xref: /freebsd/sys/kern/vfs_subr.c (revision b19cdab3456b361e1ef79651fe1437d8cab4de19)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smr.h>
80 #include <sys/smp.h>
81 #include <sys/stat.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 #include <sys/watchdog.h>
87 
88 #include <machine/stdarg.h>
89 
90 #include <security/mac/mac_framework.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_extern.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_kern.h>
99 #include <vm/uma.h>
100 
101 #ifdef DDB
102 #include <ddb/ddb.h>
103 #endif
104 
105 static void	delmntque(struct vnode *vp);
106 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
107 		    int slpflag, int slptimeo);
108 static void	syncer_shutdown(void *arg, int howto);
109 static int	vtryrecycle(struct vnode *vp);
110 static void	v_init_counters(struct vnode *);
111 static void	vgonel(struct vnode *);
112 static bool	vhold_recycle(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 static counter_u64_t deferred_inact;
209 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
210     "Number of times inactive processing was deferred");
211 
212 /* To keep more than one thread at a time from running vfs_getnewfsid */
213 static struct mtx mntid_mtx;
214 
215 /*
216  * Lock for any access to the following:
217  *	vnode_list
218  *	numvnodes
219  *	freevnodes
220  */
221 static struct mtx __exclusive_cache_line vnode_list_mtx;
222 
223 /* Publicly exported FS */
224 struct nfs_public nfs_pub;
225 
226 static uma_zone_t buf_trie_zone;
227 static smr_t buf_trie_smr;
228 
229 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
230 static uma_zone_t vnode_zone;
231 static uma_zone_t vnodepoll_zone;
232 
233 __read_frequently smr_t vfs_smr;
234 
235 /*
236  * The workitem queue.
237  *
238  * It is useful to delay writes of file data and filesystem metadata
239  * for tens of seconds so that quickly created and deleted files need
240  * not waste disk bandwidth being created and removed. To realize this,
241  * we append vnodes to a "workitem" queue. When running with a soft
242  * updates implementation, most pending metadata dependencies should
243  * not wait for more than a few seconds. Thus, mounted on block devices
244  * are delayed only about a half the time that file data is delayed.
245  * Similarly, directory updates are more critical, so are only delayed
246  * about a third the time that file data is delayed. Thus, there are
247  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
248  * one each second (driven off the filesystem syncer process). The
249  * syncer_delayno variable indicates the next queue that is to be processed.
250  * Items that need to be processed soon are placed in this queue:
251  *
252  *	syncer_workitem_pending[syncer_delayno]
253  *
254  * A delay of fifteen seconds is done by placing the request fifteen
255  * entries later in the queue:
256  *
257  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
258  *
259  */
260 static int syncer_delayno;
261 static long syncer_mask;
262 LIST_HEAD(synclist, bufobj);
263 static struct synclist *syncer_workitem_pending;
264 /*
265  * The sync_mtx protects:
266  *	bo->bo_synclist
267  *	sync_vnode_count
268  *	syncer_delayno
269  *	syncer_state
270  *	syncer_workitem_pending
271  *	syncer_worklist_len
272  *	rushjob
273  */
274 static struct mtx sync_mtx;
275 static struct cv sync_wakeup;
276 
277 #define SYNCER_MAXDELAY		32
278 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
279 static int syncdelay = 30;		/* max time to delay syncing data */
280 static int filedelay = 30;		/* time to delay syncing files */
281 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
282     "Time to delay syncing files (in seconds)");
283 static int dirdelay = 29;		/* time to delay syncing directories */
284 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
285     "Time to delay syncing directories (in seconds)");
286 static int metadelay = 28;		/* time to delay syncing metadata */
287 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
288     "Time to delay syncing metadata (in seconds)");
289 static int rushjob;		/* number of slots to run ASAP */
290 static int stat_rush_requests;	/* number of times I/O speeded up */
291 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
292     "Number of times I/O speeded up (rush requests)");
293 
294 #define	VDBATCH_SIZE 8
295 struct vdbatch {
296 	u_int index;
297 	long freevnodes;
298 	struct mtx lock;
299 	struct vnode *tab[VDBATCH_SIZE];
300 };
301 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
302 
303 static void	vdbatch_dequeue(struct vnode *vp);
304 
305 /*
306  * When shutting down the syncer, run it at four times normal speed.
307  */
308 #define SYNCER_SHUTDOWN_SPEEDUP		4
309 static int sync_vnode_count;
310 static int syncer_worklist_len;
311 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
312     syncer_state;
313 
314 /* Target for maximum number of vnodes. */
315 u_long desiredvnodes;
316 static u_long gapvnodes;		/* gap between wanted and desired */
317 static u_long vhiwat;		/* enough extras after expansion */
318 static u_long vlowat;		/* minimal extras before expansion */
319 static u_long vstir;		/* nonzero to stir non-free vnodes */
320 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
321 
322 static u_long vnlru_read_freevnodes(void);
323 
324 /*
325  * Note that no attempt is made to sanitize these parameters.
326  */
327 static int
328 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
329 {
330 	u_long val;
331 	int error;
332 
333 	val = desiredvnodes;
334 	error = sysctl_handle_long(oidp, &val, 0, req);
335 	if (error != 0 || req->newptr == NULL)
336 		return (error);
337 
338 	if (val == desiredvnodes)
339 		return (0);
340 	mtx_lock(&vnode_list_mtx);
341 	desiredvnodes = val;
342 	wantfreevnodes = desiredvnodes / 4;
343 	vnlru_recalc();
344 	mtx_unlock(&vnode_list_mtx);
345 	/*
346 	 * XXX There is no protection against multiple threads changing
347 	 * desiredvnodes at the same time. Locking above only helps vnlru and
348 	 * getnewvnode.
349 	 */
350 	vfs_hash_changesize(desiredvnodes);
351 	cache_changesize(desiredvnodes);
352 	return (0);
353 }
354 
355 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
356     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
357     "LU", "Target for maximum number of vnodes");
358 
359 static int
360 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
361 {
362 	u_long val;
363 	int error;
364 
365 	val = wantfreevnodes;
366 	error = sysctl_handle_long(oidp, &val, 0, req);
367 	if (error != 0 || req->newptr == NULL)
368 		return (error);
369 
370 	if (val == wantfreevnodes)
371 		return (0);
372 	mtx_lock(&vnode_list_mtx);
373 	wantfreevnodes = val;
374 	vnlru_recalc();
375 	mtx_unlock(&vnode_list_mtx);
376 	return (0);
377 }
378 
379 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
380     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
381     "LU", "Target for minimum number of \"free\" vnodes");
382 
383 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
384     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
385 static int vnlru_nowhere;
386 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
387     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
388 
389 static int
390 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
391 {
392 	struct vnode *vp;
393 	struct nameidata nd;
394 	char *buf;
395 	unsigned long ndflags;
396 	int error;
397 
398 	if (req->newptr == NULL)
399 		return (EINVAL);
400 	if (req->newlen >= PATH_MAX)
401 		return (E2BIG);
402 
403 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
404 	error = SYSCTL_IN(req, buf, req->newlen);
405 	if (error != 0)
406 		goto out;
407 
408 	buf[req->newlen] = '\0';
409 
410 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
411 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
412 	if ((error = namei(&nd)) != 0)
413 		goto out;
414 	vp = nd.ni_vp;
415 
416 	if (VN_IS_DOOMED(vp)) {
417 		/*
418 		 * This vnode is being recycled.  Return != 0 to let the caller
419 		 * know that the sysctl had no effect.  Return EAGAIN because a
420 		 * subsequent call will likely succeed (since namei will create
421 		 * a new vnode if necessary)
422 		 */
423 		error = EAGAIN;
424 		goto putvnode;
425 	}
426 
427 	counter_u64_add(recycles_count, 1);
428 	vgone(vp);
429 putvnode:
430 	NDFREE(&nd, 0);
431 out:
432 	free(buf, M_TEMP);
433 	return (error);
434 }
435 
436 static int
437 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
438 {
439 	struct thread *td = curthread;
440 	struct vnode *vp;
441 	struct file *fp;
442 	int error;
443 	int fd;
444 
445 	if (req->newptr == NULL)
446 		return (EBADF);
447 
448         error = sysctl_handle_int(oidp, &fd, 0, req);
449         if (error != 0)
450                 return (error);
451 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
452 	if (error != 0)
453 		return (error);
454 	vp = fp->f_vnode;
455 
456 	error = vn_lock(vp, LK_EXCLUSIVE);
457 	if (error != 0)
458 		goto drop;
459 
460 	counter_u64_add(recycles_count, 1);
461 	vgone(vp);
462 	VOP_UNLOCK(vp);
463 drop:
464 	fdrop(fp, td);
465 	return (error);
466 }
467 
468 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
469     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
470     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
471 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
472     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
473     sysctl_ftry_reclaim_vnode, "I",
474     "Try to reclaim a vnode by its file descriptor");
475 
476 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
477 static int vnsz2log;
478 
479 /*
480  * Support for the bufobj clean & dirty pctrie.
481  */
482 static void *
483 buf_trie_alloc(struct pctrie *ptree)
484 {
485 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
486 }
487 
488 static void
489 buf_trie_free(struct pctrie *ptree, void *node)
490 {
491 	uma_zfree_smr(buf_trie_zone, node);
492 }
493 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
494     buf_trie_smr);
495 
496 /*
497  * Initialize the vnode management data structures.
498  *
499  * Reevaluate the following cap on the number of vnodes after the physical
500  * memory size exceeds 512GB.  In the limit, as the physical memory size
501  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
502  */
503 #ifndef	MAXVNODES_MAX
504 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
505 #endif
506 
507 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
508 
509 static struct vnode *
510 vn_alloc_marker(struct mount *mp)
511 {
512 	struct vnode *vp;
513 
514 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
515 	vp->v_type = VMARKER;
516 	vp->v_mount = mp;
517 
518 	return (vp);
519 }
520 
521 static void
522 vn_free_marker(struct vnode *vp)
523 {
524 
525 	MPASS(vp->v_type == VMARKER);
526 	free(vp, M_VNODE_MARKER);
527 }
528 
529 /*
530  * Initialize a vnode as it first enters the zone.
531  */
532 static int
533 vnode_init(void *mem, int size, int flags)
534 {
535 	struct vnode *vp;
536 
537 	vp = mem;
538 	bzero(vp, size);
539 	/*
540 	 * Setup locks.
541 	 */
542 	vp->v_vnlock = &vp->v_lock;
543 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
544 	/*
545 	 * By default, don't allow shared locks unless filesystems opt-in.
546 	 */
547 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
548 	    LK_NOSHARE | LK_IS_VNODE);
549 	/*
550 	 * Initialize bufobj.
551 	 */
552 	bufobj_init(&vp->v_bufobj, vp);
553 	/*
554 	 * Initialize namecache.
555 	 */
556 	cache_vnode_init(vp);
557 	/*
558 	 * Initialize rangelocks.
559 	 */
560 	rangelock_init(&vp->v_rl);
561 
562 	vp->v_dbatchcpu = NOCPU;
563 
564 	mtx_lock(&vnode_list_mtx);
565 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
566 	mtx_unlock(&vnode_list_mtx);
567 	return (0);
568 }
569 
570 /*
571  * Free a vnode when it is cleared from the zone.
572  */
573 static void
574 vnode_fini(void *mem, int size)
575 {
576 	struct vnode *vp;
577 	struct bufobj *bo;
578 
579 	vp = mem;
580 	vdbatch_dequeue(vp);
581 	mtx_lock(&vnode_list_mtx);
582 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
583 	mtx_unlock(&vnode_list_mtx);
584 	rangelock_destroy(&vp->v_rl);
585 	lockdestroy(vp->v_vnlock);
586 	mtx_destroy(&vp->v_interlock);
587 	bo = &vp->v_bufobj;
588 	rw_destroy(BO_LOCKPTR(bo));
589 }
590 
591 /*
592  * Provide the size of NFS nclnode and NFS fh for calculation of the
593  * vnode memory consumption.  The size is specified directly to
594  * eliminate dependency on NFS-private header.
595  *
596  * Other filesystems may use bigger or smaller (like UFS and ZFS)
597  * private inode data, but the NFS-based estimation is ample enough.
598  * Still, we care about differences in the size between 64- and 32-bit
599  * platforms.
600  *
601  * Namecache structure size is heuristically
602  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
603  */
604 #ifdef _LP64
605 #define	NFS_NCLNODE_SZ	(528 + 64)
606 #define	NC_SZ		148
607 #else
608 #define	NFS_NCLNODE_SZ	(360 + 32)
609 #define	NC_SZ		92
610 #endif
611 
612 static void
613 vntblinit(void *dummy __unused)
614 {
615 	struct vdbatch *vd;
616 	int cpu, physvnodes, virtvnodes;
617 	u_int i;
618 
619 	/*
620 	 * Desiredvnodes is a function of the physical memory size and the
621 	 * kernel's heap size.  Generally speaking, it scales with the
622 	 * physical memory size.  The ratio of desiredvnodes to the physical
623 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
624 	 * Thereafter, the
625 	 * marginal ratio of desiredvnodes to the physical memory size is
626 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
627 	 * size.  The memory required by desiredvnodes vnodes and vm objects
628 	 * must not exceed 1/10th of the kernel's heap size.
629 	 */
630 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
631 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
632 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
633 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
634 	desiredvnodes = min(physvnodes, virtvnodes);
635 	if (desiredvnodes > MAXVNODES_MAX) {
636 		if (bootverbose)
637 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
638 			    desiredvnodes, MAXVNODES_MAX);
639 		desiredvnodes = MAXVNODES_MAX;
640 	}
641 	wantfreevnodes = desiredvnodes / 4;
642 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
643 	TAILQ_INIT(&vnode_list);
644 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
645 	/*
646 	 * The lock is taken to appease WITNESS.
647 	 */
648 	mtx_lock(&vnode_list_mtx);
649 	vnlru_recalc();
650 	mtx_unlock(&vnode_list_mtx);
651 	vnode_list_free_marker = vn_alloc_marker(NULL);
652 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
653 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
654 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
655 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
656 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
657 	uma_zone_set_smr(vnode_zone, vfs_smr);
658 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
659 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
660 	/*
661 	 * Preallocate enough nodes to support one-per buf so that
662 	 * we can not fail an insert.  reassignbuf() callers can not
663 	 * tolerate the insertion failure.
664 	 */
665 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
666 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
667 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
668 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
669 	uma_prealloc(buf_trie_zone, nbuf);
670 
671 	vnodes_created = counter_u64_alloc(M_WAITOK);
672 	recycles_count = counter_u64_alloc(M_WAITOK);
673 	recycles_free_count = counter_u64_alloc(M_WAITOK);
674 	deferred_inact = counter_u64_alloc(M_WAITOK);
675 
676 	/*
677 	 * Initialize the filesystem syncer.
678 	 */
679 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
680 	    &syncer_mask);
681 	syncer_maxdelay = syncer_mask + 1;
682 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
683 	cv_init(&sync_wakeup, "syncer");
684 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
685 		vnsz2log++;
686 	vnsz2log--;
687 
688 	CPU_FOREACH(cpu) {
689 		vd = DPCPU_ID_PTR((cpu), vd);
690 		bzero(vd, sizeof(*vd));
691 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
692 	}
693 }
694 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
695 
696 /*
697  * Mark a mount point as busy. Used to synchronize access and to delay
698  * unmounting. Eventually, mountlist_mtx is not released on failure.
699  *
700  * vfs_busy() is a custom lock, it can block the caller.
701  * vfs_busy() only sleeps if the unmount is active on the mount point.
702  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
703  * vnode belonging to mp.
704  *
705  * Lookup uses vfs_busy() to traverse mount points.
706  * root fs			var fs
707  * / vnode lock		A	/ vnode lock (/var)		D
708  * /var vnode lock	B	/log vnode lock(/var/log)	E
709  * vfs_busy lock	C	vfs_busy lock			F
710  *
711  * Within each file system, the lock order is C->A->B and F->D->E.
712  *
713  * When traversing across mounts, the system follows that lock order:
714  *
715  *        C->A->B
716  *              |
717  *              +->F->D->E
718  *
719  * The lookup() process for namei("/var") illustrates the process:
720  *  VOP_LOOKUP() obtains B while A is held
721  *  vfs_busy() obtains a shared lock on F while A and B are held
722  *  vput() releases lock on B
723  *  vput() releases lock on A
724  *  VFS_ROOT() obtains lock on D while shared lock on F is held
725  *  vfs_unbusy() releases shared lock on F
726  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
727  *    Attempt to lock A (instead of vp_crossmp) while D is held would
728  *    violate the global order, causing deadlocks.
729  *
730  * dounmount() locks B while F is drained.
731  */
732 int
733 vfs_busy(struct mount *mp, int flags)
734 {
735 
736 	MPASS((flags & ~MBF_MASK) == 0);
737 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
738 
739 	if (vfs_op_thread_enter(mp)) {
740 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
741 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
742 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
743 		vfs_mp_count_add_pcpu(mp, ref, 1);
744 		vfs_mp_count_add_pcpu(mp, lockref, 1);
745 		vfs_op_thread_exit(mp);
746 		if (flags & MBF_MNTLSTLOCK)
747 			mtx_unlock(&mountlist_mtx);
748 		return (0);
749 	}
750 
751 	MNT_ILOCK(mp);
752 	vfs_assert_mount_counters(mp);
753 	MNT_REF(mp);
754 	/*
755 	 * If mount point is currently being unmounted, sleep until the
756 	 * mount point fate is decided.  If thread doing the unmounting fails,
757 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
758 	 * that this mount point has survived the unmount attempt and vfs_busy
759 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
760 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
761 	 * about to be really destroyed.  vfs_busy needs to release its
762 	 * reference on the mount point in this case and return with ENOENT,
763 	 * telling the caller that mount mount it tried to busy is no longer
764 	 * valid.
765 	 */
766 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
767 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
768 			MNT_REL(mp);
769 			MNT_IUNLOCK(mp);
770 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
771 			    __func__);
772 			return (ENOENT);
773 		}
774 		if (flags & MBF_MNTLSTLOCK)
775 			mtx_unlock(&mountlist_mtx);
776 		mp->mnt_kern_flag |= MNTK_MWAIT;
777 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
778 		if (flags & MBF_MNTLSTLOCK)
779 			mtx_lock(&mountlist_mtx);
780 		MNT_ILOCK(mp);
781 	}
782 	if (flags & MBF_MNTLSTLOCK)
783 		mtx_unlock(&mountlist_mtx);
784 	mp->mnt_lockref++;
785 	MNT_IUNLOCK(mp);
786 	return (0);
787 }
788 
789 /*
790  * Free a busy filesystem.
791  */
792 void
793 vfs_unbusy(struct mount *mp)
794 {
795 	int c;
796 
797 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
798 
799 	if (vfs_op_thread_enter(mp)) {
800 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
801 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
802 		vfs_mp_count_sub_pcpu(mp, ref, 1);
803 		vfs_op_thread_exit(mp);
804 		return;
805 	}
806 
807 	MNT_ILOCK(mp);
808 	vfs_assert_mount_counters(mp);
809 	MNT_REL(mp);
810 	c = --mp->mnt_lockref;
811 	if (mp->mnt_vfs_ops == 0) {
812 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
813 		MNT_IUNLOCK(mp);
814 		return;
815 	}
816 	if (c < 0)
817 		vfs_dump_mount_counters(mp);
818 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
819 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
820 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
821 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
822 		wakeup(&mp->mnt_lockref);
823 	}
824 	MNT_IUNLOCK(mp);
825 }
826 
827 /*
828  * Lookup a mount point by filesystem identifier.
829  */
830 struct mount *
831 vfs_getvfs(fsid_t *fsid)
832 {
833 	struct mount *mp;
834 
835 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
836 	mtx_lock(&mountlist_mtx);
837 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
838 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
839 			vfs_ref(mp);
840 			mtx_unlock(&mountlist_mtx);
841 			return (mp);
842 		}
843 	}
844 	mtx_unlock(&mountlist_mtx);
845 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
846 	return ((struct mount *) 0);
847 }
848 
849 /*
850  * Lookup a mount point by filesystem identifier, busying it before
851  * returning.
852  *
853  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
854  * cache for popular filesystem identifiers.  The cache is lockess, using
855  * the fact that struct mount's are never freed.  In worst case we may
856  * get pointer to unmounted or even different filesystem, so we have to
857  * check what we got, and go slow way if so.
858  */
859 struct mount *
860 vfs_busyfs(fsid_t *fsid)
861 {
862 #define	FSID_CACHE_SIZE	256
863 	typedef struct mount * volatile vmp_t;
864 	static vmp_t cache[FSID_CACHE_SIZE];
865 	struct mount *mp;
866 	int error;
867 	uint32_t hash;
868 
869 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
870 	hash = fsid->val[0] ^ fsid->val[1];
871 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
872 	mp = cache[hash];
873 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
874 		goto slow;
875 	if (vfs_busy(mp, 0) != 0) {
876 		cache[hash] = NULL;
877 		goto slow;
878 	}
879 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
880 		return (mp);
881 	else
882 	    vfs_unbusy(mp);
883 
884 slow:
885 	mtx_lock(&mountlist_mtx);
886 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
887 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
888 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
889 			if (error) {
890 				cache[hash] = NULL;
891 				mtx_unlock(&mountlist_mtx);
892 				return (NULL);
893 			}
894 			cache[hash] = mp;
895 			return (mp);
896 		}
897 	}
898 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
899 	mtx_unlock(&mountlist_mtx);
900 	return ((struct mount *) 0);
901 }
902 
903 /*
904  * Check if a user can access privileged mount options.
905  */
906 int
907 vfs_suser(struct mount *mp, struct thread *td)
908 {
909 	int error;
910 
911 	if (jailed(td->td_ucred)) {
912 		/*
913 		 * If the jail of the calling thread lacks permission for
914 		 * this type of file system, deny immediately.
915 		 */
916 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
917 			return (EPERM);
918 
919 		/*
920 		 * If the file system was mounted outside the jail of the
921 		 * calling thread, deny immediately.
922 		 */
923 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
924 			return (EPERM);
925 	}
926 
927 	/*
928 	 * If file system supports delegated administration, we don't check
929 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
930 	 * by the file system itself.
931 	 * If this is not the user that did original mount, we check for
932 	 * the PRIV_VFS_MOUNT_OWNER privilege.
933 	 */
934 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
935 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
936 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
937 			return (error);
938 	}
939 	return (0);
940 }
941 
942 /*
943  * Get a new unique fsid.  Try to make its val[0] unique, since this value
944  * will be used to create fake device numbers for stat().  Also try (but
945  * not so hard) make its val[0] unique mod 2^16, since some emulators only
946  * support 16-bit device numbers.  We end up with unique val[0]'s for the
947  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
948  *
949  * Keep in mind that several mounts may be running in parallel.  Starting
950  * the search one past where the previous search terminated is both a
951  * micro-optimization and a defense against returning the same fsid to
952  * different mounts.
953  */
954 void
955 vfs_getnewfsid(struct mount *mp)
956 {
957 	static uint16_t mntid_base;
958 	struct mount *nmp;
959 	fsid_t tfsid;
960 	int mtype;
961 
962 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
963 	mtx_lock(&mntid_mtx);
964 	mtype = mp->mnt_vfc->vfc_typenum;
965 	tfsid.val[1] = mtype;
966 	mtype = (mtype & 0xFF) << 24;
967 	for (;;) {
968 		tfsid.val[0] = makedev(255,
969 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
970 		mntid_base++;
971 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
972 			break;
973 		vfs_rel(nmp);
974 	}
975 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
976 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
977 	mtx_unlock(&mntid_mtx);
978 }
979 
980 /*
981  * Knob to control the precision of file timestamps:
982  *
983  *   0 = seconds only; nanoseconds zeroed.
984  *   1 = seconds and nanoseconds, accurate within 1/HZ.
985  *   2 = seconds and nanoseconds, truncated to microseconds.
986  * >=3 = seconds and nanoseconds, maximum precision.
987  */
988 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
989 
990 static int timestamp_precision = TSP_USEC;
991 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
992     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
993     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
994     "3+: sec + ns (max. precision))");
995 
996 /*
997  * Get a current timestamp.
998  */
999 void
1000 vfs_timestamp(struct timespec *tsp)
1001 {
1002 	struct timeval tv;
1003 
1004 	switch (timestamp_precision) {
1005 	case TSP_SEC:
1006 		tsp->tv_sec = time_second;
1007 		tsp->tv_nsec = 0;
1008 		break;
1009 	case TSP_HZ:
1010 		getnanotime(tsp);
1011 		break;
1012 	case TSP_USEC:
1013 		microtime(&tv);
1014 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1015 		break;
1016 	case TSP_NSEC:
1017 	default:
1018 		nanotime(tsp);
1019 		break;
1020 	}
1021 }
1022 
1023 /*
1024  * Set vnode attributes to VNOVAL
1025  */
1026 void
1027 vattr_null(struct vattr *vap)
1028 {
1029 
1030 	vap->va_type = VNON;
1031 	vap->va_size = VNOVAL;
1032 	vap->va_bytes = VNOVAL;
1033 	vap->va_mode = VNOVAL;
1034 	vap->va_nlink = VNOVAL;
1035 	vap->va_uid = VNOVAL;
1036 	vap->va_gid = VNOVAL;
1037 	vap->va_fsid = VNOVAL;
1038 	vap->va_fileid = VNOVAL;
1039 	vap->va_blocksize = VNOVAL;
1040 	vap->va_rdev = VNOVAL;
1041 	vap->va_atime.tv_sec = VNOVAL;
1042 	vap->va_atime.tv_nsec = VNOVAL;
1043 	vap->va_mtime.tv_sec = VNOVAL;
1044 	vap->va_mtime.tv_nsec = VNOVAL;
1045 	vap->va_ctime.tv_sec = VNOVAL;
1046 	vap->va_ctime.tv_nsec = VNOVAL;
1047 	vap->va_birthtime.tv_sec = VNOVAL;
1048 	vap->va_birthtime.tv_nsec = VNOVAL;
1049 	vap->va_flags = VNOVAL;
1050 	vap->va_gen = VNOVAL;
1051 	vap->va_vaflags = 0;
1052 }
1053 
1054 /*
1055  * Try to reduce the total number of vnodes.
1056  *
1057  * This routine (and its user) are buggy in at least the following ways:
1058  * - all parameters were picked years ago when RAM sizes were significantly
1059  *   smaller
1060  * - it can pick vnodes based on pages used by the vm object, but filesystems
1061  *   like ZFS don't use it making the pick broken
1062  * - since ZFS has its own aging policy it gets partially combated by this one
1063  * - a dedicated method should be provided for filesystems to let them decide
1064  *   whether the vnode should be recycled
1065  *
1066  * This routine is called when we have too many vnodes.  It attempts
1067  * to free <count> vnodes and will potentially free vnodes that still
1068  * have VM backing store (VM backing store is typically the cause
1069  * of a vnode blowout so we want to do this).  Therefore, this operation
1070  * is not considered cheap.
1071  *
1072  * A number of conditions may prevent a vnode from being reclaimed.
1073  * the buffer cache may have references on the vnode, a directory
1074  * vnode may still have references due to the namei cache representing
1075  * underlying files, or the vnode may be in active use.   It is not
1076  * desirable to reuse such vnodes.  These conditions may cause the
1077  * number of vnodes to reach some minimum value regardless of what
1078  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1079  *
1080  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1081  * 			 entries if this argument is strue
1082  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1083  *			 pages.
1084  * @param target	 How many vnodes to reclaim.
1085  * @return		 The number of vnodes that were reclaimed.
1086  */
1087 static int
1088 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1089 {
1090 	struct vnode *vp, *mvp;
1091 	struct mount *mp;
1092 	struct vm_object *object;
1093 	u_long done;
1094 	bool retried;
1095 
1096 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1097 
1098 	retried = false;
1099 	done = 0;
1100 
1101 	mvp = vnode_list_reclaim_marker;
1102 restart:
1103 	vp = mvp;
1104 	while (done < target) {
1105 		vp = TAILQ_NEXT(vp, v_vnodelist);
1106 		if (__predict_false(vp == NULL))
1107 			break;
1108 
1109 		if (__predict_false(vp->v_type == VMARKER))
1110 			continue;
1111 
1112 		/*
1113 		 * If it's been deconstructed already, it's still
1114 		 * referenced, or it exceeds the trigger, skip it.
1115 		 * Also skip free vnodes.  We are trying to make space
1116 		 * to expand the free list, not reduce it.
1117 		 */
1118 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1119 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1120 			goto next_iter;
1121 
1122 		if (vp->v_type == VBAD || vp->v_type == VNON)
1123 			goto next_iter;
1124 
1125 		object = atomic_load_ptr(&vp->v_object);
1126 		if (object == NULL || object->resident_page_count > trigger) {
1127 			goto next_iter;
1128 		}
1129 
1130 		if (!vhold_recycle(vp))
1131 			goto next_iter;
1132 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1133 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1134 		mtx_unlock(&vnode_list_mtx);
1135 
1136 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1137 			vdrop(vp);
1138 			goto next_iter_unlocked;
1139 		}
1140 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1141 			vdrop(vp);
1142 			vn_finished_write(mp);
1143 			goto next_iter_unlocked;
1144 		}
1145 
1146 		VI_LOCK(vp);
1147 		if (vp->v_usecount > 0 ||
1148 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1149 		    (vp->v_object != NULL &&
1150 		    vp->v_object->resident_page_count > trigger)) {
1151 			VOP_UNLOCK(vp);
1152 			vdropl(vp);
1153 			vn_finished_write(mp);
1154 			goto next_iter_unlocked;
1155 		}
1156 		counter_u64_add(recycles_count, 1);
1157 		vgonel(vp);
1158 		VOP_UNLOCK(vp);
1159 		vdropl(vp);
1160 		vn_finished_write(mp);
1161 		done++;
1162 next_iter_unlocked:
1163 		if (should_yield())
1164 			kern_yield(PRI_USER);
1165 		mtx_lock(&vnode_list_mtx);
1166 		goto restart;
1167 next_iter:
1168 		MPASS(vp->v_type != VMARKER);
1169 		if (!should_yield())
1170 			continue;
1171 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1172 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1173 		mtx_unlock(&vnode_list_mtx);
1174 		kern_yield(PRI_USER);
1175 		mtx_lock(&vnode_list_mtx);
1176 		goto restart;
1177 	}
1178 	if (done == 0 && !retried) {
1179 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1180 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1181 		retried = true;
1182 		goto restart;
1183 	}
1184 	return (done);
1185 }
1186 
1187 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1188 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1189     0,
1190     "limit on vnode free requests per call to the vnlru_free routine");
1191 
1192 /*
1193  * Attempt to reduce the free list by the requested amount.
1194  */
1195 static int
1196 vnlru_free_locked(int count, struct vfsops *mnt_op)
1197 {
1198 	struct vnode *vp, *mvp;
1199 	struct mount *mp;
1200 	int ocount;
1201 
1202 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1203 	if (count > max_vnlru_free)
1204 		count = max_vnlru_free;
1205 	ocount = count;
1206 	mvp = vnode_list_free_marker;
1207 restart:
1208 	vp = mvp;
1209 	while (count > 0) {
1210 		vp = TAILQ_NEXT(vp, v_vnodelist);
1211 		if (__predict_false(vp == NULL)) {
1212 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1213 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1214 			break;
1215 		}
1216 		if (__predict_false(vp->v_type == VMARKER))
1217 			continue;
1218 
1219 		/*
1220 		 * Don't recycle if our vnode is from different type
1221 		 * of mount point.  Note that mp is type-safe, the
1222 		 * check does not reach unmapped address even if
1223 		 * vnode is reclaimed.
1224 		 * Don't recycle if we can't get the interlock without
1225 		 * blocking.
1226 		 */
1227 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1228 		    mp->mnt_op != mnt_op)) {
1229 			continue;
1230 		}
1231 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1232 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1233 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1234 			continue;
1235 		}
1236 		if (!vhold_recycle(vp))
1237 			continue;
1238 		count--;
1239 		mtx_unlock(&vnode_list_mtx);
1240 		vtryrecycle(vp);
1241 		mtx_lock(&vnode_list_mtx);
1242 		goto restart;
1243 	}
1244 	return (ocount - count);
1245 }
1246 
1247 void
1248 vnlru_free(int count, struct vfsops *mnt_op)
1249 {
1250 
1251 	mtx_lock(&vnode_list_mtx);
1252 	vnlru_free_locked(count, mnt_op);
1253 	mtx_unlock(&vnode_list_mtx);
1254 }
1255 
1256 static void
1257 vnlru_recalc(void)
1258 {
1259 
1260 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1261 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1262 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1263 	vlowat = vhiwat / 2;
1264 }
1265 
1266 /*
1267  * Attempt to recycle vnodes in a context that is always safe to block.
1268  * Calling vlrurecycle() from the bowels of filesystem code has some
1269  * interesting deadlock problems.
1270  */
1271 static struct proc *vnlruproc;
1272 static int vnlruproc_sig;
1273 
1274 /*
1275  * The main freevnodes counter is only updated when threads requeue their vnode
1276  * batches. CPUs are conditionally walked to compute a more accurate total.
1277  *
1278  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1279  * at any given moment can still exceed slop, but it should not be by significant
1280  * margin in practice.
1281  */
1282 #define VNLRU_FREEVNODES_SLOP 128
1283 
1284 static __inline void
1285 vn_freevnodes_inc(void)
1286 {
1287 	struct vdbatch *vd;
1288 
1289 	critical_enter();
1290 	vd = DPCPU_PTR(vd);
1291 	vd->freevnodes++;
1292 	critical_exit();
1293 }
1294 
1295 static __inline void
1296 vn_freevnodes_dec(void)
1297 {
1298 	struct vdbatch *vd;
1299 
1300 	critical_enter();
1301 	vd = DPCPU_PTR(vd);
1302 	vd->freevnodes--;
1303 	critical_exit();
1304 }
1305 
1306 static u_long
1307 vnlru_read_freevnodes(void)
1308 {
1309 	struct vdbatch *vd;
1310 	long slop;
1311 	int cpu;
1312 
1313 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1314 	if (freevnodes > freevnodes_old)
1315 		slop = freevnodes - freevnodes_old;
1316 	else
1317 		slop = freevnodes_old - freevnodes;
1318 	if (slop < VNLRU_FREEVNODES_SLOP)
1319 		return (freevnodes >= 0 ? freevnodes : 0);
1320 	freevnodes_old = freevnodes;
1321 	CPU_FOREACH(cpu) {
1322 		vd = DPCPU_ID_PTR((cpu), vd);
1323 		freevnodes_old += vd->freevnodes;
1324 	}
1325 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1326 }
1327 
1328 static bool
1329 vnlru_under(u_long rnumvnodes, u_long limit)
1330 {
1331 	u_long rfreevnodes, space;
1332 
1333 	if (__predict_false(rnumvnodes > desiredvnodes))
1334 		return (true);
1335 
1336 	space = desiredvnodes - rnumvnodes;
1337 	if (space < limit) {
1338 		rfreevnodes = vnlru_read_freevnodes();
1339 		if (rfreevnodes > wantfreevnodes)
1340 			space += rfreevnodes - wantfreevnodes;
1341 	}
1342 	return (space < limit);
1343 }
1344 
1345 static bool
1346 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1347 {
1348 	long rfreevnodes, space;
1349 
1350 	if (__predict_false(rnumvnodes > desiredvnodes))
1351 		return (true);
1352 
1353 	space = desiredvnodes - rnumvnodes;
1354 	if (space < limit) {
1355 		rfreevnodes = atomic_load_long(&freevnodes);
1356 		if (rfreevnodes > wantfreevnodes)
1357 			space += rfreevnodes - wantfreevnodes;
1358 	}
1359 	return (space < limit);
1360 }
1361 
1362 static void
1363 vnlru_kick(void)
1364 {
1365 
1366 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1367 	if (vnlruproc_sig == 0) {
1368 		vnlruproc_sig = 1;
1369 		wakeup(vnlruproc);
1370 	}
1371 }
1372 
1373 static void
1374 vnlru_proc(void)
1375 {
1376 	u_long rnumvnodes, rfreevnodes, target;
1377 	unsigned long onumvnodes;
1378 	int done, force, trigger, usevnodes;
1379 	bool reclaim_nc_src, want_reread;
1380 
1381 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1382 	    SHUTDOWN_PRI_FIRST);
1383 
1384 	force = 0;
1385 	want_reread = false;
1386 	for (;;) {
1387 		kproc_suspend_check(vnlruproc);
1388 		mtx_lock(&vnode_list_mtx);
1389 		rnumvnodes = atomic_load_long(&numvnodes);
1390 
1391 		if (want_reread) {
1392 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1393 			want_reread = false;
1394 		}
1395 
1396 		/*
1397 		 * If numvnodes is too large (due to desiredvnodes being
1398 		 * adjusted using its sysctl, or emergency growth), first
1399 		 * try to reduce it by discarding from the free list.
1400 		 */
1401 		if (rnumvnodes > desiredvnodes) {
1402 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1403 			rnumvnodes = atomic_load_long(&numvnodes);
1404 		}
1405 		/*
1406 		 * Sleep if the vnode cache is in a good state.  This is
1407 		 * when it is not over-full and has space for about a 4%
1408 		 * or 9% expansion (by growing its size or inexcessively
1409 		 * reducing its free list).  Otherwise, try to reclaim
1410 		 * space for a 10% expansion.
1411 		 */
1412 		if (vstir && force == 0) {
1413 			force = 1;
1414 			vstir = 0;
1415 		}
1416 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1417 			vnlruproc_sig = 0;
1418 			wakeup(&vnlruproc_sig);
1419 			msleep(vnlruproc, &vnode_list_mtx,
1420 			    PVFS|PDROP, "vlruwt", hz);
1421 			continue;
1422 		}
1423 		rfreevnodes = vnlru_read_freevnodes();
1424 
1425 		onumvnodes = rnumvnodes;
1426 		/*
1427 		 * Calculate parameters for recycling.  These are the same
1428 		 * throughout the loop to give some semblance of fairness.
1429 		 * The trigger point is to avoid recycling vnodes with lots
1430 		 * of resident pages.  We aren't trying to free memory; we
1431 		 * are trying to recycle or at least free vnodes.
1432 		 */
1433 		if (rnumvnodes <= desiredvnodes)
1434 			usevnodes = rnumvnodes - rfreevnodes;
1435 		else
1436 			usevnodes = rnumvnodes;
1437 		if (usevnodes <= 0)
1438 			usevnodes = 1;
1439 		/*
1440 		 * The trigger value is is chosen to give a conservatively
1441 		 * large value to ensure that it alone doesn't prevent
1442 		 * making progress.  The value can easily be so large that
1443 		 * it is effectively infinite in some congested and
1444 		 * misconfigured cases, and this is necessary.  Normally
1445 		 * it is about 8 to 100 (pages), which is quite large.
1446 		 */
1447 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1448 		if (force < 2)
1449 			trigger = vsmalltrigger;
1450 		reclaim_nc_src = force >= 3;
1451 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1452 		target = target / 10 + 1;
1453 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1454 		mtx_unlock(&vnode_list_mtx);
1455 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1456 			uma_reclaim(UMA_RECLAIM_DRAIN);
1457 		if (done == 0) {
1458 			if (force == 0 || force == 1) {
1459 				force = 2;
1460 				continue;
1461 			}
1462 			if (force == 2) {
1463 				force = 3;
1464 				continue;
1465 			}
1466 			want_reread = true;
1467 			force = 0;
1468 			vnlru_nowhere++;
1469 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1470 		} else {
1471 			want_reread = true;
1472 			kern_yield(PRI_USER);
1473 		}
1474 	}
1475 }
1476 
1477 static struct kproc_desc vnlru_kp = {
1478 	"vnlru",
1479 	vnlru_proc,
1480 	&vnlruproc
1481 };
1482 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1483     &vnlru_kp);
1484 
1485 /*
1486  * Routines having to do with the management of the vnode table.
1487  */
1488 
1489 /*
1490  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1491  * before we actually vgone().  This function must be called with the vnode
1492  * held to prevent the vnode from being returned to the free list midway
1493  * through vgone().
1494  */
1495 static int
1496 vtryrecycle(struct vnode *vp)
1497 {
1498 	struct mount *vnmp;
1499 
1500 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1501 	VNASSERT(vp->v_holdcnt, vp,
1502 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1503 	/*
1504 	 * This vnode may found and locked via some other list, if so we
1505 	 * can't recycle it yet.
1506 	 */
1507 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1508 		CTR2(KTR_VFS,
1509 		    "%s: impossible to recycle, vp %p lock is already held",
1510 		    __func__, vp);
1511 		vdrop(vp);
1512 		return (EWOULDBLOCK);
1513 	}
1514 	/*
1515 	 * Don't recycle if its filesystem is being suspended.
1516 	 */
1517 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1518 		VOP_UNLOCK(vp);
1519 		CTR2(KTR_VFS,
1520 		    "%s: impossible to recycle, cannot start the write for %p",
1521 		    __func__, vp);
1522 		vdrop(vp);
1523 		return (EBUSY);
1524 	}
1525 	/*
1526 	 * If we got this far, we need to acquire the interlock and see if
1527 	 * anyone picked up this vnode from another list.  If not, we will
1528 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1529 	 * will skip over it.
1530 	 */
1531 	VI_LOCK(vp);
1532 	if (vp->v_usecount) {
1533 		VOP_UNLOCK(vp);
1534 		vdropl(vp);
1535 		vn_finished_write(vnmp);
1536 		CTR2(KTR_VFS,
1537 		    "%s: impossible to recycle, %p is already referenced",
1538 		    __func__, vp);
1539 		return (EBUSY);
1540 	}
1541 	if (!VN_IS_DOOMED(vp)) {
1542 		counter_u64_add(recycles_free_count, 1);
1543 		vgonel(vp);
1544 	}
1545 	VOP_UNLOCK(vp);
1546 	vdropl(vp);
1547 	vn_finished_write(vnmp);
1548 	return (0);
1549 }
1550 
1551 /*
1552  * Allocate a new vnode.
1553  *
1554  * The operation never returns an error. Returning an error was disabled
1555  * in r145385 (dated 2005) with the following comment:
1556  *
1557  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1558  *
1559  * Given the age of this commit (almost 15 years at the time of writing this
1560  * comment) restoring the ability to fail requires a significant audit of
1561  * all codepaths.
1562  *
1563  * The routine can try to free a vnode or stall for up to 1 second waiting for
1564  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1565  */
1566 static u_long vn_alloc_cyclecount;
1567 
1568 static struct vnode * __noinline
1569 vn_alloc_hard(struct mount *mp)
1570 {
1571 	u_long rnumvnodes, rfreevnodes;
1572 
1573 	mtx_lock(&vnode_list_mtx);
1574 	rnumvnodes = atomic_load_long(&numvnodes);
1575 	if (rnumvnodes + 1 < desiredvnodes) {
1576 		vn_alloc_cyclecount = 0;
1577 		goto alloc;
1578 	}
1579 	rfreevnodes = vnlru_read_freevnodes();
1580 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1581 		vn_alloc_cyclecount = 0;
1582 		vstir = 1;
1583 	}
1584 	/*
1585 	 * Grow the vnode cache if it will not be above its target max
1586 	 * after growing.  Otherwise, if the free list is nonempty, try
1587 	 * to reclaim 1 item from it before growing the cache (possibly
1588 	 * above its target max if the reclamation failed or is delayed).
1589 	 * Otherwise, wait for some space.  In all cases, schedule
1590 	 * vnlru_proc() if we are getting short of space.  The watermarks
1591 	 * should be chosen so that we never wait or even reclaim from
1592 	 * the free list to below its target minimum.
1593 	 */
1594 	if (vnlru_free_locked(1, NULL) > 0)
1595 		goto alloc;
1596 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1597 		/*
1598 		 * Wait for space for a new vnode.
1599 		 */
1600 		vnlru_kick();
1601 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1602 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1603 		    vnlru_read_freevnodes() > 1)
1604 			vnlru_free_locked(1, NULL);
1605 	}
1606 alloc:
1607 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1608 	if (vnlru_under(rnumvnodes, vlowat))
1609 		vnlru_kick();
1610 	mtx_unlock(&vnode_list_mtx);
1611 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1612 }
1613 
1614 static struct vnode *
1615 vn_alloc(struct mount *mp)
1616 {
1617 	u_long rnumvnodes;
1618 
1619 	if (__predict_false(vn_alloc_cyclecount != 0))
1620 		return (vn_alloc_hard(mp));
1621 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1622 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1623 		atomic_subtract_long(&numvnodes, 1);
1624 		return (vn_alloc_hard(mp));
1625 	}
1626 
1627 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1628 }
1629 
1630 static void
1631 vn_free(struct vnode *vp)
1632 {
1633 
1634 	atomic_subtract_long(&numvnodes, 1);
1635 	uma_zfree_smr(vnode_zone, vp);
1636 }
1637 
1638 /*
1639  * Return the next vnode from the free list.
1640  */
1641 int
1642 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1643     struct vnode **vpp)
1644 {
1645 	struct vnode *vp;
1646 	struct thread *td;
1647 	struct lock_object *lo;
1648 
1649 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1650 
1651 	KASSERT(vops->registered,
1652 	    ("%s: not registered vector op %p\n", __func__, vops));
1653 
1654 	td = curthread;
1655 	if (td->td_vp_reserved != NULL) {
1656 		vp = td->td_vp_reserved;
1657 		td->td_vp_reserved = NULL;
1658 	} else {
1659 		vp = vn_alloc(mp);
1660 	}
1661 	counter_u64_add(vnodes_created, 1);
1662 	/*
1663 	 * Locks are given the generic name "vnode" when created.
1664 	 * Follow the historic practice of using the filesystem
1665 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1666 	 *
1667 	 * Locks live in a witness group keyed on their name. Thus,
1668 	 * when a lock is renamed, it must also move from the witness
1669 	 * group of its old name to the witness group of its new name.
1670 	 *
1671 	 * The change only needs to be made when the vnode moves
1672 	 * from one filesystem type to another. We ensure that each
1673 	 * filesystem use a single static name pointer for its tag so
1674 	 * that we can compare pointers rather than doing a strcmp().
1675 	 */
1676 	lo = &vp->v_vnlock->lock_object;
1677 #ifdef WITNESS
1678 	if (lo->lo_name != tag) {
1679 #endif
1680 		lo->lo_name = tag;
1681 #ifdef WITNESS
1682 		WITNESS_DESTROY(lo);
1683 		WITNESS_INIT(lo, tag);
1684 	}
1685 #endif
1686 	/*
1687 	 * By default, don't allow shared locks unless filesystems opt-in.
1688 	 */
1689 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1690 	/*
1691 	 * Finalize various vnode identity bits.
1692 	 */
1693 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1694 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1695 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1696 	vp->v_type = VNON;
1697 	vp->v_op = vops;
1698 	v_init_counters(vp);
1699 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1700 #ifdef DIAGNOSTIC
1701 	if (mp == NULL && vops != &dead_vnodeops)
1702 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1703 #endif
1704 #ifdef MAC
1705 	mac_vnode_init(vp);
1706 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1707 		mac_vnode_associate_singlelabel(mp, vp);
1708 #endif
1709 	if (mp != NULL) {
1710 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1711 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1712 			vp->v_vflag |= VV_NOKNOTE;
1713 	}
1714 
1715 	/*
1716 	 * For the filesystems which do not use vfs_hash_insert(),
1717 	 * still initialize v_hash to have vfs_hash_index() useful.
1718 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1719 	 * its own hashing.
1720 	 */
1721 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1722 
1723 	*vpp = vp;
1724 	return (0);
1725 }
1726 
1727 void
1728 getnewvnode_reserve(void)
1729 {
1730 	struct thread *td;
1731 
1732 	td = curthread;
1733 	MPASS(td->td_vp_reserved == NULL);
1734 	td->td_vp_reserved = vn_alloc(NULL);
1735 }
1736 
1737 void
1738 getnewvnode_drop_reserve(void)
1739 {
1740 	struct thread *td;
1741 
1742 	td = curthread;
1743 	if (td->td_vp_reserved != NULL) {
1744 		vn_free(td->td_vp_reserved);
1745 		td->td_vp_reserved = NULL;
1746 	}
1747 }
1748 
1749 static void __noinline
1750 freevnode(struct vnode *vp)
1751 {
1752 	struct bufobj *bo;
1753 
1754 	/*
1755 	 * The vnode has been marked for destruction, so free it.
1756 	 *
1757 	 * The vnode will be returned to the zone where it will
1758 	 * normally remain until it is needed for another vnode. We
1759 	 * need to cleanup (or verify that the cleanup has already
1760 	 * been done) any residual data left from its current use
1761 	 * so as not to contaminate the freshly allocated vnode.
1762 	 */
1763 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1764 	/*
1765 	 * Paired with vgone.
1766 	 */
1767 	vn_seqc_write_end_locked(vp);
1768 	VNPASS(vp->v_seqc_users == 0, vp);
1769 
1770 	bo = &vp->v_bufobj;
1771 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1772 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1773 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1774 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1775 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1776 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1777 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1778 	    ("clean blk trie not empty"));
1779 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1780 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1781 	    ("dirty blk trie not empty"));
1782 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1783 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1784 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1785 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1786 	    ("Dangling rangelock waiters"));
1787 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1788 	    ("Leaked inactivation"));
1789 	VI_UNLOCK(vp);
1790 #ifdef MAC
1791 	mac_vnode_destroy(vp);
1792 #endif
1793 	if (vp->v_pollinfo != NULL) {
1794 		destroy_vpollinfo(vp->v_pollinfo);
1795 		vp->v_pollinfo = NULL;
1796 	}
1797 #ifdef INVARIANTS
1798 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1799 	vp->v_op = NULL;
1800 #endif
1801 	vp->v_mountedhere = NULL;
1802 	vp->v_unpcb = NULL;
1803 	vp->v_rdev = NULL;
1804 	vp->v_fifoinfo = NULL;
1805 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1806 	vp->v_irflag = 0;
1807 	vp->v_iflag = 0;
1808 	vp->v_vflag = 0;
1809 	bo->bo_flag = 0;
1810 	vn_free(vp);
1811 }
1812 
1813 /*
1814  * Delete from old mount point vnode list, if on one.
1815  */
1816 static void
1817 delmntque(struct vnode *vp)
1818 {
1819 	struct mount *mp;
1820 
1821 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1822 
1823 	mp = vp->v_mount;
1824 	if (mp == NULL)
1825 		return;
1826 	MNT_ILOCK(mp);
1827 	VI_LOCK(vp);
1828 	vp->v_mount = NULL;
1829 	VI_UNLOCK(vp);
1830 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1831 		("bad mount point vnode list size"));
1832 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1833 	mp->mnt_nvnodelistsize--;
1834 	MNT_REL(mp);
1835 	MNT_IUNLOCK(mp);
1836 }
1837 
1838 static void
1839 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1840 {
1841 
1842 	vp->v_data = NULL;
1843 	vp->v_op = &dead_vnodeops;
1844 	vgone(vp);
1845 	vput(vp);
1846 }
1847 
1848 /*
1849  * Insert into list of vnodes for the new mount point, if available.
1850  */
1851 int
1852 insmntque1(struct vnode *vp, struct mount *mp,
1853 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1854 {
1855 
1856 	KASSERT(vp->v_mount == NULL,
1857 		("insmntque: vnode already on per mount vnode list"));
1858 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1859 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1860 
1861 	/*
1862 	 * We acquire the vnode interlock early to ensure that the
1863 	 * vnode cannot be recycled by another process releasing a
1864 	 * holdcnt on it before we get it on both the vnode list
1865 	 * and the active vnode list. The mount mutex protects only
1866 	 * manipulation of the vnode list and the vnode freelist
1867 	 * mutex protects only manipulation of the active vnode list.
1868 	 * Hence the need to hold the vnode interlock throughout.
1869 	 */
1870 	MNT_ILOCK(mp);
1871 	VI_LOCK(vp);
1872 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1873 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1874 	    mp->mnt_nvnodelistsize == 0)) &&
1875 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1876 		VI_UNLOCK(vp);
1877 		MNT_IUNLOCK(mp);
1878 		if (dtr != NULL)
1879 			dtr(vp, dtr_arg);
1880 		return (EBUSY);
1881 	}
1882 	vp->v_mount = mp;
1883 	MNT_REF(mp);
1884 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1885 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1886 		("neg mount point vnode list size"));
1887 	mp->mnt_nvnodelistsize++;
1888 	VI_UNLOCK(vp);
1889 	MNT_IUNLOCK(mp);
1890 	return (0);
1891 }
1892 
1893 int
1894 insmntque(struct vnode *vp, struct mount *mp)
1895 {
1896 
1897 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1898 }
1899 
1900 /*
1901  * Flush out and invalidate all buffers associated with a bufobj
1902  * Called with the underlying object locked.
1903  */
1904 int
1905 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1906 {
1907 	int error;
1908 
1909 	BO_LOCK(bo);
1910 	if (flags & V_SAVE) {
1911 		error = bufobj_wwait(bo, slpflag, slptimeo);
1912 		if (error) {
1913 			BO_UNLOCK(bo);
1914 			return (error);
1915 		}
1916 		if (bo->bo_dirty.bv_cnt > 0) {
1917 			BO_UNLOCK(bo);
1918 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1919 				return (error);
1920 			/*
1921 			 * XXX We could save a lock/unlock if this was only
1922 			 * enabled under INVARIANTS
1923 			 */
1924 			BO_LOCK(bo);
1925 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1926 				panic("vinvalbuf: dirty bufs");
1927 		}
1928 	}
1929 	/*
1930 	 * If you alter this loop please notice that interlock is dropped and
1931 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1932 	 * no race conditions occur from this.
1933 	 */
1934 	do {
1935 		error = flushbuflist(&bo->bo_clean,
1936 		    flags, bo, slpflag, slptimeo);
1937 		if (error == 0 && !(flags & V_CLEANONLY))
1938 			error = flushbuflist(&bo->bo_dirty,
1939 			    flags, bo, slpflag, slptimeo);
1940 		if (error != 0 && error != EAGAIN) {
1941 			BO_UNLOCK(bo);
1942 			return (error);
1943 		}
1944 	} while (error != 0);
1945 
1946 	/*
1947 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1948 	 * have write I/O in-progress but if there is a VM object then the
1949 	 * VM object can also have read-I/O in-progress.
1950 	 */
1951 	do {
1952 		bufobj_wwait(bo, 0, 0);
1953 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1954 			BO_UNLOCK(bo);
1955 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1956 			BO_LOCK(bo);
1957 		}
1958 	} while (bo->bo_numoutput > 0);
1959 	BO_UNLOCK(bo);
1960 
1961 	/*
1962 	 * Destroy the copy in the VM cache, too.
1963 	 */
1964 	if (bo->bo_object != NULL &&
1965 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1966 		VM_OBJECT_WLOCK(bo->bo_object);
1967 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1968 		    OBJPR_CLEANONLY : 0);
1969 		VM_OBJECT_WUNLOCK(bo->bo_object);
1970 	}
1971 
1972 #ifdef INVARIANTS
1973 	BO_LOCK(bo);
1974 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1975 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1976 	    bo->bo_clean.bv_cnt > 0))
1977 		panic("vinvalbuf: flush failed");
1978 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1979 	    bo->bo_dirty.bv_cnt > 0)
1980 		panic("vinvalbuf: flush dirty failed");
1981 	BO_UNLOCK(bo);
1982 #endif
1983 	return (0);
1984 }
1985 
1986 /*
1987  * Flush out and invalidate all buffers associated with a vnode.
1988  * Called with the underlying object locked.
1989  */
1990 int
1991 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1992 {
1993 
1994 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1995 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1996 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1997 		return (0);
1998 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1999 }
2000 
2001 /*
2002  * Flush out buffers on the specified list.
2003  *
2004  */
2005 static int
2006 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2007     int slptimeo)
2008 {
2009 	struct buf *bp, *nbp;
2010 	int retval, error;
2011 	daddr_t lblkno;
2012 	b_xflags_t xflags;
2013 
2014 	ASSERT_BO_WLOCKED(bo);
2015 
2016 	retval = 0;
2017 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2018 		/*
2019 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2020 		 * do not skip any buffers. If we are flushing only V_NORMAL
2021 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2022 		 * flushing only V_ALT buffers then skip buffers not marked
2023 		 * as BX_ALTDATA.
2024 		 */
2025 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2026 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2027 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2028 			continue;
2029 		}
2030 		if (nbp != NULL) {
2031 			lblkno = nbp->b_lblkno;
2032 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2033 		}
2034 		retval = EAGAIN;
2035 		error = BUF_TIMELOCK(bp,
2036 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2037 		    "flushbuf", slpflag, slptimeo);
2038 		if (error) {
2039 			BO_LOCK(bo);
2040 			return (error != ENOLCK ? error : EAGAIN);
2041 		}
2042 		KASSERT(bp->b_bufobj == bo,
2043 		    ("bp %p wrong b_bufobj %p should be %p",
2044 		    bp, bp->b_bufobj, bo));
2045 		/*
2046 		 * XXX Since there are no node locks for NFS, I
2047 		 * believe there is a slight chance that a delayed
2048 		 * write will occur while sleeping just above, so
2049 		 * check for it.
2050 		 */
2051 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2052 		    (flags & V_SAVE)) {
2053 			bremfree(bp);
2054 			bp->b_flags |= B_ASYNC;
2055 			bwrite(bp);
2056 			BO_LOCK(bo);
2057 			return (EAGAIN);	/* XXX: why not loop ? */
2058 		}
2059 		bremfree(bp);
2060 		bp->b_flags |= (B_INVAL | B_RELBUF);
2061 		bp->b_flags &= ~B_ASYNC;
2062 		brelse(bp);
2063 		BO_LOCK(bo);
2064 		if (nbp == NULL)
2065 			break;
2066 		nbp = gbincore(bo, lblkno);
2067 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2068 		    != xflags)
2069 			break;			/* nbp invalid */
2070 	}
2071 	return (retval);
2072 }
2073 
2074 int
2075 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2076 {
2077 	struct buf *bp;
2078 	int error;
2079 	daddr_t lblkno;
2080 
2081 	ASSERT_BO_LOCKED(bo);
2082 
2083 	for (lblkno = startn;;) {
2084 again:
2085 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2086 		if (bp == NULL || bp->b_lblkno >= endn ||
2087 		    bp->b_lblkno < startn)
2088 			break;
2089 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2090 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2091 		if (error != 0) {
2092 			BO_RLOCK(bo);
2093 			if (error == ENOLCK)
2094 				goto again;
2095 			return (error);
2096 		}
2097 		KASSERT(bp->b_bufobj == bo,
2098 		    ("bp %p wrong b_bufobj %p should be %p",
2099 		    bp, bp->b_bufobj, bo));
2100 		lblkno = bp->b_lblkno + 1;
2101 		if ((bp->b_flags & B_MANAGED) == 0)
2102 			bremfree(bp);
2103 		bp->b_flags |= B_RELBUF;
2104 		/*
2105 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2106 		 * pages backing each buffer in the range are unlikely to be
2107 		 * reused.  Dirty buffers will have the hint applied once
2108 		 * they've been written.
2109 		 */
2110 		if ((bp->b_flags & B_VMIO) != 0)
2111 			bp->b_flags |= B_NOREUSE;
2112 		brelse(bp);
2113 		BO_RLOCK(bo);
2114 	}
2115 	return (0);
2116 }
2117 
2118 /*
2119  * Truncate a file's buffer and pages to a specified length.  This
2120  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2121  * sync activity.
2122  */
2123 int
2124 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2125 {
2126 	struct buf *bp, *nbp;
2127 	struct bufobj *bo;
2128 	daddr_t startlbn;
2129 
2130 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2131 	    vp, blksize, (uintmax_t)length);
2132 
2133 	/*
2134 	 * Round up to the *next* lbn.
2135 	 */
2136 	startlbn = howmany(length, blksize);
2137 
2138 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2139 
2140 	bo = &vp->v_bufobj;
2141 restart_unlocked:
2142 	BO_LOCK(bo);
2143 
2144 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2145 		;
2146 
2147 	if (length > 0) {
2148 restartsync:
2149 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2150 			if (bp->b_lblkno > 0)
2151 				continue;
2152 			/*
2153 			 * Since we hold the vnode lock this should only
2154 			 * fail if we're racing with the buf daemon.
2155 			 */
2156 			if (BUF_LOCK(bp,
2157 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2158 			    BO_LOCKPTR(bo)) == ENOLCK)
2159 				goto restart_unlocked;
2160 
2161 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2162 			    ("buf(%p) on dirty queue without DELWRI", bp));
2163 
2164 			bremfree(bp);
2165 			bawrite(bp);
2166 			BO_LOCK(bo);
2167 			goto restartsync;
2168 		}
2169 	}
2170 
2171 	bufobj_wwait(bo, 0, 0);
2172 	BO_UNLOCK(bo);
2173 	vnode_pager_setsize(vp, length);
2174 
2175 	return (0);
2176 }
2177 
2178 /*
2179  * Invalidate the cached pages of a file's buffer within the range of block
2180  * numbers [startlbn, endlbn).
2181  */
2182 void
2183 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2184     int blksize)
2185 {
2186 	struct bufobj *bo;
2187 	off_t start, end;
2188 
2189 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2190 
2191 	start = blksize * startlbn;
2192 	end = blksize * endlbn;
2193 
2194 	bo = &vp->v_bufobj;
2195 	BO_LOCK(bo);
2196 	MPASS(blksize == bo->bo_bsize);
2197 
2198 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2199 		;
2200 
2201 	BO_UNLOCK(bo);
2202 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2203 }
2204 
2205 static int
2206 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2207     daddr_t startlbn, daddr_t endlbn)
2208 {
2209 	struct buf *bp, *nbp;
2210 	bool anyfreed;
2211 
2212 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2213 	ASSERT_BO_LOCKED(bo);
2214 
2215 	do {
2216 		anyfreed = false;
2217 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2218 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2219 				continue;
2220 			if (BUF_LOCK(bp,
2221 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2222 			    BO_LOCKPTR(bo)) == ENOLCK) {
2223 				BO_LOCK(bo);
2224 				return (EAGAIN);
2225 			}
2226 
2227 			bremfree(bp);
2228 			bp->b_flags |= B_INVAL | B_RELBUF;
2229 			bp->b_flags &= ~B_ASYNC;
2230 			brelse(bp);
2231 			anyfreed = true;
2232 
2233 			BO_LOCK(bo);
2234 			if (nbp != NULL &&
2235 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2236 			    nbp->b_vp != vp ||
2237 			    (nbp->b_flags & B_DELWRI) != 0))
2238 				return (EAGAIN);
2239 		}
2240 
2241 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2242 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2243 				continue;
2244 			if (BUF_LOCK(bp,
2245 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2246 			    BO_LOCKPTR(bo)) == ENOLCK) {
2247 				BO_LOCK(bo);
2248 				return (EAGAIN);
2249 			}
2250 			bremfree(bp);
2251 			bp->b_flags |= B_INVAL | B_RELBUF;
2252 			bp->b_flags &= ~B_ASYNC;
2253 			brelse(bp);
2254 			anyfreed = true;
2255 
2256 			BO_LOCK(bo);
2257 			if (nbp != NULL &&
2258 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2259 			    (nbp->b_vp != vp) ||
2260 			    (nbp->b_flags & B_DELWRI) == 0))
2261 				return (EAGAIN);
2262 		}
2263 	} while (anyfreed);
2264 	return (0);
2265 }
2266 
2267 static void
2268 buf_vlist_remove(struct buf *bp)
2269 {
2270 	struct bufv *bv;
2271 	b_xflags_t flags;
2272 
2273 	flags = bp->b_xflags;
2274 
2275 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2276 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2277 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2278 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2279 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2280 
2281 	if ((flags & BX_VNDIRTY) != 0)
2282 		bv = &bp->b_bufobj->bo_dirty;
2283 	else
2284 		bv = &bp->b_bufobj->bo_clean;
2285 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2286 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2287 	bv->bv_cnt--;
2288 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2289 }
2290 
2291 /*
2292  * Add the buffer to the sorted clean or dirty block list.
2293  *
2294  * NOTE: xflags is passed as a constant, optimizing this inline function!
2295  */
2296 static void
2297 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2298 {
2299 	struct bufv *bv;
2300 	struct buf *n;
2301 	int error;
2302 
2303 	ASSERT_BO_WLOCKED(bo);
2304 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2305 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2306 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2307 	    ("dead bo %p", bo));
2308 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2309 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2310 	bp->b_xflags |= xflags;
2311 	if (xflags & BX_VNDIRTY)
2312 		bv = &bo->bo_dirty;
2313 	else
2314 		bv = &bo->bo_clean;
2315 
2316 	/*
2317 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2318 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2319 	 * than _ge.
2320 	 */
2321 	if (bv->bv_cnt == 0 ||
2322 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2323 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2324 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2325 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2326 	else
2327 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2328 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2329 	if (error)
2330 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2331 	bv->bv_cnt++;
2332 }
2333 
2334 /*
2335  * Look up a buffer using the buffer tries.
2336  */
2337 struct buf *
2338 gbincore(struct bufobj *bo, daddr_t lblkno)
2339 {
2340 	struct buf *bp;
2341 
2342 	ASSERT_BO_LOCKED(bo);
2343 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2344 	if (bp != NULL)
2345 		return (bp);
2346 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2347 }
2348 
2349 /*
2350  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2351  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2352  * stability of the result.  Like other lockless lookups, the found buf may
2353  * already be invalid by the time this function returns.
2354  */
2355 struct buf *
2356 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2357 {
2358 	struct buf *bp;
2359 
2360 	ASSERT_BO_UNLOCKED(bo);
2361 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2362 	if (bp != NULL)
2363 		return (bp);
2364 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2365 }
2366 
2367 /*
2368  * Associate a buffer with a vnode.
2369  */
2370 void
2371 bgetvp(struct vnode *vp, struct buf *bp)
2372 {
2373 	struct bufobj *bo;
2374 
2375 	bo = &vp->v_bufobj;
2376 	ASSERT_BO_WLOCKED(bo);
2377 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2378 
2379 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2380 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2381 	    ("bgetvp: bp already attached! %p", bp));
2382 
2383 	vhold(vp);
2384 	bp->b_vp = vp;
2385 	bp->b_bufobj = bo;
2386 	/*
2387 	 * Insert onto list for new vnode.
2388 	 */
2389 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2390 }
2391 
2392 /*
2393  * Disassociate a buffer from a vnode.
2394  */
2395 void
2396 brelvp(struct buf *bp)
2397 {
2398 	struct bufobj *bo;
2399 	struct vnode *vp;
2400 
2401 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2402 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2403 
2404 	/*
2405 	 * Delete from old vnode list, if on one.
2406 	 */
2407 	vp = bp->b_vp;		/* XXX */
2408 	bo = bp->b_bufobj;
2409 	BO_LOCK(bo);
2410 	buf_vlist_remove(bp);
2411 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2412 		bo->bo_flag &= ~BO_ONWORKLST;
2413 		mtx_lock(&sync_mtx);
2414 		LIST_REMOVE(bo, bo_synclist);
2415 		syncer_worklist_len--;
2416 		mtx_unlock(&sync_mtx);
2417 	}
2418 	bp->b_vp = NULL;
2419 	bp->b_bufobj = NULL;
2420 	BO_UNLOCK(bo);
2421 	vdrop(vp);
2422 }
2423 
2424 /*
2425  * Add an item to the syncer work queue.
2426  */
2427 static void
2428 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2429 {
2430 	int slot;
2431 
2432 	ASSERT_BO_WLOCKED(bo);
2433 
2434 	mtx_lock(&sync_mtx);
2435 	if (bo->bo_flag & BO_ONWORKLST)
2436 		LIST_REMOVE(bo, bo_synclist);
2437 	else {
2438 		bo->bo_flag |= BO_ONWORKLST;
2439 		syncer_worklist_len++;
2440 	}
2441 
2442 	if (delay > syncer_maxdelay - 2)
2443 		delay = syncer_maxdelay - 2;
2444 	slot = (syncer_delayno + delay) & syncer_mask;
2445 
2446 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2447 	mtx_unlock(&sync_mtx);
2448 }
2449 
2450 static int
2451 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2452 {
2453 	int error, len;
2454 
2455 	mtx_lock(&sync_mtx);
2456 	len = syncer_worklist_len - sync_vnode_count;
2457 	mtx_unlock(&sync_mtx);
2458 	error = SYSCTL_OUT(req, &len, sizeof(len));
2459 	return (error);
2460 }
2461 
2462 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2463     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2464     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2465 
2466 static struct proc *updateproc;
2467 static void sched_sync(void);
2468 static struct kproc_desc up_kp = {
2469 	"syncer",
2470 	sched_sync,
2471 	&updateproc
2472 };
2473 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2474 
2475 static int
2476 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2477 {
2478 	struct vnode *vp;
2479 	struct mount *mp;
2480 
2481 	*bo = LIST_FIRST(slp);
2482 	if (*bo == NULL)
2483 		return (0);
2484 	vp = bo2vnode(*bo);
2485 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2486 		return (1);
2487 	/*
2488 	 * We use vhold in case the vnode does not
2489 	 * successfully sync.  vhold prevents the vnode from
2490 	 * going away when we unlock the sync_mtx so that
2491 	 * we can acquire the vnode interlock.
2492 	 */
2493 	vholdl(vp);
2494 	mtx_unlock(&sync_mtx);
2495 	VI_UNLOCK(vp);
2496 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2497 		vdrop(vp);
2498 		mtx_lock(&sync_mtx);
2499 		return (*bo == LIST_FIRST(slp));
2500 	}
2501 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2502 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2503 	VOP_UNLOCK(vp);
2504 	vn_finished_write(mp);
2505 	BO_LOCK(*bo);
2506 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2507 		/*
2508 		 * Put us back on the worklist.  The worklist
2509 		 * routine will remove us from our current
2510 		 * position and then add us back in at a later
2511 		 * position.
2512 		 */
2513 		vn_syncer_add_to_worklist(*bo, syncdelay);
2514 	}
2515 	BO_UNLOCK(*bo);
2516 	vdrop(vp);
2517 	mtx_lock(&sync_mtx);
2518 	return (0);
2519 }
2520 
2521 static int first_printf = 1;
2522 
2523 /*
2524  * System filesystem synchronizer daemon.
2525  */
2526 static void
2527 sched_sync(void)
2528 {
2529 	struct synclist *next, *slp;
2530 	struct bufobj *bo;
2531 	long starttime;
2532 	struct thread *td = curthread;
2533 	int last_work_seen;
2534 	int net_worklist_len;
2535 	int syncer_final_iter;
2536 	int error;
2537 
2538 	last_work_seen = 0;
2539 	syncer_final_iter = 0;
2540 	syncer_state = SYNCER_RUNNING;
2541 	starttime = time_uptime;
2542 	td->td_pflags |= TDP_NORUNNINGBUF;
2543 
2544 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2545 	    SHUTDOWN_PRI_LAST);
2546 
2547 	mtx_lock(&sync_mtx);
2548 	for (;;) {
2549 		if (syncer_state == SYNCER_FINAL_DELAY &&
2550 		    syncer_final_iter == 0) {
2551 			mtx_unlock(&sync_mtx);
2552 			kproc_suspend_check(td->td_proc);
2553 			mtx_lock(&sync_mtx);
2554 		}
2555 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2556 		if (syncer_state != SYNCER_RUNNING &&
2557 		    starttime != time_uptime) {
2558 			if (first_printf) {
2559 				printf("\nSyncing disks, vnodes remaining... ");
2560 				first_printf = 0;
2561 			}
2562 			printf("%d ", net_worklist_len);
2563 		}
2564 		starttime = time_uptime;
2565 
2566 		/*
2567 		 * Push files whose dirty time has expired.  Be careful
2568 		 * of interrupt race on slp queue.
2569 		 *
2570 		 * Skip over empty worklist slots when shutting down.
2571 		 */
2572 		do {
2573 			slp = &syncer_workitem_pending[syncer_delayno];
2574 			syncer_delayno += 1;
2575 			if (syncer_delayno == syncer_maxdelay)
2576 				syncer_delayno = 0;
2577 			next = &syncer_workitem_pending[syncer_delayno];
2578 			/*
2579 			 * If the worklist has wrapped since the
2580 			 * it was emptied of all but syncer vnodes,
2581 			 * switch to the FINAL_DELAY state and run
2582 			 * for one more second.
2583 			 */
2584 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2585 			    net_worklist_len == 0 &&
2586 			    last_work_seen == syncer_delayno) {
2587 				syncer_state = SYNCER_FINAL_DELAY;
2588 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2589 			}
2590 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2591 		    syncer_worklist_len > 0);
2592 
2593 		/*
2594 		 * Keep track of the last time there was anything
2595 		 * on the worklist other than syncer vnodes.
2596 		 * Return to the SHUTTING_DOWN state if any
2597 		 * new work appears.
2598 		 */
2599 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2600 			last_work_seen = syncer_delayno;
2601 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2602 			syncer_state = SYNCER_SHUTTING_DOWN;
2603 		while (!LIST_EMPTY(slp)) {
2604 			error = sync_vnode(slp, &bo, td);
2605 			if (error == 1) {
2606 				LIST_REMOVE(bo, bo_synclist);
2607 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2608 				continue;
2609 			}
2610 
2611 			if (first_printf == 0) {
2612 				/*
2613 				 * Drop the sync mutex, because some watchdog
2614 				 * drivers need to sleep while patting
2615 				 */
2616 				mtx_unlock(&sync_mtx);
2617 				wdog_kern_pat(WD_LASTVAL);
2618 				mtx_lock(&sync_mtx);
2619 			}
2620 		}
2621 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2622 			syncer_final_iter--;
2623 		/*
2624 		 * The variable rushjob allows the kernel to speed up the
2625 		 * processing of the filesystem syncer process. A rushjob
2626 		 * value of N tells the filesystem syncer to process the next
2627 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2628 		 * is used by the soft update code to speed up the filesystem
2629 		 * syncer process when the incore state is getting so far
2630 		 * ahead of the disk that the kernel memory pool is being
2631 		 * threatened with exhaustion.
2632 		 */
2633 		if (rushjob > 0) {
2634 			rushjob -= 1;
2635 			continue;
2636 		}
2637 		/*
2638 		 * Just sleep for a short period of time between
2639 		 * iterations when shutting down to allow some I/O
2640 		 * to happen.
2641 		 *
2642 		 * If it has taken us less than a second to process the
2643 		 * current work, then wait. Otherwise start right over
2644 		 * again. We can still lose time if any single round
2645 		 * takes more than two seconds, but it does not really
2646 		 * matter as we are just trying to generally pace the
2647 		 * filesystem activity.
2648 		 */
2649 		if (syncer_state != SYNCER_RUNNING ||
2650 		    time_uptime == starttime) {
2651 			thread_lock(td);
2652 			sched_prio(td, PPAUSE);
2653 			thread_unlock(td);
2654 		}
2655 		if (syncer_state != SYNCER_RUNNING)
2656 			cv_timedwait(&sync_wakeup, &sync_mtx,
2657 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2658 		else if (time_uptime == starttime)
2659 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2660 	}
2661 }
2662 
2663 /*
2664  * Request the syncer daemon to speed up its work.
2665  * We never push it to speed up more than half of its
2666  * normal turn time, otherwise it could take over the cpu.
2667  */
2668 int
2669 speedup_syncer(void)
2670 {
2671 	int ret = 0;
2672 
2673 	mtx_lock(&sync_mtx);
2674 	if (rushjob < syncdelay / 2) {
2675 		rushjob += 1;
2676 		stat_rush_requests += 1;
2677 		ret = 1;
2678 	}
2679 	mtx_unlock(&sync_mtx);
2680 	cv_broadcast(&sync_wakeup);
2681 	return (ret);
2682 }
2683 
2684 /*
2685  * Tell the syncer to speed up its work and run though its work
2686  * list several times, then tell it to shut down.
2687  */
2688 static void
2689 syncer_shutdown(void *arg, int howto)
2690 {
2691 
2692 	if (howto & RB_NOSYNC)
2693 		return;
2694 	mtx_lock(&sync_mtx);
2695 	syncer_state = SYNCER_SHUTTING_DOWN;
2696 	rushjob = 0;
2697 	mtx_unlock(&sync_mtx);
2698 	cv_broadcast(&sync_wakeup);
2699 	kproc_shutdown(arg, howto);
2700 }
2701 
2702 void
2703 syncer_suspend(void)
2704 {
2705 
2706 	syncer_shutdown(updateproc, 0);
2707 }
2708 
2709 void
2710 syncer_resume(void)
2711 {
2712 
2713 	mtx_lock(&sync_mtx);
2714 	first_printf = 1;
2715 	syncer_state = SYNCER_RUNNING;
2716 	mtx_unlock(&sync_mtx);
2717 	cv_broadcast(&sync_wakeup);
2718 	kproc_resume(updateproc);
2719 }
2720 
2721 /*
2722  * Move the buffer between the clean and dirty lists of its vnode.
2723  */
2724 void
2725 reassignbuf(struct buf *bp)
2726 {
2727 	struct vnode *vp;
2728 	struct bufobj *bo;
2729 	int delay;
2730 #ifdef INVARIANTS
2731 	struct bufv *bv;
2732 #endif
2733 
2734 	vp = bp->b_vp;
2735 	bo = bp->b_bufobj;
2736 
2737 	KASSERT((bp->b_flags & B_PAGING) == 0,
2738 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2739 
2740 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2741 	    bp, bp->b_vp, bp->b_flags);
2742 
2743 	BO_LOCK(bo);
2744 	buf_vlist_remove(bp);
2745 
2746 	/*
2747 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2748 	 * of clean buffers.
2749 	 */
2750 	if (bp->b_flags & B_DELWRI) {
2751 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2752 			switch (vp->v_type) {
2753 			case VDIR:
2754 				delay = dirdelay;
2755 				break;
2756 			case VCHR:
2757 				delay = metadelay;
2758 				break;
2759 			default:
2760 				delay = filedelay;
2761 			}
2762 			vn_syncer_add_to_worklist(bo, delay);
2763 		}
2764 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2765 	} else {
2766 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2767 
2768 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2769 			mtx_lock(&sync_mtx);
2770 			LIST_REMOVE(bo, bo_synclist);
2771 			syncer_worklist_len--;
2772 			mtx_unlock(&sync_mtx);
2773 			bo->bo_flag &= ~BO_ONWORKLST;
2774 		}
2775 	}
2776 #ifdef INVARIANTS
2777 	bv = &bo->bo_clean;
2778 	bp = TAILQ_FIRST(&bv->bv_hd);
2779 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2780 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2781 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2782 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2783 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2784 	bv = &bo->bo_dirty;
2785 	bp = TAILQ_FIRST(&bv->bv_hd);
2786 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2787 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2788 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2789 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2790 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2791 #endif
2792 	BO_UNLOCK(bo);
2793 }
2794 
2795 static void
2796 v_init_counters(struct vnode *vp)
2797 {
2798 
2799 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2800 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2801 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2802 
2803 	refcount_init(&vp->v_holdcnt, 1);
2804 	refcount_init(&vp->v_usecount, 1);
2805 }
2806 
2807 /*
2808  * Grab a particular vnode from the free list, increment its
2809  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2810  * is being destroyed.  Only callers who specify LK_RETRY will
2811  * see doomed vnodes.  If inactive processing was delayed in
2812  * vput try to do it here.
2813  *
2814  * usecount is manipulated using atomics without holding any locks.
2815  *
2816  * holdcnt can be manipulated using atomics without holding any locks,
2817  * except when transitioning 1<->0, in which case the interlock is held.
2818  *
2819  * Consumers which don't guarantee liveness of the vnode can use SMR to
2820  * try to get a reference. Note this operation can fail since the vnode
2821  * may be awaiting getting freed by the time they get to it.
2822  */
2823 enum vgetstate
2824 vget_prep_smr(struct vnode *vp)
2825 {
2826 	enum vgetstate vs;
2827 
2828 	VFS_SMR_ASSERT_ENTERED();
2829 
2830 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2831 		vs = VGET_USECOUNT;
2832 	} else {
2833 		if (vhold_smr(vp))
2834 			vs = VGET_HOLDCNT;
2835 		else
2836 			vs = VGET_NONE;
2837 	}
2838 	return (vs);
2839 }
2840 
2841 enum vgetstate
2842 vget_prep(struct vnode *vp)
2843 {
2844 	enum vgetstate vs;
2845 
2846 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2847 		vs = VGET_USECOUNT;
2848 	} else {
2849 		vhold(vp);
2850 		vs = VGET_HOLDCNT;
2851 	}
2852 	return (vs);
2853 }
2854 
2855 void
2856 vget_abort(struct vnode *vp, enum vgetstate vs)
2857 {
2858 
2859 	switch (vs) {
2860 	case VGET_USECOUNT:
2861 		vrele(vp);
2862 		break;
2863 	case VGET_HOLDCNT:
2864 		vdrop(vp);
2865 		break;
2866 	default:
2867 		__assert_unreachable();
2868 	}
2869 }
2870 
2871 int
2872 vget(struct vnode *vp, int flags)
2873 {
2874 	enum vgetstate vs;
2875 
2876 	vs = vget_prep(vp);
2877 	return (vget_finish(vp, flags, vs));
2878 }
2879 
2880 int
2881 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2882 {
2883 	int error;
2884 
2885 	if ((flags & LK_INTERLOCK) != 0)
2886 		ASSERT_VI_LOCKED(vp, __func__);
2887 	else
2888 		ASSERT_VI_UNLOCKED(vp, __func__);
2889 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2890 	VNPASS(vp->v_holdcnt > 0, vp);
2891 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2892 
2893 	error = vn_lock(vp, flags);
2894 	if (__predict_false(error != 0)) {
2895 		vget_abort(vp, vs);
2896 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2897 		    vp);
2898 		return (error);
2899 	}
2900 
2901 	vget_finish_ref(vp, vs);
2902 	return (0);
2903 }
2904 
2905 void
2906 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
2907 {
2908 	int old;
2909 
2910 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2911 	VNPASS(vp->v_holdcnt > 0, vp);
2912 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2913 
2914 	if (vs == VGET_USECOUNT)
2915 		return;
2916 
2917 	/*
2918 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2919 	 * the vnode around. Otherwise someone else lended their hold count and
2920 	 * we have to drop ours.
2921 	 */
2922 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2923 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2924 	if (old != 0) {
2925 #ifdef INVARIANTS
2926 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2927 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2928 #else
2929 		refcount_release(&vp->v_holdcnt);
2930 #endif
2931 	}
2932 }
2933 
2934 void
2935 vref(struct vnode *vp)
2936 {
2937 	enum vgetstate vs;
2938 
2939 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2940 	vs = vget_prep(vp);
2941 	vget_finish_ref(vp, vs);
2942 }
2943 
2944 void
2945 vrefact(struct vnode *vp)
2946 {
2947 
2948 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2949 #ifdef INVARIANTS
2950 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2951 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
2952 #else
2953 	refcount_acquire(&vp->v_usecount);
2954 #endif
2955 }
2956 
2957 void
2958 vlazy(struct vnode *vp)
2959 {
2960 	struct mount *mp;
2961 
2962 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2963 
2964 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
2965 		return;
2966 	/*
2967 	 * We may get here for inactive routines after the vnode got doomed.
2968 	 */
2969 	if (VN_IS_DOOMED(vp))
2970 		return;
2971 	mp = vp->v_mount;
2972 	mtx_lock(&mp->mnt_listmtx);
2973 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
2974 		vp->v_mflag |= VMP_LAZYLIST;
2975 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
2976 		mp->mnt_lazyvnodelistsize++;
2977 	}
2978 	mtx_unlock(&mp->mnt_listmtx);
2979 }
2980 
2981 /*
2982  * This routine is only meant to be called from vgonel prior to dooming
2983  * the vnode.
2984  */
2985 static void
2986 vunlazy_gone(struct vnode *vp)
2987 {
2988 	struct mount *mp;
2989 
2990 	ASSERT_VOP_ELOCKED(vp, __func__);
2991 	ASSERT_VI_LOCKED(vp, __func__);
2992 	VNPASS(!VN_IS_DOOMED(vp), vp);
2993 
2994 	if (vp->v_mflag & VMP_LAZYLIST) {
2995 		mp = vp->v_mount;
2996 		mtx_lock(&mp->mnt_listmtx);
2997 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
2998 		vp->v_mflag &= ~VMP_LAZYLIST;
2999 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3000 		mp->mnt_lazyvnodelistsize--;
3001 		mtx_unlock(&mp->mnt_listmtx);
3002 	}
3003 }
3004 
3005 static void
3006 vdefer_inactive(struct vnode *vp)
3007 {
3008 
3009 	ASSERT_VI_LOCKED(vp, __func__);
3010 	VNASSERT(vp->v_holdcnt > 0, vp,
3011 	    ("%s: vnode without hold count", __func__));
3012 	if (VN_IS_DOOMED(vp)) {
3013 		vdropl(vp);
3014 		return;
3015 	}
3016 	if (vp->v_iflag & VI_DEFINACT) {
3017 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3018 		vdropl(vp);
3019 		return;
3020 	}
3021 	if (vp->v_usecount > 0) {
3022 		vp->v_iflag &= ~VI_OWEINACT;
3023 		vdropl(vp);
3024 		return;
3025 	}
3026 	vlazy(vp);
3027 	vp->v_iflag |= VI_DEFINACT;
3028 	VI_UNLOCK(vp);
3029 	counter_u64_add(deferred_inact, 1);
3030 }
3031 
3032 static void
3033 vdefer_inactive_unlocked(struct vnode *vp)
3034 {
3035 
3036 	VI_LOCK(vp);
3037 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3038 		vdropl(vp);
3039 		return;
3040 	}
3041 	vdefer_inactive(vp);
3042 }
3043 
3044 enum vput_op { VRELE, VPUT, VUNREF };
3045 
3046 /*
3047  * Handle ->v_usecount transitioning to 0.
3048  *
3049  * By releasing the last usecount we take ownership of the hold count which
3050  * provides liveness of the vnode, meaning we have to vdrop.
3051  *
3052  * For all vnodes we may need to perform inactive processing. It requires an
3053  * exclusive lock on the vnode, while it is legal to call here with only a
3054  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3055  * inactive processing gets deferred to the syncer.
3056  *
3057  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3058  * on the lock being held all the way until VOP_INACTIVE. This in particular
3059  * happens with UFS which adds half-constructed vnodes to the hash, where they
3060  * can be found by other code.
3061  */
3062 static void
3063 vput_final(struct vnode *vp, enum vput_op func)
3064 {
3065 	int error;
3066 	bool want_unlock;
3067 
3068 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3069 	VNPASS(vp->v_holdcnt > 0, vp);
3070 
3071 	VI_LOCK(vp);
3072 
3073 	/*
3074 	 * By the time we got here someone else might have transitioned
3075 	 * the count back to > 0.
3076 	 */
3077 	if (vp->v_usecount > 0)
3078 		goto out;
3079 
3080 	/*
3081 	 * If the vnode is doomed vgone already performed inactive processing
3082 	 * (if needed).
3083 	 */
3084 	if (VN_IS_DOOMED(vp))
3085 		goto out;
3086 
3087 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3088 		goto out;
3089 
3090 	if (vp->v_iflag & VI_DOINGINACT)
3091 		goto out;
3092 
3093 	/*
3094 	 * Locking operations here will drop the interlock and possibly the
3095 	 * vnode lock, opening a window where the vnode can get doomed all the
3096 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3097 	 * perform inactive.
3098 	 */
3099 	vp->v_iflag |= VI_OWEINACT;
3100 	want_unlock = false;
3101 	error = 0;
3102 	switch (func) {
3103 	case VRELE:
3104 		switch (VOP_ISLOCKED(vp)) {
3105 		case LK_EXCLUSIVE:
3106 			break;
3107 		case LK_EXCLOTHER:
3108 		case 0:
3109 			want_unlock = true;
3110 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3111 			VI_LOCK(vp);
3112 			break;
3113 		default:
3114 			/*
3115 			 * The lock has at least one sharer, but we have no way
3116 			 * to conclude whether this is us. Play it safe and
3117 			 * defer processing.
3118 			 */
3119 			error = EAGAIN;
3120 			break;
3121 		}
3122 		break;
3123 	case VPUT:
3124 		want_unlock = true;
3125 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3126 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3127 			    LK_NOWAIT);
3128 			VI_LOCK(vp);
3129 		}
3130 		break;
3131 	case VUNREF:
3132 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3133 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3134 			VI_LOCK(vp);
3135 		}
3136 		break;
3137 	}
3138 	if (error == 0) {
3139 		vinactive(vp);
3140 		if (want_unlock)
3141 			VOP_UNLOCK(vp);
3142 		vdropl(vp);
3143 	} else {
3144 		vdefer_inactive(vp);
3145 	}
3146 	return;
3147 out:
3148 	if (func == VPUT)
3149 		VOP_UNLOCK(vp);
3150 	vdropl(vp);
3151 }
3152 
3153 /*
3154  * Decrement ->v_usecount for a vnode.
3155  *
3156  * Releasing the last use count requires additional processing, see vput_final
3157  * above for details.
3158  *
3159  * Comment above each variant denotes lock state on entry and exit.
3160  */
3161 
3162 /*
3163  * in: any
3164  * out: same as passed in
3165  */
3166 void
3167 vrele(struct vnode *vp)
3168 {
3169 
3170 	ASSERT_VI_UNLOCKED(vp, __func__);
3171 	if (!refcount_release(&vp->v_usecount))
3172 		return;
3173 	vput_final(vp, VRELE);
3174 }
3175 
3176 /*
3177  * in: locked
3178  * out: unlocked
3179  */
3180 void
3181 vput(struct vnode *vp)
3182 {
3183 
3184 	ASSERT_VOP_LOCKED(vp, __func__);
3185 	ASSERT_VI_UNLOCKED(vp, __func__);
3186 	if (!refcount_release(&vp->v_usecount)) {
3187 		VOP_UNLOCK(vp);
3188 		return;
3189 	}
3190 	vput_final(vp, VPUT);
3191 }
3192 
3193 /*
3194  * in: locked
3195  * out: locked
3196  */
3197 void
3198 vunref(struct vnode *vp)
3199 {
3200 
3201 	ASSERT_VOP_LOCKED(vp, __func__);
3202 	ASSERT_VI_UNLOCKED(vp, __func__);
3203 	if (!refcount_release(&vp->v_usecount))
3204 		return;
3205 	vput_final(vp, VUNREF);
3206 }
3207 
3208 void
3209 vhold(struct vnode *vp)
3210 {
3211 	int old;
3212 
3213 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3214 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3215 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3216 	    ("%s: wrong hold count %d", __func__, old));
3217 	if (old == 0)
3218 		vn_freevnodes_dec();
3219 }
3220 
3221 void
3222 vholdnz(struct vnode *vp)
3223 {
3224 
3225 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3226 #ifdef INVARIANTS
3227 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3228 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3229 	    ("%s: wrong hold count %d", __func__, old));
3230 #else
3231 	atomic_add_int(&vp->v_holdcnt, 1);
3232 #endif
3233 }
3234 
3235 /*
3236  * Grab a hold count unless the vnode is freed.
3237  *
3238  * Only use this routine if vfs smr is the only protection you have against
3239  * freeing the vnode.
3240  *
3241  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3242  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3243  * the thread which managed to set the flag.
3244  *
3245  * It may be tempting to replace the loop with:
3246  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3247  * if (count & VHOLD_NO_SMR) {
3248  *     backpedal and error out;
3249  * }
3250  *
3251  * However, while this is more performant, it hinders debugging by eliminating
3252  * the previously mentioned invariant.
3253  */
3254 static bool __always_inline
3255 _vhold_cond(struct vnode *vp)
3256 {
3257 	int count;
3258 
3259 	count = atomic_load_int(&vp->v_holdcnt);
3260 	for (;;) {
3261 		if (count & VHOLD_NO_SMR) {
3262 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3263 			    ("non-zero hold count with flags %d\n", count));
3264 			return (false);
3265 		}
3266 
3267 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3268 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3269 			if (count == 0)
3270 				vn_freevnodes_dec();
3271 			return (true);
3272 		}
3273 	}
3274 }
3275 
3276 bool
3277 vhold_smr(struct vnode *vp)
3278 {
3279 
3280 	VFS_SMR_ASSERT_ENTERED();
3281 	return (_vhold_cond(vp));
3282 }
3283 
3284 /*
3285  * Special case for vnode recycling.
3286  *
3287  * Vnodes are present on the global list until UMA takes them out.
3288  * Attempts to recycle only need the relevant lock and have no use for SMR.
3289  */
3290 static bool
3291 vhold_recycle(struct vnode *vp)
3292 {
3293 
3294 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3295 	return (_vhold_cond(vp));
3296 }
3297 
3298 static void __noinline
3299 vdbatch_process(struct vdbatch *vd)
3300 {
3301 	struct vnode *vp;
3302 	int i;
3303 
3304 	mtx_assert(&vd->lock, MA_OWNED);
3305 	MPASS(curthread->td_pinned > 0);
3306 	MPASS(vd->index == VDBATCH_SIZE);
3307 
3308 	mtx_lock(&vnode_list_mtx);
3309 	critical_enter();
3310 	freevnodes += vd->freevnodes;
3311 	for (i = 0; i < VDBATCH_SIZE; i++) {
3312 		vp = vd->tab[i];
3313 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3314 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3315 		MPASS(vp->v_dbatchcpu != NOCPU);
3316 		vp->v_dbatchcpu = NOCPU;
3317 	}
3318 	mtx_unlock(&vnode_list_mtx);
3319 	vd->freevnodes = 0;
3320 	bzero(vd->tab, sizeof(vd->tab));
3321 	vd->index = 0;
3322 	critical_exit();
3323 }
3324 
3325 static void
3326 vdbatch_enqueue(struct vnode *vp)
3327 {
3328 	struct vdbatch *vd;
3329 
3330 	ASSERT_VI_LOCKED(vp, __func__);
3331 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3332 	    ("%s: deferring requeue of a doomed vnode", __func__));
3333 
3334 	if (vp->v_dbatchcpu != NOCPU) {
3335 		VI_UNLOCK(vp);
3336 		return;
3337 	}
3338 
3339 	sched_pin();
3340 	vd = DPCPU_PTR(vd);
3341 	mtx_lock(&vd->lock);
3342 	MPASS(vd->index < VDBATCH_SIZE);
3343 	MPASS(vd->tab[vd->index] == NULL);
3344 	/*
3345 	 * A hack: we depend on being pinned so that we know what to put in
3346 	 * ->v_dbatchcpu.
3347 	 */
3348 	vp->v_dbatchcpu = curcpu;
3349 	vd->tab[vd->index] = vp;
3350 	vd->index++;
3351 	VI_UNLOCK(vp);
3352 	if (vd->index == VDBATCH_SIZE)
3353 		vdbatch_process(vd);
3354 	mtx_unlock(&vd->lock);
3355 	sched_unpin();
3356 }
3357 
3358 /*
3359  * This routine must only be called for vnodes which are about to be
3360  * deallocated. Supporting dequeue for arbitrary vndoes would require
3361  * validating that the locked batch matches.
3362  */
3363 static void
3364 vdbatch_dequeue(struct vnode *vp)
3365 {
3366 	struct vdbatch *vd;
3367 	int i;
3368 	short cpu;
3369 
3370 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3371 	    ("%s: called for a used vnode\n", __func__));
3372 
3373 	cpu = vp->v_dbatchcpu;
3374 	if (cpu == NOCPU)
3375 		return;
3376 
3377 	vd = DPCPU_ID_PTR(cpu, vd);
3378 	mtx_lock(&vd->lock);
3379 	for (i = 0; i < vd->index; i++) {
3380 		if (vd->tab[i] != vp)
3381 			continue;
3382 		vp->v_dbatchcpu = NOCPU;
3383 		vd->index--;
3384 		vd->tab[i] = vd->tab[vd->index];
3385 		vd->tab[vd->index] = NULL;
3386 		break;
3387 	}
3388 	mtx_unlock(&vd->lock);
3389 	/*
3390 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3391 	 */
3392 	MPASS(vp->v_dbatchcpu == NOCPU);
3393 }
3394 
3395 /*
3396  * Drop the hold count of the vnode.  If this is the last reference to
3397  * the vnode we place it on the free list unless it has been vgone'd
3398  * (marked VIRF_DOOMED) in which case we will free it.
3399  *
3400  * Because the vnode vm object keeps a hold reference on the vnode if
3401  * there is at least one resident non-cached page, the vnode cannot
3402  * leave the active list without the page cleanup done.
3403  */
3404 static void
3405 vdrop_deactivate(struct vnode *vp)
3406 {
3407 	struct mount *mp;
3408 
3409 	ASSERT_VI_LOCKED(vp, __func__);
3410 	/*
3411 	 * Mark a vnode as free: remove it from its active list
3412 	 * and put it up for recycling on the freelist.
3413 	 */
3414 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3415 	    ("vdrop: returning doomed vnode"));
3416 	VNASSERT(vp->v_op != NULL, vp,
3417 	    ("vdrop: vnode already reclaimed."));
3418 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3419 	    ("vnode with VI_OWEINACT set"));
3420 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3421 	    ("vnode with VI_DEFINACT set"));
3422 	if (vp->v_mflag & VMP_LAZYLIST) {
3423 		mp = vp->v_mount;
3424 		mtx_lock(&mp->mnt_listmtx);
3425 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3426 		/*
3427 		 * Don't remove the vnode from the lazy list if another thread
3428 		 * has increased the hold count. It may have re-enqueued the
3429 		 * vnode to the lazy list and is now responsible for its
3430 		 * removal.
3431 		 */
3432 		if (vp->v_holdcnt == 0) {
3433 			vp->v_mflag &= ~VMP_LAZYLIST;
3434 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3435 			mp->mnt_lazyvnodelistsize--;
3436 		}
3437 		mtx_unlock(&mp->mnt_listmtx);
3438 	}
3439 	vdbatch_enqueue(vp);
3440 }
3441 
3442 static void __noinline
3443 vdropl_final(struct vnode *vp)
3444 {
3445 
3446 	ASSERT_VI_LOCKED(vp, __func__);
3447 	VNPASS(VN_IS_DOOMED(vp), vp);
3448 	/*
3449 	 * Set the VHOLD_NO_SMR flag.
3450 	 *
3451 	 * We may be racing against vhold_smr. If they win we can just pretend
3452 	 * we never got this far, they will vdrop later.
3453 	 */
3454 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3455 		vn_freevnodes_inc();
3456 		VI_UNLOCK(vp);
3457 		/*
3458 		 * We lost the aforementioned race. Any subsequent access is
3459 		 * invalid as they might have managed to vdropl on their own.
3460 		 */
3461 		return;
3462 	}
3463 	/*
3464 	 * Don't bump freevnodes as this one is going away.
3465 	 */
3466 	freevnode(vp);
3467 }
3468 
3469 void
3470 vdrop(struct vnode *vp)
3471 {
3472 
3473 	ASSERT_VI_UNLOCKED(vp, __func__);
3474 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3475 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3476 		return;
3477 	VI_LOCK(vp);
3478 	vdropl(vp);
3479 }
3480 
3481 void
3482 vdropl(struct vnode *vp)
3483 {
3484 
3485 	ASSERT_VI_LOCKED(vp, __func__);
3486 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3487 	if (!refcount_release(&vp->v_holdcnt)) {
3488 		VI_UNLOCK(vp);
3489 		return;
3490 	}
3491 	if (!VN_IS_DOOMED(vp)) {
3492 		vn_freevnodes_inc();
3493 		vdrop_deactivate(vp);
3494 		/*
3495 		 * Also unlocks the interlock. We can't assert on it as we
3496 		 * released our hold and by now the vnode might have been
3497 		 * freed.
3498 		 */
3499 		return;
3500 	}
3501 	vdropl_final(vp);
3502 }
3503 
3504 /*
3505  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3506  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3507  */
3508 static void
3509 vinactivef(struct vnode *vp)
3510 {
3511 	struct vm_object *obj;
3512 
3513 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3514 	ASSERT_VI_LOCKED(vp, "vinactive");
3515 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3516 	    ("vinactive: recursed on VI_DOINGINACT"));
3517 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3518 	vp->v_iflag |= VI_DOINGINACT;
3519 	vp->v_iflag &= ~VI_OWEINACT;
3520 	VI_UNLOCK(vp);
3521 	/*
3522 	 * Before moving off the active list, we must be sure that any
3523 	 * modified pages are converted into the vnode's dirty
3524 	 * buffers, since these will no longer be checked once the
3525 	 * vnode is on the inactive list.
3526 	 *
3527 	 * The write-out of the dirty pages is asynchronous.  At the
3528 	 * point that VOP_INACTIVE() is called, there could still be
3529 	 * pending I/O and dirty pages in the object.
3530 	 */
3531 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3532 	    vm_object_mightbedirty(obj)) {
3533 		VM_OBJECT_WLOCK(obj);
3534 		vm_object_page_clean(obj, 0, 0, 0);
3535 		VM_OBJECT_WUNLOCK(obj);
3536 	}
3537 	VOP_INACTIVE(vp);
3538 	VI_LOCK(vp);
3539 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3540 	    ("vinactive: lost VI_DOINGINACT"));
3541 	vp->v_iflag &= ~VI_DOINGINACT;
3542 }
3543 
3544 void
3545 vinactive(struct vnode *vp)
3546 {
3547 
3548 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3549 	ASSERT_VI_LOCKED(vp, "vinactive");
3550 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3551 
3552 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3553 		return;
3554 	if (vp->v_iflag & VI_DOINGINACT)
3555 		return;
3556 	if (vp->v_usecount > 0) {
3557 		vp->v_iflag &= ~VI_OWEINACT;
3558 		return;
3559 	}
3560 	vinactivef(vp);
3561 }
3562 
3563 /*
3564  * Remove any vnodes in the vnode table belonging to mount point mp.
3565  *
3566  * If FORCECLOSE is not specified, there should not be any active ones,
3567  * return error if any are found (nb: this is a user error, not a
3568  * system error). If FORCECLOSE is specified, detach any active vnodes
3569  * that are found.
3570  *
3571  * If WRITECLOSE is set, only flush out regular file vnodes open for
3572  * writing.
3573  *
3574  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3575  *
3576  * `rootrefs' specifies the base reference count for the root vnode
3577  * of this filesystem. The root vnode is considered busy if its
3578  * v_usecount exceeds this value. On a successful return, vflush(, td)
3579  * will call vrele() on the root vnode exactly rootrefs times.
3580  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3581  * be zero.
3582  */
3583 #ifdef DIAGNOSTIC
3584 static int busyprt = 0;		/* print out busy vnodes */
3585 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3586 #endif
3587 
3588 int
3589 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3590 {
3591 	struct vnode *vp, *mvp, *rootvp = NULL;
3592 	struct vattr vattr;
3593 	int busy = 0, error;
3594 
3595 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3596 	    rootrefs, flags);
3597 	if (rootrefs > 0) {
3598 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3599 		    ("vflush: bad args"));
3600 		/*
3601 		 * Get the filesystem root vnode. We can vput() it
3602 		 * immediately, since with rootrefs > 0, it won't go away.
3603 		 */
3604 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3605 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3606 			    __func__, error);
3607 			return (error);
3608 		}
3609 		vput(rootvp);
3610 	}
3611 loop:
3612 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3613 		vholdl(vp);
3614 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3615 		if (error) {
3616 			vdrop(vp);
3617 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3618 			goto loop;
3619 		}
3620 		/*
3621 		 * Skip over a vnodes marked VV_SYSTEM.
3622 		 */
3623 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3624 			VOP_UNLOCK(vp);
3625 			vdrop(vp);
3626 			continue;
3627 		}
3628 		/*
3629 		 * If WRITECLOSE is set, flush out unlinked but still open
3630 		 * files (even if open only for reading) and regular file
3631 		 * vnodes open for writing.
3632 		 */
3633 		if (flags & WRITECLOSE) {
3634 			if (vp->v_object != NULL) {
3635 				VM_OBJECT_WLOCK(vp->v_object);
3636 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3637 				VM_OBJECT_WUNLOCK(vp->v_object);
3638 			}
3639 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3640 			if (error != 0) {
3641 				VOP_UNLOCK(vp);
3642 				vdrop(vp);
3643 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3644 				return (error);
3645 			}
3646 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3647 			VI_LOCK(vp);
3648 
3649 			if ((vp->v_type == VNON ||
3650 			    (error == 0 && vattr.va_nlink > 0)) &&
3651 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3652 				VOP_UNLOCK(vp);
3653 				vdropl(vp);
3654 				continue;
3655 			}
3656 		} else
3657 			VI_LOCK(vp);
3658 		/*
3659 		 * With v_usecount == 0, all we need to do is clear out the
3660 		 * vnode data structures and we are done.
3661 		 *
3662 		 * If FORCECLOSE is set, forcibly close the vnode.
3663 		 */
3664 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3665 			vgonel(vp);
3666 		} else {
3667 			busy++;
3668 #ifdef DIAGNOSTIC
3669 			if (busyprt)
3670 				vn_printf(vp, "vflush: busy vnode ");
3671 #endif
3672 		}
3673 		VOP_UNLOCK(vp);
3674 		vdropl(vp);
3675 	}
3676 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3677 		/*
3678 		 * If just the root vnode is busy, and if its refcount
3679 		 * is equal to `rootrefs', then go ahead and kill it.
3680 		 */
3681 		VI_LOCK(rootvp);
3682 		KASSERT(busy > 0, ("vflush: not busy"));
3683 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3684 		    ("vflush: usecount %d < rootrefs %d",
3685 		     rootvp->v_usecount, rootrefs));
3686 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3687 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3688 			vgone(rootvp);
3689 			VOP_UNLOCK(rootvp);
3690 			busy = 0;
3691 		} else
3692 			VI_UNLOCK(rootvp);
3693 	}
3694 	if (busy) {
3695 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3696 		    busy);
3697 		return (EBUSY);
3698 	}
3699 	for (; rootrefs > 0; rootrefs--)
3700 		vrele(rootvp);
3701 	return (0);
3702 }
3703 
3704 /*
3705  * Recycle an unused vnode to the front of the free list.
3706  */
3707 int
3708 vrecycle(struct vnode *vp)
3709 {
3710 	int recycled;
3711 
3712 	VI_LOCK(vp);
3713 	recycled = vrecyclel(vp);
3714 	VI_UNLOCK(vp);
3715 	return (recycled);
3716 }
3717 
3718 /*
3719  * vrecycle, with the vp interlock held.
3720  */
3721 int
3722 vrecyclel(struct vnode *vp)
3723 {
3724 	int recycled;
3725 
3726 	ASSERT_VOP_ELOCKED(vp, __func__);
3727 	ASSERT_VI_LOCKED(vp, __func__);
3728 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3729 	recycled = 0;
3730 	if (vp->v_usecount == 0) {
3731 		recycled = 1;
3732 		vgonel(vp);
3733 	}
3734 	return (recycled);
3735 }
3736 
3737 /*
3738  * Eliminate all activity associated with a vnode
3739  * in preparation for reuse.
3740  */
3741 void
3742 vgone(struct vnode *vp)
3743 {
3744 	VI_LOCK(vp);
3745 	vgonel(vp);
3746 	VI_UNLOCK(vp);
3747 }
3748 
3749 static void
3750 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3751     struct vnode *lowervp __unused)
3752 {
3753 }
3754 
3755 /*
3756  * Notify upper mounts about reclaimed or unlinked vnode.
3757  */
3758 void
3759 vfs_notify_upper(struct vnode *vp, int event)
3760 {
3761 	static struct vfsops vgonel_vfsops = {
3762 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3763 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3764 	};
3765 	struct mount *mp, *ump, *mmp;
3766 
3767 	mp = vp->v_mount;
3768 	if (mp == NULL)
3769 		return;
3770 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3771 		return;
3772 
3773 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3774 	mmp->mnt_op = &vgonel_vfsops;
3775 	mmp->mnt_kern_flag |= MNTK_MARKER;
3776 	MNT_ILOCK(mp);
3777 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3778 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3779 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3780 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3781 			continue;
3782 		}
3783 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3784 		MNT_IUNLOCK(mp);
3785 		switch (event) {
3786 		case VFS_NOTIFY_UPPER_RECLAIM:
3787 			VFS_RECLAIM_LOWERVP(ump, vp);
3788 			break;
3789 		case VFS_NOTIFY_UPPER_UNLINK:
3790 			VFS_UNLINK_LOWERVP(ump, vp);
3791 			break;
3792 		default:
3793 			KASSERT(0, ("invalid event %d", event));
3794 			break;
3795 		}
3796 		MNT_ILOCK(mp);
3797 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3798 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3799 	}
3800 	free(mmp, M_TEMP);
3801 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3802 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3803 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3804 		wakeup(&mp->mnt_uppers);
3805 	}
3806 	MNT_IUNLOCK(mp);
3807 }
3808 
3809 /*
3810  * vgone, with the vp interlock held.
3811  */
3812 static void
3813 vgonel(struct vnode *vp)
3814 {
3815 	struct thread *td;
3816 	struct mount *mp;
3817 	vm_object_t object;
3818 	bool active, doinginact, oweinact;
3819 
3820 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3821 	ASSERT_VI_LOCKED(vp, "vgonel");
3822 	VNASSERT(vp->v_holdcnt, vp,
3823 	    ("vgonel: vp %p has no reference.", vp));
3824 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3825 	td = curthread;
3826 
3827 	/*
3828 	 * Don't vgonel if we're already doomed.
3829 	 */
3830 	if (vp->v_irflag & VIRF_DOOMED)
3831 		return;
3832 	/*
3833 	 * Paired with freevnode.
3834 	 */
3835 	vn_seqc_write_begin_locked(vp);
3836 	vunlazy_gone(vp);
3837 	vp->v_irflag |= VIRF_DOOMED;
3838 
3839 	/*
3840 	 * Check to see if the vnode is in use.  If so, we have to
3841 	 * call VOP_CLOSE() and VOP_INACTIVE().
3842 	 *
3843 	 * It could be that VOP_INACTIVE() requested reclamation, in
3844 	 * which case we should avoid recursion, so check
3845 	 * VI_DOINGINACT.  This is not precise but good enough.
3846 	 */
3847 	active = vp->v_usecount > 0;
3848 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3849 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3850 
3851 	/*
3852 	 * If we need to do inactive VI_OWEINACT will be set.
3853 	 */
3854 	if (vp->v_iflag & VI_DEFINACT) {
3855 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3856 		vp->v_iflag &= ~VI_DEFINACT;
3857 		vdropl(vp);
3858 	} else {
3859 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3860 		VI_UNLOCK(vp);
3861 	}
3862 	cache_purge_vgone(vp);
3863 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3864 
3865 	/*
3866 	 * If purging an active vnode, it must be closed and
3867 	 * deactivated before being reclaimed.
3868 	 */
3869 	if (active)
3870 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3871 	if ((oweinact || active) && !doinginact) {
3872 		VI_LOCK(vp);
3873 		vinactivef(vp);
3874 		VI_UNLOCK(vp);
3875 	}
3876 	if (vp->v_type == VSOCK)
3877 		vfs_unp_reclaim(vp);
3878 
3879 	/*
3880 	 * Clean out any buffers associated with the vnode.
3881 	 * If the flush fails, just toss the buffers.
3882 	 */
3883 	mp = NULL;
3884 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3885 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3886 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3887 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3888 			;
3889 	}
3890 
3891 	BO_LOCK(&vp->v_bufobj);
3892 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3893 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3894 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3895 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3896 	    ("vp %p bufobj not invalidated", vp));
3897 
3898 	/*
3899 	 * For VMIO bufobj, BO_DEAD is set later, or in
3900 	 * vm_object_terminate() after the object's page queue is
3901 	 * flushed.
3902 	 */
3903 	object = vp->v_bufobj.bo_object;
3904 	if (object == NULL)
3905 		vp->v_bufobj.bo_flag |= BO_DEAD;
3906 	BO_UNLOCK(&vp->v_bufobj);
3907 
3908 	/*
3909 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3910 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3911 	 * should not touch the object borrowed from the lower vnode
3912 	 * (the handle check).
3913 	 */
3914 	if (object != NULL && object->type == OBJT_VNODE &&
3915 	    object->handle == vp)
3916 		vnode_destroy_vobject(vp);
3917 
3918 	/*
3919 	 * Reclaim the vnode.
3920 	 */
3921 	if (VOP_RECLAIM(vp))
3922 		panic("vgone: cannot reclaim");
3923 	if (mp != NULL)
3924 		vn_finished_secondary_write(mp);
3925 	VNASSERT(vp->v_object == NULL, vp,
3926 	    ("vop_reclaim left v_object vp=%p", vp));
3927 	/*
3928 	 * Clear the advisory locks and wake up waiting threads.
3929 	 */
3930 	(void)VOP_ADVLOCKPURGE(vp);
3931 	vp->v_lockf = NULL;
3932 	/*
3933 	 * Delete from old mount point vnode list.
3934 	 */
3935 	delmntque(vp);
3936 	/*
3937 	 * Done with purge, reset to the standard lock and invalidate
3938 	 * the vnode.
3939 	 */
3940 	VI_LOCK(vp);
3941 	vp->v_vnlock = &vp->v_lock;
3942 	vp->v_op = &dead_vnodeops;
3943 	vp->v_type = VBAD;
3944 }
3945 
3946 /*
3947  * Print out a description of a vnode.
3948  */
3949 static const char * const typename[] =
3950 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3951  "VMARKER"};
3952 
3953 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
3954     "new hold count flag not added to vn_printf");
3955 
3956 void
3957 vn_printf(struct vnode *vp, const char *fmt, ...)
3958 {
3959 	va_list ap;
3960 	char buf[256], buf2[16];
3961 	u_long flags;
3962 	u_int holdcnt;
3963 
3964 	va_start(ap, fmt);
3965 	vprintf(fmt, ap);
3966 	va_end(ap);
3967 	printf("%p: ", (void *)vp);
3968 	printf("type %s\n", typename[vp->v_type]);
3969 	holdcnt = atomic_load_int(&vp->v_holdcnt);
3970 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
3971 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
3972 	    vp->v_seqc_users);
3973 	switch (vp->v_type) {
3974 	case VDIR:
3975 		printf(" mountedhere %p\n", vp->v_mountedhere);
3976 		break;
3977 	case VCHR:
3978 		printf(" rdev %p\n", vp->v_rdev);
3979 		break;
3980 	case VSOCK:
3981 		printf(" socket %p\n", vp->v_unpcb);
3982 		break;
3983 	case VFIFO:
3984 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3985 		break;
3986 	default:
3987 		printf("\n");
3988 		break;
3989 	}
3990 	buf[0] = '\0';
3991 	buf[1] = '\0';
3992 	if (holdcnt & VHOLD_NO_SMR)
3993 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
3994 	printf("    hold count flags (%s)\n", buf + 1);
3995 
3996 	buf[0] = '\0';
3997 	buf[1] = '\0';
3998 	if (vp->v_irflag & VIRF_DOOMED)
3999 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4000 	if (vp->v_irflag & VIRF_PGREAD)
4001 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4002 	flags = vp->v_irflag & ~(VIRF_DOOMED | VIRF_PGREAD);
4003 	if (flags != 0) {
4004 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4005 		strlcat(buf, buf2, sizeof(buf));
4006 	}
4007 	if (vp->v_vflag & VV_ROOT)
4008 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4009 	if (vp->v_vflag & VV_ISTTY)
4010 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4011 	if (vp->v_vflag & VV_NOSYNC)
4012 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4013 	if (vp->v_vflag & VV_ETERNALDEV)
4014 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4015 	if (vp->v_vflag & VV_CACHEDLABEL)
4016 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4017 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4018 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4019 	if (vp->v_vflag & VV_COPYONWRITE)
4020 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4021 	if (vp->v_vflag & VV_SYSTEM)
4022 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4023 	if (vp->v_vflag & VV_PROCDEP)
4024 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4025 	if (vp->v_vflag & VV_NOKNOTE)
4026 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4027 	if (vp->v_vflag & VV_DELETED)
4028 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4029 	if (vp->v_vflag & VV_MD)
4030 		strlcat(buf, "|VV_MD", sizeof(buf));
4031 	if (vp->v_vflag & VV_FORCEINSMQ)
4032 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4033 	if (vp->v_vflag & VV_READLINK)
4034 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4035 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4036 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4037 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4038 	if (flags != 0) {
4039 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4040 		strlcat(buf, buf2, sizeof(buf));
4041 	}
4042 	if (vp->v_iflag & VI_TEXT_REF)
4043 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4044 	if (vp->v_iflag & VI_MOUNT)
4045 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4046 	if (vp->v_iflag & VI_DOINGINACT)
4047 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4048 	if (vp->v_iflag & VI_OWEINACT)
4049 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4050 	if (vp->v_iflag & VI_DEFINACT)
4051 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4052 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4053 	    VI_OWEINACT | VI_DEFINACT);
4054 	if (flags != 0) {
4055 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4056 		strlcat(buf, buf2, sizeof(buf));
4057 	}
4058 	if (vp->v_mflag & VMP_LAZYLIST)
4059 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4060 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4061 	if (flags != 0) {
4062 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4063 		strlcat(buf, buf2, sizeof(buf));
4064 	}
4065 	printf("    flags (%s)\n", buf + 1);
4066 	if (mtx_owned(VI_MTX(vp)))
4067 		printf(" VI_LOCKed");
4068 	if (vp->v_object != NULL)
4069 		printf("    v_object %p ref %d pages %d "
4070 		    "cleanbuf %d dirtybuf %d\n",
4071 		    vp->v_object, vp->v_object->ref_count,
4072 		    vp->v_object->resident_page_count,
4073 		    vp->v_bufobj.bo_clean.bv_cnt,
4074 		    vp->v_bufobj.bo_dirty.bv_cnt);
4075 	printf("    ");
4076 	lockmgr_printinfo(vp->v_vnlock);
4077 	if (vp->v_data != NULL)
4078 		VOP_PRINT(vp);
4079 }
4080 
4081 #ifdef DDB
4082 /*
4083  * List all of the locked vnodes in the system.
4084  * Called when debugging the kernel.
4085  */
4086 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4087 {
4088 	struct mount *mp;
4089 	struct vnode *vp;
4090 
4091 	/*
4092 	 * Note: because this is DDB, we can't obey the locking semantics
4093 	 * for these structures, which means we could catch an inconsistent
4094 	 * state and dereference a nasty pointer.  Not much to be done
4095 	 * about that.
4096 	 */
4097 	db_printf("Locked vnodes\n");
4098 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4099 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4100 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4101 				vn_printf(vp, "vnode ");
4102 		}
4103 	}
4104 }
4105 
4106 /*
4107  * Show details about the given vnode.
4108  */
4109 DB_SHOW_COMMAND(vnode, db_show_vnode)
4110 {
4111 	struct vnode *vp;
4112 
4113 	if (!have_addr)
4114 		return;
4115 	vp = (struct vnode *)addr;
4116 	vn_printf(vp, "vnode ");
4117 }
4118 
4119 /*
4120  * Show details about the given mount point.
4121  */
4122 DB_SHOW_COMMAND(mount, db_show_mount)
4123 {
4124 	struct mount *mp;
4125 	struct vfsopt *opt;
4126 	struct statfs *sp;
4127 	struct vnode *vp;
4128 	char buf[512];
4129 	uint64_t mflags;
4130 	u_int flags;
4131 
4132 	if (!have_addr) {
4133 		/* No address given, print short info about all mount points. */
4134 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4135 			db_printf("%p %s on %s (%s)\n", mp,
4136 			    mp->mnt_stat.f_mntfromname,
4137 			    mp->mnt_stat.f_mntonname,
4138 			    mp->mnt_stat.f_fstypename);
4139 			if (db_pager_quit)
4140 				break;
4141 		}
4142 		db_printf("\nMore info: show mount <addr>\n");
4143 		return;
4144 	}
4145 
4146 	mp = (struct mount *)addr;
4147 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4148 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4149 
4150 	buf[0] = '\0';
4151 	mflags = mp->mnt_flag;
4152 #define	MNT_FLAG(flag)	do {						\
4153 	if (mflags & (flag)) {						\
4154 		if (buf[0] != '\0')					\
4155 			strlcat(buf, ", ", sizeof(buf));		\
4156 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4157 		mflags &= ~(flag);					\
4158 	}								\
4159 } while (0)
4160 	MNT_FLAG(MNT_RDONLY);
4161 	MNT_FLAG(MNT_SYNCHRONOUS);
4162 	MNT_FLAG(MNT_NOEXEC);
4163 	MNT_FLAG(MNT_NOSUID);
4164 	MNT_FLAG(MNT_NFS4ACLS);
4165 	MNT_FLAG(MNT_UNION);
4166 	MNT_FLAG(MNT_ASYNC);
4167 	MNT_FLAG(MNT_SUIDDIR);
4168 	MNT_FLAG(MNT_SOFTDEP);
4169 	MNT_FLAG(MNT_NOSYMFOLLOW);
4170 	MNT_FLAG(MNT_GJOURNAL);
4171 	MNT_FLAG(MNT_MULTILABEL);
4172 	MNT_FLAG(MNT_ACLS);
4173 	MNT_FLAG(MNT_NOATIME);
4174 	MNT_FLAG(MNT_NOCLUSTERR);
4175 	MNT_FLAG(MNT_NOCLUSTERW);
4176 	MNT_FLAG(MNT_SUJ);
4177 	MNT_FLAG(MNT_EXRDONLY);
4178 	MNT_FLAG(MNT_EXPORTED);
4179 	MNT_FLAG(MNT_DEFEXPORTED);
4180 	MNT_FLAG(MNT_EXPORTANON);
4181 	MNT_FLAG(MNT_EXKERB);
4182 	MNT_FLAG(MNT_EXPUBLIC);
4183 	MNT_FLAG(MNT_LOCAL);
4184 	MNT_FLAG(MNT_QUOTA);
4185 	MNT_FLAG(MNT_ROOTFS);
4186 	MNT_FLAG(MNT_USER);
4187 	MNT_FLAG(MNT_IGNORE);
4188 	MNT_FLAG(MNT_UPDATE);
4189 	MNT_FLAG(MNT_DELEXPORT);
4190 	MNT_FLAG(MNT_RELOAD);
4191 	MNT_FLAG(MNT_FORCE);
4192 	MNT_FLAG(MNT_SNAPSHOT);
4193 	MNT_FLAG(MNT_BYFSID);
4194 #undef MNT_FLAG
4195 	if (mflags != 0) {
4196 		if (buf[0] != '\0')
4197 			strlcat(buf, ", ", sizeof(buf));
4198 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4199 		    "0x%016jx", mflags);
4200 	}
4201 	db_printf("    mnt_flag = %s\n", buf);
4202 
4203 	buf[0] = '\0';
4204 	flags = mp->mnt_kern_flag;
4205 #define	MNT_KERN_FLAG(flag)	do {					\
4206 	if (flags & (flag)) {						\
4207 		if (buf[0] != '\0')					\
4208 			strlcat(buf, ", ", sizeof(buf));		\
4209 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4210 		flags &= ~(flag);					\
4211 	}								\
4212 } while (0)
4213 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4214 	MNT_KERN_FLAG(MNTK_ASYNC);
4215 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4216 	MNT_KERN_FLAG(MNTK_DRAINING);
4217 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4218 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4219 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4220 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4221 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4222 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4223 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4224 	MNT_KERN_FLAG(MNTK_MARKER);
4225 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4226 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4227 	MNT_KERN_FLAG(MNTK_NOASYNC);
4228 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4229 	MNT_KERN_FLAG(MNTK_MWAIT);
4230 	MNT_KERN_FLAG(MNTK_SUSPEND);
4231 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4232 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4233 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4234 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4235 #undef MNT_KERN_FLAG
4236 	if (flags != 0) {
4237 		if (buf[0] != '\0')
4238 			strlcat(buf, ", ", sizeof(buf));
4239 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4240 		    "0x%08x", flags);
4241 	}
4242 	db_printf("    mnt_kern_flag = %s\n", buf);
4243 
4244 	db_printf("    mnt_opt = ");
4245 	opt = TAILQ_FIRST(mp->mnt_opt);
4246 	if (opt != NULL) {
4247 		db_printf("%s", opt->name);
4248 		opt = TAILQ_NEXT(opt, link);
4249 		while (opt != NULL) {
4250 			db_printf(", %s", opt->name);
4251 			opt = TAILQ_NEXT(opt, link);
4252 		}
4253 	}
4254 	db_printf("\n");
4255 
4256 	sp = &mp->mnt_stat;
4257 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4258 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4259 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4260 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4261 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4262 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4263 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4264 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4265 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4266 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4267 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4268 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4269 
4270 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4271 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4272 	if (jailed(mp->mnt_cred))
4273 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4274 	db_printf(" }\n");
4275 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4276 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4277 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4278 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4279 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4280 	    mp->mnt_lazyvnodelistsize);
4281 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4282 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4283 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4284 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4285 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4286 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4287 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4288 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4289 	db_printf("    mnt_secondary_accwrites = %d\n",
4290 	    mp->mnt_secondary_accwrites);
4291 	db_printf("    mnt_gjprovider = %s\n",
4292 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4293 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4294 
4295 	db_printf("\n\nList of active vnodes\n");
4296 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4297 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4298 			vn_printf(vp, "vnode ");
4299 			if (db_pager_quit)
4300 				break;
4301 		}
4302 	}
4303 	db_printf("\n\nList of inactive vnodes\n");
4304 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4305 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4306 			vn_printf(vp, "vnode ");
4307 			if (db_pager_quit)
4308 				break;
4309 		}
4310 	}
4311 }
4312 #endif	/* DDB */
4313 
4314 /*
4315  * Fill in a struct xvfsconf based on a struct vfsconf.
4316  */
4317 static int
4318 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4319 {
4320 	struct xvfsconf xvfsp;
4321 
4322 	bzero(&xvfsp, sizeof(xvfsp));
4323 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4324 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4325 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4326 	xvfsp.vfc_flags = vfsp->vfc_flags;
4327 	/*
4328 	 * These are unused in userland, we keep them
4329 	 * to not break binary compatibility.
4330 	 */
4331 	xvfsp.vfc_vfsops = NULL;
4332 	xvfsp.vfc_next = NULL;
4333 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4334 }
4335 
4336 #ifdef COMPAT_FREEBSD32
4337 struct xvfsconf32 {
4338 	uint32_t	vfc_vfsops;
4339 	char		vfc_name[MFSNAMELEN];
4340 	int32_t		vfc_typenum;
4341 	int32_t		vfc_refcount;
4342 	int32_t		vfc_flags;
4343 	uint32_t	vfc_next;
4344 };
4345 
4346 static int
4347 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4348 {
4349 	struct xvfsconf32 xvfsp;
4350 
4351 	bzero(&xvfsp, sizeof(xvfsp));
4352 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4353 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4354 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4355 	xvfsp.vfc_flags = vfsp->vfc_flags;
4356 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4357 }
4358 #endif
4359 
4360 /*
4361  * Top level filesystem related information gathering.
4362  */
4363 static int
4364 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4365 {
4366 	struct vfsconf *vfsp;
4367 	int error;
4368 
4369 	error = 0;
4370 	vfsconf_slock();
4371 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4372 #ifdef COMPAT_FREEBSD32
4373 		if (req->flags & SCTL_MASK32)
4374 			error = vfsconf2x32(req, vfsp);
4375 		else
4376 #endif
4377 			error = vfsconf2x(req, vfsp);
4378 		if (error)
4379 			break;
4380 	}
4381 	vfsconf_sunlock();
4382 	return (error);
4383 }
4384 
4385 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4386     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4387     "S,xvfsconf", "List of all configured filesystems");
4388 
4389 #ifndef BURN_BRIDGES
4390 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4391 
4392 static int
4393 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4394 {
4395 	int *name = (int *)arg1 - 1;	/* XXX */
4396 	u_int namelen = arg2 + 1;	/* XXX */
4397 	struct vfsconf *vfsp;
4398 
4399 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4400 	    "please rebuild world\n");
4401 
4402 #if 1 || defined(COMPAT_PRELITE2)
4403 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4404 	if (namelen == 1)
4405 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4406 #endif
4407 
4408 	switch (name[1]) {
4409 	case VFS_MAXTYPENUM:
4410 		if (namelen != 2)
4411 			return (ENOTDIR);
4412 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4413 	case VFS_CONF:
4414 		if (namelen != 3)
4415 			return (ENOTDIR);	/* overloaded */
4416 		vfsconf_slock();
4417 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4418 			if (vfsp->vfc_typenum == name[2])
4419 				break;
4420 		}
4421 		vfsconf_sunlock();
4422 		if (vfsp == NULL)
4423 			return (EOPNOTSUPP);
4424 #ifdef COMPAT_FREEBSD32
4425 		if (req->flags & SCTL_MASK32)
4426 			return (vfsconf2x32(req, vfsp));
4427 		else
4428 #endif
4429 			return (vfsconf2x(req, vfsp));
4430 	}
4431 	return (EOPNOTSUPP);
4432 }
4433 
4434 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4435     CTLFLAG_MPSAFE, vfs_sysctl,
4436     "Generic filesystem");
4437 
4438 #if 1 || defined(COMPAT_PRELITE2)
4439 
4440 static int
4441 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4442 {
4443 	int error;
4444 	struct vfsconf *vfsp;
4445 	struct ovfsconf ovfs;
4446 
4447 	vfsconf_slock();
4448 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4449 		bzero(&ovfs, sizeof(ovfs));
4450 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4451 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4452 		ovfs.vfc_index = vfsp->vfc_typenum;
4453 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4454 		ovfs.vfc_flags = vfsp->vfc_flags;
4455 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4456 		if (error != 0) {
4457 			vfsconf_sunlock();
4458 			return (error);
4459 		}
4460 	}
4461 	vfsconf_sunlock();
4462 	return (0);
4463 }
4464 
4465 #endif /* 1 || COMPAT_PRELITE2 */
4466 #endif /* !BURN_BRIDGES */
4467 
4468 #define KINFO_VNODESLOP		10
4469 #ifdef notyet
4470 /*
4471  * Dump vnode list (via sysctl).
4472  */
4473 /* ARGSUSED */
4474 static int
4475 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4476 {
4477 	struct xvnode *xvn;
4478 	struct mount *mp;
4479 	struct vnode *vp;
4480 	int error, len, n;
4481 
4482 	/*
4483 	 * Stale numvnodes access is not fatal here.
4484 	 */
4485 	req->lock = 0;
4486 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4487 	if (!req->oldptr)
4488 		/* Make an estimate */
4489 		return (SYSCTL_OUT(req, 0, len));
4490 
4491 	error = sysctl_wire_old_buffer(req, 0);
4492 	if (error != 0)
4493 		return (error);
4494 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4495 	n = 0;
4496 	mtx_lock(&mountlist_mtx);
4497 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4498 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4499 			continue;
4500 		MNT_ILOCK(mp);
4501 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4502 			if (n == len)
4503 				break;
4504 			vref(vp);
4505 			xvn[n].xv_size = sizeof *xvn;
4506 			xvn[n].xv_vnode = vp;
4507 			xvn[n].xv_id = 0;	/* XXX compat */
4508 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4509 			XV_COPY(usecount);
4510 			XV_COPY(writecount);
4511 			XV_COPY(holdcnt);
4512 			XV_COPY(mount);
4513 			XV_COPY(numoutput);
4514 			XV_COPY(type);
4515 #undef XV_COPY
4516 			xvn[n].xv_flag = vp->v_vflag;
4517 
4518 			switch (vp->v_type) {
4519 			case VREG:
4520 			case VDIR:
4521 			case VLNK:
4522 				break;
4523 			case VBLK:
4524 			case VCHR:
4525 				if (vp->v_rdev == NULL) {
4526 					vrele(vp);
4527 					continue;
4528 				}
4529 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4530 				break;
4531 			case VSOCK:
4532 				xvn[n].xv_socket = vp->v_socket;
4533 				break;
4534 			case VFIFO:
4535 				xvn[n].xv_fifo = vp->v_fifoinfo;
4536 				break;
4537 			case VNON:
4538 			case VBAD:
4539 			default:
4540 				/* shouldn't happen? */
4541 				vrele(vp);
4542 				continue;
4543 			}
4544 			vrele(vp);
4545 			++n;
4546 		}
4547 		MNT_IUNLOCK(mp);
4548 		mtx_lock(&mountlist_mtx);
4549 		vfs_unbusy(mp);
4550 		if (n == len)
4551 			break;
4552 	}
4553 	mtx_unlock(&mountlist_mtx);
4554 
4555 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4556 	free(xvn, M_TEMP);
4557 	return (error);
4558 }
4559 
4560 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4561     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4562     "");
4563 #endif
4564 
4565 static void
4566 unmount_or_warn(struct mount *mp)
4567 {
4568 	int error;
4569 
4570 	error = dounmount(mp, MNT_FORCE, curthread);
4571 	if (error != 0) {
4572 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4573 		if (error == EBUSY)
4574 			printf("BUSY)\n");
4575 		else
4576 			printf("%d)\n", error);
4577 	}
4578 }
4579 
4580 /*
4581  * Unmount all filesystems. The list is traversed in reverse order
4582  * of mounting to avoid dependencies.
4583  */
4584 void
4585 vfs_unmountall(void)
4586 {
4587 	struct mount *mp, *tmp;
4588 
4589 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4590 
4591 	/*
4592 	 * Since this only runs when rebooting, it is not interlocked.
4593 	 */
4594 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4595 		vfs_ref(mp);
4596 
4597 		/*
4598 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4599 		 * unmount of the latter.
4600 		 */
4601 		if (mp == rootdevmp)
4602 			continue;
4603 
4604 		unmount_or_warn(mp);
4605 	}
4606 
4607 	if (rootdevmp != NULL)
4608 		unmount_or_warn(rootdevmp);
4609 }
4610 
4611 static void
4612 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4613 {
4614 
4615 	ASSERT_VI_LOCKED(vp, __func__);
4616 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4617 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4618 		vdropl(vp);
4619 		return;
4620 	}
4621 	if (vn_lock(vp, lkflags) == 0) {
4622 		VI_LOCK(vp);
4623 		vinactive(vp);
4624 		VOP_UNLOCK(vp);
4625 		vdropl(vp);
4626 		return;
4627 	}
4628 	vdefer_inactive_unlocked(vp);
4629 }
4630 
4631 static int
4632 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4633 {
4634 
4635 	return (vp->v_iflag & VI_DEFINACT);
4636 }
4637 
4638 static void __noinline
4639 vfs_periodic_inactive(struct mount *mp, int flags)
4640 {
4641 	struct vnode *vp, *mvp;
4642 	int lkflags;
4643 
4644 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4645 	if (flags != MNT_WAIT)
4646 		lkflags |= LK_NOWAIT;
4647 
4648 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4649 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4650 			VI_UNLOCK(vp);
4651 			continue;
4652 		}
4653 		vp->v_iflag &= ~VI_DEFINACT;
4654 		vfs_deferred_inactive(vp, lkflags);
4655 	}
4656 }
4657 
4658 static inline bool
4659 vfs_want_msync(struct vnode *vp)
4660 {
4661 	struct vm_object *obj;
4662 
4663 	/*
4664 	 * This test may be performed without any locks held.
4665 	 * We rely on vm_object's type stability.
4666 	 */
4667 	if (vp->v_vflag & VV_NOSYNC)
4668 		return (false);
4669 	obj = vp->v_object;
4670 	return (obj != NULL && vm_object_mightbedirty(obj));
4671 }
4672 
4673 static int
4674 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4675 {
4676 
4677 	if (vp->v_vflag & VV_NOSYNC)
4678 		return (false);
4679 	if (vp->v_iflag & VI_DEFINACT)
4680 		return (true);
4681 	return (vfs_want_msync(vp));
4682 }
4683 
4684 static void __noinline
4685 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4686 {
4687 	struct vnode *vp, *mvp;
4688 	struct vm_object *obj;
4689 	int lkflags, objflags;
4690 	bool seen_defer;
4691 
4692 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4693 	if (flags != MNT_WAIT) {
4694 		lkflags |= LK_NOWAIT;
4695 		objflags = OBJPC_NOSYNC;
4696 	} else {
4697 		objflags = OBJPC_SYNC;
4698 	}
4699 
4700 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4701 		seen_defer = false;
4702 		if (vp->v_iflag & VI_DEFINACT) {
4703 			vp->v_iflag &= ~VI_DEFINACT;
4704 			seen_defer = true;
4705 		}
4706 		if (!vfs_want_msync(vp)) {
4707 			if (seen_defer)
4708 				vfs_deferred_inactive(vp, lkflags);
4709 			else
4710 				VI_UNLOCK(vp);
4711 			continue;
4712 		}
4713 		if (vget(vp, lkflags) == 0) {
4714 			obj = vp->v_object;
4715 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4716 				VM_OBJECT_WLOCK(obj);
4717 				vm_object_page_clean(obj, 0, 0, objflags);
4718 				VM_OBJECT_WUNLOCK(obj);
4719 			}
4720 			vput(vp);
4721 			if (seen_defer)
4722 				vdrop(vp);
4723 		} else {
4724 			if (seen_defer)
4725 				vdefer_inactive_unlocked(vp);
4726 		}
4727 	}
4728 }
4729 
4730 void
4731 vfs_periodic(struct mount *mp, int flags)
4732 {
4733 
4734 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4735 
4736 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4737 		vfs_periodic_inactive(mp, flags);
4738 	else
4739 		vfs_periodic_msync_inactive(mp, flags);
4740 }
4741 
4742 static void
4743 destroy_vpollinfo_free(struct vpollinfo *vi)
4744 {
4745 
4746 	knlist_destroy(&vi->vpi_selinfo.si_note);
4747 	mtx_destroy(&vi->vpi_lock);
4748 	uma_zfree(vnodepoll_zone, vi);
4749 }
4750 
4751 static void
4752 destroy_vpollinfo(struct vpollinfo *vi)
4753 {
4754 
4755 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4756 	seldrain(&vi->vpi_selinfo);
4757 	destroy_vpollinfo_free(vi);
4758 }
4759 
4760 /*
4761  * Initialize per-vnode helper structure to hold poll-related state.
4762  */
4763 void
4764 v_addpollinfo(struct vnode *vp)
4765 {
4766 	struct vpollinfo *vi;
4767 
4768 	if (vp->v_pollinfo != NULL)
4769 		return;
4770 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4771 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4772 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4773 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4774 	VI_LOCK(vp);
4775 	if (vp->v_pollinfo != NULL) {
4776 		VI_UNLOCK(vp);
4777 		destroy_vpollinfo_free(vi);
4778 		return;
4779 	}
4780 	vp->v_pollinfo = vi;
4781 	VI_UNLOCK(vp);
4782 }
4783 
4784 /*
4785  * Record a process's interest in events which might happen to
4786  * a vnode.  Because poll uses the historic select-style interface
4787  * internally, this routine serves as both the ``check for any
4788  * pending events'' and the ``record my interest in future events''
4789  * functions.  (These are done together, while the lock is held,
4790  * to avoid race conditions.)
4791  */
4792 int
4793 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4794 {
4795 
4796 	v_addpollinfo(vp);
4797 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4798 	if (vp->v_pollinfo->vpi_revents & events) {
4799 		/*
4800 		 * This leaves events we are not interested
4801 		 * in available for the other process which
4802 		 * which presumably had requested them
4803 		 * (otherwise they would never have been
4804 		 * recorded).
4805 		 */
4806 		events &= vp->v_pollinfo->vpi_revents;
4807 		vp->v_pollinfo->vpi_revents &= ~events;
4808 
4809 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4810 		return (events);
4811 	}
4812 	vp->v_pollinfo->vpi_events |= events;
4813 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4814 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4815 	return (0);
4816 }
4817 
4818 /*
4819  * Routine to create and manage a filesystem syncer vnode.
4820  */
4821 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4822 static int	sync_fsync(struct  vop_fsync_args *);
4823 static int	sync_inactive(struct  vop_inactive_args *);
4824 static int	sync_reclaim(struct  vop_reclaim_args *);
4825 
4826 static struct vop_vector sync_vnodeops = {
4827 	.vop_bypass =	VOP_EOPNOTSUPP,
4828 	.vop_close =	sync_close,		/* close */
4829 	.vop_fsync =	sync_fsync,		/* fsync */
4830 	.vop_inactive =	sync_inactive,	/* inactive */
4831 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4832 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4833 	.vop_lock1 =	vop_stdlock,	/* lock */
4834 	.vop_unlock =	vop_stdunlock,	/* unlock */
4835 	.vop_islocked =	vop_stdislocked,	/* islocked */
4836 };
4837 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4838 
4839 /*
4840  * Create a new filesystem syncer vnode for the specified mount point.
4841  */
4842 void
4843 vfs_allocate_syncvnode(struct mount *mp)
4844 {
4845 	struct vnode *vp;
4846 	struct bufobj *bo;
4847 	static long start, incr, next;
4848 	int error;
4849 
4850 	/* Allocate a new vnode */
4851 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4852 	if (error != 0)
4853 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4854 	vp->v_type = VNON;
4855 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4856 	vp->v_vflag |= VV_FORCEINSMQ;
4857 	error = insmntque(vp, mp);
4858 	if (error != 0)
4859 		panic("vfs_allocate_syncvnode: insmntque() failed");
4860 	vp->v_vflag &= ~VV_FORCEINSMQ;
4861 	VOP_UNLOCK(vp);
4862 	/*
4863 	 * Place the vnode onto the syncer worklist. We attempt to
4864 	 * scatter them about on the list so that they will go off
4865 	 * at evenly distributed times even if all the filesystems
4866 	 * are mounted at once.
4867 	 */
4868 	next += incr;
4869 	if (next == 0 || next > syncer_maxdelay) {
4870 		start /= 2;
4871 		incr /= 2;
4872 		if (start == 0) {
4873 			start = syncer_maxdelay / 2;
4874 			incr = syncer_maxdelay;
4875 		}
4876 		next = start;
4877 	}
4878 	bo = &vp->v_bufobj;
4879 	BO_LOCK(bo);
4880 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4881 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4882 	mtx_lock(&sync_mtx);
4883 	sync_vnode_count++;
4884 	if (mp->mnt_syncer == NULL) {
4885 		mp->mnt_syncer = vp;
4886 		vp = NULL;
4887 	}
4888 	mtx_unlock(&sync_mtx);
4889 	BO_UNLOCK(bo);
4890 	if (vp != NULL) {
4891 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4892 		vgone(vp);
4893 		vput(vp);
4894 	}
4895 }
4896 
4897 void
4898 vfs_deallocate_syncvnode(struct mount *mp)
4899 {
4900 	struct vnode *vp;
4901 
4902 	mtx_lock(&sync_mtx);
4903 	vp = mp->mnt_syncer;
4904 	if (vp != NULL)
4905 		mp->mnt_syncer = NULL;
4906 	mtx_unlock(&sync_mtx);
4907 	if (vp != NULL)
4908 		vrele(vp);
4909 }
4910 
4911 /*
4912  * Do a lazy sync of the filesystem.
4913  */
4914 static int
4915 sync_fsync(struct vop_fsync_args *ap)
4916 {
4917 	struct vnode *syncvp = ap->a_vp;
4918 	struct mount *mp = syncvp->v_mount;
4919 	int error, save;
4920 	struct bufobj *bo;
4921 
4922 	/*
4923 	 * We only need to do something if this is a lazy evaluation.
4924 	 */
4925 	if (ap->a_waitfor != MNT_LAZY)
4926 		return (0);
4927 
4928 	/*
4929 	 * Move ourselves to the back of the sync list.
4930 	 */
4931 	bo = &syncvp->v_bufobj;
4932 	BO_LOCK(bo);
4933 	vn_syncer_add_to_worklist(bo, syncdelay);
4934 	BO_UNLOCK(bo);
4935 
4936 	/*
4937 	 * Walk the list of vnodes pushing all that are dirty and
4938 	 * not already on the sync list.
4939 	 */
4940 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4941 		return (0);
4942 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4943 		vfs_unbusy(mp);
4944 		return (0);
4945 	}
4946 	save = curthread_pflags_set(TDP_SYNCIO);
4947 	/*
4948 	 * The filesystem at hand may be idle with free vnodes stored in the
4949 	 * batch.  Return them instead of letting them stay there indefinitely.
4950 	 */
4951 	vfs_periodic(mp, MNT_NOWAIT);
4952 	error = VFS_SYNC(mp, MNT_LAZY);
4953 	curthread_pflags_restore(save);
4954 	vn_finished_write(mp);
4955 	vfs_unbusy(mp);
4956 	return (error);
4957 }
4958 
4959 /*
4960  * The syncer vnode is no referenced.
4961  */
4962 static int
4963 sync_inactive(struct vop_inactive_args *ap)
4964 {
4965 
4966 	vgone(ap->a_vp);
4967 	return (0);
4968 }
4969 
4970 /*
4971  * The syncer vnode is no longer needed and is being decommissioned.
4972  *
4973  * Modifications to the worklist must be protected by sync_mtx.
4974  */
4975 static int
4976 sync_reclaim(struct vop_reclaim_args *ap)
4977 {
4978 	struct vnode *vp = ap->a_vp;
4979 	struct bufobj *bo;
4980 
4981 	bo = &vp->v_bufobj;
4982 	BO_LOCK(bo);
4983 	mtx_lock(&sync_mtx);
4984 	if (vp->v_mount->mnt_syncer == vp)
4985 		vp->v_mount->mnt_syncer = NULL;
4986 	if (bo->bo_flag & BO_ONWORKLST) {
4987 		LIST_REMOVE(bo, bo_synclist);
4988 		syncer_worklist_len--;
4989 		sync_vnode_count--;
4990 		bo->bo_flag &= ~BO_ONWORKLST;
4991 	}
4992 	mtx_unlock(&sync_mtx);
4993 	BO_UNLOCK(bo);
4994 
4995 	return (0);
4996 }
4997 
4998 int
4999 vn_need_pageq_flush(struct vnode *vp)
5000 {
5001 	struct vm_object *obj;
5002 	int need;
5003 
5004 	MPASS(mtx_owned(VI_MTX(vp)));
5005 	need = 0;
5006 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5007 	    vm_object_mightbedirty(obj))
5008 		need = 1;
5009 	return (need);
5010 }
5011 
5012 /*
5013  * Check if vnode represents a disk device
5014  */
5015 bool
5016 vn_isdisk_error(struct vnode *vp, int *errp)
5017 {
5018 	int error;
5019 
5020 	if (vp->v_type != VCHR) {
5021 		error = ENOTBLK;
5022 		goto out;
5023 	}
5024 	error = 0;
5025 	dev_lock();
5026 	if (vp->v_rdev == NULL)
5027 		error = ENXIO;
5028 	else if (vp->v_rdev->si_devsw == NULL)
5029 		error = ENXIO;
5030 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5031 		error = ENOTBLK;
5032 	dev_unlock();
5033 out:
5034 	*errp = error;
5035 	return (error == 0);
5036 }
5037 
5038 bool
5039 vn_isdisk(struct vnode *vp)
5040 {
5041 	int error;
5042 
5043 	return (vn_isdisk_error(vp, &error));
5044 }
5045 
5046 /*
5047  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5048  * the comment above cache_fplookup for details.
5049  */
5050 int
5051 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5052 {
5053 	int error;
5054 
5055 	VFS_SMR_ASSERT_ENTERED();
5056 
5057 	/* Check the owner. */
5058 	if (cred->cr_uid == file_uid) {
5059 		if (file_mode & S_IXUSR)
5060 			return (0);
5061 		goto out_error;
5062 	}
5063 
5064 	/* Otherwise, check the groups (first match) */
5065 	if (groupmember(file_gid, cred)) {
5066 		if (file_mode & S_IXGRP)
5067 			return (0);
5068 		goto out_error;
5069 	}
5070 
5071 	/* Otherwise, check everyone else. */
5072 	if (file_mode & S_IXOTH)
5073 		return (0);
5074 out_error:
5075 	/*
5076 	 * Permission check failed, but it is possible denial will get overwritten
5077 	 * (e.g., when root is traversing through a 700 directory owned by someone
5078 	 * else).
5079 	 *
5080 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5081 	 * modules overriding this result. It's quite unclear what semantics
5082 	 * are allowed for them to operate, thus for safety we don't call them
5083 	 * from within the SMR section. This also means if any such modules
5084 	 * are present, we have to let the regular lookup decide.
5085 	 */
5086 	error = priv_check_cred_vfs_lookup_nomac(cred);
5087 	switch (error) {
5088 	case 0:
5089 		return (0);
5090 	case EAGAIN:
5091 		/*
5092 		 * MAC modules present.
5093 		 */
5094 		return (EAGAIN);
5095 	case EPERM:
5096 		return (EACCES);
5097 	default:
5098 		return (error);
5099 	}
5100 }
5101 
5102 /*
5103  * Common filesystem object access control check routine.  Accepts a
5104  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5105  * Returns 0 on success, or an errno on failure.
5106  */
5107 int
5108 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5109     accmode_t accmode, struct ucred *cred)
5110 {
5111 	accmode_t dac_granted;
5112 	accmode_t priv_granted;
5113 
5114 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5115 	    ("invalid bit in accmode"));
5116 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5117 	    ("VAPPEND without VWRITE"));
5118 
5119 	/*
5120 	 * Look for a normal, non-privileged way to access the file/directory
5121 	 * as requested.  If it exists, go with that.
5122 	 */
5123 
5124 	dac_granted = 0;
5125 
5126 	/* Check the owner. */
5127 	if (cred->cr_uid == file_uid) {
5128 		dac_granted |= VADMIN;
5129 		if (file_mode & S_IXUSR)
5130 			dac_granted |= VEXEC;
5131 		if (file_mode & S_IRUSR)
5132 			dac_granted |= VREAD;
5133 		if (file_mode & S_IWUSR)
5134 			dac_granted |= (VWRITE | VAPPEND);
5135 
5136 		if ((accmode & dac_granted) == accmode)
5137 			return (0);
5138 
5139 		goto privcheck;
5140 	}
5141 
5142 	/* Otherwise, check the groups (first match) */
5143 	if (groupmember(file_gid, cred)) {
5144 		if (file_mode & S_IXGRP)
5145 			dac_granted |= VEXEC;
5146 		if (file_mode & S_IRGRP)
5147 			dac_granted |= VREAD;
5148 		if (file_mode & S_IWGRP)
5149 			dac_granted |= (VWRITE | VAPPEND);
5150 
5151 		if ((accmode & dac_granted) == accmode)
5152 			return (0);
5153 
5154 		goto privcheck;
5155 	}
5156 
5157 	/* Otherwise, check everyone else. */
5158 	if (file_mode & S_IXOTH)
5159 		dac_granted |= VEXEC;
5160 	if (file_mode & S_IROTH)
5161 		dac_granted |= VREAD;
5162 	if (file_mode & S_IWOTH)
5163 		dac_granted |= (VWRITE | VAPPEND);
5164 	if ((accmode & dac_granted) == accmode)
5165 		return (0);
5166 
5167 privcheck:
5168 	/*
5169 	 * Build a privilege mask to determine if the set of privileges
5170 	 * satisfies the requirements when combined with the granted mask
5171 	 * from above.  For each privilege, if the privilege is required,
5172 	 * bitwise or the request type onto the priv_granted mask.
5173 	 */
5174 	priv_granted = 0;
5175 
5176 	if (type == VDIR) {
5177 		/*
5178 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5179 		 * requests, instead of PRIV_VFS_EXEC.
5180 		 */
5181 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5182 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5183 			priv_granted |= VEXEC;
5184 	} else {
5185 		/*
5186 		 * Ensure that at least one execute bit is on. Otherwise,
5187 		 * a privileged user will always succeed, and we don't want
5188 		 * this to happen unless the file really is executable.
5189 		 */
5190 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5191 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5192 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5193 			priv_granted |= VEXEC;
5194 	}
5195 
5196 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5197 	    !priv_check_cred(cred, PRIV_VFS_READ))
5198 		priv_granted |= VREAD;
5199 
5200 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5201 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5202 		priv_granted |= (VWRITE | VAPPEND);
5203 
5204 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5205 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5206 		priv_granted |= VADMIN;
5207 
5208 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5209 		return (0);
5210 	}
5211 
5212 	return ((accmode & VADMIN) ? EPERM : EACCES);
5213 }
5214 
5215 /*
5216  * Credential check based on process requesting service, and per-attribute
5217  * permissions.
5218  */
5219 int
5220 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5221     struct thread *td, accmode_t accmode)
5222 {
5223 
5224 	/*
5225 	 * Kernel-invoked always succeeds.
5226 	 */
5227 	if (cred == NOCRED)
5228 		return (0);
5229 
5230 	/*
5231 	 * Do not allow privileged processes in jail to directly manipulate
5232 	 * system attributes.
5233 	 */
5234 	switch (attrnamespace) {
5235 	case EXTATTR_NAMESPACE_SYSTEM:
5236 		/* Potentially should be: return (EPERM); */
5237 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5238 	case EXTATTR_NAMESPACE_USER:
5239 		return (VOP_ACCESS(vp, accmode, cred, td));
5240 	default:
5241 		return (EPERM);
5242 	}
5243 }
5244 
5245 #ifdef DEBUG_VFS_LOCKS
5246 /*
5247  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5248  * no longer ok to have an unlocked VFS.
5249  */
5250 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5251 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5252 
5253 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5254 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5255     "Drop into debugger on lock violation");
5256 
5257 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5258 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5259     0, "Check for interlock across VOPs");
5260 
5261 int vfs_badlock_print = 1;	/* Print lock violations. */
5262 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5263     0, "Print lock violations");
5264 
5265 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5266 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5267     0, "Print vnode details on lock violations");
5268 
5269 #ifdef KDB
5270 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5271 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5272     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5273 #endif
5274 
5275 static void
5276 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5277 {
5278 
5279 #ifdef KDB
5280 	if (vfs_badlock_backtrace)
5281 		kdb_backtrace();
5282 #endif
5283 	if (vfs_badlock_vnode)
5284 		vn_printf(vp, "vnode ");
5285 	if (vfs_badlock_print)
5286 		printf("%s: %p %s\n", str, (void *)vp, msg);
5287 	if (vfs_badlock_ddb)
5288 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5289 }
5290 
5291 void
5292 assert_vi_locked(struct vnode *vp, const char *str)
5293 {
5294 
5295 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5296 		vfs_badlock("interlock is not locked but should be", str, vp);
5297 }
5298 
5299 void
5300 assert_vi_unlocked(struct vnode *vp, const char *str)
5301 {
5302 
5303 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5304 		vfs_badlock("interlock is locked but should not be", str, vp);
5305 }
5306 
5307 void
5308 assert_vop_locked(struct vnode *vp, const char *str)
5309 {
5310 	int locked;
5311 
5312 	if (!IGNORE_LOCK(vp)) {
5313 		locked = VOP_ISLOCKED(vp);
5314 		if (locked == 0 || locked == LK_EXCLOTHER)
5315 			vfs_badlock("is not locked but should be", str, vp);
5316 	}
5317 }
5318 
5319 void
5320 assert_vop_unlocked(struct vnode *vp, const char *str)
5321 {
5322 
5323 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5324 		vfs_badlock("is locked but should not be", str, vp);
5325 }
5326 
5327 void
5328 assert_vop_elocked(struct vnode *vp, const char *str)
5329 {
5330 
5331 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5332 		vfs_badlock("is not exclusive locked but should be", str, vp);
5333 }
5334 #endif /* DEBUG_VFS_LOCKS */
5335 
5336 void
5337 vop_rename_fail(struct vop_rename_args *ap)
5338 {
5339 
5340 	if (ap->a_tvp != NULL)
5341 		vput(ap->a_tvp);
5342 	if (ap->a_tdvp == ap->a_tvp)
5343 		vrele(ap->a_tdvp);
5344 	else
5345 		vput(ap->a_tdvp);
5346 	vrele(ap->a_fdvp);
5347 	vrele(ap->a_fvp);
5348 }
5349 
5350 void
5351 vop_rename_pre(void *ap)
5352 {
5353 	struct vop_rename_args *a = ap;
5354 
5355 #ifdef DEBUG_VFS_LOCKS
5356 	if (a->a_tvp)
5357 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5358 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5359 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5360 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5361 
5362 	/* Check the source (from). */
5363 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5364 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5365 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5366 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5367 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5368 
5369 	/* Check the target. */
5370 	if (a->a_tvp)
5371 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5372 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5373 #endif
5374 	/*
5375 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5376 	 * in vop_rename_post but that's not going to work out since some
5377 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5378 	 *
5379 	 * For now filesystems are expected to do the relevant calls after they
5380 	 * decide what vnodes to operate on.
5381 	 */
5382 	if (a->a_tdvp != a->a_fdvp)
5383 		vhold(a->a_fdvp);
5384 	if (a->a_tvp != a->a_fvp)
5385 		vhold(a->a_fvp);
5386 	vhold(a->a_tdvp);
5387 	if (a->a_tvp)
5388 		vhold(a->a_tvp);
5389 }
5390 
5391 #ifdef DEBUG_VFS_LOCKS
5392 void
5393 vop_fplookup_vexec_debugpre(void *ap __unused)
5394 {
5395 
5396 	VFS_SMR_ASSERT_ENTERED();
5397 }
5398 
5399 void
5400 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5401 {
5402 
5403 	VFS_SMR_ASSERT_ENTERED();
5404 }
5405 
5406 void
5407 vop_strategy_debugpre(void *ap)
5408 {
5409 	struct vop_strategy_args *a;
5410 	struct buf *bp;
5411 
5412 	a = ap;
5413 	bp = a->a_bp;
5414 
5415 	/*
5416 	 * Cluster ops lock their component buffers but not the IO container.
5417 	 */
5418 	if ((bp->b_flags & B_CLUSTER) != 0)
5419 		return;
5420 
5421 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5422 		if (vfs_badlock_print)
5423 			printf(
5424 			    "VOP_STRATEGY: bp is not locked but should be\n");
5425 		if (vfs_badlock_ddb)
5426 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5427 	}
5428 }
5429 
5430 void
5431 vop_lock_debugpre(void *ap)
5432 {
5433 	struct vop_lock1_args *a = ap;
5434 
5435 	if ((a->a_flags & LK_INTERLOCK) == 0)
5436 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5437 	else
5438 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5439 }
5440 
5441 void
5442 vop_lock_debugpost(void *ap, int rc)
5443 {
5444 	struct vop_lock1_args *a = ap;
5445 
5446 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5447 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5448 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5449 }
5450 
5451 void
5452 vop_unlock_debugpre(void *ap)
5453 {
5454 	struct vop_unlock_args *a = ap;
5455 
5456 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5457 }
5458 
5459 void
5460 vop_need_inactive_debugpre(void *ap)
5461 {
5462 	struct vop_need_inactive_args *a = ap;
5463 
5464 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5465 }
5466 
5467 void
5468 vop_need_inactive_debugpost(void *ap, int rc)
5469 {
5470 	struct vop_need_inactive_args *a = ap;
5471 
5472 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5473 }
5474 #endif
5475 
5476 void
5477 vop_create_pre(void *ap)
5478 {
5479 	struct vop_create_args *a;
5480 	struct vnode *dvp;
5481 
5482 	a = ap;
5483 	dvp = a->a_dvp;
5484 	vn_seqc_write_begin(dvp);
5485 }
5486 
5487 void
5488 vop_create_post(void *ap, int rc)
5489 {
5490 	struct vop_create_args *a;
5491 	struct vnode *dvp;
5492 
5493 	a = ap;
5494 	dvp = a->a_dvp;
5495 	vn_seqc_write_end(dvp);
5496 	if (!rc)
5497 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5498 }
5499 
5500 void
5501 vop_whiteout_pre(void *ap)
5502 {
5503 	struct vop_whiteout_args *a;
5504 	struct vnode *dvp;
5505 
5506 	a = ap;
5507 	dvp = a->a_dvp;
5508 	vn_seqc_write_begin(dvp);
5509 }
5510 
5511 void
5512 vop_whiteout_post(void *ap, int rc)
5513 {
5514 	struct vop_whiteout_args *a;
5515 	struct vnode *dvp;
5516 
5517 	a = ap;
5518 	dvp = a->a_dvp;
5519 	vn_seqc_write_end(dvp);
5520 }
5521 
5522 void
5523 vop_deleteextattr_pre(void *ap)
5524 {
5525 	struct vop_deleteextattr_args *a;
5526 	struct vnode *vp;
5527 
5528 	a = ap;
5529 	vp = a->a_vp;
5530 	vn_seqc_write_begin(vp);
5531 }
5532 
5533 void
5534 vop_deleteextattr_post(void *ap, int rc)
5535 {
5536 	struct vop_deleteextattr_args *a;
5537 	struct vnode *vp;
5538 
5539 	a = ap;
5540 	vp = a->a_vp;
5541 	vn_seqc_write_end(vp);
5542 	if (!rc)
5543 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5544 }
5545 
5546 void
5547 vop_link_pre(void *ap)
5548 {
5549 	struct vop_link_args *a;
5550 	struct vnode *vp, *tdvp;
5551 
5552 	a = ap;
5553 	vp = a->a_vp;
5554 	tdvp = a->a_tdvp;
5555 	vn_seqc_write_begin(vp);
5556 	vn_seqc_write_begin(tdvp);
5557 }
5558 
5559 void
5560 vop_link_post(void *ap, int rc)
5561 {
5562 	struct vop_link_args *a;
5563 	struct vnode *vp, *tdvp;
5564 
5565 	a = ap;
5566 	vp = a->a_vp;
5567 	tdvp = a->a_tdvp;
5568 	vn_seqc_write_end(vp);
5569 	vn_seqc_write_end(tdvp);
5570 	if (!rc) {
5571 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5572 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5573 	}
5574 }
5575 
5576 void
5577 vop_mkdir_pre(void *ap)
5578 {
5579 	struct vop_mkdir_args *a;
5580 	struct vnode *dvp;
5581 
5582 	a = ap;
5583 	dvp = a->a_dvp;
5584 	vn_seqc_write_begin(dvp);
5585 }
5586 
5587 void
5588 vop_mkdir_post(void *ap, int rc)
5589 {
5590 	struct vop_mkdir_args *a;
5591 	struct vnode *dvp;
5592 
5593 	a = ap;
5594 	dvp = a->a_dvp;
5595 	vn_seqc_write_end(dvp);
5596 	if (!rc)
5597 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5598 }
5599 
5600 #ifdef DEBUG_VFS_LOCKS
5601 void
5602 vop_mkdir_debugpost(void *ap, int rc)
5603 {
5604 	struct vop_mkdir_args *a;
5605 
5606 	a = ap;
5607 	if (!rc)
5608 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5609 }
5610 #endif
5611 
5612 void
5613 vop_mknod_pre(void *ap)
5614 {
5615 	struct vop_mknod_args *a;
5616 	struct vnode *dvp;
5617 
5618 	a = ap;
5619 	dvp = a->a_dvp;
5620 	vn_seqc_write_begin(dvp);
5621 }
5622 
5623 void
5624 vop_mknod_post(void *ap, int rc)
5625 {
5626 	struct vop_mknod_args *a;
5627 	struct vnode *dvp;
5628 
5629 	a = ap;
5630 	dvp = a->a_dvp;
5631 	vn_seqc_write_end(dvp);
5632 	if (!rc)
5633 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5634 }
5635 
5636 void
5637 vop_reclaim_post(void *ap, int rc)
5638 {
5639 	struct vop_reclaim_args *a;
5640 	struct vnode *vp;
5641 
5642 	a = ap;
5643 	vp = a->a_vp;
5644 	ASSERT_VOP_IN_SEQC(vp);
5645 	if (!rc)
5646 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5647 }
5648 
5649 void
5650 vop_remove_pre(void *ap)
5651 {
5652 	struct vop_remove_args *a;
5653 	struct vnode *dvp, *vp;
5654 
5655 	a = ap;
5656 	dvp = a->a_dvp;
5657 	vp = a->a_vp;
5658 	vn_seqc_write_begin(dvp);
5659 	vn_seqc_write_begin(vp);
5660 }
5661 
5662 void
5663 vop_remove_post(void *ap, int rc)
5664 {
5665 	struct vop_remove_args *a;
5666 	struct vnode *dvp, *vp;
5667 
5668 	a = ap;
5669 	dvp = a->a_dvp;
5670 	vp = a->a_vp;
5671 	vn_seqc_write_end(dvp);
5672 	vn_seqc_write_end(vp);
5673 	if (!rc) {
5674 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5675 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5676 	}
5677 }
5678 
5679 void
5680 vop_rename_post(void *ap, int rc)
5681 {
5682 	struct vop_rename_args *a = ap;
5683 	long hint;
5684 
5685 	if (!rc) {
5686 		hint = NOTE_WRITE;
5687 		if (a->a_fdvp == a->a_tdvp) {
5688 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5689 				hint |= NOTE_LINK;
5690 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5691 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5692 		} else {
5693 			hint |= NOTE_EXTEND;
5694 			if (a->a_fvp->v_type == VDIR)
5695 				hint |= NOTE_LINK;
5696 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5697 
5698 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5699 			    a->a_tvp->v_type == VDIR)
5700 				hint &= ~NOTE_LINK;
5701 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5702 		}
5703 
5704 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5705 		if (a->a_tvp)
5706 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5707 	}
5708 	if (a->a_tdvp != a->a_fdvp)
5709 		vdrop(a->a_fdvp);
5710 	if (a->a_tvp != a->a_fvp)
5711 		vdrop(a->a_fvp);
5712 	vdrop(a->a_tdvp);
5713 	if (a->a_tvp)
5714 		vdrop(a->a_tvp);
5715 }
5716 
5717 void
5718 vop_rmdir_pre(void *ap)
5719 {
5720 	struct vop_rmdir_args *a;
5721 	struct vnode *dvp, *vp;
5722 
5723 	a = ap;
5724 	dvp = a->a_dvp;
5725 	vp = a->a_vp;
5726 	vn_seqc_write_begin(dvp);
5727 	vn_seqc_write_begin(vp);
5728 }
5729 
5730 void
5731 vop_rmdir_post(void *ap, int rc)
5732 {
5733 	struct vop_rmdir_args *a;
5734 	struct vnode *dvp, *vp;
5735 
5736 	a = ap;
5737 	dvp = a->a_dvp;
5738 	vp = a->a_vp;
5739 	vn_seqc_write_end(dvp);
5740 	vn_seqc_write_end(vp);
5741 	if (!rc) {
5742 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5743 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5744 	}
5745 }
5746 
5747 void
5748 vop_setattr_pre(void *ap)
5749 {
5750 	struct vop_setattr_args *a;
5751 	struct vnode *vp;
5752 
5753 	a = ap;
5754 	vp = a->a_vp;
5755 	vn_seqc_write_begin(vp);
5756 }
5757 
5758 void
5759 vop_setattr_post(void *ap, int rc)
5760 {
5761 	struct vop_setattr_args *a;
5762 	struct vnode *vp;
5763 
5764 	a = ap;
5765 	vp = a->a_vp;
5766 	vn_seqc_write_end(vp);
5767 	if (!rc)
5768 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5769 }
5770 
5771 void
5772 vop_setacl_pre(void *ap)
5773 {
5774 	struct vop_setacl_args *a;
5775 	struct vnode *vp;
5776 
5777 	a = ap;
5778 	vp = a->a_vp;
5779 	vn_seqc_write_begin(vp);
5780 }
5781 
5782 void
5783 vop_setacl_post(void *ap, int rc __unused)
5784 {
5785 	struct vop_setacl_args *a;
5786 	struct vnode *vp;
5787 
5788 	a = ap;
5789 	vp = a->a_vp;
5790 	vn_seqc_write_end(vp);
5791 }
5792 
5793 void
5794 vop_setextattr_pre(void *ap)
5795 {
5796 	struct vop_setextattr_args *a;
5797 	struct vnode *vp;
5798 
5799 	a = ap;
5800 	vp = a->a_vp;
5801 	vn_seqc_write_begin(vp);
5802 }
5803 
5804 void
5805 vop_setextattr_post(void *ap, int rc)
5806 {
5807 	struct vop_setextattr_args *a;
5808 	struct vnode *vp;
5809 
5810 	a = ap;
5811 	vp = a->a_vp;
5812 	vn_seqc_write_end(vp);
5813 	if (!rc)
5814 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5815 }
5816 
5817 void
5818 vop_symlink_pre(void *ap)
5819 {
5820 	struct vop_symlink_args *a;
5821 	struct vnode *dvp;
5822 
5823 	a = ap;
5824 	dvp = a->a_dvp;
5825 	vn_seqc_write_begin(dvp);
5826 }
5827 
5828 void
5829 vop_symlink_post(void *ap, int rc)
5830 {
5831 	struct vop_symlink_args *a;
5832 	struct vnode *dvp;
5833 
5834 	a = ap;
5835 	dvp = a->a_dvp;
5836 	vn_seqc_write_end(dvp);
5837 	if (!rc)
5838 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5839 }
5840 
5841 void
5842 vop_open_post(void *ap, int rc)
5843 {
5844 	struct vop_open_args *a = ap;
5845 
5846 	if (!rc)
5847 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5848 }
5849 
5850 void
5851 vop_close_post(void *ap, int rc)
5852 {
5853 	struct vop_close_args *a = ap;
5854 
5855 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5856 	    !VN_IS_DOOMED(a->a_vp))) {
5857 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5858 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5859 	}
5860 }
5861 
5862 void
5863 vop_read_post(void *ap, int rc)
5864 {
5865 	struct vop_read_args *a = ap;
5866 
5867 	if (!rc)
5868 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5869 }
5870 
5871 void
5872 vop_read_pgcache_post(void *ap, int rc)
5873 {
5874 	struct vop_read_pgcache_args *a = ap;
5875 
5876 	if (!rc)
5877 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
5878 }
5879 
5880 void
5881 vop_readdir_post(void *ap, int rc)
5882 {
5883 	struct vop_readdir_args *a = ap;
5884 
5885 	if (!rc)
5886 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5887 }
5888 
5889 static struct knlist fs_knlist;
5890 
5891 static void
5892 vfs_event_init(void *arg)
5893 {
5894 	knlist_init_mtx(&fs_knlist, NULL);
5895 }
5896 /* XXX - correct order? */
5897 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5898 
5899 void
5900 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5901 {
5902 
5903 	KNOTE_UNLOCKED(&fs_knlist, event);
5904 }
5905 
5906 static int	filt_fsattach(struct knote *kn);
5907 static void	filt_fsdetach(struct knote *kn);
5908 static int	filt_fsevent(struct knote *kn, long hint);
5909 
5910 struct filterops fs_filtops = {
5911 	.f_isfd = 0,
5912 	.f_attach = filt_fsattach,
5913 	.f_detach = filt_fsdetach,
5914 	.f_event = filt_fsevent
5915 };
5916 
5917 static int
5918 filt_fsattach(struct knote *kn)
5919 {
5920 
5921 	kn->kn_flags |= EV_CLEAR;
5922 	knlist_add(&fs_knlist, kn, 0);
5923 	return (0);
5924 }
5925 
5926 static void
5927 filt_fsdetach(struct knote *kn)
5928 {
5929 
5930 	knlist_remove(&fs_knlist, kn, 0);
5931 }
5932 
5933 static int
5934 filt_fsevent(struct knote *kn, long hint)
5935 {
5936 
5937 	kn->kn_fflags |= hint;
5938 	return (kn->kn_fflags != 0);
5939 }
5940 
5941 static int
5942 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5943 {
5944 	struct vfsidctl vc;
5945 	int error;
5946 	struct mount *mp;
5947 
5948 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5949 	if (error)
5950 		return (error);
5951 	if (vc.vc_vers != VFS_CTL_VERS1)
5952 		return (EINVAL);
5953 	mp = vfs_getvfs(&vc.vc_fsid);
5954 	if (mp == NULL)
5955 		return (ENOENT);
5956 	/* ensure that a specific sysctl goes to the right filesystem. */
5957 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5958 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5959 		vfs_rel(mp);
5960 		return (EINVAL);
5961 	}
5962 	VCTLTOREQ(&vc, req);
5963 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5964 	vfs_rel(mp);
5965 	return (error);
5966 }
5967 
5968 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5969     NULL, 0, sysctl_vfs_ctl, "",
5970     "Sysctl by fsid");
5971 
5972 /*
5973  * Function to initialize a va_filerev field sensibly.
5974  * XXX: Wouldn't a random number make a lot more sense ??
5975  */
5976 u_quad_t
5977 init_va_filerev(void)
5978 {
5979 	struct bintime bt;
5980 
5981 	getbinuptime(&bt);
5982 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5983 }
5984 
5985 static int	filt_vfsread(struct knote *kn, long hint);
5986 static int	filt_vfswrite(struct knote *kn, long hint);
5987 static int	filt_vfsvnode(struct knote *kn, long hint);
5988 static void	filt_vfsdetach(struct knote *kn);
5989 static struct filterops vfsread_filtops = {
5990 	.f_isfd = 1,
5991 	.f_detach = filt_vfsdetach,
5992 	.f_event = filt_vfsread
5993 };
5994 static struct filterops vfswrite_filtops = {
5995 	.f_isfd = 1,
5996 	.f_detach = filt_vfsdetach,
5997 	.f_event = filt_vfswrite
5998 };
5999 static struct filterops vfsvnode_filtops = {
6000 	.f_isfd = 1,
6001 	.f_detach = filt_vfsdetach,
6002 	.f_event = filt_vfsvnode
6003 };
6004 
6005 static void
6006 vfs_knllock(void *arg)
6007 {
6008 	struct vnode *vp = arg;
6009 
6010 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6011 }
6012 
6013 static void
6014 vfs_knlunlock(void *arg)
6015 {
6016 	struct vnode *vp = arg;
6017 
6018 	VOP_UNLOCK(vp);
6019 }
6020 
6021 static void
6022 vfs_knl_assert_locked(void *arg)
6023 {
6024 #ifdef DEBUG_VFS_LOCKS
6025 	struct vnode *vp = arg;
6026 
6027 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6028 #endif
6029 }
6030 
6031 static void
6032 vfs_knl_assert_unlocked(void *arg)
6033 {
6034 #ifdef DEBUG_VFS_LOCKS
6035 	struct vnode *vp = arg;
6036 
6037 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6038 #endif
6039 }
6040 
6041 int
6042 vfs_kqfilter(struct vop_kqfilter_args *ap)
6043 {
6044 	struct vnode *vp = ap->a_vp;
6045 	struct knote *kn = ap->a_kn;
6046 	struct knlist *knl;
6047 
6048 	switch (kn->kn_filter) {
6049 	case EVFILT_READ:
6050 		kn->kn_fop = &vfsread_filtops;
6051 		break;
6052 	case EVFILT_WRITE:
6053 		kn->kn_fop = &vfswrite_filtops;
6054 		break;
6055 	case EVFILT_VNODE:
6056 		kn->kn_fop = &vfsvnode_filtops;
6057 		break;
6058 	default:
6059 		return (EINVAL);
6060 	}
6061 
6062 	kn->kn_hook = (caddr_t)vp;
6063 
6064 	v_addpollinfo(vp);
6065 	if (vp->v_pollinfo == NULL)
6066 		return (ENOMEM);
6067 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6068 	vhold(vp);
6069 	knlist_add(knl, kn, 0);
6070 
6071 	return (0);
6072 }
6073 
6074 /*
6075  * Detach knote from vnode
6076  */
6077 static void
6078 filt_vfsdetach(struct knote *kn)
6079 {
6080 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6081 
6082 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6083 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6084 	vdrop(vp);
6085 }
6086 
6087 /*ARGSUSED*/
6088 static int
6089 filt_vfsread(struct knote *kn, long hint)
6090 {
6091 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6092 	struct vattr va;
6093 	int res;
6094 
6095 	/*
6096 	 * filesystem is gone, so set the EOF flag and schedule
6097 	 * the knote for deletion.
6098 	 */
6099 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6100 		VI_LOCK(vp);
6101 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6102 		VI_UNLOCK(vp);
6103 		return (1);
6104 	}
6105 
6106 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6107 		return (0);
6108 
6109 	VI_LOCK(vp);
6110 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6111 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6112 	VI_UNLOCK(vp);
6113 	return (res);
6114 }
6115 
6116 /*ARGSUSED*/
6117 static int
6118 filt_vfswrite(struct knote *kn, long hint)
6119 {
6120 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6121 
6122 	VI_LOCK(vp);
6123 
6124 	/*
6125 	 * filesystem is gone, so set the EOF flag and schedule
6126 	 * the knote for deletion.
6127 	 */
6128 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6129 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6130 
6131 	kn->kn_data = 0;
6132 	VI_UNLOCK(vp);
6133 	return (1);
6134 }
6135 
6136 static int
6137 filt_vfsvnode(struct knote *kn, long hint)
6138 {
6139 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6140 	int res;
6141 
6142 	VI_LOCK(vp);
6143 	if (kn->kn_sfflags & hint)
6144 		kn->kn_fflags |= hint;
6145 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6146 		kn->kn_flags |= EV_EOF;
6147 		VI_UNLOCK(vp);
6148 		return (1);
6149 	}
6150 	res = (kn->kn_fflags != 0);
6151 	VI_UNLOCK(vp);
6152 	return (res);
6153 }
6154 
6155 /*
6156  * Returns whether the directory is empty or not.
6157  * If it is empty, the return value is 0; otherwise
6158  * the return value is an error value (which may
6159  * be ENOTEMPTY).
6160  */
6161 int
6162 vfs_emptydir(struct vnode *vp)
6163 {
6164 	struct uio uio;
6165 	struct iovec iov;
6166 	struct dirent *dirent, *dp, *endp;
6167 	int error, eof;
6168 
6169 	error = 0;
6170 	eof = 0;
6171 
6172 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6173 
6174 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6175 	iov.iov_base = dirent;
6176 	iov.iov_len = sizeof(struct dirent);
6177 
6178 	uio.uio_iov = &iov;
6179 	uio.uio_iovcnt = 1;
6180 	uio.uio_offset = 0;
6181 	uio.uio_resid = sizeof(struct dirent);
6182 	uio.uio_segflg = UIO_SYSSPACE;
6183 	uio.uio_rw = UIO_READ;
6184 	uio.uio_td = curthread;
6185 
6186 	while (eof == 0 && error == 0) {
6187 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6188 		    NULL, NULL);
6189 		if (error != 0)
6190 			break;
6191 		endp = (void *)((uint8_t *)dirent +
6192 		    sizeof(struct dirent) - uio.uio_resid);
6193 		for (dp = dirent; dp < endp;
6194 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6195 			if (dp->d_type == DT_WHT)
6196 				continue;
6197 			if (dp->d_namlen == 0)
6198 				continue;
6199 			if (dp->d_type != DT_DIR &&
6200 			    dp->d_type != DT_UNKNOWN) {
6201 				error = ENOTEMPTY;
6202 				break;
6203 			}
6204 			if (dp->d_namlen > 2) {
6205 				error = ENOTEMPTY;
6206 				break;
6207 			}
6208 			if (dp->d_namlen == 1 &&
6209 			    dp->d_name[0] != '.') {
6210 				error = ENOTEMPTY;
6211 				break;
6212 			}
6213 			if (dp->d_namlen == 2 &&
6214 			    dp->d_name[1] != '.') {
6215 				error = ENOTEMPTY;
6216 				break;
6217 			}
6218 			uio.uio_resid = sizeof(struct dirent);
6219 		}
6220 	}
6221 	free(dirent, M_TEMP);
6222 	return (error);
6223 }
6224 
6225 int
6226 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6227 {
6228 	int error;
6229 
6230 	if (dp->d_reclen > ap->a_uio->uio_resid)
6231 		return (ENAMETOOLONG);
6232 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6233 	if (error) {
6234 		if (ap->a_ncookies != NULL) {
6235 			if (ap->a_cookies != NULL)
6236 				free(ap->a_cookies, M_TEMP);
6237 			ap->a_cookies = NULL;
6238 			*ap->a_ncookies = 0;
6239 		}
6240 		return (error);
6241 	}
6242 	if (ap->a_ncookies == NULL)
6243 		return (0);
6244 
6245 	KASSERT(ap->a_cookies,
6246 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6247 
6248 	*ap->a_cookies = realloc(*ap->a_cookies,
6249 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6250 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6251 	*ap->a_ncookies += 1;
6252 	return (0);
6253 }
6254 
6255 /*
6256  * The purpose of this routine is to remove granularity from accmode_t,
6257  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6258  * VADMIN and VAPPEND.
6259  *
6260  * If it returns 0, the caller is supposed to continue with the usual
6261  * access checks using 'accmode' as modified by this routine.  If it
6262  * returns nonzero value, the caller is supposed to return that value
6263  * as errno.
6264  *
6265  * Note that after this routine runs, accmode may be zero.
6266  */
6267 int
6268 vfs_unixify_accmode(accmode_t *accmode)
6269 {
6270 	/*
6271 	 * There is no way to specify explicit "deny" rule using
6272 	 * file mode or POSIX.1e ACLs.
6273 	 */
6274 	if (*accmode & VEXPLICIT_DENY) {
6275 		*accmode = 0;
6276 		return (0);
6277 	}
6278 
6279 	/*
6280 	 * None of these can be translated into usual access bits.
6281 	 * Also, the common case for NFSv4 ACLs is to not contain
6282 	 * either of these bits. Caller should check for VWRITE
6283 	 * on the containing directory instead.
6284 	 */
6285 	if (*accmode & (VDELETE_CHILD | VDELETE))
6286 		return (EPERM);
6287 
6288 	if (*accmode & VADMIN_PERMS) {
6289 		*accmode &= ~VADMIN_PERMS;
6290 		*accmode |= VADMIN;
6291 	}
6292 
6293 	/*
6294 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6295 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6296 	 */
6297 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6298 
6299 	return (0);
6300 }
6301 
6302 /*
6303  * Clear out a doomed vnode (if any) and replace it with a new one as long
6304  * as the fs is not being unmounted. Return the root vnode to the caller.
6305  */
6306 static int __noinline
6307 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6308 {
6309 	struct vnode *vp;
6310 	int error;
6311 
6312 restart:
6313 	if (mp->mnt_rootvnode != NULL) {
6314 		MNT_ILOCK(mp);
6315 		vp = mp->mnt_rootvnode;
6316 		if (vp != NULL) {
6317 			if (!VN_IS_DOOMED(vp)) {
6318 				vrefact(vp);
6319 				MNT_IUNLOCK(mp);
6320 				error = vn_lock(vp, flags);
6321 				if (error == 0) {
6322 					*vpp = vp;
6323 					return (0);
6324 				}
6325 				vrele(vp);
6326 				goto restart;
6327 			}
6328 			/*
6329 			 * Clear the old one.
6330 			 */
6331 			mp->mnt_rootvnode = NULL;
6332 		}
6333 		MNT_IUNLOCK(mp);
6334 		if (vp != NULL) {
6335 			vfs_op_barrier_wait(mp);
6336 			vrele(vp);
6337 		}
6338 	}
6339 	error = VFS_CACHEDROOT(mp, flags, vpp);
6340 	if (error != 0)
6341 		return (error);
6342 	if (mp->mnt_vfs_ops == 0) {
6343 		MNT_ILOCK(mp);
6344 		if (mp->mnt_vfs_ops != 0) {
6345 			MNT_IUNLOCK(mp);
6346 			return (0);
6347 		}
6348 		if (mp->mnt_rootvnode == NULL) {
6349 			vrefact(*vpp);
6350 			mp->mnt_rootvnode = *vpp;
6351 		} else {
6352 			if (mp->mnt_rootvnode != *vpp) {
6353 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6354 					panic("%s: mismatch between vnode returned "
6355 					    " by VFS_CACHEDROOT and the one cached "
6356 					    " (%p != %p)",
6357 					    __func__, *vpp, mp->mnt_rootvnode);
6358 				}
6359 			}
6360 		}
6361 		MNT_IUNLOCK(mp);
6362 	}
6363 	return (0);
6364 }
6365 
6366 int
6367 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6368 {
6369 	struct vnode *vp;
6370 	int error;
6371 
6372 	if (!vfs_op_thread_enter(mp))
6373 		return (vfs_cache_root_fallback(mp, flags, vpp));
6374 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6375 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6376 		vfs_op_thread_exit(mp);
6377 		return (vfs_cache_root_fallback(mp, flags, vpp));
6378 	}
6379 	vrefact(vp);
6380 	vfs_op_thread_exit(mp);
6381 	error = vn_lock(vp, flags);
6382 	if (error != 0) {
6383 		vrele(vp);
6384 		return (vfs_cache_root_fallback(mp, flags, vpp));
6385 	}
6386 	*vpp = vp;
6387 	return (0);
6388 }
6389 
6390 struct vnode *
6391 vfs_cache_root_clear(struct mount *mp)
6392 {
6393 	struct vnode *vp;
6394 
6395 	/*
6396 	 * ops > 0 guarantees there is nobody who can see this vnode
6397 	 */
6398 	MPASS(mp->mnt_vfs_ops > 0);
6399 	vp = mp->mnt_rootvnode;
6400 	if (vp != NULL)
6401 		vn_seqc_write_begin(vp);
6402 	mp->mnt_rootvnode = NULL;
6403 	return (vp);
6404 }
6405 
6406 void
6407 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6408 {
6409 
6410 	MPASS(mp->mnt_vfs_ops > 0);
6411 	vrefact(vp);
6412 	mp->mnt_rootvnode = vp;
6413 }
6414 
6415 /*
6416  * These are helper functions for filesystems to traverse all
6417  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6418  *
6419  * This interface replaces MNT_VNODE_FOREACH.
6420  */
6421 
6422 struct vnode *
6423 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6424 {
6425 	struct vnode *vp;
6426 
6427 	if (should_yield())
6428 		kern_yield(PRI_USER);
6429 	MNT_ILOCK(mp);
6430 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6431 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6432 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6433 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6434 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6435 			continue;
6436 		VI_LOCK(vp);
6437 		if (VN_IS_DOOMED(vp)) {
6438 			VI_UNLOCK(vp);
6439 			continue;
6440 		}
6441 		break;
6442 	}
6443 	if (vp == NULL) {
6444 		__mnt_vnode_markerfree_all(mvp, mp);
6445 		/* MNT_IUNLOCK(mp); -- done in above function */
6446 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6447 		return (NULL);
6448 	}
6449 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6450 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6451 	MNT_IUNLOCK(mp);
6452 	return (vp);
6453 }
6454 
6455 struct vnode *
6456 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6457 {
6458 	struct vnode *vp;
6459 
6460 	*mvp = vn_alloc_marker(mp);
6461 	MNT_ILOCK(mp);
6462 	MNT_REF(mp);
6463 
6464 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6465 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6466 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6467 			continue;
6468 		VI_LOCK(vp);
6469 		if (VN_IS_DOOMED(vp)) {
6470 			VI_UNLOCK(vp);
6471 			continue;
6472 		}
6473 		break;
6474 	}
6475 	if (vp == NULL) {
6476 		MNT_REL(mp);
6477 		MNT_IUNLOCK(mp);
6478 		vn_free_marker(*mvp);
6479 		*mvp = NULL;
6480 		return (NULL);
6481 	}
6482 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6483 	MNT_IUNLOCK(mp);
6484 	return (vp);
6485 }
6486 
6487 void
6488 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6489 {
6490 
6491 	if (*mvp == NULL) {
6492 		MNT_IUNLOCK(mp);
6493 		return;
6494 	}
6495 
6496 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6497 
6498 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6499 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6500 	MNT_REL(mp);
6501 	MNT_IUNLOCK(mp);
6502 	vn_free_marker(*mvp);
6503 	*mvp = NULL;
6504 }
6505 
6506 /*
6507  * These are helper functions for filesystems to traverse their
6508  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6509  */
6510 static void
6511 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6512 {
6513 
6514 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6515 
6516 	MNT_ILOCK(mp);
6517 	MNT_REL(mp);
6518 	MNT_IUNLOCK(mp);
6519 	vn_free_marker(*mvp);
6520 	*mvp = NULL;
6521 }
6522 
6523 /*
6524  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6525  * conventional lock order during mnt_vnode_next_lazy iteration.
6526  *
6527  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6528  * The list lock is dropped and reacquired.  On success, both locks are held.
6529  * On failure, the mount vnode list lock is held but the vnode interlock is
6530  * not, and the procedure may have yielded.
6531  */
6532 static bool
6533 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6534     struct vnode *vp)
6535 {
6536 
6537 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6538 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6539 	    ("%s: bad marker", __func__));
6540 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6541 	    ("%s: inappropriate vnode", __func__));
6542 	ASSERT_VI_UNLOCKED(vp, __func__);
6543 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6544 
6545 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6546 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6547 
6548 	/*
6549 	 * Note we may be racing against vdrop which transitioned the hold
6550 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6551 	 * if we are the only user after we get the interlock we will just
6552 	 * vdrop.
6553 	 */
6554 	vhold(vp);
6555 	mtx_unlock(&mp->mnt_listmtx);
6556 	VI_LOCK(vp);
6557 	if (VN_IS_DOOMED(vp)) {
6558 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6559 		goto out_lost;
6560 	}
6561 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6562 	/*
6563 	 * There is nothing to do if we are the last user.
6564 	 */
6565 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6566 		goto out_lost;
6567 	mtx_lock(&mp->mnt_listmtx);
6568 	return (true);
6569 out_lost:
6570 	vdropl(vp);
6571 	maybe_yield();
6572 	mtx_lock(&mp->mnt_listmtx);
6573 	return (false);
6574 }
6575 
6576 static struct vnode *
6577 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6578     void *cbarg)
6579 {
6580 	struct vnode *vp;
6581 
6582 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6583 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6584 restart:
6585 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6586 	while (vp != NULL) {
6587 		if (vp->v_type == VMARKER) {
6588 			vp = TAILQ_NEXT(vp, v_lazylist);
6589 			continue;
6590 		}
6591 		/*
6592 		 * See if we want to process the vnode. Note we may encounter a
6593 		 * long string of vnodes we don't care about and hog the list
6594 		 * as a result. Check for it and requeue the marker.
6595 		 */
6596 		VNPASS(!VN_IS_DOOMED(vp), vp);
6597 		if (!cb(vp, cbarg)) {
6598 			if (!should_yield()) {
6599 				vp = TAILQ_NEXT(vp, v_lazylist);
6600 				continue;
6601 			}
6602 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6603 			    v_lazylist);
6604 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6605 			    v_lazylist);
6606 			mtx_unlock(&mp->mnt_listmtx);
6607 			kern_yield(PRI_USER);
6608 			mtx_lock(&mp->mnt_listmtx);
6609 			goto restart;
6610 		}
6611 		/*
6612 		 * Try-lock because this is the wrong lock order.
6613 		 */
6614 		if (!VI_TRYLOCK(vp) &&
6615 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6616 			goto restart;
6617 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6618 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6619 		    ("alien vnode on the lazy list %p %p", vp, mp));
6620 		VNPASS(vp->v_mount == mp, vp);
6621 		VNPASS(!VN_IS_DOOMED(vp), vp);
6622 		break;
6623 	}
6624 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6625 
6626 	/* Check if we are done */
6627 	if (vp == NULL) {
6628 		mtx_unlock(&mp->mnt_listmtx);
6629 		mnt_vnode_markerfree_lazy(mvp, mp);
6630 		return (NULL);
6631 	}
6632 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6633 	mtx_unlock(&mp->mnt_listmtx);
6634 	ASSERT_VI_LOCKED(vp, "lazy iter");
6635 	return (vp);
6636 }
6637 
6638 struct vnode *
6639 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6640     void *cbarg)
6641 {
6642 
6643 	if (should_yield())
6644 		kern_yield(PRI_USER);
6645 	mtx_lock(&mp->mnt_listmtx);
6646 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6647 }
6648 
6649 struct vnode *
6650 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6651     void *cbarg)
6652 {
6653 	struct vnode *vp;
6654 
6655 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6656 		return (NULL);
6657 
6658 	*mvp = vn_alloc_marker(mp);
6659 	MNT_ILOCK(mp);
6660 	MNT_REF(mp);
6661 	MNT_IUNLOCK(mp);
6662 
6663 	mtx_lock(&mp->mnt_listmtx);
6664 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6665 	if (vp == NULL) {
6666 		mtx_unlock(&mp->mnt_listmtx);
6667 		mnt_vnode_markerfree_lazy(mvp, mp);
6668 		return (NULL);
6669 	}
6670 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6671 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6672 }
6673 
6674 void
6675 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6676 {
6677 
6678 	if (*mvp == NULL)
6679 		return;
6680 
6681 	mtx_lock(&mp->mnt_listmtx);
6682 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6683 	mtx_unlock(&mp->mnt_listmtx);
6684 	mnt_vnode_markerfree_lazy(mvp, mp);
6685 }
6686 
6687 int
6688 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6689 {
6690 
6691 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6692 		cnp->cn_flags &= ~NOEXECCHECK;
6693 		return (0);
6694 	}
6695 
6696 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6697 }
6698 
6699 /*
6700  * Do not use this variant unless you have means other than the hold count
6701  * to prevent the vnode from getting freed.
6702  */
6703 void
6704 vn_seqc_write_begin_unheld_locked(struct vnode *vp)
6705 {
6706 
6707 	ASSERT_VI_LOCKED(vp, __func__);
6708 	VNPASS(vp->v_seqc_users >= 0, vp);
6709 	vp->v_seqc_users++;
6710 	if (vp->v_seqc_users == 1)
6711 		seqc_sleepable_write_begin(&vp->v_seqc);
6712 }
6713 
6714 void
6715 vn_seqc_write_begin_locked(struct vnode *vp)
6716 {
6717 
6718 	ASSERT_VI_LOCKED(vp, __func__);
6719 	VNPASS(vp->v_holdcnt > 0, vp);
6720 	vn_seqc_write_begin_unheld_locked(vp);
6721 }
6722 
6723 void
6724 vn_seqc_write_begin(struct vnode *vp)
6725 {
6726 
6727 	VI_LOCK(vp);
6728 	vn_seqc_write_begin_locked(vp);
6729 	VI_UNLOCK(vp);
6730 }
6731 
6732 void
6733 vn_seqc_write_begin_unheld(struct vnode *vp)
6734 {
6735 
6736 	VI_LOCK(vp);
6737 	vn_seqc_write_begin_unheld_locked(vp);
6738 	VI_UNLOCK(vp);
6739 }
6740 
6741 void
6742 vn_seqc_write_end_locked(struct vnode *vp)
6743 {
6744 
6745 	ASSERT_VI_LOCKED(vp, __func__);
6746 	VNPASS(vp->v_seqc_users > 0, vp);
6747 	vp->v_seqc_users--;
6748 	if (vp->v_seqc_users == 0)
6749 		seqc_sleepable_write_end(&vp->v_seqc);
6750 }
6751 
6752 void
6753 vn_seqc_write_end(struct vnode *vp)
6754 {
6755 
6756 	VI_LOCK(vp);
6757 	vn_seqc_write_end_locked(vp);
6758 	VI_UNLOCK(vp);
6759 }
6760