xref: /freebsd/sys/kern/vfs_subr.c (revision a33774adb3e7e902ca1742227ee070b93d602fc8)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/stat.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/uma.h>
99 
100 #ifdef DDB
101 #include <ddb/ddb.h>
102 #endif
103 
104 static void	delmntque(struct vnode *vp);
105 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 		    int slpflag, int slptimeo);
107 static void	syncer_shutdown(void *arg, int howto);
108 static int	vtryrecycle(struct vnode *vp);
109 static void	v_init_counters(struct vnode *);
110 static void	v_incr_devcount(struct vnode *);
111 static void	v_decr_devcount(struct vnode *);
112 static void	vgonel(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_locked(void *arg);
116 static void	vfs_knl_assert_unlocked(void *arg);
117 static void	vnlru_return_batches(struct vfsops *mnt_op);
118 static void	destroy_vpollinfo(struct vpollinfo *vi);
119 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
120 		    daddr_t startlbn, daddr_t endlbn);
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 unsigned long	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 static u_long mnt_free_list_batch = 128;
152 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
153     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
154 
155 /*
156  * Conversion tables for conversion from vnode types to inode formats
157  * and back.
158  */
159 enum vtype iftovt_tab[16] = {
160 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
161 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
162 };
163 int vttoif_tab[10] = {
164 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
165 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
166 };
167 
168 /*
169  * List of vnodes that are ready for recycling.
170  */
171 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
172 
173 /*
174  * "Free" vnode target.  Free vnodes are rarely completely free, but are
175  * just ones that are cheap to recycle.  Usually they are for files which
176  * have been stat'd but not read; these usually have inode and namecache
177  * data attached to them.  This target is the preferred minimum size of a
178  * sub-cache consisting mostly of such files. The system balances the size
179  * of this sub-cache with its complement to try to prevent either from
180  * thrashing while the other is relatively inactive.  The targets express
181  * a preference for the best balance.
182  *
183  * "Above" this target there are 2 further targets (watermarks) related
184  * to recyling of free vnodes.  In the best-operating case, the cache is
185  * exactly full, the free list has size between vlowat and vhiwat above the
186  * free target, and recycling from it and normal use maintains this state.
187  * Sometimes the free list is below vlowat or even empty, but this state
188  * is even better for immediate use provided the cache is not full.
189  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
190  * ones) to reach one of these states.  The watermarks are currently hard-
191  * coded as 4% and 9% of the available space higher.  These and the default
192  * of 25% for wantfreevnodes are too large if the memory size is large.
193  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
194  * whenever vnlru_proc() becomes active.
195  */
196 static u_long wantfreevnodes;
197 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
198     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
199 static u_long freevnodes;
200 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
201     &freevnodes, 0, "Number of \"free\" vnodes");
202 
203 static counter_u64_t recycles_count;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
205     "Number of vnodes recycled to meet vnode cache targets");
206 
207 /*
208  * Various variables used for debugging the new implementation of
209  * reassignbuf().
210  * XXX these are probably of (very) limited utility now.
211  */
212 static int reassignbufcalls;
213 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
214     &reassignbufcalls, 0, "Number of calls to reassignbuf");
215 
216 static counter_u64_t free_owe_inact;
217 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
218     "Number of times free vnodes kept on active list due to VFS "
219     "owing inactivation");
220 
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
223 
224 /*
225  * Lock for any access to the following:
226  *	vnode_free_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx vnode_free_list_mtx;
231 
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
234 
235 static uma_zone_t buf_trie_zone;
236 
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
240 
241 /*
242  * The workitem queue.
243  *
244  * It is useful to delay writes of file data and filesystem metadata
245  * for tens of seconds so that quickly created and deleted files need
246  * not waste disk bandwidth being created and removed. To realize this,
247  * we append vnodes to a "workitem" queue. When running with a soft
248  * updates implementation, most pending metadata dependencies should
249  * not wait for more than a few seconds. Thus, mounted on block devices
250  * are delayed only about a half the time that file data is delayed.
251  * Similarly, directory updates are more critical, so are only delayed
252  * about a third the time that file data is delayed. Thus, there are
253  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254  * one each second (driven off the filesystem syncer process). The
255  * syncer_delayno variable indicates the next queue that is to be processed.
256  * Items that need to be processed soon are placed in this queue:
257  *
258  *	syncer_workitem_pending[syncer_delayno]
259  *
260  * A delay of fifteen seconds is done by placing the request fifteen
261  * entries later in the queue:
262  *
263  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264  *
265  */
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
270 /*
271  * The sync_mtx protects:
272  *	bo->bo_synclist
273  *	sync_vnode_count
274  *	syncer_delayno
275  *	syncer_state
276  *	syncer_workitem_pending
277  *	syncer_worklist_len
278  *	rushjob
279  */
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
282 
283 #define SYNCER_MAXDELAY		32
284 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285 static int syncdelay = 30;		/* max time to delay syncing data */
286 static int filedelay = 30;		/* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288     "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29;		/* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291     "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28;		/* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294     "Time to delay syncing metadata (in seconds)");
295 static int rushjob;		/* number of slots to run ASAP */
296 static int stat_rush_requests;	/* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298     "Number of times I/O speeded up (rush requests)");
299 
300 /*
301  * When shutting down the syncer, run it at four times normal speed.
302  */
303 #define SYNCER_SHUTDOWN_SPEEDUP		4
304 static int sync_vnode_count;
305 static int syncer_worklist_len;
306 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
307     syncer_state;
308 
309 /* Target for maximum number of vnodes. */
310 int desiredvnodes;
311 static int gapvnodes;		/* gap between wanted and desired */
312 static int vhiwat;		/* enough extras after expansion */
313 static int vlowat;		/* minimal extras before expansion */
314 static int vstir;		/* nonzero to stir non-free vnodes */
315 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
316 
317 static int
318 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
319 {
320 	int error, old_desiredvnodes;
321 
322 	old_desiredvnodes = desiredvnodes;
323 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
324 		return (error);
325 	if (old_desiredvnodes != desiredvnodes) {
326 		wantfreevnodes = desiredvnodes / 4;
327 		/* XXX locking seems to be incomplete. */
328 		vfs_hash_changesize(desiredvnodes);
329 		cache_changesize(desiredvnodes);
330 	}
331 	return (0);
332 }
333 
334 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
335     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
336     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
337 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
338     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
339 static int vnlru_nowhere;
340 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
341     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
342 
343 static int
344 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
345 {
346 	struct vnode *vp;
347 	struct nameidata nd;
348 	char *buf;
349 	unsigned long ndflags;
350 	int error;
351 
352 	if (req->newptr == NULL)
353 		return (EINVAL);
354 	if (req->newlen >= PATH_MAX)
355 		return (E2BIG);
356 
357 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
358 	error = SYSCTL_IN(req, buf, req->newlen);
359 	if (error != 0)
360 		goto out;
361 
362 	buf[req->newlen] = '\0';
363 
364 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
365 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
366 	if ((error = namei(&nd)) != 0)
367 		goto out;
368 	vp = nd.ni_vp;
369 
370 	if (VN_IS_DOOMED(vp)) {
371 		/*
372 		 * This vnode is being recycled.  Return != 0 to let the caller
373 		 * know that the sysctl had no effect.  Return EAGAIN because a
374 		 * subsequent call will likely succeed (since namei will create
375 		 * a new vnode if necessary)
376 		 */
377 		error = EAGAIN;
378 		goto putvnode;
379 	}
380 
381 	counter_u64_add(recycles_count, 1);
382 	vgone(vp);
383 putvnode:
384 	NDFREE(&nd, 0);
385 out:
386 	free(buf, M_TEMP);
387 	return (error);
388 }
389 
390 static int
391 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
392 {
393 	struct thread *td = curthread;
394 	struct vnode *vp;
395 	struct file *fp;
396 	int error;
397 	int fd;
398 
399 	if (req->newptr == NULL)
400 		return (EBADF);
401 
402         error = sysctl_handle_int(oidp, &fd, 0, req);
403         if (error != 0)
404                 return (error);
405 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
406 	if (error != 0)
407 		return (error);
408 	vp = fp->f_vnode;
409 
410 	error = vn_lock(vp, LK_EXCLUSIVE);
411 	if (error != 0)
412 		goto drop;
413 
414 	counter_u64_add(recycles_count, 1);
415 	vgone(vp);
416 	VOP_UNLOCK(vp);
417 drop:
418 	fdrop(fp, td);
419 	return (error);
420 }
421 
422 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
423     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
424     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
425 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
426     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
427     sysctl_ftry_reclaim_vnode, "I",
428     "Try to reclaim a vnode by its file descriptor");
429 
430 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
431 static int vnsz2log;
432 
433 /*
434  * Support for the bufobj clean & dirty pctrie.
435  */
436 static void *
437 buf_trie_alloc(struct pctrie *ptree)
438 {
439 
440 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
441 }
442 
443 static void
444 buf_trie_free(struct pctrie *ptree, void *node)
445 {
446 
447 	uma_zfree(buf_trie_zone, node);
448 }
449 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
450 
451 /*
452  * Initialize the vnode management data structures.
453  *
454  * Reevaluate the following cap on the number of vnodes after the physical
455  * memory size exceeds 512GB.  In the limit, as the physical memory size
456  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
457  */
458 #ifndef	MAXVNODES_MAX
459 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
460 #endif
461 
462 /*
463  * Initialize a vnode as it first enters the zone.
464  */
465 static int
466 vnode_init(void *mem, int size, int flags)
467 {
468 	struct vnode *vp;
469 
470 	vp = mem;
471 	bzero(vp, size);
472 	/*
473 	 * Setup locks.
474 	 */
475 	vp->v_vnlock = &vp->v_lock;
476 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
477 	/*
478 	 * By default, don't allow shared locks unless filesystems opt-in.
479 	 */
480 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
481 	    LK_NOSHARE | LK_IS_VNODE);
482 	/*
483 	 * Initialize bufobj.
484 	 */
485 	bufobj_init(&vp->v_bufobj, vp);
486 	/*
487 	 * Initialize namecache.
488 	 */
489 	LIST_INIT(&vp->v_cache_src);
490 	TAILQ_INIT(&vp->v_cache_dst);
491 	/*
492 	 * Initialize rangelocks.
493 	 */
494 	rangelock_init(&vp->v_rl);
495 	return (0);
496 }
497 
498 /*
499  * Free a vnode when it is cleared from the zone.
500  */
501 static void
502 vnode_fini(void *mem, int size)
503 {
504 	struct vnode *vp;
505 	struct bufobj *bo;
506 
507 	vp = mem;
508 	rangelock_destroy(&vp->v_rl);
509 	lockdestroy(vp->v_vnlock);
510 	mtx_destroy(&vp->v_interlock);
511 	bo = &vp->v_bufobj;
512 	rw_destroy(BO_LOCKPTR(bo));
513 }
514 
515 /*
516  * Provide the size of NFS nclnode and NFS fh for calculation of the
517  * vnode memory consumption.  The size is specified directly to
518  * eliminate dependency on NFS-private header.
519  *
520  * Other filesystems may use bigger or smaller (like UFS and ZFS)
521  * private inode data, but the NFS-based estimation is ample enough.
522  * Still, we care about differences in the size between 64- and 32-bit
523  * platforms.
524  *
525  * Namecache structure size is heuristically
526  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
527  */
528 #ifdef _LP64
529 #define	NFS_NCLNODE_SZ	(528 + 64)
530 #define	NC_SZ		148
531 #else
532 #define	NFS_NCLNODE_SZ	(360 + 32)
533 #define	NC_SZ		92
534 #endif
535 
536 static void
537 vntblinit(void *dummy __unused)
538 {
539 	u_int i;
540 	int physvnodes, virtvnodes;
541 
542 	/*
543 	 * Desiredvnodes is a function of the physical memory size and the
544 	 * kernel's heap size.  Generally speaking, it scales with the
545 	 * physical memory size.  The ratio of desiredvnodes to the physical
546 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
547 	 * Thereafter, the
548 	 * marginal ratio of desiredvnodes to the physical memory size is
549 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
550 	 * size.  The memory required by desiredvnodes vnodes and vm objects
551 	 * must not exceed 1/10th of the kernel's heap size.
552 	 */
553 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
554 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
555 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
556 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
557 	desiredvnodes = min(physvnodes, virtvnodes);
558 	if (desiredvnodes > MAXVNODES_MAX) {
559 		if (bootverbose)
560 			printf("Reducing kern.maxvnodes %d -> %d\n",
561 			    desiredvnodes, MAXVNODES_MAX);
562 		desiredvnodes = MAXVNODES_MAX;
563 	}
564 	wantfreevnodes = desiredvnodes / 4;
565 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
566 	TAILQ_INIT(&vnode_free_list);
567 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
568 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
569 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
570 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
571 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
572 	/*
573 	 * Preallocate enough nodes to support one-per buf so that
574 	 * we can not fail an insert.  reassignbuf() callers can not
575 	 * tolerate the insertion failure.
576 	 */
577 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
578 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
579 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
580 	uma_prealloc(buf_trie_zone, nbuf);
581 
582 	vnodes_created = counter_u64_alloc(M_WAITOK);
583 	recycles_count = counter_u64_alloc(M_WAITOK);
584 	free_owe_inact = counter_u64_alloc(M_WAITOK);
585 
586 	/*
587 	 * Initialize the filesystem syncer.
588 	 */
589 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
590 	    &syncer_mask);
591 	syncer_maxdelay = syncer_mask + 1;
592 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
593 	cv_init(&sync_wakeup, "syncer");
594 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
595 		vnsz2log++;
596 	vnsz2log--;
597 }
598 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
599 
600 
601 /*
602  * Mark a mount point as busy. Used to synchronize access and to delay
603  * unmounting. Eventually, mountlist_mtx is not released on failure.
604  *
605  * vfs_busy() is a custom lock, it can block the caller.
606  * vfs_busy() only sleeps if the unmount is active on the mount point.
607  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
608  * vnode belonging to mp.
609  *
610  * Lookup uses vfs_busy() to traverse mount points.
611  * root fs			var fs
612  * / vnode lock		A	/ vnode lock (/var)		D
613  * /var vnode lock	B	/log vnode lock(/var/log)	E
614  * vfs_busy lock	C	vfs_busy lock			F
615  *
616  * Within each file system, the lock order is C->A->B and F->D->E.
617  *
618  * When traversing across mounts, the system follows that lock order:
619  *
620  *        C->A->B
621  *              |
622  *              +->F->D->E
623  *
624  * The lookup() process for namei("/var") illustrates the process:
625  *  VOP_LOOKUP() obtains B while A is held
626  *  vfs_busy() obtains a shared lock on F while A and B are held
627  *  vput() releases lock on B
628  *  vput() releases lock on A
629  *  VFS_ROOT() obtains lock on D while shared lock on F is held
630  *  vfs_unbusy() releases shared lock on F
631  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
632  *    Attempt to lock A (instead of vp_crossmp) while D is held would
633  *    violate the global order, causing deadlocks.
634  *
635  * dounmount() locks B while F is drained.
636  */
637 int
638 vfs_busy(struct mount *mp, int flags)
639 {
640 
641 	MPASS((flags & ~MBF_MASK) == 0);
642 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
643 
644 	if (vfs_op_thread_enter(mp)) {
645 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
646 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
647 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
648 		vfs_mp_count_add_pcpu(mp, ref, 1);
649 		vfs_mp_count_add_pcpu(mp, lockref, 1);
650 		vfs_op_thread_exit(mp);
651 		if (flags & MBF_MNTLSTLOCK)
652 			mtx_unlock(&mountlist_mtx);
653 		return (0);
654 	}
655 
656 	MNT_ILOCK(mp);
657 	vfs_assert_mount_counters(mp);
658 	MNT_REF(mp);
659 	/*
660 	 * If mount point is currently being unmounted, sleep until the
661 	 * mount point fate is decided.  If thread doing the unmounting fails,
662 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
663 	 * that this mount point has survived the unmount attempt and vfs_busy
664 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
665 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
666 	 * about to be really destroyed.  vfs_busy needs to release its
667 	 * reference on the mount point in this case and return with ENOENT,
668 	 * telling the caller that mount mount it tried to busy is no longer
669 	 * valid.
670 	 */
671 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
672 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
673 			MNT_REL(mp);
674 			MNT_IUNLOCK(mp);
675 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
676 			    __func__);
677 			return (ENOENT);
678 		}
679 		if (flags & MBF_MNTLSTLOCK)
680 			mtx_unlock(&mountlist_mtx);
681 		mp->mnt_kern_flag |= MNTK_MWAIT;
682 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
683 		if (flags & MBF_MNTLSTLOCK)
684 			mtx_lock(&mountlist_mtx);
685 		MNT_ILOCK(mp);
686 	}
687 	if (flags & MBF_MNTLSTLOCK)
688 		mtx_unlock(&mountlist_mtx);
689 	mp->mnt_lockref++;
690 	MNT_IUNLOCK(mp);
691 	return (0);
692 }
693 
694 /*
695  * Free a busy filesystem.
696  */
697 void
698 vfs_unbusy(struct mount *mp)
699 {
700 	int c;
701 
702 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
703 
704 	if (vfs_op_thread_enter(mp)) {
705 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
706 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
707 		vfs_mp_count_sub_pcpu(mp, ref, 1);
708 		vfs_op_thread_exit(mp);
709 		return;
710 	}
711 
712 	MNT_ILOCK(mp);
713 	vfs_assert_mount_counters(mp);
714 	MNT_REL(mp);
715 	c = --mp->mnt_lockref;
716 	if (mp->mnt_vfs_ops == 0) {
717 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
718 		MNT_IUNLOCK(mp);
719 		return;
720 	}
721 	if (c < 0)
722 		vfs_dump_mount_counters(mp);
723 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
724 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
725 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
726 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
727 		wakeup(&mp->mnt_lockref);
728 	}
729 	MNT_IUNLOCK(mp);
730 }
731 
732 /*
733  * Lookup a mount point by filesystem identifier.
734  */
735 struct mount *
736 vfs_getvfs(fsid_t *fsid)
737 {
738 	struct mount *mp;
739 
740 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
741 	mtx_lock(&mountlist_mtx);
742 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
743 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
744 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
745 			vfs_ref(mp);
746 			mtx_unlock(&mountlist_mtx);
747 			return (mp);
748 		}
749 	}
750 	mtx_unlock(&mountlist_mtx);
751 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
752 	return ((struct mount *) 0);
753 }
754 
755 /*
756  * Lookup a mount point by filesystem identifier, busying it before
757  * returning.
758  *
759  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
760  * cache for popular filesystem identifiers.  The cache is lockess, using
761  * the fact that struct mount's are never freed.  In worst case we may
762  * get pointer to unmounted or even different filesystem, so we have to
763  * check what we got, and go slow way if so.
764  */
765 struct mount *
766 vfs_busyfs(fsid_t *fsid)
767 {
768 #define	FSID_CACHE_SIZE	256
769 	typedef struct mount * volatile vmp_t;
770 	static vmp_t cache[FSID_CACHE_SIZE];
771 	struct mount *mp;
772 	int error;
773 	uint32_t hash;
774 
775 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
776 	hash = fsid->val[0] ^ fsid->val[1];
777 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
778 	mp = cache[hash];
779 	if (mp == NULL ||
780 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
781 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
782 		goto slow;
783 	if (vfs_busy(mp, 0) != 0) {
784 		cache[hash] = NULL;
785 		goto slow;
786 	}
787 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
788 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
789 		return (mp);
790 	else
791 	    vfs_unbusy(mp);
792 
793 slow:
794 	mtx_lock(&mountlist_mtx);
795 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
796 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
797 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
798 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
799 			if (error) {
800 				cache[hash] = NULL;
801 				mtx_unlock(&mountlist_mtx);
802 				return (NULL);
803 			}
804 			cache[hash] = mp;
805 			return (mp);
806 		}
807 	}
808 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
809 	mtx_unlock(&mountlist_mtx);
810 	return ((struct mount *) 0);
811 }
812 
813 /*
814  * Check if a user can access privileged mount options.
815  */
816 int
817 vfs_suser(struct mount *mp, struct thread *td)
818 {
819 	int error;
820 
821 	if (jailed(td->td_ucred)) {
822 		/*
823 		 * If the jail of the calling thread lacks permission for
824 		 * this type of file system, deny immediately.
825 		 */
826 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
827 			return (EPERM);
828 
829 		/*
830 		 * If the file system was mounted outside the jail of the
831 		 * calling thread, deny immediately.
832 		 */
833 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
834 			return (EPERM);
835 	}
836 
837 	/*
838 	 * If file system supports delegated administration, we don't check
839 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
840 	 * by the file system itself.
841 	 * If this is not the user that did original mount, we check for
842 	 * the PRIV_VFS_MOUNT_OWNER privilege.
843 	 */
844 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
845 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
846 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
847 			return (error);
848 	}
849 	return (0);
850 }
851 
852 /*
853  * Get a new unique fsid.  Try to make its val[0] unique, since this value
854  * will be used to create fake device numbers for stat().  Also try (but
855  * not so hard) make its val[0] unique mod 2^16, since some emulators only
856  * support 16-bit device numbers.  We end up with unique val[0]'s for the
857  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
858  *
859  * Keep in mind that several mounts may be running in parallel.  Starting
860  * the search one past where the previous search terminated is both a
861  * micro-optimization and a defense against returning the same fsid to
862  * different mounts.
863  */
864 void
865 vfs_getnewfsid(struct mount *mp)
866 {
867 	static uint16_t mntid_base;
868 	struct mount *nmp;
869 	fsid_t tfsid;
870 	int mtype;
871 
872 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
873 	mtx_lock(&mntid_mtx);
874 	mtype = mp->mnt_vfc->vfc_typenum;
875 	tfsid.val[1] = mtype;
876 	mtype = (mtype & 0xFF) << 24;
877 	for (;;) {
878 		tfsid.val[0] = makedev(255,
879 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
880 		mntid_base++;
881 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
882 			break;
883 		vfs_rel(nmp);
884 	}
885 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
886 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
887 	mtx_unlock(&mntid_mtx);
888 }
889 
890 /*
891  * Knob to control the precision of file timestamps:
892  *
893  *   0 = seconds only; nanoseconds zeroed.
894  *   1 = seconds and nanoseconds, accurate within 1/HZ.
895  *   2 = seconds and nanoseconds, truncated to microseconds.
896  * >=3 = seconds and nanoseconds, maximum precision.
897  */
898 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
899 
900 static int timestamp_precision = TSP_USEC;
901 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
902     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
903     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
904     "3+: sec + ns (max. precision))");
905 
906 /*
907  * Get a current timestamp.
908  */
909 void
910 vfs_timestamp(struct timespec *tsp)
911 {
912 	struct timeval tv;
913 
914 	switch (timestamp_precision) {
915 	case TSP_SEC:
916 		tsp->tv_sec = time_second;
917 		tsp->tv_nsec = 0;
918 		break;
919 	case TSP_HZ:
920 		getnanotime(tsp);
921 		break;
922 	case TSP_USEC:
923 		microtime(&tv);
924 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
925 		break;
926 	case TSP_NSEC:
927 	default:
928 		nanotime(tsp);
929 		break;
930 	}
931 }
932 
933 /*
934  * Set vnode attributes to VNOVAL
935  */
936 void
937 vattr_null(struct vattr *vap)
938 {
939 
940 	vap->va_type = VNON;
941 	vap->va_size = VNOVAL;
942 	vap->va_bytes = VNOVAL;
943 	vap->va_mode = VNOVAL;
944 	vap->va_nlink = VNOVAL;
945 	vap->va_uid = VNOVAL;
946 	vap->va_gid = VNOVAL;
947 	vap->va_fsid = VNOVAL;
948 	vap->va_fileid = VNOVAL;
949 	vap->va_blocksize = VNOVAL;
950 	vap->va_rdev = VNOVAL;
951 	vap->va_atime.tv_sec = VNOVAL;
952 	vap->va_atime.tv_nsec = VNOVAL;
953 	vap->va_mtime.tv_sec = VNOVAL;
954 	vap->va_mtime.tv_nsec = VNOVAL;
955 	vap->va_ctime.tv_sec = VNOVAL;
956 	vap->va_ctime.tv_nsec = VNOVAL;
957 	vap->va_birthtime.tv_sec = VNOVAL;
958 	vap->va_birthtime.tv_nsec = VNOVAL;
959 	vap->va_flags = VNOVAL;
960 	vap->va_gen = VNOVAL;
961 	vap->va_vaflags = 0;
962 }
963 
964 /*
965  * This routine is called when we have too many vnodes.  It attempts
966  * to free <count> vnodes and will potentially free vnodes that still
967  * have VM backing store (VM backing store is typically the cause
968  * of a vnode blowout so we want to do this).  Therefore, this operation
969  * is not considered cheap.
970  *
971  * A number of conditions may prevent a vnode from being reclaimed.
972  * the buffer cache may have references on the vnode, a directory
973  * vnode may still have references due to the namei cache representing
974  * underlying files, or the vnode may be in active use.   It is not
975  * desirable to reuse such vnodes.  These conditions may cause the
976  * number of vnodes to reach some minimum value regardless of what
977  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
978  *
979  * @param mp		 Try to reclaim vnodes from this mountpoint
980  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
981  * 			 entries if this argument is strue
982  * @param trigger	 Only reclaim vnodes with fewer than this many resident
983  *			 pages.
984  * @return		 The number of vnodes that were reclaimed.
985  */
986 static int
987 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
988 {
989 	struct vnode *vp;
990 	int count, done, target;
991 
992 	done = 0;
993 	vn_start_write(NULL, &mp, V_WAIT);
994 	MNT_ILOCK(mp);
995 	count = mp->mnt_nvnodelistsize;
996 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
997 	target = target / 10 + 1;
998 	while (count != 0 && done < target) {
999 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1000 		while (vp != NULL && vp->v_type == VMARKER)
1001 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
1002 		if (vp == NULL)
1003 			break;
1004 		/*
1005 		 * XXX LRU is completely broken for non-free vnodes.  First
1006 		 * by calling here in mountpoint order, then by moving
1007 		 * unselected vnodes to the end here, and most grossly by
1008 		 * removing the vlruvp() function that was supposed to
1009 		 * maintain the order.  (This function was born broken
1010 		 * since syncer problems prevented it doing anything.)  The
1011 		 * order is closer to LRC (C = Created).
1012 		 *
1013 		 * LRU reclaiming of vnodes seems to have last worked in
1014 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
1015 		 * Then there was no hold count, and inactive vnodes were
1016 		 * simply put on the free list in LRU order.  The separate
1017 		 * lists also break LRU.  We prefer to reclaim from the
1018 		 * free list for technical reasons.  This tends to thrash
1019 		 * the free list to keep very unrecently used held vnodes.
1020 		 * The problem is mitigated by keeping the free list large.
1021 		 */
1022 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1023 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1024 		--count;
1025 		if (!VI_TRYLOCK(vp))
1026 			goto next_iter;
1027 		/*
1028 		 * If it's been deconstructed already, it's still
1029 		 * referenced, or it exceeds the trigger, skip it.
1030 		 * Also skip free vnodes.  We are trying to make space
1031 		 * to expand the free list, not reduce it.
1032 		 */
1033 		if (vp->v_usecount ||
1034 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1035 		    ((vp->v_iflag & VI_FREE) != 0) ||
1036 		    VN_IS_DOOMED(vp) || (vp->v_object != NULL &&
1037 		    vp->v_object->resident_page_count > trigger)) {
1038 			VI_UNLOCK(vp);
1039 			goto next_iter;
1040 		}
1041 		MNT_IUNLOCK(mp);
1042 		vholdl(vp);
1043 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1044 			vdrop(vp);
1045 			goto next_iter_mntunlocked;
1046 		}
1047 		VI_LOCK(vp);
1048 		/*
1049 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1050 		 * the vnode interlock and before it was locked again.
1051 		 *
1052 		 * It is not necessary to recheck VIRF_DOOMED because it can
1053 		 * only be set by another thread that holds both the vnode
1054 		 * lock and vnode interlock.  If another thread has the
1055 		 * vnode lock before we get to VOP_LOCK() and obtains the
1056 		 * vnode interlock after VOP_LOCK() drops the vnode
1057 		 * interlock, the other thread will be unable to drop the
1058 		 * vnode lock before our VOP_LOCK() call fails.
1059 		 */
1060 		if (vp->v_usecount ||
1061 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1062 		    (vp->v_object != NULL &&
1063 		    vp->v_object->resident_page_count > trigger)) {
1064 			VOP_UNLOCK(vp);
1065 			vdropl(vp);
1066 			goto next_iter_mntunlocked;
1067 		}
1068 		KASSERT(!VN_IS_DOOMED(vp),
1069 		    ("VIRF_DOOMED unexpectedly detected in vlrureclaim()"));
1070 		counter_u64_add(recycles_count, 1);
1071 		vgonel(vp);
1072 		VOP_UNLOCK(vp);
1073 		vdropl(vp);
1074 		done++;
1075 next_iter_mntunlocked:
1076 		if (!should_yield())
1077 			goto relock_mnt;
1078 		goto yield;
1079 next_iter:
1080 		if (!should_yield())
1081 			continue;
1082 		MNT_IUNLOCK(mp);
1083 yield:
1084 		kern_yield(PRI_USER);
1085 relock_mnt:
1086 		MNT_ILOCK(mp);
1087 	}
1088 	MNT_IUNLOCK(mp);
1089 	vn_finished_write(mp);
1090 	return done;
1091 }
1092 
1093 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1094 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1095     0,
1096     "limit on vnode free requests per call to the vnlru_free routine");
1097 
1098 /*
1099  * Attempt to reduce the free list by the requested amount.
1100  */
1101 static void
1102 vnlru_free_locked(int count, struct vfsops *mnt_op)
1103 {
1104 	struct vnode *vp;
1105 	struct mount *mp;
1106 	bool tried_batches;
1107 
1108 	tried_batches = false;
1109 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1110 	if (count > max_vnlru_free)
1111 		count = max_vnlru_free;
1112 	for (; count > 0; count--) {
1113 		vp = TAILQ_FIRST(&vnode_free_list);
1114 		/*
1115 		 * The list can be modified while the free_list_mtx
1116 		 * has been dropped and vp could be NULL here.
1117 		 */
1118 		if (vp == NULL) {
1119 			if (tried_batches)
1120 				break;
1121 			mtx_unlock(&vnode_free_list_mtx);
1122 			vnlru_return_batches(mnt_op);
1123 			tried_batches = true;
1124 			mtx_lock(&vnode_free_list_mtx);
1125 			continue;
1126 		}
1127 
1128 		VNASSERT(vp->v_op != NULL, vp,
1129 		    ("vnlru_free: vnode already reclaimed."));
1130 		KASSERT((vp->v_iflag & VI_FREE) != 0,
1131 		    ("Removing vnode not on freelist"));
1132 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1133 		    ("Mangling active vnode"));
1134 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1135 
1136 		/*
1137 		 * Don't recycle if our vnode is from different type
1138 		 * of mount point.  Note that mp is type-safe, the
1139 		 * check does not reach unmapped address even if
1140 		 * vnode is reclaimed.
1141 		 * Don't recycle if we can't get the interlock without
1142 		 * blocking.
1143 		 */
1144 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1145 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1146 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1147 			continue;
1148 		}
1149 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1150 		    vp, ("vp inconsistent on freelist"));
1151 
1152 		/*
1153 		 * The clear of VI_FREE prevents activation of the
1154 		 * vnode.  There is no sense in putting the vnode on
1155 		 * the mount point active list, only to remove it
1156 		 * later during recycling.  Inline the relevant part
1157 		 * of vholdl(), to avoid triggering assertions or
1158 		 * activating.
1159 		 */
1160 		freevnodes--;
1161 		vp->v_iflag &= ~VI_FREE;
1162 		VNODE_REFCOUNT_FENCE_REL();
1163 		refcount_acquire(&vp->v_holdcnt);
1164 
1165 		mtx_unlock(&vnode_free_list_mtx);
1166 		VI_UNLOCK(vp);
1167 		vtryrecycle(vp);
1168 		/*
1169 		 * If the recycled succeeded this vdrop will actually free
1170 		 * the vnode.  If not it will simply place it back on
1171 		 * the free list.
1172 		 */
1173 		vdrop(vp);
1174 		mtx_lock(&vnode_free_list_mtx);
1175 	}
1176 }
1177 
1178 void
1179 vnlru_free(int count, struct vfsops *mnt_op)
1180 {
1181 
1182 	mtx_lock(&vnode_free_list_mtx);
1183 	vnlru_free_locked(count, mnt_op);
1184 	mtx_unlock(&vnode_free_list_mtx);
1185 }
1186 
1187 
1188 /* XXX some names and initialization are bad for limits and watermarks. */
1189 static int
1190 vspace(void)
1191 {
1192 	int space;
1193 
1194 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1195 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1196 	vlowat = vhiwat / 2;
1197 	if (numvnodes > desiredvnodes)
1198 		return (0);
1199 	space = desiredvnodes - numvnodes;
1200 	if (freevnodes > wantfreevnodes)
1201 		space += freevnodes - wantfreevnodes;
1202 	return (space);
1203 }
1204 
1205 static void
1206 vnlru_return_batch_locked(struct mount *mp)
1207 {
1208 	struct vnode *vp;
1209 
1210 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1211 
1212 	if (mp->mnt_tmpfreevnodelistsize == 0)
1213 		return;
1214 
1215 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1216 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1217 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1218 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1219 	}
1220 	mtx_lock(&vnode_free_list_mtx);
1221 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1222 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1223 	mtx_unlock(&vnode_free_list_mtx);
1224 	mp->mnt_tmpfreevnodelistsize = 0;
1225 }
1226 
1227 static void
1228 vnlru_return_batch(struct mount *mp)
1229 {
1230 
1231 	mtx_lock(&mp->mnt_listmtx);
1232 	vnlru_return_batch_locked(mp);
1233 	mtx_unlock(&mp->mnt_listmtx);
1234 }
1235 
1236 static void
1237 vnlru_return_batches(struct vfsops *mnt_op)
1238 {
1239 	struct mount *mp, *nmp;
1240 	bool need_unbusy;
1241 
1242 	mtx_lock(&mountlist_mtx);
1243 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1244 		need_unbusy = false;
1245 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1246 			goto next;
1247 		if (mp->mnt_tmpfreevnodelistsize == 0)
1248 			goto next;
1249 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1250 			vnlru_return_batch(mp);
1251 			need_unbusy = true;
1252 			mtx_lock(&mountlist_mtx);
1253 		}
1254 next:
1255 		nmp = TAILQ_NEXT(mp, mnt_list);
1256 		if (need_unbusy)
1257 			vfs_unbusy(mp);
1258 	}
1259 	mtx_unlock(&mountlist_mtx);
1260 }
1261 
1262 /*
1263  * Attempt to recycle vnodes in a context that is always safe to block.
1264  * Calling vlrurecycle() from the bowels of filesystem code has some
1265  * interesting deadlock problems.
1266  */
1267 static struct proc *vnlruproc;
1268 static int vnlruproc_sig;
1269 
1270 static void
1271 vnlru_proc(void)
1272 {
1273 	struct mount *mp, *nmp;
1274 	unsigned long onumvnodes;
1275 	int done, force, trigger, usevnodes, vsp;
1276 	bool reclaim_nc_src;
1277 
1278 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1279 	    SHUTDOWN_PRI_FIRST);
1280 
1281 	force = 0;
1282 	for (;;) {
1283 		kproc_suspend_check(vnlruproc);
1284 		mtx_lock(&vnode_free_list_mtx);
1285 		/*
1286 		 * If numvnodes is too large (due to desiredvnodes being
1287 		 * adjusted using its sysctl, or emergency growth), first
1288 		 * try to reduce it by discarding from the free list.
1289 		 */
1290 		if (numvnodes > desiredvnodes)
1291 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1292 		/*
1293 		 * Sleep if the vnode cache is in a good state.  This is
1294 		 * when it is not over-full and has space for about a 4%
1295 		 * or 9% expansion (by growing its size or inexcessively
1296 		 * reducing its free list).  Otherwise, try to reclaim
1297 		 * space for a 10% expansion.
1298 		 */
1299 		if (vstir && force == 0) {
1300 			force = 1;
1301 			vstir = 0;
1302 		}
1303 		vsp = vspace();
1304 		if (vsp >= vlowat && force == 0) {
1305 			vnlruproc_sig = 0;
1306 			wakeup(&vnlruproc_sig);
1307 			msleep(vnlruproc, &vnode_free_list_mtx,
1308 			    PVFS|PDROP, "vlruwt", hz);
1309 			continue;
1310 		}
1311 		mtx_unlock(&vnode_free_list_mtx);
1312 		done = 0;
1313 		onumvnodes = numvnodes;
1314 		/*
1315 		 * Calculate parameters for recycling.  These are the same
1316 		 * throughout the loop to give some semblance of fairness.
1317 		 * The trigger point is to avoid recycling vnodes with lots
1318 		 * of resident pages.  We aren't trying to free memory; we
1319 		 * are trying to recycle or at least free vnodes.
1320 		 */
1321 		if (numvnodes <= desiredvnodes)
1322 			usevnodes = numvnodes - freevnodes;
1323 		else
1324 			usevnodes = numvnodes;
1325 		if (usevnodes <= 0)
1326 			usevnodes = 1;
1327 		/*
1328 		 * The trigger value is is chosen to give a conservatively
1329 		 * large value to ensure that it alone doesn't prevent
1330 		 * making progress.  The value can easily be so large that
1331 		 * it is effectively infinite in some congested and
1332 		 * misconfigured cases, and this is necessary.  Normally
1333 		 * it is about 8 to 100 (pages), which is quite large.
1334 		 */
1335 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1336 		if (force < 2)
1337 			trigger = vsmalltrigger;
1338 		reclaim_nc_src = force >= 3;
1339 		mtx_lock(&mountlist_mtx);
1340 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1341 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1342 				nmp = TAILQ_NEXT(mp, mnt_list);
1343 				continue;
1344 			}
1345 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1346 			mtx_lock(&mountlist_mtx);
1347 			nmp = TAILQ_NEXT(mp, mnt_list);
1348 			vfs_unbusy(mp);
1349 		}
1350 		mtx_unlock(&mountlist_mtx);
1351 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1352 			uma_reclaim(UMA_RECLAIM_DRAIN);
1353 		if (done == 0) {
1354 			if (force == 0 || force == 1) {
1355 				force = 2;
1356 				continue;
1357 			}
1358 			if (force == 2) {
1359 				force = 3;
1360 				continue;
1361 			}
1362 			force = 0;
1363 			vnlru_nowhere++;
1364 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1365 		} else
1366 			kern_yield(PRI_USER);
1367 		/*
1368 		 * After becoming active to expand above low water, keep
1369 		 * active until above high water.
1370 		 */
1371 		vsp = vspace();
1372 		force = vsp < vhiwat;
1373 	}
1374 }
1375 
1376 static struct kproc_desc vnlru_kp = {
1377 	"vnlru",
1378 	vnlru_proc,
1379 	&vnlruproc
1380 };
1381 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1382     &vnlru_kp);
1383 
1384 /*
1385  * Routines having to do with the management of the vnode table.
1386  */
1387 
1388 /*
1389  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1390  * before we actually vgone().  This function must be called with the vnode
1391  * held to prevent the vnode from being returned to the free list midway
1392  * through vgone().
1393  */
1394 static int
1395 vtryrecycle(struct vnode *vp)
1396 {
1397 	struct mount *vnmp;
1398 
1399 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1400 	VNASSERT(vp->v_holdcnt, vp,
1401 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1402 	/*
1403 	 * This vnode may found and locked via some other list, if so we
1404 	 * can't recycle it yet.
1405 	 */
1406 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1407 		CTR2(KTR_VFS,
1408 		    "%s: impossible to recycle, vp %p lock is already held",
1409 		    __func__, vp);
1410 		return (EWOULDBLOCK);
1411 	}
1412 	/*
1413 	 * Don't recycle if its filesystem is being suspended.
1414 	 */
1415 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1416 		VOP_UNLOCK(vp);
1417 		CTR2(KTR_VFS,
1418 		    "%s: impossible to recycle, cannot start the write for %p",
1419 		    __func__, vp);
1420 		return (EBUSY);
1421 	}
1422 	/*
1423 	 * If we got this far, we need to acquire the interlock and see if
1424 	 * anyone picked up this vnode from another list.  If not, we will
1425 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1426 	 * will skip over it.
1427 	 */
1428 	VI_LOCK(vp);
1429 	if (vp->v_usecount) {
1430 		VOP_UNLOCK(vp);
1431 		VI_UNLOCK(vp);
1432 		vn_finished_write(vnmp);
1433 		CTR2(KTR_VFS,
1434 		    "%s: impossible to recycle, %p is already referenced",
1435 		    __func__, vp);
1436 		return (EBUSY);
1437 	}
1438 	if (!VN_IS_DOOMED(vp)) {
1439 		counter_u64_add(recycles_count, 1);
1440 		vgonel(vp);
1441 	}
1442 	VOP_UNLOCK(vp);
1443 	VI_UNLOCK(vp);
1444 	vn_finished_write(vnmp);
1445 	return (0);
1446 }
1447 
1448 static void
1449 vcheckspace(void)
1450 {
1451 	int vsp;
1452 
1453 	vsp = vspace();
1454 	if (vsp < vlowat && vnlruproc_sig == 0) {
1455 		vnlruproc_sig = 1;
1456 		wakeup(vnlruproc);
1457 	}
1458 }
1459 
1460 /*
1461  * Wait if necessary for space for a new vnode.
1462  */
1463 static int
1464 getnewvnode_wait(int suspended)
1465 {
1466 
1467 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1468 	if (numvnodes >= desiredvnodes) {
1469 		if (suspended) {
1470 			/*
1471 			 * The file system is being suspended.  We cannot
1472 			 * risk a deadlock here, so allow allocation of
1473 			 * another vnode even if this would give too many.
1474 			 */
1475 			return (0);
1476 		}
1477 		if (vnlruproc_sig == 0) {
1478 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1479 			wakeup(vnlruproc);
1480 		}
1481 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1482 		    "vlruwk", hz);
1483 	}
1484 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1485 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1486 		vnlru_free_locked(1, NULL);
1487 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1488 }
1489 
1490 /*
1491  * This hack is fragile, and probably not needed any more now that the
1492  * watermark handling works.
1493  */
1494 void
1495 getnewvnode_reserve(u_int count)
1496 {
1497 	struct thread *td;
1498 
1499 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1500 	/* XXX no longer so quick, but this part is not racy. */
1501 	mtx_lock(&vnode_free_list_mtx);
1502 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1503 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1504 		    freevnodes - wantfreevnodes), NULL);
1505 	mtx_unlock(&vnode_free_list_mtx);
1506 
1507 	td = curthread;
1508 	/* First try to be quick and racy. */
1509 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1510 		td->td_vp_reserv += count;
1511 		vcheckspace();	/* XXX no longer so quick, but more racy */
1512 		return;
1513 	} else
1514 		atomic_subtract_long(&numvnodes, count);
1515 
1516 	mtx_lock(&vnode_free_list_mtx);
1517 	while (count > 0) {
1518 		if (getnewvnode_wait(0) == 0) {
1519 			count--;
1520 			td->td_vp_reserv++;
1521 			atomic_add_long(&numvnodes, 1);
1522 		}
1523 	}
1524 	vcheckspace();
1525 	mtx_unlock(&vnode_free_list_mtx);
1526 }
1527 
1528 /*
1529  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1530  * misconfgured or changed significantly.  Reducing desiredvnodes below
1531  * the reserved amount should cause bizarre behaviour like reducing it
1532  * below the number of active vnodes -- the system will try to reduce
1533  * numvnodes to match, but should fail, so the subtraction below should
1534  * not overflow.
1535  */
1536 void
1537 getnewvnode_drop_reserve(void)
1538 {
1539 	struct thread *td;
1540 
1541 	td = curthread;
1542 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1543 	td->td_vp_reserv = 0;
1544 }
1545 
1546 /*
1547  * Return the next vnode from the free list.
1548  */
1549 int
1550 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1551     struct vnode **vpp)
1552 {
1553 	struct vnode *vp;
1554 	struct thread *td;
1555 	struct lock_object *lo;
1556 	static int cyclecount;
1557 	int error __unused;
1558 
1559 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1560 
1561 	KASSERT(vops->registered,
1562 	    ("%s: not registered vector op %p\n", __func__, vops));
1563 
1564 	vp = NULL;
1565 	td = curthread;
1566 	if (td->td_vp_reserv > 0) {
1567 		td->td_vp_reserv -= 1;
1568 		goto alloc;
1569 	}
1570 	mtx_lock(&vnode_free_list_mtx);
1571 	if (numvnodes < desiredvnodes)
1572 		cyclecount = 0;
1573 	else if (cyclecount++ >= freevnodes) {
1574 		cyclecount = 0;
1575 		vstir = 1;
1576 	}
1577 	/*
1578 	 * Grow the vnode cache if it will not be above its target max
1579 	 * after growing.  Otherwise, if the free list is nonempty, try
1580 	 * to reclaim 1 item from it before growing the cache (possibly
1581 	 * above its target max if the reclamation failed or is delayed).
1582 	 * Otherwise, wait for some space.  In all cases, schedule
1583 	 * vnlru_proc() if we are getting short of space.  The watermarks
1584 	 * should be chosen so that we never wait or even reclaim from
1585 	 * the free list to below its target minimum.
1586 	 */
1587 	if (numvnodes + 1 <= desiredvnodes)
1588 		;
1589 	else if (freevnodes > 0)
1590 		vnlru_free_locked(1, NULL);
1591 	else {
1592 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1593 		    MNTK_SUSPEND));
1594 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1595 		if (error != 0) {
1596 			mtx_unlock(&vnode_free_list_mtx);
1597 			return (error);
1598 		}
1599 #endif
1600 	}
1601 	vcheckspace();
1602 	atomic_add_long(&numvnodes, 1);
1603 	mtx_unlock(&vnode_free_list_mtx);
1604 alloc:
1605 	counter_u64_add(vnodes_created, 1);
1606 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1607 	/*
1608 	 * Locks are given the generic name "vnode" when created.
1609 	 * Follow the historic practice of using the filesystem
1610 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1611 	 *
1612 	 * Locks live in a witness group keyed on their name. Thus,
1613 	 * when a lock is renamed, it must also move from the witness
1614 	 * group of its old name to the witness group of its new name.
1615 	 *
1616 	 * The change only needs to be made when the vnode moves
1617 	 * from one filesystem type to another. We ensure that each
1618 	 * filesystem use a single static name pointer for its tag so
1619 	 * that we can compare pointers rather than doing a strcmp().
1620 	 */
1621 	lo = &vp->v_vnlock->lock_object;
1622 	if (lo->lo_name != tag) {
1623 		lo->lo_name = tag;
1624 		WITNESS_DESTROY(lo);
1625 		WITNESS_INIT(lo, tag);
1626 	}
1627 	/*
1628 	 * By default, don't allow shared locks unless filesystems opt-in.
1629 	 */
1630 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1631 	/*
1632 	 * Finalize various vnode identity bits.
1633 	 */
1634 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1635 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1636 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1637 	vp->v_type = VNON;
1638 	vp->v_tag = tag;
1639 	vp->v_op = vops;
1640 	v_init_counters(vp);
1641 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1642 #ifdef DIAGNOSTIC
1643 	if (mp == NULL && vops != &dead_vnodeops)
1644 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1645 #endif
1646 #ifdef MAC
1647 	mac_vnode_init(vp);
1648 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1649 		mac_vnode_associate_singlelabel(mp, vp);
1650 #endif
1651 	if (mp != NULL) {
1652 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1653 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1654 			vp->v_vflag |= VV_NOKNOTE;
1655 	}
1656 
1657 	/*
1658 	 * For the filesystems which do not use vfs_hash_insert(),
1659 	 * still initialize v_hash to have vfs_hash_index() useful.
1660 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1661 	 * its own hashing.
1662 	 */
1663 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1664 
1665 	*vpp = vp;
1666 	return (0);
1667 }
1668 
1669 static void
1670 freevnode(struct vnode *vp)
1671 {
1672 	struct bufobj *bo;
1673 
1674 	/*
1675 	 * The vnode has been marked for destruction, so free it.
1676 	 *
1677 	 * The vnode will be returned to the zone where it will
1678 	 * normally remain until it is needed for another vnode. We
1679 	 * need to cleanup (or verify that the cleanup has already
1680 	 * been done) any residual data left from its current use
1681 	 * so as not to contaminate the freshly allocated vnode.
1682 	 */
1683 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1684 	atomic_subtract_long(&numvnodes, 1);
1685 	bo = &vp->v_bufobj;
1686 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
1687 	    ("cleaned vnode still on the free list."));
1688 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1689 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1690 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1691 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1692 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1693 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1694 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1695 	    ("clean blk trie not empty"));
1696 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1697 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1698 	    ("dirty blk trie not empty"));
1699 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1700 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1701 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1702 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1703 	    ("Dangling rangelock waiters"));
1704 	VI_UNLOCK(vp);
1705 #ifdef MAC
1706 	mac_vnode_destroy(vp);
1707 #endif
1708 	if (vp->v_pollinfo != NULL) {
1709 		destroy_vpollinfo(vp->v_pollinfo);
1710 		vp->v_pollinfo = NULL;
1711 	}
1712 #ifdef INVARIANTS
1713 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1714 	vp->v_op = NULL;
1715 #endif
1716 	vp->v_mountedhere = NULL;
1717 	vp->v_unpcb = NULL;
1718 	vp->v_rdev = NULL;
1719 	vp->v_fifoinfo = NULL;
1720 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1721 	vp->v_irflag = 0;
1722 	vp->v_iflag = 0;
1723 	vp->v_vflag = 0;
1724 	bo->bo_flag = 0;
1725 	uma_zfree(vnode_zone, vp);
1726 }
1727 
1728 /*
1729  * Delete from old mount point vnode list, if on one.
1730  */
1731 static void
1732 delmntque(struct vnode *vp)
1733 {
1734 	struct mount *mp;
1735 
1736 	mp = vp->v_mount;
1737 	if (mp == NULL)
1738 		return;
1739 	MNT_ILOCK(mp);
1740 	VI_LOCK(vp);
1741 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1742 	    ("Active vnode list size %d > Vnode list size %d",
1743 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1744 	if (vp->v_iflag & VI_ACTIVE) {
1745 		vp->v_iflag &= ~VI_ACTIVE;
1746 		mtx_lock(&mp->mnt_listmtx);
1747 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1748 		mp->mnt_activevnodelistsize--;
1749 		mtx_unlock(&mp->mnt_listmtx);
1750 	}
1751 	vp->v_mount = NULL;
1752 	VI_UNLOCK(vp);
1753 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1754 		("bad mount point vnode list size"));
1755 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1756 	mp->mnt_nvnodelistsize--;
1757 	MNT_REL(mp);
1758 	MNT_IUNLOCK(mp);
1759 }
1760 
1761 static void
1762 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1763 {
1764 
1765 	vp->v_data = NULL;
1766 	vp->v_op = &dead_vnodeops;
1767 	vgone(vp);
1768 	vput(vp);
1769 }
1770 
1771 /*
1772  * Insert into list of vnodes for the new mount point, if available.
1773  */
1774 int
1775 insmntque1(struct vnode *vp, struct mount *mp,
1776 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1777 {
1778 
1779 	KASSERT(vp->v_mount == NULL,
1780 		("insmntque: vnode already on per mount vnode list"));
1781 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1782 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1783 
1784 	/*
1785 	 * We acquire the vnode interlock early to ensure that the
1786 	 * vnode cannot be recycled by another process releasing a
1787 	 * holdcnt on it before we get it on both the vnode list
1788 	 * and the active vnode list. The mount mutex protects only
1789 	 * manipulation of the vnode list and the vnode freelist
1790 	 * mutex protects only manipulation of the active vnode list.
1791 	 * Hence the need to hold the vnode interlock throughout.
1792 	 */
1793 	MNT_ILOCK(mp);
1794 	VI_LOCK(vp);
1795 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1796 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1797 	    mp->mnt_nvnodelistsize == 0)) &&
1798 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1799 		VI_UNLOCK(vp);
1800 		MNT_IUNLOCK(mp);
1801 		if (dtr != NULL)
1802 			dtr(vp, dtr_arg);
1803 		return (EBUSY);
1804 	}
1805 	vp->v_mount = mp;
1806 	MNT_REF(mp);
1807 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1808 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1809 		("neg mount point vnode list size"));
1810 	mp->mnt_nvnodelistsize++;
1811 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1812 	    ("Activating already active vnode"));
1813 	vp->v_iflag |= VI_ACTIVE;
1814 	mtx_lock(&mp->mnt_listmtx);
1815 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1816 	mp->mnt_activevnodelistsize++;
1817 	mtx_unlock(&mp->mnt_listmtx);
1818 	VI_UNLOCK(vp);
1819 	MNT_IUNLOCK(mp);
1820 	return (0);
1821 }
1822 
1823 int
1824 insmntque(struct vnode *vp, struct mount *mp)
1825 {
1826 
1827 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1828 }
1829 
1830 /*
1831  * Flush out and invalidate all buffers associated with a bufobj
1832  * Called with the underlying object locked.
1833  */
1834 int
1835 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1836 {
1837 	int error;
1838 
1839 	BO_LOCK(bo);
1840 	if (flags & V_SAVE) {
1841 		error = bufobj_wwait(bo, slpflag, slptimeo);
1842 		if (error) {
1843 			BO_UNLOCK(bo);
1844 			return (error);
1845 		}
1846 		if (bo->bo_dirty.bv_cnt > 0) {
1847 			BO_UNLOCK(bo);
1848 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1849 				return (error);
1850 			/*
1851 			 * XXX We could save a lock/unlock if this was only
1852 			 * enabled under INVARIANTS
1853 			 */
1854 			BO_LOCK(bo);
1855 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1856 				panic("vinvalbuf: dirty bufs");
1857 		}
1858 	}
1859 	/*
1860 	 * If you alter this loop please notice that interlock is dropped and
1861 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1862 	 * no race conditions occur from this.
1863 	 */
1864 	do {
1865 		error = flushbuflist(&bo->bo_clean,
1866 		    flags, bo, slpflag, slptimeo);
1867 		if (error == 0 && !(flags & V_CLEANONLY))
1868 			error = flushbuflist(&bo->bo_dirty,
1869 			    flags, bo, slpflag, slptimeo);
1870 		if (error != 0 && error != EAGAIN) {
1871 			BO_UNLOCK(bo);
1872 			return (error);
1873 		}
1874 	} while (error != 0);
1875 
1876 	/*
1877 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1878 	 * have write I/O in-progress but if there is a VM object then the
1879 	 * VM object can also have read-I/O in-progress.
1880 	 */
1881 	do {
1882 		bufobj_wwait(bo, 0, 0);
1883 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1884 			BO_UNLOCK(bo);
1885 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1886 			BO_LOCK(bo);
1887 		}
1888 	} while (bo->bo_numoutput > 0);
1889 	BO_UNLOCK(bo);
1890 
1891 	/*
1892 	 * Destroy the copy in the VM cache, too.
1893 	 */
1894 	if (bo->bo_object != NULL &&
1895 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1896 		VM_OBJECT_WLOCK(bo->bo_object);
1897 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1898 		    OBJPR_CLEANONLY : 0);
1899 		VM_OBJECT_WUNLOCK(bo->bo_object);
1900 	}
1901 
1902 #ifdef INVARIANTS
1903 	BO_LOCK(bo);
1904 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1905 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1906 	    bo->bo_clean.bv_cnt > 0))
1907 		panic("vinvalbuf: flush failed");
1908 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1909 	    bo->bo_dirty.bv_cnt > 0)
1910 		panic("vinvalbuf: flush dirty failed");
1911 	BO_UNLOCK(bo);
1912 #endif
1913 	return (0);
1914 }
1915 
1916 /*
1917  * Flush out and invalidate all buffers associated with a vnode.
1918  * Called with the underlying object locked.
1919  */
1920 int
1921 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1922 {
1923 
1924 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1925 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1926 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1927 		return (0);
1928 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1929 }
1930 
1931 /*
1932  * Flush out buffers on the specified list.
1933  *
1934  */
1935 static int
1936 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1937     int slptimeo)
1938 {
1939 	struct buf *bp, *nbp;
1940 	int retval, error;
1941 	daddr_t lblkno;
1942 	b_xflags_t xflags;
1943 
1944 	ASSERT_BO_WLOCKED(bo);
1945 
1946 	retval = 0;
1947 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1948 		/*
1949 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1950 		 * do not skip any buffers. If we are flushing only V_NORMAL
1951 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1952 		 * flushing only V_ALT buffers then skip buffers not marked
1953 		 * as BX_ALTDATA.
1954 		 */
1955 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1956 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1957 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1958 			continue;
1959 		}
1960 		if (nbp != NULL) {
1961 			lblkno = nbp->b_lblkno;
1962 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1963 		}
1964 		retval = EAGAIN;
1965 		error = BUF_TIMELOCK(bp,
1966 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1967 		    "flushbuf", slpflag, slptimeo);
1968 		if (error) {
1969 			BO_LOCK(bo);
1970 			return (error != ENOLCK ? error : EAGAIN);
1971 		}
1972 		KASSERT(bp->b_bufobj == bo,
1973 		    ("bp %p wrong b_bufobj %p should be %p",
1974 		    bp, bp->b_bufobj, bo));
1975 		/*
1976 		 * XXX Since there are no node locks for NFS, I
1977 		 * believe there is a slight chance that a delayed
1978 		 * write will occur while sleeping just above, so
1979 		 * check for it.
1980 		 */
1981 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1982 		    (flags & V_SAVE)) {
1983 			bremfree(bp);
1984 			bp->b_flags |= B_ASYNC;
1985 			bwrite(bp);
1986 			BO_LOCK(bo);
1987 			return (EAGAIN);	/* XXX: why not loop ? */
1988 		}
1989 		bremfree(bp);
1990 		bp->b_flags |= (B_INVAL | B_RELBUF);
1991 		bp->b_flags &= ~B_ASYNC;
1992 		brelse(bp);
1993 		BO_LOCK(bo);
1994 		if (nbp == NULL)
1995 			break;
1996 		nbp = gbincore(bo, lblkno);
1997 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1998 		    != xflags)
1999 			break;			/* nbp invalid */
2000 	}
2001 	return (retval);
2002 }
2003 
2004 int
2005 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2006 {
2007 	struct buf *bp;
2008 	int error;
2009 	daddr_t lblkno;
2010 
2011 	ASSERT_BO_LOCKED(bo);
2012 
2013 	for (lblkno = startn;;) {
2014 again:
2015 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2016 		if (bp == NULL || bp->b_lblkno >= endn ||
2017 		    bp->b_lblkno < startn)
2018 			break;
2019 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2020 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2021 		if (error != 0) {
2022 			BO_RLOCK(bo);
2023 			if (error == ENOLCK)
2024 				goto again;
2025 			return (error);
2026 		}
2027 		KASSERT(bp->b_bufobj == bo,
2028 		    ("bp %p wrong b_bufobj %p should be %p",
2029 		    bp, bp->b_bufobj, bo));
2030 		lblkno = bp->b_lblkno + 1;
2031 		if ((bp->b_flags & B_MANAGED) == 0)
2032 			bremfree(bp);
2033 		bp->b_flags |= B_RELBUF;
2034 		/*
2035 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2036 		 * pages backing each buffer in the range are unlikely to be
2037 		 * reused.  Dirty buffers will have the hint applied once
2038 		 * they've been written.
2039 		 */
2040 		if ((bp->b_flags & B_VMIO) != 0)
2041 			bp->b_flags |= B_NOREUSE;
2042 		brelse(bp);
2043 		BO_RLOCK(bo);
2044 	}
2045 	return (0);
2046 }
2047 
2048 /*
2049  * Truncate a file's buffer and pages to a specified length.  This
2050  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2051  * sync activity.
2052  */
2053 int
2054 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2055 {
2056 	struct buf *bp, *nbp;
2057 	struct bufobj *bo;
2058 	daddr_t startlbn;
2059 
2060 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2061 	    vp, blksize, (uintmax_t)length);
2062 
2063 	/*
2064 	 * Round up to the *next* lbn.
2065 	 */
2066 	startlbn = howmany(length, blksize);
2067 
2068 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2069 
2070 	bo = &vp->v_bufobj;
2071 restart_unlocked:
2072 	BO_LOCK(bo);
2073 
2074 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2075 		;
2076 
2077 	if (length > 0) {
2078 restartsync:
2079 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2080 			if (bp->b_lblkno > 0)
2081 				continue;
2082 			/*
2083 			 * Since we hold the vnode lock this should only
2084 			 * fail if we're racing with the buf daemon.
2085 			 */
2086 			if (BUF_LOCK(bp,
2087 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2088 			    BO_LOCKPTR(bo)) == ENOLCK)
2089 				goto restart_unlocked;
2090 
2091 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2092 			    ("buf(%p) on dirty queue without DELWRI", bp));
2093 
2094 			bremfree(bp);
2095 			bawrite(bp);
2096 			BO_LOCK(bo);
2097 			goto restartsync;
2098 		}
2099 	}
2100 
2101 	bufobj_wwait(bo, 0, 0);
2102 	BO_UNLOCK(bo);
2103 	vnode_pager_setsize(vp, length);
2104 
2105 	return (0);
2106 }
2107 
2108 /*
2109  * Invalidate the cached pages of a file's buffer within the range of block
2110  * numbers [startlbn, endlbn).
2111  */
2112 void
2113 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2114     int blksize)
2115 {
2116 	struct bufobj *bo;
2117 	off_t start, end;
2118 
2119 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2120 
2121 	start = blksize * startlbn;
2122 	end = blksize * endlbn;
2123 
2124 	bo = &vp->v_bufobj;
2125 	BO_LOCK(bo);
2126 	MPASS(blksize == bo->bo_bsize);
2127 
2128 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2129 		;
2130 
2131 	BO_UNLOCK(bo);
2132 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2133 }
2134 
2135 static int
2136 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2137     daddr_t startlbn, daddr_t endlbn)
2138 {
2139 	struct buf *bp, *nbp;
2140 	bool anyfreed;
2141 
2142 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2143 	ASSERT_BO_LOCKED(bo);
2144 
2145 	do {
2146 		anyfreed = false;
2147 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2148 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2149 				continue;
2150 			if (BUF_LOCK(bp,
2151 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2152 			    BO_LOCKPTR(bo)) == ENOLCK) {
2153 				BO_LOCK(bo);
2154 				return (EAGAIN);
2155 			}
2156 
2157 			bremfree(bp);
2158 			bp->b_flags |= B_INVAL | B_RELBUF;
2159 			bp->b_flags &= ~B_ASYNC;
2160 			brelse(bp);
2161 			anyfreed = true;
2162 
2163 			BO_LOCK(bo);
2164 			if (nbp != NULL &&
2165 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2166 			    nbp->b_vp != vp ||
2167 			    (nbp->b_flags & B_DELWRI) != 0))
2168 				return (EAGAIN);
2169 		}
2170 
2171 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2172 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2173 				continue;
2174 			if (BUF_LOCK(bp,
2175 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2176 			    BO_LOCKPTR(bo)) == ENOLCK) {
2177 				BO_LOCK(bo);
2178 				return (EAGAIN);
2179 			}
2180 			bremfree(bp);
2181 			bp->b_flags |= B_INVAL | B_RELBUF;
2182 			bp->b_flags &= ~B_ASYNC;
2183 			brelse(bp);
2184 			anyfreed = true;
2185 
2186 			BO_LOCK(bo);
2187 			if (nbp != NULL &&
2188 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2189 			    (nbp->b_vp != vp) ||
2190 			    (nbp->b_flags & B_DELWRI) == 0))
2191 				return (EAGAIN);
2192 		}
2193 	} while (anyfreed);
2194 	return (0);
2195 }
2196 
2197 static void
2198 buf_vlist_remove(struct buf *bp)
2199 {
2200 	struct bufv *bv;
2201 
2202 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2203 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2204 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2205 	    (BX_VNDIRTY|BX_VNCLEAN),
2206 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2207 	if (bp->b_xflags & BX_VNDIRTY)
2208 		bv = &bp->b_bufobj->bo_dirty;
2209 	else
2210 		bv = &bp->b_bufobj->bo_clean;
2211 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2212 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2213 	bv->bv_cnt--;
2214 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2215 }
2216 
2217 /*
2218  * Add the buffer to the sorted clean or dirty block list.
2219  *
2220  * NOTE: xflags is passed as a constant, optimizing this inline function!
2221  */
2222 static void
2223 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2224 {
2225 	struct bufv *bv;
2226 	struct buf *n;
2227 	int error;
2228 
2229 	ASSERT_BO_WLOCKED(bo);
2230 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2231 	    ("dead bo %p", bo));
2232 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2233 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2234 	bp->b_xflags |= xflags;
2235 	if (xflags & BX_VNDIRTY)
2236 		bv = &bo->bo_dirty;
2237 	else
2238 		bv = &bo->bo_clean;
2239 
2240 	/*
2241 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2242 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2243 	 * than _ge.
2244 	 */
2245 	if (bv->bv_cnt == 0 ||
2246 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2247 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2248 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2249 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2250 	else
2251 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2252 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2253 	if (error)
2254 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2255 	bv->bv_cnt++;
2256 }
2257 
2258 /*
2259  * Look up a buffer using the buffer tries.
2260  */
2261 struct buf *
2262 gbincore(struct bufobj *bo, daddr_t lblkno)
2263 {
2264 	struct buf *bp;
2265 
2266 	ASSERT_BO_LOCKED(bo);
2267 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2268 	if (bp != NULL)
2269 		return (bp);
2270 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2271 }
2272 
2273 /*
2274  * Associate a buffer with a vnode.
2275  */
2276 void
2277 bgetvp(struct vnode *vp, struct buf *bp)
2278 {
2279 	struct bufobj *bo;
2280 
2281 	bo = &vp->v_bufobj;
2282 	ASSERT_BO_WLOCKED(bo);
2283 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2284 
2285 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2286 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2287 	    ("bgetvp: bp already attached! %p", bp));
2288 
2289 	vhold(vp);
2290 	bp->b_vp = vp;
2291 	bp->b_bufobj = bo;
2292 	/*
2293 	 * Insert onto list for new vnode.
2294 	 */
2295 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2296 }
2297 
2298 /*
2299  * Disassociate a buffer from a vnode.
2300  */
2301 void
2302 brelvp(struct buf *bp)
2303 {
2304 	struct bufobj *bo;
2305 	struct vnode *vp;
2306 
2307 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2308 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2309 
2310 	/*
2311 	 * Delete from old vnode list, if on one.
2312 	 */
2313 	vp = bp->b_vp;		/* XXX */
2314 	bo = bp->b_bufobj;
2315 	BO_LOCK(bo);
2316 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2317 		buf_vlist_remove(bp);
2318 	else
2319 		panic("brelvp: Buffer %p not on queue.", bp);
2320 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2321 		bo->bo_flag &= ~BO_ONWORKLST;
2322 		mtx_lock(&sync_mtx);
2323 		LIST_REMOVE(bo, bo_synclist);
2324 		syncer_worklist_len--;
2325 		mtx_unlock(&sync_mtx);
2326 	}
2327 	bp->b_vp = NULL;
2328 	bp->b_bufobj = NULL;
2329 	BO_UNLOCK(bo);
2330 	vdrop(vp);
2331 }
2332 
2333 /*
2334  * Add an item to the syncer work queue.
2335  */
2336 static void
2337 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2338 {
2339 	int slot;
2340 
2341 	ASSERT_BO_WLOCKED(bo);
2342 
2343 	mtx_lock(&sync_mtx);
2344 	if (bo->bo_flag & BO_ONWORKLST)
2345 		LIST_REMOVE(bo, bo_synclist);
2346 	else {
2347 		bo->bo_flag |= BO_ONWORKLST;
2348 		syncer_worklist_len++;
2349 	}
2350 
2351 	if (delay > syncer_maxdelay - 2)
2352 		delay = syncer_maxdelay - 2;
2353 	slot = (syncer_delayno + delay) & syncer_mask;
2354 
2355 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2356 	mtx_unlock(&sync_mtx);
2357 }
2358 
2359 static int
2360 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2361 {
2362 	int error, len;
2363 
2364 	mtx_lock(&sync_mtx);
2365 	len = syncer_worklist_len - sync_vnode_count;
2366 	mtx_unlock(&sync_mtx);
2367 	error = SYSCTL_OUT(req, &len, sizeof(len));
2368 	return (error);
2369 }
2370 
2371 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2372     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2373 
2374 static struct proc *updateproc;
2375 static void sched_sync(void);
2376 static struct kproc_desc up_kp = {
2377 	"syncer",
2378 	sched_sync,
2379 	&updateproc
2380 };
2381 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2382 
2383 static int
2384 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2385 {
2386 	struct vnode *vp;
2387 	struct mount *mp;
2388 
2389 	*bo = LIST_FIRST(slp);
2390 	if (*bo == NULL)
2391 		return (0);
2392 	vp = bo2vnode(*bo);
2393 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2394 		return (1);
2395 	/*
2396 	 * We use vhold in case the vnode does not
2397 	 * successfully sync.  vhold prevents the vnode from
2398 	 * going away when we unlock the sync_mtx so that
2399 	 * we can acquire the vnode interlock.
2400 	 */
2401 	vholdl(vp);
2402 	mtx_unlock(&sync_mtx);
2403 	VI_UNLOCK(vp);
2404 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2405 		vdrop(vp);
2406 		mtx_lock(&sync_mtx);
2407 		return (*bo == LIST_FIRST(slp));
2408 	}
2409 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2410 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2411 	VOP_UNLOCK(vp);
2412 	vn_finished_write(mp);
2413 	BO_LOCK(*bo);
2414 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2415 		/*
2416 		 * Put us back on the worklist.  The worklist
2417 		 * routine will remove us from our current
2418 		 * position and then add us back in at a later
2419 		 * position.
2420 		 */
2421 		vn_syncer_add_to_worklist(*bo, syncdelay);
2422 	}
2423 	BO_UNLOCK(*bo);
2424 	vdrop(vp);
2425 	mtx_lock(&sync_mtx);
2426 	return (0);
2427 }
2428 
2429 static int first_printf = 1;
2430 
2431 /*
2432  * System filesystem synchronizer daemon.
2433  */
2434 static void
2435 sched_sync(void)
2436 {
2437 	struct synclist *next, *slp;
2438 	struct bufobj *bo;
2439 	long starttime;
2440 	struct thread *td = curthread;
2441 	int last_work_seen;
2442 	int net_worklist_len;
2443 	int syncer_final_iter;
2444 	int error;
2445 
2446 	last_work_seen = 0;
2447 	syncer_final_iter = 0;
2448 	syncer_state = SYNCER_RUNNING;
2449 	starttime = time_uptime;
2450 	td->td_pflags |= TDP_NORUNNINGBUF;
2451 
2452 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2453 	    SHUTDOWN_PRI_LAST);
2454 
2455 	mtx_lock(&sync_mtx);
2456 	for (;;) {
2457 		if (syncer_state == SYNCER_FINAL_DELAY &&
2458 		    syncer_final_iter == 0) {
2459 			mtx_unlock(&sync_mtx);
2460 			kproc_suspend_check(td->td_proc);
2461 			mtx_lock(&sync_mtx);
2462 		}
2463 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2464 		if (syncer_state != SYNCER_RUNNING &&
2465 		    starttime != time_uptime) {
2466 			if (first_printf) {
2467 				printf("\nSyncing disks, vnodes remaining... ");
2468 				first_printf = 0;
2469 			}
2470 			printf("%d ", net_worklist_len);
2471 		}
2472 		starttime = time_uptime;
2473 
2474 		/*
2475 		 * Push files whose dirty time has expired.  Be careful
2476 		 * of interrupt race on slp queue.
2477 		 *
2478 		 * Skip over empty worklist slots when shutting down.
2479 		 */
2480 		do {
2481 			slp = &syncer_workitem_pending[syncer_delayno];
2482 			syncer_delayno += 1;
2483 			if (syncer_delayno == syncer_maxdelay)
2484 				syncer_delayno = 0;
2485 			next = &syncer_workitem_pending[syncer_delayno];
2486 			/*
2487 			 * If the worklist has wrapped since the
2488 			 * it was emptied of all but syncer vnodes,
2489 			 * switch to the FINAL_DELAY state and run
2490 			 * for one more second.
2491 			 */
2492 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2493 			    net_worklist_len == 0 &&
2494 			    last_work_seen == syncer_delayno) {
2495 				syncer_state = SYNCER_FINAL_DELAY;
2496 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2497 			}
2498 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2499 		    syncer_worklist_len > 0);
2500 
2501 		/*
2502 		 * Keep track of the last time there was anything
2503 		 * on the worklist other than syncer vnodes.
2504 		 * Return to the SHUTTING_DOWN state if any
2505 		 * new work appears.
2506 		 */
2507 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2508 			last_work_seen = syncer_delayno;
2509 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2510 			syncer_state = SYNCER_SHUTTING_DOWN;
2511 		while (!LIST_EMPTY(slp)) {
2512 			error = sync_vnode(slp, &bo, td);
2513 			if (error == 1) {
2514 				LIST_REMOVE(bo, bo_synclist);
2515 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2516 				continue;
2517 			}
2518 
2519 			if (first_printf == 0) {
2520 				/*
2521 				 * Drop the sync mutex, because some watchdog
2522 				 * drivers need to sleep while patting
2523 				 */
2524 				mtx_unlock(&sync_mtx);
2525 				wdog_kern_pat(WD_LASTVAL);
2526 				mtx_lock(&sync_mtx);
2527 			}
2528 
2529 		}
2530 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2531 			syncer_final_iter--;
2532 		/*
2533 		 * The variable rushjob allows the kernel to speed up the
2534 		 * processing of the filesystem syncer process. A rushjob
2535 		 * value of N tells the filesystem syncer to process the next
2536 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2537 		 * is used by the soft update code to speed up the filesystem
2538 		 * syncer process when the incore state is getting so far
2539 		 * ahead of the disk that the kernel memory pool is being
2540 		 * threatened with exhaustion.
2541 		 */
2542 		if (rushjob > 0) {
2543 			rushjob -= 1;
2544 			continue;
2545 		}
2546 		/*
2547 		 * Just sleep for a short period of time between
2548 		 * iterations when shutting down to allow some I/O
2549 		 * to happen.
2550 		 *
2551 		 * If it has taken us less than a second to process the
2552 		 * current work, then wait. Otherwise start right over
2553 		 * again. We can still lose time if any single round
2554 		 * takes more than two seconds, but it does not really
2555 		 * matter as we are just trying to generally pace the
2556 		 * filesystem activity.
2557 		 */
2558 		if (syncer_state != SYNCER_RUNNING ||
2559 		    time_uptime == starttime) {
2560 			thread_lock(td);
2561 			sched_prio(td, PPAUSE);
2562 			thread_unlock(td);
2563 		}
2564 		if (syncer_state != SYNCER_RUNNING)
2565 			cv_timedwait(&sync_wakeup, &sync_mtx,
2566 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2567 		else if (time_uptime == starttime)
2568 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2569 	}
2570 }
2571 
2572 /*
2573  * Request the syncer daemon to speed up its work.
2574  * We never push it to speed up more than half of its
2575  * normal turn time, otherwise it could take over the cpu.
2576  */
2577 int
2578 speedup_syncer(void)
2579 {
2580 	int ret = 0;
2581 
2582 	mtx_lock(&sync_mtx);
2583 	if (rushjob < syncdelay / 2) {
2584 		rushjob += 1;
2585 		stat_rush_requests += 1;
2586 		ret = 1;
2587 	}
2588 	mtx_unlock(&sync_mtx);
2589 	cv_broadcast(&sync_wakeup);
2590 	return (ret);
2591 }
2592 
2593 /*
2594  * Tell the syncer to speed up its work and run though its work
2595  * list several times, then tell it to shut down.
2596  */
2597 static void
2598 syncer_shutdown(void *arg, int howto)
2599 {
2600 
2601 	if (howto & RB_NOSYNC)
2602 		return;
2603 	mtx_lock(&sync_mtx);
2604 	syncer_state = SYNCER_SHUTTING_DOWN;
2605 	rushjob = 0;
2606 	mtx_unlock(&sync_mtx);
2607 	cv_broadcast(&sync_wakeup);
2608 	kproc_shutdown(arg, howto);
2609 }
2610 
2611 void
2612 syncer_suspend(void)
2613 {
2614 
2615 	syncer_shutdown(updateproc, 0);
2616 }
2617 
2618 void
2619 syncer_resume(void)
2620 {
2621 
2622 	mtx_lock(&sync_mtx);
2623 	first_printf = 1;
2624 	syncer_state = SYNCER_RUNNING;
2625 	mtx_unlock(&sync_mtx);
2626 	cv_broadcast(&sync_wakeup);
2627 	kproc_resume(updateproc);
2628 }
2629 
2630 /*
2631  * Reassign a buffer from one vnode to another.
2632  * Used to assign file specific control information
2633  * (indirect blocks) to the vnode to which they belong.
2634  */
2635 void
2636 reassignbuf(struct buf *bp)
2637 {
2638 	struct vnode *vp;
2639 	struct bufobj *bo;
2640 	int delay;
2641 #ifdef INVARIANTS
2642 	struct bufv *bv;
2643 #endif
2644 
2645 	vp = bp->b_vp;
2646 	bo = bp->b_bufobj;
2647 	++reassignbufcalls;
2648 
2649 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2650 	    bp, bp->b_vp, bp->b_flags);
2651 	/*
2652 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2653 	 * is not fully linked in.
2654 	 */
2655 	if (bp->b_flags & B_PAGING)
2656 		panic("cannot reassign paging buffer");
2657 
2658 	/*
2659 	 * Delete from old vnode list, if on one.
2660 	 */
2661 	BO_LOCK(bo);
2662 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2663 		buf_vlist_remove(bp);
2664 	else
2665 		panic("reassignbuf: Buffer %p not on queue.", bp);
2666 	/*
2667 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2668 	 * of clean buffers.
2669 	 */
2670 	if (bp->b_flags & B_DELWRI) {
2671 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2672 			switch (vp->v_type) {
2673 			case VDIR:
2674 				delay = dirdelay;
2675 				break;
2676 			case VCHR:
2677 				delay = metadelay;
2678 				break;
2679 			default:
2680 				delay = filedelay;
2681 			}
2682 			vn_syncer_add_to_worklist(bo, delay);
2683 		}
2684 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2685 	} else {
2686 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2687 
2688 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2689 			mtx_lock(&sync_mtx);
2690 			LIST_REMOVE(bo, bo_synclist);
2691 			syncer_worklist_len--;
2692 			mtx_unlock(&sync_mtx);
2693 			bo->bo_flag &= ~BO_ONWORKLST;
2694 		}
2695 	}
2696 #ifdef INVARIANTS
2697 	bv = &bo->bo_clean;
2698 	bp = TAILQ_FIRST(&bv->bv_hd);
2699 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2700 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2701 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2702 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2703 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2704 	bv = &bo->bo_dirty;
2705 	bp = TAILQ_FIRST(&bv->bv_hd);
2706 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2707 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2708 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2709 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2710 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2711 #endif
2712 	BO_UNLOCK(bo);
2713 }
2714 
2715 static void
2716 v_init_counters(struct vnode *vp)
2717 {
2718 
2719 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2720 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2721 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2722 
2723 	refcount_init(&vp->v_holdcnt, 1);
2724 	refcount_init(&vp->v_usecount, 1);
2725 }
2726 
2727 /*
2728  * Increment si_usecount of the associated device, if any.
2729  */
2730 static void
2731 v_incr_devcount(struct vnode *vp)
2732 {
2733 
2734 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2735 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2736 		dev_lock();
2737 		vp->v_rdev->si_usecount++;
2738 		dev_unlock();
2739 	}
2740 }
2741 
2742 /*
2743  * Decrement si_usecount of the associated device, if any.
2744  */
2745 static void
2746 v_decr_devcount(struct vnode *vp)
2747 {
2748 
2749 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2750 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2751 		dev_lock();
2752 		vp->v_rdev->si_usecount--;
2753 		dev_unlock();
2754 	}
2755 }
2756 
2757 /*
2758  * Grab a particular vnode from the free list, increment its
2759  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2760  * is being destroyed.  Only callers who specify LK_RETRY will
2761  * see doomed vnodes.  If inactive processing was delayed in
2762  * vput try to do it here.
2763  *
2764  * Both holdcnt and usecount can be manipulated using atomics without holding
2765  * any locks except in these cases which require the vnode interlock:
2766  * holdcnt: 1->0 and 0->1
2767  * usecount: 0->1
2768  *
2769  * usecount is permitted to transition 1->0 without the interlock because
2770  * vnode is kept live by holdcnt.
2771  */
2772 static enum vgetstate __always_inline
2773 _vget_prep(struct vnode *vp, bool interlock)
2774 {
2775 	enum vgetstate vs;
2776 
2777 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2778 		vs = VGET_USECOUNT;
2779 	} else {
2780 		if (interlock)
2781 			vholdl(vp);
2782 		else
2783 			vhold(vp);
2784 		vs = VGET_HOLDCNT;
2785 	}
2786 	return (vs);
2787 }
2788 
2789 enum vgetstate
2790 vget_prep(struct vnode *vp)
2791 {
2792 
2793 	return (_vget_prep(vp, false));
2794 }
2795 
2796 int
2797 vget(struct vnode *vp, int flags, struct thread *td)
2798 {
2799 	enum vgetstate vs;
2800 
2801 	MPASS(td == curthread);
2802 
2803 	vs = _vget_prep(vp, (flags & LK_INTERLOCK) != 0);
2804 	return (vget_finish(vp, flags, vs));
2805 }
2806 
2807 int
2808 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2809 {
2810 	int error, oweinact;
2811 
2812 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2813 	    ("%s: invalid lock operation", __func__));
2814 
2815 	if ((flags & LK_INTERLOCK) != 0)
2816 		ASSERT_VI_LOCKED(vp, __func__);
2817 	else
2818 		ASSERT_VI_UNLOCKED(vp, __func__);
2819 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2820 	if (vs == VGET_USECOUNT) {
2821 		VNASSERT(vp->v_usecount > 0, vp,
2822 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2823 		    __func__));
2824 	}
2825 
2826 	if ((error = vn_lock(vp, flags)) != 0) {
2827 		if (vs == VGET_USECOUNT)
2828 			vrele(vp);
2829 		else
2830 			vdrop(vp);
2831 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2832 		    vp);
2833 		return (error);
2834 	}
2835 
2836 	if (vs == VGET_USECOUNT) {
2837 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2838 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2839 		return (0);
2840 	}
2841 
2842 	/*
2843 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2844 	 * the vnode around. Otherwise someone else lended their hold count and
2845 	 * we have to drop ours.
2846 	 */
2847 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2848 #ifdef INVARIANTS
2849 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2850 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2851 #else
2852 		refcount_release(&vp->v_holdcnt);
2853 #endif
2854 		VNODE_REFCOUNT_FENCE_ACQ();
2855 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2856 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2857 		return (0);
2858 	}
2859 
2860 	/*
2861 	 * We don't guarantee that any particular close will
2862 	 * trigger inactive processing so just make a best effort
2863 	 * here at preventing a reference to a removed file.  If
2864 	 * we don't succeed no harm is done.
2865 	 *
2866 	 * Upgrade our holdcnt to a usecount.
2867 	 */
2868 	VI_LOCK(vp);
2869 	/*
2870 	 * See the previous section. By the time we get here we may find
2871 	 * ourselves in the same spot.
2872 	 */
2873 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2874 #ifdef INVARIANTS
2875 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2876 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2877 #else
2878 		refcount_release(&vp->v_holdcnt);
2879 #endif
2880 		VNODE_REFCOUNT_FENCE_ACQ();
2881 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2882 		    ("%s: vnode with usecount and VI_OWEINACT set",
2883 		    __func__));
2884 		VI_UNLOCK(vp);
2885 		return (0);
2886 	}
2887 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
2888 		oweinact = 0;
2889 	} else {
2890 		oweinact = 1;
2891 		vp->v_iflag &= ~VI_OWEINACT;
2892 		VNODE_REFCOUNT_FENCE_REL();
2893 	}
2894 	v_incr_devcount(vp);
2895 	refcount_acquire(&vp->v_usecount);
2896 	if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2897 	    (flags & LK_NOWAIT) == 0)
2898 		vinactive(vp);
2899 	VI_UNLOCK(vp);
2900 	return (0);
2901 }
2902 
2903 /*
2904  * Increase the reference (use) and hold count of a vnode.
2905  * This will also remove the vnode from the free list if it is presently free.
2906  */
2907 void
2908 vref(struct vnode *vp)
2909 {
2910 
2911 	ASSERT_VI_UNLOCKED(vp, __func__);
2912 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2913 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2914 		VNODE_REFCOUNT_FENCE_ACQ();
2915 		VNASSERT(vp->v_holdcnt > 0, vp,
2916 		    ("%s: active vnode not held", __func__));
2917 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2918 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2919 		return;
2920 	}
2921 	VI_LOCK(vp);
2922 	vrefl(vp);
2923 	VI_UNLOCK(vp);
2924 }
2925 
2926 void
2927 vrefl(struct vnode *vp)
2928 {
2929 
2930 	ASSERT_VI_LOCKED(vp, __func__);
2931 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2932 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2933 		VNODE_REFCOUNT_FENCE_ACQ();
2934 		VNASSERT(vp->v_holdcnt > 0, vp,
2935 		    ("%s: active vnode not held", __func__));
2936 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2937 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2938 		return;
2939 	}
2940 	vholdl(vp);
2941 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2942 		vp->v_iflag &= ~VI_OWEINACT;
2943 		VNODE_REFCOUNT_FENCE_REL();
2944 	}
2945 	v_incr_devcount(vp);
2946 	refcount_acquire(&vp->v_usecount);
2947 }
2948 
2949 void
2950 vrefact(struct vnode *vp)
2951 {
2952 
2953 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2954 #ifdef INVARIANTS
2955 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2956 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
2957 #else
2958 	refcount_acquire(&vp->v_usecount);
2959 #endif
2960 }
2961 
2962 /*
2963  * Return reference count of a vnode.
2964  *
2965  * The results of this call are only guaranteed when some mechanism is used to
2966  * stop other processes from gaining references to the vnode.  This may be the
2967  * case if the caller holds the only reference.  This is also useful when stale
2968  * data is acceptable as race conditions may be accounted for by some other
2969  * means.
2970  */
2971 int
2972 vrefcnt(struct vnode *vp)
2973 {
2974 
2975 	return (vp->v_usecount);
2976 }
2977 
2978 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
2979 
2980 /*
2981  * Decrement the use and hold counts for a vnode.
2982  *
2983  * See an explanation near vget() as to why atomic operation is safe.
2984  */
2985 static void
2986 vputx(struct vnode *vp, enum vputx_op func)
2987 {
2988 	int error;
2989 
2990 	KASSERT(vp != NULL, ("vputx: null vp"));
2991 	if (func == VPUTX_VUNREF)
2992 		ASSERT_VOP_LOCKED(vp, "vunref");
2993 	ASSERT_VI_UNLOCKED(vp, __func__);
2994 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2995 	    ("%s: wrong ref counts", __func__));
2996 
2997 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2998 
2999 	/*
3000 	 * We want to hold the vnode until the inactive finishes to
3001 	 * prevent vgone() races.  We drop the use count here and the
3002 	 * hold count below when we're done.
3003 	 *
3004 	 * If we release the last usecount we take ownership of the hold
3005 	 * count which provides liveness of the vnode, in which case we
3006 	 * have to vdrop.
3007 	 */
3008 	if (!refcount_release(&vp->v_usecount))
3009 		return;
3010 	VI_LOCK(vp);
3011 	v_decr_devcount(vp);
3012 	/*
3013 	 * By the time we got here someone else might have transitioned
3014 	 * the count back to > 0.
3015 	 */
3016 	if (vp->v_usecount > 0) {
3017 		vdropl(vp);
3018 		return;
3019 	}
3020 	if (vp->v_iflag & VI_DOINGINACT) {
3021 		vdropl(vp);
3022 		return;
3023 	}
3024 
3025 	/*
3026 	 * Check if the fs wants to perform inactive processing. Note we
3027 	 * may be only holding the interlock, in which case it is possible
3028 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3029 	 * Since vgone performs inactive on its own there is nothing to do
3030 	 * here but to drop our hold count.
3031 	 */
3032 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3033 	    VOP_NEED_INACTIVE(vp) == 0) {
3034 		vdropl(vp);
3035 		return;
3036 	}
3037 
3038 	/*
3039 	 * We must call VOP_INACTIVE with the node locked. Mark
3040 	 * as VI_DOINGINACT to avoid recursion.
3041 	 */
3042 	vp->v_iflag |= VI_OWEINACT;
3043 	switch (func) {
3044 	case VPUTX_VRELE:
3045 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3046 		VI_LOCK(vp);
3047 		break;
3048 	case VPUTX_VPUT:
3049 		error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT);
3050 		VI_LOCK(vp);
3051 		break;
3052 	case VPUTX_VUNREF:
3053 		error = 0;
3054 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3055 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3056 			VI_LOCK(vp);
3057 		}
3058 		break;
3059 	}
3060 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
3061 	    ("vnode with usecount and VI_OWEINACT set"));
3062 	if (error == 0) {
3063 		if (vp->v_iflag & VI_OWEINACT)
3064 			vinactive(vp);
3065 		if (func != VPUTX_VUNREF)
3066 			VOP_UNLOCK(vp);
3067 	}
3068 	vdropl(vp);
3069 }
3070 
3071 /*
3072  * Vnode put/release.
3073  * If count drops to zero, call inactive routine and return to freelist.
3074  */
3075 void
3076 vrele(struct vnode *vp)
3077 {
3078 
3079 	vputx(vp, VPUTX_VRELE);
3080 }
3081 
3082 /*
3083  * Release an already locked vnode.  This give the same effects as
3084  * unlock+vrele(), but takes less time and avoids releasing and
3085  * re-aquiring the lock (as vrele() acquires the lock internally.)
3086  *
3087  * It is an invariant that all VOP_* calls operate on a held vnode.
3088  * We may be only having an implicit hold stemming from our usecount,
3089  * which we are about to release. If we unlock the vnode afterwards we
3090  * open a time window where someone else dropped the last usecount and
3091  * proceeded to free the vnode before our unlock finished. For this
3092  * reason we unlock the vnode early. This is a little bit wasteful as
3093  * it may be the vnode is exclusively locked and inactive processing is
3094  * needed, in which case we are adding work.
3095  */
3096 void
3097 vput(struct vnode *vp)
3098 {
3099 
3100 	VOP_UNLOCK(vp);
3101 	vputx(vp, VPUTX_VPUT);
3102 }
3103 
3104 /*
3105  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3106  */
3107 void
3108 vunref(struct vnode *vp)
3109 {
3110 
3111 	vputx(vp, VPUTX_VUNREF);
3112 }
3113 
3114 /*
3115  * Increase the hold count and activate if this is the first reference.
3116  */
3117 static void
3118 vhold_activate(struct vnode *vp)
3119 {
3120 	struct mount *mp;
3121 
3122 	ASSERT_VI_LOCKED(vp, __func__);
3123 	VNASSERT(vp->v_holdcnt == 0, vp,
3124 	    ("%s: wrong hold count", __func__));
3125 	VNASSERT(vp->v_op != NULL, vp,
3126 	    ("%s: vnode already reclaimed.", __func__));
3127 	/*
3128 	 * Remove a vnode from the free list, mark it as in use,
3129 	 * and put it on the active list.
3130 	 */
3131 	VNASSERT(vp->v_mount != NULL, vp,
3132 	    ("_vhold: vnode not on per mount vnode list"));
3133 	mp = vp->v_mount;
3134 	mtx_lock(&mp->mnt_listmtx);
3135 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
3136 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3137 		mp->mnt_tmpfreevnodelistsize--;
3138 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
3139 	} else {
3140 		mtx_lock(&vnode_free_list_mtx);
3141 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3142 		freevnodes--;
3143 		mtx_unlock(&vnode_free_list_mtx);
3144 	}
3145 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3146 	    ("Activating already active vnode"));
3147 	vp->v_iflag &= ~VI_FREE;
3148 	vp->v_iflag |= VI_ACTIVE;
3149 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3150 	mp->mnt_activevnodelistsize++;
3151 	mtx_unlock(&mp->mnt_listmtx);
3152 	refcount_acquire(&vp->v_holdcnt);
3153 }
3154 
3155 void
3156 vhold(struct vnode *vp)
3157 {
3158 
3159 	ASSERT_VI_UNLOCKED(vp, __func__);
3160 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3161 	if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
3162 		VNODE_REFCOUNT_FENCE_ACQ();
3163 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3164 		    ("vhold: vnode with holdcnt is free"));
3165 		return;
3166 	}
3167 	VI_LOCK(vp);
3168 	vholdl(vp);
3169 	VI_UNLOCK(vp);
3170 }
3171 
3172 void
3173 vholdl(struct vnode *vp)
3174 {
3175 
3176 	ASSERT_VI_LOCKED(vp, __func__);
3177 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3178 	if ((vp->v_iflag & VI_FREE) == 0) {
3179 		refcount_acquire(&vp->v_holdcnt);
3180 		return;
3181 	}
3182 	vhold_activate(vp);
3183 }
3184 
3185 void
3186 vholdnz(struct vnode *vp)
3187 {
3188 
3189 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3190 #ifdef INVARIANTS
3191 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3192 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3193 #else
3194 	atomic_add_int(&vp->v_holdcnt, 1);
3195 #endif
3196 }
3197 
3198 /*
3199  * Drop the hold count of the vnode.  If this is the last reference to
3200  * the vnode we place it on the free list unless it has been vgone'd
3201  * (marked VIRF_DOOMED) in which case we will free it.
3202  *
3203  * Because the vnode vm object keeps a hold reference on the vnode if
3204  * there is at least one resident non-cached page, the vnode cannot
3205  * leave the active list without the page cleanup done.
3206  */
3207 static void
3208 vdrop_deactivate(struct vnode *vp)
3209 {
3210 	struct mount *mp;
3211 
3212 	ASSERT_VI_LOCKED(vp, __func__);
3213 	/*
3214 	 * Mark a vnode as free: remove it from its active list
3215 	 * and put it up for recycling on the freelist.
3216 	 */
3217 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3218 	    ("vdrop: returning doomed vnode"));
3219 	VNASSERT(vp->v_op != NULL, vp,
3220 	    ("vdrop: vnode already reclaimed."));
3221 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3222 	    ("vnode already free"));
3223 	VNASSERT(vp->v_holdcnt == 0, vp,
3224 	    ("vdrop: freeing when we shouldn't"));
3225 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3226 		mp = vp->v_mount;
3227 		mtx_lock(&mp->mnt_listmtx);
3228 		if (vp->v_iflag & VI_ACTIVE) {
3229 			vp->v_iflag &= ~VI_ACTIVE;
3230 			TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
3231 			mp->mnt_activevnodelistsize--;
3232 		}
3233 		TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3234 		mp->mnt_tmpfreevnodelistsize++;
3235 		vp->v_iflag |= VI_FREE;
3236 		vp->v_mflag |= VMP_TMPMNTFREELIST;
3237 		VI_UNLOCK(vp);
3238 		if (mp->mnt_tmpfreevnodelistsize >= mnt_free_list_batch)
3239 			vnlru_return_batch_locked(mp);
3240 		mtx_unlock(&mp->mnt_listmtx);
3241 	} else {
3242 		VI_UNLOCK(vp);
3243 		counter_u64_add(free_owe_inact, 1);
3244 	}
3245 }
3246 
3247 void
3248 vdrop(struct vnode *vp)
3249 {
3250 
3251 	ASSERT_VI_UNLOCKED(vp, __func__);
3252 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3253 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3254 		return;
3255 	VI_LOCK(vp);
3256 	vdropl(vp);
3257 }
3258 
3259 void
3260 vdropl(struct vnode *vp)
3261 {
3262 
3263 	ASSERT_VI_LOCKED(vp, __func__);
3264 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3265 	if (!refcount_release(&vp->v_holdcnt)) {
3266 		VI_UNLOCK(vp);
3267 		return;
3268 	}
3269 	if (VN_IS_DOOMED(vp)) {
3270 		freevnode(vp);
3271 		return;
3272 	}
3273 	vdrop_deactivate(vp);
3274 }
3275 
3276 /*
3277  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3278  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3279  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3280  * failed lock upgrade.
3281  */
3282 void
3283 vinactive(struct vnode *vp)
3284 {
3285 	struct vm_object *obj;
3286 
3287 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3288 	ASSERT_VI_LOCKED(vp, "vinactive");
3289 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3290 	    ("vinactive: recursed on VI_DOINGINACT"));
3291 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3292 	vp->v_iflag |= VI_DOINGINACT;
3293 	vp->v_iflag &= ~VI_OWEINACT;
3294 	VI_UNLOCK(vp);
3295 	/*
3296 	 * Before moving off the active list, we must be sure that any
3297 	 * modified pages are converted into the vnode's dirty
3298 	 * buffers, since these will no longer be checked once the
3299 	 * vnode is on the inactive list.
3300 	 *
3301 	 * The write-out of the dirty pages is asynchronous.  At the
3302 	 * point that VOP_INACTIVE() is called, there could still be
3303 	 * pending I/O and dirty pages in the object.
3304 	 */
3305 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3306 	    vm_object_mightbedirty(obj)) {
3307 		VM_OBJECT_WLOCK(obj);
3308 		vm_object_page_clean(obj, 0, 0, 0);
3309 		VM_OBJECT_WUNLOCK(obj);
3310 	}
3311 	VOP_INACTIVE(vp, curthread);
3312 	VI_LOCK(vp);
3313 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3314 	    ("vinactive: lost VI_DOINGINACT"));
3315 	vp->v_iflag &= ~VI_DOINGINACT;
3316 }
3317 
3318 /*
3319  * Remove any vnodes in the vnode table belonging to mount point mp.
3320  *
3321  * If FORCECLOSE is not specified, there should not be any active ones,
3322  * return error if any are found (nb: this is a user error, not a
3323  * system error). If FORCECLOSE is specified, detach any active vnodes
3324  * that are found.
3325  *
3326  * If WRITECLOSE is set, only flush out regular file vnodes open for
3327  * writing.
3328  *
3329  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3330  *
3331  * `rootrefs' specifies the base reference count for the root vnode
3332  * of this filesystem. The root vnode is considered busy if its
3333  * v_usecount exceeds this value. On a successful return, vflush(, td)
3334  * will call vrele() on the root vnode exactly rootrefs times.
3335  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3336  * be zero.
3337  */
3338 #ifdef DIAGNOSTIC
3339 static int busyprt = 0;		/* print out busy vnodes */
3340 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3341 #endif
3342 
3343 int
3344 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3345 {
3346 	struct vnode *vp, *mvp, *rootvp = NULL;
3347 	struct vattr vattr;
3348 	int busy = 0, error;
3349 
3350 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3351 	    rootrefs, flags);
3352 	if (rootrefs > 0) {
3353 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3354 		    ("vflush: bad args"));
3355 		/*
3356 		 * Get the filesystem root vnode. We can vput() it
3357 		 * immediately, since with rootrefs > 0, it won't go away.
3358 		 */
3359 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3360 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3361 			    __func__, error);
3362 			return (error);
3363 		}
3364 		vput(rootvp);
3365 	}
3366 loop:
3367 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3368 		vholdl(vp);
3369 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3370 		if (error) {
3371 			vdrop(vp);
3372 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3373 			goto loop;
3374 		}
3375 		/*
3376 		 * Skip over a vnodes marked VV_SYSTEM.
3377 		 */
3378 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3379 			VOP_UNLOCK(vp);
3380 			vdrop(vp);
3381 			continue;
3382 		}
3383 		/*
3384 		 * If WRITECLOSE is set, flush out unlinked but still open
3385 		 * files (even if open only for reading) and regular file
3386 		 * vnodes open for writing.
3387 		 */
3388 		if (flags & WRITECLOSE) {
3389 			if (vp->v_object != NULL) {
3390 				VM_OBJECT_WLOCK(vp->v_object);
3391 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3392 				VM_OBJECT_WUNLOCK(vp->v_object);
3393 			}
3394 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3395 			if (error != 0) {
3396 				VOP_UNLOCK(vp);
3397 				vdrop(vp);
3398 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3399 				return (error);
3400 			}
3401 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3402 			VI_LOCK(vp);
3403 
3404 			if ((vp->v_type == VNON ||
3405 			    (error == 0 && vattr.va_nlink > 0)) &&
3406 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3407 				VOP_UNLOCK(vp);
3408 				vdropl(vp);
3409 				continue;
3410 			}
3411 		} else
3412 			VI_LOCK(vp);
3413 		/*
3414 		 * With v_usecount == 0, all we need to do is clear out the
3415 		 * vnode data structures and we are done.
3416 		 *
3417 		 * If FORCECLOSE is set, forcibly close the vnode.
3418 		 */
3419 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3420 			vgonel(vp);
3421 		} else {
3422 			busy++;
3423 #ifdef DIAGNOSTIC
3424 			if (busyprt)
3425 				vn_printf(vp, "vflush: busy vnode ");
3426 #endif
3427 		}
3428 		VOP_UNLOCK(vp);
3429 		vdropl(vp);
3430 	}
3431 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3432 		/*
3433 		 * If just the root vnode is busy, and if its refcount
3434 		 * is equal to `rootrefs', then go ahead and kill it.
3435 		 */
3436 		VI_LOCK(rootvp);
3437 		KASSERT(busy > 0, ("vflush: not busy"));
3438 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3439 		    ("vflush: usecount %d < rootrefs %d",
3440 		     rootvp->v_usecount, rootrefs));
3441 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3442 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3443 			vgone(rootvp);
3444 			VOP_UNLOCK(rootvp);
3445 			busy = 0;
3446 		} else
3447 			VI_UNLOCK(rootvp);
3448 	}
3449 	if (busy) {
3450 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3451 		    busy);
3452 		return (EBUSY);
3453 	}
3454 	for (; rootrefs > 0; rootrefs--)
3455 		vrele(rootvp);
3456 	return (0);
3457 }
3458 
3459 /*
3460  * Recycle an unused vnode to the front of the free list.
3461  */
3462 int
3463 vrecycle(struct vnode *vp)
3464 {
3465 	int recycled;
3466 
3467 	VI_LOCK(vp);
3468 	recycled = vrecyclel(vp);
3469 	VI_UNLOCK(vp);
3470 	return (recycled);
3471 }
3472 
3473 /*
3474  * vrecycle, with the vp interlock held.
3475  */
3476 int
3477 vrecyclel(struct vnode *vp)
3478 {
3479 	int recycled;
3480 
3481 	ASSERT_VOP_ELOCKED(vp, __func__);
3482 	ASSERT_VI_LOCKED(vp, __func__);
3483 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3484 	recycled = 0;
3485 	if (vp->v_usecount == 0) {
3486 		recycled = 1;
3487 		vgonel(vp);
3488 	}
3489 	return (recycled);
3490 }
3491 
3492 /*
3493  * Eliminate all activity associated with a vnode
3494  * in preparation for reuse.
3495  */
3496 void
3497 vgone(struct vnode *vp)
3498 {
3499 	VI_LOCK(vp);
3500 	vgonel(vp);
3501 	VI_UNLOCK(vp);
3502 }
3503 
3504 static void
3505 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3506     struct vnode *lowervp __unused)
3507 {
3508 }
3509 
3510 /*
3511  * Notify upper mounts about reclaimed or unlinked vnode.
3512  */
3513 void
3514 vfs_notify_upper(struct vnode *vp, int event)
3515 {
3516 	static struct vfsops vgonel_vfsops = {
3517 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3518 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3519 	};
3520 	struct mount *mp, *ump, *mmp;
3521 
3522 	mp = vp->v_mount;
3523 	if (mp == NULL)
3524 		return;
3525 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3526 		return;
3527 
3528 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3529 	mmp->mnt_op = &vgonel_vfsops;
3530 	mmp->mnt_kern_flag |= MNTK_MARKER;
3531 	MNT_ILOCK(mp);
3532 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3533 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3534 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3535 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3536 			continue;
3537 		}
3538 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3539 		MNT_IUNLOCK(mp);
3540 		switch (event) {
3541 		case VFS_NOTIFY_UPPER_RECLAIM:
3542 			VFS_RECLAIM_LOWERVP(ump, vp);
3543 			break;
3544 		case VFS_NOTIFY_UPPER_UNLINK:
3545 			VFS_UNLINK_LOWERVP(ump, vp);
3546 			break;
3547 		default:
3548 			KASSERT(0, ("invalid event %d", event));
3549 			break;
3550 		}
3551 		MNT_ILOCK(mp);
3552 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3553 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3554 	}
3555 	free(mmp, M_TEMP);
3556 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3557 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3558 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3559 		wakeup(&mp->mnt_uppers);
3560 	}
3561 	MNT_IUNLOCK(mp);
3562 }
3563 
3564 /*
3565  * vgone, with the vp interlock held.
3566  */
3567 static void
3568 vgonel(struct vnode *vp)
3569 {
3570 	struct thread *td;
3571 	struct mount *mp;
3572 	vm_object_t object;
3573 	bool active, oweinact;
3574 
3575 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3576 	ASSERT_VI_LOCKED(vp, "vgonel");
3577 	VNASSERT(vp->v_holdcnt, vp,
3578 	    ("vgonel: vp %p has no reference.", vp));
3579 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3580 	td = curthread;
3581 
3582 	/*
3583 	 * Don't vgonel if we're already doomed.
3584 	 */
3585 	if (vp->v_irflag & VIRF_DOOMED)
3586 		return;
3587 	vp->v_irflag |= VIRF_DOOMED;
3588 
3589 	/*
3590 	 * Check to see if the vnode is in use.  If so, we have to call
3591 	 * VOP_CLOSE() and VOP_INACTIVE().
3592 	 */
3593 	active = vp->v_usecount > 0;
3594 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3595 	VI_UNLOCK(vp);
3596 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3597 
3598 	/*
3599 	 * If purging an active vnode, it must be closed and
3600 	 * deactivated before being reclaimed.
3601 	 */
3602 	if (active)
3603 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3604 	if (oweinact || active) {
3605 		VI_LOCK(vp);
3606 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3607 			vinactive(vp);
3608 		VI_UNLOCK(vp);
3609 	}
3610 	if (vp->v_type == VSOCK)
3611 		vfs_unp_reclaim(vp);
3612 
3613 	/*
3614 	 * Clean out any buffers associated with the vnode.
3615 	 * If the flush fails, just toss the buffers.
3616 	 */
3617 	mp = NULL;
3618 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3619 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3620 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3621 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3622 			;
3623 	}
3624 
3625 	BO_LOCK(&vp->v_bufobj);
3626 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3627 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3628 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3629 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3630 	    ("vp %p bufobj not invalidated", vp));
3631 
3632 	/*
3633 	 * For VMIO bufobj, BO_DEAD is set later, or in
3634 	 * vm_object_terminate() after the object's page queue is
3635 	 * flushed.
3636 	 */
3637 	object = vp->v_bufobj.bo_object;
3638 	if (object == NULL)
3639 		vp->v_bufobj.bo_flag |= BO_DEAD;
3640 	BO_UNLOCK(&vp->v_bufobj);
3641 
3642 	/*
3643 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3644 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3645 	 * should not touch the object borrowed from the lower vnode
3646 	 * (the handle check).
3647 	 */
3648 	if (object != NULL && object->type == OBJT_VNODE &&
3649 	    object->handle == vp)
3650 		vnode_destroy_vobject(vp);
3651 
3652 	/*
3653 	 * Reclaim the vnode.
3654 	 */
3655 	if (VOP_RECLAIM(vp, td))
3656 		panic("vgone: cannot reclaim");
3657 	if (mp != NULL)
3658 		vn_finished_secondary_write(mp);
3659 	VNASSERT(vp->v_object == NULL, vp,
3660 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3661 	/*
3662 	 * Clear the advisory locks and wake up waiting threads.
3663 	 */
3664 	(void)VOP_ADVLOCKPURGE(vp);
3665 	vp->v_lockf = NULL;
3666 	/*
3667 	 * Delete from old mount point vnode list.
3668 	 */
3669 	delmntque(vp);
3670 	cache_purge(vp);
3671 	/*
3672 	 * Done with purge, reset to the standard lock and invalidate
3673 	 * the vnode.
3674 	 */
3675 	VI_LOCK(vp);
3676 	vp->v_vnlock = &vp->v_lock;
3677 	vp->v_op = &dead_vnodeops;
3678 	vp->v_tag = "none";
3679 	vp->v_type = VBAD;
3680 }
3681 
3682 /*
3683  * Calculate the total number of references to a special device.
3684  */
3685 int
3686 vcount(struct vnode *vp)
3687 {
3688 	int count;
3689 
3690 	dev_lock();
3691 	count = vp->v_rdev->si_usecount;
3692 	dev_unlock();
3693 	return (count);
3694 }
3695 
3696 /*
3697  * Print out a description of a vnode.
3698  */
3699 static char *typename[] =
3700 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3701  "VMARKER"};
3702 
3703 void
3704 vn_printf(struct vnode *vp, const char *fmt, ...)
3705 {
3706 	va_list ap;
3707 	char buf[256], buf2[16];
3708 	u_long flags;
3709 
3710 	va_start(ap, fmt);
3711 	vprintf(fmt, ap);
3712 	va_end(ap);
3713 	printf("%p: ", (void *)vp);
3714 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3715 	printf("    usecount %d, writecount %d, refcount %d",
3716 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3717 	switch (vp->v_type) {
3718 	case VDIR:
3719 		printf(" mountedhere %p\n", vp->v_mountedhere);
3720 		break;
3721 	case VCHR:
3722 		printf(" rdev %p\n", vp->v_rdev);
3723 		break;
3724 	case VSOCK:
3725 		printf(" socket %p\n", vp->v_unpcb);
3726 		break;
3727 	case VFIFO:
3728 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3729 		break;
3730 	default:
3731 		printf("\n");
3732 		break;
3733 	}
3734 	buf[0] = '\0';
3735 	buf[1] = '\0';
3736 	if (vp->v_irflag & VIRF_DOOMED)
3737 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3738 	flags = vp->v_irflag & ~(VIRF_DOOMED);
3739 	if (flags != 0) {
3740 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3741 		strlcat(buf, buf2, sizeof(buf));
3742 	}
3743 	if (vp->v_vflag & VV_ROOT)
3744 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3745 	if (vp->v_vflag & VV_ISTTY)
3746 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3747 	if (vp->v_vflag & VV_NOSYNC)
3748 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3749 	if (vp->v_vflag & VV_ETERNALDEV)
3750 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3751 	if (vp->v_vflag & VV_CACHEDLABEL)
3752 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3753 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3754 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3755 	if (vp->v_vflag & VV_COPYONWRITE)
3756 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3757 	if (vp->v_vflag & VV_SYSTEM)
3758 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3759 	if (vp->v_vflag & VV_PROCDEP)
3760 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3761 	if (vp->v_vflag & VV_NOKNOTE)
3762 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3763 	if (vp->v_vflag & VV_DELETED)
3764 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3765 	if (vp->v_vflag & VV_MD)
3766 		strlcat(buf, "|VV_MD", sizeof(buf));
3767 	if (vp->v_vflag & VV_FORCEINSMQ)
3768 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3769 	if (vp->v_vflag & VV_READLINK)
3770 		strlcat(buf, "|VV_READLINK", sizeof(buf));
3771 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3772 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3773 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3774 	if (flags != 0) {
3775 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3776 		strlcat(buf, buf2, sizeof(buf));
3777 	}
3778 	if (vp->v_iflag & VI_TEXT_REF)
3779 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
3780 	if (vp->v_iflag & VI_MOUNT)
3781 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3782 	if (vp->v_iflag & VI_FREE)
3783 		strlcat(buf, "|VI_FREE", sizeof(buf));
3784 	if (vp->v_iflag & VI_ACTIVE)
3785 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3786 	if (vp->v_iflag & VI_DOINGINACT)
3787 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3788 	if (vp->v_iflag & VI_OWEINACT)
3789 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3790 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_FREE | VI_ACTIVE |
3791 	    VI_DOINGINACT | VI_OWEINACT);
3792 	if (flags != 0) {
3793 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3794 		strlcat(buf, buf2, sizeof(buf));
3795 	}
3796 	if (vp->v_mflag & VMP_TMPMNTFREELIST)
3797 		strlcat(buf, "|VMP_TMPMNTFREELIST", sizeof(buf));
3798 	flags = vp->v_mflag & ~(VMP_TMPMNTFREELIST);
3799 	if (flags != 0) {
3800 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
3801 		strlcat(buf, buf2, sizeof(buf));
3802 	}
3803 	printf("    flags (%s)\n", buf + 1);
3804 	if (mtx_owned(VI_MTX(vp)))
3805 		printf(" VI_LOCKed");
3806 	if (vp->v_object != NULL)
3807 		printf("    v_object %p ref %d pages %d "
3808 		    "cleanbuf %d dirtybuf %d\n",
3809 		    vp->v_object, vp->v_object->ref_count,
3810 		    vp->v_object->resident_page_count,
3811 		    vp->v_bufobj.bo_clean.bv_cnt,
3812 		    vp->v_bufobj.bo_dirty.bv_cnt);
3813 	printf("    ");
3814 	lockmgr_printinfo(vp->v_vnlock);
3815 	if (vp->v_data != NULL)
3816 		VOP_PRINT(vp);
3817 }
3818 
3819 #ifdef DDB
3820 /*
3821  * List all of the locked vnodes in the system.
3822  * Called when debugging the kernel.
3823  */
3824 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3825 {
3826 	struct mount *mp;
3827 	struct vnode *vp;
3828 
3829 	/*
3830 	 * Note: because this is DDB, we can't obey the locking semantics
3831 	 * for these structures, which means we could catch an inconsistent
3832 	 * state and dereference a nasty pointer.  Not much to be done
3833 	 * about that.
3834 	 */
3835 	db_printf("Locked vnodes\n");
3836 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3837 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3838 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3839 				vn_printf(vp, "vnode ");
3840 		}
3841 	}
3842 }
3843 
3844 /*
3845  * Show details about the given vnode.
3846  */
3847 DB_SHOW_COMMAND(vnode, db_show_vnode)
3848 {
3849 	struct vnode *vp;
3850 
3851 	if (!have_addr)
3852 		return;
3853 	vp = (struct vnode *)addr;
3854 	vn_printf(vp, "vnode ");
3855 }
3856 
3857 /*
3858  * Show details about the given mount point.
3859  */
3860 DB_SHOW_COMMAND(mount, db_show_mount)
3861 {
3862 	struct mount *mp;
3863 	struct vfsopt *opt;
3864 	struct statfs *sp;
3865 	struct vnode *vp;
3866 	char buf[512];
3867 	uint64_t mflags;
3868 	u_int flags;
3869 
3870 	if (!have_addr) {
3871 		/* No address given, print short info about all mount points. */
3872 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3873 			db_printf("%p %s on %s (%s)\n", mp,
3874 			    mp->mnt_stat.f_mntfromname,
3875 			    mp->mnt_stat.f_mntonname,
3876 			    mp->mnt_stat.f_fstypename);
3877 			if (db_pager_quit)
3878 				break;
3879 		}
3880 		db_printf("\nMore info: show mount <addr>\n");
3881 		return;
3882 	}
3883 
3884 	mp = (struct mount *)addr;
3885 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3886 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3887 
3888 	buf[0] = '\0';
3889 	mflags = mp->mnt_flag;
3890 #define	MNT_FLAG(flag)	do {						\
3891 	if (mflags & (flag)) {						\
3892 		if (buf[0] != '\0')					\
3893 			strlcat(buf, ", ", sizeof(buf));		\
3894 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3895 		mflags &= ~(flag);					\
3896 	}								\
3897 } while (0)
3898 	MNT_FLAG(MNT_RDONLY);
3899 	MNT_FLAG(MNT_SYNCHRONOUS);
3900 	MNT_FLAG(MNT_NOEXEC);
3901 	MNT_FLAG(MNT_NOSUID);
3902 	MNT_FLAG(MNT_NFS4ACLS);
3903 	MNT_FLAG(MNT_UNION);
3904 	MNT_FLAG(MNT_ASYNC);
3905 	MNT_FLAG(MNT_SUIDDIR);
3906 	MNT_FLAG(MNT_SOFTDEP);
3907 	MNT_FLAG(MNT_NOSYMFOLLOW);
3908 	MNT_FLAG(MNT_GJOURNAL);
3909 	MNT_FLAG(MNT_MULTILABEL);
3910 	MNT_FLAG(MNT_ACLS);
3911 	MNT_FLAG(MNT_NOATIME);
3912 	MNT_FLAG(MNT_NOCLUSTERR);
3913 	MNT_FLAG(MNT_NOCLUSTERW);
3914 	MNT_FLAG(MNT_SUJ);
3915 	MNT_FLAG(MNT_EXRDONLY);
3916 	MNT_FLAG(MNT_EXPORTED);
3917 	MNT_FLAG(MNT_DEFEXPORTED);
3918 	MNT_FLAG(MNT_EXPORTANON);
3919 	MNT_FLAG(MNT_EXKERB);
3920 	MNT_FLAG(MNT_EXPUBLIC);
3921 	MNT_FLAG(MNT_LOCAL);
3922 	MNT_FLAG(MNT_QUOTA);
3923 	MNT_FLAG(MNT_ROOTFS);
3924 	MNT_FLAG(MNT_USER);
3925 	MNT_FLAG(MNT_IGNORE);
3926 	MNT_FLAG(MNT_UPDATE);
3927 	MNT_FLAG(MNT_DELEXPORT);
3928 	MNT_FLAG(MNT_RELOAD);
3929 	MNT_FLAG(MNT_FORCE);
3930 	MNT_FLAG(MNT_SNAPSHOT);
3931 	MNT_FLAG(MNT_BYFSID);
3932 #undef MNT_FLAG
3933 	if (mflags != 0) {
3934 		if (buf[0] != '\0')
3935 			strlcat(buf, ", ", sizeof(buf));
3936 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3937 		    "0x%016jx", mflags);
3938 	}
3939 	db_printf("    mnt_flag = %s\n", buf);
3940 
3941 	buf[0] = '\0';
3942 	flags = mp->mnt_kern_flag;
3943 #define	MNT_KERN_FLAG(flag)	do {					\
3944 	if (flags & (flag)) {						\
3945 		if (buf[0] != '\0')					\
3946 			strlcat(buf, ", ", sizeof(buf));		\
3947 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3948 		flags &= ~(flag);					\
3949 	}								\
3950 } while (0)
3951 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3952 	MNT_KERN_FLAG(MNTK_ASYNC);
3953 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3954 	MNT_KERN_FLAG(MNTK_DRAINING);
3955 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3956 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3957 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3958 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3959 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3960 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3961 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3962 	MNT_KERN_FLAG(MNTK_MARKER);
3963 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3964 	MNT_KERN_FLAG(MNTK_NOASYNC);
3965 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3966 	MNT_KERN_FLAG(MNTK_MWAIT);
3967 	MNT_KERN_FLAG(MNTK_SUSPEND);
3968 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3969 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3970 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3971 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3972 #undef MNT_KERN_FLAG
3973 	if (flags != 0) {
3974 		if (buf[0] != '\0')
3975 			strlcat(buf, ", ", sizeof(buf));
3976 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3977 		    "0x%08x", flags);
3978 	}
3979 	db_printf("    mnt_kern_flag = %s\n", buf);
3980 
3981 	db_printf("    mnt_opt = ");
3982 	opt = TAILQ_FIRST(mp->mnt_opt);
3983 	if (opt != NULL) {
3984 		db_printf("%s", opt->name);
3985 		opt = TAILQ_NEXT(opt, link);
3986 		while (opt != NULL) {
3987 			db_printf(", %s", opt->name);
3988 			opt = TAILQ_NEXT(opt, link);
3989 		}
3990 	}
3991 	db_printf("\n");
3992 
3993 	sp = &mp->mnt_stat;
3994 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3995 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3996 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3997 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3998 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3999 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4000 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4001 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4002 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4003 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4004 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4005 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4006 
4007 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4008 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4009 	if (jailed(mp->mnt_cred))
4010 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4011 	db_printf(" }\n");
4012 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4013 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4014 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4015 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4016 	db_printf("    mnt_activevnodelistsize = %d\n",
4017 	    mp->mnt_activevnodelistsize);
4018 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4019 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4020 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4021 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4022 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4023 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4024 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4025 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4026 	db_printf("    mnt_secondary_accwrites = %d\n",
4027 	    mp->mnt_secondary_accwrites);
4028 	db_printf("    mnt_gjprovider = %s\n",
4029 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4030 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4031 
4032 	db_printf("\n\nList of active vnodes\n");
4033 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
4034 		if (vp->v_type != VMARKER) {
4035 			vn_printf(vp, "vnode ");
4036 			if (db_pager_quit)
4037 				break;
4038 		}
4039 	}
4040 	db_printf("\n\nList of inactive vnodes\n");
4041 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4042 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
4043 			vn_printf(vp, "vnode ");
4044 			if (db_pager_quit)
4045 				break;
4046 		}
4047 	}
4048 }
4049 #endif	/* DDB */
4050 
4051 /*
4052  * Fill in a struct xvfsconf based on a struct vfsconf.
4053  */
4054 static int
4055 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4056 {
4057 	struct xvfsconf xvfsp;
4058 
4059 	bzero(&xvfsp, sizeof(xvfsp));
4060 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4061 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4062 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4063 	xvfsp.vfc_flags = vfsp->vfc_flags;
4064 	/*
4065 	 * These are unused in userland, we keep them
4066 	 * to not break binary compatibility.
4067 	 */
4068 	xvfsp.vfc_vfsops = NULL;
4069 	xvfsp.vfc_next = NULL;
4070 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4071 }
4072 
4073 #ifdef COMPAT_FREEBSD32
4074 struct xvfsconf32 {
4075 	uint32_t	vfc_vfsops;
4076 	char		vfc_name[MFSNAMELEN];
4077 	int32_t		vfc_typenum;
4078 	int32_t		vfc_refcount;
4079 	int32_t		vfc_flags;
4080 	uint32_t	vfc_next;
4081 };
4082 
4083 static int
4084 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4085 {
4086 	struct xvfsconf32 xvfsp;
4087 
4088 	bzero(&xvfsp, sizeof(xvfsp));
4089 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4090 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4091 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4092 	xvfsp.vfc_flags = vfsp->vfc_flags;
4093 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4094 }
4095 #endif
4096 
4097 /*
4098  * Top level filesystem related information gathering.
4099  */
4100 static int
4101 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4102 {
4103 	struct vfsconf *vfsp;
4104 	int error;
4105 
4106 	error = 0;
4107 	vfsconf_slock();
4108 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4109 #ifdef COMPAT_FREEBSD32
4110 		if (req->flags & SCTL_MASK32)
4111 			error = vfsconf2x32(req, vfsp);
4112 		else
4113 #endif
4114 			error = vfsconf2x(req, vfsp);
4115 		if (error)
4116 			break;
4117 	}
4118 	vfsconf_sunlock();
4119 	return (error);
4120 }
4121 
4122 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4123     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4124     "S,xvfsconf", "List of all configured filesystems");
4125 
4126 #ifndef BURN_BRIDGES
4127 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4128 
4129 static int
4130 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4131 {
4132 	int *name = (int *)arg1 - 1;	/* XXX */
4133 	u_int namelen = arg2 + 1;	/* XXX */
4134 	struct vfsconf *vfsp;
4135 
4136 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4137 	    "please rebuild world\n");
4138 
4139 #if 1 || defined(COMPAT_PRELITE2)
4140 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4141 	if (namelen == 1)
4142 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4143 #endif
4144 
4145 	switch (name[1]) {
4146 	case VFS_MAXTYPENUM:
4147 		if (namelen != 2)
4148 			return (ENOTDIR);
4149 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4150 	case VFS_CONF:
4151 		if (namelen != 3)
4152 			return (ENOTDIR);	/* overloaded */
4153 		vfsconf_slock();
4154 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4155 			if (vfsp->vfc_typenum == name[2])
4156 				break;
4157 		}
4158 		vfsconf_sunlock();
4159 		if (vfsp == NULL)
4160 			return (EOPNOTSUPP);
4161 #ifdef COMPAT_FREEBSD32
4162 		if (req->flags & SCTL_MASK32)
4163 			return (vfsconf2x32(req, vfsp));
4164 		else
4165 #endif
4166 			return (vfsconf2x(req, vfsp));
4167 	}
4168 	return (EOPNOTSUPP);
4169 }
4170 
4171 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4172     CTLFLAG_MPSAFE, vfs_sysctl,
4173     "Generic filesystem");
4174 
4175 #if 1 || defined(COMPAT_PRELITE2)
4176 
4177 static int
4178 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4179 {
4180 	int error;
4181 	struct vfsconf *vfsp;
4182 	struct ovfsconf ovfs;
4183 
4184 	vfsconf_slock();
4185 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4186 		bzero(&ovfs, sizeof(ovfs));
4187 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4188 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4189 		ovfs.vfc_index = vfsp->vfc_typenum;
4190 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4191 		ovfs.vfc_flags = vfsp->vfc_flags;
4192 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4193 		if (error != 0) {
4194 			vfsconf_sunlock();
4195 			return (error);
4196 		}
4197 	}
4198 	vfsconf_sunlock();
4199 	return (0);
4200 }
4201 
4202 #endif /* 1 || COMPAT_PRELITE2 */
4203 #endif /* !BURN_BRIDGES */
4204 
4205 #define KINFO_VNODESLOP		10
4206 #ifdef notyet
4207 /*
4208  * Dump vnode list (via sysctl).
4209  */
4210 /* ARGSUSED */
4211 static int
4212 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4213 {
4214 	struct xvnode *xvn;
4215 	struct mount *mp;
4216 	struct vnode *vp;
4217 	int error, len, n;
4218 
4219 	/*
4220 	 * Stale numvnodes access is not fatal here.
4221 	 */
4222 	req->lock = 0;
4223 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4224 	if (!req->oldptr)
4225 		/* Make an estimate */
4226 		return (SYSCTL_OUT(req, 0, len));
4227 
4228 	error = sysctl_wire_old_buffer(req, 0);
4229 	if (error != 0)
4230 		return (error);
4231 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4232 	n = 0;
4233 	mtx_lock(&mountlist_mtx);
4234 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4235 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4236 			continue;
4237 		MNT_ILOCK(mp);
4238 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4239 			if (n == len)
4240 				break;
4241 			vref(vp);
4242 			xvn[n].xv_size = sizeof *xvn;
4243 			xvn[n].xv_vnode = vp;
4244 			xvn[n].xv_id = 0;	/* XXX compat */
4245 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4246 			XV_COPY(usecount);
4247 			XV_COPY(writecount);
4248 			XV_COPY(holdcnt);
4249 			XV_COPY(mount);
4250 			XV_COPY(numoutput);
4251 			XV_COPY(type);
4252 #undef XV_COPY
4253 			xvn[n].xv_flag = vp->v_vflag;
4254 
4255 			switch (vp->v_type) {
4256 			case VREG:
4257 			case VDIR:
4258 			case VLNK:
4259 				break;
4260 			case VBLK:
4261 			case VCHR:
4262 				if (vp->v_rdev == NULL) {
4263 					vrele(vp);
4264 					continue;
4265 				}
4266 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4267 				break;
4268 			case VSOCK:
4269 				xvn[n].xv_socket = vp->v_socket;
4270 				break;
4271 			case VFIFO:
4272 				xvn[n].xv_fifo = vp->v_fifoinfo;
4273 				break;
4274 			case VNON:
4275 			case VBAD:
4276 			default:
4277 				/* shouldn't happen? */
4278 				vrele(vp);
4279 				continue;
4280 			}
4281 			vrele(vp);
4282 			++n;
4283 		}
4284 		MNT_IUNLOCK(mp);
4285 		mtx_lock(&mountlist_mtx);
4286 		vfs_unbusy(mp);
4287 		if (n == len)
4288 			break;
4289 	}
4290 	mtx_unlock(&mountlist_mtx);
4291 
4292 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4293 	free(xvn, M_TEMP);
4294 	return (error);
4295 }
4296 
4297 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4298     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4299     "");
4300 #endif
4301 
4302 static void
4303 unmount_or_warn(struct mount *mp)
4304 {
4305 	int error;
4306 
4307 	error = dounmount(mp, MNT_FORCE, curthread);
4308 	if (error != 0) {
4309 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4310 		if (error == EBUSY)
4311 			printf("BUSY)\n");
4312 		else
4313 			printf("%d)\n", error);
4314 	}
4315 }
4316 
4317 /*
4318  * Unmount all filesystems. The list is traversed in reverse order
4319  * of mounting to avoid dependencies.
4320  */
4321 void
4322 vfs_unmountall(void)
4323 {
4324 	struct mount *mp, *tmp;
4325 
4326 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4327 
4328 	/*
4329 	 * Since this only runs when rebooting, it is not interlocked.
4330 	 */
4331 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4332 		vfs_ref(mp);
4333 
4334 		/*
4335 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4336 		 * unmount of the latter.
4337 		 */
4338 		if (mp == rootdevmp)
4339 			continue;
4340 
4341 		unmount_or_warn(mp);
4342 	}
4343 
4344 	if (rootdevmp != NULL)
4345 		unmount_or_warn(rootdevmp);
4346 }
4347 
4348 /*
4349  * perform msync on all vnodes under a mount point
4350  * the mount point must be locked.
4351  */
4352 void
4353 vfs_msync(struct mount *mp, int flags)
4354 {
4355 	struct vnode *vp, *mvp;
4356 	struct vm_object *obj;
4357 
4358 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4359 
4360 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4361 		return;
4362 
4363 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4364 		obj = vp->v_object;
4365 		if (obj != NULL && vm_object_mightbedirty(obj) &&
4366 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4367 			if (!vget(vp,
4368 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4369 			    curthread)) {
4370 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4371 					vput(vp);
4372 					continue;
4373 				}
4374 
4375 				obj = vp->v_object;
4376 				if (obj != NULL) {
4377 					VM_OBJECT_WLOCK(obj);
4378 					vm_object_page_clean(obj, 0, 0,
4379 					    flags == MNT_WAIT ?
4380 					    OBJPC_SYNC : OBJPC_NOSYNC);
4381 					VM_OBJECT_WUNLOCK(obj);
4382 				}
4383 				vput(vp);
4384 			}
4385 		} else
4386 			VI_UNLOCK(vp);
4387 	}
4388 }
4389 
4390 static void
4391 destroy_vpollinfo_free(struct vpollinfo *vi)
4392 {
4393 
4394 	knlist_destroy(&vi->vpi_selinfo.si_note);
4395 	mtx_destroy(&vi->vpi_lock);
4396 	uma_zfree(vnodepoll_zone, vi);
4397 }
4398 
4399 static void
4400 destroy_vpollinfo(struct vpollinfo *vi)
4401 {
4402 
4403 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4404 	seldrain(&vi->vpi_selinfo);
4405 	destroy_vpollinfo_free(vi);
4406 }
4407 
4408 /*
4409  * Initialize per-vnode helper structure to hold poll-related state.
4410  */
4411 void
4412 v_addpollinfo(struct vnode *vp)
4413 {
4414 	struct vpollinfo *vi;
4415 
4416 	if (vp->v_pollinfo != NULL)
4417 		return;
4418 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4419 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4420 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4421 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4422 	VI_LOCK(vp);
4423 	if (vp->v_pollinfo != NULL) {
4424 		VI_UNLOCK(vp);
4425 		destroy_vpollinfo_free(vi);
4426 		return;
4427 	}
4428 	vp->v_pollinfo = vi;
4429 	VI_UNLOCK(vp);
4430 }
4431 
4432 /*
4433  * Record a process's interest in events which might happen to
4434  * a vnode.  Because poll uses the historic select-style interface
4435  * internally, this routine serves as both the ``check for any
4436  * pending events'' and the ``record my interest in future events''
4437  * functions.  (These are done together, while the lock is held,
4438  * to avoid race conditions.)
4439  */
4440 int
4441 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4442 {
4443 
4444 	v_addpollinfo(vp);
4445 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4446 	if (vp->v_pollinfo->vpi_revents & events) {
4447 		/*
4448 		 * This leaves events we are not interested
4449 		 * in available for the other process which
4450 		 * which presumably had requested them
4451 		 * (otherwise they would never have been
4452 		 * recorded).
4453 		 */
4454 		events &= vp->v_pollinfo->vpi_revents;
4455 		vp->v_pollinfo->vpi_revents &= ~events;
4456 
4457 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4458 		return (events);
4459 	}
4460 	vp->v_pollinfo->vpi_events |= events;
4461 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4462 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4463 	return (0);
4464 }
4465 
4466 /*
4467  * Routine to create and manage a filesystem syncer vnode.
4468  */
4469 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4470 static int	sync_fsync(struct  vop_fsync_args *);
4471 static int	sync_inactive(struct  vop_inactive_args *);
4472 static int	sync_reclaim(struct  vop_reclaim_args *);
4473 
4474 static struct vop_vector sync_vnodeops = {
4475 	.vop_bypass =	VOP_EOPNOTSUPP,
4476 	.vop_close =	sync_close,		/* close */
4477 	.vop_fsync =	sync_fsync,		/* fsync */
4478 	.vop_inactive =	sync_inactive,	/* inactive */
4479 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4480 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4481 	.vop_lock1 =	vop_stdlock,	/* lock */
4482 	.vop_unlock =	vop_stdunlock,	/* unlock */
4483 	.vop_islocked =	vop_stdislocked,	/* islocked */
4484 };
4485 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4486 
4487 /*
4488  * Create a new filesystem syncer vnode for the specified mount point.
4489  */
4490 void
4491 vfs_allocate_syncvnode(struct mount *mp)
4492 {
4493 	struct vnode *vp;
4494 	struct bufobj *bo;
4495 	static long start, incr, next;
4496 	int error;
4497 
4498 	/* Allocate a new vnode */
4499 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4500 	if (error != 0)
4501 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4502 	vp->v_type = VNON;
4503 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4504 	vp->v_vflag |= VV_FORCEINSMQ;
4505 	error = insmntque(vp, mp);
4506 	if (error != 0)
4507 		panic("vfs_allocate_syncvnode: insmntque() failed");
4508 	vp->v_vflag &= ~VV_FORCEINSMQ;
4509 	VOP_UNLOCK(vp);
4510 	/*
4511 	 * Place the vnode onto the syncer worklist. We attempt to
4512 	 * scatter them about on the list so that they will go off
4513 	 * at evenly distributed times even if all the filesystems
4514 	 * are mounted at once.
4515 	 */
4516 	next += incr;
4517 	if (next == 0 || next > syncer_maxdelay) {
4518 		start /= 2;
4519 		incr /= 2;
4520 		if (start == 0) {
4521 			start = syncer_maxdelay / 2;
4522 			incr = syncer_maxdelay;
4523 		}
4524 		next = start;
4525 	}
4526 	bo = &vp->v_bufobj;
4527 	BO_LOCK(bo);
4528 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4529 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4530 	mtx_lock(&sync_mtx);
4531 	sync_vnode_count++;
4532 	if (mp->mnt_syncer == NULL) {
4533 		mp->mnt_syncer = vp;
4534 		vp = NULL;
4535 	}
4536 	mtx_unlock(&sync_mtx);
4537 	BO_UNLOCK(bo);
4538 	if (vp != NULL) {
4539 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4540 		vgone(vp);
4541 		vput(vp);
4542 	}
4543 }
4544 
4545 void
4546 vfs_deallocate_syncvnode(struct mount *mp)
4547 {
4548 	struct vnode *vp;
4549 
4550 	mtx_lock(&sync_mtx);
4551 	vp = mp->mnt_syncer;
4552 	if (vp != NULL)
4553 		mp->mnt_syncer = NULL;
4554 	mtx_unlock(&sync_mtx);
4555 	if (vp != NULL)
4556 		vrele(vp);
4557 }
4558 
4559 /*
4560  * Do a lazy sync of the filesystem.
4561  */
4562 static int
4563 sync_fsync(struct vop_fsync_args *ap)
4564 {
4565 	struct vnode *syncvp = ap->a_vp;
4566 	struct mount *mp = syncvp->v_mount;
4567 	int error, save;
4568 	struct bufobj *bo;
4569 
4570 	/*
4571 	 * We only need to do something if this is a lazy evaluation.
4572 	 */
4573 	if (ap->a_waitfor != MNT_LAZY)
4574 		return (0);
4575 
4576 	/*
4577 	 * Move ourselves to the back of the sync list.
4578 	 */
4579 	bo = &syncvp->v_bufobj;
4580 	BO_LOCK(bo);
4581 	vn_syncer_add_to_worklist(bo, syncdelay);
4582 	BO_UNLOCK(bo);
4583 
4584 	/*
4585 	 * Walk the list of vnodes pushing all that are dirty and
4586 	 * not already on the sync list.
4587 	 */
4588 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4589 		return (0);
4590 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4591 		vfs_unbusy(mp);
4592 		return (0);
4593 	}
4594 	save = curthread_pflags_set(TDP_SYNCIO);
4595 	/*
4596 	 * The filesystem at hand may be idle with free vnodes stored in the
4597 	 * batch.  Return them instead of letting them stay there indefinitely.
4598 	 */
4599 	vnlru_return_batch(mp);
4600 	vfs_msync(mp, MNT_NOWAIT);
4601 	error = VFS_SYNC(mp, MNT_LAZY);
4602 	curthread_pflags_restore(save);
4603 	vn_finished_write(mp);
4604 	vfs_unbusy(mp);
4605 	return (error);
4606 }
4607 
4608 /*
4609  * The syncer vnode is no referenced.
4610  */
4611 static int
4612 sync_inactive(struct vop_inactive_args *ap)
4613 {
4614 
4615 	vgone(ap->a_vp);
4616 	return (0);
4617 }
4618 
4619 /*
4620  * The syncer vnode is no longer needed and is being decommissioned.
4621  *
4622  * Modifications to the worklist must be protected by sync_mtx.
4623  */
4624 static int
4625 sync_reclaim(struct vop_reclaim_args *ap)
4626 {
4627 	struct vnode *vp = ap->a_vp;
4628 	struct bufobj *bo;
4629 
4630 	bo = &vp->v_bufobj;
4631 	BO_LOCK(bo);
4632 	mtx_lock(&sync_mtx);
4633 	if (vp->v_mount->mnt_syncer == vp)
4634 		vp->v_mount->mnt_syncer = NULL;
4635 	if (bo->bo_flag & BO_ONWORKLST) {
4636 		LIST_REMOVE(bo, bo_synclist);
4637 		syncer_worklist_len--;
4638 		sync_vnode_count--;
4639 		bo->bo_flag &= ~BO_ONWORKLST;
4640 	}
4641 	mtx_unlock(&sync_mtx);
4642 	BO_UNLOCK(bo);
4643 
4644 	return (0);
4645 }
4646 
4647 int
4648 vn_need_pageq_flush(struct vnode *vp)
4649 {
4650 	struct vm_object *obj;
4651 	int need;
4652 
4653 	MPASS(mtx_owned(VI_MTX(vp)));
4654 	need = 0;
4655 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4656 	    vm_object_mightbedirty(obj))
4657 		need = 1;
4658 	return (need);
4659 }
4660 
4661 /*
4662  * Check if vnode represents a disk device
4663  */
4664 int
4665 vn_isdisk(struct vnode *vp, int *errp)
4666 {
4667 	int error;
4668 
4669 	if (vp->v_type != VCHR) {
4670 		error = ENOTBLK;
4671 		goto out;
4672 	}
4673 	error = 0;
4674 	dev_lock();
4675 	if (vp->v_rdev == NULL)
4676 		error = ENXIO;
4677 	else if (vp->v_rdev->si_devsw == NULL)
4678 		error = ENXIO;
4679 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4680 		error = ENOTBLK;
4681 	dev_unlock();
4682 out:
4683 	if (errp != NULL)
4684 		*errp = error;
4685 	return (error == 0);
4686 }
4687 
4688 /*
4689  * Common filesystem object access control check routine.  Accepts a
4690  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4691  * and optional call-by-reference privused argument allowing vaccess()
4692  * to indicate to the caller whether privilege was used to satisfy the
4693  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4694  */
4695 int
4696 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4697     accmode_t accmode, struct ucred *cred, int *privused)
4698 {
4699 	accmode_t dac_granted;
4700 	accmode_t priv_granted;
4701 
4702 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4703 	    ("invalid bit in accmode"));
4704 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4705 	    ("VAPPEND without VWRITE"));
4706 
4707 	/*
4708 	 * Look for a normal, non-privileged way to access the file/directory
4709 	 * as requested.  If it exists, go with that.
4710 	 */
4711 
4712 	if (privused != NULL)
4713 		*privused = 0;
4714 
4715 	dac_granted = 0;
4716 
4717 	/* Check the owner. */
4718 	if (cred->cr_uid == file_uid) {
4719 		dac_granted |= VADMIN;
4720 		if (file_mode & S_IXUSR)
4721 			dac_granted |= VEXEC;
4722 		if (file_mode & S_IRUSR)
4723 			dac_granted |= VREAD;
4724 		if (file_mode & S_IWUSR)
4725 			dac_granted |= (VWRITE | VAPPEND);
4726 
4727 		if ((accmode & dac_granted) == accmode)
4728 			return (0);
4729 
4730 		goto privcheck;
4731 	}
4732 
4733 	/* Otherwise, check the groups (first match) */
4734 	if (groupmember(file_gid, cred)) {
4735 		if (file_mode & S_IXGRP)
4736 			dac_granted |= VEXEC;
4737 		if (file_mode & S_IRGRP)
4738 			dac_granted |= VREAD;
4739 		if (file_mode & S_IWGRP)
4740 			dac_granted |= (VWRITE | VAPPEND);
4741 
4742 		if ((accmode & dac_granted) == accmode)
4743 			return (0);
4744 
4745 		goto privcheck;
4746 	}
4747 
4748 	/* Otherwise, check everyone else. */
4749 	if (file_mode & S_IXOTH)
4750 		dac_granted |= VEXEC;
4751 	if (file_mode & S_IROTH)
4752 		dac_granted |= VREAD;
4753 	if (file_mode & S_IWOTH)
4754 		dac_granted |= (VWRITE | VAPPEND);
4755 	if ((accmode & dac_granted) == accmode)
4756 		return (0);
4757 
4758 privcheck:
4759 	/*
4760 	 * Build a privilege mask to determine if the set of privileges
4761 	 * satisfies the requirements when combined with the granted mask
4762 	 * from above.  For each privilege, if the privilege is required,
4763 	 * bitwise or the request type onto the priv_granted mask.
4764 	 */
4765 	priv_granted = 0;
4766 
4767 	if (type == VDIR) {
4768 		/*
4769 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4770 		 * requests, instead of PRIV_VFS_EXEC.
4771 		 */
4772 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4773 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4774 			priv_granted |= VEXEC;
4775 	} else {
4776 		/*
4777 		 * Ensure that at least one execute bit is on. Otherwise,
4778 		 * a privileged user will always succeed, and we don't want
4779 		 * this to happen unless the file really is executable.
4780 		 */
4781 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4782 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4783 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4784 			priv_granted |= VEXEC;
4785 	}
4786 
4787 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4788 	    !priv_check_cred(cred, PRIV_VFS_READ))
4789 		priv_granted |= VREAD;
4790 
4791 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4792 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4793 		priv_granted |= (VWRITE | VAPPEND);
4794 
4795 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4796 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4797 		priv_granted |= VADMIN;
4798 
4799 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4800 		/* XXX audit: privilege used */
4801 		if (privused != NULL)
4802 			*privused = 1;
4803 		return (0);
4804 	}
4805 
4806 	return ((accmode & VADMIN) ? EPERM : EACCES);
4807 }
4808 
4809 /*
4810  * Credential check based on process requesting service, and per-attribute
4811  * permissions.
4812  */
4813 int
4814 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4815     struct thread *td, accmode_t accmode)
4816 {
4817 
4818 	/*
4819 	 * Kernel-invoked always succeeds.
4820 	 */
4821 	if (cred == NOCRED)
4822 		return (0);
4823 
4824 	/*
4825 	 * Do not allow privileged processes in jail to directly manipulate
4826 	 * system attributes.
4827 	 */
4828 	switch (attrnamespace) {
4829 	case EXTATTR_NAMESPACE_SYSTEM:
4830 		/* Potentially should be: return (EPERM); */
4831 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4832 	case EXTATTR_NAMESPACE_USER:
4833 		return (VOP_ACCESS(vp, accmode, cred, td));
4834 	default:
4835 		return (EPERM);
4836 	}
4837 }
4838 
4839 #ifdef DEBUG_VFS_LOCKS
4840 /*
4841  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4842  * no longer ok to have an unlocked VFS.
4843  */
4844 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4845 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4846 
4847 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4848 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4849     "Drop into debugger on lock violation");
4850 
4851 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4852 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4853     0, "Check for interlock across VOPs");
4854 
4855 int vfs_badlock_print = 1;	/* Print lock violations. */
4856 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4857     0, "Print lock violations");
4858 
4859 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4860 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4861     0, "Print vnode details on lock violations");
4862 
4863 #ifdef KDB
4864 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4865 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4866     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4867 #endif
4868 
4869 static void
4870 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4871 {
4872 
4873 #ifdef KDB
4874 	if (vfs_badlock_backtrace)
4875 		kdb_backtrace();
4876 #endif
4877 	if (vfs_badlock_vnode)
4878 		vn_printf(vp, "vnode ");
4879 	if (vfs_badlock_print)
4880 		printf("%s: %p %s\n", str, (void *)vp, msg);
4881 	if (vfs_badlock_ddb)
4882 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4883 }
4884 
4885 void
4886 assert_vi_locked(struct vnode *vp, const char *str)
4887 {
4888 
4889 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4890 		vfs_badlock("interlock is not locked but should be", str, vp);
4891 }
4892 
4893 void
4894 assert_vi_unlocked(struct vnode *vp, const char *str)
4895 {
4896 
4897 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4898 		vfs_badlock("interlock is locked but should not be", str, vp);
4899 }
4900 
4901 void
4902 assert_vop_locked(struct vnode *vp, const char *str)
4903 {
4904 	int locked;
4905 
4906 	if (!IGNORE_LOCK(vp)) {
4907 		locked = VOP_ISLOCKED(vp);
4908 		if (locked == 0 || locked == LK_EXCLOTHER)
4909 			vfs_badlock("is not locked but should be", str, vp);
4910 	}
4911 }
4912 
4913 void
4914 assert_vop_unlocked(struct vnode *vp, const char *str)
4915 {
4916 
4917 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4918 		vfs_badlock("is locked but should not be", str, vp);
4919 }
4920 
4921 void
4922 assert_vop_elocked(struct vnode *vp, const char *str)
4923 {
4924 
4925 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4926 		vfs_badlock("is not exclusive locked but should be", str, vp);
4927 }
4928 #endif /* DEBUG_VFS_LOCKS */
4929 
4930 void
4931 vop_rename_fail(struct vop_rename_args *ap)
4932 {
4933 
4934 	if (ap->a_tvp != NULL)
4935 		vput(ap->a_tvp);
4936 	if (ap->a_tdvp == ap->a_tvp)
4937 		vrele(ap->a_tdvp);
4938 	else
4939 		vput(ap->a_tdvp);
4940 	vrele(ap->a_fdvp);
4941 	vrele(ap->a_fvp);
4942 }
4943 
4944 void
4945 vop_rename_pre(void *ap)
4946 {
4947 	struct vop_rename_args *a = ap;
4948 
4949 #ifdef DEBUG_VFS_LOCKS
4950 	if (a->a_tvp)
4951 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4952 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4953 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4954 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4955 
4956 	/* Check the source (from). */
4957 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4958 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4959 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4960 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4961 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4962 
4963 	/* Check the target. */
4964 	if (a->a_tvp)
4965 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4966 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4967 #endif
4968 	if (a->a_tdvp != a->a_fdvp)
4969 		vhold(a->a_fdvp);
4970 	if (a->a_tvp != a->a_fvp)
4971 		vhold(a->a_fvp);
4972 	vhold(a->a_tdvp);
4973 	if (a->a_tvp)
4974 		vhold(a->a_tvp);
4975 }
4976 
4977 #ifdef DEBUG_VFS_LOCKS
4978 void
4979 vop_strategy_pre(void *ap)
4980 {
4981 	struct vop_strategy_args *a;
4982 	struct buf *bp;
4983 
4984 	a = ap;
4985 	bp = a->a_bp;
4986 
4987 	/*
4988 	 * Cluster ops lock their component buffers but not the IO container.
4989 	 */
4990 	if ((bp->b_flags & B_CLUSTER) != 0)
4991 		return;
4992 
4993 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4994 		if (vfs_badlock_print)
4995 			printf(
4996 			    "VOP_STRATEGY: bp is not locked but should be\n");
4997 		if (vfs_badlock_ddb)
4998 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4999 	}
5000 }
5001 
5002 void
5003 vop_lock_pre(void *ap)
5004 {
5005 	struct vop_lock1_args *a = ap;
5006 
5007 	if ((a->a_flags & LK_INTERLOCK) == 0)
5008 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5009 	else
5010 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5011 }
5012 
5013 void
5014 vop_lock_post(void *ap, int rc)
5015 {
5016 	struct vop_lock1_args *a = ap;
5017 
5018 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5019 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5020 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5021 }
5022 
5023 void
5024 vop_unlock_pre(void *ap)
5025 {
5026 	struct vop_unlock_args *a = ap;
5027 
5028 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5029 }
5030 
5031 void
5032 vop_unlock_post(void *ap, int rc)
5033 {
5034 	return;
5035 }
5036 
5037 void
5038 vop_need_inactive_pre(void *ap)
5039 {
5040 	struct vop_need_inactive_args *a = ap;
5041 
5042 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5043 }
5044 
5045 void
5046 vop_need_inactive_post(void *ap, int rc)
5047 {
5048 	struct vop_need_inactive_args *a = ap;
5049 
5050 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5051 }
5052 #endif
5053 
5054 void
5055 vop_create_post(void *ap, int rc)
5056 {
5057 	struct vop_create_args *a = ap;
5058 
5059 	if (!rc)
5060 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5061 }
5062 
5063 void
5064 vop_deleteextattr_post(void *ap, int rc)
5065 {
5066 	struct vop_deleteextattr_args *a = ap;
5067 
5068 	if (!rc)
5069 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5070 }
5071 
5072 void
5073 vop_link_post(void *ap, int rc)
5074 {
5075 	struct vop_link_args *a = ap;
5076 
5077 	if (!rc) {
5078 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5079 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5080 	}
5081 }
5082 
5083 void
5084 vop_mkdir_post(void *ap, int rc)
5085 {
5086 	struct vop_mkdir_args *a = ap;
5087 
5088 	if (!rc)
5089 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5090 }
5091 
5092 void
5093 vop_mknod_post(void *ap, int rc)
5094 {
5095 	struct vop_mknod_args *a = ap;
5096 
5097 	if (!rc)
5098 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5099 }
5100 
5101 void
5102 vop_reclaim_post(void *ap, int rc)
5103 {
5104 	struct vop_reclaim_args *a = ap;
5105 
5106 	if (!rc)
5107 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5108 }
5109 
5110 void
5111 vop_remove_post(void *ap, int rc)
5112 {
5113 	struct vop_remove_args *a = ap;
5114 
5115 	if (!rc) {
5116 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5117 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5118 	}
5119 }
5120 
5121 void
5122 vop_rename_post(void *ap, int rc)
5123 {
5124 	struct vop_rename_args *a = ap;
5125 	long hint;
5126 
5127 	if (!rc) {
5128 		hint = NOTE_WRITE;
5129 		if (a->a_fdvp == a->a_tdvp) {
5130 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5131 				hint |= NOTE_LINK;
5132 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5133 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5134 		} else {
5135 			hint |= NOTE_EXTEND;
5136 			if (a->a_fvp->v_type == VDIR)
5137 				hint |= NOTE_LINK;
5138 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5139 
5140 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5141 			    a->a_tvp->v_type == VDIR)
5142 				hint &= ~NOTE_LINK;
5143 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5144 		}
5145 
5146 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5147 		if (a->a_tvp)
5148 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5149 	}
5150 	if (a->a_tdvp != a->a_fdvp)
5151 		vdrop(a->a_fdvp);
5152 	if (a->a_tvp != a->a_fvp)
5153 		vdrop(a->a_fvp);
5154 	vdrop(a->a_tdvp);
5155 	if (a->a_tvp)
5156 		vdrop(a->a_tvp);
5157 }
5158 
5159 void
5160 vop_rmdir_post(void *ap, int rc)
5161 {
5162 	struct vop_rmdir_args *a = ap;
5163 
5164 	if (!rc) {
5165 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5166 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5167 	}
5168 }
5169 
5170 void
5171 vop_setattr_post(void *ap, int rc)
5172 {
5173 	struct vop_setattr_args *a = ap;
5174 
5175 	if (!rc)
5176 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5177 }
5178 
5179 void
5180 vop_setextattr_post(void *ap, int rc)
5181 {
5182 	struct vop_setextattr_args *a = ap;
5183 
5184 	if (!rc)
5185 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5186 }
5187 
5188 void
5189 vop_symlink_post(void *ap, int rc)
5190 {
5191 	struct vop_symlink_args *a = ap;
5192 
5193 	if (!rc)
5194 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5195 }
5196 
5197 void
5198 vop_open_post(void *ap, int rc)
5199 {
5200 	struct vop_open_args *a = ap;
5201 
5202 	if (!rc)
5203 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5204 }
5205 
5206 void
5207 vop_close_post(void *ap, int rc)
5208 {
5209 	struct vop_close_args *a = ap;
5210 
5211 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5212 	    !VN_IS_DOOMED(a->a_vp))) {
5213 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5214 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5215 	}
5216 }
5217 
5218 void
5219 vop_read_post(void *ap, int rc)
5220 {
5221 	struct vop_read_args *a = ap;
5222 
5223 	if (!rc)
5224 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5225 }
5226 
5227 void
5228 vop_readdir_post(void *ap, int rc)
5229 {
5230 	struct vop_readdir_args *a = ap;
5231 
5232 	if (!rc)
5233 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5234 }
5235 
5236 static struct knlist fs_knlist;
5237 
5238 static void
5239 vfs_event_init(void *arg)
5240 {
5241 	knlist_init_mtx(&fs_knlist, NULL);
5242 }
5243 /* XXX - correct order? */
5244 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5245 
5246 void
5247 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5248 {
5249 
5250 	KNOTE_UNLOCKED(&fs_knlist, event);
5251 }
5252 
5253 static int	filt_fsattach(struct knote *kn);
5254 static void	filt_fsdetach(struct knote *kn);
5255 static int	filt_fsevent(struct knote *kn, long hint);
5256 
5257 struct filterops fs_filtops = {
5258 	.f_isfd = 0,
5259 	.f_attach = filt_fsattach,
5260 	.f_detach = filt_fsdetach,
5261 	.f_event = filt_fsevent
5262 };
5263 
5264 static int
5265 filt_fsattach(struct knote *kn)
5266 {
5267 
5268 	kn->kn_flags |= EV_CLEAR;
5269 	knlist_add(&fs_knlist, kn, 0);
5270 	return (0);
5271 }
5272 
5273 static void
5274 filt_fsdetach(struct knote *kn)
5275 {
5276 
5277 	knlist_remove(&fs_knlist, kn, 0);
5278 }
5279 
5280 static int
5281 filt_fsevent(struct knote *kn, long hint)
5282 {
5283 
5284 	kn->kn_fflags |= hint;
5285 	return (kn->kn_fflags != 0);
5286 }
5287 
5288 static int
5289 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5290 {
5291 	struct vfsidctl vc;
5292 	int error;
5293 	struct mount *mp;
5294 
5295 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5296 	if (error)
5297 		return (error);
5298 	if (vc.vc_vers != VFS_CTL_VERS1)
5299 		return (EINVAL);
5300 	mp = vfs_getvfs(&vc.vc_fsid);
5301 	if (mp == NULL)
5302 		return (ENOENT);
5303 	/* ensure that a specific sysctl goes to the right filesystem. */
5304 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5305 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5306 		vfs_rel(mp);
5307 		return (EINVAL);
5308 	}
5309 	VCTLTOREQ(&vc, req);
5310 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5311 	vfs_rel(mp);
5312 	return (error);
5313 }
5314 
5315 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5316     NULL, 0, sysctl_vfs_ctl, "",
5317     "Sysctl by fsid");
5318 
5319 /*
5320  * Function to initialize a va_filerev field sensibly.
5321  * XXX: Wouldn't a random number make a lot more sense ??
5322  */
5323 u_quad_t
5324 init_va_filerev(void)
5325 {
5326 	struct bintime bt;
5327 
5328 	getbinuptime(&bt);
5329 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5330 }
5331 
5332 static int	filt_vfsread(struct knote *kn, long hint);
5333 static int	filt_vfswrite(struct knote *kn, long hint);
5334 static int	filt_vfsvnode(struct knote *kn, long hint);
5335 static void	filt_vfsdetach(struct knote *kn);
5336 static struct filterops vfsread_filtops = {
5337 	.f_isfd = 1,
5338 	.f_detach = filt_vfsdetach,
5339 	.f_event = filt_vfsread
5340 };
5341 static struct filterops vfswrite_filtops = {
5342 	.f_isfd = 1,
5343 	.f_detach = filt_vfsdetach,
5344 	.f_event = filt_vfswrite
5345 };
5346 static struct filterops vfsvnode_filtops = {
5347 	.f_isfd = 1,
5348 	.f_detach = filt_vfsdetach,
5349 	.f_event = filt_vfsvnode
5350 };
5351 
5352 static void
5353 vfs_knllock(void *arg)
5354 {
5355 	struct vnode *vp = arg;
5356 
5357 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5358 }
5359 
5360 static void
5361 vfs_knlunlock(void *arg)
5362 {
5363 	struct vnode *vp = arg;
5364 
5365 	VOP_UNLOCK(vp);
5366 }
5367 
5368 static void
5369 vfs_knl_assert_locked(void *arg)
5370 {
5371 #ifdef DEBUG_VFS_LOCKS
5372 	struct vnode *vp = arg;
5373 
5374 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5375 #endif
5376 }
5377 
5378 static void
5379 vfs_knl_assert_unlocked(void *arg)
5380 {
5381 #ifdef DEBUG_VFS_LOCKS
5382 	struct vnode *vp = arg;
5383 
5384 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5385 #endif
5386 }
5387 
5388 int
5389 vfs_kqfilter(struct vop_kqfilter_args *ap)
5390 {
5391 	struct vnode *vp = ap->a_vp;
5392 	struct knote *kn = ap->a_kn;
5393 	struct knlist *knl;
5394 
5395 	switch (kn->kn_filter) {
5396 	case EVFILT_READ:
5397 		kn->kn_fop = &vfsread_filtops;
5398 		break;
5399 	case EVFILT_WRITE:
5400 		kn->kn_fop = &vfswrite_filtops;
5401 		break;
5402 	case EVFILT_VNODE:
5403 		kn->kn_fop = &vfsvnode_filtops;
5404 		break;
5405 	default:
5406 		return (EINVAL);
5407 	}
5408 
5409 	kn->kn_hook = (caddr_t)vp;
5410 
5411 	v_addpollinfo(vp);
5412 	if (vp->v_pollinfo == NULL)
5413 		return (ENOMEM);
5414 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5415 	vhold(vp);
5416 	knlist_add(knl, kn, 0);
5417 
5418 	return (0);
5419 }
5420 
5421 /*
5422  * Detach knote from vnode
5423  */
5424 static void
5425 filt_vfsdetach(struct knote *kn)
5426 {
5427 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5428 
5429 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5430 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5431 	vdrop(vp);
5432 }
5433 
5434 /*ARGSUSED*/
5435 static int
5436 filt_vfsread(struct knote *kn, long hint)
5437 {
5438 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5439 	struct vattr va;
5440 	int res;
5441 
5442 	/*
5443 	 * filesystem is gone, so set the EOF flag and schedule
5444 	 * the knote for deletion.
5445 	 */
5446 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5447 		VI_LOCK(vp);
5448 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5449 		VI_UNLOCK(vp);
5450 		return (1);
5451 	}
5452 
5453 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5454 		return (0);
5455 
5456 	VI_LOCK(vp);
5457 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5458 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5459 	VI_UNLOCK(vp);
5460 	return (res);
5461 }
5462 
5463 /*ARGSUSED*/
5464 static int
5465 filt_vfswrite(struct knote *kn, long hint)
5466 {
5467 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5468 
5469 	VI_LOCK(vp);
5470 
5471 	/*
5472 	 * filesystem is gone, so set the EOF flag and schedule
5473 	 * the knote for deletion.
5474 	 */
5475 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5476 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5477 
5478 	kn->kn_data = 0;
5479 	VI_UNLOCK(vp);
5480 	return (1);
5481 }
5482 
5483 static int
5484 filt_vfsvnode(struct knote *kn, long hint)
5485 {
5486 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5487 	int res;
5488 
5489 	VI_LOCK(vp);
5490 	if (kn->kn_sfflags & hint)
5491 		kn->kn_fflags |= hint;
5492 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5493 		kn->kn_flags |= EV_EOF;
5494 		VI_UNLOCK(vp);
5495 		return (1);
5496 	}
5497 	res = (kn->kn_fflags != 0);
5498 	VI_UNLOCK(vp);
5499 	return (res);
5500 }
5501 
5502 /*
5503  * Returns whether the directory is empty or not.
5504  * If it is empty, the return value is 0; otherwise
5505  * the return value is an error value (which may
5506  * be ENOTEMPTY).
5507  */
5508 int
5509 vfs_emptydir(struct vnode *vp)
5510 {
5511 	struct uio uio;
5512 	struct iovec iov;
5513 	struct dirent *dirent, *dp, *endp;
5514 	int error, eof;
5515 
5516 	error = 0;
5517 	eof = 0;
5518 
5519 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5520 
5521 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5522 	iov.iov_base = dirent;
5523 	iov.iov_len = sizeof(struct dirent);
5524 
5525 	uio.uio_iov = &iov;
5526 	uio.uio_iovcnt = 1;
5527 	uio.uio_offset = 0;
5528 	uio.uio_resid = sizeof(struct dirent);
5529 	uio.uio_segflg = UIO_SYSSPACE;
5530 	uio.uio_rw = UIO_READ;
5531 	uio.uio_td = curthread;
5532 
5533 	while (eof == 0 && error == 0) {
5534 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5535 		    NULL, NULL);
5536 		if (error != 0)
5537 			break;
5538 		endp = (void *)((uint8_t *)dirent +
5539 		    sizeof(struct dirent) - uio.uio_resid);
5540 		for (dp = dirent; dp < endp;
5541 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5542 			if (dp->d_type == DT_WHT)
5543 				continue;
5544 			if (dp->d_namlen == 0)
5545 				continue;
5546 			if (dp->d_type != DT_DIR &&
5547 			    dp->d_type != DT_UNKNOWN) {
5548 				error = ENOTEMPTY;
5549 				break;
5550 			}
5551 			if (dp->d_namlen > 2) {
5552 				error = ENOTEMPTY;
5553 				break;
5554 			}
5555 			if (dp->d_namlen == 1 &&
5556 			    dp->d_name[0] != '.') {
5557 				error = ENOTEMPTY;
5558 				break;
5559 			}
5560 			if (dp->d_namlen == 2 &&
5561 			    dp->d_name[1] != '.') {
5562 				error = ENOTEMPTY;
5563 				break;
5564 			}
5565 			uio.uio_resid = sizeof(struct dirent);
5566 		}
5567 	}
5568 	free(dirent, M_TEMP);
5569 	return (error);
5570 }
5571 
5572 int
5573 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5574 {
5575 	int error;
5576 
5577 	if (dp->d_reclen > ap->a_uio->uio_resid)
5578 		return (ENAMETOOLONG);
5579 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5580 	if (error) {
5581 		if (ap->a_ncookies != NULL) {
5582 			if (ap->a_cookies != NULL)
5583 				free(ap->a_cookies, M_TEMP);
5584 			ap->a_cookies = NULL;
5585 			*ap->a_ncookies = 0;
5586 		}
5587 		return (error);
5588 	}
5589 	if (ap->a_ncookies == NULL)
5590 		return (0);
5591 
5592 	KASSERT(ap->a_cookies,
5593 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5594 
5595 	*ap->a_cookies = realloc(*ap->a_cookies,
5596 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5597 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5598 	*ap->a_ncookies += 1;
5599 	return (0);
5600 }
5601 
5602 /*
5603  * Mark for update the access time of the file if the filesystem
5604  * supports VOP_MARKATIME.  This functionality is used by execve and
5605  * mmap, so we want to avoid the I/O implied by directly setting
5606  * va_atime for the sake of efficiency.
5607  */
5608 void
5609 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5610 {
5611 	struct mount *mp;
5612 
5613 	mp = vp->v_mount;
5614 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5615 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5616 		(void)VOP_MARKATIME(vp);
5617 }
5618 
5619 /*
5620  * The purpose of this routine is to remove granularity from accmode_t,
5621  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5622  * VADMIN and VAPPEND.
5623  *
5624  * If it returns 0, the caller is supposed to continue with the usual
5625  * access checks using 'accmode' as modified by this routine.  If it
5626  * returns nonzero value, the caller is supposed to return that value
5627  * as errno.
5628  *
5629  * Note that after this routine runs, accmode may be zero.
5630  */
5631 int
5632 vfs_unixify_accmode(accmode_t *accmode)
5633 {
5634 	/*
5635 	 * There is no way to specify explicit "deny" rule using
5636 	 * file mode or POSIX.1e ACLs.
5637 	 */
5638 	if (*accmode & VEXPLICIT_DENY) {
5639 		*accmode = 0;
5640 		return (0);
5641 	}
5642 
5643 	/*
5644 	 * None of these can be translated into usual access bits.
5645 	 * Also, the common case for NFSv4 ACLs is to not contain
5646 	 * either of these bits. Caller should check for VWRITE
5647 	 * on the containing directory instead.
5648 	 */
5649 	if (*accmode & (VDELETE_CHILD | VDELETE))
5650 		return (EPERM);
5651 
5652 	if (*accmode & VADMIN_PERMS) {
5653 		*accmode &= ~VADMIN_PERMS;
5654 		*accmode |= VADMIN;
5655 	}
5656 
5657 	/*
5658 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5659 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5660 	 */
5661 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5662 
5663 	return (0);
5664 }
5665 
5666 /*
5667  * Clear out a doomed vnode (if any) and replace it with a new one as long
5668  * as the fs is not being unmounted. Return the root vnode to the caller.
5669  */
5670 static int __noinline
5671 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
5672 {
5673 	struct vnode *vp;
5674 	int error;
5675 
5676 restart:
5677 	if (mp->mnt_rootvnode != NULL) {
5678 		MNT_ILOCK(mp);
5679 		vp = mp->mnt_rootvnode;
5680 		if (vp != NULL) {
5681 			if (!VN_IS_DOOMED(vp)) {
5682 				vrefact(vp);
5683 				MNT_IUNLOCK(mp);
5684 				error = vn_lock(vp, flags);
5685 				if (error == 0) {
5686 					*vpp = vp;
5687 					return (0);
5688 				}
5689 				vrele(vp);
5690 				goto restart;
5691 			}
5692 			/*
5693 			 * Clear the old one.
5694 			 */
5695 			mp->mnt_rootvnode = NULL;
5696 		}
5697 		MNT_IUNLOCK(mp);
5698 		if (vp != NULL) {
5699 			/*
5700 			 * Paired with a fence in vfs_op_thread_exit().
5701 			 */
5702 			atomic_thread_fence_acq();
5703 			vfs_op_barrier_wait(mp);
5704 			vrele(vp);
5705 		}
5706 	}
5707 	error = VFS_CACHEDROOT(mp, flags, vpp);
5708 	if (error != 0)
5709 		return (error);
5710 	if (mp->mnt_vfs_ops == 0) {
5711 		MNT_ILOCK(mp);
5712 		if (mp->mnt_vfs_ops != 0) {
5713 			MNT_IUNLOCK(mp);
5714 			return (0);
5715 		}
5716 		if (mp->mnt_rootvnode == NULL) {
5717 			vrefact(*vpp);
5718 			mp->mnt_rootvnode = *vpp;
5719 		} else {
5720 			if (mp->mnt_rootvnode != *vpp) {
5721 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
5722 					panic("%s: mismatch between vnode returned "
5723 					    " by VFS_CACHEDROOT and the one cached "
5724 					    " (%p != %p)",
5725 					    __func__, *vpp, mp->mnt_rootvnode);
5726 				}
5727 			}
5728 		}
5729 		MNT_IUNLOCK(mp);
5730 	}
5731 	return (0);
5732 }
5733 
5734 int
5735 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
5736 {
5737 	struct vnode *vp;
5738 	int error;
5739 
5740 	if (!vfs_op_thread_enter(mp))
5741 		return (vfs_cache_root_fallback(mp, flags, vpp));
5742 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
5743 	if (vp == NULL || VN_IS_DOOMED(vp)) {
5744 		vfs_op_thread_exit(mp);
5745 		return (vfs_cache_root_fallback(mp, flags, vpp));
5746 	}
5747 	vrefact(vp);
5748 	vfs_op_thread_exit(mp);
5749 	error = vn_lock(vp, flags);
5750 	if (error != 0) {
5751 		vrele(vp);
5752 		return (vfs_cache_root_fallback(mp, flags, vpp));
5753 	}
5754 	*vpp = vp;
5755 	return (0);
5756 }
5757 
5758 struct vnode *
5759 vfs_cache_root_clear(struct mount *mp)
5760 {
5761 	struct vnode *vp;
5762 
5763 	/*
5764 	 * ops > 0 guarantees there is nobody who can see this vnode
5765 	 */
5766 	MPASS(mp->mnt_vfs_ops > 0);
5767 	vp = mp->mnt_rootvnode;
5768 	mp->mnt_rootvnode = NULL;
5769 	return (vp);
5770 }
5771 
5772 void
5773 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
5774 {
5775 
5776 	MPASS(mp->mnt_vfs_ops > 0);
5777 	vrefact(vp);
5778 	mp->mnt_rootvnode = vp;
5779 }
5780 
5781 /*
5782  * These are helper functions for filesystems to traverse all
5783  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5784  *
5785  * This interface replaces MNT_VNODE_FOREACH.
5786  */
5787 
5788 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5789 
5790 struct vnode *
5791 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5792 {
5793 	struct vnode *vp;
5794 
5795 	if (should_yield())
5796 		kern_yield(PRI_USER);
5797 	MNT_ILOCK(mp);
5798 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5799 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5800 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5801 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
5802 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
5803 			continue;
5804 		VI_LOCK(vp);
5805 		if (VN_IS_DOOMED(vp)) {
5806 			VI_UNLOCK(vp);
5807 			continue;
5808 		}
5809 		break;
5810 	}
5811 	if (vp == NULL) {
5812 		__mnt_vnode_markerfree_all(mvp, mp);
5813 		/* MNT_IUNLOCK(mp); -- done in above function */
5814 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5815 		return (NULL);
5816 	}
5817 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5818 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5819 	MNT_IUNLOCK(mp);
5820 	return (vp);
5821 }
5822 
5823 struct vnode *
5824 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5825 {
5826 	struct vnode *vp;
5827 
5828 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5829 	MNT_ILOCK(mp);
5830 	MNT_REF(mp);
5831 	(*mvp)->v_mount = mp;
5832 	(*mvp)->v_type = VMARKER;
5833 
5834 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5835 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
5836 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
5837 			continue;
5838 		VI_LOCK(vp);
5839 		if (VN_IS_DOOMED(vp)) {
5840 			VI_UNLOCK(vp);
5841 			continue;
5842 		}
5843 		break;
5844 	}
5845 	if (vp == NULL) {
5846 		MNT_REL(mp);
5847 		MNT_IUNLOCK(mp);
5848 		free(*mvp, M_VNODE_MARKER);
5849 		*mvp = NULL;
5850 		return (NULL);
5851 	}
5852 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5853 	MNT_IUNLOCK(mp);
5854 	return (vp);
5855 }
5856 
5857 void
5858 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5859 {
5860 
5861 	if (*mvp == NULL) {
5862 		MNT_IUNLOCK(mp);
5863 		return;
5864 	}
5865 
5866 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5867 
5868 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5869 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5870 	MNT_REL(mp);
5871 	MNT_IUNLOCK(mp);
5872 	free(*mvp, M_VNODE_MARKER);
5873 	*mvp = NULL;
5874 }
5875 
5876 /*
5877  * These are helper functions for filesystems to traverse their
5878  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5879  */
5880 static void
5881 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5882 {
5883 
5884 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5885 
5886 	MNT_ILOCK(mp);
5887 	MNT_REL(mp);
5888 	MNT_IUNLOCK(mp);
5889 	free(*mvp, M_VNODE_MARKER);
5890 	*mvp = NULL;
5891 }
5892 
5893 /*
5894  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5895  * conventional lock order during mnt_vnode_next_active iteration.
5896  *
5897  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5898  * The list lock is dropped and reacquired.  On success, both locks are held.
5899  * On failure, the mount vnode list lock is held but the vnode interlock is
5900  * not, and the procedure may have yielded.
5901  */
5902 static bool
5903 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5904     struct vnode *vp)
5905 {
5906 	const struct vnode *tmp;
5907 	bool held, ret;
5908 
5909 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5910 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5911 	    ("%s: bad marker", __func__));
5912 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5913 	    ("%s: inappropriate vnode", __func__));
5914 	ASSERT_VI_UNLOCKED(vp, __func__);
5915 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5916 
5917 	ret = false;
5918 
5919 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5920 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5921 
5922 	/*
5923 	 * Use a hold to prevent vp from disappearing while the mount vnode
5924 	 * list lock is dropped and reacquired.  Normally a hold would be
5925 	 * acquired with vhold(), but that might try to acquire the vnode
5926 	 * interlock, which would be a LOR with the mount vnode list lock.
5927 	 */
5928 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5929 	mtx_unlock(&mp->mnt_listmtx);
5930 	if (!held)
5931 		goto abort;
5932 	VI_LOCK(vp);
5933 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5934 		vdropl(vp);
5935 		goto abort;
5936 	}
5937 	mtx_lock(&mp->mnt_listmtx);
5938 
5939 	/*
5940 	 * Determine whether the vnode is still the next one after the marker,
5941 	 * excepting any other markers.  If the vnode has not been doomed by
5942 	 * vgone() then the hold should have ensured that it remained on the
5943 	 * active list.  If it has been doomed but is still on the active list,
5944 	 * don't abort, but rather skip over it (avoid spinning on doomed
5945 	 * vnodes).
5946 	 */
5947 	tmp = mvp;
5948 	do {
5949 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5950 	} while (tmp != NULL && tmp->v_type == VMARKER);
5951 	if (tmp != vp) {
5952 		mtx_unlock(&mp->mnt_listmtx);
5953 		VI_UNLOCK(vp);
5954 		goto abort;
5955 	}
5956 
5957 	ret = true;
5958 	goto out;
5959 abort:
5960 	maybe_yield();
5961 	mtx_lock(&mp->mnt_listmtx);
5962 out:
5963 	if (ret)
5964 		ASSERT_VI_LOCKED(vp, __func__);
5965 	else
5966 		ASSERT_VI_UNLOCKED(vp, __func__);
5967 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5968 	return (ret);
5969 }
5970 
5971 static struct vnode *
5972 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5973 {
5974 	struct vnode *vp, *nvp;
5975 
5976 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5977 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5978 restart:
5979 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5980 	while (vp != NULL) {
5981 		if (vp->v_type == VMARKER) {
5982 			vp = TAILQ_NEXT(vp, v_actfreelist);
5983 			continue;
5984 		}
5985 		/*
5986 		 * Try-lock because this is the wrong lock order.  If that does
5987 		 * not succeed, drop the mount vnode list lock and try to
5988 		 * reacquire it and the vnode interlock in the right order.
5989 		 */
5990 		if (!VI_TRYLOCK(vp) &&
5991 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5992 			goto restart;
5993 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5994 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5995 		    ("alien vnode on the active list %p %p", vp, mp));
5996 		if (vp->v_mount == mp && !VN_IS_DOOMED(vp))
5997 			break;
5998 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5999 		VI_UNLOCK(vp);
6000 		vp = nvp;
6001 	}
6002 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
6003 
6004 	/* Check if we are done */
6005 	if (vp == NULL) {
6006 		mtx_unlock(&mp->mnt_listmtx);
6007 		mnt_vnode_markerfree_active(mvp, mp);
6008 		return (NULL);
6009 	}
6010 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
6011 	mtx_unlock(&mp->mnt_listmtx);
6012 	ASSERT_VI_LOCKED(vp, "active iter");
6013 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
6014 	return (vp);
6015 }
6016 
6017 struct vnode *
6018 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
6019 {
6020 
6021 	if (should_yield())
6022 		kern_yield(PRI_USER);
6023 	mtx_lock(&mp->mnt_listmtx);
6024 	return (mnt_vnode_next_active(mvp, mp));
6025 }
6026 
6027 struct vnode *
6028 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
6029 {
6030 	struct vnode *vp;
6031 
6032 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
6033 	MNT_ILOCK(mp);
6034 	MNT_REF(mp);
6035 	MNT_IUNLOCK(mp);
6036 	(*mvp)->v_type = VMARKER;
6037 	(*mvp)->v_mount = mp;
6038 
6039 	mtx_lock(&mp->mnt_listmtx);
6040 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
6041 	if (vp == NULL) {
6042 		mtx_unlock(&mp->mnt_listmtx);
6043 		mnt_vnode_markerfree_active(mvp, mp);
6044 		return (NULL);
6045 	}
6046 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
6047 	return (mnt_vnode_next_active(mvp, mp));
6048 }
6049 
6050 void
6051 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
6052 {
6053 
6054 	if (*mvp == NULL)
6055 		return;
6056 
6057 	mtx_lock(&mp->mnt_listmtx);
6058 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
6059 	mtx_unlock(&mp->mnt_listmtx);
6060 	mnt_vnode_markerfree_active(mvp, mp);
6061 }
6062