 		}
2016 		if (nbp != NULL) {
2017 			lblkno = nbp->b_lblkno;
2018 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2019 		}
2020 		retval = EAGAIN;
2021 		error = BUF_TIMELOCK(bp,
2022 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2023 		    "flushbuf", slpflag, slptimeo);
2024 		if (error) {
2025 			BO_LOCK(bo);
2026 			return (error != ENOLCK ? error : EAGAIN);
2027 		}
2028 		KASSERT(bp->b_bufobj == bo,
2029 		    ("bp %p wrong b_bufobj %p should be %p",
2030 		    bp, bp->b_bufobj, bo));
2031 		/*
2032 		 * XXX Since there are no node locks for NFS, I
2033 		 * believe there is a slight chance that a delayed
2034 		 * write will occur while sleeping just above, so
2035 		 * check for it.
2036 		 */
2037 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2038 		    (flags & V_SAVE)) {
2039 			bremfree(bp);
2040 			bp->b_flags |= B_ASYNC;
2041 			bwrite(bp);
2042 			BO_LOCK(bo);
2043 			return (EAGAIN);	/* XXX: why not loop ? */
2044 		}
2045 		bremfree(bp);
2046 		bp->b_flags |= (B_INVAL | B_RELBUF);
2047 		bp->b_flags &= ~B_ASYNC;
2048 		brelse(bp);
2049 		BO_LOCK(bo);
2050 		if (nbp == NULL)
2051 			break;
2052 		nbp = gbincore(bo, lblkno);
2053 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2054 		    != xflags)
2055 			break;			/* nbp invalid */
2056 	}
2057 	return (retval);
2058 }
2059 
2060 int
2061 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2062 {
2063 	struct buf *bp;
2064 	int error;
2065 	daddr_t lblkno;
2066 
2067 	ASSERT_BO_LOCKED(bo);
2068 
2069 	for (lblkno = startn;;) {
2070 again:
2071 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2072 		if (bp == NULL || bp->b_lblkno >= endn ||
2073 		    bp->b_lblkno < startn)
2074 			break;
2075 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2076 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2077 		if (error != 0) {
2078 			BO_RLOCK(bo);
2079 			if (error == ENOLCK)
2080 				goto again;
2081 			return (error);
2082 		}
2083 		KASSERT(bp->b_bufobj == bo,
2084 		    ("bp %p wrong b_bufobj %p should be %p",
2085 		    bp, bp->b_bufobj, bo));
2086 		lblkno = bp->b_lblkno + 1;
2087 		if ((bp->b_flags & B_MANAGED) == 0)
2088 			bremfree(bp);
2089 		bp->b_flags |= B_RELBUF;
2090 		/*
2091 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2092 		 * pages backing each buffer in the range are unlikely to be
2093 		 * reused.  Dirty buffers will have the hint applied once
2094 		 * they've been written.
2095 		 */
2096 		if ((bp->b_flags & B_VMIO) != 0)
2097 			bp->b_flags |= B_NOREUSE;
2098 		brelse(bp);
2099 		BO_RLOCK(bo);
2100 	}
2101 	return (0);
2102 }
2103 
2104 /*
2105  * Truncate a file's buffer and pages to a specified length.  This
2106  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2107  * sync activity.
2108  */
2109 int
2110 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2111 {
2112 	struct buf *bp, *nbp;
2113 	struct bufobj *bo;
2114 	daddr_t startlbn;
2115 
2116 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2117 	    vp, blksize, (uintmax_t)length);
2118 
2119 	/*
2120 	 * Round up to the *next* lbn.
2121 	 */
2122 	startlbn = howmany(length, blksize);
2123 
2124 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2125 
2126 	bo = &vp->v_bufobj;
2127 restart_unlocked:
2128 	BO_LOCK(bo);
2129 
2130 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2131 		;
2132 
2133 	if (length > 0) {
2134 restartsync:
2135 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2136 			if (bp->b_lblkno > 0)
2137 				continue;
2138 			/*
2139 			 * Since we hold the vnode lock this should only
2140 			 * fail if we're racing with the buf daemon.
2141 			 */
2142 			if (BUF_LOCK(bp,
2143 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2144 			    BO_LOCKPTR(bo)) == ENOLCK)
2145 				goto restart_unlocked;
2146 
2147 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2148 			    ("buf(%p) on dirty queue without DELWRI", bp));
2149 
2150 			bremfree(bp);
2151 			bawrite(bp);
2152 			BO_LOCK(bo);
2153 			goto restartsync;
2154 		}
2155 	}
2156 
2157 	bufobj_wwait(bo, 0, 0);
2158 	BO_UNLOCK(bo);
2159 	vnode_pager_setsize(vp, length);
2160 
2161 	return (0);
2162 }
2163 
2164 /*
2165  * Invalidate the cached pages of a file's buffer within the range of block
2166  * numbers [startlbn, endlbn).
2167  */
2168 void
2169 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2170     int blksize)
2171 {
2172 	struct bufobj *bo;
2173 	off_t start, end;
2174 
2175 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2176 
2177 	start = blksize * startlbn;
2178 	end = blksize * endlbn;
2179 
2180 	bo = &vp->v_bufobj;
2181 	BO_LOCK(bo);
2182 	MPASS(blksize == bo->bo_bsize);
2183 
2184 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2185 		;
2186 
2187 	BO_UNLOCK(bo);
2188 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2189 }
2190 
2191 static int
2192 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2193     daddr_t startlbn, daddr_t endlbn)
2194 {
2195 	struct buf *bp, *nbp;
2196 	bool anyfreed;
2197 
2198 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2199 	ASSERT_BO_LOCKED(bo);
2200 
2201 	do {
2202 		anyfreed = false;
2203 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2204 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2205 				continue;
2206 			if (BUF_LOCK(bp,
2207 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2208 			    BO_LOCKPTR(bo)) == ENOLCK) {
2209 				BO_LOCK(bo);
2210 				return (EAGAIN);
2211 			}
2212 
2213 			bremfree(bp);
2214 			bp->b_flags |= B_INVAL | B_RELBUF;
2215 			bp->b_flags &= ~B_ASYNC;
2216 			brelse(bp);
2217 			anyfreed = true;
2218 
2219 			BO_LOCK(bo);
2220 			if (nbp != NULL &&
2221 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2222 			    nbp->b_vp != vp ||
2223 			    (nbp->b_flags & B_DELWRI) != 0))
2224 				return (EAGAIN);
2225 		}
2226 
2227 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2228 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2229 				continue;
2230 			if (BUF_LOCK(bp,
2231 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2232 			    BO_LOCKPTR(bo)) == ENOLCK) {
2233 				BO_LOCK(bo);
2234 				return (EAGAIN);
2235 			}
2236 			bremfree(bp);
2237 			bp->b_flags |= B_INVAL | B_RELBUF;
2238 			bp->b_flags &= ~B_ASYNC;
2239 			brelse(bp);
2240 			anyfreed = true;
2241 
2242 			BO_LOCK(bo);
2243 			if (nbp != NULL &&
2244 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2245 			    (nbp->b_vp != vp) ||
2246 			    (nbp->b_flags & B_DELWRI) == 0))
2247 				return (EAGAIN);
2248 		}
2249 	} while (anyfreed);
2250 	return (0);
2251 }
2252 
2253 static void
2254 buf_vlist_remove(struct buf *bp)
2255 {
2256 	struct bufv *bv;
2257 
2258 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2259 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2260 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2261 	    (BX_VNDIRTY|BX_VNCLEAN),
2262 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2263 	if (bp->b_xflags & BX_VNDIRTY)
2264 		bv = &bp->b_bufobj->bo_dirty;
2265 	else
2266 		bv = &bp->b_bufobj->bo_clean;
2267 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2268 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2269 	bv->bv_cnt--;
2270 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2271 }
2272 
2273 /*
2274  * Add the buffer to the sorted clean or dirty block list.
2275  *
2276  * NOTE: xflags is passed as a constant, optimizing this inline function!
2277  */
2278 static void
2279 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2280 {
2281 	struct bufv *bv;
2282 	struct buf *n;
2283 	int error;
2284 
2285 	ASSERT_BO_WLOCKED(bo);
2286 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2287 	    ("dead bo %p", bo));
2288 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2289 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2290 	bp->b_xflags |= xflags;
2291 	if (xflags & BX_VNDIRTY)
2292 		bv = &bo->bo_dirty;
2293 	else
2294 		bv = &bo->bo_clean;
2295 
2296 	/*
2297 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2298 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2299 	 * than _ge.
2300 	 */
2301 	if (bv->bv_cnt == 0 ||
2302 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2303 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2304 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2305 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2306 	else
2307 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2308 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2309 	if (error)
2310 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2311 	bv->bv_cnt++;
2312 }
2313 
2314 /*
2315  * Look up a buffer using the buffer tries.
2316  */
2317 struct buf *
2318 gbincore(struct bufobj *bo, daddr_t lblkno)
2319 {
2320 	struct buf *bp;
2321 
2322 	ASSERT_BO_LOCKED(bo);
2323 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2324 	if (bp != NULL)
2325 		return (bp);
2326 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2327 }
2328 
2329 /*
2330  * Associate a buffer with a vnode.
2331  */
2332 void
2333 bgetvp(struct vnode *vp, struct buf *bp)
2334 {
2335 	struct bufobj *bo;
2336 
2337 	bo = &vp->v_bufobj;
2338 	ASSERT_BO_WLOCKED(bo);
2339 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2340 
2341 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2342 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2343 	    ("bgetvp: bp already attached! %p", bp));
2344 
2345 	vhold(vp);
2346 	bp->b_vp = vp;
2347 	bp->b_bufobj = bo;
2348 	/*
2349 	 * Insert onto list for new vnode.
2350 	 */
2351 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2352 }
2353 
2354 /*
2355  * Disassociate a buffer from a vnode.
2356  */
2357 void
2358 brelvp(struct buf *bp)
2359 {
2360 	struct bufobj *bo;
2361 	struct vnode *vp;
2362 
2363 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2364 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2365 
2366 	/*
2367 	 * Delete from old vnode list, if on one.
2368 	 */
2369 	vp = bp->b_vp;		/* XXX */
2370 	bo = bp->b_bufobj;
2371 	BO_LOCK(bo);
2372 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2373 		buf_vlist_remove(bp);
2374 	else
2375 		panic("brelvp: Buffer %p not on queue.", bp);
2376 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2377 		bo->bo_flag &= ~BO_ONWORKLST;
2378 		mtx_lock(&sync_mtx);
2379 		LIST_REMOVE(bo, bo_synclist);
2380 		syncer_worklist_len--;
2381 		mtx_unlock(&sync_mtx);
2382 	}
2383 	bp->b_vp = NULL;
2384 	bp->b_bufobj = NULL;
2385 	BO_UNLOCK(bo);
2386 	vdrop(vp);
2387 }
2388 
2389 /*
2390  * Add an item to the syncer work queue.
2391  */
2392 static void
2393 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2394 {
2395 	int slot;
2396 
2397 	ASSERT_BO_WLOCKED(bo);
2398 
2399 	mtx_lock(&sync_mtx);
2400 	if (bo->bo_flag & BO_ONWORKLST)
2401 		LIST_REMOVE(bo, bo_synclist);
2402 	else {
2403 		bo->bo_flag |= BO_ONWORKLST;
2404 		syncer_worklist_len++;
2405 	}
2406 
2407 	if (delay > syncer_maxdelay - 2)
2408 		delay = syncer_maxdelay - 2;
2409 	slot = (syncer_delayno + delay) & syncer_mask;
2410 
2411 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2412 	mtx_unlock(&sync_mtx);
2413 }
2414 
2415 static int
2416 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2417 {
2418 	int error, len;
2419 
2420 	mtx_lock(&sync_mtx);
2421 	len = syncer_worklist_len - sync_vnode_count;
2422 	mtx_unlock(&sync_mtx);
2423 	error = SYSCTL_OUT(req, &len, sizeof(len));
2424 	return (error);
2425 }
2426 
2427 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2428     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2429     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2430 
2431 static struct proc *updateproc;
2432 static void sched_sync(void);
2433 static struct kproc_desc up_kp = {
2434 	"syncer",
2435 	sched_sync,
2436 	&updateproc
2437 };
2438 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2439 
2440 static int
2441 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2442 {
2443 	struct vnode *vp;
2444 	struct mount *mp;
2445 
2446 	*bo = LIST_FIRST(slp);
2447 	if (*bo == NULL)
2448 		return (0);
2449 	vp = bo2vnode(*bo);
2450 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2451 		return (1);
2452 	/*
2453 	 * We use vhold in case the vnode does not
2454 	 * successfully sync.  vhold prevents the vnode from
2455 	 * going away when we unlock the sync_mtx so that
2456 	 * we can acquire the vnode interlock.
2457 	 */
2458 	vholdl(vp);
2459 	mtx_unlock(&sync_mtx);
2460 	VI_UNLOCK(vp);
2461 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2462 		vdrop(vp);
2463 		mtx_lock(&sync_mtx);
2464 		return (*bo == LIST_FIRST(slp));
2465 	}
2466 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2467 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2468 	VOP_UNLOCK(vp);
2469 	vn_finished_write(mp);
2470 	BO_LOCK(*bo);
2471 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2472 		/*
2473 		 * Put us back on the worklist.  The worklist
2474 		 * routine will remove us from our current
2475 		 * position and then add us back in at a later
2476 		 * position.
2477 		 */
2478 		vn_syncer_add_to_worklist(*bo, syncdelay);
2479 	}
2480 	BO_UNLOCK(*bo);
2481 	vdrop(vp);
2482 	mtx_lock(&sync_mtx);
2483 	return (0);
2484 }
2485 
2486 static int first_printf = 1;
2487 
2488 /*
2489  * System filesystem synchronizer daemon.
2490  */
2491 static void
2492 sched_sync(void)
2493 {
2494 	struct synclist *next, *slp;
2495 	struct bufobj *bo;
2496 	long starttime;
2497 	struct thread *td = curthread;
2498 	int last_work_seen;
2499 	int net_worklist_len;
2500 	int syncer_final_iter;
2501 	int error;
2502 
2503 	last_work_seen = 0;
2504 	syncer_final_iter = 0;
2505 	syncer_state = SYNCER_RUNNING;
2506 	starttime = time_uptime;
2507 	td->td_pflags |= TDP_NORUNNINGBUF;
2508 
2509 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2510 	    SHUTDOWN_PRI_LAST);
2511 
2512 	mtx_lock(&sync_mtx);
2513 	for (;;) {
2514 		if (syncer_state == SYNCER_FINAL_DELAY &&
2515 		    syncer_final_iter == 0) {
2516 			mtx_unlock(&sync_mtx);
2517 			kproc_suspend_check(td->td_proc);
2518 			mtx_lock(&sync_mtx);
2519 		}
2520 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2521 		if (syncer_state != SYNCER_RUNNING &&
2522 		    starttime != time_uptime) {
2523 			if (first_printf) {
2524 				printf("\nSyncing disks, vnodes remaining... ");
2525 				first_printf = 0;
2526 			}
2527 			printf("%d ", net_worklist_len);
2528 		}
2529 		starttime = time_uptime;
2530 
2531 		/*
2532 		 * Push files whose dirty time has expired.  Be careful
2533 		 * of interrupt race on slp queue.
2534 		 *
2535 		 * Skip over empty worklist slots when shutting down.
2536 		 */
2537 		do {
2538 			slp = &syncer_workitem_pending[syncer_delayno];
2539 			syncer_delayno += 1;
2540 			if (syncer_delayno == syncer_maxdelay)
2541 				syncer_delayno = 0;
2542 			next = &syncer_workitem_pending[syncer_delayno];
2543 			/*
2544 			 * If the worklist has wrapped since the
2545 			 * it was emptied of all but syncer vnodes,
2546 			 * switch to the FINAL_DELAY state and run
2547 			 * for one more second.
2548 			 */
2549 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2550 			    net_worklist_len == 0 &&
2551 			    last_work_seen == syncer_delayno) {
2552 				syncer_state = SYNCER_FINAL_DELAY;
2553 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2554 			}
2555 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2556 		    syncer_worklist_len > 0);
2557 
2558 		/*
2559 		 * Keep track of the last time there was anything
2560 		 * on the worklist other than syncer vnodes.
2561 		 * Return to the SHUTTING_DOWN state if any
2562 		 * new work appears.
2563 		 */
2564 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2565 			last_work_seen = syncer_delayno;
2566 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2567 			syncer_state = SYNCER_SHUTTING_DOWN;
2568 		while (!LIST_EMPTY(slp)) {
2569 			error = sync_vnode(slp, &bo, td);
2570 			if (error == 1) {
2571 				LIST_REMOVE(bo, bo_synclist);
2572 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2573 				continue;
2574 			}
2575 
2576 			if (first_printf == 0) {
2577 				/*
2578 				 * Drop the sync mutex, because some watchdog
2579 				 * drivers need to sleep while patting
2580 				 */
2581 				mtx_unlock(&sync_mtx);
2582 				wdog_kern_pat(WD_LASTVAL);
2583 				mtx_lock(&sync_mtx);
2584 			}
2585 
2586 		}
2587 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2588 			syncer_final_iter--;
2589 		/*
2590 		 * The variable rushjob allows the kernel to speed up the
2591 		 * processing of the filesystem syncer process. A rushjob
2592 		 * value of N tells the filesystem syncer to process the next
2593 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2594 		 * is used by the soft update code to speed up the filesystem
2595 		 * syncer process when the incore state is getting so far
2596 		 * ahead of the disk that the kernel memory pool is being
2597 		 * threatened with exhaustion.
2598 		 */
2599 		if (rushjob > 0) {
2600 			rushjob -= 1;
2601 			continue;
2602 		}
2603 		/*
2604 		 * Just sleep for a short period of time between
2605 		 * iterations when shutting down to allow some I/O
2606 		 * to happen.
2607 		 *
2608 		 * If it has taken us less than a second to process the
2609 		 * current work, then wait. Otherwise start right over
2610 		 * again. We can still lose time if any single round
2611 		 * takes more than two seconds, but it does not really
2612 		 * matter as we are just trying to generally pace the
2613 		 * filesystem activity.
2614 		 */
2615 		if (syncer_state != SYNCER_RUNNING ||
2616 		    time_uptime == starttime) {
2617 			thread_lock(td);
2618 			sched_prio(td, PPAUSE);
2619 			thread_unlock(td);
2620 		}
2621 		if (syncer_state != SYNCER_RUNNING)
2622 			cv_timedwait(&sync_wakeup, &sync_mtx,
2623 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2624 		else if (time_uptime == starttime)
2625 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2626 	}
2627 }
2628 
2629 /*
2630  * Request the syncer daemon to speed up its work.
2631  * We never push it to speed up more than half of its
2632  * normal turn time, otherwise it could take over the cpu.
2633  */
2634 int
2635 speedup_syncer(void)
2636 {
2637 	int ret = 0;
2638 
2639 	mtx_lock(&sync_mtx);
2640 	if (rushjob < syncdelay / 2) {
2641 		rushjob += 1;
2642 		stat_rush_requests += 1;
2643 		ret = 1;
2644 	}
2645 	mtx_unlock(&sync_mtx);
2646 	cv_broadcast(&sync_wakeup);
2647 	return (ret);
2648 }
2649 
2650 /*
2651  * Tell the syncer to speed up its work and run though its work
2652  * list several times, then tell it to shut down.
2653  */
2654 static void
2655 syncer_shutdown(void *arg, int howto)
2656 {
2657 
2658 	if (howto & RB_NOSYNC)
2659 		return;
2660 	mtx_lock(&sync_mtx);
2661 	syncer_state = SYNCER_SHUTTING_DOWN;
2662 	rushjob = 0;
2663 	mtx_unlock(&sync_mtx);
2664 	cv_broadcast(&sync_wakeup);
2665 	kproc_shutdown(arg, howto);
2666 }
2667 
2668 void
2669 syncer_suspend(void)
2670 {
2671 
2672 	syncer_shutdown(updateproc, 0);
2673 }
2674 
2675 void
2676 syncer_resume(void)
2677 {
2678 
2679 	mtx_lock(&sync_mtx);
2680 	first_printf = 1;
2681 	syncer_state = SYNCER_RUNNING;
2682 	mtx_unlock(&sync_mtx);
2683 	cv_broadcast(&sync_wakeup);
2684 	kproc_resume(updateproc);
2685 }
2686 
2687 /*
2688  * Reassign a buffer from one vnode to another.
2689  * Used to assign file specific control information
2690  * (indirect blocks) to the vnode to which they belong.
2691  */
2692 void
2693 reassignbuf(struct buf *bp)
2694 {
2695 	struct vnode *vp;
2696 	struct bufobj *bo;
2697 	int delay;
2698 #ifdef INVARIANTS
2699 	struct bufv *bv;
2700 #endif
2701 
2702 	vp = bp->b_vp;
2703 	bo = bp->b_bufobj;
2704 	++reassignbufcalls;
2705 
2706 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2707 	    bp, bp->b_vp, bp->b_flags);
2708 	/*
2709 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2710 	 * is not fully linked in.
2711 	 */
2712 	if (bp->b_flags & B_PAGING)
2713 		panic("cannot reassign paging buffer");
2714 
2715 	/*
2716 	 * Delete from old vnode list, if on one.
2717 	 */
2718 	BO_LOCK(bo);
2719 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2720 		buf_vlist_remove(bp);
2721 	else
2722 		panic("reassignbuf: Buffer %p not on queue.", bp);
2723 	/*
2724 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2725 	 * of clean buffers.
2726 	 */
2727 	if (bp->b_flags & B_DELWRI) {
2728 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2729 			switch (vp->v_type) {
2730 			case VDIR:
2731 				delay = dirdelay;
2732 				break;
2733 			case VCHR:
2734 				delay = metadelay;
2735 				break;
2736 			default:
2737 				delay = filedelay;
2738 			}
2739 			vn_syncer_add_to_worklist(bo, delay);
2740 		}
2741 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2742 	} else {
2743 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2744 
2745 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2746 			mtx_lock(&sync_mtx);
2747 			LIST_REMOVE(bo, bo_synclist);
2748 			syncer_worklist_len--;
2749 			mtx_unlock(&sync_mtx);
2750 			bo->bo_flag &= ~BO_ONWORKLST;
2751 		}
2752 	}
2753 #ifdef INVARIANTS
2754 	bv = &bo->bo_clean;
2755 	bp = TAILQ_FIRST(&bv->bv_hd);
2756 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2757 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2758 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2759 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2760 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2761 	bv = &bo->bo_dirty;
2762 	bp = TAILQ_FIRST(&bv->bv_hd);
2763 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2764 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2765 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2766 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2767 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2768 #endif
2769 	BO_UNLOCK(bo);
2770 }
2771 
2772 static void
2773 v_init_counters(struct vnode *vp)
2774 {
2775 
2776 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2777 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2778 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2779 
2780 	refcount_init(&vp->v_holdcnt, 1);
2781 	refcount_init(&vp->v_usecount, 1);
2782 }
2783 
2784 /*
2785  * Increment si_usecount of the associated device, if any.
2786  */
2787 static void
2788 v_incr_devcount(struct vnode *vp)
2789 {
2790 
2791 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2792 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2793 		dev_lock();
2794 		vp->v_rdev->si_usecount++;
2795 		dev_unlock();
2796 	}
2797 }
2798 
2799 /*
2800  * Decrement si_usecount of the associated device, if any.
2801  */
2802 static void
2803 v_decr_devcount(struct vnode *vp)
2804 {
2805 
2806 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2807 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2808 		dev_lock();
2809 		vp->v_rdev->si_usecount--;
2810 		dev_unlock();
2811 	}
2812 }
2813 
2814 /*
2815  * Grab a particular vnode from the free list, increment its
2816  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2817  * is being destroyed.  Only callers who specify LK_RETRY will
2818  * see doomed vnodes.  If inactive processing was delayed in
2819  * vput try to do it here.
2820  *
2821  * usecount is manipulated using atomics without holding any locks.
2822  *
2823  * holdcnt can be manipulated using atomics without holding any locks,
2824  * except when transitioning 1<->0, in which case the interlock is held.
2825  */
2826 enum vgetstate
2827 vget_prep(struct vnode *vp)
2828 {
2829 	enum vgetstate vs;
2830 
2831 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2832 		vs = VGET_USECOUNT;
2833 	} else {
2834 		vhold(vp);
2835 		vs = VGET_HOLDCNT;
2836 	}
2837 	return (vs);
2838 }
2839 
2840 int
2841 vget(struct vnode *vp, int flags, struct thread *td)
2842 {
2843 	enum vgetstate vs;
2844 
2845 	MPASS(td == curthread);
2846 
2847 	vs = vget_prep(vp);
2848 	return (vget_finish(vp, flags, vs));
2849 }
2850 
2851 static int __noinline
2852 vget_finish_vchr(struct vnode *vp)
2853 {
2854 
2855 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2856 
2857 	/*
2858 	 * See the comment in vget_finish before usecount bump.
2859 	 */
2860 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2861 #ifdef INVARIANTS
2862 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2863 		VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2864 #else
2865 		refcount_release(&vp->v_holdcnt);
2866 #endif
2867 		return (0);
2868 	}
2869 
2870 	VI_LOCK(vp);
2871 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2872 #ifdef INVARIANTS
2873 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2874 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2875 #else
2876 		refcount_release(&vp->v_holdcnt);
2877 #endif
2878 		VI_UNLOCK(vp);
2879 		return (0);
2880 	}
2881 	v_incr_devcount(vp);
2882 	refcount_acquire(&vp->v_usecount);
2883 	VI_UNLOCK(vp);
2884 	return (0);
2885 }
2886 
2887 int
2888 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2889 {
2890 	int error, old;
2891 
2892 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2893 	    ("%s: invalid lock operation", __func__));
2894 
2895 	if ((flags & LK_INTERLOCK) != 0)
2896 		ASSERT_VI_LOCKED(vp, __func__);
2897 	else
2898 		ASSERT_VI_UNLOCKED(vp, __func__);
2899 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2900 	if (vs == VGET_USECOUNT) {
2901 		VNASSERT(vp->v_usecount > 0, vp,
2902 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2903 		    __func__));
2904 	}
2905 
2906 	error = vn_lock(vp, flags);
2907 	if (__predict_false(error != 0)) {
2908 		if (vs == VGET_USECOUNT)
2909 			vrele(vp);
2910 		else
2911 			vdrop(vp);
2912 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2913 		    vp);
2914 		return (error);
2915 	}
2916 
2917 	if (vs == VGET_USECOUNT) {
2918 		return (0);
2919 	}
2920 
2921 	if (__predict_false(vp->v_type == VCHR))
2922 		return (vget_finish_vchr(vp));
2923 
2924 	/*
2925 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2926 	 * the vnode around. Otherwise someone else lended their hold count and
2927 	 * we have to drop ours.
2928 	 */
2929 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2930 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2931 	if (old != 0) {
2932 #ifdef INVARIANTS
2933 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2934 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2935 #else
2936 		refcount_release(&vp->v_holdcnt);
2937 #endif
2938 	}
2939 	return (0);
2940 }
2941 
2942 /*
2943  * Increase the reference (use) and hold count of a vnode.
2944  * This will also remove the vnode from the free list if it is presently free.
2945  */
2946 static void __noinline
2947 vref_vchr(struct vnode *vp, bool interlock)
2948 {
2949 
2950 	/*
2951 	 * See the comment in vget_finish before usecount bump.
2952 	 */
2953 	if (!interlock) {
2954 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2955 			VNODE_REFCOUNT_FENCE_ACQ();
2956 			VNASSERT(vp->v_holdcnt > 0, vp,
2957 			    ("%s: active vnode not held", __func__));
2958 			return;
2959 		}
2960 		VI_LOCK(vp);
2961 		/*
2962 		 * By the time we get here the vnode might have been doomed, at
2963 		 * which point the 0->1 use count transition is no longer
2964 		 * protected by the interlock. Since it can't bounce back to
2965 		 * VCHR and requires vref semantics, punt it back
2966 		 */
2967 		if (__predict_false(vp->v_type == VBAD)) {
2968 			VI_UNLOCK(vp);
2969 			vref(vp);
2970 			return;
2971 		}
2972 	}
2973 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2974 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2975 		VNODE_REFCOUNT_FENCE_ACQ();
2976 		VNASSERT(vp->v_holdcnt > 0, vp,
2977 		    ("%s: active vnode not held", __func__));
2978 		if (!interlock)
2979 			VI_UNLOCK(vp);
2980 		return;
2981 	}
2982 	vhold(vp);
2983 	v_incr_devcount(vp);
2984 	refcount_acquire(&vp->v_usecount);
2985 	if (!interlock)
2986 		VI_UNLOCK(vp);
2987 	return;
2988 }
2989 
2990 void
2991 vref(struct vnode *vp)
2992 {
2993 	int old;
2994 
2995 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2996 	if (__predict_false(vp->v_type == VCHR)) {
2997 		 vref_vchr(vp, false);
2998 		 return;
2999 	}
3000 
3001 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3002 		VNODE_REFCOUNT_FENCE_ACQ();
3003 		VNASSERT(vp->v_holdcnt > 0, vp,
3004 		    ("%s: active vnode not held", __func__));
3005 		return;
3006 	}
3007 	vhold(vp);
3008 	/*
3009 	 * See the comment in vget_finish.
3010 	 */
3011 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3012 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3013 	if (old != 0) {
3014 #ifdef INVARIANTS
3015 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3016 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3017 #else
3018 		refcount_release(&vp->v_holdcnt);
3019 #endif
3020 	}
3021 }
3022 
3023 void
3024 vrefl(struct vnode *vp)
3025 {
3026 
3027 	ASSERT_VI_LOCKED(vp, __func__);
3028 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3029 	if (__predict_false(vp->v_type == VCHR)) {
3030 		vref_vchr(vp, true);
3031 		return;
3032 	}
3033 	vref(vp);
3034 }
3035 
3036 void
3037 vrefact(struct vnode *vp)
3038 {
3039 
3040 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3041 #ifdef INVARIANTS
3042 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3043 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3044 #else
3045 	refcount_acquire(&vp->v_usecount);
3046 #endif
3047 }
3048 
3049 void
3050 vrefactn(struct vnode *vp, u_int n)
3051 {
3052 
3053 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3054 #ifdef INVARIANTS
3055 	int old = atomic_fetchadd_int(&vp->v_usecount, n);
3056 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3057 #else
3058 	atomic_add_int(&vp->v_usecount, n);
3059 #endif
3060 }
3061 
3062 /*
3063  * Return reference count of a vnode.
3064  *
3065  * The results of this call are only guaranteed when some mechanism is used to
3066  * stop other processes from gaining references to the vnode.  This may be the
3067  * case if the caller holds the only reference.  This is also useful when stale
3068  * data is acceptable as race conditions may be accounted for by some other
3069  * means.
3070  */
3071 int
3072 vrefcnt(struct vnode *vp)
3073 {
3074 
3075 	return (vp->v_usecount);
3076 }
3077 
3078 void
3079 vlazy(struct vnode *vp)
3080 {
3081 	struct mount *mp;
3082 
3083 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3084 
3085 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3086 		return;
3087 	/*
3088 	 * We may get here for inactive routines after the vnode got doomed.
3089 	 */
3090 	if (VN_IS_DOOMED(vp))
3091 		return;
3092 	mp = vp->v_mount;
3093 	mtx_lock(&mp->mnt_listmtx);
3094 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3095 		vp->v_mflag |= VMP_LAZYLIST;
3096 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3097 		mp->mnt_lazyvnodelistsize++;
3098 	}
3099 	mtx_unlock(&mp->mnt_listmtx);
3100 }
3101 
3102 /*
3103  * This routine is only meant to be called from vgonel prior to dooming
3104  * the vnode.
3105  */
3106 static void
3107 vunlazy_gone(struct vnode *vp)
3108 {
3109 	struct mount *mp;
3110 
3111 	ASSERT_VOP_ELOCKED(vp, __func__);
3112 	ASSERT_VI_LOCKED(vp, __func__);
3113 	VNPASS(!VN_IS_DOOMED(vp), vp);
3114 
3115 	if (vp->v_mflag & VMP_LAZYLIST) {
3116 		mp = vp->v_mount;
3117 		mtx_lock(&mp->mnt_listmtx);
3118 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3119 		vp->v_mflag &= ~VMP_LAZYLIST;
3120 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3121 		mp->mnt_lazyvnodelistsize--;
3122 		mtx_unlock(&mp->mnt_listmtx);
3123 	}
3124 }
3125 
3126 static void
3127 vdefer_inactive(struct vnode *vp)
3128 {
3129 
3130 	ASSERT_VI_LOCKED(vp, __func__);
3131 	VNASSERT(vp->v_holdcnt > 0, vp,
3132 	    ("%s: vnode without hold count", __func__));
3133 	if (VN_IS_DOOMED(vp)) {
3134 		vdropl(vp);
3135 		return;
3136 	}
3137 	if (vp->v_iflag & VI_DEFINACT) {
3138 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3139 		vdropl(vp);
3140 		return;
3141 	}
3142 	if (vp->v_usecount > 0) {
3143 		vp->v_iflag &= ~VI_OWEINACT;
3144 		vdropl(vp);
3145 		return;
3146 	}
3147 	vlazy(vp);
3148 	vp->v_iflag |= VI_DEFINACT;
3149 	VI_UNLOCK(vp);
3150 	counter_u64_add(deferred_inact, 1);
3151 }
3152 
3153 static void
3154 vdefer_inactive_unlocked(struct vnode *vp)
3155 {
3156 
3157 	VI_LOCK(vp);
3158 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3159 		vdropl(vp);
3160 		return;
3161 	}
3162 	vdefer_inactive(vp);
3163 }
3164 
3165 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
3166 
3167 /*
3168  * Decrement the use and hold counts for a vnode.
3169  *
3170  * See an explanation near vget() as to why atomic operation is safe.
3171  *
3172  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3173  * on the lock being held all the way until VOP_INACTIVE. This in particular
3174  * happens with UFS which adds half-constructed vnodes to the hash, where they
3175  * can be found by other code.
3176  */
3177 static void
3178 vputx(struct vnode *vp, enum vputx_op func)
3179 {
3180 	int error;
3181 
3182 	KASSERT(vp != NULL, ("vputx: null vp"));
3183 	if (func == VPUTX_VUNREF)
3184 		ASSERT_VOP_LOCKED(vp, "vunref");
3185 	else if (func == VPUTX_VPUT)
3186 		ASSERT_VOP_LOCKED(vp, "vput");
3187 	ASSERT_VI_UNLOCKED(vp, __func__);
3188 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
3189 	    ("%s: wrong ref counts", __func__));
3190 
3191 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3192 
3193 	/*
3194 	 * We want to hold the vnode until the inactive finishes to
3195 	 * prevent vgone() races.  We drop the use count here and the
3196 	 * hold count below when we're done.
3197 	 *
3198 	 * If we release the last usecount we take ownership of the hold
3199 	 * count which provides liveness of the vnode, in which case we
3200 	 * have to vdrop.
3201 	 */
3202 	if (!refcount_release(&vp->v_usecount)) {
3203 		if (func == VPUTX_VPUT)
3204 			VOP_UNLOCK(vp);
3205 		return;
3206 	}
3207 	VI_LOCK(vp);
3208 	v_decr_devcount(vp);
3209 	/*
3210 	 * By the time we got here someone else might have transitioned
3211 	 * the count back to > 0.
3212 	 */
3213 	if (vp->v_usecount > 0 || vp->v_iflag & VI_DOINGINACT)
3214 		goto out;
3215 
3216 	/*
3217 	 * Check if the fs wants to perform inactive processing. Note we
3218 	 * may be only holding the interlock, in which case it is possible
3219 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3220 	 * Since vgone performs inactive on its own there is nothing to do
3221 	 * here but to drop our hold count.
3222 	 */
3223 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3224 	    VOP_NEED_INACTIVE(vp) == 0)
3225 		goto out;
3226 
3227 	/*
3228 	 * We must call VOP_INACTIVE with the node locked. Mark
3229 	 * as VI_DOINGINACT to avoid recursion.
3230 	 */
3231 	vp->v_iflag |= VI_OWEINACT;
3232 	switch (func) {
3233 	case VPUTX_VRELE:
3234 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3235 		VI_LOCK(vp);
3236 		break;
3237 	case VPUTX_VPUT:
3238 		error = 0;
3239 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3240 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3241 			    LK_NOWAIT);
3242 			VI_LOCK(vp);
3243 		}
3244 		break;
3245 	case VPUTX_VUNREF:
3246 		error = 0;
3247 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3248 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3249 			VI_LOCK(vp);
3250 		}
3251 		break;
3252 	}
3253 	if (error == 0) {
3254 		vinactive(vp);
3255 		if (func != VPUTX_VUNREF)
3256 			VOP_UNLOCK(vp);
3257 		vdropl(vp);
3258 	} else {
3259 		vdefer_inactive(vp);
3260 	}
3261 	return;
3262 out:
3263 	if (func == VPUTX_VPUT)
3264 		VOP_UNLOCK(vp);
3265 	vdropl(vp);
3266 }
3267 
3268 /*
3269  * Vnode put/release.
3270  * If count drops to zero, call inactive routine and return to freelist.
3271  */
3272 void
3273 vrele(struct vnode *vp)
3274 {
3275 
3276 	vputx(vp, VPUTX_VRELE);
3277 }
3278 
3279 /*
3280  * Release an already locked vnode.  This give the same effects as
3281  * unlock+vrele(), but takes less time and avoids releasing and
3282  * re-aquiring the lock (as vrele() acquires the lock internally.)
3283  */
3284 void
3285 vput(struct vnode *vp)
3286 {
3287 
3288 	vputx(vp, VPUTX_VPUT);
3289 }
3290 
3291 /*
3292  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3293  */
3294 void
3295 vunref(struct vnode *vp)
3296 {
3297 
3298 	vputx(vp, VPUTX_VUNREF);
3299 }
3300 
3301 void
3302 vhold(struct vnode *vp)
3303 {
3304 	struct vdbatch *vd;
3305 	int old;
3306 
3307 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3308 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3309 	VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3310 	if (old != 0)
3311 		return;
3312 	critical_enter();
3313 	vd = DPCPU_PTR(vd);
3314 	vd->freevnodes--;
3315 	critical_exit();
3316 }
3317 
3318 void
3319 vholdl(struct vnode *vp)
3320 {
3321 
3322 	ASSERT_VI_LOCKED(vp, __func__);
3323 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3324 	vhold(vp);
3325 }
3326 
3327 void
3328 vholdnz(struct vnode *vp)
3329 {
3330 
3331 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3332 #ifdef INVARIANTS
3333 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3334 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3335 #else
3336 	atomic_add_int(&vp->v_holdcnt, 1);
3337 #endif
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 	critical_enter();
3377 	vd = DPCPU_PTR(vd);
3378 	vd->freevnodes++;
3379 	if (vp->v_dbatchcpu != NOCPU) {
3380 		VI_UNLOCK(vp);
3381 		critical_exit();
3382 		return;
3383 	}
3384 
3385 	sched_pin();
3386 	critical_exit();
3387 	mtx_lock(&vd->lock);
3388 	MPASS(vd->index < VDBATCH_SIZE);
3389 	MPASS(vd->tab[vd->index] == NULL);
3390 	/*
3391 	 * A hack: we depend on being pinned so that we know what to put in
3392 	 * ->v_dbatchcpu.
3393 	 */
3394 	vp->v_dbatchcpu = curcpu;
3395 	vd->tab[vd->index] = vp;
3396 	vd->index++;
3397 	VI_UNLOCK(vp);
3398 	if (vd->index == VDBATCH_SIZE)
3399 		vdbatch_process(vd);
3400 	mtx_unlock(&vd->lock);
3401 	sched_unpin();
3402 }
3403 
3404 /*
3405  * This routine must only be called for vnodes which are about to be
3406  * deallocated. Supporting dequeue for arbitrary vndoes would require
3407  * validating that the locked batch matches.
3408  */
3409 static void
3410 vdbatch_dequeue(struct vnode *vp)
3411 {
3412 	struct vdbatch *vd;
3413 	int i;
3414 	short cpu;
3415 
3416 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3417 	    ("%s: called for a used vnode\n", __func__));
3418 
3419 	cpu = vp->v_dbatchcpu;
3420 	if (cpu == NOCPU)
3421 		return;
3422 
3423 	vd = DPCPU_ID_PTR(cpu, vd);
3424 	mtx_lock(&vd->lock);
3425 	for (i = 0; i < vd->index; i++) {
3426 		if (vd->tab[i] != vp)
3427 			continue;
3428 		vp->v_dbatchcpu = NOCPU;
3429 		vd->index--;
3430 		vd->tab[i] = vd->tab[vd->index];
3431 		vd->tab[vd->index] = NULL;
3432 		break;
3433 	}
3434 	mtx_unlock(&vd->lock);
3435 	/*
3436 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3437 	 */
3438 	MPASS(vp->v_dbatchcpu == NOCPU);
3439 }
3440 
3441 /*
3442  * Drop the hold count of the vnode.  If this is the last reference to
3443  * the vnode we place it on the free list unless it has been vgone'd
3444  * (marked VIRF_DOOMED) in which case we will free it.
3445  *
3446  * Because the vnode vm object keeps a hold reference on the vnode if
3447  * there is at least one resident non-cached page, the vnode cannot
3448  * leave the active list without the page cleanup done.
3449  */
3450 static void
3451 vdrop_deactivate(struct vnode *vp)
3452 {
3453 	struct mount *mp;
3454 
3455 	ASSERT_VI_LOCKED(vp, __func__);
3456 	/*
3457 	 * Mark a vnode as free: remove it from its active list
3458 	 * and put it up for recycling on the freelist.
3459 	 */
3460 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3461 	    ("vdrop: returning doomed vnode"));
3462 	VNASSERT(vp->v_op != NULL, vp,
3463 	    ("vdrop: vnode already reclaimed."));
3464 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3465 	    ("vnode with VI_OWEINACT set"));
3466 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3467 	    ("vnode with VI_DEFINACT set"));
3468 	if (vp->v_mflag & VMP_LAZYLIST) {
3469 		mp = vp->v_mount;
3470 		mtx_lock(&mp->mnt_listmtx);
3471 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3472 		/*
3473 		 * Don't remove the vnode from the lazy list if another thread
3474 		 * has increased the hold count. It may have re-enqueued the
3475 		 * vnode to the lazy list and is now responsible for its
3476 		 * removal.
3477 		 */
3478 		if (vp->v_holdcnt == 0) {
3479 			vp->v_mflag &= ~VMP_LAZYLIST;
3480 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3481 			mp->mnt_lazyvnodelistsize--;
3482 		}
3483 		mtx_unlock(&mp->mnt_listmtx);
3484 	}
3485 	vdbatch_enqueue(vp);
3486 }
3487 
3488 void
3489 vdrop(struct vnode *vp)
3490 {
3491 
3492 	ASSERT_VI_UNLOCKED(vp, __func__);
3493 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3494 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3495 		return;
3496 	VI_LOCK(vp);
3497 	vdropl(vp);
3498 }
3499 
3500 void
3501 vdropl(struct vnode *vp)
3502 {
3503 
3504 	ASSERT_VI_LOCKED(vp, __func__);
3505 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3506 	if (!refcount_release(&vp->v_holdcnt)) {
3507 		VI_UNLOCK(vp);
3508 		return;
3509 	}
3510 	if (VN_IS_DOOMED(vp)) {
3511 		freevnode(vp);
3512 		return;
3513 	}
3514 	vdrop_deactivate(vp);
3515 }
3516 
3517 /*
3518  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3519  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3520  */
3521 static void
3522 vinactivef(struct vnode *vp)
3523 {
3524 	struct vm_object *obj;
3525 
3526 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3527 	ASSERT_VI_LOCKED(vp, "vinactive");
3528 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3529 	    ("vinactive: recursed on VI_DOINGINACT"));
3530 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3531 	vp->v_iflag |= VI_DOINGINACT;
3532 	vp->v_iflag &= ~VI_OWEINACT;
3533 	VI_UNLOCK(vp);
3534 	/*
3535 	 * Before moving off the active list, we must be sure that any
3536 	 * modified pages are converted into the vnode's dirty
3537 	 * buffers, since these will no longer be checked once the
3538 	 * vnode is on the inactive list.
3539 	 *
3540 	 * The write-out of the dirty pages is asynchronous.  At the
3541 	 * point that VOP_INACTIVE() is called, there could still be
3542 	 * pending I/O and dirty pages in the object.
3543 	 */
3544 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3545 	    vm_object_mightbedirty(obj)) {
3546 		VM_OBJECT_WLOCK(obj);
3547 		vm_object_page_clean(obj, 0, 0, 0);
3548 		VM_OBJECT_WUNLOCK(obj);
3549 	}
3550 	VOP_INACTIVE(vp, curthread);
3551 	VI_LOCK(vp);
3552 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3553 	    ("vinactive: lost VI_DOINGINACT"));
3554 	vp->v_iflag &= ~VI_DOINGINACT;
3555 }
3556 
3557 void
3558 vinactive(struct vnode *vp)
3559 {
3560 
3561 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3562 	ASSERT_VI_LOCKED(vp, "vinactive");
3563 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3564 
3565 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3566 		return;
3567 	if (vp->v_iflag & VI_DOINGINACT)
3568 		return;
3569 	if (vp->v_usecount > 0) {
3570 		vp->v_iflag &= ~VI_OWEINACT;
3571 		return;
3572 	}
3573 	vinactivef(vp);
3574 }
3575 
3576 /*
3577  * Remove any vnodes in the vnode table belonging to mount point mp.
3578  *
3579  * If FORCECLOSE is not specified, there should not be any active ones,
3580  * return error if any are found (nb: this is a user error, not a
3581  * system error). If FORCECLOSE is specified, detach any active vnodes
3582  * that are found.
3583  *
3584  * If WRITECLOSE is set, only flush out regular file vnodes open for
3585  * writing.
3586  *
3587  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3588  *
3589  * `rootrefs' specifies the base reference count for the root vnode
3590  * of this filesystem. The root vnode is considered busy if its
3591  * v_usecount exceeds this value. On a successful return, vflush(, td)
3592  * will call vrele() on the root vnode exactly rootrefs times.
3593  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3594  * be zero.
3595  */
3596 #ifdef DIAGNOSTIC
3597 static int busyprt = 0;		/* print out busy vnodes */
3598 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3599 #endif
3600 
3601 int
3602 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3603 {
3604 	struct vnode *vp, *mvp, *rootvp = NULL;
3605 	struct vattr vattr;
3606 	int busy = 0, error;
3607 
3608 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3609 	    rootrefs, flags);
3610 	if (rootrefs > 0) {
3611 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3612 		    ("vflush: bad args"));
3613 		/*
3614 		 * Get the filesystem root vnode. We can vput() it
3615 		 * immediately, since with rootrefs > 0, it won't go away.
3616 		 */
3617 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3618 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3619 			    __func__, error);
3620 			return (error);
3621 		}
3622 		vput(rootvp);
3623 	}
3624 loop:
3625 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3626 		vholdl(vp);
3627 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3628 		if (error) {
3629 			vdrop(vp);
3630 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3631 			goto loop;
3632 		}
3633 		/*
3634 		 * Skip over a vnodes marked VV_SYSTEM.
3635 		 */
3636 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3637 			VOP_UNLOCK(vp);
3638 			vdrop(vp);
3639 			continue;
3640 		}
3641 		/*
3642 		 * If WRITECLOSE is set, flush out unlinked but still open
3643 		 * files (even if open only for reading) and regular file
3644 		 * vnodes open for writing.
3645 		 */
3646 		if (flags & WRITECLOSE) {
3647 			if (vp->v_object != NULL) {
3648 				VM_OBJECT_WLOCK(vp->v_object);
3649 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3650 				VM_OBJECT_WUNLOCK(vp->v_object);
3651 			}
3652 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3653 			if (error != 0) {
3654 				VOP_UNLOCK(vp);
3655 				vdrop(vp);
3656 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3657 				return (error);
3658 			}
3659 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3660 			VI_LOCK(vp);
3661 
3662 			if ((vp->v_type == VNON ||
3663 			    (error == 0 && vattr.va_nlink > 0)) &&
3664 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3665 				VOP_UNLOCK(vp);
3666 				vdropl(vp);
3667 				continue;
3668 			}
3669 		} else
3670 			VI_LOCK(vp);
3671 		/*
3672 		 * With v_usecount == 0, all we need to do is clear out the
3673 		 * vnode data structures and we are done.
3674 		 *
3675 		 * If FORCECLOSE is set, forcibly close the vnode.
3676 		 */
3677 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3678 			vgonel(vp);
3679 		} else {
3680 			busy++;
3681 #ifdef DIAGNOSTIC
3682 			if (busyprt)
3683 				vn_printf(vp, "vflush: busy vnode ");
3684 #endif
3685 		}
3686 		VOP_UNLOCK(vp);
3687 		vdropl(vp);
3688 	}
3689 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3690 		/*
3691 		 * If just the root vnode is busy, and if its refcount
3692 		 * is equal to `rootrefs', then go ahead and kill it.
3693 		 */
3694 		VI_LOCK(rootvp);
3695 		KASSERT(busy > 0, ("vflush: not busy"));
3696 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3697 		    ("vflush: usecount %d < rootrefs %d",
3698 		     rootvp->v_usecount, rootrefs));
3699 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3700 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3701 			vgone(rootvp);
3702 			VOP_UNLOCK(rootvp);
3703 			busy = 0;
3704 		} else
3705 			VI_UNLOCK(rootvp);
3706 	}
3707 	if (busy) {
3708 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3709 		    busy);
3710 		return (EBUSY);
3711 	}
3712 	for (; rootrefs > 0; rootrefs--)
3713 		vrele(rootvp);
3714 	return (0);
3715 }
3716 
3717 /*
3718  * Recycle an unused vnode to the front of the free list.
3719  */
3720 int
3721 vrecycle(struct vnode *vp)
3722 {
3723 	int recycled;
3724 
3725 	VI_LOCK(vp);
3726 	recycled = vrecyclel(vp);
3727 	VI_UNLOCK(vp);
3728 	return (recycled);
3729 }
3730 
3731 /*
3732  * vrecycle, with the vp interlock held.
3733  */
3734 int
3735 vrecyclel(struct vnode *vp)
3736 {
3737 	int recycled;
3738 
3739 	ASSERT_VOP_ELOCKED(vp, __func__);
3740 	ASSERT_VI_LOCKED(vp, __func__);
3741 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3742 	recycled = 0;
3743 	if (vp->v_usecount == 0) {
3744 		recycled = 1;
3745 		vgonel(vp);
3746 	}
3747 	return (recycled);
3748 }
3749 
3750 /*
3751  * Eliminate all activity associated with a vnode
3752  * in preparation for reuse.
3753  */
3754 void
3755 vgone(struct vnode *vp)
3756 {
3757 	VI_LOCK(vp);
3758 	vgonel(vp);
3759 	VI_UNLOCK(vp);
3760 }
3761 
3762 static void
3763 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3764     struct vnode *lowervp __unused)
3765 {
3766 }
3767 
3768 /*
3769  * Notify upper mounts about reclaimed or unlinked vnode.
3770  */
3771 void
3772 vfs_notify_upper(struct vnode *vp, int event)
3773 {
3774 	static struct vfsops vgonel_vfsops = {
3775 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3776 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3777 	};
3778 	struct mount *mp, *ump, *mmp;
3779 
3780 	mp = vp->v_mount;
3781 	if (mp == NULL)
3782 		return;
3783 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3784 		return;
3785 
3786 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3787 	mmp->mnt_op = &vgonel_vfsops;
3788 	mmp->mnt_kern_flag |= MNTK_MARKER;
3789 	MNT_ILOCK(mp);
3790 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3791 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3792 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3793 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3794 			continue;
3795 		}
3796 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3797 		MNT_IUNLOCK(mp);
3798 		switch (event) {
3799 		case VFS_NOTIFY_UPPER_RECLAIM:
3800 			VFS_RECLAIM_LOWERVP(ump, vp);
3801 			break;
3802 		case VFS_NOTIFY_UPPER_UNLINK:
3803 			VFS_UNLINK_LOWERVP(ump, vp);
3804 			break;
3805 		default:
3806 			KASSERT(0, ("invalid event %d", event));
3807 			break;
3808 		}
3809 		MNT_ILOCK(mp);
3810 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3811 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3812 	}
3813 	free(mmp, M_TEMP);
3814 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3815 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3816 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3817 		wakeup(&mp->mnt_uppers);
3818 	}
3819 	MNT_IUNLOCK(mp);
3820 }
3821 
3822 /*
3823  * vgone, with the vp interlock held.
3824  */
3825 static void
3826 vgonel(struct vnode *vp)
3827 {
3828 	struct thread *td;
3829 	struct mount *mp;
3830 	vm_object_t object;
3831 	bool active, oweinact;
3832 
3833 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3834 	ASSERT_VI_LOCKED(vp, "vgonel");
3835 	VNASSERT(vp->v_holdcnt, vp,
3836 	    ("vgonel: vp %p has no reference.", vp));
3837 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3838 	td = curthread;
3839 
3840 	/*
3841 	 * Don't vgonel if we're already doomed.
3842 	 */
3843 	if (vp->v_irflag & VIRF_DOOMED)
3844 		return;
3845 	vunlazy_gone(vp);
3846 	vp->v_irflag |= VIRF_DOOMED;
3847 
3848 	/*
3849 	 * Check to see if the vnode is in use.  If so, we have to call
3850 	 * VOP_CLOSE() and VOP_INACTIVE().
3851 	 */
3852 	active = vp->v_usecount > 0;
3853 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3854 	/*
3855 	 * If we need to do inactive VI_OWEINACT will be set.
3856 	 */
3857 	if (vp->v_iflag & VI_DEFINACT) {
3858 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3859 		vp->v_iflag &= ~VI_DEFINACT;
3860 		vdropl(vp);
3861 	} else {
3862 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3863 		VI_UNLOCK(vp);
3864 	}
3865 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3866 
3867 	/*
3868 	 * If purging an active vnode, it must be closed and
3869 	 * deactivated before being reclaimed.
3870 	 */
3871 	if (active)
3872 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3873 	if (oweinact || active) {
3874 		VI_LOCK(vp);
3875 		vinactivef(vp);
3876 		VI_UNLOCK(vp);
3877 	}
3878 	if (vp->v_type == VSOCK)
3879 		vfs_unp_reclaim(vp);
3880 
3881 	/*
3882 	 * Clean out any buffers associated with the vnode.
3883 	 * If the flush fails, just toss the buffers.
3884 	 */
3885 	mp = NULL;
3886 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3887 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3888 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3889 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3890 			;
3891 	}
3892 
3893 	BO_LOCK(&vp->v_bufobj);
3894 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3895 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3896 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3897 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3898 	    ("vp %p bufobj not invalidated", vp));
3899 
3900 	/*
3901 	 * For VMIO bufobj, BO_DEAD is set later, or in
3902 	 * vm_object_terminate() after the object's page queue is
3903 	 * flushed.
3904 	 */
3905 	object = vp->v_bufobj.bo_object;
3906 	if (object == NULL)
3907 		vp->v_bufobj.bo_flag |= BO_DEAD;
3908 	BO_UNLOCK(&vp->v_bufobj);
3909 
3910 	/*
3911 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3912 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3913 	 * should not touch the object borrowed from the lower vnode
3914 	 * (the handle check).
3915 	 */
3916 	if (object != NULL && object->type == OBJT_VNODE &&
3917 	    object->handle == vp)
3918 		vnode_destroy_vobject(vp);
3919 
3920 	/*
3921 	 * Reclaim the vnode.
3922 	 */
3923 	if (VOP_RECLAIM(vp, td))
3924 		panic("vgone: cannot reclaim");
3925 	if (mp != NULL)
3926 		vn_finished_secondary_write(mp);
3927 	VNASSERT(vp->v_object == NULL, vp,
3928 	    ("vop_reclaim left v_object vp=%p", vp));
3929 	/*
3930 	 * Clear the advisory locks and wake up waiting threads.
3931 	 */
3932 	(void)VOP_ADVLOCKPURGE(vp);
3933 	vp->v_lockf = NULL;
3934 	/*
3935 	 * Delete from old mount point vnode list.
3936 	 */
3937 	delmntque(vp);
3938 	cache_purge(vp);
3939 	/*
3940 	 * Done with purge, reset to the standard lock and invalidate
3941 	 * the vnode.
3942 	 */
3943 	VI_LOCK(vp);
3944 	vp->v_vnlock = &vp->v_lock;
3945 	vp->v_op = &dead_vnodeops;
3946 	vp->v_type = VBAD;
3947 }
3948 
3949 /*
3950  * Calculate the total number of references to a special device.
3951  */
3952 int
3953 vcount(struct vnode *vp)
3954 {
3955 	int count;
3956 
3957 	dev_lock();
3958 	count = vp->v_rdev->si_usecount;
3959 	dev_unlock();
3960 	return (count);
3961 }
3962 
3963 /*
3964  * Print out a description of a vnode.
3965  */
3966 static char *typename[] =
3967 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3968  "VMARKER"};
3969 
3970 void
3971 vn_printf(struct vnode *vp, const char *fmt, ...)
3972 {
3973 	va_list ap;
3974 	char buf[256], buf2[16];
3975 	u_long flags;
3976 
3977 	va_start(ap, fmt);
3978 	vprintf(fmt, ap);
3979 	va_end(ap);
3980 	printf("%p: ", (void *)vp);
3981 	printf("type %s\n", typename[vp->v_type]);
3982 	printf("    usecount %d, writecount %d, refcount %d",
3983 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3984 	switch (vp->v_type) {
3985 	case VDIR:
3986 		printf(" mountedhere %p\n", vp->v_mountedhere);
3987 		break;
3988 	case VCHR:
3989 		printf(" rdev %p\n", vp->v_rdev);
3990 		break;
3991 	case VSOCK:
3992 		printf(" socket %p\n", vp->v_unpcb);
3993 		break;
3994 	case VFIFO:
3995 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3996 		break;
3997 	default:
3998 		printf("\n");
3999 		break;
4000 	}
4001 	buf[0] = '\0';
4002 	buf[1] = '\0';
4003 	if (vp->v_irflag & VIRF_DOOMED)
4004 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4005 	flags = vp->v_irflag & ~(VIRF_DOOMED);
4006 	if (flags != 0) {
4007 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4008 		strlcat(buf, buf2, sizeof(buf));
4009 	}
4010 	if (vp->v_vflag & VV_ROOT)
4011 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4012 	if (vp->v_vflag & VV_ISTTY)
4013 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4014 	if (vp->v_vflag & VV_NOSYNC)
4015 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4016 	if (vp->v_vflag & VV_ETERNALDEV)
4017 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4018 	if (vp->v_vflag & VV_CACHEDLABEL)
4019 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4020 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4021 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4022 	if (vp->v_vflag & VV_COPYONWRITE)
4023 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4024 	if (vp->v_vflag & VV_SYSTEM)
4025 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4026 	if (vp->v_vflag & VV_PROCDEP)
4027 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4028 	if (vp->v_vflag & VV_NOKNOTE)
4029 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4030 	if (vp->v_vflag & VV_DELETED)
4031 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4032 	if (vp->v_vflag & VV_MD)
4033 		strlcat(buf, "|VV_MD", sizeof(buf));
4034 	if (vp->v_vflag & VV_FORCEINSMQ)
4035 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4036 	if (vp->v_vflag & VV_READLINK)
4037 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4038 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4039 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4040 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4041 	if (flags != 0) {
4042 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4043 		strlcat(buf, buf2, sizeof(buf));
4044 	}
4045 	if (vp->v_iflag & VI_TEXT_REF)
4046 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4047 	if (vp->v_iflag & VI_MOUNT)
4048 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4049 	if (vp->v_iflag & VI_DOINGINACT)
4050 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4051 	if (vp->v_iflag & VI_OWEINACT)
4052 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4053 	if (vp->v_iflag & VI_DEFINACT)
4054 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4055 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4056 	    VI_OWEINACT | VI_DEFINACT);
4057 	if (flags != 0) {
4058 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4059 		strlcat(buf, buf2, sizeof(buf));
4060 	}
4061 	if (vp->v_mflag & VMP_LAZYLIST)
4062 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4063 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4064 	if (flags != 0) {
4065 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4066 		strlcat(buf, buf2, sizeof(buf));
4067 	}
4068 	printf("    flags (%s)\n", buf + 1);
4069 	if (mtx_owned(VI_MTX(vp)))
4070 		printf(" VI_LOCKed");
4071 	if (vp->v_object != NULL)
4072 		printf("    v_object %p ref %d pages %d "
4073 		    "cleanbuf %d dirtybuf %d\n",
4074 		    vp->v_object, vp->v_object->ref_count,
4075 		    vp->v_object->resident_page_count,
4076 		    vp->v_bufobj.bo_clean.bv_cnt,
4077 		    vp->v_bufobj.bo_dirty.bv_cnt);
4078 	printf("    ");
4079 	lockmgr_printinfo(vp->v_vnlock);
4080 	if (vp->v_data != NULL)
4081 		VOP_PRINT(vp);
4082 }
4083 
4084 #ifdef DDB
4085 /*
4086  * List all of the locked vnodes in the system.
4087  * Called when debugging the kernel.
4088  */
4089 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4090 {
4091 	struct mount *mp;
4092 	struct vnode *vp;
4093 
4094 	/*
4095 	 * Note: because this is DDB, we can't obey the locking semantics
4096 	 * for these structures, which means we could catch an inconsistent
4097 	 * state and dereference a nasty pointer.  Not much to be done
4098 	 * about that.
4099 	 */
4100 	db_printf("Locked vnodes\n");
4101 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4102 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4103 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4104 				vn_printf(vp, "vnode ");
4105 		}
4106 	}
4107 }
4108 
4109 /*
4110  * Show details about the given vnode.
4111  */
4112 DB_SHOW_COMMAND(vnode, db_show_vnode)
4113 {
4114 	struct vnode *vp;
4115 
4116 	if (!have_addr)
4117 		return;
4118 	vp = (struct vnode *)addr;
4119 	vn_printf(vp, "vnode ");
4120 }
4121 
4122 /*
4123  * Show details about the given mount point.
4124  */
4125 DB_SHOW_COMMAND(mount, db_show_mount)
4126 {
4127 	struct mount *mp;
4128 	struct vfsopt *opt;
4129 	struct statfs *sp;
4130 	struct vnode *vp;
4131 	char buf[512];
4132 	uint64_t mflags;
4133 	u_int flags;
4134 
4135 	if (!have_addr) {
4136 		/* No address given, print short info about all mount points. */
4137 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4138 			db_printf("%p %s on %s (%s)\n", mp,
4139 			    mp->mnt_stat.f_mntfromname,
4140 			    mp->mnt_stat.f_mntonname,
4141 			    mp->mnt_stat.f_fstypename);
4142 			if (db_pager_quit)
4143 				break;
4144 		}
4145 		db_printf("\nMore info: show mount <addr>\n");
4146 		return;
4147 	}
4148 
4149 	mp = (struct mount *)addr;
4150 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4151 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4152 
4153 	buf[0] = '\0';
4154 	mflags = mp->mnt_flag;
4155 #define	MNT_FLAG(flag)	do {						\
4156 	if (mflags & (flag)) {						\
4157 		if (buf[0] != '\0')					\
4158 			strlcat(buf, ", ", sizeof(buf));		\
4159 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4160 		mflags &= ~(flag);					\
4161 	}								\
4162 } while (0)
4163 	MNT_FLAG(MNT_RDONLY);
4164 	MNT_FLAG(MNT_SYNCHRONOUS);
4165 	MNT_FLAG(MNT_NOEXEC);
4166 	MNT_FLAG(MNT_NOSUID);
4167 	MNT_FLAG(MNT_NFS4ACLS);
4168 	MNT_FLAG(MNT_UNION);
4169 	MNT_FLAG(MNT_ASYNC);
4170 	MNT_FLAG(MNT_SUIDDIR);
4171 	MNT_FLAG(MNT_SOFTDEP);
4172 	MNT_FLAG(MNT_NOSYMFOLLOW);
4173 	MNT_FLAG(MNT_GJOURNAL);
4174 	MNT_FLAG(MNT_MULTILABEL);
4175 	MNT_FLAG(MNT_ACLS);
4176 	MNT_FLAG(MNT_NOATIME);
4177 	MNT_FLAG(MNT_NOCLUSTERR);
4178 	MNT_FLAG(MNT_NOCLUSTERW);
4179 	MNT_FLAG(MNT_SUJ);
4180 	MNT_FLAG(MNT_EXRDONLY);
4181 	MNT_FLAG(MNT_EXPORTED);
4182 	MNT_FLAG(MNT_DEFEXPORTED);
4183 	MNT_FLAG(MNT_EXPORTANON);
4184 	MNT_FLAG(MNT_EXKERB);
4185 	MNT_FLAG(MNT_EXPUBLIC);
4186 	MNT_FLAG(MNT_LOCAL);
4187 	MNT_FLAG(MNT_QUOTA);
4188 	MNT_FLAG(MNT_ROOTFS);
4189 	MNT_FLAG(MNT_USER);
4190 	MNT_FLAG(MNT_IGNORE);
4191 	MNT_FLAG(MNT_UPDATE);
4192 	MNT_FLAG(MNT_DELEXPORT);
4193 	MNT_FLAG(MNT_RELOAD);
4194 	MNT_FLAG(MNT_FORCE);
4195 	MNT_FLAG(MNT_SNAPSHOT);
4196 	MNT_FLAG(MNT_BYFSID);
4197 #undef MNT_FLAG
4198 	if (mflags != 0) {
4199 		if (buf[0] != '\0')
4200 			strlcat(buf, ", ", sizeof(buf));
4201 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4202 		    "0x%016jx", mflags);
4203 	}
4204 	db_printf("    mnt_flag = %s\n", buf);
4205 
4206 	buf[0] = '\0';
4207 	flags = mp->mnt_kern_flag;
4208 #define	MNT_KERN_FLAG(flag)	do {					\
4209 	if (flags & (flag)) {						\
4210 		if (buf[0] != '\0')					\
4211 			strlcat(buf, ", ", sizeof(buf));		\
4212 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4213 		flags &= ~(flag);					\
4214 	}								\
4215 } while (0)
4216 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4217 	MNT_KERN_FLAG(MNTK_ASYNC);
4218 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4219 	MNT_KERN_FLAG(MNTK_DRAINING);
4220 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4221 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4222 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4223 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4224 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4225 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4226 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4227 	MNT_KERN_FLAG(MNTK_MARKER);
4228 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4229 	MNT_KERN_FLAG(MNTK_NOASYNC);
4230 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4231 	MNT_KERN_FLAG(MNTK_MWAIT);
4232 	MNT_KERN_FLAG(MNTK_SUSPEND);
4233 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4234 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4235 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4236 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4237 #undef MNT_KERN_FLAG
4238 	if (flags != 0) {
4239 		if (buf[0] != '\0')
4240 			strlcat(buf, ", ", sizeof(buf));
4241 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4242 		    "0x%08x", flags);
4243 	}
4244 	db_printf("    mnt_kern_flag = %s\n", buf);
4245 
4246 	db_printf("    mnt_opt = ");
4247 	opt = TAILQ_FIRST(mp->mnt_opt);
4248 	if (opt != NULL) {
4249 		db_printf("%s", opt->name);
4250 		opt = TAILQ_NEXT(opt, link);
4251 		while (opt != NULL) {
4252 			db_printf(", %s", opt->name);
4253 			opt = TAILQ_NEXT(opt, link);
4254 		}
4255 	}
4256 	db_printf("\n");
4257 
4258 	sp = &mp->mnt_stat;
4259 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4260 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4261 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4262 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4263 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4264 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4265 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4266 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4267 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4268 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4269 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4270 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4271 
4272 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4273 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4274 	if (jailed(mp->mnt_cred))
4275 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4276 	db_printf(" }\n");
4277 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4278 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4279 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4280 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4281 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4282 	    mp->mnt_lazyvnodelistsize);
4283 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4284 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4285 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4286 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4287 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4288 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4289 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4290 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4291 	db_printf("    mnt_secondary_accwrites = %d\n",
4292 	    mp->mnt_secondary_accwrites);
4293 	db_printf("    mnt_gjprovider = %s\n",
4294 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4295 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4296 
4297 	db_printf("\n\nList of active vnodes\n");
4298 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4299 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4300 			vn_printf(vp, "vnode ");
4301 			if (db_pager_quit)
4302 				break;
4303 		}
4304 	}
4305 	db_printf("\n\nList of inactive vnodes\n");
4306 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4307 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4308 			vn_printf(vp, "vnode ");
4309 			if (db_pager_quit)
4310 				break;
4311 		}
4312 	}
4313 }
4314 #endif	/* DDB */
4315 
4316 /*
4317  * Fill in a struct xvfsconf based on a struct vfsconf.
4318  */
4319 static int
4320 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4321 {
4322 	struct xvfsconf xvfsp;
4323 
4324 	bzero(&xvfsp, sizeof(xvfsp));
4325 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4326 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4327 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4328 	xvfsp.vfc_flags = vfsp->vfc_flags;
4329 	/*
4330 	 * These are unused in userland, we keep them
4331 	 * to not break binary compatibility.
4332 	 */
4333 	xvfsp.vfc_vfsops = NULL;
4334 	xvfsp.vfc_next = NULL;
4335 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4336 }
4337 
4338 #ifdef COMPAT_FREEBSD32
4339 struct xvfsconf32 {
4340 	uint32_t	vfc_vfsops;
4341 	char		vfc_name[MFSNAMELEN];
4342 	int32_t		vfc_typenum;
4343 	int32_t		vfc_refcount;
4344 	int32_t		vfc_flags;
4345 	uint32_t	vfc_next;
4346 };
4347 
4348 static int
4349 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4350 {
4351 	struct xvfsconf32 xvfsp;
4352 
4353 	bzero(&xvfsp, sizeof(xvfsp));
4354 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4355 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4356 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4357 	xvfsp.vfc_flags = vfsp->vfc_flags;
4358 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4359 }
4360 #endif
4361 
4362 /*
4363  * Top level filesystem related information gathering.
4364  */
4365 static int
4366 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4367 {
4368 	struct vfsconf *vfsp;
4369 	int error;
4370 
4371 	error = 0;
4372 	vfsconf_slock();
4373 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4374 #ifdef COMPAT_FREEBSD32
4375 		if (req->flags & SCTL_MASK32)
4376 			error = vfsconf2x32(req, vfsp);
4377 		else
4378 #endif
4379 			error = vfsconf2x(req, vfsp);
4380 		if (error)
4381 			break;
4382 	}
4383 	vfsconf_sunlock();
4384 	return (error);
4385 }
4386 
4387 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4388     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4389     "S,xvfsconf", "List of all configured filesystems");
4390 
4391 #ifndef BURN_BRIDGES
4392 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4393 
4394 static int
4395 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4396 {
4397 	int *name = (int *)arg1 - 1;	/* XXX */
4398 	u_int namelen = arg2 + 1;	/* XXX */
4399 	struct vfsconf *vfsp;
4400 
4401 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4402 	    "please rebuild world\n");
4403 
4404 #if 1 || defined(COMPAT_PRELITE2)
4405 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4406 	if (namelen == 1)
4407 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4408 #endif
4409 
4410 	switch (name[1]) {
4411 	case VFS_MAXTYPENUM:
4412 		if (namelen != 2)
4413 			return (ENOTDIR);
4414 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4415 	case VFS_CONF:
4416 		if (namelen != 3)
4417 			return (ENOTDIR);	/* overloaded */
4418 		vfsconf_slock();
4419 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4420 			if (vfsp->vfc_typenum == name[2])
4421 				break;
4422 		}
4423 		vfsconf_sunlock();
4424 		if (vfsp == NULL)
4425 			return (EOPNOTSUPP);
4426 #ifdef COMPAT_FREEBSD32
4427 		if (req->flags & SCTL_MASK32)
4428 			return (vfsconf2x32(req, vfsp));
4429 		else
4430 #endif
4431 			return (vfsconf2x(req, vfsp));
4432 	}
4433 	return (EOPNOTSUPP);
4434 }
4435 
4436 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4437     CTLFLAG_MPSAFE, vfs_sysctl,
4438     "Generic filesystem");
4439 
4440 #if 1 || defined(COMPAT_PRELITE2)
4441 
4442 static int
4443 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4444 {
4445 	int error;
4446 	struct vfsconf *vfsp;
4447 	struct ovfsconf ovfs;
4448 
4449 	vfsconf_slock();
4450 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4451 		bzero(&ovfs, sizeof(ovfs));
4452 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4453 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4454 		ovfs.vfc_index = vfsp->vfc_typenum;
4455 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4456 		ovfs.vfc_flags = vfsp->vfc_flags;
4457 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4458 		if (error != 0) {
4459 			vfsconf_sunlock();
4460 			return (error);
4461 		}
4462 	}
4463 	vfsconf_sunlock();
4464 	return (0);
4465 }
4466 
4467 #endif /* 1 || COMPAT_PRELITE2 */
4468 #endif /* !BURN_BRIDGES */
4469 
4470 #define KINFO_VNODESLOP		10
4471 #ifdef notyet
4472 /*
4473  * Dump vnode list (via sysctl).
4474  */
4475 /* ARGSUSED */
4476 static int
4477 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4478 {
4479 	struct xvnode *xvn;
4480 	struct mount *mp;
4481 	struct vnode *vp;
4482 	int error, len, n;
4483 
4484 	/*
4485 	 * Stale numvnodes access is not fatal here.
4486 	 */
4487 	req->lock = 0;
4488 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4489 	if (!req->oldptr)
4490 		/* Make an estimate */
4491 		return (SYSCTL_OUT(req, 0, len));
4492 
4493 	error = sysctl_wire_old_buffer(req, 0);
4494 	if (error != 0)
4495 		return (error);
4496 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4497 	n = 0;
4498 	mtx_lock(&mountlist_mtx);
4499 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4500 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4501 			continue;
4502 		MNT_ILOCK(mp);
4503 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4504 			if (n == len)
4505 				break;
4506 			vref(vp);
4507 			xvn[n].xv_size = sizeof *xvn;
4508 			xvn[n].xv_vnode = vp;
4509 			xvn[n].xv_id = 0;	/* XXX compat */
4510 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4511 			XV_COPY(usecount);
4512 			XV_COPY(writecount);
4513 			XV_COPY(holdcnt);
4514 			XV_COPY(mount);
4515 			XV_COPY(numoutput);
4516 			XV_COPY(type);
4517 #undef XV_COPY
4518 			xvn[n].xv_flag = vp->v_vflag;
4519 
4520 			switch (vp->v_type) {
4521 			case VREG:
4522 			case VDIR:
4523 			case VLNK:
4524 				break;
4525 			case VBLK:
4526 			case VCHR:
4527 				if (vp->v_rdev == NULL) {
4528 					vrele(vp);
4529 					continue;
4530 				}
4531 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4532 				break;
4533 			case VSOCK:
4534 				xvn[n].xv_socket = vp->v_socket;
4535 				break;
4536 			case VFIFO:
4537 				xvn[n].xv_fifo = vp->v_fifoinfo;
4538 				break;
4539 			case VNON:
4540 			case VBAD:
4541 			default:
4542 				/* shouldn't happen? */
4543 				vrele(vp);
4544 				continue;
4545 			}
4546 			vrele(vp);
4547 			++n;
4548 		}
4549 		MNT_IUNLOCK(mp);
4550 		mtx_lock(&mountlist_mtx);
4551 		vfs_unbusy(mp);
4552 		if (n == len)
4553 			break;
4554 	}
4555 	mtx_unlock(&mountlist_mtx);
4556 
4557 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4558 	free(xvn, M_TEMP);
4559 	return (error);
4560 }
4561 
4562 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4563     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4564     "");
4565 #endif
4566 
4567 static void
4568 unmount_or_warn(struct mount *mp)
4569 {
4570 	int error;
4571 
4572 	error = dounmount(mp, MNT_FORCE, curthread);
4573 	if (error != 0) {
4574 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4575 		if (error == EBUSY)
4576 			printf("BUSY)\n");
4577 		else
4578 			printf("%d)\n", error);
4579 	}
4580 }
4581 
4582 /*
4583  * Unmount all filesystems. The list is traversed in reverse order
4584  * of mounting to avoid dependencies.
4585  */
4586 void
4587 vfs_unmountall(void)
4588 {
4589 	struct mount *mp, *tmp;
4590 
4591 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4592 
4593 	/*
4594 	 * Since this only runs when rebooting, it is not interlocked.
4595 	 */
4596 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4597 		vfs_ref(mp);
4598 
4599 		/*
4600 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4601 		 * unmount of the latter.
4602 		 */
4603 		if (mp == rootdevmp)
4604 			continue;
4605 
4606 		unmount_or_warn(mp);
4607 	}
4608 
4609 	if (rootdevmp != NULL)
4610 		unmount_or_warn(rootdevmp);
4611 }
4612 
4613 static void
4614 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4615 {
4616 
4617 	ASSERT_VI_LOCKED(vp, __func__);
4618 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4619 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4620 		vdropl(vp);
4621 		return;
4622 	}
4623 	if (vn_lock(vp, lkflags) == 0) {
4624 		VI_LOCK(vp);
4625 		vinactive(vp);
4626 		VOP_UNLOCK(vp);
4627 		vdropl(vp);
4628 		return;
4629 	}
4630 	vdefer_inactive_unlocked(vp);
4631 }
4632 
4633 static int
4634 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4635 {
4636 
4637 	return (vp->v_iflag & VI_DEFINACT);
4638 }
4639 
4640 static void __noinline
4641 vfs_periodic_inactive(struct mount *mp, int flags)
4642 {
4643 	struct vnode *vp, *mvp;
4644 	int lkflags;
4645 
4646 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4647 	if (flags != MNT_WAIT)
4648 		lkflags |= LK_NOWAIT;
4649 
4650 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4651 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4652 			VI_UNLOCK(vp);
4653 			continue;
4654 		}
4655 		vp->v_iflag &= ~VI_DEFINACT;
4656 		vfs_deferred_inactive(vp, lkflags);
4657 	}
4658 }
4659 
4660 static inline bool
4661 vfs_want_msync(struct vnode *vp)
4662 {
4663 	struct vm_object *obj;
4664 
4665 	/*
4666 	 * This test may be performed without any locks held.
4667 	 * We rely on vm_object's type stability.
4668 	 */
4669 	if (vp->v_vflag & VV_NOSYNC)
4670 		return (false);
4671 	obj = vp->v_object;
4672 	return (obj != NULL && vm_object_mightbedirty(obj));
4673 }
4674 
4675 static int
4676 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4677 {
4678 
4679 	if (vp->v_vflag & VV_NOSYNC)
4680 		return (false);
4681 	if (vp->v_iflag & VI_DEFINACT)
4682 		return (true);
4683 	return (vfs_want_msync(vp));
4684 }
4685 
4686 static void __noinline
4687 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4688 {
4689 	struct vnode *vp, *mvp;
4690 	struct vm_object *obj;
4691 	struct thread *td;
4692 	int lkflags, objflags;
4693 	bool seen_defer;
4694 
4695 	td = curthread;
4696 
4697 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4698 	if (flags != MNT_WAIT) {
4699 		lkflags |= LK_NOWAIT;
4700 		objflags = OBJPC_NOSYNC;
4701 	} else {
4702 		objflags = OBJPC_SYNC;
4703 	}
4704 
4705 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4706 		seen_defer = false;
4707 		if (vp->v_iflag & VI_DEFINACT) {
4708 			vp->v_iflag &= ~VI_DEFINACT;
4709 			seen_defer = true;
4710 		}
4711 		if (!vfs_want_msync(vp)) {
4712 			if (seen_defer)
4713 				vfs_deferred_inactive(vp, lkflags);
4714 			else
4715 				VI_UNLOCK(vp);
4716 			continue;
4717 		}
4718 		if (vget(vp, lkflags, td) == 0) {
4719 			obj = vp->v_object;
4720 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4721 				VM_OBJECT_WLOCK(obj);
4722 				vm_object_page_clean(obj, 0, 0, objflags);
4723 				VM_OBJECT_WUNLOCK(obj);
4724 			}
4725 			vput(vp);
4726 			if (seen_defer)
4727 				vdrop(vp);
4728 		} else {
4729 			if (seen_defer)
4730 				vdefer_inactive_unlocked(vp);
4731 		}
4732 	}
4733 }
4734 
4735 void
4736 vfs_periodic(struct mount *mp, int flags)
4737 {
4738 
4739 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4740 
4741 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4742 		vfs_periodic_inactive(mp, flags);
4743 	else
4744 		vfs_periodic_msync_inactive(mp, flags);
4745 }
4746 
4747 static void
4748 destroy_vpollinfo_free(struct vpollinfo *vi)
4749 {
4750 
4751 	knlist_destroy(&vi->vpi_selinfo.si_note);
4752 	mtx_destroy(&vi->vpi_lock);
4753 	uma_zfree(vnodepoll_zone, vi);
4754 }
4755 
4756 static void
4757 destroy_vpollinfo(struct vpollinfo *vi)
4758 {
4759 
4760 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4761 	seldrain(&vi->vpi_selinfo);
4762 	destroy_vpollinfo_free(vi);
4763 }
4764 
4765 /*
4766  * Initialize per-vnode helper structure to hold poll-related state.
4767  */
4768 void
4769 v_addpollinfo(struct vnode *vp)
4770 {
4771 	struct vpollinfo *vi;
4772 
4773 	if (vp->v_pollinfo != NULL)
4774 		return;
4775 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4776 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4777 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4778 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4779 	VI_LOCK(vp);
4780 	if (vp->v_pollinfo != NULL) {
4781 		VI_UNLOCK(vp);
4782 		destroy_vpollinfo_free(vi);
4783 		return;
4784 	}
4785 	vp->v_pollinfo = vi;
4786 	VI_UNLOCK(vp);
4787 }
4788 
4789 /*
4790  * Record a process's interest in events which might happen to
4791  * a vnode.  Because poll uses the historic select-style interface
4792  * internally, this routine serves as both the ``check for any
4793  * pending events'' and the ``record my interest in future events''
4794  * functions.  (These are done together, while the lock is held,
4795  * to avoid race conditions.)
4796  */
4797 int
4798 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4799 {
4800 
4801 	v_addpollinfo(vp);
4802 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4803 	if (vp->v_pollinfo->vpi_revents & events) {
4804 		/*
4805 		 * This leaves events we are not interested
4806 		 * in available for the other process which
4807 		 * which presumably had requested them
4808 		 * (otherwise they would never have been
4809 		 * recorded).
4810 		 */
4811 		events &= vp->v_pollinfo->vpi_revents;
4812 		vp->v_pollinfo->vpi_revents &= ~events;
4813 
4814 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4815 		return (events);
4816 	}
4817 	vp->v_pollinfo->vpi_events |= events;
4818 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4819 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4820 	return (0);
4821 }
4822 
4823 /*
4824  * Routine to create and manage a filesystem syncer vnode.
4825  */
4826 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4827 static int	sync_fsync(struct  vop_fsync_args *);
4828 static int	sync_inactive(struct  vop_inactive_args *);
4829 static int	sync_reclaim(struct  vop_reclaim_args *);
4830 
4831 static struct vop_vector sync_vnodeops = {
4832 	.vop_bypass =	VOP_EOPNOTSUPP,
4833 	.vop_close =	sync_close,		/* close */
4834 	.vop_fsync =	sync_fsync,		/* fsync */
4835 	.vop_inactive =	sync_inactive,	/* inactive */
4836 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4837 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4838 	.vop_lock1 =	vop_stdlock,	/* lock */
4839 	.vop_unlock =	vop_stdunlock,	/* unlock */
4840 	.vop_islocked =	vop_stdislocked,	/* islocked */
4841 };
4842 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4843 
4844 /*
4845  * Create a new filesystem syncer vnode for the specified mount point.
4846  */
4847 void
4848 vfs_allocate_syncvnode(struct mount *mp)
4849 {
4850 	struct vnode *vp;
4851 	struct bufobj *bo;
4852 	static long start, incr, next;
4853 	int error;
4854 
4855 	/* Allocate a new vnode */
4856 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4857 	if (error != 0)
4858 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4859 	vp->v_type = VNON;
4860 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4861 	vp->v_vflag |= VV_FORCEINSMQ;
4862 	error = insmntque(vp, mp);
4863 	if (error != 0)
4864 		panic("vfs_allocate_syncvnode: insmntque() failed");
4865 	vp->v_vflag &= ~VV_FORCEINSMQ;
4866 	VOP_UNLOCK(vp);
4867 	/*
4868 	 * Place the vnode onto the syncer worklist. We attempt to
4869 	 * scatter them about on the list so that they will go off
4870 	 * at evenly distributed times even if all the filesystems
4871 	 * are mounted at once.
4872 	 */
4873 	next += incr;
4874 	if (next == 0 || next > syncer_maxdelay) {
4875 		start /= 2;
4876 		incr /= 2;
4877 		if (start == 0) {
4878 			start = syncer_maxdelay / 2;
4879 			incr = syncer_maxdelay;
4880 		}
4881 		next = start;
4882 	}
4883 	bo = &vp->v_bufobj;
4884 	BO_LOCK(bo);
4885 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4886 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4887 	mtx_lock(&sync_mtx);
4888 	sync_vnode_count++;
4889 	if (mp->mnt_syncer == NULL) {
4890 		mp->mnt_syncer = vp;
4891 		vp = NULL;
4892 	}
4893 	mtx_unlock(&sync_mtx);
4894 	BO_UNLOCK(bo);
4895 	if (vp != NULL) {
4896 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4897 		vgone(vp);
4898 		vput(vp);
4899 	}
4900 }
4901 
4902 void
4903 vfs_deallocate_syncvnode(struct mount *mp)
4904 {
4905 	struct vnode *vp;
4906 
4907 	mtx_lock(&sync_mtx);
4908 	vp = mp->mnt_syncer;
4909 	if (vp != NULL)
4910 		mp->mnt_syncer = NULL;
4911 	mtx_unlock(&sync_mtx);
4912 	if (vp != NULL)
4913 		vrele(vp);
4914 }
4915 
4916 /*
4917  * Do a lazy sync of the filesystem.
4918  */
4919 static int
4920 sync_fsync(struct vop_fsync_args *ap)
4921 {
4922 	struct vnode *syncvp = ap->a_vp;
4923 	struct mount *mp = syncvp->v_mount;
4924 	int error, save;
4925 	struct bufobj *bo;
4926 
4927 	/*
4928 	 * We only need to do something if this is a lazy evaluation.
4929 	 */
4930 	if (ap->a_waitfor != MNT_LAZY)
4931 		return (0);
4932 
4933 	/*
4934 	 * Move ourselves to the back of the sync list.
4935 	 */
4936 	bo = &syncvp->v_bufobj;
4937 	BO_LOCK(bo);
4938 	vn_syncer_add_to_worklist(bo, syncdelay);
4939 	BO_UNLOCK(bo);
4940 
4941 	/*
4942 	 * Walk the list of vnodes pushing all that are dirty and
4943 	 * not already on the sync list.
4944 	 */
4945 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4946 		return (0);
4947 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4948 		vfs_unbusy(mp);
4949 		return (0);
4950 	}
4951 	save = curthread_pflags_set(TDP_SYNCIO);
4952 	/*
4953 	 * The filesystem at hand may be idle with free vnodes stored in the
4954 	 * batch.  Return them instead of letting them stay there indefinitely.
4955 	 */
4956 	vfs_periodic(mp, MNT_NOWAIT);
4957 	error = VFS_SYNC(mp, MNT_LAZY);
4958 	curthread_pflags_restore(save);
4959 	vn_finished_write(mp);
4960 	vfs_unbusy(mp);
4961 	return (error);
4962 }
4963 
4964 /*
4965  * The syncer vnode is no referenced.
4966  */
4967 static int
4968 sync_inactive(struct vop_inactive_args *ap)
4969 {
4970 
4971 	vgone(ap->a_vp);
4972 	return (0);
4973 }
4974 
4975 /*
4976  * The syncer vnode is no longer needed and is being decommissioned.
4977  *
4978  * Modifications to the worklist must be protected by sync_mtx.
4979  */
4980 static int
4981 sync_reclaim(struct vop_reclaim_args *ap)
4982 {
4983 	struct vnode *vp = ap->a_vp;
4984 	struct bufobj *bo;
4985 
4986 	bo = &vp->v_bufobj;
4987 	BO_LOCK(bo);
4988 	mtx_lock(&sync_mtx);
4989 	if (vp->v_mount->mnt_syncer == vp)
4990 		vp->v_mount->mnt_syncer = NULL;
4991 	if (bo->bo_flag & BO_ONWORKLST) {
4992 		LIST_REMOVE(bo, bo_synclist);
4993 		syncer_worklist_len--;
4994 		sync_vnode_count--;
4995 		bo->bo_flag &= ~BO_ONWORKLST;
4996 	}
4997 	mtx_unlock(&sync_mtx);
4998 	BO_UNLOCK(bo);
4999 
5000 	return (0);
5001 }
5002 
5003 int
5004 vn_need_pageq_flush(struct vnode *vp)
5005 {
5006 	struct vm_object *obj;
5007 	int need;
5008 
5009 	MPASS(mtx_owned(VI_MTX(vp)));
5010 	need = 0;
5011 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5012 	    vm_object_mightbedirty(obj))
5013 		need = 1;
5014 	return (need);
5015 }
5016 
5017 /*
5018  * Check if vnode represents a disk device
5019  */
5020 int
5021 vn_isdisk(struct vnode *vp, int *errp)
5022 {
5023 	int error;
5024 
5025 	if (vp->v_type != VCHR) {
5026 		error = ENOTBLK;
5027 		goto out;
5028 	}
5029 	error = 0;
5030 	dev_lock();
5031 	if (vp->v_rdev == NULL)
5032 		error = ENXIO;
5033 	else if (vp->v_rdev->si_devsw == NULL)
5034 		error = ENXIO;
5035 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5036 		error = ENOTBLK;
5037 	dev_unlock();
5038 out:
5039 	if (errp != NULL)
5040 		*errp = error;
5041 	return (error == 0);
5042 }
5043 
5044 /*
5045  * Common filesystem object access control check routine.  Accepts a
5046  * vnode's type, "mode", uid and gid, requested access mode, credentials,
5047  * and optional call-by-reference privused argument allowing vaccess()
5048  * to indicate to the caller whether privilege was used to satisfy the
5049  * request (obsoleted).  Returns 0 on success, or an errno on failure.
5050  */
5051 int
5052 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5053     accmode_t accmode, struct ucred *cred, int *privused)
5054 {
5055 	accmode_t dac_granted;
5056 	accmode_t priv_granted;
5057 
5058 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5059 	    ("invalid bit in accmode"));
5060 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5061 	    ("VAPPEND without VWRITE"));
5062 
5063 	/*
5064 	 * Look for a normal, non-privileged way to access the file/directory
5065 	 * as requested.  If it exists, go with that.
5066 	 */
5067 
5068 	if (privused != NULL)
5069 		*privused = 0;
5070 
5071 	dac_granted = 0;
5072 
5073 	/* Check the owner. */
5074 	if (cred->cr_uid == file_uid) {
5075 		dac_granted |= VADMIN;
5076 		if (file_mode & S_IXUSR)
5077 			dac_granted |= VEXEC;
5078 		if (file_mode & S_IRUSR)
5079 			dac_granted |= VREAD;
5080 		if (file_mode & S_IWUSR)
5081 			dac_granted |= (VWRITE | VAPPEND);
5082 
5083 		if ((accmode & dac_granted) == accmode)
5084 			return (0);
5085 
5086 		goto privcheck;
5087 	}
5088 
5089 	/* Otherwise, check the groups (first match) */
5090 	if (groupmember(file_gid, cred)) {
5091 		if (file_mode & S_IXGRP)
5092 			dac_granted |= VEXEC;
5093 		if (file_mode & S_IRGRP)
5094 			dac_granted |= VREAD;
5095 		if (file_mode & S_IWGRP)
5096 			dac_granted |= (VWRITE | VAPPEND);
5097 
5098 		if ((accmode & dac_granted) == accmode)
5099 			return (0);
5100 
5101 		goto privcheck;
5102 	}
5103 
5104 	/* Otherwise, check everyone else. */
5105 	if (file_mode & S_IXOTH)
5106 		dac_granted |= VEXEC;
5107 	if (file_mode & S_IROTH)
5108 		dac_granted |= VREAD;
5109 	if (file_mode & S_IWOTH)
5110 		dac_granted |= (VWRITE | VAPPEND);
5111 	if ((accmode & dac_granted) == accmode)
5112 		return (0);
5113 
5114 privcheck:
5115 	/*
5116 	 * Build a privilege mask to determine if the set of privileges
5117 	 * satisfies the requirements when combined with the granted mask
5118 	 * from above.  For each privilege, if the privilege is required,
5119 	 * bitwise or the request type onto the priv_granted mask.
5120 	 */
5121 	priv_granted = 0;
5122 
5123 	if (type == VDIR) {
5124 		/*
5125 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5126 		 * requests, instead of PRIV_VFS_EXEC.
5127 		 */
5128 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5129 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5130 			priv_granted |= VEXEC;
5131 	} else {
5132 		/*
5133 		 * Ensure that at least one execute bit is on. Otherwise,
5134 		 * a privileged user will always succeed, and we don't want
5135 		 * this to happen unless the file really is executable.
5136 		 */
5137 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5138 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5139 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5140 			priv_granted |= VEXEC;
5141 	}
5142 
5143 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5144 	    !priv_check_cred(cred, PRIV_VFS_READ))
5145 		priv_granted |= VREAD;
5146 
5147 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5148 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5149 		priv_granted |= (VWRITE | VAPPEND);
5150 
5151 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5152 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5153 		priv_granted |= VADMIN;
5154 
5155 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5156 		/* XXX audit: privilege used */
5157 		if (privused != NULL)
5158 			*privused = 1;
5159 		return (0);
5160 	}
5161 
5162 	return ((accmode & VADMIN) ? EPERM : EACCES);
5163 }
5164 
5165 /*
5166  * Credential check based on process requesting service, and per-attribute
5167  * permissions.
5168  */
5169 int
5170 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5171     struct thread *td, accmode_t accmode)
5172 {
5173 
5174 	/*
5175 	 * Kernel-invoked always succeeds.
5176 	 */
5177 	if (cred == NOCRED)
5178 		return (0);
5179 
5180 	/*
5181 	 * Do not allow privileged processes in jail to directly manipulate
5182 	 * system attributes.
5183 	 */
5184 	switch (attrnamespace) {
5185 	case EXTATTR_NAMESPACE_SYSTEM:
5186 		/* Potentially should be: return (EPERM); */
5187 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5188 	case EXTATTR_NAMESPACE_USER:
5189 		return (VOP_ACCESS(vp, accmode, cred, td));
5190 	default:
5191 		return (EPERM);
5192 	}
5193 }
5194 
5195 #ifdef DEBUG_VFS_LOCKS
5196 /*
5197  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5198  * no longer ok to have an unlocked VFS.
5199  */
5200 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5201 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5202 
5203 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5204 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5205     "Drop into debugger on lock violation");
5206 
5207 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5208 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5209     0, "Check for interlock across VOPs");
5210 
5211 int vfs_badlock_print = 1;	/* Print lock violations. */
5212 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5213     0, "Print lock violations");
5214 
5215 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5216 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5217     0, "Print vnode details on lock violations");
5218 
5219 #ifdef KDB
5220 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5221 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5222     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5223 #endif
5224 
5225 static void
5226 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5227 {
5228 
5229 #ifdef KDB
5230 	if (vfs_badlock_backtrace)
5231 		kdb_backtrace();
5232 #endif
5233 	if (vfs_badlock_vnode)
5234 		vn_printf(vp, "vnode ");
5235 	if (vfs_badlock_print)
5236 		printf("%s: %p %s\n", str, (void *)vp, msg);
5237 	if (vfs_badlock_ddb)
5238 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5239 }
5240 
5241 void
5242 assert_vi_locked(struct vnode *vp, const char *str)
5243 {
5244 
5245 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5246 		vfs_badlock("interlock is not locked but should be", str, vp);
5247 }
5248 
5249 void
5250 assert_vi_unlocked(struct vnode *vp, const char *str)
5251 {
5252 
5253 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5254 		vfs_badlock("interlock is locked but should not be", str, vp);
5255 }
5256 
5257 void
5258 assert_vop_locked(struct vnode *vp, const char *str)
5259 {
5260 	int locked;
5261 
5262 	if (!IGNORE_LOCK(vp)) {
5263 		locked = VOP_ISLOCKED(vp);
5264 		if (locked == 0 || locked == LK_EXCLOTHER)
5265 			vfs_badlock("is not locked but should be", str, vp);
5266 	}
5267 }
5268 
5269 void
5270 assert_vop_unlocked(struct vnode *vp, const char *str)
5271 {
5272 
5273 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5274 		vfs_badlock("is locked but should not be", str, vp);
5275 }
5276 
5277 void
5278 assert_vop_elocked(struct vnode *vp, const char *str)
5279 {
5280 
5281 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5282 		vfs_badlock("is not exclusive locked but should be", str, vp);
5283 }
5284 #endif /* DEBUG_VFS_LOCKS */
5285 
5286 void
5287 vop_rename_fail(struct vop_rename_args *ap)
5288 {
5289 
5290 	if (ap->a_tvp != NULL)
5291 		vput(ap->a_tvp);
5292 	if (ap->a_tdvp == ap->a_tvp)
5293 		vrele(ap->a_tdvp);
5294 	else
5295 		vput(ap->a_tdvp);
5296 	vrele(ap->a_fdvp);
5297 	vrele(ap->a_fvp);
5298 }
5299 
5300 void
5301 vop_rename_pre(void *ap)
5302 {
5303 	struct vop_rename_args *a = ap;
5304 
5305 #ifdef DEBUG_VFS_LOCKS
5306 	if (a->a_tvp)
5307 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5308 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5309 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5310 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5311 
5312 	/* Check the source (from). */
5313 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5314 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5315 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5316 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5317 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5318 
5319 	/* Check the target. */
5320 	if (a->a_tvp)
5321 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5322 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5323 #endif
5324 	if (a->a_tdvp != a->a_fdvp)
5325 		vhold(a->a_fdvp);
5326 	if (a->a_tvp != a->a_fvp)
5327 		vhold(a->a_fvp);
5328 	vhold(a->a_tdvp);
5329 	if (a->a_tvp)
5330 		vhold(a->a_tvp);
5331 }
5332 
5333 #ifdef DEBUG_VFS_LOCKS
5334 void
5335 vop_strategy_pre(void *ap)
5336 {
5337 	struct vop_strategy_args *a;
5338 	struct buf *bp;
5339 
5340 	a = ap;
5341 	bp = a->a_bp;
5342 
5343 	/*
5344 	 * Cluster ops lock their component buffers but not the IO container.
5345 	 */
5346 	if ((bp->b_flags & B_CLUSTER) != 0)
5347 		return;
5348 
5349 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5350 		if (vfs_badlock_print)
5351 			printf(
5352 			    "VOP_STRATEGY: bp is not locked but should be\n");
5353 		if (vfs_badlock_ddb)
5354 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5355 	}
5356 }
5357 
5358 void
5359 vop_lock_pre(void *ap)
5360 {
5361 	struct vop_lock1_args *a = ap;
5362 
5363 	if ((a->a_flags & LK_INTERLOCK) == 0)
5364 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5365 	else
5366 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5367 }
5368 
5369 void
5370 vop_lock_post(void *ap, int rc)
5371 {
5372 	struct vop_lock1_args *a = ap;
5373 
5374 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5375 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5376 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5377 }
5378 
5379 void
5380 vop_unlock_pre(void *ap)
5381 {
5382 	struct vop_unlock_args *a = ap;
5383 
5384 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5385 }
5386 
5387 void
5388 vop_need_inactive_pre(void *ap)
5389 {
5390 	struct vop_need_inactive_args *a = ap;
5391 
5392 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5393 }
5394 
5395 void
5396 vop_need_inactive_post(void *ap, int rc)
5397 {
5398 	struct vop_need_inactive_args *a = ap;
5399 
5400 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5401 }
5402 #endif
5403 
5404 void
5405 vop_create_post(void *ap, int rc)
5406 {
5407 	struct vop_create_args *a = ap;
5408 
5409 	if (!rc)
5410 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5411 }
5412 
5413 void
5414 vop_deleteextattr_post(void *ap, int rc)
5415 {
5416 	struct vop_deleteextattr_args *a = ap;
5417 
5418 	if (!rc)
5419 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5420 }
5421 
5422 void
5423 vop_link_post(void *ap, int rc)
5424 {
5425 	struct vop_link_args *a = ap;
5426 
5427 	if (!rc) {
5428 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5429 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5430 	}
5431 }
5432 
5433 void
5434 vop_mkdir_post(void *ap, int rc)
5435 {
5436 	struct vop_mkdir_args *a = ap;
5437 
5438 	if (!rc)
5439 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5440 }
5441 
5442 void
5443 vop_mknod_post(void *ap, int rc)
5444 {
5445 	struct vop_mknod_args *a = ap;
5446 
5447 	if (!rc)
5448 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5449 }
5450 
5451 void
5452 vop_reclaim_post(void *ap, int rc)
5453 {
5454 	struct vop_reclaim_args *a = ap;
5455 
5456 	if (!rc)
5457 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5458 }
5459 
5460 void
5461 vop_remove_post(void *ap, int rc)
5462 {
5463 	struct vop_remove_args *a = ap;
5464 
5465 	if (!rc) {
5466 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5467 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5468 	}
5469 }
5470 
5471 void
5472 vop_rename_post(void *ap, int rc)
5473 {
5474 	struct vop_rename_args *a = ap;
5475 	long hint;
5476 
5477 	if (!rc) {
5478 		hint = NOTE_WRITE;
5479 		if (a->a_fdvp == a->a_tdvp) {
5480 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5481 				hint |= NOTE_LINK;
5482 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5483 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5484 		} else {
5485 			hint |= NOTE_EXTEND;
5486 			if (a->a_fvp->v_type == VDIR)
5487 				hint |= NOTE_LINK;
5488 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5489 
5490 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5491 			    a->a_tvp->v_type == VDIR)
5492 				hint &= ~NOTE_LINK;
5493 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5494 		}
5495 
5496 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5497 		if (a->a_tvp)
5498 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5499 	}
5500 	if (a->a_tdvp != a->a_fdvp)
5501 		vdrop(a->a_fdvp);
5502 	if (a->a_tvp != a->a_fvp)
5503 		vdrop(a->a_fvp);
5504 	vdrop(a->a_tdvp);
5505 	if (a->a_tvp)
5506 		vdrop(a->a_tvp);
5507 }
5508 
5509 void
5510 vop_rmdir_post(void *ap, int rc)
5511 {
5512 	struct vop_rmdir_args *a = ap;
5513 
5514 	if (!rc) {
5515 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5516 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5517 	}
5518 }
5519 
5520 void
5521 vop_setattr_post(void *ap, int rc)
5522 {
5523 	struct vop_setattr_args *a = ap;
5524 
5525 	if (!rc)
5526 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5527 }
5528 
5529 void
5530 vop_setextattr_post(void *ap, int rc)
5531 {
5532 	struct vop_setextattr_args *a = ap;
5533 
5534 	if (!rc)
5535 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5536 }
5537 
5538 void
5539 vop_symlink_post(void *ap, int rc)
5540 {
5541 	struct vop_symlink_args *a = ap;
5542 
5543 	if (!rc)
5544 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5545 }
5546 
5547 void
5548 vop_open_post(void *ap, int rc)
5549 {
5550 	struct vop_open_args *a = ap;
5551 
5552 	if (!rc)
5553 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5554 }
5555 
5556 void
5557 vop_close_post(void *ap, int rc)
5558 {
5559 	struct vop_close_args *a = ap;
5560 
5561 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5562 	    !VN_IS_DOOMED(a->a_vp))) {
5563 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5564 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5565 	}
5566 }
5567 
5568 void
5569 vop_read_post(void *ap, int rc)
5570 {
5571 	struct vop_read_args *a = ap;
5572 
5573 	if (!rc)
5574 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5575 }
5576 
5577 void
5578 vop_readdir_post(void *ap, int rc)
5579 {
5580 	struct vop_readdir_args *a = ap;
5581 
5582 	if (!rc)
5583 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5584 }
5585 
5586 static struct knlist fs_knlist;
5587 
5588 static void
5589 vfs_event_init(void *arg)
5590 {
5591 	knlist_init_mtx(&fs_knlist, NULL);
5592 }
5593 /* XXX - correct order? */
5594 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5595 
5596 void
5597 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5598 {
5599 
5600 	KNOTE_UNLOCKED(&fs_knlist, event);
5601 }
5602 
5603 static int	filt_fsattach(struct knote *kn);
5604 static void	filt_fsdetach(struct knote *kn);
5605 static int	filt_fsevent(struct knote *kn, long hint);
5606 
5607 struct filterops fs_filtops = {
5608 	.f_isfd = 0,
5609 	.f_attach = filt_fsattach,
5610 	.f_detach = filt_fsdetach,
5611 	.f_event = filt_fsevent
5612 };
5613 
5614 static int
5615 filt_fsattach(struct knote *kn)
5616 {
5617 
5618 	kn->kn_flags |= EV_CLEAR;
5619 	knlist_add(&fs_knlist, kn, 0);
5620 	return (0);
5621 }
5622 
5623 static void
5624 filt_fsdetach(struct knote *kn)
5625 {
5626 
5627 	knlist_remove(&fs_knlist, kn, 0);
5628 }
5629 
5630 static int
5631 filt_fsevent(struct knote *kn, long hint)
5632 {
5633 
5634 	kn->kn_fflags |= hint;
5635 	return (kn->kn_fflags != 0);
5636 }
5637 
5638 static int
5639 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5640 {
5641 	struct vfsidctl vc;
5642 	int error;
5643 	struct mount *mp;
5644 
5645 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5646 	if (error)
5647 		return (error);
5648 	if (vc.vc_vers != VFS_CTL_VERS1)
5649 		return (EINVAL);
5650 	mp = vfs_getvfs(&vc.vc_fsid);
5651 	if (mp == NULL)
5652 		return (ENOENT);
5653 	/* ensure that a specific sysctl goes to the right filesystem. */
5654 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5655 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5656 		vfs_rel(mp);
5657 		return (EINVAL);
5658 	}
5659 	VCTLTOREQ(&vc, req);
5660 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5661 	vfs_rel(mp);
5662 	return (error);
5663 }
5664 
5665 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5666     NULL, 0, sysctl_vfs_ctl, "",
5667     "Sysctl by fsid");
5668 
5669 /*
5670  * Function to initialize a va_filerev field sensibly.
5671  * XXX: Wouldn't a random number make a lot more sense ??
5672  */
5673 u_quad_t
5674 init_va_filerev(void)
5675 {
5676 	struct bintime bt;
5677 
5678 	getbinuptime(&bt);
5679 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5680 }
5681 
5682 static int	filt_vfsread(struct knote *kn, long hint);
5683 static int	filt_vfswrite(struct knote *kn, long hint);
5684 static int	filt_vfsvnode(struct knote *kn, long hint);
5685 static void	filt_vfsdetach(struct knote *kn);
5686 static struct filterops vfsread_filtops = {
5687 	.f_isfd = 1,
5688 	.f_detach = filt_vfsdetach,
5689 	.f_event = filt_vfsread
5690 };
5691 static struct filterops vfswrite_filtops = {
5692 	.f_isfd = 1,
5693 	.f_detach = filt_vfsdetach,
5694 	.f_event = filt_vfswrite
5695 };
5696 static struct filterops vfsvnode_filtops = {
5697 	.f_isfd = 1,
5698 	.f_detach = filt_vfsdetach,
5699 	.f_event = filt_vfsvnode
5700 };
5701 
5702 static void
5703 vfs_knllock(void *arg)
5704 {
5705 	struct vnode *vp = arg;
5706 
5707 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5708 }
5709 
5710 static void
5711 vfs_knlunlock(void *arg)
5712 {
5713 	struct vnode *vp = arg;
5714 
5715 	VOP_UNLOCK(vp);
5716 }
5717 
5718 static void
5719 vfs_knl_assert_locked(void *arg)
5720 {
5721 #ifdef DEBUG_VFS_LOCKS
5722 	struct vnode *vp = arg;
5723 
5724 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5725 #endif
5726 }
5727 
5728 static void
5729 vfs_knl_assert_unlocked(void *arg)
5730 {
5731 #ifdef DEBUG_VFS_LOCKS
5732 	struct vnode *vp = arg;
5733 
5734 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5735 #endif
5736 }
5737 
5738 int
5739 vfs_kqfilter(struct vop_kqfilter_args *ap)
5740 {
5741 	struct vnode *vp = ap->a_vp;
5742 	struct knote *kn = ap->a_kn;
5743 	struct knlist *knl;
5744 
5745 	switch (kn->kn_filter) {
5746 	case EVFILT_READ:
5747 		kn->kn_fop = &vfsread_filtops;
5748 		break;
5749 	case EVFILT_WRITE:
5750 		kn->kn_fop = &vfswrite_filtops;
5751 		break;
5752 	case EVFILT_VNODE:
5753 		kn->kn_fop = &vfsvnode_filtops;
5754 		break;
5755 	default:
5756 		return (EINVAL);
5757 	}
5758 
5759 	kn->kn_hook = (caddr_t)vp;
5760 
5761 	v_addpollinfo(vp);
5762 	if (vp->v_pollinfo == NULL)
5763 		return (ENOMEM);
5764 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5765 	vhold(vp);
5766 	knlist_add(knl, kn, 0);
5767 
5768 	return (0);
5769 }
5770 
5771 /*
5772  * Detach knote from vnode
5773  */
5774 static void
5775 filt_vfsdetach(struct knote *kn)
5776 {
5777 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5778 
5779 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5780 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5781 	vdrop(vp);
5782 }
5783 
5784 /*ARGSUSED*/
5785 static int
5786 filt_vfsread(struct knote *kn, long hint)
5787 {
5788 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5789 	struct vattr va;
5790 	int res;
5791 
5792 	/*
5793 	 * filesystem is gone, so set the EOF flag and schedule
5794 	 * the knote for deletion.
5795 	 */
5796 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5797 		VI_LOCK(vp);
5798 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5799 		VI_UNLOCK(vp);
5800 		return (1);
5801 	}
5802 
5803 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5804 		return (0);
5805 
5806 	VI_LOCK(vp);
5807 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5808 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5809 	VI_UNLOCK(vp);
5810 	return (res);
5811 }
5812 
5813 /*ARGSUSED*/
5814 static int
5815 filt_vfswrite(struct knote *kn, long hint)
5816 {
5817 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5818 
5819 	VI_LOCK(vp);
5820 
5821 	/*
5822 	 * filesystem is gone, so set the EOF flag and schedule
5823 	 * the knote for deletion.
5824 	 */
5825 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5826 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5827 
5828 	kn->kn_data = 0;
5829 	VI_UNLOCK(vp);
5830 	return (1);
5831 }
5832 
5833 static int
5834 filt_vfsvnode(struct knote *kn, long hint)
5835 {
5836 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5837 	int res;
5838 
5839 	VI_LOCK(vp);
5840 	if (kn->kn_sfflags & hint)
5841 		kn->kn_fflags |= hint;
5842 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5843 		kn->kn_flags |= EV_EOF;
5844 		VI_UNLOCK(vp);
5845 		return (1);
5846 	}
5847 	res = (kn->kn_fflags != 0);
5848 	VI_UNLOCK(vp);
5849 	return (res);
5850 }
5851 
5852 /*
5853  * Returns whether the directory is empty or not.
5854  * If it is empty, the return value is 0; otherwise
5855  * the return value is an error value (which may
5856  * be ENOTEMPTY).
5857  */
5858 int
5859 vfs_emptydir(struct vnode *vp)
5860 {
5861 	struct uio uio;
5862 	struct iovec iov;
5863 	struct dirent *dirent, *dp, *endp;
5864 	int error, eof;
5865 
5866 	error = 0;
5867 	eof = 0;
5868 
5869 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5870 
5871 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5872 	iov.iov_base = dirent;
5873 	iov.iov_len = sizeof(struct dirent);
5874 
5875 	uio.uio_iov = &iov;
5876 	uio.uio_iovcnt = 1;
5877 	uio.uio_offset = 0;
5878 	uio.uio_resid = sizeof(struct dirent);
5879 	uio.uio_segflg = UIO_SYSSPACE;
5880 	uio.uio_rw = UIO_READ;
5881 	uio.uio_td = curthread;
5882 
5883 	while (eof == 0 && error == 0) {
5884 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5885 		    NULL, NULL);
5886 		if (error != 0)
5887 			break;
5888 		endp = (void *)((uint8_t *)dirent +
5889 		    sizeof(struct dirent) - uio.uio_resid);
5890 		for (dp = dirent; dp < endp;
5891 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5892 			if (dp->d_type == DT_WHT)
5893 				continue;
5894 			if (dp->d_namlen == 0)
5895 				continue;
5896 			if (dp->d_type != DT_DIR &&
5897 			    dp->d_type != DT_UNKNOWN) {
5898 				error = ENOTEMPTY;
5899 				break;
5900 			}
5901 			if (dp->d_namlen > 2) {
5902 				error = ENOTEMPTY;
5903 				break;
5904 			}
5905 			if (dp->d_namlen == 1 &&
5906 			    dp->d_name[0] != '.') {
5907 				error = ENOTEMPTY;
5908 				break;
5909 			}
5910 			if (dp->d_namlen == 2 &&
5911 			    dp->d_name[1] != '.') {
5912 				error = ENOTEMPTY;
5913 				break;
5914 			}
5915 			uio.uio_resid = sizeof(struct dirent);
5916 		}
5917 	}
5918 	free(dirent, M_TEMP);
5919 	return (error);
5920 }
5921 
5922 int
5923 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5924 {
5925 	int error;
5926 
5927 	if (dp->d_reclen > ap->a_uio->uio_resid)
5928 		return (ENAMETOOLONG);
5929 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5930 	if (error) {
5931 		if (ap->a_ncookies != NULL) {
5932 			if (ap->a_cookies != NULL)
5933 				free(ap->a_cookies, M_TEMP);
5934 			ap->a_cookies = NULL;
5935 			*ap->a_ncookies = 0;
5936 		}
5937 		return (error);
5938 	}
5939 	if (ap->a_ncookies == NULL)
5940 		return (0);
5941 
5942 	KASSERT(ap->a_cookies,
5943 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5944 
5945 	*ap->a_cookies = realloc(*ap->a_cookies,
5946 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5947 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5948 	*ap->a_ncookies += 1;
5949 	return (0);
5950 }
5951 
5952 /*
5953  * The purpose of this routine is to remove granularity from accmode_t,
5954  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5955  * VADMIN and VAPPEND.
5956  *
5957  * If it returns 0, the caller is supposed to continue with the usual
5958  * access checks using 'accmode' as modified by this routine.  If it
5959  * returns nonzero value, the caller is supposed to return that value
5960  * as errno.
5961  *
5962  * Note that after this routine runs, accmode may be zero.
5963  */
5964 int
5965 vfs_unixify_accmode(accmode_t *accmode)
5966 {
5967 	/*
5968 	 * There is no way to specify explicit "deny" rule using
5969 	 * file mode or POSIX.1e ACLs.
5970 	 */
5971 	if (*accmode & VEXPLICIT_DENY) {
5972 		*accmode = 0;
5973 		return (0);
5974 	}
5975 
5976 	/*
5977 	 * None of these can be translated into usual access bits.
5978 	 * Also, the common case for NFSv4 ACLs is to not contain
5979 	 * either of these bits. Caller should check for VWRITE
5980 	 * on the containing directory instead.
5981 	 */
5982 	if (*accmode & (VDELETE_CHILD | VDELETE))
5983 		return (EPERM);
5984 
5985 	if (*accmode & VADMIN_PERMS) {
5986 		*accmode &= ~VADMIN_PERMS;
5987 		*accmode |= VADMIN;
5988 	}
5989 
5990 	/*
5991 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5992 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5993 	 */
5994 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5995 
5996 	return (0);
5997 }
5998 
5999 /*
6000  * Clear out a doomed vnode (if any) and replace it with a new one as long
6001  * as the fs is not being unmounted. Return the root vnode to the caller.
6002  */
6003 static int __noinline
6004 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6005 {
6006 	struct vnode *vp;
6007 	int error;
6008 
6009 restart:
6010 	if (mp->mnt_rootvnode != NULL) {
6011 		MNT_ILOCK(mp);
6012 		vp = mp->mnt_rootvnode;
6013 		if (vp != NULL) {
6014 			if (!VN_IS_DOOMED(vp)) {
6015 				vrefact(vp);
6016 				MNT_IUNLOCK(mp);
6017 				error = vn_lock(vp, flags);
6018 				if (error == 0) {
6019 					*vpp = vp;
6020 					return (0);
6021 				}
6022 				vrele(vp);
6023 				goto restart;
6024 			}
6025 			/*
6026 			 * Clear the old one.
6027 			 */
6028 			mp->mnt_rootvnode = NULL;
6029 		}
6030 		MNT_IUNLOCK(mp);
6031 		if (vp != NULL) {
6032 			/*
6033 			 * Paired with a fence in vfs_op_thread_exit().
6034 			 */
6035 			atomic_thread_fence_acq();
6036 			vfs_op_barrier_wait(mp);
6037 			vrele(vp);
6038 		}
6039 	}
6040 	error = VFS_CACHEDROOT(mp, flags, vpp);
6041 	if (error != 0)
6042 		return (error);
6043 	if (mp->mnt_vfs_ops == 0) {
6044 		MNT_ILOCK(mp);
6045 		if (mp->mnt_vfs_ops != 0) {
6046 			MNT_IUNLOCK(mp);
6047 			return (0);
6048 		}
6049 		if (mp->mnt_rootvnode == NULL) {
6050 			vrefact(*vpp);
6051 			mp->mnt_rootvnode = *vpp;
6052 		} else {
6053 			if (mp->mnt_rootvnode != *vpp) {
6054 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6055 					panic("%s: mismatch between vnode returned "
6056 					    " by VFS_CACHEDROOT and the one cached "
6057 					    " (%p != %p)",
6058 					    __func__, *vpp, mp->mnt_rootvnode);
6059 				}
6060 			}
6061 		}
6062 		MNT_IUNLOCK(mp);
6063 	}
6064 	return (0);
6065 }
6066 
6067 int
6068 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6069 {
6070 	struct vnode *vp;
6071 	int error;
6072 
6073 	if (!vfs_op_thread_enter(mp))
6074 		return (vfs_cache_root_fallback(mp, flags, vpp));
6075 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
6076 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6077 		vfs_op_thread_exit(mp);
6078 		return (vfs_cache_root_fallback(mp, flags, vpp));
6079 	}
6080 	vrefact(vp);
6081 	vfs_op_thread_exit(mp);
6082 	error = vn_lock(vp, flags);
6083 	if (error != 0) {
6084 		vrele(vp);
6085 		return (vfs_cache_root_fallback(mp, flags, vpp));
6086 	}
6087 	*vpp = vp;
6088 	return (0);
6089 }
6090 
6091 struct vnode *
6092 vfs_cache_root_clear(struct mount *mp)
6093 {
6094 	struct vnode *vp;
6095 
6096 	/*
6097 	 * ops > 0 guarantees there is nobody who can see this vnode
6098 	 */
6099 	MPASS(mp->mnt_vfs_ops > 0);
6100 	vp = mp->mnt_rootvnode;
6101 	mp->mnt_rootvnode = NULL;
6102 	return (vp);
6103 }
6104 
6105 void
6106 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6107 {
6108 
6109 	MPASS(mp->mnt_vfs_ops > 0);
6110 	vrefact(vp);
6111 	mp->mnt_rootvnode = vp;
6112 }
6113 
6114 /*
6115  * These are helper functions for filesystems to traverse all
6116  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6117  *
6118  * This interface replaces MNT_VNODE_FOREACH.
6119  */
6120 
6121 struct vnode *
6122 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6123 {
6124 	struct vnode *vp;
6125 
6126 	if (should_yield())
6127 		kern_yield(PRI_USER);
6128 	MNT_ILOCK(mp);
6129 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6130 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6131 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6132 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6133 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6134 			continue;
6135 		VI_LOCK(vp);
6136 		if (VN_IS_DOOMED(vp)) {
6137 			VI_UNLOCK(vp);
6138 			continue;
6139 		}
6140 		break;
6141 	}
6142 	if (vp == NULL) {
6143 		__mnt_vnode_markerfree_all(mvp, mp);
6144 		/* MNT_IUNLOCK(mp); -- done in above function */
6145 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6146 		return (NULL);
6147 	}
6148 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6149 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6150 	MNT_IUNLOCK(mp);
6151 	return (vp);
6152 }
6153 
6154 struct vnode *
6155 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6156 {
6157 	struct vnode *vp;
6158 
6159 	*mvp = vn_alloc_marker(mp);
6160 	MNT_ILOCK(mp);
6161 	MNT_REF(mp);
6162 
6163 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6164 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6165 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6166 			continue;
6167 		VI_LOCK(vp);
6168 		if (VN_IS_DOOMED(vp)) {
6169 			VI_UNLOCK(vp);
6170 			continue;
6171 		}
6172 		break;
6173 	}
6174 	if (vp == NULL) {
6175 		MNT_REL(mp);
6176 		MNT_IUNLOCK(mp);
6177 		vn_free_marker(*mvp);
6178 		*mvp = NULL;
6179 		return (NULL);
6180 	}
6181 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6182 	MNT_IUNLOCK(mp);
6183 	return (vp);
6184 }
6185 
6186 void
6187 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6188 {
6189 
6190 	if (*mvp == NULL) {
6191 		MNT_IUNLOCK(mp);
6192 		return;
6193 	}
6194 
6195 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6196 
6197 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6198 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6199 	MNT_REL(mp);
6200 	MNT_IUNLOCK(mp);
6201 	vn_free_marker(*mvp);
6202 	*mvp = NULL;
6203 }
6204 
6205 /*
6206  * These are helper functions for filesystems to traverse their
6207  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6208  */
6209 static void
6210 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6211 {
6212 
6213 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6214 
6215 	MNT_ILOCK(mp);
6216 	MNT_REL(mp);
6217 	MNT_IUNLOCK(mp);
6218 	vn_free_marker(*mvp);
6219 	*mvp = NULL;
6220 }
6221 
6222 /*
6223  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6224  * conventional lock order during mnt_vnode_next_lazy iteration.
6225  *
6226  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6227  * The list lock is dropped and reacquired.  On success, both locks are held.
6228  * On failure, the mount vnode list lock is held but the vnode interlock is
6229  * not, and the procedure may have yielded.
6230  */
6231 static bool
6232 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6233     struct vnode *vp)
6234 {
6235 
6236 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6237 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6238 	    ("%s: bad marker", __func__));
6239 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6240 	    ("%s: inappropriate vnode", __func__));
6241 	ASSERT_VI_UNLOCKED(vp, __func__);
6242 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6243 
6244 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6245 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6246 
6247 	vholdnz(vp);
6248 	mtx_unlock(&mp->mnt_listmtx);
6249 	VI_LOCK(vp);
6250 	if (VN_IS_DOOMED(vp)) {
6251 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6252 		goto out_lost;
6253 	}
6254 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6255 	/*
6256 	 * Since we had a period with no locks held we may be the last
6257 	 * remaining user, in which case there is nothing to do.
6258 	 */
6259 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6260 		goto out_lost;
6261 	mtx_lock(&mp->mnt_listmtx);
6262 	return (true);
6263 out_lost:
6264 	vdropl(vp);
6265 	maybe_yield();
6266 	mtx_lock(&mp->mnt_listmtx);
6267 	return (false);
6268 }
6269 
6270 static struct vnode *
6271 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6272     void *cbarg)
6273 {
6274 	struct vnode *vp;
6275 
6276 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6277 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6278 restart:
6279 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6280 	while (vp != NULL) {
6281 		if (vp->v_type == VMARKER) {
6282 			vp = TAILQ_NEXT(vp, v_lazylist);
6283 			continue;
6284 		}
6285 		/*
6286 		 * See if we want to process the vnode. Note we may encounter a
6287 		 * long string of vnodes we don't care about and hog the list
6288 		 * as a result. Check for it and requeue the marker.
6289 		 */
6290 		VNPASS(!VN_IS_DOOMED(vp), vp);
6291 		if (!cb(vp, cbarg)) {
6292 			if (!should_yield()) {
6293 				vp = TAILQ_NEXT(vp, v_lazylist);
6294 				continue;
6295 			}
6296 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6297 			    v_lazylist);
6298 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6299 			    v_lazylist);
6300 			mtx_unlock(&mp->mnt_listmtx);
6301 			kern_yield(PRI_USER);
6302 			mtx_lock(&mp->mnt_listmtx);
6303 			goto restart;
6304 		}
6305 		/*
6306 		 * Try-lock because this is the wrong lock order.
6307 		 */
6308 		if (!VI_TRYLOCK(vp) &&
6309 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6310 			goto restart;
6311 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6312 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6313 		    ("alien vnode on the lazy list %p %p", vp, mp));
6314 		VNPASS(vp->v_mount == mp, vp);
6315 		VNPASS(!VN_IS_DOOMED(vp), vp);
6316 		break;
6317 	}
6318 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6319 
6320 	/* Check if we are done */
6321 	if (vp == NULL) {
6322 		mtx_unlock(&mp->mnt_listmtx);
6323 		mnt_vnode_markerfree_lazy(mvp, mp);
6324 		return (NULL);
6325 	}
6326 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6327 	mtx_unlock(&mp->mnt_listmtx);
6328 	ASSERT_VI_LOCKED(vp, "lazy iter");
6329 	return (vp);
6330 }
6331 
6332 struct vnode *
6333 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6334     void *cbarg)
6335 {
6336 
6337 	if (should_yield())
6338 		kern_yield(PRI_USER);
6339 	mtx_lock(&mp->mnt_listmtx);
6340 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6341 }
6342 
6343 struct vnode *
6344 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6345     void *cbarg)
6346 {
6347 	struct vnode *vp;
6348 
6349 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6350 		return (NULL);
6351 
6352 	*mvp = vn_alloc_marker(mp);
6353 	MNT_ILOCK(mp);
6354 	MNT_REF(mp);
6355 	MNT_IUNLOCK(mp);
6356 
6357 	mtx_lock(&mp->mnt_listmtx);
6358 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6359 	if (vp == NULL) {
6360 		mtx_unlock(&mp->mnt_listmtx);
6361 		mnt_vnode_markerfree_lazy(mvp, mp);
6362 		return (NULL);
6363 	}
6364 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6365 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6366 }
6367 
6368 void
6369 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6370 {
6371 
6372 	if (*mvp == NULL)
6373 		return;
6374 
6375 	mtx_lock(&mp->mnt_listmtx);
6376 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6377 	mtx_unlock(&mp->mnt_listmtx);
6378 	mnt_vnode_markerfree_lazy(mvp, mp);
6379 }
6380 
6381 int
6382 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6383 {
6384 
6385 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6386 		cnp->cn_flags &= ~NOEXECCHECK;
6387 		return (0);
6388 	}
6389 
6390 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6391 }
6392