xref: /freebsd/sys/kern/vfs_subr.c (revision 2c3f47a727372086b41ef9ce06eb1f1eb83a67d3)
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 VI_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, &reassignbufcalls, 0,
214     "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 + 1, 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 ((vp->v_iflag & VI_DOOMED) != 0) {
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, 0);
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 	MNT_ILOCK(mp);
645 	MNT_REF(mp);
646 	/*
647 	 * If mount point is currently being unmounted, sleep until the
648 	 * mount point fate is decided.  If thread doing the unmounting fails,
649 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
650 	 * that this mount point has survived the unmount attempt and vfs_busy
651 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
652 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
653 	 * about to be really destroyed.  vfs_busy needs to release its
654 	 * reference on the mount point in this case and return with ENOENT,
655 	 * telling the caller that mount mount it tried to busy is no longer
656 	 * valid.
657 	 */
658 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
659 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
660 			MNT_REL(mp);
661 			MNT_IUNLOCK(mp);
662 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
663 			    __func__);
664 			return (ENOENT);
665 		}
666 		if (flags & MBF_MNTLSTLOCK)
667 			mtx_unlock(&mountlist_mtx);
668 		mp->mnt_kern_flag |= MNTK_MWAIT;
669 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
670 		if (flags & MBF_MNTLSTLOCK)
671 			mtx_lock(&mountlist_mtx);
672 		MNT_ILOCK(mp);
673 	}
674 	if (flags & MBF_MNTLSTLOCK)
675 		mtx_unlock(&mountlist_mtx);
676 	mp->mnt_lockref++;
677 	MNT_IUNLOCK(mp);
678 	return (0);
679 }
680 
681 /*
682  * Free a busy filesystem.
683  */
684 void
685 vfs_unbusy(struct mount *mp)
686 {
687 
688 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
689 	MNT_ILOCK(mp);
690 	MNT_REL(mp);
691 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
692 	mp->mnt_lockref--;
693 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
694 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
695 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
696 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
697 		wakeup(&mp->mnt_lockref);
698 	}
699 	MNT_IUNLOCK(mp);
700 }
701 
702 /*
703  * Lookup a mount point by filesystem identifier.
704  */
705 struct mount *
706 vfs_getvfs(fsid_t *fsid)
707 {
708 	struct mount *mp;
709 
710 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
711 	mtx_lock(&mountlist_mtx);
712 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
713 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
714 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
715 			vfs_ref(mp);
716 			mtx_unlock(&mountlist_mtx);
717 			return (mp);
718 		}
719 	}
720 	mtx_unlock(&mountlist_mtx);
721 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
722 	return ((struct mount *) 0);
723 }
724 
725 /*
726  * Lookup a mount point by filesystem identifier, busying it before
727  * returning.
728  *
729  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
730  * cache for popular filesystem identifiers.  The cache is lockess, using
731  * the fact that struct mount's are never freed.  In worst case we may
732  * get pointer to unmounted or even different filesystem, so we have to
733  * check what we got, and go slow way if so.
734  */
735 struct mount *
736 vfs_busyfs(fsid_t *fsid)
737 {
738 #define	FSID_CACHE_SIZE	256
739 	typedef struct mount * volatile vmp_t;
740 	static vmp_t cache[FSID_CACHE_SIZE];
741 	struct mount *mp;
742 	int error;
743 	uint32_t hash;
744 
745 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
746 	hash = fsid->val[0] ^ fsid->val[1];
747 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
748 	mp = cache[hash];
749 	if (mp == NULL ||
750 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
751 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
752 		goto slow;
753 	if (vfs_busy(mp, 0) != 0) {
754 		cache[hash] = NULL;
755 		goto slow;
756 	}
757 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
758 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
759 		return (mp);
760 	else
761 	    vfs_unbusy(mp);
762 
763 slow:
764 	mtx_lock(&mountlist_mtx);
765 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
766 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
767 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
768 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
769 			if (error) {
770 				cache[hash] = NULL;
771 				mtx_unlock(&mountlist_mtx);
772 				return (NULL);
773 			}
774 			cache[hash] = mp;
775 			return (mp);
776 		}
777 	}
778 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
779 	mtx_unlock(&mountlist_mtx);
780 	return ((struct mount *) 0);
781 }
782 
783 /*
784  * Check if a user can access privileged mount options.
785  */
786 int
787 vfs_suser(struct mount *mp, struct thread *td)
788 {
789 	int error;
790 
791 	if (jailed(td->td_ucred)) {
792 		/*
793 		 * If the jail of the calling thread lacks permission for
794 		 * this type of file system, deny immediately.
795 		 */
796 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
797 			return (EPERM);
798 
799 		/*
800 		 * If the file system was mounted outside the jail of the
801 		 * calling thread, deny immediately.
802 		 */
803 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
804 			return (EPERM);
805 	}
806 
807 	/*
808 	 * If file system supports delegated administration, we don't check
809 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
810 	 * by the file system itself.
811 	 * If this is not the user that did original mount, we check for
812 	 * the PRIV_VFS_MOUNT_OWNER privilege.
813 	 */
814 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
815 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
816 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
817 			return (error);
818 	}
819 	return (0);
820 }
821 
822 /*
823  * Get a new unique fsid.  Try to make its val[0] unique, since this value
824  * will be used to create fake device numbers for stat().  Also try (but
825  * not so hard) make its val[0] unique mod 2^16, since some emulators only
826  * support 16-bit device numbers.  We end up with unique val[0]'s for the
827  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
828  *
829  * Keep in mind that several mounts may be running in parallel.  Starting
830  * the search one past where the previous search terminated is both a
831  * micro-optimization and a defense against returning the same fsid to
832  * different mounts.
833  */
834 void
835 vfs_getnewfsid(struct mount *mp)
836 {
837 	static uint16_t mntid_base;
838 	struct mount *nmp;
839 	fsid_t tfsid;
840 	int mtype;
841 
842 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
843 	mtx_lock(&mntid_mtx);
844 	mtype = mp->mnt_vfc->vfc_typenum;
845 	tfsid.val[1] = mtype;
846 	mtype = (mtype & 0xFF) << 24;
847 	for (;;) {
848 		tfsid.val[0] = makedev(255,
849 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
850 		mntid_base++;
851 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
852 			break;
853 		vfs_rel(nmp);
854 	}
855 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
856 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
857 	mtx_unlock(&mntid_mtx);
858 }
859 
860 /*
861  * Knob to control the precision of file timestamps:
862  *
863  *   0 = seconds only; nanoseconds zeroed.
864  *   1 = seconds and nanoseconds, accurate within 1/HZ.
865  *   2 = seconds and nanoseconds, truncated to microseconds.
866  * >=3 = seconds and nanoseconds, maximum precision.
867  */
868 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
869 
870 static int timestamp_precision = TSP_USEC;
871 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
872     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
873     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
874     "3+: sec + ns (max. precision))");
875 
876 /*
877  * Get a current timestamp.
878  */
879 void
880 vfs_timestamp(struct timespec *tsp)
881 {
882 	struct timeval tv;
883 
884 	switch (timestamp_precision) {
885 	case TSP_SEC:
886 		tsp->tv_sec = time_second;
887 		tsp->tv_nsec = 0;
888 		break;
889 	case TSP_HZ:
890 		getnanotime(tsp);
891 		break;
892 	case TSP_USEC:
893 		microtime(&tv);
894 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
895 		break;
896 	case TSP_NSEC:
897 	default:
898 		nanotime(tsp);
899 		break;
900 	}
901 }
902 
903 /*
904  * Set vnode attributes to VNOVAL
905  */
906 void
907 vattr_null(struct vattr *vap)
908 {
909 
910 	vap->va_type = VNON;
911 	vap->va_size = VNOVAL;
912 	vap->va_bytes = VNOVAL;
913 	vap->va_mode = VNOVAL;
914 	vap->va_nlink = VNOVAL;
915 	vap->va_uid = VNOVAL;
916 	vap->va_gid = VNOVAL;
917 	vap->va_fsid = VNOVAL;
918 	vap->va_fileid = VNOVAL;
919 	vap->va_blocksize = VNOVAL;
920 	vap->va_rdev = VNOVAL;
921 	vap->va_atime.tv_sec = VNOVAL;
922 	vap->va_atime.tv_nsec = VNOVAL;
923 	vap->va_mtime.tv_sec = VNOVAL;
924 	vap->va_mtime.tv_nsec = VNOVAL;
925 	vap->va_ctime.tv_sec = VNOVAL;
926 	vap->va_ctime.tv_nsec = VNOVAL;
927 	vap->va_birthtime.tv_sec = VNOVAL;
928 	vap->va_birthtime.tv_nsec = VNOVAL;
929 	vap->va_flags = VNOVAL;
930 	vap->va_gen = VNOVAL;
931 	vap->va_vaflags = 0;
932 }
933 
934 /*
935  * This routine is called when we have too many vnodes.  It attempts
936  * to free <count> vnodes and will potentially free vnodes that still
937  * have VM backing store (VM backing store is typically the cause
938  * of a vnode blowout so we want to do this).  Therefore, this operation
939  * is not considered cheap.
940  *
941  * A number of conditions may prevent a vnode from being reclaimed.
942  * the buffer cache may have references on the vnode, a directory
943  * vnode may still have references due to the namei cache representing
944  * underlying files, or the vnode may be in active use.   It is not
945  * desirable to reuse such vnodes.  These conditions may cause the
946  * number of vnodes to reach some minimum value regardless of what
947  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
948  *
949  * @param mp		 Try to reclaim vnodes from this mountpoint
950  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
951  * 			 entries if this argument is strue
952  * @param trigger	 Only reclaim vnodes with fewer than this many resident
953  *			 pages.
954  * @return		 The number of vnodes that were reclaimed.
955  */
956 static int
957 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
958 {
959 	struct vnode *vp;
960 	int count, done, target;
961 
962 	done = 0;
963 	vn_start_write(NULL, &mp, V_WAIT);
964 	MNT_ILOCK(mp);
965 	count = mp->mnt_nvnodelistsize;
966 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
967 	target = target / 10 + 1;
968 	while (count != 0 && done < target) {
969 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
970 		while (vp != NULL && vp->v_type == VMARKER)
971 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
972 		if (vp == NULL)
973 			break;
974 		/*
975 		 * XXX LRU is completely broken for non-free vnodes.  First
976 		 * by calling here in mountpoint order, then by moving
977 		 * unselected vnodes to the end here, and most grossly by
978 		 * removing the vlruvp() function that was supposed to
979 		 * maintain the order.  (This function was born broken
980 		 * since syncer problems prevented it doing anything.)  The
981 		 * order is closer to LRC (C = Created).
982 		 *
983 		 * LRU reclaiming of vnodes seems to have last worked in
984 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
985 		 * Then there was no hold count, and inactive vnodes were
986 		 * simply put on the free list in LRU order.  The separate
987 		 * lists also break LRU.  We prefer to reclaim from the
988 		 * free list for technical reasons.  This tends to thrash
989 		 * the free list to keep very unrecently used held vnodes.
990 		 * The problem is mitigated by keeping the free list large.
991 		 */
992 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
993 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
994 		--count;
995 		if (!VI_TRYLOCK(vp))
996 			goto next_iter;
997 		/*
998 		 * If it's been deconstructed already, it's still
999 		 * referenced, or it exceeds the trigger, skip it.
1000 		 * Also skip free vnodes.  We are trying to make space
1001 		 * to expand the free list, not reduce it.
1002 		 */
1003 		if (vp->v_usecount ||
1004 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1005 		    ((vp->v_iflag & VI_FREE) != 0) ||
1006 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
1007 		    vp->v_object->resident_page_count > trigger)) {
1008 			VI_UNLOCK(vp);
1009 			goto next_iter;
1010 		}
1011 		MNT_IUNLOCK(mp);
1012 		vholdl(vp);
1013 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1014 			vdrop(vp);
1015 			goto next_iter_mntunlocked;
1016 		}
1017 		VI_LOCK(vp);
1018 		/*
1019 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1020 		 * the vnode interlock and before it was locked again.
1021 		 *
1022 		 * It is not necessary to recheck VI_DOOMED because it can
1023 		 * only be set by another thread that holds both the vnode
1024 		 * lock and vnode interlock.  If another thread has the
1025 		 * vnode lock before we get to VOP_LOCK() and obtains the
1026 		 * vnode interlock after VOP_LOCK() drops the vnode
1027 		 * interlock, the other thread will be unable to drop the
1028 		 * vnode lock before our VOP_LOCK() call fails.
1029 		 */
1030 		if (vp->v_usecount ||
1031 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1032 		    (vp->v_iflag & VI_FREE) != 0 ||
1033 		    (vp->v_object != NULL &&
1034 		    vp->v_object->resident_page_count > trigger)) {
1035 			VOP_UNLOCK(vp, 0);
1036 			vdropl(vp);
1037 			goto next_iter_mntunlocked;
1038 		}
1039 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
1040 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
1041 		counter_u64_add(recycles_count, 1);
1042 		vgonel(vp);
1043 		VOP_UNLOCK(vp, 0);
1044 		vdropl(vp);
1045 		done++;
1046 next_iter_mntunlocked:
1047 		if (!should_yield())
1048 			goto relock_mnt;
1049 		goto yield;
1050 next_iter:
1051 		if (!should_yield())
1052 			continue;
1053 		MNT_IUNLOCK(mp);
1054 yield:
1055 		kern_yield(PRI_USER);
1056 relock_mnt:
1057 		MNT_ILOCK(mp);
1058 	}
1059 	MNT_IUNLOCK(mp);
1060 	vn_finished_write(mp);
1061 	return done;
1062 }
1063 
1064 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1065 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1066     0,
1067     "limit on vnode free requests per call to the vnlru_free routine");
1068 
1069 /*
1070  * Attempt to reduce the free list by the requested amount.
1071  */
1072 static void
1073 vnlru_free_locked(int count, struct vfsops *mnt_op)
1074 {
1075 	struct vnode *vp;
1076 	struct mount *mp;
1077 	bool tried_batches;
1078 
1079 	tried_batches = false;
1080 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1081 	if (count > max_vnlru_free)
1082 		count = max_vnlru_free;
1083 	for (; count > 0; count--) {
1084 		vp = TAILQ_FIRST(&vnode_free_list);
1085 		/*
1086 		 * The list can be modified while the free_list_mtx
1087 		 * has been dropped and vp could be NULL here.
1088 		 */
1089 		if (vp == NULL) {
1090 			if (tried_batches)
1091 				break;
1092 			mtx_unlock(&vnode_free_list_mtx);
1093 			vnlru_return_batches(mnt_op);
1094 			tried_batches = true;
1095 			mtx_lock(&vnode_free_list_mtx);
1096 			continue;
1097 		}
1098 
1099 		VNASSERT(vp->v_op != NULL, vp,
1100 		    ("vnlru_free: vnode already reclaimed."));
1101 		KASSERT((vp->v_iflag & VI_FREE) != 0,
1102 		    ("Removing vnode not on freelist"));
1103 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1104 		    ("Mangling active vnode"));
1105 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1106 
1107 		/*
1108 		 * Don't recycle if our vnode is from different type
1109 		 * of mount point.  Note that mp is type-safe, the
1110 		 * check does not reach unmapped address even if
1111 		 * vnode is reclaimed.
1112 		 * Don't recycle if we can't get the interlock without
1113 		 * blocking.
1114 		 */
1115 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1116 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1117 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1118 			continue;
1119 		}
1120 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1121 		    vp, ("vp inconsistent on freelist"));
1122 
1123 		/*
1124 		 * The clear of VI_FREE prevents activation of the
1125 		 * vnode.  There is no sense in putting the vnode on
1126 		 * the mount point active list, only to remove it
1127 		 * later during recycling.  Inline the relevant part
1128 		 * of vholdl(), to avoid triggering assertions or
1129 		 * activating.
1130 		 */
1131 		freevnodes--;
1132 		vp->v_iflag &= ~VI_FREE;
1133 		VNODE_REFCOUNT_FENCE_REL();
1134 		refcount_acquire(&vp->v_holdcnt);
1135 
1136 		mtx_unlock(&vnode_free_list_mtx);
1137 		VI_UNLOCK(vp);
1138 		vtryrecycle(vp);
1139 		/*
1140 		 * If the recycled succeeded this vdrop will actually free
1141 		 * the vnode.  If not it will simply place it back on
1142 		 * the free list.
1143 		 */
1144 		vdrop(vp);
1145 		mtx_lock(&vnode_free_list_mtx);
1146 	}
1147 }
1148 
1149 void
1150 vnlru_free(int count, struct vfsops *mnt_op)
1151 {
1152 
1153 	mtx_lock(&vnode_free_list_mtx);
1154 	vnlru_free_locked(count, mnt_op);
1155 	mtx_unlock(&vnode_free_list_mtx);
1156 }
1157 
1158 
1159 /* XXX some names and initialization are bad for limits and watermarks. */
1160 static int
1161 vspace(void)
1162 {
1163 	int space;
1164 
1165 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1166 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1167 	vlowat = vhiwat / 2;
1168 	if (numvnodes > desiredvnodes)
1169 		return (0);
1170 	space = desiredvnodes - numvnodes;
1171 	if (freevnodes > wantfreevnodes)
1172 		space += freevnodes - wantfreevnodes;
1173 	return (space);
1174 }
1175 
1176 static void
1177 vnlru_return_batch_locked(struct mount *mp)
1178 {
1179 	struct vnode *vp;
1180 
1181 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1182 
1183 	if (mp->mnt_tmpfreevnodelistsize == 0)
1184 		return;
1185 
1186 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1187 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1188 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1189 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1190 	}
1191 	mtx_lock(&vnode_free_list_mtx);
1192 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1193 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1194 	mtx_unlock(&vnode_free_list_mtx);
1195 	mp->mnt_tmpfreevnodelistsize = 0;
1196 }
1197 
1198 static void
1199 vnlru_return_batch(struct mount *mp)
1200 {
1201 
1202 	mtx_lock(&mp->mnt_listmtx);
1203 	vnlru_return_batch_locked(mp);
1204 	mtx_unlock(&mp->mnt_listmtx);
1205 }
1206 
1207 static void
1208 vnlru_return_batches(struct vfsops *mnt_op)
1209 {
1210 	struct mount *mp, *nmp;
1211 	bool need_unbusy;
1212 
1213 	mtx_lock(&mountlist_mtx);
1214 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1215 		need_unbusy = false;
1216 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1217 			goto next;
1218 		if (mp->mnt_tmpfreevnodelistsize == 0)
1219 			goto next;
1220 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1221 			vnlru_return_batch(mp);
1222 			need_unbusy = true;
1223 			mtx_lock(&mountlist_mtx);
1224 		}
1225 next:
1226 		nmp = TAILQ_NEXT(mp, mnt_list);
1227 		if (need_unbusy)
1228 			vfs_unbusy(mp);
1229 	}
1230 	mtx_unlock(&mountlist_mtx);
1231 }
1232 
1233 /*
1234  * Attempt to recycle vnodes in a context that is always safe to block.
1235  * Calling vlrurecycle() from the bowels of filesystem code has some
1236  * interesting deadlock problems.
1237  */
1238 static struct proc *vnlruproc;
1239 static int vnlruproc_sig;
1240 
1241 static void
1242 vnlru_proc(void)
1243 {
1244 	struct mount *mp, *nmp;
1245 	unsigned long onumvnodes;
1246 	int done, force, trigger, usevnodes;
1247 	bool reclaim_nc_src;
1248 
1249 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1250 	    SHUTDOWN_PRI_FIRST);
1251 
1252 	force = 0;
1253 	for (;;) {
1254 		kproc_suspend_check(vnlruproc);
1255 		mtx_lock(&vnode_free_list_mtx);
1256 		/*
1257 		 * If numvnodes is too large (due to desiredvnodes being
1258 		 * adjusted using its sysctl, or emergency growth), first
1259 		 * try to reduce it by discarding from the free list.
1260 		 */
1261 		if (numvnodes > desiredvnodes)
1262 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1263 		/*
1264 		 * Sleep if the vnode cache is in a good state.  This is
1265 		 * when it is not over-full and has space for about a 4%
1266 		 * or 9% expansion (by growing its size or inexcessively
1267 		 * reducing its free list).  Otherwise, try to reclaim
1268 		 * space for a 10% expansion.
1269 		 */
1270 		if (vstir && force == 0) {
1271 			force = 1;
1272 			vstir = 0;
1273 		}
1274 		if (vspace() >= vlowat && force == 0) {
1275 			vnlruproc_sig = 0;
1276 			wakeup(&vnlruproc_sig);
1277 			msleep(vnlruproc, &vnode_free_list_mtx,
1278 			    PVFS|PDROP, "vlruwt", hz);
1279 			continue;
1280 		}
1281 		mtx_unlock(&vnode_free_list_mtx);
1282 		done = 0;
1283 		onumvnodes = numvnodes;
1284 		/*
1285 		 * Calculate parameters for recycling.  These are the same
1286 		 * throughout the loop to give some semblance of fairness.
1287 		 * The trigger point is to avoid recycling vnodes with lots
1288 		 * of resident pages.  We aren't trying to free memory; we
1289 		 * are trying to recycle or at least free vnodes.
1290 		 */
1291 		if (numvnodes <= desiredvnodes)
1292 			usevnodes = numvnodes - freevnodes;
1293 		else
1294 			usevnodes = numvnodes;
1295 		if (usevnodes <= 0)
1296 			usevnodes = 1;
1297 		/*
1298 		 * The trigger value is is chosen to give a conservatively
1299 		 * large value to ensure that it alone doesn't prevent
1300 		 * making progress.  The value can easily be so large that
1301 		 * it is effectively infinite in some congested and
1302 		 * misconfigured cases, and this is necessary.  Normally
1303 		 * it is about 8 to 100 (pages), which is quite large.
1304 		 */
1305 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1306 		if (force < 2)
1307 			trigger = vsmalltrigger;
1308 		reclaim_nc_src = force >= 3;
1309 		mtx_lock(&mountlist_mtx);
1310 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1311 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1312 				nmp = TAILQ_NEXT(mp, mnt_list);
1313 				continue;
1314 			}
1315 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1316 			mtx_lock(&mountlist_mtx);
1317 			nmp = TAILQ_NEXT(mp, mnt_list);
1318 			vfs_unbusy(mp);
1319 		}
1320 		mtx_unlock(&mountlist_mtx);
1321 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1322 			uma_reclaim(UMA_RECLAIM_DRAIN);
1323 		if (done == 0) {
1324 			if (force == 0 || force == 1) {
1325 				force = 2;
1326 				continue;
1327 			}
1328 			if (force == 2) {
1329 				force = 3;
1330 				continue;
1331 			}
1332 			force = 0;
1333 			vnlru_nowhere++;
1334 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1335 		} else
1336 			kern_yield(PRI_USER);
1337 		/*
1338 		 * After becoming active to expand above low water, keep
1339 		 * active until above high water.
1340 		 */
1341 		force = vspace() < vhiwat;
1342 	}
1343 }
1344 
1345 static struct kproc_desc vnlru_kp = {
1346 	"vnlru",
1347 	vnlru_proc,
1348 	&vnlruproc
1349 };
1350 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1351     &vnlru_kp);
1352 
1353 /*
1354  * Routines having to do with the management of the vnode table.
1355  */
1356 
1357 /*
1358  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1359  * before we actually vgone().  This function must be called with the vnode
1360  * held to prevent the vnode from being returned to the free list midway
1361  * through vgone().
1362  */
1363 static int
1364 vtryrecycle(struct vnode *vp)
1365 {
1366 	struct mount *vnmp;
1367 
1368 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1369 	VNASSERT(vp->v_holdcnt, vp,
1370 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1371 	/*
1372 	 * This vnode may found and locked via some other list, if so we
1373 	 * can't recycle it yet.
1374 	 */
1375 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1376 		CTR2(KTR_VFS,
1377 		    "%s: impossible to recycle, vp %p lock is already held",
1378 		    __func__, vp);
1379 		return (EWOULDBLOCK);
1380 	}
1381 	/*
1382 	 * Don't recycle if its filesystem is being suspended.
1383 	 */
1384 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1385 		VOP_UNLOCK(vp, 0);
1386 		CTR2(KTR_VFS,
1387 		    "%s: impossible to recycle, cannot start the write for %p",
1388 		    __func__, vp);
1389 		return (EBUSY);
1390 	}
1391 	/*
1392 	 * If we got this far, we need to acquire the interlock and see if
1393 	 * anyone picked up this vnode from another list.  If not, we will
1394 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1395 	 * will skip over it.
1396 	 */
1397 	VI_LOCK(vp);
1398 	if (vp->v_usecount) {
1399 		VOP_UNLOCK(vp, 0);
1400 		VI_UNLOCK(vp);
1401 		vn_finished_write(vnmp);
1402 		CTR2(KTR_VFS,
1403 		    "%s: impossible to recycle, %p is already referenced",
1404 		    __func__, vp);
1405 		return (EBUSY);
1406 	}
1407 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1408 		counter_u64_add(recycles_count, 1);
1409 		vgonel(vp);
1410 	}
1411 	VOP_UNLOCK(vp, 0);
1412 	VI_UNLOCK(vp);
1413 	vn_finished_write(vnmp);
1414 	return (0);
1415 }
1416 
1417 static void
1418 vcheckspace(void)
1419 {
1420 
1421 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1422 		vnlruproc_sig = 1;
1423 		wakeup(vnlruproc);
1424 	}
1425 }
1426 
1427 /*
1428  * Wait if necessary for space for a new vnode.
1429  */
1430 static int
1431 getnewvnode_wait(int suspended)
1432 {
1433 
1434 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1435 	if (numvnodes >= desiredvnodes) {
1436 		if (suspended) {
1437 			/*
1438 			 * The file system is being suspended.  We cannot
1439 			 * risk a deadlock here, so allow allocation of
1440 			 * another vnode even if this would give too many.
1441 			 */
1442 			return (0);
1443 		}
1444 		if (vnlruproc_sig == 0) {
1445 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1446 			wakeup(vnlruproc);
1447 		}
1448 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1449 		    "vlruwk", hz);
1450 	}
1451 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1452 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1453 		vnlru_free_locked(1, NULL);
1454 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1455 }
1456 
1457 /*
1458  * This hack is fragile, and probably not needed any more now that the
1459  * watermark handling works.
1460  */
1461 void
1462 getnewvnode_reserve(u_int count)
1463 {
1464 	struct thread *td;
1465 
1466 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1467 	/* XXX no longer so quick, but this part is not racy. */
1468 	mtx_lock(&vnode_free_list_mtx);
1469 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1470 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1471 		    freevnodes - wantfreevnodes), NULL);
1472 	mtx_unlock(&vnode_free_list_mtx);
1473 
1474 	td = curthread;
1475 	/* First try to be quick and racy. */
1476 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1477 		td->td_vp_reserv += count;
1478 		vcheckspace();	/* XXX no longer so quick, but more racy */
1479 		return;
1480 	} else
1481 		atomic_subtract_long(&numvnodes, count);
1482 
1483 	mtx_lock(&vnode_free_list_mtx);
1484 	while (count > 0) {
1485 		if (getnewvnode_wait(0) == 0) {
1486 			count--;
1487 			td->td_vp_reserv++;
1488 			atomic_add_long(&numvnodes, 1);
1489 		}
1490 	}
1491 	vcheckspace();
1492 	mtx_unlock(&vnode_free_list_mtx);
1493 }
1494 
1495 /*
1496  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1497  * misconfgured or changed significantly.  Reducing desiredvnodes below
1498  * the reserved amount should cause bizarre behaviour like reducing it
1499  * below the number of active vnodes -- the system will try to reduce
1500  * numvnodes to match, but should fail, so the subtraction below should
1501  * not overflow.
1502  */
1503 void
1504 getnewvnode_drop_reserve(void)
1505 {
1506 	struct thread *td;
1507 
1508 	td = curthread;
1509 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1510 	td->td_vp_reserv = 0;
1511 }
1512 
1513 /*
1514  * Return the next vnode from the free list.
1515  */
1516 int
1517 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1518     struct vnode **vpp)
1519 {
1520 	struct vnode *vp;
1521 	struct thread *td;
1522 	struct lock_object *lo;
1523 	static int cyclecount;
1524 	int error __unused;
1525 
1526 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1527 	vp = NULL;
1528 	td = curthread;
1529 	if (td->td_vp_reserv > 0) {
1530 		td->td_vp_reserv -= 1;
1531 		goto alloc;
1532 	}
1533 	mtx_lock(&vnode_free_list_mtx);
1534 	if (numvnodes < desiredvnodes)
1535 		cyclecount = 0;
1536 	else if (cyclecount++ >= freevnodes) {
1537 		cyclecount = 0;
1538 		vstir = 1;
1539 	}
1540 	/*
1541 	 * Grow the vnode cache if it will not be above its target max
1542 	 * after growing.  Otherwise, if the free list is nonempty, try
1543 	 * to reclaim 1 item from it before growing the cache (possibly
1544 	 * above its target max if the reclamation failed or is delayed).
1545 	 * Otherwise, wait for some space.  In all cases, schedule
1546 	 * vnlru_proc() if we are getting short of space.  The watermarks
1547 	 * should be chosen so that we never wait or even reclaim from
1548 	 * the free list to below its target minimum.
1549 	 */
1550 	if (numvnodes + 1 <= desiredvnodes)
1551 		;
1552 	else if (freevnodes > 0)
1553 		vnlru_free_locked(1, NULL);
1554 	else {
1555 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1556 		    MNTK_SUSPEND));
1557 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1558 		if (error != 0) {
1559 			mtx_unlock(&vnode_free_list_mtx);
1560 			return (error);
1561 		}
1562 #endif
1563 	}
1564 	vcheckspace();
1565 	atomic_add_long(&numvnodes, 1);
1566 	mtx_unlock(&vnode_free_list_mtx);
1567 alloc:
1568 	counter_u64_add(vnodes_created, 1);
1569 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1570 	/*
1571 	 * Locks are given the generic name "vnode" when created.
1572 	 * Follow the historic practice of using the filesystem
1573 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1574 	 *
1575 	 * Locks live in a witness group keyed on their name. Thus,
1576 	 * when a lock is renamed, it must also move from the witness
1577 	 * group of its old name to the witness group of its new name.
1578 	 *
1579 	 * The change only needs to be made when the vnode moves
1580 	 * from one filesystem type to another. We ensure that each
1581 	 * filesystem use a single static name pointer for its tag so
1582 	 * that we can compare pointers rather than doing a strcmp().
1583 	 */
1584 	lo = &vp->v_vnlock->lock_object;
1585 	if (lo->lo_name != tag) {
1586 		lo->lo_name = tag;
1587 		WITNESS_DESTROY(lo);
1588 		WITNESS_INIT(lo, tag);
1589 	}
1590 	/*
1591 	 * By default, don't allow shared locks unless filesystems opt-in.
1592 	 */
1593 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1594 	/*
1595 	 * Finalize various vnode identity bits.
1596 	 */
1597 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1598 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1599 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1600 	vp->v_type = VNON;
1601 	vp->v_tag = tag;
1602 	vp->v_op = vops;
1603 	v_init_counters(vp);
1604 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1605 #ifdef DIAGNOSTIC
1606 	if (mp == NULL && vops != &dead_vnodeops)
1607 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1608 #endif
1609 #ifdef MAC
1610 	mac_vnode_init(vp);
1611 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1612 		mac_vnode_associate_singlelabel(mp, vp);
1613 #endif
1614 	if (mp != NULL) {
1615 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1616 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1617 			vp->v_vflag |= VV_NOKNOTE;
1618 	}
1619 
1620 	/*
1621 	 * For the filesystems which do not use vfs_hash_insert(),
1622 	 * still initialize v_hash to have vfs_hash_index() useful.
1623 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1624 	 * its own hashing.
1625 	 */
1626 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1627 
1628 	*vpp = vp;
1629 	return (0);
1630 }
1631 
1632 /*
1633  * Delete from old mount point vnode list, if on one.
1634  */
1635 static void
1636 delmntque(struct vnode *vp)
1637 {
1638 	struct mount *mp;
1639 	int active;
1640 
1641 	mp = vp->v_mount;
1642 	if (mp == NULL)
1643 		return;
1644 	MNT_ILOCK(mp);
1645 	VI_LOCK(vp);
1646 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1647 	    ("Active vnode list size %d > Vnode list size %d",
1648 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1649 	active = vp->v_iflag & VI_ACTIVE;
1650 	vp->v_iflag &= ~VI_ACTIVE;
1651 	if (active) {
1652 		mtx_lock(&mp->mnt_listmtx);
1653 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1654 		mp->mnt_activevnodelistsize--;
1655 		mtx_unlock(&mp->mnt_listmtx);
1656 	}
1657 	vp->v_mount = NULL;
1658 	VI_UNLOCK(vp);
1659 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1660 		("bad mount point vnode list size"));
1661 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1662 	mp->mnt_nvnodelistsize--;
1663 	MNT_REL(mp);
1664 	MNT_IUNLOCK(mp);
1665 }
1666 
1667 static void
1668 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1669 {
1670 
1671 	vp->v_data = NULL;
1672 	vp->v_op = &dead_vnodeops;
1673 	vgone(vp);
1674 	vput(vp);
1675 }
1676 
1677 /*
1678  * Insert into list of vnodes for the new mount point, if available.
1679  */
1680 int
1681 insmntque1(struct vnode *vp, struct mount *mp,
1682 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1683 {
1684 
1685 	KASSERT(vp->v_mount == NULL,
1686 		("insmntque: vnode already on per mount vnode list"));
1687 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1688 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1689 
1690 	/*
1691 	 * We acquire the vnode interlock early to ensure that the
1692 	 * vnode cannot be recycled by another process releasing a
1693 	 * holdcnt on it before we get it on both the vnode list
1694 	 * and the active vnode list. The mount mutex protects only
1695 	 * manipulation of the vnode list and the vnode freelist
1696 	 * mutex protects only manipulation of the active vnode list.
1697 	 * Hence the need to hold the vnode interlock throughout.
1698 	 */
1699 	MNT_ILOCK(mp);
1700 	VI_LOCK(vp);
1701 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1702 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1703 	    mp->mnt_nvnodelistsize == 0)) &&
1704 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1705 		VI_UNLOCK(vp);
1706 		MNT_IUNLOCK(mp);
1707 		if (dtr != NULL)
1708 			dtr(vp, dtr_arg);
1709 		return (EBUSY);
1710 	}
1711 	vp->v_mount = mp;
1712 	MNT_REF(mp);
1713 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1714 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1715 		("neg mount point vnode list size"));
1716 	mp->mnt_nvnodelistsize++;
1717 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1718 	    ("Activating already active vnode"));
1719 	vp->v_iflag |= VI_ACTIVE;
1720 	mtx_lock(&mp->mnt_listmtx);
1721 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1722 	mp->mnt_activevnodelistsize++;
1723 	mtx_unlock(&mp->mnt_listmtx);
1724 	VI_UNLOCK(vp);
1725 	MNT_IUNLOCK(mp);
1726 	return (0);
1727 }
1728 
1729 int
1730 insmntque(struct vnode *vp, struct mount *mp)
1731 {
1732 
1733 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1734 }
1735 
1736 /*
1737  * Flush out and invalidate all buffers associated with a bufobj
1738  * Called with the underlying object locked.
1739  */
1740 int
1741 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1742 {
1743 	int error;
1744 
1745 	BO_LOCK(bo);
1746 	if (flags & V_SAVE) {
1747 		error = bufobj_wwait(bo, slpflag, slptimeo);
1748 		if (error) {
1749 			BO_UNLOCK(bo);
1750 			return (error);
1751 		}
1752 		if (bo->bo_dirty.bv_cnt > 0) {
1753 			BO_UNLOCK(bo);
1754 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1755 				return (error);
1756 			/*
1757 			 * XXX We could save a lock/unlock if this was only
1758 			 * enabled under INVARIANTS
1759 			 */
1760 			BO_LOCK(bo);
1761 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1762 				panic("vinvalbuf: dirty bufs");
1763 		}
1764 	}
1765 	/*
1766 	 * If you alter this loop please notice that interlock is dropped and
1767 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1768 	 * no race conditions occur from this.
1769 	 */
1770 	do {
1771 		error = flushbuflist(&bo->bo_clean,
1772 		    flags, bo, slpflag, slptimeo);
1773 		if (error == 0 && !(flags & V_CLEANONLY))
1774 			error = flushbuflist(&bo->bo_dirty,
1775 			    flags, bo, slpflag, slptimeo);
1776 		if (error != 0 && error != EAGAIN) {
1777 			BO_UNLOCK(bo);
1778 			return (error);
1779 		}
1780 	} while (error != 0);
1781 
1782 	/*
1783 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1784 	 * have write I/O in-progress but if there is a VM object then the
1785 	 * VM object can also have read-I/O in-progress.
1786 	 */
1787 	do {
1788 		bufobj_wwait(bo, 0, 0);
1789 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1790 			BO_UNLOCK(bo);
1791 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1792 			BO_LOCK(bo);
1793 		}
1794 	} while (bo->bo_numoutput > 0);
1795 	BO_UNLOCK(bo);
1796 
1797 	/*
1798 	 * Destroy the copy in the VM cache, too.
1799 	 */
1800 	if (bo->bo_object != NULL &&
1801 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1802 		VM_OBJECT_WLOCK(bo->bo_object);
1803 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1804 		    OBJPR_CLEANONLY : 0);
1805 		VM_OBJECT_WUNLOCK(bo->bo_object);
1806 	}
1807 
1808 #ifdef INVARIANTS
1809 	BO_LOCK(bo);
1810 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1811 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1812 	    bo->bo_clean.bv_cnt > 0))
1813 		panic("vinvalbuf: flush failed");
1814 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1815 	    bo->bo_dirty.bv_cnt > 0)
1816 		panic("vinvalbuf: flush dirty failed");
1817 	BO_UNLOCK(bo);
1818 #endif
1819 	return (0);
1820 }
1821 
1822 /*
1823  * Flush out and invalidate all buffers associated with a vnode.
1824  * Called with the underlying object locked.
1825  */
1826 int
1827 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1828 {
1829 
1830 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1831 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1832 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1833 		return (0);
1834 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1835 }
1836 
1837 /*
1838  * Flush out buffers on the specified list.
1839  *
1840  */
1841 static int
1842 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1843     int slptimeo)
1844 {
1845 	struct buf *bp, *nbp;
1846 	int retval, error;
1847 	daddr_t lblkno;
1848 	b_xflags_t xflags;
1849 
1850 	ASSERT_BO_WLOCKED(bo);
1851 
1852 	retval = 0;
1853 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1854 		/*
1855 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1856 		 * do not skip any buffers. If we are flushing only V_NORMAL
1857 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1858 		 * flushing only V_ALT buffers then skip buffers not marked
1859 		 * as BX_ALTDATA.
1860 		 */
1861 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1862 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1863 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1864 			continue;
1865 		}
1866 		if (nbp != NULL) {
1867 			lblkno = nbp->b_lblkno;
1868 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1869 		}
1870 		retval = EAGAIN;
1871 		error = BUF_TIMELOCK(bp,
1872 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1873 		    "flushbuf", slpflag, slptimeo);
1874 		if (error) {
1875 			BO_LOCK(bo);
1876 			return (error != ENOLCK ? error : EAGAIN);
1877 		}
1878 		KASSERT(bp->b_bufobj == bo,
1879 		    ("bp %p wrong b_bufobj %p should be %p",
1880 		    bp, bp->b_bufobj, bo));
1881 		/*
1882 		 * XXX Since there are no node locks for NFS, I
1883 		 * believe there is a slight chance that a delayed
1884 		 * write will occur while sleeping just above, so
1885 		 * check for it.
1886 		 */
1887 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1888 		    (flags & V_SAVE)) {
1889 			bremfree(bp);
1890 			bp->b_flags |= B_ASYNC;
1891 			bwrite(bp);
1892 			BO_LOCK(bo);
1893 			return (EAGAIN);	/* XXX: why not loop ? */
1894 		}
1895 		bremfree(bp);
1896 		bp->b_flags |= (B_INVAL | B_RELBUF);
1897 		bp->b_flags &= ~B_ASYNC;
1898 		brelse(bp);
1899 		BO_LOCK(bo);
1900 		if (nbp == NULL)
1901 			break;
1902 		nbp = gbincore(bo, lblkno);
1903 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1904 		    != xflags)
1905 			break;			/* nbp invalid */
1906 	}
1907 	return (retval);
1908 }
1909 
1910 int
1911 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1912 {
1913 	struct buf *bp;
1914 	int error;
1915 	daddr_t lblkno;
1916 
1917 	ASSERT_BO_LOCKED(bo);
1918 
1919 	for (lblkno = startn;;) {
1920 again:
1921 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1922 		if (bp == NULL || bp->b_lblkno >= endn ||
1923 		    bp->b_lblkno < startn)
1924 			break;
1925 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1926 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1927 		if (error != 0) {
1928 			BO_RLOCK(bo);
1929 			if (error == ENOLCK)
1930 				goto again;
1931 			return (error);
1932 		}
1933 		KASSERT(bp->b_bufobj == bo,
1934 		    ("bp %p wrong b_bufobj %p should be %p",
1935 		    bp, bp->b_bufobj, bo));
1936 		lblkno = bp->b_lblkno + 1;
1937 		if ((bp->b_flags & B_MANAGED) == 0)
1938 			bremfree(bp);
1939 		bp->b_flags |= B_RELBUF;
1940 		/*
1941 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1942 		 * pages backing each buffer in the range are unlikely to be
1943 		 * reused.  Dirty buffers will have the hint applied once
1944 		 * they've been written.
1945 		 */
1946 		if ((bp->b_flags & B_VMIO) != 0)
1947 			bp->b_flags |= B_NOREUSE;
1948 		brelse(bp);
1949 		BO_RLOCK(bo);
1950 	}
1951 	return (0);
1952 }
1953 
1954 /*
1955  * Truncate a file's buffer and pages to a specified length.  This
1956  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1957  * sync activity.
1958  */
1959 int
1960 vtruncbuf(struct vnode *vp, off_t length, int blksize)
1961 {
1962 	struct buf *bp, *nbp;
1963 	struct bufobj *bo;
1964 	daddr_t startlbn;
1965 
1966 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
1967 	    vp, blksize, (uintmax_t)length);
1968 
1969 	/*
1970 	 * Round up to the *next* lbn.
1971 	 */
1972 	startlbn = howmany(length, blksize);
1973 
1974 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1975 
1976 	bo = &vp->v_bufobj;
1977 restart_unlocked:
1978 	BO_LOCK(bo);
1979 
1980 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
1981 		;
1982 
1983 	if (length > 0) {
1984 restartsync:
1985 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1986 			if (bp->b_lblkno > 0)
1987 				continue;
1988 			/*
1989 			 * Since we hold the vnode lock this should only
1990 			 * fail if we're racing with the buf daemon.
1991 			 */
1992 			if (BUF_LOCK(bp,
1993 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1994 			    BO_LOCKPTR(bo)) == ENOLCK)
1995 				goto restart_unlocked;
1996 
1997 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1998 			    ("buf(%p) on dirty queue without DELWRI", bp));
1999 
2000 			bremfree(bp);
2001 			bawrite(bp);
2002 			BO_LOCK(bo);
2003 			goto restartsync;
2004 		}
2005 	}
2006 
2007 	bufobj_wwait(bo, 0, 0);
2008 	BO_UNLOCK(bo);
2009 	vnode_pager_setsize(vp, length);
2010 
2011 	return (0);
2012 }
2013 
2014 /*
2015  * Invalidate the cached pages of a file's buffer within the range of block
2016  * numbers [startlbn, endlbn).
2017  */
2018 void
2019 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2020     int blksize)
2021 {
2022 	struct bufobj *bo;
2023 	off_t start, end;
2024 
2025 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2026 
2027 	start = blksize * startlbn;
2028 	end = blksize * endlbn;
2029 
2030 	bo = &vp->v_bufobj;
2031 	BO_LOCK(bo);
2032 	MPASS(blksize == bo->bo_bsize);
2033 
2034 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2035 		;
2036 
2037 	BO_UNLOCK(bo);
2038 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2039 }
2040 
2041 static int
2042 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2043     daddr_t startlbn, daddr_t endlbn)
2044 {
2045 	struct buf *bp, *nbp;
2046 	bool anyfreed;
2047 
2048 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2049 	ASSERT_BO_LOCKED(bo);
2050 
2051 	do {
2052 		anyfreed = false;
2053 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2054 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2055 				continue;
2056 			if (BUF_LOCK(bp,
2057 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2058 			    BO_LOCKPTR(bo)) == ENOLCK) {
2059 				BO_LOCK(bo);
2060 				return (EAGAIN);
2061 			}
2062 
2063 			bremfree(bp);
2064 			bp->b_flags |= B_INVAL | B_RELBUF;
2065 			bp->b_flags &= ~B_ASYNC;
2066 			brelse(bp);
2067 			anyfreed = true;
2068 
2069 			BO_LOCK(bo);
2070 			if (nbp != NULL &&
2071 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2072 			    nbp->b_vp != vp ||
2073 			    (nbp->b_flags & B_DELWRI) != 0))
2074 				return (EAGAIN);
2075 		}
2076 
2077 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2078 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2079 				continue;
2080 			if (BUF_LOCK(bp,
2081 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2082 			    BO_LOCKPTR(bo)) == ENOLCK) {
2083 				BO_LOCK(bo);
2084 				return (EAGAIN);
2085 			}
2086 			bremfree(bp);
2087 			bp->b_flags |= B_INVAL | B_RELBUF;
2088 			bp->b_flags &= ~B_ASYNC;
2089 			brelse(bp);
2090 			anyfreed = true;
2091 
2092 			BO_LOCK(bo);
2093 			if (nbp != NULL &&
2094 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2095 			    (nbp->b_vp != vp) ||
2096 			    (nbp->b_flags & B_DELWRI) == 0))
2097 				return (EAGAIN);
2098 		}
2099 	} while (anyfreed);
2100 	return (0);
2101 }
2102 
2103 static void
2104 buf_vlist_remove(struct buf *bp)
2105 {
2106 	struct bufv *bv;
2107 
2108 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2109 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2110 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2111 	    (BX_VNDIRTY|BX_VNCLEAN),
2112 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2113 	if (bp->b_xflags & BX_VNDIRTY)
2114 		bv = &bp->b_bufobj->bo_dirty;
2115 	else
2116 		bv = &bp->b_bufobj->bo_clean;
2117 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2118 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2119 	bv->bv_cnt--;
2120 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2121 }
2122 
2123 /*
2124  * Add the buffer to the sorted clean or dirty block list.
2125  *
2126  * NOTE: xflags is passed as a constant, optimizing this inline function!
2127  */
2128 static void
2129 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2130 {
2131 	struct bufv *bv;
2132 	struct buf *n;
2133 	int error;
2134 
2135 	ASSERT_BO_WLOCKED(bo);
2136 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2137 	    ("dead bo %p", bo));
2138 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2139 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2140 	bp->b_xflags |= xflags;
2141 	if (xflags & BX_VNDIRTY)
2142 		bv = &bo->bo_dirty;
2143 	else
2144 		bv = &bo->bo_clean;
2145 
2146 	/*
2147 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2148 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2149 	 * than _ge.
2150 	 */
2151 	if (bv->bv_cnt == 0 ||
2152 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2153 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2154 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2155 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2156 	else
2157 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2158 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2159 	if (error)
2160 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2161 	bv->bv_cnt++;
2162 }
2163 
2164 /*
2165  * Look up a buffer using the buffer tries.
2166  */
2167 struct buf *
2168 gbincore(struct bufobj *bo, daddr_t lblkno)
2169 {
2170 	struct buf *bp;
2171 
2172 	ASSERT_BO_LOCKED(bo);
2173 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2174 	if (bp != NULL)
2175 		return (bp);
2176 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2177 }
2178 
2179 /*
2180  * Associate a buffer with a vnode.
2181  */
2182 void
2183 bgetvp(struct vnode *vp, struct buf *bp)
2184 {
2185 	struct bufobj *bo;
2186 
2187 	bo = &vp->v_bufobj;
2188 	ASSERT_BO_WLOCKED(bo);
2189 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2190 
2191 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2192 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2193 	    ("bgetvp: bp already attached! %p", bp));
2194 
2195 	vhold(vp);
2196 	bp->b_vp = vp;
2197 	bp->b_bufobj = bo;
2198 	/*
2199 	 * Insert onto list for new vnode.
2200 	 */
2201 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2202 }
2203 
2204 /*
2205  * Disassociate a buffer from a vnode.
2206  */
2207 void
2208 brelvp(struct buf *bp)
2209 {
2210 	struct bufobj *bo;
2211 	struct vnode *vp;
2212 
2213 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2214 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2215 
2216 	/*
2217 	 * Delete from old vnode list, if on one.
2218 	 */
2219 	vp = bp->b_vp;		/* XXX */
2220 	bo = bp->b_bufobj;
2221 	BO_LOCK(bo);
2222 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2223 		buf_vlist_remove(bp);
2224 	else
2225 		panic("brelvp: Buffer %p not on queue.", bp);
2226 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2227 		bo->bo_flag &= ~BO_ONWORKLST;
2228 		mtx_lock(&sync_mtx);
2229 		LIST_REMOVE(bo, bo_synclist);
2230 		syncer_worklist_len--;
2231 		mtx_unlock(&sync_mtx);
2232 	}
2233 	bp->b_vp = NULL;
2234 	bp->b_bufobj = NULL;
2235 	BO_UNLOCK(bo);
2236 	vdrop(vp);
2237 }
2238 
2239 /*
2240  * Add an item to the syncer work queue.
2241  */
2242 static void
2243 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2244 {
2245 	int slot;
2246 
2247 	ASSERT_BO_WLOCKED(bo);
2248 
2249 	mtx_lock(&sync_mtx);
2250 	if (bo->bo_flag & BO_ONWORKLST)
2251 		LIST_REMOVE(bo, bo_synclist);
2252 	else {
2253 		bo->bo_flag |= BO_ONWORKLST;
2254 		syncer_worklist_len++;
2255 	}
2256 
2257 	if (delay > syncer_maxdelay - 2)
2258 		delay = syncer_maxdelay - 2;
2259 	slot = (syncer_delayno + delay) & syncer_mask;
2260 
2261 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2262 	mtx_unlock(&sync_mtx);
2263 }
2264 
2265 static int
2266 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2267 {
2268 	int error, len;
2269 
2270 	mtx_lock(&sync_mtx);
2271 	len = syncer_worklist_len - sync_vnode_count;
2272 	mtx_unlock(&sync_mtx);
2273 	error = SYSCTL_OUT(req, &len, sizeof(len));
2274 	return (error);
2275 }
2276 
2277 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2278     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2279 
2280 static struct proc *updateproc;
2281 static void sched_sync(void);
2282 static struct kproc_desc up_kp = {
2283 	"syncer",
2284 	sched_sync,
2285 	&updateproc
2286 };
2287 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2288 
2289 static int
2290 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2291 {
2292 	struct vnode *vp;
2293 	struct mount *mp;
2294 
2295 	*bo = LIST_FIRST(slp);
2296 	if (*bo == NULL)
2297 		return (0);
2298 	vp = bo2vnode(*bo);
2299 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2300 		return (1);
2301 	/*
2302 	 * We use vhold in case the vnode does not
2303 	 * successfully sync.  vhold prevents the vnode from
2304 	 * going away when we unlock the sync_mtx so that
2305 	 * we can acquire the vnode interlock.
2306 	 */
2307 	vholdl(vp);
2308 	mtx_unlock(&sync_mtx);
2309 	VI_UNLOCK(vp);
2310 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2311 		vdrop(vp);
2312 		mtx_lock(&sync_mtx);
2313 		return (*bo == LIST_FIRST(slp));
2314 	}
2315 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2316 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2317 	VOP_UNLOCK(vp, 0);
2318 	vn_finished_write(mp);
2319 	BO_LOCK(*bo);
2320 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2321 		/*
2322 		 * Put us back on the worklist.  The worklist
2323 		 * routine will remove us from our current
2324 		 * position and then add us back in at a later
2325 		 * position.
2326 		 */
2327 		vn_syncer_add_to_worklist(*bo, syncdelay);
2328 	}
2329 	BO_UNLOCK(*bo);
2330 	vdrop(vp);
2331 	mtx_lock(&sync_mtx);
2332 	return (0);
2333 }
2334 
2335 static int first_printf = 1;
2336 
2337 /*
2338  * System filesystem synchronizer daemon.
2339  */
2340 static void
2341 sched_sync(void)
2342 {
2343 	struct synclist *next, *slp;
2344 	struct bufobj *bo;
2345 	long starttime;
2346 	struct thread *td = curthread;
2347 	int last_work_seen;
2348 	int net_worklist_len;
2349 	int syncer_final_iter;
2350 	int error;
2351 
2352 	last_work_seen = 0;
2353 	syncer_final_iter = 0;
2354 	syncer_state = SYNCER_RUNNING;
2355 	starttime = time_uptime;
2356 	td->td_pflags |= TDP_NORUNNINGBUF;
2357 
2358 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2359 	    SHUTDOWN_PRI_LAST);
2360 
2361 	mtx_lock(&sync_mtx);
2362 	for (;;) {
2363 		if (syncer_state == SYNCER_FINAL_DELAY &&
2364 		    syncer_final_iter == 0) {
2365 			mtx_unlock(&sync_mtx);
2366 			kproc_suspend_check(td->td_proc);
2367 			mtx_lock(&sync_mtx);
2368 		}
2369 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2370 		if (syncer_state != SYNCER_RUNNING &&
2371 		    starttime != time_uptime) {
2372 			if (first_printf) {
2373 				printf("\nSyncing disks, vnodes remaining... ");
2374 				first_printf = 0;
2375 			}
2376 			printf("%d ", net_worklist_len);
2377 		}
2378 		starttime = time_uptime;
2379 
2380 		/*
2381 		 * Push files whose dirty time has expired.  Be careful
2382 		 * of interrupt race on slp queue.
2383 		 *
2384 		 * Skip over empty worklist slots when shutting down.
2385 		 */
2386 		do {
2387 			slp = &syncer_workitem_pending[syncer_delayno];
2388 			syncer_delayno += 1;
2389 			if (syncer_delayno == syncer_maxdelay)
2390 				syncer_delayno = 0;
2391 			next = &syncer_workitem_pending[syncer_delayno];
2392 			/*
2393 			 * If the worklist has wrapped since the
2394 			 * it was emptied of all but syncer vnodes,
2395 			 * switch to the FINAL_DELAY state and run
2396 			 * for one more second.
2397 			 */
2398 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2399 			    net_worklist_len == 0 &&
2400 			    last_work_seen == syncer_delayno) {
2401 				syncer_state = SYNCER_FINAL_DELAY;
2402 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2403 			}
2404 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2405 		    syncer_worklist_len > 0);
2406 
2407 		/*
2408 		 * Keep track of the last time there was anything
2409 		 * on the worklist other than syncer vnodes.
2410 		 * Return to the SHUTTING_DOWN state if any
2411 		 * new work appears.
2412 		 */
2413 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2414 			last_work_seen = syncer_delayno;
2415 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2416 			syncer_state = SYNCER_SHUTTING_DOWN;
2417 		while (!LIST_EMPTY(slp)) {
2418 			error = sync_vnode(slp, &bo, td);
2419 			if (error == 1) {
2420 				LIST_REMOVE(bo, bo_synclist);
2421 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2422 				continue;
2423 			}
2424 
2425 			if (first_printf == 0) {
2426 				/*
2427 				 * Drop the sync mutex, because some watchdog
2428 				 * drivers need to sleep while patting
2429 				 */
2430 				mtx_unlock(&sync_mtx);
2431 				wdog_kern_pat(WD_LASTVAL);
2432 				mtx_lock(&sync_mtx);
2433 			}
2434 
2435 		}
2436 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2437 			syncer_final_iter--;
2438 		/*
2439 		 * The variable rushjob allows the kernel to speed up the
2440 		 * processing of the filesystem syncer process. A rushjob
2441 		 * value of N tells the filesystem syncer to process the next
2442 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2443 		 * is used by the soft update code to speed up the filesystem
2444 		 * syncer process when the incore state is getting so far
2445 		 * ahead of the disk that the kernel memory pool is being
2446 		 * threatened with exhaustion.
2447 		 */
2448 		if (rushjob > 0) {
2449 			rushjob -= 1;
2450 			continue;
2451 		}
2452 		/*
2453 		 * Just sleep for a short period of time between
2454 		 * iterations when shutting down to allow some I/O
2455 		 * to happen.
2456 		 *
2457 		 * If it has taken us less than a second to process the
2458 		 * current work, then wait. Otherwise start right over
2459 		 * again. We can still lose time if any single round
2460 		 * takes more than two seconds, but it does not really
2461 		 * matter as we are just trying to generally pace the
2462 		 * filesystem activity.
2463 		 */
2464 		if (syncer_state != SYNCER_RUNNING ||
2465 		    time_uptime == starttime) {
2466 			thread_lock(td);
2467 			sched_prio(td, PPAUSE);
2468 			thread_unlock(td);
2469 		}
2470 		if (syncer_state != SYNCER_RUNNING)
2471 			cv_timedwait(&sync_wakeup, &sync_mtx,
2472 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2473 		else if (time_uptime == starttime)
2474 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2475 	}
2476 }
2477 
2478 /*
2479  * Request the syncer daemon to speed up its work.
2480  * We never push it to speed up more than half of its
2481  * normal turn time, otherwise it could take over the cpu.
2482  */
2483 int
2484 speedup_syncer(void)
2485 {
2486 	int ret = 0;
2487 
2488 	mtx_lock(&sync_mtx);
2489 	if (rushjob < syncdelay / 2) {
2490 		rushjob += 1;
2491 		stat_rush_requests += 1;
2492 		ret = 1;
2493 	}
2494 	mtx_unlock(&sync_mtx);
2495 	cv_broadcast(&sync_wakeup);
2496 	return (ret);
2497 }
2498 
2499 /*
2500  * Tell the syncer to speed up its work and run though its work
2501  * list several times, then tell it to shut down.
2502  */
2503 static void
2504 syncer_shutdown(void *arg, int howto)
2505 {
2506 
2507 	if (howto & RB_NOSYNC)
2508 		return;
2509 	mtx_lock(&sync_mtx);
2510 	syncer_state = SYNCER_SHUTTING_DOWN;
2511 	rushjob = 0;
2512 	mtx_unlock(&sync_mtx);
2513 	cv_broadcast(&sync_wakeup);
2514 	kproc_shutdown(arg, howto);
2515 }
2516 
2517 void
2518 syncer_suspend(void)
2519 {
2520 
2521 	syncer_shutdown(updateproc, 0);
2522 }
2523 
2524 void
2525 syncer_resume(void)
2526 {
2527 
2528 	mtx_lock(&sync_mtx);
2529 	first_printf = 1;
2530 	syncer_state = SYNCER_RUNNING;
2531 	mtx_unlock(&sync_mtx);
2532 	cv_broadcast(&sync_wakeup);
2533 	kproc_resume(updateproc);
2534 }
2535 
2536 /*
2537  * Reassign a buffer from one vnode to another.
2538  * Used to assign file specific control information
2539  * (indirect blocks) to the vnode to which they belong.
2540  */
2541 void
2542 reassignbuf(struct buf *bp)
2543 {
2544 	struct vnode *vp;
2545 	struct bufobj *bo;
2546 	int delay;
2547 #ifdef INVARIANTS
2548 	struct bufv *bv;
2549 #endif
2550 
2551 	vp = bp->b_vp;
2552 	bo = bp->b_bufobj;
2553 	++reassignbufcalls;
2554 
2555 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2556 	    bp, bp->b_vp, bp->b_flags);
2557 	/*
2558 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2559 	 * is not fully linked in.
2560 	 */
2561 	if (bp->b_flags & B_PAGING)
2562 		panic("cannot reassign paging buffer");
2563 
2564 	/*
2565 	 * Delete from old vnode list, if on one.
2566 	 */
2567 	BO_LOCK(bo);
2568 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2569 		buf_vlist_remove(bp);
2570 	else
2571 		panic("reassignbuf: Buffer %p not on queue.", bp);
2572 	/*
2573 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2574 	 * of clean buffers.
2575 	 */
2576 	if (bp->b_flags & B_DELWRI) {
2577 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2578 			switch (vp->v_type) {
2579 			case VDIR:
2580 				delay = dirdelay;
2581 				break;
2582 			case VCHR:
2583 				delay = metadelay;
2584 				break;
2585 			default:
2586 				delay = filedelay;
2587 			}
2588 			vn_syncer_add_to_worklist(bo, delay);
2589 		}
2590 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2591 	} else {
2592 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2593 
2594 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2595 			mtx_lock(&sync_mtx);
2596 			LIST_REMOVE(bo, bo_synclist);
2597 			syncer_worklist_len--;
2598 			mtx_unlock(&sync_mtx);
2599 			bo->bo_flag &= ~BO_ONWORKLST;
2600 		}
2601 	}
2602 #ifdef INVARIANTS
2603 	bv = &bo->bo_clean;
2604 	bp = TAILQ_FIRST(&bv->bv_hd);
2605 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2606 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2607 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2608 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2609 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2610 	bv = &bo->bo_dirty;
2611 	bp = TAILQ_FIRST(&bv->bv_hd);
2612 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2613 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2614 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2615 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2616 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2617 #endif
2618 	BO_UNLOCK(bo);
2619 }
2620 
2621 static void
2622 v_init_counters(struct vnode *vp)
2623 {
2624 
2625 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2626 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2627 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2628 
2629 	refcount_init(&vp->v_holdcnt, 1);
2630 	refcount_init(&vp->v_usecount, 1);
2631 }
2632 
2633 /*
2634  * Increment si_usecount of the associated device, if any.
2635  */
2636 static void
2637 v_incr_devcount(struct vnode *vp)
2638 {
2639 
2640 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2641 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2642 		dev_lock();
2643 		vp->v_rdev->si_usecount++;
2644 		dev_unlock();
2645 	}
2646 }
2647 
2648 /*
2649  * Decrement si_usecount of the associated device, if any.
2650  */
2651 static void
2652 v_decr_devcount(struct vnode *vp)
2653 {
2654 
2655 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2656 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2657 		dev_lock();
2658 		vp->v_rdev->si_usecount--;
2659 		dev_unlock();
2660 	}
2661 }
2662 
2663 /*
2664  * Grab a particular vnode from the free list, increment its
2665  * reference count and lock it.  VI_DOOMED is set if the vnode
2666  * is being destroyed.  Only callers who specify LK_RETRY will
2667  * see doomed vnodes.  If inactive processing was delayed in
2668  * vput try to do it here.
2669  *
2670  * Both holdcnt and usecount can be manipulated using atomics without holding
2671  * any locks except in these cases which require the vnode interlock:
2672  * holdcnt: 1->0 and 0->1
2673  * usecount: 0->1
2674  *
2675  * usecount is permitted to transition 1->0 without the interlock because
2676  * vnode is kept live by holdcnt.
2677  */
2678 static enum vgetstate
2679 _vget_prep(struct vnode *vp, bool interlock)
2680 {
2681 	enum vgetstate vs;
2682 
2683 	if (__predict_true(vp->v_type != VCHR)) {
2684 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2685 			vs = VGET_USECOUNT;
2686 		} else {
2687 			_vhold(vp, interlock);
2688 			vs = VGET_HOLDCNT;
2689 		}
2690 	} else {
2691 		if (!interlock)
2692 			VI_LOCK(vp);
2693 		if (vp->v_usecount == 0) {
2694 			vholdl(vp);
2695 			vs = VGET_HOLDCNT;
2696 		} else {
2697 			v_incr_devcount(vp);
2698 			refcount_acquire(&vp->v_usecount);
2699 			vs = VGET_USECOUNT;
2700 		}
2701 		if (!interlock)
2702 			VI_UNLOCK(vp);
2703 	}
2704 	return (vs);
2705 }
2706 
2707 enum vgetstate
2708 vget_prep(struct vnode *vp)
2709 {
2710 
2711 	return (_vget_prep(vp, false));
2712 }
2713 
2714 int
2715 vget(struct vnode *vp, int flags, struct thread *td)
2716 {
2717 	enum vgetstate vs;
2718 
2719 	MPASS(td == curthread);
2720 
2721 	vs = _vget_prep(vp, (flags & LK_INTERLOCK) != 0);
2722 	return (vget_finish(vp, flags, vs));
2723 }
2724 
2725 int
2726 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2727 {
2728 	int error, oweinact;
2729 
2730 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2731 	    ("%s: invalid lock operation", __func__));
2732 
2733 	if ((flags & LK_INTERLOCK) != 0)
2734 		ASSERT_VI_LOCKED(vp, __func__);
2735 	else
2736 		ASSERT_VI_UNLOCKED(vp, __func__);
2737 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2738 	if (vs == VGET_USECOUNT) {
2739 		VNASSERT(vp->v_usecount > 0, vp,
2740 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2741 		    __func__));
2742 	}
2743 
2744 	if ((error = vn_lock(vp, flags)) != 0) {
2745 		if (vs == VGET_USECOUNT)
2746 			vrele(vp);
2747 		else
2748 			vdrop(vp);
2749 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2750 		    vp);
2751 		return (error);
2752 	}
2753 
2754 	if (vs == VGET_USECOUNT) {
2755 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2756 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2757 		return (0);
2758 	}
2759 
2760 	/*
2761 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2762 	 * the vnode around. Otherwise someone else lended their hold count and
2763 	 * we have to drop ours.
2764 	 */
2765 	if (vp->v_type != VCHR &&
2766 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2767 #ifdef INVARIANTS
2768 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1) - 1;
2769 		VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2770 #else
2771 		refcount_release(&vp->v_holdcnt);
2772 #endif
2773 		VNODE_REFCOUNT_FENCE_ACQ();
2774 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2775 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2776 		return (0);
2777 	}
2778 
2779 	/*
2780 	 * We don't guarantee that any particular close will
2781 	 * trigger inactive processing so just make a best effort
2782 	 * here at preventing a reference to a removed file.  If
2783 	 * we don't succeed no harm is done.
2784 	 *
2785 	 * Upgrade our holdcnt to a usecount.
2786 	 */
2787 	VI_LOCK(vp);
2788 	/*
2789 	 * See the previous section. By the time we get here we may find
2790 	 * ourselves in the same spot.
2791 	 */
2792 	if (vp->v_type != VCHR) {
2793 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2794 #ifdef INVARIANTS
2795 			int old = atomic_fetchadd_int(&vp->v_holdcnt, -1) - 1;
2796 			VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2797 #else
2798 			refcount_release(&vp->v_holdcnt);
2799 #endif
2800 			VNODE_REFCOUNT_FENCE_ACQ();
2801 			VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2802 			    ("%s: vnode with usecount and VI_OWEINACT set",
2803 			    __func__));
2804 			VI_UNLOCK(vp);
2805 			return (0);
2806 		}
2807 	} else {
2808 		if (vp->v_usecount > 0)
2809 			refcount_release(&vp->v_holdcnt);
2810 	}
2811 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
2812 		oweinact = 0;
2813 	} else {
2814 		oweinact = 1;
2815 		vp->v_iflag &= ~VI_OWEINACT;
2816 		VNODE_REFCOUNT_FENCE_REL();
2817 	}
2818 	v_incr_devcount(vp);
2819 	refcount_acquire(&vp->v_usecount);
2820 	if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2821 	    (flags & LK_NOWAIT) == 0)
2822 		vinactive(vp, curthread);
2823 	VI_UNLOCK(vp);
2824 	return (0);
2825 }
2826 
2827 /*
2828  * Increase the reference (use) and hold count of a vnode.
2829  * This will also remove the vnode from the free list if it is presently free.
2830  */
2831 void
2832 vref(struct vnode *vp)
2833 {
2834 
2835 	ASSERT_VI_UNLOCKED(vp, __func__);
2836 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2837 	if (vp->v_type != VCHR &&
2838 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2839 		VNODE_REFCOUNT_FENCE_ACQ();
2840 		VNASSERT(vp->v_holdcnt > 0, vp,
2841 		    ("%s: active vnode not held", __func__));
2842 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2843 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2844 		return;
2845 	}
2846 	VI_LOCK(vp);
2847 	vrefl(vp);
2848 	VI_UNLOCK(vp);
2849 }
2850 
2851 void
2852 vrefl(struct vnode *vp)
2853 {
2854 
2855 	ASSERT_VI_LOCKED(vp, __func__);
2856 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2857 	if (vp->v_type != VCHR &&
2858 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2859 		VNODE_REFCOUNT_FENCE_ACQ();
2860 		VNASSERT(vp->v_holdcnt > 0, vp,
2861 		    ("%s: active vnode not held", __func__));
2862 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2863 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2864 		return;
2865 	}
2866 	if (vp->v_usecount == 0)
2867 		vholdl(vp);
2868 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2869 		vp->v_iflag &= ~VI_OWEINACT;
2870 		VNODE_REFCOUNT_FENCE_REL();
2871 	}
2872 	v_incr_devcount(vp);
2873 	refcount_acquire(&vp->v_usecount);
2874 }
2875 
2876 void
2877 vrefact(struct vnode *vp)
2878 {
2879 
2880 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2881 	if (__predict_false(vp->v_type == VCHR)) {
2882 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2883 		    ("%s: wrong ref counts", __func__));
2884 		vref(vp);
2885 		return;
2886 	}
2887 #ifdef INVARIANTS
2888 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2889 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2890 #else
2891 	refcount_acquire(&vp->v_usecount);
2892 #endif
2893 }
2894 
2895 /*
2896  * Return reference count of a vnode.
2897  *
2898  * The results of this call are only guaranteed when some mechanism is used to
2899  * stop other processes from gaining references to the vnode.  This may be the
2900  * case if the caller holds the only reference.  This is also useful when stale
2901  * data is acceptable as race conditions may be accounted for by some other
2902  * means.
2903  */
2904 int
2905 vrefcnt(struct vnode *vp)
2906 {
2907 
2908 	return (vp->v_usecount);
2909 }
2910 
2911 #define	VPUTX_VRELE	1
2912 #define	VPUTX_VPUT	2
2913 #define	VPUTX_VUNREF	3
2914 
2915 /*
2916  * Decrement the use and hold counts for a vnode.
2917  *
2918  * See an explanation near vget() as to why atomic operation is safe.
2919  */
2920 static void
2921 vputx(struct vnode *vp, int func)
2922 {
2923 	int error;
2924 
2925 	KASSERT(vp != NULL, ("vputx: null vp"));
2926 	if (func == VPUTX_VUNREF)
2927 		ASSERT_VOP_LOCKED(vp, "vunref");
2928 	else if (func == VPUTX_VPUT)
2929 		ASSERT_VOP_LOCKED(vp, "vput");
2930 	else
2931 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2932 	ASSERT_VI_UNLOCKED(vp, __func__);
2933 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2934 	    ("%s: wrong ref counts", __func__));
2935 
2936 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2937 
2938 	/*
2939 	 * It is an invariant that all VOP_* calls operate on a held vnode.
2940 	 * We may be only having an implicit hold stemming from our usecount,
2941 	 * which we are about to release. If we unlock the vnode afterwards we
2942 	 * open a time window where someone else dropped the last usecount and
2943 	 * proceeded to free the vnode before our unlock finished. For this
2944 	 * reason we unlock the vnode early. This is a little bit wasteful as
2945 	 * it may be the vnode is exclusively locked and inactive processing is
2946 	 * needed, in which case we are adding work.
2947 	 */
2948 	if (func == VPUTX_VPUT)
2949 		VOP_UNLOCK(vp, 0);
2950 
2951 	/*
2952 	 * We want to hold the vnode until the inactive finishes to
2953 	 * prevent vgone() races.  We drop the use count here and the
2954 	 * hold count below when we're done.
2955 	 */
2956 	if (vp->v_type != VCHR) {
2957 		/*
2958 		 * If we release the last usecount we take ownership of the hold
2959 		 * count which provides liveness of the vnode, in which case we
2960 		 * have to vdrop.
2961 		 */
2962 		if (!refcount_release(&vp->v_usecount))
2963 			return;
2964 		VI_LOCK(vp);
2965 		/*
2966 		 * By the time we got here someone else might have transitioned
2967 		 * the count back to > 0.
2968 		 */
2969 		if (vp->v_usecount > 0) {
2970 			vdropl(vp);
2971 			return;
2972 		}
2973 	} else {
2974 		VI_LOCK(vp);
2975 		v_decr_devcount(vp);
2976 		if (!refcount_release(&vp->v_usecount)) {
2977 			VI_UNLOCK(vp);
2978 			return;
2979 		}
2980 	}
2981 	if (vp->v_iflag & VI_DOINGINACT) {
2982 		vdropl(vp);
2983 		return;
2984 	}
2985 
2986 	error = 0;
2987 
2988 	if (vp->v_usecount != 0) {
2989 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2990 		panic("vputx: usecount not zero");
2991 	}
2992 
2993 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2994 
2995 	/*
2996 	 * Check if the fs wants to perform inactive processing. Note we
2997 	 * may be only holding the interlock, in which case it is possible
2998 	 * someone else called vgone on the vnode and ->v_data is now NULL.
2999 	 * Since vgone performs inactive on its own there is nothing to do
3000 	 * here but to drop our hold count.
3001 	 */
3002 	if (__predict_false(vp->v_iflag & VI_DOOMED) ||
3003 	    VOP_NEED_INACTIVE(vp) == 0) {
3004 		vdropl(vp);
3005 		return;
3006 	}
3007 
3008 	/*
3009 	 * We must call VOP_INACTIVE with the node locked. Mark
3010 	 * as VI_DOINGINACT to avoid recursion.
3011 	 */
3012 	vp->v_iflag |= VI_OWEINACT;
3013 	switch (func) {
3014 	case VPUTX_VRELE:
3015 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3016 		VI_LOCK(vp);
3017 		break;
3018 	case VPUTX_VPUT:
3019 		error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT);
3020 		VI_LOCK(vp);
3021 		break;
3022 	case VPUTX_VUNREF:
3023 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3024 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3025 			VI_LOCK(vp);
3026 		}
3027 		break;
3028 	}
3029 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
3030 	    ("vnode with usecount and VI_OWEINACT set"));
3031 	if (error == 0) {
3032 		if (vp->v_iflag & VI_OWEINACT)
3033 			vinactive(vp, curthread);
3034 		if (func != VPUTX_VUNREF)
3035 			VOP_UNLOCK(vp, 0);
3036 	}
3037 	vdropl(vp);
3038 }
3039 
3040 /*
3041  * Vnode put/release.
3042  * If count drops to zero, call inactive routine and return to freelist.
3043  */
3044 void
3045 vrele(struct vnode *vp)
3046 {
3047 
3048 	vputx(vp, VPUTX_VRELE);
3049 }
3050 
3051 /*
3052  * Release an already locked vnode.  This give the same effects as
3053  * unlock+vrele(), but takes less time and avoids releasing and
3054  * re-aquiring the lock (as vrele() acquires the lock internally.)
3055  */
3056 void
3057 vput(struct vnode *vp)
3058 {
3059 
3060 	vputx(vp, VPUTX_VPUT);
3061 }
3062 
3063 /*
3064  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3065  */
3066 void
3067 vunref(struct vnode *vp)
3068 {
3069 
3070 	vputx(vp, VPUTX_VUNREF);
3071 }
3072 
3073 /*
3074  * Increase the hold count and activate if this is the first reference.
3075  */
3076 void
3077 _vhold(struct vnode *vp, bool locked)
3078 {
3079 	struct mount *mp;
3080 
3081 	if (locked)
3082 		ASSERT_VI_LOCKED(vp, __func__);
3083 	else
3084 		ASSERT_VI_UNLOCKED(vp, __func__);
3085 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3086 	if (!locked) {
3087 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
3088 			VNODE_REFCOUNT_FENCE_ACQ();
3089 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3090 			    ("_vhold: vnode with holdcnt is free"));
3091 			return;
3092 		}
3093 		VI_LOCK(vp);
3094 	}
3095 	if ((vp->v_iflag & VI_FREE) == 0) {
3096 		refcount_acquire(&vp->v_holdcnt);
3097 		if (!locked)
3098 			VI_UNLOCK(vp);
3099 		return;
3100 	}
3101 	VNASSERT(vp->v_holdcnt == 0, vp,
3102 	    ("%s: wrong hold count", __func__));
3103 	VNASSERT(vp->v_op != NULL, vp,
3104 	    ("%s: vnode already reclaimed.", __func__));
3105 	/*
3106 	 * Remove a vnode from the free list, mark it as in use,
3107 	 * and put it on the active list.
3108 	 */
3109 	VNASSERT(vp->v_mount != NULL, vp,
3110 	    ("_vhold: vnode not on per mount vnode list"));
3111 	mp = vp->v_mount;
3112 	mtx_lock(&mp->mnt_listmtx);
3113 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
3114 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3115 		mp->mnt_tmpfreevnodelistsize--;
3116 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
3117 	} else {
3118 		mtx_lock(&vnode_free_list_mtx);
3119 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3120 		freevnodes--;
3121 		mtx_unlock(&vnode_free_list_mtx);
3122 	}
3123 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3124 	    ("Activating already active vnode"));
3125 	vp->v_iflag &= ~VI_FREE;
3126 	vp->v_iflag |= VI_ACTIVE;
3127 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3128 	mp->mnt_activevnodelistsize++;
3129 	mtx_unlock(&mp->mnt_listmtx);
3130 	refcount_acquire(&vp->v_holdcnt);
3131 	if (!locked)
3132 		VI_UNLOCK(vp);
3133 }
3134 
3135 void
3136 vholdnz(struct vnode *vp)
3137 {
3138 
3139 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3140 #ifdef INVARIANTS
3141 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3142 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
3143 #else
3144 	atomic_add_int(&vp->v_holdcnt, 1);
3145 #endif
3146 }
3147 
3148 /*
3149  * Drop the hold count of the vnode.  If this is the last reference to
3150  * the vnode we place it on the free list unless it has been vgone'd
3151  * (marked VI_DOOMED) in which case we will free it.
3152  *
3153  * Because the vnode vm object keeps a hold reference on the vnode if
3154  * there is at least one resident non-cached page, the vnode cannot
3155  * leave the active list without the page cleanup done.
3156  */
3157 void
3158 _vdrop(struct vnode *vp, bool locked)
3159 {
3160 	struct bufobj *bo;
3161 	struct mount *mp;
3162 	int active;
3163 
3164 	if (locked)
3165 		ASSERT_VI_LOCKED(vp, __func__);
3166 	else
3167 		ASSERT_VI_UNLOCKED(vp, __func__);
3168 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3169 	if (__predict_false((int)vp->v_holdcnt <= 0)) {
3170 		vn_printf(vp, "vdrop: holdcnt %d", vp->v_holdcnt);
3171 		panic("vdrop: wrong holdcnt");
3172 	}
3173 	if (!locked) {
3174 		if (refcount_release_if_not_last(&vp->v_holdcnt))
3175 			return;
3176 		VI_LOCK(vp);
3177 	}
3178 	if (refcount_release(&vp->v_holdcnt) == 0) {
3179 		VI_UNLOCK(vp);
3180 		return;
3181 	}
3182 	if ((vp->v_iflag & VI_DOOMED) == 0) {
3183 		/*
3184 		 * Mark a vnode as free: remove it from its active list
3185 		 * and put it up for recycling on the freelist.
3186 		 */
3187 		VNASSERT(vp->v_op != NULL, vp,
3188 		    ("vdropl: vnode already reclaimed."));
3189 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3190 		    ("vnode already free"));
3191 		VNASSERT(vp->v_holdcnt == 0, vp,
3192 		    ("vdropl: freeing when we shouldn't"));
3193 		active = vp->v_iflag & VI_ACTIVE;
3194 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
3195 			vp->v_iflag &= ~VI_ACTIVE;
3196 			mp = vp->v_mount;
3197 			if (mp != NULL) {
3198 				mtx_lock(&mp->mnt_listmtx);
3199 				if (active) {
3200 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
3201 					    vp, v_actfreelist);
3202 					mp->mnt_activevnodelistsize--;
3203 				}
3204 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3205 				    vp, v_actfreelist);
3206 				mp->mnt_tmpfreevnodelistsize++;
3207 				vp->v_iflag |= VI_FREE;
3208 				vp->v_mflag |= VMP_TMPMNTFREELIST;
3209 				VI_UNLOCK(vp);
3210 				if (mp->mnt_tmpfreevnodelistsize >=
3211 				    mnt_free_list_batch)
3212 					vnlru_return_batch_locked(mp);
3213 				mtx_unlock(&mp->mnt_listmtx);
3214 			} else {
3215 				VNASSERT(active == 0, vp,
3216 				    ("vdropl: active vnode not on per mount "
3217 				    "vnode list"));
3218 				mtx_lock(&vnode_free_list_mtx);
3219 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3220 				    v_actfreelist);
3221 				freevnodes++;
3222 				vp->v_iflag |= VI_FREE;
3223 				VI_UNLOCK(vp);
3224 				mtx_unlock(&vnode_free_list_mtx);
3225 			}
3226 		} else {
3227 			VI_UNLOCK(vp);
3228 			counter_u64_add(free_owe_inact, 1);
3229 		}
3230 		return;
3231 	}
3232 	/*
3233 	 * The vnode has been marked for destruction, so free it.
3234 	 *
3235 	 * The vnode will be returned to the zone where it will
3236 	 * normally remain until it is needed for another vnode. We
3237 	 * need to cleanup (or verify that the cleanup has already
3238 	 * been done) any residual data left from its current use
3239 	 * so as not to contaminate the freshly allocated vnode.
3240 	 */
3241 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
3242 	atomic_subtract_long(&numvnodes, 1);
3243 	bo = &vp->v_bufobj;
3244 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3245 	    ("cleaned vnode still on the free list."));
3246 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
3247 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
3248 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
3249 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
3250 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
3251 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
3252 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
3253 	    ("clean blk trie not empty"));
3254 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3255 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3256 	    ("dirty blk trie not empty"));
3257 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3258 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3259 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3260 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3261 	    ("Dangling rangelock waiters"));
3262 	VI_UNLOCK(vp);
3263 #ifdef MAC
3264 	mac_vnode_destroy(vp);
3265 #endif
3266 	if (vp->v_pollinfo != NULL) {
3267 		destroy_vpollinfo(vp->v_pollinfo);
3268 		vp->v_pollinfo = NULL;
3269 	}
3270 #ifdef INVARIANTS
3271 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3272 	vp->v_op = NULL;
3273 #endif
3274 	vp->v_mountedhere = NULL;
3275 	vp->v_unpcb = NULL;
3276 	vp->v_rdev = NULL;
3277 	vp->v_fifoinfo = NULL;
3278 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3279 	vp->v_iflag = 0;
3280 	vp->v_vflag = 0;
3281 	bo->bo_flag = 0;
3282 	uma_zfree(vnode_zone, vp);
3283 }
3284 
3285 /*
3286  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3287  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3288  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3289  * failed lock upgrade.
3290  */
3291 void
3292 vinactive(struct vnode *vp, struct thread *td)
3293 {
3294 	struct vm_object *obj;
3295 
3296 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3297 	ASSERT_VI_LOCKED(vp, "vinactive");
3298 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3299 	    ("vinactive: recursed on VI_DOINGINACT"));
3300 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3301 	vp->v_iflag |= VI_DOINGINACT;
3302 	vp->v_iflag &= ~VI_OWEINACT;
3303 	VI_UNLOCK(vp);
3304 	/*
3305 	 * Before moving off the active list, we must be sure that any
3306 	 * modified pages are converted into the vnode's dirty
3307 	 * buffers, since these will no longer be checked once the
3308 	 * vnode is on the inactive list.
3309 	 *
3310 	 * The write-out of the dirty pages is asynchronous.  At the
3311 	 * point that VOP_INACTIVE() is called, there could still be
3312 	 * pending I/O and dirty pages in the object.
3313 	 */
3314 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3315 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3316 		VM_OBJECT_WLOCK(obj);
3317 		vm_object_page_clean(obj, 0, 0, 0);
3318 		VM_OBJECT_WUNLOCK(obj);
3319 	}
3320 	VOP_INACTIVE(vp, td);
3321 	VI_LOCK(vp);
3322 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3323 	    ("vinactive: lost VI_DOINGINACT"));
3324 	vp->v_iflag &= ~VI_DOINGINACT;
3325 }
3326 
3327 /*
3328  * Remove any vnodes in the vnode table belonging to mount point mp.
3329  *
3330  * If FORCECLOSE is not specified, there should not be any active ones,
3331  * return error if any are found (nb: this is a user error, not a
3332  * system error). If FORCECLOSE is specified, detach any active vnodes
3333  * that are found.
3334  *
3335  * If WRITECLOSE is set, only flush out regular file vnodes open for
3336  * writing.
3337  *
3338  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3339  *
3340  * `rootrefs' specifies the base reference count for the root vnode
3341  * of this filesystem. The root vnode is considered busy if its
3342  * v_usecount exceeds this value. On a successful return, vflush(, td)
3343  * will call vrele() on the root vnode exactly rootrefs times.
3344  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3345  * be zero.
3346  */
3347 #ifdef DIAGNOSTIC
3348 static int busyprt = 0;		/* print out busy vnodes */
3349 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3350 #endif
3351 
3352 int
3353 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3354 {
3355 	struct vnode *vp, *mvp, *rootvp = NULL;
3356 	struct vattr vattr;
3357 	int busy = 0, error;
3358 
3359 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3360 	    rootrefs, flags);
3361 	if (rootrefs > 0) {
3362 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3363 		    ("vflush: bad args"));
3364 		/*
3365 		 * Get the filesystem root vnode. We can vput() it
3366 		 * immediately, since with rootrefs > 0, it won't go away.
3367 		 */
3368 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3369 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3370 			    __func__, error);
3371 			return (error);
3372 		}
3373 		vput(rootvp);
3374 	}
3375 loop:
3376 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3377 		vholdl(vp);
3378 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3379 		if (error) {
3380 			vdrop(vp);
3381 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3382 			goto loop;
3383 		}
3384 		/*
3385 		 * Skip over a vnodes marked VV_SYSTEM.
3386 		 */
3387 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3388 			VOP_UNLOCK(vp, 0);
3389 			vdrop(vp);
3390 			continue;
3391 		}
3392 		/*
3393 		 * If WRITECLOSE is set, flush out unlinked but still open
3394 		 * files (even if open only for reading) and regular file
3395 		 * vnodes open for writing.
3396 		 */
3397 		if (flags & WRITECLOSE) {
3398 			if (vp->v_object != NULL) {
3399 				VM_OBJECT_WLOCK(vp->v_object);
3400 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3401 				VM_OBJECT_WUNLOCK(vp->v_object);
3402 			}
3403 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3404 			if (error != 0) {
3405 				VOP_UNLOCK(vp, 0);
3406 				vdrop(vp);
3407 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3408 				return (error);
3409 			}
3410 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3411 			VI_LOCK(vp);
3412 
3413 			if ((vp->v_type == VNON ||
3414 			    (error == 0 && vattr.va_nlink > 0)) &&
3415 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3416 				VOP_UNLOCK(vp, 0);
3417 				vdropl(vp);
3418 				continue;
3419 			}
3420 		} else
3421 			VI_LOCK(vp);
3422 		/*
3423 		 * With v_usecount == 0, all we need to do is clear out the
3424 		 * vnode data structures and we are done.
3425 		 *
3426 		 * If FORCECLOSE is set, forcibly close the vnode.
3427 		 */
3428 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3429 			vgonel(vp);
3430 		} else {
3431 			busy++;
3432 #ifdef DIAGNOSTIC
3433 			if (busyprt)
3434 				vn_printf(vp, "vflush: busy vnode ");
3435 #endif
3436 		}
3437 		VOP_UNLOCK(vp, 0);
3438 		vdropl(vp);
3439 	}
3440 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3441 		/*
3442 		 * If just the root vnode is busy, and if its refcount
3443 		 * is equal to `rootrefs', then go ahead and kill it.
3444 		 */
3445 		VI_LOCK(rootvp);
3446 		KASSERT(busy > 0, ("vflush: not busy"));
3447 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3448 		    ("vflush: usecount %d < rootrefs %d",
3449 		     rootvp->v_usecount, rootrefs));
3450 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3451 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3452 			vgone(rootvp);
3453 			VOP_UNLOCK(rootvp, 0);
3454 			busy = 0;
3455 		} else
3456 			VI_UNLOCK(rootvp);
3457 	}
3458 	if (busy) {
3459 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3460 		    busy);
3461 		return (EBUSY);
3462 	}
3463 	for (; rootrefs > 0; rootrefs--)
3464 		vrele(rootvp);
3465 	return (0);
3466 }
3467 
3468 /*
3469  * Recycle an unused vnode to the front of the free list.
3470  */
3471 int
3472 vrecycle(struct vnode *vp)
3473 {
3474 	int recycled;
3475 
3476 	VI_LOCK(vp);
3477 	recycled = vrecyclel(vp);
3478 	VI_UNLOCK(vp);
3479 	return (recycled);
3480 }
3481 
3482 /*
3483  * vrecycle, with the vp interlock held.
3484  */
3485 int
3486 vrecyclel(struct vnode *vp)
3487 {
3488 	int recycled;
3489 
3490 	ASSERT_VOP_ELOCKED(vp, __func__);
3491 	ASSERT_VI_LOCKED(vp, __func__);
3492 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3493 	recycled = 0;
3494 	if (vp->v_usecount == 0) {
3495 		recycled = 1;
3496 		vgonel(vp);
3497 	}
3498 	return (recycled);
3499 }
3500 
3501 /*
3502  * Eliminate all activity associated with a vnode
3503  * in preparation for reuse.
3504  */
3505 void
3506 vgone(struct vnode *vp)
3507 {
3508 	VI_LOCK(vp);
3509 	vgonel(vp);
3510 	VI_UNLOCK(vp);
3511 }
3512 
3513 static void
3514 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3515     struct vnode *lowervp __unused)
3516 {
3517 }
3518 
3519 /*
3520  * Notify upper mounts about reclaimed or unlinked vnode.
3521  */
3522 void
3523 vfs_notify_upper(struct vnode *vp, int event)
3524 {
3525 	static struct vfsops vgonel_vfsops = {
3526 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3527 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3528 	};
3529 	struct mount *mp, *ump, *mmp;
3530 
3531 	mp = vp->v_mount;
3532 	if (mp == NULL)
3533 		return;
3534 
3535 	MNT_ILOCK(mp);
3536 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3537 		goto unlock;
3538 	MNT_IUNLOCK(mp);
3539 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3540 	mmp->mnt_op = &vgonel_vfsops;
3541 	mmp->mnt_kern_flag |= MNTK_MARKER;
3542 	MNT_ILOCK(mp);
3543 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3544 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3545 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3546 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3547 			continue;
3548 		}
3549 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3550 		MNT_IUNLOCK(mp);
3551 		switch (event) {
3552 		case VFS_NOTIFY_UPPER_RECLAIM:
3553 			VFS_RECLAIM_LOWERVP(ump, vp);
3554 			break;
3555 		case VFS_NOTIFY_UPPER_UNLINK:
3556 			VFS_UNLINK_LOWERVP(ump, vp);
3557 			break;
3558 		default:
3559 			KASSERT(0, ("invalid event %d", event));
3560 			break;
3561 		}
3562 		MNT_ILOCK(mp);
3563 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3564 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3565 	}
3566 	free(mmp, M_TEMP);
3567 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3568 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3569 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3570 		wakeup(&mp->mnt_uppers);
3571 	}
3572 unlock:
3573 	MNT_IUNLOCK(mp);
3574 }
3575 
3576 /*
3577  * vgone, with the vp interlock held.
3578  */
3579 static void
3580 vgonel(struct vnode *vp)
3581 {
3582 	struct thread *td;
3583 	struct mount *mp;
3584 	vm_object_t object;
3585 	bool active, oweinact;
3586 
3587 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3588 	ASSERT_VI_LOCKED(vp, "vgonel");
3589 	VNASSERT(vp->v_holdcnt, vp,
3590 	    ("vgonel: vp %p has no reference.", vp));
3591 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3592 	td = curthread;
3593 
3594 	/*
3595 	 * Don't vgonel if we're already doomed.
3596 	 */
3597 	if (vp->v_iflag & VI_DOOMED)
3598 		return;
3599 	vp->v_iflag |= VI_DOOMED;
3600 
3601 	/*
3602 	 * Check to see if the vnode is in use.  If so, we have to call
3603 	 * VOP_CLOSE() and VOP_INACTIVE().
3604 	 */
3605 	active = vp->v_usecount > 0;
3606 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3607 	VI_UNLOCK(vp);
3608 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3609 
3610 	/*
3611 	 * If purging an active vnode, it must be closed and
3612 	 * deactivated before being reclaimed.
3613 	 */
3614 	if (active)
3615 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3616 	if (oweinact || active) {
3617 		VI_LOCK(vp);
3618 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3619 			vinactive(vp, td);
3620 		VI_UNLOCK(vp);
3621 	}
3622 	if (vp->v_type == VSOCK)
3623 		vfs_unp_reclaim(vp);
3624 
3625 	/*
3626 	 * Clean out any buffers associated with the vnode.
3627 	 * If the flush fails, just toss the buffers.
3628 	 */
3629 	mp = NULL;
3630 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3631 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3632 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3633 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3634 			;
3635 	}
3636 
3637 	BO_LOCK(&vp->v_bufobj);
3638 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3639 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3640 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3641 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3642 	    ("vp %p bufobj not invalidated", vp));
3643 
3644 	/*
3645 	 * For VMIO bufobj, BO_DEAD is set later, or in
3646 	 * vm_object_terminate() after the object's page queue is
3647 	 * flushed.
3648 	 */
3649 	object = vp->v_bufobj.bo_object;
3650 	if (object == NULL)
3651 		vp->v_bufobj.bo_flag |= BO_DEAD;
3652 	BO_UNLOCK(&vp->v_bufobj);
3653 
3654 	/*
3655 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3656 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3657 	 * should not touch the object borrowed from the lower vnode
3658 	 * (the handle check).
3659 	 */
3660 	if (object != NULL && object->type == OBJT_VNODE &&
3661 	    object->handle == vp)
3662 		vnode_destroy_vobject(vp);
3663 
3664 	/*
3665 	 * Reclaim the vnode.
3666 	 */
3667 	if (VOP_RECLAIM(vp, td))
3668 		panic("vgone: cannot reclaim");
3669 	if (mp != NULL)
3670 		vn_finished_secondary_write(mp);
3671 	VNASSERT(vp->v_object == NULL, vp,
3672 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3673 	/*
3674 	 * Clear the advisory locks and wake up waiting threads.
3675 	 */
3676 	(void)VOP_ADVLOCKPURGE(vp);
3677 	vp->v_lockf = NULL;
3678 	/*
3679 	 * Delete from old mount point vnode list.
3680 	 */
3681 	delmntque(vp);
3682 	cache_purge(vp);
3683 	/*
3684 	 * Done with purge, reset to the standard lock and invalidate
3685 	 * the vnode.
3686 	 */
3687 	VI_LOCK(vp);
3688 	vp->v_vnlock = &vp->v_lock;
3689 	vp->v_op = &dead_vnodeops;
3690 	vp->v_tag = "none";
3691 	vp->v_type = VBAD;
3692 }
3693 
3694 /*
3695  * Calculate the total number of references to a special device.
3696  */
3697 int
3698 vcount(struct vnode *vp)
3699 {
3700 	int count;
3701 
3702 	dev_lock();
3703 	count = vp->v_rdev->si_usecount;
3704 	dev_unlock();
3705 	return (count);
3706 }
3707 
3708 /*
3709  * Same as above, but using the struct cdev *as argument
3710  */
3711 int
3712 count_dev(struct cdev *dev)
3713 {
3714 	int count;
3715 
3716 	dev_lock();
3717 	count = dev->si_usecount;
3718 	dev_unlock();
3719 	return(count);
3720 }
3721 
3722 /*
3723  * Print out a description of a vnode.
3724  */
3725 static char *typename[] =
3726 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3727  "VMARKER"};
3728 
3729 void
3730 vn_printf(struct vnode *vp, const char *fmt, ...)
3731 {
3732 	va_list ap;
3733 	char buf[256], buf2[16];
3734 	u_long flags;
3735 
3736 	va_start(ap, fmt);
3737 	vprintf(fmt, ap);
3738 	va_end(ap);
3739 	printf("%p: ", (void *)vp);
3740 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3741 	printf("    usecount %d, writecount %d, refcount %d",
3742 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3743 	switch (vp->v_type) {
3744 	case VDIR:
3745 		printf(" mountedhere %p\n", vp->v_mountedhere);
3746 		break;
3747 	case VCHR:
3748 		printf(" rdev %p\n", vp->v_rdev);
3749 		break;
3750 	case VSOCK:
3751 		printf(" socket %p\n", vp->v_unpcb);
3752 		break;
3753 	case VFIFO:
3754 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3755 		break;
3756 	default:
3757 		printf("\n");
3758 		break;
3759 	}
3760 	buf[0] = '\0';
3761 	buf[1] = '\0';
3762 	if (vp->v_vflag & VV_ROOT)
3763 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3764 	if (vp->v_vflag & VV_ISTTY)
3765 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3766 	if (vp->v_vflag & VV_NOSYNC)
3767 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3768 	if (vp->v_vflag & VV_ETERNALDEV)
3769 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3770 	if (vp->v_vflag & VV_CACHEDLABEL)
3771 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3772 	if (vp->v_vflag & VV_COPYONWRITE)
3773 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3774 	if (vp->v_vflag & VV_SYSTEM)
3775 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3776 	if (vp->v_vflag & VV_PROCDEP)
3777 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3778 	if (vp->v_vflag & VV_NOKNOTE)
3779 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3780 	if (vp->v_vflag & VV_DELETED)
3781 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3782 	if (vp->v_vflag & VV_MD)
3783 		strlcat(buf, "|VV_MD", sizeof(buf));
3784 	if (vp->v_vflag & VV_FORCEINSMQ)
3785 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3786 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3787 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3788 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3789 	if (flags != 0) {
3790 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3791 		strlcat(buf, buf2, sizeof(buf));
3792 	}
3793 	if (vp->v_iflag & VI_MOUNT)
3794 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3795 	if (vp->v_iflag & VI_DOOMED)
3796 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3797 	if (vp->v_iflag & VI_FREE)
3798 		strlcat(buf, "|VI_FREE", sizeof(buf));
3799 	if (vp->v_iflag & VI_ACTIVE)
3800 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3801 	if (vp->v_iflag & VI_DOINGINACT)
3802 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3803 	if (vp->v_iflag & VI_OWEINACT)
3804 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3805 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3806 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3807 	if (flags != 0) {
3808 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3809 		strlcat(buf, buf2, sizeof(buf));
3810 	}
3811 	printf("    flags (%s)\n", buf + 1);
3812 	if (mtx_owned(VI_MTX(vp)))
3813 		printf(" VI_LOCKed");
3814 	if (vp->v_object != NULL)
3815 		printf("    v_object %p ref %d pages %d "
3816 		    "cleanbuf %d dirtybuf %d\n",
3817 		    vp->v_object, vp->v_object->ref_count,
3818 		    vp->v_object->resident_page_count,
3819 		    vp->v_bufobj.bo_clean.bv_cnt,
3820 		    vp->v_bufobj.bo_dirty.bv_cnt);
3821 	printf("    ");
3822 	lockmgr_printinfo(vp->v_vnlock);
3823 	if (vp->v_data != NULL)
3824 		VOP_PRINT(vp);
3825 }
3826 
3827 #ifdef DDB
3828 /*
3829  * List all of the locked vnodes in the system.
3830  * Called when debugging the kernel.
3831  */
3832 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3833 {
3834 	struct mount *mp;
3835 	struct vnode *vp;
3836 
3837 	/*
3838 	 * Note: because this is DDB, we can't obey the locking semantics
3839 	 * for these structures, which means we could catch an inconsistent
3840 	 * state and dereference a nasty pointer.  Not much to be done
3841 	 * about that.
3842 	 */
3843 	db_printf("Locked vnodes\n");
3844 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3845 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3846 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3847 				vn_printf(vp, "vnode ");
3848 		}
3849 	}
3850 }
3851 
3852 /*
3853  * Show details about the given vnode.
3854  */
3855 DB_SHOW_COMMAND(vnode, db_show_vnode)
3856 {
3857 	struct vnode *vp;
3858 
3859 	if (!have_addr)
3860 		return;
3861 	vp = (struct vnode *)addr;
3862 	vn_printf(vp, "vnode ");
3863 }
3864 
3865 /*
3866  * Show details about the given mount point.
3867  */
3868 DB_SHOW_COMMAND(mount, db_show_mount)
3869 {
3870 	struct mount *mp;
3871 	struct vfsopt *opt;
3872 	struct statfs *sp;
3873 	struct vnode *vp;
3874 	char buf[512];
3875 	uint64_t mflags;
3876 	u_int flags;
3877 
3878 	if (!have_addr) {
3879 		/* No address given, print short info about all mount points. */
3880 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3881 			db_printf("%p %s on %s (%s)\n", mp,
3882 			    mp->mnt_stat.f_mntfromname,
3883 			    mp->mnt_stat.f_mntonname,
3884 			    mp->mnt_stat.f_fstypename);
3885 			if (db_pager_quit)
3886 				break;
3887 		}
3888 		db_printf("\nMore info: show mount <addr>\n");
3889 		return;
3890 	}
3891 
3892 	mp = (struct mount *)addr;
3893 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3894 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3895 
3896 	buf[0] = '\0';
3897 	mflags = mp->mnt_flag;
3898 #define	MNT_FLAG(flag)	do {						\
3899 	if (mflags & (flag)) {						\
3900 		if (buf[0] != '\0')					\
3901 			strlcat(buf, ", ", sizeof(buf));		\
3902 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3903 		mflags &= ~(flag);					\
3904 	}								\
3905 } while (0)
3906 	MNT_FLAG(MNT_RDONLY);
3907 	MNT_FLAG(MNT_SYNCHRONOUS);
3908 	MNT_FLAG(MNT_NOEXEC);
3909 	MNT_FLAG(MNT_NOSUID);
3910 	MNT_FLAG(MNT_NFS4ACLS);
3911 	MNT_FLAG(MNT_UNION);
3912 	MNT_FLAG(MNT_ASYNC);
3913 	MNT_FLAG(MNT_SUIDDIR);
3914 	MNT_FLAG(MNT_SOFTDEP);
3915 	MNT_FLAG(MNT_NOSYMFOLLOW);
3916 	MNT_FLAG(MNT_GJOURNAL);
3917 	MNT_FLAG(MNT_MULTILABEL);
3918 	MNT_FLAG(MNT_ACLS);
3919 	MNT_FLAG(MNT_NOATIME);
3920 	MNT_FLAG(MNT_NOCLUSTERR);
3921 	MNT_FLAG(MNT_NOCLUSTERW);
3922 	MNT_FLAG(MNT_SUJ);
3923 	MNT_FLAG(MNT_EXRDONLY);
3924 	MNT_FLAG(MNT_EXPORTED);
3925 	MNT_FLAG(MNT_DEFEXPORTED);
3926 	MNT_FLAG(MNT_EXPORTANON);
3927 	MNT_FLAG(MNT_EXKERB);
3928 	MNT_FLAG(MNT_EXPUBLIC);
3929 	MNT_FLAG(MNT_LOCAL);
3930 	MNT_FLAG(MNT_QUOTA);
3931 	MNT_FLAG(MNT_ROOTFS);
3932 	MNT_FLAG(MNT_USER);
3933 	MNT_FLAG(MNT_IGNORE);
3934 	MNT_FLAG(MNT_UPDATE);
3935 	MNT_FLAG(MNT_DELEXPORT);
3936 	MNT_FLAG(MNT_RELOAD);
3937 	MNT_FLAG(MNT_FORCE);
3938 	MNT_FLAG(MNT_SNAPSHOT);
3939 	MNT_FLAG(MNT_BYFSID);
3940 #undef MNT_FLAG
3941 	if (mflags != 0) {
3942 		if (buf[0] != '\0')
3943 			strlcat(buf, ", ", sizeof(buf));
3944 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3945 		    "0x%016jx", mflags);
3946 	}
3947 	db_printf("    mnt_flag = %s\n", buf);
3948 
3949 	buf[0] = '\0';
3950 	flags = mp->mnt_kern_flag;
3951 #define	MNT_KERN_FLAG(flag)	do {					\
3952 	if (flags & (flag)) {						\
3953 		if (buf[0] != '\0')					\
3954 			strlcat(buf, ", ", sizeof(buf));		\
3955 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3956 		flags &= ~(flag);					\
3957 	}								\
3958 } while (0)
3959 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3960 	MNT_KERN_FLAG(MNTK_ASYNC);
3961 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3962 	MNT_KERN_FLAG(MNTK_DRAINING);
3963 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3964 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3965 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3966 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3967 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3968 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3969 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3970 	MNT_KERN_FLAG(MNTK_MARKER);
3971 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3972 	MNT_KERN_FLAG(MNTK_NOASYNC);
3973 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3974 	MNT_KERN_FLAG(MNTK_MWAIT);
3975 	MNT_KERN_FLAG(MNTK_SUSPEND);
3976 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3977 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3978 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3979 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3980 #undef MNT_KERN_FLAG
3981 	if (flags != 0) {
3982 		if (buf[0] != '\0')
3983 			strlcat(buf, ", ", sizeof(buf));
3984 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3985 		    "0x%08x", flags);
3986 	}
3987 	db_printf("    mnt_kern_flag = %s\n", buf);
3988 
3989 	db_printf("    mnt_opt = ");
3990 	opt = TAILQ_FIRST(mp->mnt_opt);
3991 	if (opt != NULL) {
3992 		db_printf("%s", opt->name);
3993 		opt = TAILQ_NEXT(opt, link);
3994 		while (opt != NULL) {
3995 			db_printf(", %s", opt->name);
3996 			opt = TAILQ_NEXT(opt, link);
3997 		}
3998 	}
3999 	db_printf("\n");
4000 
4001 	sp = &mp->mnt_stat;
4002 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4003 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4004 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4005 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4006 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4007 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4008 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4009 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4010 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4011 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4012 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4013 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4014 
4015 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4016 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4017 	if (jailed(mp->mnt_cred))
4018 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4019 	db_printf(" }\n");
4020 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
4021 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4022 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4023 	db_printf("    mnt_activevnodelistsize = %d\n",
4024 	    mp->mnt_activevnodelistsize);
4025 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
4026 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4027 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4028 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4029 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
4030 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4031 	db_printf("    mnt_secondary_accwrites = %d\n",
4032 	    mp->mnt_secondary_accwrites);
4033 	db_printf("    mnt_gjprovider = %s\n",
4034 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4035 
4036 	db_printf("\n\nList of active vnodes\n");
4037 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
4038 		if (vp->v_type != VMARKER) {
4039 			vn_printf(vp, "vnode ");
4040 			if (db_pager_quit)
4041 				break;
4042 		}
4043 	}
4044 	db_printf("\n\nList of inactive vnodes\n");
4045 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4046 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
4047 			vn_printf(vp, "vnode ");
4048 			if (db_pager_quit)
4049 				break;
4050 		}
4051 	}
4052 }
4053 #endif	/* DDB */
4054 
4055 /*
4056  * Fill in a struct xvfsconf based on a struct vfsconf.
4057  */
4058 static int
4059 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4060 {
4061 	struct xvfsconf xvfsp;
4062 
4063 	bzero(&xvfsp, sizeof(xvfsp));
4064 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4065 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4066 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4067 	xvfsp.vfc_flags = vfsp->vfc_flags;
4068 	/*
4069 	 * These are unused in userland, we keep them
4070 	 * to not break binary compatibility.
4071 	 */
4072 	xvfsp.vfc_vfsops = NULL;
4073 	xvfsp.vfc_next = NULL;
4074 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4075 }
4076 
4077 #ifdef COMPAT_FREEBSD32
4078 struct xvfsconf32 {
4079 	uint32_t	vfc_vfsops;
4080 	char		vfc_name[MFSNAMELEN];
4081 	int32_t		vfc_typenum;
4082 	int32_t		vfc_refcount;
4083 	int32_t		vfc_flags;
4084 	uint32_t	vfc_next;
4085 };
4086 
4087 static int
4088 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4089 {
4090 	struct xvfsconf32 xvfsp;
4091 
4092 	bzero(&xvfsp, sizeof(xvfsp));
4093 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4094 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4095 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4096 	xvfsp.vfc_flags = vfsp->vfc_flags;
4097 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4098 }
4099 #endif
4100 
4101 /*
4102  * Top level filesystem related information gathering.
4103  */
4104 static int
4105 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4106 {
4107 	struct vfsconf *vfsp;
4108 	int error;
4109 
4110 	error = 0;
4111 	vfsconf_slock();
4112 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4113 #ifdef COMPAT_FREEBSD32
4114 		if (req->flags & SCTL_MASK32)
4115 			error = vfsconf2x32(req, vfsp);
4116 		else
4117 #endif
4118 			error = vfsconf2x(req, vfsp);
4119 		if (error)
4120 			break;
4121 	}
4122 	vfsconf_sunlock();
4123 	return (error);
4124 }
4125 
4126 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4127     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4128     "S,xvfsconf", "List of all configured filesystems");
4129 
4130 #ifndef BURN_BRIDGES
4131 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4132 
4133 static int
4134 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4135 {
4136 	int *name = (int *)arg1 - 1;	/* XXX */
4137 	u_int namelen = arg2 + 1;	/* XXX */
4138 	struct vfsconf *vfsp;
4139 
4140 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4141 	    "please rebuild world\n");
4142 
4143 #if 1 || defined(COMPAT_PRELITE2)
4144 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4145 	if (namelen == 1)
4146 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4147 #endif
4148 
4149 	switch (name[1]) {
4150 	case VFS_MAXTYPENUM:
4151 		if (namelen != 2)
4152 			return (ENOTDIR);
4153 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4154 	case VFS_CONF:
4155 		if (namelen != 3)
4156 			return (ENOTDIR);	/* overloaded */
4157 		vfsconf_slock();
4158 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4159 			if (vfsp->vfc_typenum == name[2])
4160 				break;
4161 		}
4162 		vfsconf_sunlock();
4163 		if (vfsp == NULL)
4164 			return (EOPNOTSUPP);
4165 #ifdef COMPAT_FREEBSD32
4166 		if (req->flags & SCTL_MASK32)
4167 			return (vfsconf2x32(req, vfsp));
4168 		else
4169 #endif
4170 			return (vfsconf2x(req, vfsp));
4171 	}
4172 	return (EOPNOTSUPP);
4173 }
4174 
4175 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4176     CTLFLAG_MPSAFE, vfs_sysctl,
4177     "Generic filesystem");
4178 
4179 #if 1 || defined(COMPAT_PRELITE2)
4180 
4181 static int
4182 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4183 {
4184 	int error;
4185 	struct vfsconf *vfsp;
4186 	struct ovfsconf ovfs;
4187 
4188 	vfsconf_slock();
4189 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4190 		bzero(&ovfs, sizeof(ovfs));
4191 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4192 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4193 		ovfs.vfc_index = vfsp->vfc_typenum;
4194 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4195 		ovfs.vfc_flags = vfsp->vfc_flags;
4196 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4197 		if (error != 0) {
4198 			vfsconf_sunlock();
4199 			return (error);
4200 		}
4201 	}
4202 	vfsconf_sunlock();
4203 	return (0);
4204 }
4205 
4206 #endif /* 1 || COMPAT_PRELITE2 */
4207 #endif /* !BURN_BRIDGES */
4208 
4209 #define KINFO_VNODESLOP		10
4210 #ifdef notyet
4211 /*
4212  * Dump vnode list (via sysctl).
4213  */
4214 /* ARGSUSED */
4215 static int
4216 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4217 {
4218 	struct xvnode *xvn;
4219 	struct mount *mp;
4220 	struct vnode *vp;
4221 	int error, len, n;
4222 
4223 	/*
4224 	 * Stale numvnodes access is not fatal here.
4225 	 */
4226 	req->lock = 0;
4227 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4228 	if (!req->oldptr)
4229 		/* Make an estimate */
4230 		return (SYSCTL_OUT(req, 0, len));
4231 
4232 	error = sysctl_wire_old_buffer(req, 0);
4233 	if (error != 0)
4234 		return (error);
4235 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4236 	n = 0;
4237 	mtx_lock(&mountlist_mtx);
4238 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4239 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4240 			continue;
4241 		MNT_ILOCK(mp);
4242 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4243 			if (n == len)
4244 				break;
4245 			vref(vp);
4246 			xvn[n].xv_size = sizeof *xvn;
4247 			xvn[n].xv_vnode = vp;
4248 			xvn[n].xv_id = 0;	/* XXX compat */
4249 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4250 			XV_COPY(usecount);
4251 			XV_COPY(writecount);
4252 			XV_COPY(holdcnt);
4253 			XV_COPY(mount);
4254 			XV_COPY(numoutput);
4255 			XV_COPY(type);
4256 #undef XV_COPY
4257 			xvn[n].xv_flag = vp->v_vflag;
4258 
4259 			switch (vp->v_type) {
4260 			case VREG:
4261 			case VDIR:
4262 			case VLNK:
4263 				break;
4264 			case VBLK:
4265 			case VCHR:
4266 				if (vp->v_rdev == NULL) {
4267 					vrele(vp);
4268 					continue;
4269 				}
4270 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4271 				break;
4272 			case VSOCK:
4273 				xvn[n].xv_socket = vp->v_socket;
4274 				break;
4275 			case VFIFO:
4276 				xvn[n].xv_fifo = vp->v_fifoinfo;
4277 				break;
4278 			case VNON:
4279 			case VBAD:
4280 			default:
4281 				/* shouldn't happen? */
4282 				vrele(vp);
4283 				continue;
4284 			}
4285 			vrele(vp);
4286 			++n;
4287 		}
4288 		MNT_IUNLOCK(mp);
4289 		mtx_lock(&mountlist_mtx);
4290 		vfs_unbusy(mp);
4291 		if (n == len)
4292 			break;
4293 	}
4294 	mtx_unlock(&mountlist_mtx);
4295 
4296 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4297 	free(xvn, M_TEMP);
4298 	return (error);
4299 }
4300 
4301 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4302     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4303     "");
4304 #endif
4305 
4306 static void
4307 unmount_or_warn(struct mount *mp)
4308 {
4309 	int error;
4310 
4311 	error = dounmount(mp, MNT_FORCE, curthread);
4312 	if (error != 0) {
4313 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4314 		if (error == EBUSY)
4315 			printf("BUSY)\n");
4316 		else
4317 			printf("%d)\n", error);
4318 	}
4319 }
4320 
4321 /*
4322  * Unmount all filesystems. The list is traversed in reverse order
4323  * of mounting to avoid dependencies.
4324  */
4325 void
4326 vfs_unmountall(void)
4327 {
4328 	struct mount *mp, *tmp;
4329 
4330 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4331 
4332 	/*
4333 	 * Since this only runs when rebooting, it is not interlocked.
4334 	 */
4335 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4336 		vfs_ref(mp);
4337 
4338 		/*
4339 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4340 		 * unmount of the latter.
4341 		 */
4342 		if (mp == rootdevmp)
4343 			continue;
4344 
4345 		unmount_or_warn(mp);
4346 	}
4347 
4348 	if (rootdevmp != NULL)
4349 		unmount_or_warn(rootdevmp);
4350 }
4351 
4352 /*
4353  * perform msync on all vnodes under a mount point
4354  * the mount point must be locked.
4355  */
4356 void
4357 vfs_msync(struct mount *mp, int flags)
4358 {
4359 	struct vnode *vp, *mvp;
4360 	struct vm_object *obj;
4361 
4362 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4363 
4364 	vnlru_return_batch(mp);
4365 
4366 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4367 		obj = vp->v_object;
4368 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4369 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4370 			if (!vget(vp,
4371 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4372 			    curthread)) {
4373 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4374 					vput(vp);
4375 					continue;
4376 				}
4377 
4378 				obj = vp->v_object;
4379 				if (obj != NULL) {
4380 					VM_OBJECT_WLOCK(obj);
4381 					vm_object_page_clean(obj, 0, 0,
4382 					    flags == MNT_WAIT ?
4383 					    OBJPC_SYNC : OBJPC_NOSYNC);
4384 					VM_OBJECT_WUNLOCK(obj);
4385 				}
4386 				vput(vp);
4387 			}
4388 		} else
4389 			VI_UNLOCK(vp);
4390 	}
4391 }
4392 
4393 static void
4394 destroy_vpollinfo_free(struct vpollinfo *vi)
4395 {
4396 
4397 	knlist_destroy(&vi->vpi_selinfo.si_note);
4398 	mtx_destroy(&vi->vpi_lock);
4399 	uma_zfree(vnodepoll_zone, vi);
4400 }
4401 
4402 static void
4403 destroy_vpollinfo(struct vpollinfo *vi)
4404 {
4405 
4406 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4407 	seldrain(&vi->vpi_selinfo);
4408 	destroy_vpollinfo_free(vi);
4409 }
4410 
4411 /*
4412  * Initialize per-vnode helper structure to hold poll-related state.
4413  */
4414 void
4415 v_addpollinfo(struct vnode *vp)
4416 {
4417 	struct vpollinfo *vi;
4418 
4419 	if (vp->v_pollinfo != NULL)
4420 		return;
4421 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4422 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4423 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4424 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4425 	VI_LOCK(vp);
4426 	if (vp->v_pollinfo != NULL) {
4427 		VI_UNLOCK(vp);
4428 		destroy_vpollinfo_free(vi);
4429 		return;
4430 	}
4431 	vp->v_pollinfo = vi;
4432 	VI_UNLOCK(vp);
4433 }
4434 
4435 /*
4436  * Record a process's interest in events which might happen to
4437  * a vnode.  Because poll uses the historic select-style interface
4438  * internally, this routine serves as both the ``check for any
4439  * pending events'' and the ``record my interest in future events''
4440  * functions.  (These are done together, while the lock is held,
4441  * to avoid race conditions.)
4442  */
4443 int
4444 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4445 {
4446 
4447 	v_addpollinfo(vp);
4448 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4449 	if (vp->v_pollinfo->vpi_revents & events) {
4450 		/*
4451 		 * This leaves events we are not interested
4452 		 * in available for the other process which
4453 		 * which presumably had requested them
4454 		 * (otherwise they would never have been
4455 		 * recorded).
4456 		 */
4457 		events &= vp->v_pollinfo->vpi_revents;
4458 		vp->v_pollinfo->vpi_revents &= ~events;
4459 
4460 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4461 		return (events);
4462 	}
4463 	vp->v_pollinfo->vpi_events |= events;
4464 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4465 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4466 	return (0);
4467 }
4468 
4469 /*
4470  * Routine to create and manage a filesystem syncer vnode.
4471  */
4472 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4473 static int	sync_fsync(struct  vop_fsync_args *);
4474 static int	sync_inactive(struct  vop_inactive_args *);
4475 static int	sync_reclaim(struct  vop_reclaim_args *);
4476 
4477 static struct vop_vector sync_vnodeops = {
4478 	.vop_bypass =	VOP_EOPNOTSUPP,
4479 	.vop_close =	sync_close,		/* close */
4480 	.vop_fsync =	sync_fsync,		/* fsync */
4481 	.vop_inactive =	sync_inactive,	/* inactive */
4482 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4483 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4484 	.vop_lock1 =	vop_stdlock,	/* lock */
4485 	.vop_unlock =	vop_stdunlock,	/* unlock */
4486 	.vop_islocked =	vop_stdislocked,	/* islocked */
4487 };
4488 
4489 /*
4490  * Create a new filesystem syncer vnode for the specified mount point.
4491  */
4492 void
4493 vfs_allocate_syncvnode(struct mount *mp)
4494 {
4495 	struct vnode *vp;
4496 	struct bufobj *bo;
4497 	static long start, incr, next;
4498 	int error;
4499 
4500 	/* Allocate a new vnode */
4501 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4502 	if (error != 0)
4503 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4504 	vp->v_type = VNON;
4505 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4506 	vp->v_vflag |= VV_FORCEINSMQ;
4507 	error = insmntque(vp, mp);
4508 	if (error != 0)
4509 		panic("vfs_allocate_syncvnode: insmntque() failed");
4510 	vp->v_vflag &= ~VV_FORCEINSMQ;
4511 	VOP_UNLOCK(vp, 0);
4512 	/*
4513 	 * Place the vnode onto the syncer worklist. We attempt to
4514 	 * scatter them about on the list so that they will go off
4515 	 * at evenly distributed times even if all the filesystems
4516 	 * are mounted at once.
4517 	 */
4518 	next += incr;
4519 	if (next == 0 || next > syncer_maxdelay) {
4520 		start /= 2;
4521 		incr /= 2;
4522 		if (start == 0) {
4523 			start = syncer_maxdelay / 2;
4524 			incr = syncer_maxdelay;
4525 		}
4526 		next = start;
4527 	}
4528 	bo = &vp->v_bufobj;
4529 	BO_LOCK(bo);
4530 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4531 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4532 	mtx_lock(&sync_mtx);
4533 	sync_vnode_count++;
4534 	if (mp->mnt_syncer == NULL) {
4535 		mp->mnt_syncer = vp;
4536 		vp = NULL;
4537 	}
4538 	mtx_unlock(&sync_mtx);
4539 	BO_UNLOCK(bo);
4540 	if (vp != NULL) {
4541 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4542 		vgone(vp);
4543 		vput(vp);
4544 	}
4545 }
4546 
4547 void
4548 vfs_deallocate_syncvnode(struct mount *mp)
4549 {
4550 	struct vnode *vp;
4551 
4552 	mtx_lock(&sync_mtx);
4553 	vp = mp->mnt_syncer;
4554 	if (vp != NULL)
4555 		mp->mnt_syncer = NULL;
4556 	mtx_unlock(&sync_mtx);
4557 	if (vp != NULL)
4558 		vrele(vp);
4559 }
4560 
4561 /*
4562  * Do a lazy sync of the filesystem.
4563  */
4564 static int
4565 sync_fsync(struct vop_fsync_args *ap)
4566 {
4567 	struct vnode *syncvp = ap->a_vp;
4568 	struct mount *mp = syncvp->v_mount;
4569 	int error, save;
4570 	struct bufobj *bo;
4571 
4572 	/*
4573 	 * We only need to do something if this is a lazy evaluation.
4574 	 */
4575 	if (ap->a_waitfor != MNT_LAZY)
4576 		return (0);
4577 
4578 	/*
4579 	 * Move ourselves to the back of the sync list.
4580 	 */
4581 	bo = &syncvp->v_bufobj;
4582 	BO_LOCK(bo);
4583 	vn_syncer_add_to_worklist(bo, syncdelay);
4584 	BO_UNLOCK(bo);
4585 
4586 	/*
4587 	 * Walk the list of vnodes pushing all that are dirty and
4588 	 * not already on the sync list.
4589 	 */
4590 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4591 		return (0);
4592 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4593 		vfs_unbusy(mp);
4594 		return (0);
4595 	}
4596 	save = curthread_pflags_set(TDP_SYNCIO);
4597 	vfs_msync(mp, MNT_NOWAIT);
4598 	error = VFS_SYNC(mp, MNT_LAZY);
4599 	curthread_pflags_restore(save);
4600 	vn_finished_write(mp);
4601 	vfs_unbusy(mp);
4602 	return (error);
4603 }
4604 
4605 /*
4606  * The syncer vnode is no referenced.
4607  */
4608 static int
4609 sync_inactive(struct vop_inactive_args *ap)
4610 {
4611 
4612 	vgone(ap->a_vp);
4613 	return (0);
4614 }
4615 
4616 /*
4617  * The syncer vnode is no longer needed and is being decommissioned.
4618  *
4619  * Modifications to the worklist must be protected by sync_mtx.
4620  */
4621 static int
4622 sync_reclaim(struct vop_reclaim_args *ap)
4623 {
4624 	struct vnode *vp = ap->a_vp;
4625 	struct bufobj *bo;
4626 
4627 	bo = &vp->v_bufobj;
4628 	BO_LOCK(bo);
4629 	mtx_lock(&sync_mtx);
4630 	if (vp->v_mount->mnt_syncer == vp)
4631 		vp->v_mount->mnt_syncer = NULL;
4632 	if (bo->bo_flag & BO_ONWORKLST) {
4633 		LIST_REMOVE(bo, bo_synclist);
4634 		syncer_worklist_len--;
4635 		sync_vnode_count--;
4636 		bo->bo_flag &= ~BO_ONWORKLST;
4637 	}
4638 	mtx_unlock(&sync_mtx);
4639 	BO_UNLOCK(bo);
4640 
4641 	return (0);
4642 }
4643 
4644 int
4645 vn_need_pageq_flush(struct vnode *vp)
4646 {
4647 	struct vm_object *obj;
4648 	int need;
4649 
4650 	MPASS(mtx_owned(VI_MTX(vp)));
4651 	need = 0;
4652 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4653 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0)
4654 		need = 1;
4655 	return (need);
4656 }
4657 
4658 /*
4659  * Check if vnode represents a disk device
4660  */
4661 int
4662 vn_isdisk(struct vnode *vp, int *errp)
4663 {
4664 	int error;
4665 
4666 	if (vp->v_type != VCHR) {
4667 		error = ENOTBLK;
4668 		goto out;
4669 	}
4670 	error = 0;
4671 	dev_lock();
4672 	if (vp->v_rdev == NULL)
4673 		error = ENXIO;
4674 	else if (vp->v_rdev->si_devsw == NULL)
4675 		error = ENXIO;
4676 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4677 		error = ENOTBLK;
4678 	dev_unlock();
4679 out:
4680 	if (errp != NULL)
4681 		*errp = error;
4682 	return (error == 0);
4683 }
4684 
4685 /*
4686  * Common filesystem object access control check routine.  Accepts a
4687  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4688  * and optional call-by-reference privused argument allowing vaccess()
4689  * to indicate to the caller whether privilege was used to satisfy the
4690  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4691  */
4692 int
4693 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4694     accmode_t accmode, struct ucred *cred, int *privused)
4695 {
4696 	accmode_t dac_granted;
4697 	accmode_t priv_granted;
4698 
4699 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4700 	    ("invalid bit in accmode"));
4701 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4702 	    ("VAPPEND without VWRITE"));
4703 
4704 	/*
4705 	 * Look for a normal, non-privileged way to access the file/directory
4706 	 * as requested.  If it exists, go with that.
4707 	 */
4708 
4709 	if (privused != NULL)
4710 		*privused = 0;
4711 
4712 	dac_granted = 0;
4713 
4714 	/* Check the owner. */
4715 	if (cred->cr_uid == file_uid) {
4716 		dac_granted |= VADMIN;
4717 		if (file_mode & S_IXUSR)
4718 			dac_granted |= VEXEC;
4719 		if (file_mode & S_IRUSR)
4720 			dac_granted |= VREAD;
4721 		if (file_mode & S_IWUSR)
4722 			dac_granted |= (VWRITE | VAPPEND);
4723 
4724 		if ((accmode & dac_granted) == accmode)
4725 			return (0);
4726 
4727 		goto privcheck;
4728 	}
4729 
4730 	/* Otherwise, check the groups (first match) */
4731 	if (groupmember(file_gid, cred)) {
4732 		if (file_mode & S_IXGRP)
4733 			dac_granted |= VEXEC;
4734 		if (file_mode & S_IRGRP)
4735 			dac_granted |= VREAD;
4736 		if (file_mode & S_IWGRP)
4737 			dac_granted |= (VWRITE | VAPPEND);
4738 
4739 		if ((accmode & dac_granted) == accmode)
4740 			return (0);
4741 
4742 		goto privcheck;
4743 	}
4744 
4745 	/* Otherwise, check everyone else. */
4746 	if (file_mode & S_IXOTH)
4747 		dac_granted |= VEXEC;
4748 	if (file_mode & S_IROTH)
4749 		dac_granted |= VREAD;
4750 	if (file_mode & S_IWOTH)
4751 		dac_granted |= (VWRITE | VAPPEND);
4752 	if ((accmode & dac_granted) == accmode)
4753 		return (0);
4754 
4755 privcheck:
4756 	/*
4757 	 * Build a privilege mask to determine if the set of privileges
4758 	 * satisfies the requirements when combined with the granted mask
4759 	 * from above.  For each privilege, if the privilege is required,
4760 	 * bitwise or the request type onto the priv_granted mask.
4761 	 */
4762 	priv_granted = 0;
4763 
4764 	if (type == VDIR) {
4765 		/*
4766 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4767 		 * requests, instead of PRIV_VFS_EXEC.
4768 		 */
4769 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4770 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4771 			priv_granted |= VEXEC;
4772 	} else {
4773 		/*
4774 		 * Ensure that at least one execute bit is on. Otherwise,
4775 		 * a privileged user will always succeed, and we don't want
4776 		 * this to happen unless the file really is executable.
4777 		 */
4778 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4779 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4780 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4781 			priv_granted |= VEXEC;
4782 	}
4783 
4784 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4785 	    !priv_check_cred(cred, PRIV_VFS_READ))
4786 		priv_granted |= VREAD;
4787 
4788 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4789 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4790 		priv_granted |= (VWRITE | VAPPEND);
4791 
4792 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4793 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4794 		priv_granted |= VADMIN;
4795 
4796 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4797 		/* XXX audit: privilege used */
4798 		if (privused != NULL)
4799 			*privused = 1;
4800 		return (0);
4801 	}
4802 
4803 	return ((accmode & VADMIN) ? EPERM : EACCES);
4804 }
4805 
4806 /*
4807  * Credential check based on process requesting service, and per-attribute
4808  * permissions.
4809  */
4810 int
4811 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4812     struct thread *td, accmode_t accmode)
4813 {
4814 
4815 	/*
4816 	 * Kernel-invoked always succeeds.
4817 	 */
4818 	if (cred == NOCRED)
4819 		return (0);
4820 
4821 	/*
4822 	 * Do not allow privileged processes in jail to directly manipulate
4823 	 * system attributes.
4824 	 */
4825 	switch (attrnamespace) {
4826 	case EXTATTR_NAMESPACE_SYSTEM:
4827 		/* Potentially should be: return (EPERM); */
4828 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4829 	case EXTATTR_NAMESPACE_USER:
4830 		return (VOP_ACCESS(vp, accmode, cred, td));
4831 	default:
4832 		return (EPERM);
4833 	}
4834 }
4835 
4836 #ifdef DEBUG_VFS_LOCKS
4837 /*
4838  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4839  * no longer ok to have an unlocked VFS.
4840  */
4841 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4842 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4843 
4844 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4845 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4846     "Drop into debugger on lock violation");
4847 
4848 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4849 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4850     0, "Check for interlock across VOPs");
4851 
4852 int vfs_badlock_print = 1;	/* Print lock violations. */
4853 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4854     0, "Print lock violations");
4855 
4856 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4857 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4858     0, "Print vnode details on lock violations");
4859 
4860 #ifdef KDB
4861 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4862 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4863     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4864 #endif
4865 
4866 static void
4867 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4868 {
4869 
4870 #ifdef KDB
4871 	if (vfs_badlock_backtrace)
4872 		kdb_backtrace();
4873 #endif
4874 	if (vfs_badlock_vnode)
4875 		vn_printf(vp, "vnode ");
4876 	if (vfs_badlock_print)
4877 		printf("%s: %p %s\n", str, (void *)vp, msg);
4878 	if (vfs_badlock_ddb)
4879 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4880 }
4881 
4882 void
4883 assert_vi_locked(struct vnode *vp, const char *str)
4884 {
4885 
4886 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4887 		vfs_badlock("interlock is not locked but should be", str, vp);
4888 }
4889 
4890 void
4891 assert_vi_unlocked(struct vnode *vp, const char *str)
4892 {
4893 
4894 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4895 		vfs_badlock("interlock is locked but should not be", str, vp);
4896 }
4897 
4898 void
4899 assert_vop_locked(struct vnode *vp, const char *str)
4900 {
4901 	int locked;
4902 
4903 	if (!IGNORE_LOCK(vp)) {
4904 		locked = VOP_ISLOCKED(vp);
4905 		if (locked == 0 || locked == LK_EXCLOTHER)
4906 			vfs_badlock("is not locked but should be", str, vp);
4907 	}
4908 }
4909 
4910 void
4911 assert_vop_unlocked(struct vnode *vp, const char *str)
4912 {
4913 
4914 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4915 		vfs_badlock("is locked but should not be", str, vp);
4916 }
4917 
4918 void
4919 assert_vop_elocked(struct vnode *vp, const char *str)
4920 {
4921 
4922 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4923 		vfs_badlock("is not exclusive locked but should be", str, vp);
4924 }
4925 #endif /* DEBUG_VFS_LOCKS */
4926 
4927 void
4928 vop_rename_fail(struct vop_rename_args *ap)
4929 {
4930 
4931 	if (ap->a_tvp != NULL)
4932 		vput(ap->a_tvp);
4933 	if (ap->a_tdvp == ap->a_tvp)
4934 		vrele(ap->a_tdvp);
4935 	else
4936 		vput(ap->a_tdvp);
4937 	vrele(ap->a_fdvp);
4938 	vrele(ap->a_fvp);
4939 }
4940 
4941 void
4942 vop_rename_pre(void *ap)
4943 {
4944 	struct vop_rename_args *a = ap;
4945 
4946 #ifdef DEBUG_VFS_LOCKS
4947 	if (a->a_tvp)
4948 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4949 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4950 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4951 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4952 
4953 	/* Check the source (from). */
4954 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4955 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4956 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4957 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4958 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4959 
4960 	/* Check the target. */
4961 	if (a->a_tvp)
4962 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4963 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4964 #endif
4965 	if (a->a_tdvp != a->a_fdvp)
4966 		vhold(a->a_fdvp);
4967 	if (a->a_tvp != a->a_fvp)
4968 		vhold(a->a_fvp);
4969 	vhold(a->a_tdvp);
4970 	if (a->a_tvp)
4971 		vhold(a->a_tvp);
4972 }
4973 
4974 #ifdef DEBUG_VFS_LOCKS
4975 void
4976 vop_strategy_pre(void *ap)
4977 {
4978 	struct vop_strategy_args *a;
4979 	struct buf *bp;
4980 
4981 	a = ap;
4982 	bp = a->a_bp;
4983 
4984 	/*
4985 	 * Cluster ops lock their component buffers but not the IO container.
4986 	 */
4987 	if ((bp->b_flags & B_CLUSTER) != 0)
4988 		return;
4989 
4990 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4991 		if (vfs_badlock_print)
4992 			printf(
4993 			    "VOP_STRATEGY: bp is not locked but should be\n");
4994 		if (vfs_badlock_ddb)
4995 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4996 	}
4997 }
4998 
4999 void
5000 vop_lock_pre(void *ap)
5001 {
5002 	struct vop_lock1_args *a = ap;
5003 
5004 	if ((a->a_flags & LK_INTERLOCK) == 0)
5005 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5006 	else
5007 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5008 }
5009 
5010 void
5011 vop_lock_post(void *ap, int rc)
5012 {
5013 	struct vop_lock1_args *a = ap;
5014 
5015 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5016 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5017 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5018 }
5019 
5020 void
5021 vop_unlock_pre(void *ap)
5022 {
5023 	struct vop_unlock_args *a = ap;
5024 
5025 	if (a->a_flags & LK_INTERLOCK)
5026 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
5027 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5028 }
5029 
5030 void
5031 vop_unlock_post(void *ap, int rc)
5032 {
5033 	struct vop_unlock_args *a = ap;
5034 
5035 	if (a->a_flags & LK_INTERLOCK)
5036 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
5037 }
5038 
5039 void
5040 vop_need_inactive_pre(void *ap)
5041 {
5042 	struct vop_need_inactive_args *a = ap;
5043 
5044 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5045 }
5046 
5047 void
5048 vop_need_inactive_post(void *ap, int rc)
5049 {
5050 	struct vop_need_inactive_args *a = ap;
5051 
5052 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5053 }
5054 #endif
5055 
5056 void
5057 vop_create_post(void *ap, int rc)
5058 {
5059 	struct vop_create_args *a = ap;
5060 
5061 	if (!rc)
5062 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5063 }
5064 
5065 void
5066 vop_deleteextattr_post(void *ap, int rc)
5067 {
5068 	struct vop_deleteextattr_args *a = ap;
5069 
5070 	if (!rc)
5071 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5072 }
5073 
5074 void
5075 vop_link_post(void *ap, int rc)
5076 {
5077 	struct vop_link_args *a = ap;
5078 
5079 	if (!rc) {
5080 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5081 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5082 	}
5083 }
5084 
5085 void
5086 vop_mkdir_post(void *ap, int rc)
5087 {
5088 	struct vop_mkdir_args *a = ap;
5089 
5090 	if (!rc)
5091 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5092 }
5093 
5094 void
5095 vop_mknod_post(void *ap, int rc)
5096 {
5097 	struct vop_mknod_args *a = ap;
5098 
5099 	if (!rc)
5100 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5101 }
5102 
5103 void
5104 vop_reclaim_post(void *ap, int rc)
5105 {
5106 	struct vop_reclaim_args *a = ap;
5107 
5108 	if (!rc)
5109 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5110 }
5111 
5112 void
5113 vop_remove_post(void *ap, int rc)
5114 {
5115 	struct vop_remove_args *a = ap;
5116 
5117 	if (!rc) {
5118 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5119 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5120 	}
5121 }
5122 
5123 void
5124 vop_rename_post(void *ap, int rc)
5125 {
5126 	struct vop_rename_args *a = ap;
5127 	long hint;
5128 
5129 	if (!rc) {
5130 		hint = NOTE_WRITE;
5131 		if (a->a_fdvp == a->a_tdvp) {
5132 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5133 				hint |= NOTE_LINK;
5134 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5135 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5136 		} else {
5137 			hint |= NOTE_EXTEND;
5138 			if (a->a_fvp->v_type == VDIR)
5139 				hint |= NOTE_LINK;
5140 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5141 
5142 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5143 			    a->a_tvp->v_type == VDIR)
5144 				hint &= ~NOTE_LINK;
5145 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5146 		}
5147 
5148 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5149 		if (a->a_tvp)
5150 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5151 	}
5152 	if (a->a_tdvp != a->a_fdvp)
5153 		vdrop(a->a_fdvp);
5154 	if (a->a_tvp != a->a_fvp)
5155 		vdrop(a->a_fvp);
5156 	vdrop(a->a_tdvp);
5157 	if (a->a_tvp)
5158 		vdrop(a->a_tvp);
5159 }
5160 
5161 void
5162 vop_rmdir_post(void *ap, int rc)
5163 {
5164 	struct vop_rmdir_args *a = ap;
5165 
5166 	if (!rc) {
5167 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5168 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5169 	}
5170 }
5171 
5172 void
5173 vop_setattr_post(void *ap, int rc)
5174 {
5175 	struct vop_setattr_args *a = ap;
5176 
5177 	if (!rc)
5178 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5179 }
5180 
5181 void
5182 vop_setextattr_post(void *ap, int rc)
5183 {
5184 	struct vop_setextattr_args *a = ap;
5185 
5186 	if (!rc)
5187 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5188 }
5189 
5190 void
5191 vop_symlink_post(void *ap, int rc)
5192 {
5193 	struct vop_symlink_args *a = ap;
5194 
5195 	if (!rc)
5196 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5197 }
5198 
5199 void
5200 vop_open_post(void *ap, int rc)
5201 {
5202 	struct vop_open_args *a = ap;
5203 
5204 	if (!rc)
5205 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5206 }
5207 
5208 void
5209 vop_close_post(void *ap, int rc)
5210 {
5211 	struct vop_close_args *a = ap;
5212 
5213 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5214 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
5215 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5216 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5217 	}
5218 }
5219 
5220 void
5221 vop_read_post(void *ap, int rc)
5222 {
5223 	struct vop_read_args *a = ap;
5224 
5225 	if (!rc)
5226 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5227 }
5228 
5229 void
5230 vop_readdir_post(void *ap, int rc)
5231 {
5232 	struct vop_readdir_args *a = ap;
5233 
5234 	if (!rc)
5235 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5236 }
5237 
5238 static struct knlist fs_knlist;
5239 
5240 static void
5241 vfs_event_init(void *arg)
5242 {
5243 	knlist_init_mtx(&fs_knlist, NULL);
5244 }
5245 /* XXX - correct order? */
5246 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5247 
5248 void
5249 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5250 {
5251 
5252 	KNOTE_UNLOCKED(&fs_knlist, event);
5253 }
5254 
5255 static int	filt_fsattach(struct knote *kn);
5256 static void	filt_fsdetach(struct knote *kn);
5257 static int	filt_fsevent(struct knote *kn, long hint);
5258 
5259 struct filterops fs_filtops = {
5260 	.f_isfd = 0,
5261 	.f_attach = filt_fsattach,
5262 	.f_detach = filt_fsdetach,
5263 	.f_event = filt_fsevent
5264 };
5265 
5266 static int
5267 filt_fsattach(struct knote *kn)
5268 {
5269 
5270 	kn->kn_flags |= EV_CLEAR;
5271 	knlist_add(&fs_knlist, kn, 0);
5272 	return (0);
5273 }
5274 
5275 static void
5276 filt_fsdetach(struct knote *kn)
5277 {
5278 
5279 	knlist_remove(&fs_knlist, kn, 0);
5280 }
5281 
5282 static int
5283 filt_fsevent(struct knote *kn, long hint)
5284 {
5285 
5286 	kn->kn_fflags |= hint;
5287 	return (kn->kn_fflags != 0);
5288 }
5289 
5290 static int
5291 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5292 {
5293 	struct vfsidctl vc;
5294 	int error;
5295 	struct mount *mp;
5296 
5297 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5298 	if (error)
5299 		return (error);
5300 	if (vc.vc_vers != VFS_CTL_VERS1)
5301 		return (EINVAL);
5302 	mp = vfs_getvfs(&vc.vc_fsid);
5303 	if (mp == NULL)
5304 		return (ENOENT);
5305 	/* ensure that a specific sysctl goes to the right filesystem. */
5306 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5307 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5308 		vfs_rel(mp);
5309 		return (EINVAL);
5310 	}
5311 	VCTLTOREQ(&vc, req);
5312 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5313 	vfs_rel(mp);
5314 	return (error);
5315 }
5316 
5317 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5318     NULL, 0, sysctl_vfs_ctl, "",
5319     "Sysctl by fsid");
5320 
5321 /*
5322  * Function to initialize a va_filerev field sensibly.
5323  * XXX: Wouldn't a random number make a lot more sense ??
5324  */
5325 u_quad_t
5326 init_va_filerev(void)
5327 {
5328 	struct bintime bt;
5329 
5330 	getbinuptime(&bt);
5331 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5332 }
5333 
5334 static int	filt_vfsread(struct knote *kn, long hint);
5335 static int	filt_vfswrite(struct knote *kn, long hint);
5336 static int	filt_vfsvnode(struct knote *kn, long hint);
5337 static void	filt_vfsdetach(struct knote *kn);
5338 static struct filterops vfsread_filtops = {
5339 	.f_isfd = 1,
5340 	.f_detach = filt_vfsdetach,
5341 	.f_event = filt_vfsread
5342 };
5343 static struct filterops vfswrite_filtops = {
5344 	.f_isfd = 1,
5345 	.f_detach = filt_vfsdetach,
5346 	.f_event = filt_vfswrite
5347 };
5348 static struct filterops vfsvnode_filtops = {
5349 	.f_isfd = 1,
5350 	.f_detach = filt_vfsdetach,
5351 	.f_event = filt_vfsvnode
5352 };
5353 
5354 static void
5355 vfs_knllock(void *arg)
5356 {
5357 	struct vnode *vp = arg;
5358 
5359 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5360 }
5361 
5362 static void
5363 vfs_knlunlock(void *arg)
5364 {
5365 	struct vnode *vp = arg;
5366 
5367 	VOP_UNLOCK(vp, 0);
5368 }
5369 
5370 static void
5371 vfs_knl_assert_locked(void *arg)
5372 {
5373 #ifdef DEBUG_VFS_LOCKS
5374 	struct vnode *vp = arg;
5375 
5376 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5377 #endif
5378 }
5379 
5380 static void
5381 vfs_knl_assert_unlocked(void *arg)
5382 {
5383 #ifdef DEBUG_VFS_LOCKS
5384 	struct vnode *vp = arg;
5385 
5386 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5387 #endif
5388 }
5389 
5390 int
5391 vfs_kqfilter(struct vop_kqfilter_args *ap)
5392 {
5393 	struct vnode *vp = ap->a_vp;
5394 	struct knote *kn = ap->a_kn;
5395 	struct knlist *knl;
5396 
5397 	switch (kn->kn_filter) {
5398 	case EVFILT_READ:
5399 		kn->kn_fop = &vfsread_filtops;
5400 		break;
5401 	case EVFILT_WRITE:
5402 		kn->kn_fop = &vfswrite_filtops;
5403 		break;
5404 	case EVFILT_VNODE:
5405 		kn->kn_fop = &vfsvnode_filtops;
5406 		break;
5407 	default:
5408 		return (EINVAL);
5409 	}
5410 
5411 	kn->kn_hook = (caddr_t)vp;
5412 
5413 	v_addpollinfo(vp);
5414 	if (vp->v_pollinfo == NULL)
5415 		return (ENOMEM);
5416 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5417 	vhold(vp);
5418 	knlist_add(knl, kn, 0);
5419 
5420 	return (0);
5421 }
5422 
5423 /*
5424  * Detach knote from vnode
5425  */
5426 static void
5427 filt_vfsdetach(struct knote *kn)
5428 {
5429 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5430 
5431 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5432 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5433 	vdrop(vp);
5434 }
5435 
5436 /*ARGSUSED*/
5437 static int
5438 filt_vfsread(struct knote *kn, long hint)
5439 {
5440 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5441 	struct vattr va;
5442 	int res;
5443 
5444 	/*
5445 	 * filesystem is gone, so set the EOF flag and schedule
5446 	 * the knote for deletion.
5447 	 */
5448 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5449 		VI_LOCK(vp);
5450 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5451 		VI_UNLOCK(vp);
5452 		return (1);
5453 	}
5454 
5455 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5456 		return (0);
5457 
5458 	VI_LOCK(vp);
5459 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5460 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5461 	VI_UNLOCK(vp);
5462 	return (res);
5463 }
5464 
5465 /*ARGSUSED*/
5466 static int
5467 filt_vfswrite(struct knote *kn, long hint)
5468 {
5469 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5470 
5471 	VI_LOCK(vp);
5472 
5473 	/*
5474 	 * filesystem is gone, so set the EOF flag and schedule
5475 	 * the knote for deletion.
5476 	 */
5477 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5478 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5479 
5480 	kn->kn_data = 0;
5481 	VI_UNLOCK(vp);
5482 	return (1);
5483 }
5484 
5485 static int
5486 filt_vfsvnode(struct knote *kn, long hint)
5487 {
5488 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5489 	int res;
5490 
5491 	VI_LOCK(vp);
5492 	if (kn->kn_sfflags & hint)
5493 		kn->kn_fflags |= hint;
5494 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5495 		kn->kn_flags |= EV_EOF;
5496 		VI_UNLOCK(vp);
5497 		return (1);
5498 	}
5499 	res = (kn->kn_fflags != 0);
5500 	VI_UNLOCK(vp);
5501 	return (res);
5502 }
5503 
5504 int
5505 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5506 {
5507 	int error;
5508 
5509 	if (dp->d_reclen > ap->a_uio->uio_resid)
5510 		return (ENAMETOOLONG);
5511 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5512 	if (error) {
5513 		if (ap->a_ncookies != NULL) {
5514 			if (ap->a_cookies != NULL)
5515 				free(ap->a_cookies, M_TEMP);
5516 			ap->a_cookies = NULL;
5517 			*ap->a_ncookies = 0;
5518 		}
5519 		return (error);
5520 	}
5521 	if (ap->a_ncookies == NULL)
5522 		return (0);
5523 
5524 	KASSERT(ap->a_cookies,
5525 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5526 
5527 	*ap->a_cookies = realloc(*ap->a_cookies,
5528 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5529 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5530 	*ap->a_ncookies += 1;
5531 	return (0);
5532 }
5533 
5534 /*
5535  * Mark for update the access time of the file if the filesystem
5536  * supports VOP_MARKATIME.  This functionality is used by execve and
5537  * mmap, so we want to avoid the I/O implied by directly setting
5538  * va_atime for the sake of efficiency.
5539  */
5540 void
5541 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5542 {
5543 	struct mount *mp;
5544 
5545 	mp = vp->v_mount;
5546 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5547 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5548 		(void)VOP_MARKATIME(vp);
5549 }
5550 
5551 /*
5552  * The purpose of this routine is to remove granularity from accmode_t,
5553  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5554  * VADMIN and VAPPEND.
5555  *
5556  * If it returns 0, the caller is supposed to continue with the usual
5557  * access checks using 'accmode' as modified by this routine.  If it
5558  * returns nonzero value, the caller is supposed to return that value
5559  * as errno.
5560  *
5561  * Note that after this routine runs, accmode may be zero.
5562  */
5563 int
5564 vfs_unixify_accmode(accmode_t *accmode)
5565 {
5566 	/*
5567 	 * There is no way to specify explicit "deny" rule using
5568 	 * file mode or POSIX.1e ACLs.
5569 	 */
5570 	if (*accmode & VEXPLICIT_DENY) {
5571 		*accmode = 0;
5572 		return (0);
5573 	}
5574 
5575 	/*
5576 	 * None of these can be translated into usual access bits.
5577 	 * Also, the common case for NFSv4 ACLs is to not contain
5578 	 * either of these bits. Caller should check for VWRITE
5579 	 * on the containing directory instead.
5580 	 */
5581 	if (*accmode & (VDELETE_CHILD | VDELETE))
5582 		return (EPERM);
5583 
5584 	if (*accmode & VADMIN_PERMS) {
5585 		*accmode &= ~VADMIN_PERMS;
5586 		*accmode |= VADMIN;
5587 	}
5588 
5589 	/*
5590 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5591 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5592 	 */
5593 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5594 
5595 	return (0);
5596 }
5597 
5598 /*
5599  * These are helper functions for filesystems to traverse all
5600  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5601  *
5602  * This interface replaces MNT_VNODE_FOREACH.
5603  */
5604 
5605 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5606 
5607 struct vnode *
5608 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5609 {
5610 	struct vnode *vp;
5611 
5612 	if (should_yield())
5613 		kern_yield(PRI_USER);
5614 	MNT_ILOCK(mp);
5615 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5616 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5617 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5618 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5619 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5620 			continue;
5621 		VI_LOCK(vp);
5622 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5623 			VI_UNLOCK(vp);
5624 			continue;
5625 		}
5626 		break;
5627 	}
5628 	if (vp == NULL) {
5629 		__mnt_vnode_markerfree_all(mvp, mp);
5630 		/* MNT_IUNLOCK(mp); -- done in above function */
5631 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5632 		return (NULL);
5633 	}
5634 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5635 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5636 	MNT_IUNLOCK(mp);
5637 	return (vp);
5638 }
5639 
5640 struct vnode *
5641 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5642 {
5643 	struct vnode *vp;
5644 
5645 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5646 	MNT_ILOCK(mp);
5647 	MNT_REF(mp);
5648 	(*mvp)->v_mount = mp;
5649 	(*mvp)->v_type = VMARKER;
5650 
5651 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5652 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5653 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5654 			continue;
5655 		VI_LOCK(vp);
5656 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5657 			VI_UNLOCK(vp);
5658 			continue;
5659 		}
5660 		break;
5661 	}
5662 	if (vp == NULL) {
5663 		MNT_REL(mp);
5664 		MNT_IUNLOCK(mp);
5665 		free(*mvp, M_VNODE_MARKER);
5666 		*mvp = NULL;
5667 		return (NULL);
5668 	}
5669 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5670 	MNT_IUNLOCK(mp);
5671 	return (vp);
5672 }
5673 
5674 void
5675 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5676 {
5677 
5678 	if (*mvp == NULL) {
5679 		MNT_IUNLOCK(mp);
5680 		return;
5681 	}
5682 
5683 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5684 
5685 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5686 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5687 	MNT_REL(mp);
5688 	MNT_IUNLOCK(mp);
5689 	free(*mvp, M_VNODE_MARKER);
5690 	*mvp = NULL;
5691 }
5692 
5693 /*
5694  * These are helper functions for filesystems to traverse their
5695  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5696  */
5697 static void
5698 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5699 {
5700 
5701 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5702 
5703 	MNT_ILOCK(mp);
5704 	MNT_REL(mp);
5705 	MNT_IUNLOCK(mp);
5706 	free(*mvp, M_VNODE_MARKER);
5707 	*mvp = NULL;
5708 }
5709 
5710 /*
5711  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5712  * conventional lock order during mnt_vnode_next_active iteration.
5713  *
5714  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5715  * The list lock is dropped and reacquired.  On success, both locks are held.
5716  * On failure, the mount vnode list lock is held but the vnode interlock is
5717  * not, and the procedure may have yielded.
5718  */
5719 static bool
5720 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5721     struct vnode *vp)
5722 {
5723 	const struct vnode *tmp;
5724 	bool held, ret;
5725 
5726 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5727 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5728 	    ("%s: bad marker", __func__));
5729 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5730 	    ("%s: inappropriate vnode", __func__));
5731 	ASSERT_VI_UNLOCKED(vp, __func__);
5732 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5733 
5734 	ret = false;
5735 
5736 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5737 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5738 
5739 	/*
5740 	 * Use a hold to prevent vp from disappearing while the mount vnode
5741 	 * list lock is dropped and reacquired.  Normally a hold would be
5742 	 * acquired with vhold(), but that might try to acquire the vnode
5743 	 * interlock, which would be a LOR with the mount vnode list lock.
5744 	 */
5745 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5746 	mtx_unlock(&mp->mnt_listmtx);
5747 	if (!held)
5748 		goto abort;
5749 	VI_LOCK(vp);
5750 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5751 		vdropl(vp);
5752 		goto abort;
5753 	}
5754 	mtx_lock(&mp->mnt_listmtx);
5755 
5756 	/*
5757 	 * Determine whether the vnode is still the next one after the marker,
5758 	 * excepting any other markers.  If the vnode has not been doomed by
5759 	 * vgone() then the hold should have ensured that it remained on the
5760 	 * active list.  If it has been doomed but is still on the active list,
5761 	 * don't abort, but rather skip over it (avoid spinning on doomed
5762 	 * vnodes).
5763 	 */
5764 	tmp = mvp;
5765 	do {
5766 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5767 	} while (tmp != NULL && tmp->v_type == VMARKER);
5768 	if (tmp != vp) {
5769 		mtx_unlock(&mp->mnt_listmtx);
5770 		VI_UNLOCK(vp);
5771 		goto abort;
5772 	}
5773 
5774 	ret = true;
5775 	goto out;
5776 abort:
5777 	maybe_yield();
5778 	mtx_lock(&mp->mnt_listmtx);
5779 out:
5780 	if (ret)
5781 		ASSERT_VI_LOCKED(vp, __func__);
5782 	else
5783 		ASSERT_VI_UNLOCKED(vp, __func__);
5784 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5785 	return (ret);
5786 }
5787 
5788 static struct vnode *
5789 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5790 {
5791 	struct vnode *vp, *nvp;
5792 
5793 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5794 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5795 restart:
5796 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5797 	while (vp != NULL) {
5798 		if (vp->v_type == VMARKER) {
5799 			vp = TAILQ_NEXT(vp, v_actfreelist);
5800 			continue;
5801 		}
5802 		/*
5803 		 * Try-lock because this is the wrong lock order.  If that does
5804 		 * not succeed, drop the mount vnode list lock and try to
5805 		 * reacquire it and the vnode interlock in the right order.
5806 		 */
5807 		if (!VI_TRYLOCK(vp) &&
5808 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5809 			goto restart;
5810 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5811 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5812 		    ("alien vnode on the active list %p %p", vp, mp));
5813 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5814 			break;
5815 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5816 		VI_UNLOCK(vp);
5817 		vp = nvp;
5818 	}
5819 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5820 
5821 	/* Check if we are done */
5822 	if (vp == NULL) {
5823 		mtx_unlock(&mp->mnt_listmtx);
5824 		mnt_vnode_markerfree_active(mvp, mp);
5825 		return (NULL);
5826 	}
5827 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5828 	mtx_unlock(&mp->mnt_listmtx);
5829 	ASSERT_VI_LOCKED(vp, "active iter");
5830 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5831 	return (vp);
5832 }
5833 
5834 struct vnode *
5835 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5836 {
5837 
5838 	if (should_yield())
5839 		kern_yield(PRI_USER);
5840 	mtx_lock(&mp->mnt_listmtx);
5841 	return (mnt_vnode_next_active(mvp, mp));
5842 }
5843 
5844 struct vnode *
5845 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5846 {
5847 	struct vnode *vp;
5848 
5849 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5850 	MNT_ILOCK(mp);
5851 	MNT_REF(mp);
5852 	MNT_IUNLOCK(mp);
5853 	(*mvp)->v_type = VMARKER;
5854 	(*mvp)->v_mount = mp;
5855 
5856 	mtx_lock(&mp->mnt_listmtx);
5857 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5858 	if (vp == NULL) {
5859 		mtx_unlock(&mp->mnt_listmtx);
5860 		mnt_vnode_markerfree_active(mvp, mp);
5861 		return (NULL);
5862 	}
5863 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5864 	return (mnt_vnode_next_active(mvp, mp));
5865 }
5866 
5867 void
5868 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5869 {
5870 
5871 	if (*mvp == NULL)
5872 		return;
5873 
5874 	mtx_lock(&mp->mnt_listmtx);
5875 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5876 	mtx_unlock(&mp->mnt_listmtx);
5877 	mnt_vnode_markerfree_active(mvp, mp);
5878 }
5879