xref: /freebsd/sys/kern/vfs_subr.c (revision 86aa9539fef591a363b06a0ebd3aa7a07f4c1579)
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_usecount(struct vnode *);
111 static void	v_incr_usecount_locked(struct vnode *);
112 static void	v_incr_devcount(struct vnode *);
113 static void	v_decr_devcount(struct vnode *);
114 static void	vgonel(struct vnode *);
115 static void	vfs_knllock(void *arg);
116 static void	vfs_knlunlock(void *arg);
117 static void	vfs_knl_assert_locked(void *arg);
118 static void	vfs_knl_assert_unlocked(void *arg);
119 static void	vnlru_return_batches(struct vfsops *mnt_op);
120 static void	destroy_vpollinfo(struct vpollinfo *vi);
121 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
122 		    daddr_t startlbn, daddr_t endlbn);
123 
124 /*
125  * These fences are intended for cases where some synchronization is
126  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
127  * and v_usecount) updates.  Access to v_iflags is generally synchronized
128  * by the interlock, but we have some internal assertions that check vnode
129  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
130  * for now.
131  */
132 #ifdef INVARIANTS
133 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
134 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
135 #else
136 #define	VNODE_REFCOUNT_FENCE_ACQ()
137 #define	VNODE_REFCOUNT_FENCE_REL()
138 #endif
139 
140 /*
141  * Number of vnodes in existence.  Increased whenever getnewvnode()
142  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
143  */
144 static unsigned long	numvnodes;
145 
146 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
147     "Number of vnodes in existence");
148 
149 static counter_u64_t vnodes_created;
150 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
151     "Number of vnodes created by getnewvnode");
152 
153 static u_long mnt_free_list_batch = 128;
154 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
155     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
156 
157 /*
158  * Conversion tables for conversion from vnode types to inode formats
159  * and back.
160  */
161 enum vtype iftovt_tab[16] = {
162 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
163 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
164 };
165 int vttoif_tab[10] = {
166 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
167 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
168 };
169 
170 /*
171  * List of vnodes that are ready for recycling.
172  */
173 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
174 
175 /*
176  * "Free" vnode target.  Free vnodes are rarely completely free, but are
177  * just ones that are cheap to recycle.  Usually they are for files which
178  * have been stat'd but not read; these usually have inode and namecache
179  * data attached to them.  This target is the preferred minimum size of a
180  * sub-cache consisting mostly of such files. The system balances the size
181  * of this sub-cache with its complement to try to prevent either from
182  * thrashing while the other is relatively inactive.  The targets express
183  * a preference for the best balance.
184  *
185  * "Above" this target there are 2 further targets (watermarks) related
186  * to recyling of free vnodes.  In the best-operating case, the cache is
187  * exactly full, the free list has size between vlowat and vhiwat above the
188  * free target, and recycling from it and normal use maintains this state.
189  * Sometimes the free list is below vlowat or even empty, but this state
190  * is even better for immediate use provided the cache is not full.
191  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
192  * ones) to reach one of these states.  The watermarks are currently hard-
193  * coded as 4% and 9% of the available space higher.  These and the default
194  * of 25% for wantfreevnodes are too large if the memory size is large.
195  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
196  * whenever vnlru_proc() becomes active.
197  */
198 static u_long wantfreevnodes;
199 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
200     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
201 static u_long freevnodes;
202 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
203     &freevnodes, 0, "Number of \"free\" vnodes");
204 
205 static counter_u64_t recycles_count;
206 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
207     "Number of vnodes recycled to meet vnode cache targets");
208 
209 /*
210  * Various variables used for debugging the new implementation of
211  * reassignbuf().
212  * XXX these are probably of (very) limited utility now.
213  */
214 static int reassignbufcalls;
215 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
216     "Number of calls to reassignbuf");
217 
218 static counter_u64_t free_owe_inact;
219 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
220     "Number of times free vnodes kept on active list due to VFS "
221     "owing inactivation");
222 
223 /* To keep more than one thread at a time from running vfs_getnewfsid */
224 static struct mtx mntid_mtx;
225 
226 /*
227  * Lock for any access to the following:
228  *	vnode_free_list
229  *	numvnodes
230  *	freevnodes
231  */
232 static struct mtx vnode_free_list_mtx;
233 
234 /* Publicly exported FS */
235 struct nfs_public nfs_pub;
236 
237 static uma_zone_t buf_trie_zone;
238 
239 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
240 static uma_zone_t vnode_zone;
241 static uma_zone_t vnodepoll_zone;
242 
243 /*
244  * The workitem queue.
245  *
246  * It is useful to delay writes of file data and filesystem metadata
247  * for tens of seconds so that quickly created and deleted files need
248  * not waste disk bandwidth being created and removed. To realize this,
249  * we append vnodes to a "workitem" queue. When running with a soft
250  * updates implementation, most pending metadata dependencies should
251  * not wait for more than a few seconds. Thus, mounted on block devices
252  * are delayed only about a half the time that file data is delayed.
253  * Similarly, directory updates are more critical, so are only delayed
254  * about a third the time that file data is delayed. Thus, there are
255  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
256  * one each second (driven off the filesystem syncer process). The
257  * syncer_delayno variable indicates the next queue that is to be processed.
258  * Items that need to be processed soon are placed in this queue:
259  *
260  *	syncer_workitem_pending[syncer_delayno]
261  *
262  * A delay of fifteen seconds is done by placing the request fifteen
263  * entries later in the queue:
264  *
265  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
266  *
267  */
268 static int syncer_delayno;
269 static long syncer_mask;
270 LIST_HEAD(synclist, bufobj);
271 static struct synclist *syncer_workitem_pending;
272 /*
273  * The sync_mtx protects:
274  *	bo->bo_synclist
275  *	sync_vnode_count
276  *	syncer_delayno
277  *	syncer_state
278  *	syncer_workitem_pending
279  *	syncer_worklist_len
280  *	rushjob
281  */
282 static struct mtx sync_mtx;
283 static struct cv sync_wakeup;
284 
285 #define SYNCER_MAXDELAY		32
286 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
287 static int syncdelay = 30;		/* max time to delay syncing data */
288 static int filedelay = 30;		/* time to delay syncing files */
289 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
290     "Time to delay syncing files (in seconds)");
291 static int dirdelay = 29;		/* time to delay syncing directories */
292 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
293     "Time to delay syncing directories (in seconds)");
294 static int metadelay = 28;		/* time to delay syncing metadata */
295 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
296     "Time to delay syncing metadata (in seconds)");
297 static int rushjob;		/* number of slots to run ASAP */
298 static int stat_rush_requests;	/* number of times I/O speeded up */
299 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
300     "Number of times I/O speeded up (rush requests)");
301 
302 /*
303  * When shutting down the syncer, run it at four times normal speed.
304  */
305 #define SYNCER_SHUTDOWN_SPEEDUP		4
306 static int sync_vnode_count;
307 static int syncer_worklist_len;
308 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
309     syncer_state;
310 
311 /* Target for maximum number of vnodes. */
312 int desiredvnodes;
313 static int gapvnodes;		/* gap between wanted and desired */
314 static int vhiwat;		/* enough extras after expansion */
315 static int vlowat;		/* minimal extras before expansion */
316 static int vstir;		/* nonzero to stir non-free vnodes */
317 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
318 
319 static int
320 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
321 {
322 	int error, old_desiredvnodes;
323 
324 	old_desiredvnodes = desiredvnodes;
325 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
326 		return (error);
327 	if (old_desiredvnodes != desiredvnodes) {
328 		wantfreevnodes = desiredvnodes / 4;
329 		/* XXX locking seems to be incomplete. */
330 		vfs_hash_changesize(desiredvnodes);
331 		cache_changesize(desiredvnodes);
332 	}
333 	return (0);
334 }
335 
336 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
337     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
338     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
339 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
340     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
341 static int vnlru_nowhere;
342 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
343     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
344 
345 static int
346 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
347 {
348 	struct vnode *vp;
349 	struct nameidata nd;
350 	char *buf;
351 	unsigned long ndflags;
352 	int error;
353 
354 	if (req->newptr == NULL)
355 		return (EINVAL);
356 	if (req->newlen > PATH_MAX)
357 		return (E2BIG);
358 
359 	buf = malloc(PATH_MAX + 1, M_TEMP, M_WAITOK);
360 	error = SYSCTL_IN(req, buf, req->newlen);
361 	if (error != 0)
362 		goto out;
363 
364 	buf[req->newlen] = '\0';
365 
366 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
367 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
368 	if ((error = namei(&nd)) != 0)
369 		goto out;
370 	vp = nd.ni_vp;
371 
372 	if ((vp->v_iflag & VI_DOOMED) != 0) {
373 		/*
374 		 * This vnode is being recycled.  Return != 0 to let the caller
375 		 * know that the sysctl had no effect.  Return EAGAIN because a
376 		 * subsequent call will likely succeed (since namei will create
377 		 * a new vnode if necessary)
378 		 */
379 		error = EAGAIN;
380 		goto putvnode;
381 	}
382 
383 	counter_u64_add(recycles_count, 1);
384 	vgone(vp);
385 putvnode:
386 	NDFREE(&nd, 0);
387 out:
388 	free(buf, M_TEMP);
389 	return (error);
390 }
391 
392 static int
393 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
394 {
395 	struct thread *td = curthread;
396 	struct vnode *vp;
397 	struct file *fp;
398 	int error;
399 	int fd;
400 
401 	if (req->newptr == NULL)
402 		return (EBADF);
403 
404         error = sysctl_handle_int(oidp, &fd, 0, req);
405         if (error != 0)
406                 return (error);
407 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
408 	if (error != 0)
409 		return (error);
410 	vp = fp->f_vnode;
411 
412 	error = vn_lock(vp, LK_EXCLUSIVE);
413 	if (error != 0)
414 		goto drop;
415 
416 	counter_u64_add(recycles_count, 1);
417 	vgone(vp);
418 	VOP_UNLOCK(vp, 0);
419 drop:
420 	fdrop(fp, td);
421 	return (error);
422 }
423 
424 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
425     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
426     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
427 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
428     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
429     sysctl_ftry_reclaim_vnode, "I",
430     "Try to reclaim a vnode by its file descriptor");
431 
432 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
433 static int vnsz2log;
434 
435 /*
436  * Support for the bufobj clean & dirty pctrie.
437  */
438 static void *
439 buf_trie_alloc(struct pctrie *ptree)
440 {
441 
442 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
443 }
444 
445 static void
446 buf_trie_free(struct pctrie *ptree, void *node)
447 {
448 
449 	uma_zfree(buf_trie_zone, node);
450 }
451 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
452 
453 /*
454  * Initialize the vnode management data structures.
455  *
456  * Reevaluate the following cap on the number of vnodes after the physical
457  * memory size exceeds 512GB.  In the limit, as the physical memory size
458  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
459  */
460 #ifndef	MAXVNODES_MAX
461 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
462 #endif
463 
464 /*
465  * Initialize a vnode as it first enters the zone.
466  */
467 static int
468 vnode_init(void *mem, int size, int flags)
469 {
470 	struct vnode *vp;
471 
472 	vp = mem;
473 	bzero(vp, size);
474 	/*
475 	 * Setup locks.
476 	 */
477 	vp->v_vnlock = &vp->v_lock;
478 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
479 	/*
480 	 * By default, don't allow shared locks unless filesystems opt-in.
481 	 */
482 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
483 	    LK_NOSHARE | LK_IS_VNODE);
484 	/*
485 	 * Initialize bufobj.
486 	 */
487 	bufobj_init(&vp->v_bufobj, vp);
488 	/*
489 	 * Initialize namecache.
490 	 */
491 	LIST_INIT(&vp->v_cache_src);
492 	TAILQ_INIT(&vp->v_cache_dst);
493 	/*
494 	 * Initialize rangelocks.
495 	 */
496 	rangelock_init(&vp->v_rl);
497 	return (0);
498 }
499 
500 /*
501  * Free a vnode when it is cleared from the zone.
502  */
503 static void
504 vnode_fini(void *mem, int size)
505 {
506 	struct vnode *vp;
507 	struct bufobj *bo;
508 
509 	vp = mem;
510 	rangelock_destroy(&vp->v_rl);
511 	lockdestroy(vp->v_vnlock);
512 	mtx_destroy(&vp->v_interlock);
513 	bo = &vp->v_bufobj;
514 	rw_destroy(BO_LOCKPTR(bo));
515 }
516 
517 /*
518  * Provide the size of NFS nclnode and NFS fh for calculation of the
519  * vnode memory consumption.  The size is specified directly to
520  * eliminate dependency on NFS-private header.
521  *
522  * Other filesystems may use bigger or smaller (like UFS and ZFS)
523  * private inode data, but the NFS-based estimation is ample enough.
524  * Still, we care about differences in the size between 64- and 32-bit
525  * platforms.
526  *
527  * Namecache structure size is heuristically
528  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
529  */
530 #ifdef _LP64
531 #define	NFS_NCLNODE_SZ	(528 + 64)
532 #define	NC_SZ		148
533 #else
534 #define	NFS_NCLNODE_SZ	(360 + 32)
535 #define	NC_SZ		92
536 #endif
537 
538 static void
539 vntblinit(void *dummy __unused)
540 {
541 	u_int i;
542 	int physvnodes, virtvnodes;
543 
544 	/*
545 	 * Desiredvnodes is a function of the physical memory size and the
546 	 * kernel's heap size.  Generally speaking, it scales with the
547 	 * physical memory size.  The ratio of desiredvnodes to the physical
548 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
549 	 * Thereafter, the
550 	 * marginal ratio of desiredvnodes to the physical memory size is
551 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
552 	 * size.  The memory required by desiredvnodes vnodes and vm objects
553 	 * must not exceed 1/10th of the kernel's heap size.
554 	 */
555 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
556 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
557 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
558 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
559 	desiredvnodes = min(physvnodes, virtvnodes);
560 	if (desiredvnodes > MAXVNODES_MAX) {
561 		if (bootverbose)
562 			printf("Reducing kern.maxvnodes %d -> %d\n",
563 			    desiredvnodes, MAXVNODES_MAX);
564 		desiredvnodes = MAXVNODES_MAX;
565 	}
566 	wantfreevnodes = desiredvnodes / 4;
567 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
568 	TAILQ_INIT(&vnode_free_list);
569 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
570 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
571 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
572 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
573 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
574 	/*
575 	 * Preallocate enough nodes to support one-per buf so that
576 	 * we can not fail an insert.  reassignbuf() callers can not
577 	 * tolerate the insertion failure.
578 	 */
579 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
580 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
581 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
582 	uma_prealloc(buf_trie_zone, nbuf);
583 
584 	vnodes_created = counter_u64_alloc(M_WAITOK);
585 	recycles_count = counter_u64_alloc(M_WAITOK);
586 	free_owe_inact = counter_u64_alloc(M_WAITOK);
587 
588 	/*
589 	 * Initialize the filesystem syncer.
590 	 */
591 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
592 	    &syncer_mask);
593 	syncer_maxdelay = syncer_mask + 1;
594 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
595 	cv_init(&sync_wakeup, "syncer");
596 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
597 		vnsz2log++;
598 	vnsz2log--;
599 }
600 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
601 
602 
603 /*
604  * Mark a mount point as busy. Used to synchronize access and to delay
605  * unmounting. Eventually, mountlist_mtx is not released on failure.
606  *
607  * vfs_busy() is a custom lock, it can block the caller.
608  * vfs_busy() only sleeps if the unmount is active on the mount point.
609  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
610  * vnode belonging to mp.
611  *
612  * Lookup uses vfs_busy() to traverse mount points.
613  * root fs			var fs
614  * / vnode lock		A	/ vnode lock (/var)		D
615  * /var vnode lock	B	/log vnode lock(/var/log)	E
616  * vfs_busy lock	C	vfs_busy lock			F
617  *
618  * Within each file system, the lock order is C->A->B and F->D->E.
619  *
620  * When traversing across mounts, the system follows that lock order:
621  *
622  *        C->A->B
623  *              |
624  *              +->F->D->E
625  *
626  * The lookup() process for namei("/var") illustrates the process:
627  *  VOP_LOOKUP() obtains B while A is held
628  *  vfs_busy() obtains a shared lock on F while A and B are held
629  *  vput() releases lock on B
630  *  vput() releases lock on A
631  *  VFS_ROOT() obtains lock on D while shared lock on F is held
632  *  vfs_unbusy() releases shared lock on F
633  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
634  *    Attempt to lock A (instead of vp_crossmp) while D is held would
635  *    violate the global order, causing deadlocks.
636  *
637  * dounmount() locks B while F is drained.
638  */
639 int
640 vfs_busy(struct mount *mp, int flags)
641 {
642 
643 	MPASS((flags & ~MBF_MASK) == 0);
644 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
645 
646 	MNT_ILOCK(mp);
647 	MNT_REF(mp);
648 	/*
649 	 * If mount point is currently being unmounted, sleep until the
650 	 * mount point fate is decided.  If thread doing the unmounting fails,
651 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
652 	 * that this mount point has survived the unmount attempt and vfs_busy
653 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
654 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
655 	 * about to be really destroyed.  vfs_busy needs to release its
656 	 * reference on the mount point in this case and return with ENOENT,
657 	 * telling the caller that mount mount it tried to busy is no longer
658 	 * valid.
659 	 */
660 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
661 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
662 			MNT_REL(mp);
663 			MNT_IUNLOCK(mp);
664 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
665 			    __func__);
666 			return (ENOENT);
667 		}
668 		if (flags & MBF_MNTLSTLOCK)
669 			mtx_unlock(&mountlist_mtx);
670 		mp->mnt_kern_flag |= MNTK_MWAIT;
671 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
672 		if (flags & MBF_MNTLSTLOCK)
673 			mtx_lock(&mountlist_mtx);
674 		MNT_ILOCK(mp);
675 	}
676 	if (flags & MBF_MNTLSTLOCK)
677 		mtx_unlock(&mountlist_mtx);
678 	mp->mnt_lockref++;
679 	MNT_IUNLOCK(mp);
680 	return (0);
681 }
682 
683 /*
684  * Free a busy filesystem.
685  */
686 void
687 vfs_unbusy(struct mount *mp)
688 {
689 
690 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
691 	MNT_ILOCK(mp);
692 	MNT_REL(mp);
693 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
694 	mp->mnt_lockref--;
695 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
696 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
697 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
698 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
699 		wakeup(&mp->mnt_lockref);
700 	}
701 	MNT_IUNLOCK(mp);
702 }
703 
704 /*
705  * Lookup a mount point by filesystem identifier.
706  */
707 struct mount *
708 vfs_getvfs(fsid_t *fsid)
709 {
710 	struct mount *mp;
711 
712 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
713 	mtx_lock(&mountlist_mtx);
714 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
715 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
716 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
717 			vfs_ref(mp);
718 			mtx_unlock(&mountlist_mtx);
719 			return (mp);
720 		}
721 	}
722 	mtx_unlock(&mountlist_mtx);
723 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
724 	return ((struct mount *) 0);
725 }
726 
727 /*
728  * Lookup a mount point by filesystem identifier, busying it before
729  * returning.
730  *
731  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
732  * cache for popular filesystem identifiers.  The cache is lockess, using
733  * the fact that struct mount's are never freed.  In worst case we may
734  * get pointer to unmounted or even different filesystem, so we have to
735  * check what we got, and go slow way if so.
736  */
737 struct mount *
738 vfs_busyfs(fsid_t *fsid)
739 {
740 #define	FSID_CACHE_SIZE	256
741 	typedef struct mount * volatile vmp_t;
742 	static vmp_t cache[FSID_CACHE_SIZE];
743 	struct mount *mp;
744 	int error;
745 	uint32_t hash;
746 
747 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
748 	hash = fsid->val[0] ^ fsid->val[1];
749 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
750 	mp = cache[hash];
751 	if (mp == NULL ||
752 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
753 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
754 		goto slow;
755 	if (vfs_busy(mp, 0) != 0) {
756 		cache[hash] = NULL;
757 		goto slow;
758 	}
759 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
760 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
761 		return (mp);
762 	else
763 	    vfs_unbusy(mp);
764 
765 slow:
766 	mtx_lock(&mountlist_mtx);
767 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
768 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
769 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
770 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
771 			if (error) {
772 				cache[hash] = NULL;
773 				mtx_unlock(&mountlist_mtx);
774 				return (NULL);
775 			}
776 			cache[hash] = mp;
777 			return (mp);
778 		}
779 	}
780 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
781 	mtx_unlock(&mountlist_mtx);
782 	return ((struct mount *) 0);
783 }
784 
785 /*
786  * Check if a user can access privileged mount options.
787  */
788 int
789 vfs_suser(struct mount *mp, struct thread *td)
790 {
791 	int error;
792 
793 	if (jailed(td->td_ucred)) {
794 		/*
795 		 * If the jail of the calling thread lacks permission for
796 		 * this type of file system, deny immediately.
797 		 */
798 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
799 			return (EPERM);
800 
801 		/*
802 		 * If the file system was mounted outside the jail of the
803 		 * calling thread, deny immediately.
804 		 */
805 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
806 			return (EPERM);
807 	}
808 
809 	/*
810 	 * If file system supports delegated administration, we don't check
811 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
812 	 * by the file system itself.
813 	 * If this is not the user that did original mount, we check for
814 	 * the PRIV_VFS_MOUNT_OWNER privilege.
815 	 */
816 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
817 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
818 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
819 			return (error);
820 	}
821 	return (0);
822 }
823 
824 /*
825  * Get a new unique fsid.  Try to make its val[0] unique, since this value
826  * will be used to create fake device numbers for stat().  Also try (but
827  * not so hard) make its val[0] unique mod 2^16, since some emulators only
828  * support 16-bit device numbers.  We end up with unique val[0]'s for the
829  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
830  *
831  * Keep in mind that several mounts may be running in parallel.  Starting
832  * the search one past where the previous search terminated is both a
833  * micro-optimization and a defense against returning the same fsid to
834  * different mounts.
835  */
836 void
837 vfs_getnewfsid(struct mount *mp)
838 {
839 	static uint16_t mntid_base;
840 	struct mount *nmp;
841 	fsid_t tfsid;
842 	int mtype;
843 
844 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
845 	mtx_lock(&mntid_mtx);
846 	mtype = mp->mnt_vfc->vfc_typenum;
847 	tfsid.val[1] = mtype;
848 	mtype = (mtype & 0xFF) << 24;
849 	for (;;) {
850 		tfsid.val[0] = makedev(255,
851 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
852 		mntid_base++;
853 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
854 			break;
855 		vfs_rel(nmp);
856 	}
857 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
858 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
859 	mtx_unlock(&mntid_mtx);
860 }
861 
862 /*
863  * Knob to control the precision of file timestamps:
864  *
865  *   0 = seconds only; nanoseconds zeroed.
866  *   1 = seconds and nanoseconds, accurate within 1/HZ.
867  *   2 = seconds and nanoseconds, truncated to microseconds.
868  * >=3 = seconds and nanoseconds, maximum precision.
869  */
870 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
871 
872 static int timestamp_precision = TSP_USEC;
873 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
874     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
875     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
876     "3+: sec + ns (max. precision))");
877 
878 /*
879  * Get a current timestamp.
880  */
881 void
882 vfs_timestamp(struct timespec *tsp)
883 {
884 	struct timeval tv;
885 
886 	switch (timestamp_precision) {
887 	case TSP_SEC:
888 		tsp->tv_sec = time_second;
889 		tsp->tv_nsec = 0;
890 		break;
891 	case TSP_HZ:
892 		getnanotime(tsp);
893 		break;
894 	case TSP_USEC:
895 		microtime(&tv);
896 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
897 		break;
898 	case TSP_NSEC:
899 	default:
900 		nanotime(tsp);
901 		break;
902 	}
903 }
904 
905 /*
906  * Set vnode attributes to VNOVAL
907  */
908 void
909 vattr_null(struct vattr *vap)
910 {
911 
912 	vap->va_type = VNON;
913 	vap->va_size = VNOVAL;
914 	vap->va_bytes = VNOVAL;
915 	vap->va_mode = VNOVAL;
916 	vap->va_nlink = VNOVAL;
917 	vap->va_uid = VNOVAL;
918 	vap->va_gid = VNOVAL;
919 	vap->va_fsid = VNOVAL;
920 	vap->va_fileid = VNOVAL;
921 	vap->va_blocksize = VNOVAL;
922 	vap->va_rdev = VNOVAL;
923 	vap->va_atime.tv_sec = VNOVAL;
924 	vap->va_atime.tv_nsec = VNOVAL;
925 	vap->va_mtime.tv_sec = VNOVAL;
926 	vap->va_mtime.tv_nsec = VNOVAL;
927 	vap->va_ctime.tv_sec = VNOVAL;
928 	vap->va_ctime.tv_nsec = VNOVAL;
929 	vap->va_birthtime.tv_sec = VNOVAL;
930 	vap->va_birthtime.tv_nsec = VNOVAL;
931 	vap->va_flags = VNOVAL;
932 	vap->va_gen = VNOVAL;
933 	vap->va_vaflags = 0;
934 }
935 
936 /*
937  * This routine is called when we have too many vnodes.  It attempts
938  * to free <count> vnodes and will potentially free vnodes that still
939  * have VM backing store (VM backing store is typically the cause
940  * of a vnode blowout so we want to do this).  Therefore, this operation
941  * is not considered cheap.
942  *
943  * A number of conditions may prevent a vnode from being reclaimed.
944  * the buffer cache may have references on the vnode, a directory
945  * vnode may still have references due to the namei cache representing
946  * underlying files, or the vnode may be in active use.   It is not
947  * desirable to reuse such vnodes.  These conditions may cause the
948  * number of vnodes to reach some minimum value regardless of what
949  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
950  *
951  * @param mp		 Try to reclaim vnodes from this mountpoint
952  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
953  * 			 entries if this argument is strue
954  * @param trigger	 Only reclaim vnodes with fewer than this many resident
955  *			 pages.
956  * @return		 The number of vnodes that were reclaimed.
957  */
958 static int
959 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
960 {
961 	struct vnode *vp;
962 	int count, done, target;
963 
964 	done = 0;
965 	vn_start_write(NULL, &mp, V_WAIT);
966 	MNT_ILOCK(mp);
967 	count = mp->mnt_nvnodelistsize;
968 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
969 	target = target / 10 + 1;
970 	while (count != 0 && done < target) {
971 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
972 		while (vp != NULL && vp->v_type == VMARKER)
973 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
974 		if (vp == NULL)
975 			break;
976 		/*
977 		 * XXX LRU is completely broken for non-free vnodes.  First
978 		 * by calling here in mountpoint order, then by moving
979 		 * unselected vnodes to the end here, and most grossly by
980 		 * removing the vlruvp() function that was supposed to
981 		 * maintain the order.  (This function was born broken
982 		 * since syncer problems prevented it doing anything.)  The
983 		 * order is closer to LRC (C = Created).
984 		 *
985 		 * LRU reclaiming of vnodes seems to have last worked in
986 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
987 		 * Then there was no hold count, and inactive vnodes were
988 		 * simply put on the free list in LRU order.  The separate
989 		 * lists also break LRU.  We prefer to reclaim from the
990 		 * free list for technical reasons.  This tends to thrash
991 		 * the free list to keep very unrecently used held vnodes.
992 		 * The problem is mitigated by keeping the free list large.
993 		 */
994 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
995 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
996 		--count;
997 		if (!VI_TRYLOCK(vp))
998 			goto next_iter;
999 		/*
1000 		 * If it's been deconstructed already, it's still
1001 		 * referenced, or it exceeds the trigger, skip it.
1002 		 * Also skip free vnodes.  We are trying to make space
1003 		 * to expand the free list, not reduce it.
1004 		 */
1005 		if (vp->v_usecount ||
1006 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1007 		    ((vp->v_iflag & VI_FREE) != 0) ||
1008 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
1009 		    vp->v_object->resident_page_count > trigger)) {
1010 			VI_UNLOCK(vp);
1011 			goto next_iter;
1012 		}
1013 		MNT_IUNLOCK(mp);
1014 		vholdl(vp);
1015 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1016 			vdrop(vp);
1017 			goto next_iter_mntunlocked;
1018 		}
1019 		VI_LOCK(vp);
1020 		/*
1021 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1022 		 * the vnode interlock and before it was locked again.
1023 		 *
1024 		 * It is not necessary to recheck VI_DOOMED because it can
1025 		 * only be set by another thread that holds both the vnode
1026 		 * lock and vnode interlock.  If another thread has the
1027 		 * vnode lock before we get to VOP_LOCK() and obtains the
1028 		 * vnode interlock after VOP_LOCK() drops the vnode
1029 		 * interlock, the other thread will be unable to drop the
1030 		 * vnode lock before our VOP_LOCK() call fails.
1031 		 */
1032 		if (vp->v_usecount ||
1033 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1034 		    (vp->v_iflag & VI_FREE) != 0 ||
1035 		    (vp->v_object != NULL &&
1036 		    vp->v_object->resident_page_count > trigger)) {
1037 			VOP_UNLOCK(vp, LK_INTERLOCK);
1038 			vdrop(vp);
1039 			goto next_iter_mntunlocked;
1040 		}
1041 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
1042 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
1043 		counter_u64_add(recycles_count, 1);
1044 		vgonel(vp);
1045 		VOP_UNLOCK(vp, 0);
1046 		vdropl(vp);
1047 		done++;
1048 next_iter_mntunlocked:
1049 		if (!should_yield())
1050 			goto relock_mnt;
1051 		goto yield;
1052 next_iter:
1053 		if (!should_yield())
1054 			continue;
1055 		MNT_IUNLOCK(mp);
1056 yield:
1057 		kern_yield(PRI_USER);
1058 relock_mnt:
1059 		MNT_ILOCK(mp);
1060 	}
1061 	MNT_IUNLOCK(mp);
1062 	vn_finished_write(mp);
1063 	return done;
1064 }
1065 
1066 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1067 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1068     0,
1069     "limit on vnode free requests per call to the vnlru_free routine");
1070 
1071 /*
1072  * Attempt to reduce the free list by the requested amount.
1073  */
1074 static void
1075 vnlru_free_locked(int count, struct vfsops *mnt_op)
1076 {
1077 	struct vnode *vp;
1078 	struct mount *mp;
1079 	bool tried_batches;
1080 
1081 	tried_batches = false;
1082 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1083 	if (count > max_vnlru_free)
1084 		count = max_vnlru_free;
1085 	for (; count > 0; count--) {
1086 		vp = TAILQ_FIRST(&vnode_free_list);
1087 		/*
1088 		 * The list can be modified while the free_list_mtx
1089 		 * has been dropped and vp could be NULL here.
1090 		 */
1091 		if (vp == NULL) {
1092 			if (tried_batches)
1093 				break;
1094 			mtx_unlock(&vnode_free_list_mtx);
1095 			vnlru_return_batches(mnt_op);
1096 			tried_batches = true;
1097 			mtx_lock(&vnode_free_list_mtx);
1098 			continue;
1099 		}
1100 
1101 		VNASSERT(vp->v_op != NULL, vp,
1102 		    ("vnlru_free: vnode already reclaimed."));
1103 		KASSERT((vp->v_iflag & VI_FREE) != 0,
1104 		    ("Removing vnode not on freelist"));
1105 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1106 		    ("Mangling active vnode"));
1107 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1108 
1109 		/*
1110 		 * Don't recycle if our vnode is from different type
1111 		 * of mount point.  Note that mp is type-safe, the
1112 		 * check does not reach unmapped address even if
1113 		 * vnode is reclaimed.
1114 		 * Don't recycle if we can't get the interlock without
1115 		 * blocking.
1116 		 */
1117 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1118 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1119 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1120 			continue;
1121 		}
1122 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1123 		    vp, ("vp inconsistent on freelist"));
1124 
1125 		/*
1126 		 * The clear of VI_FREE prevents activation of the
1127 		 * vnode.  There is no sense in putting the vnode on
1128 		 * the mount point active list, only to remove it
1129 		 * later during recycling.  Inline the relevant part
1130 		 * of vholdl(), to avoid triggering assertions or
1131 		 * activating.
1132 		 */
1133 		freevnodes--;
1134 		vp->v_iflag &= ~VI_FREE;
1135 		VNODE_REFCOUNT_FENCE_REL();
1136 		refcount_acquire(&vp->v_holdcnt);
1137 
1138 		mtx_unlock(&vnode_free_list_mtx);
1139 		VI_UNLOCK(vp);
1140 		vtryrecycle(vp);
1141 		/*
1142 		 * If the recycled succeeded this vdrop will actually free
1143 		 * the vnode.  If not it will simply place it back on
1144 		 * the free list.
1145 		 */
1146 		vdrop(vp);
1147 		mtx_lock(&vnode_free_list_mtx);
1148 	}
1149 }
1150 
1151 void
1152 vnlru_free(int count, struct vfsops *mnt_op)
1153 {
1154 
1155 	mtx_lock(&vnode_free_list_mtx);
1156 	vnlru_free_locked(count, mnt_op);
1157 	mtx_unlock(&vnode_free_list_mtx);
1158 }
1159 
1160 
1161 /* XXX some names and initialization are bad for limits and watermarks. */
1162 static int
1163 vspace(void)
1164 {
1165 	int space;
1166 
1167 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1168 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1169 	vlowat = vhiwat / 2;
1170 	if (numvnodes > desiredvnodes)
1171 		return (0);
1172 	space = desiredvnodes - numvnodes;
1173 	if (freevnodes > wantfreevnodes)
1174 		space += freevnodes - wantfreevnodes;
1175 	return (space);
1176 }
1177 
1178 static void
1179 vnlru_return_batch_locked(struct mount *mp)
1180 {
1181 	struct vnode *vp;
1182 
1183 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1184 
1185 	if (mp->mnt_tmpfreevnodelistsize == 0)
1186 		return;
1187 
1188 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1189 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1190 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1191 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1192 	}
1193 	mtx_lock(&vnode_free_list_mtx);
1194 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1195 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1196 	mtx_unlock(&vnode_free_list_mtx);
1197 	mp->mnt_tmpfreevnodelistsize = 0;
1198 }
1199 
1200 static void
1201 vnlru_return_batch(struct mount *mp)
1202 {
1203 
1204 	mtx_lock(&mp->mnt_listmtx);
1205 	vnlru_return_batch_locked(mp);
1206 	mtx_unlock(&mp->mnt_listmtx);
1207 }
1208 
1209 static void
1210 vnlru_return_batches(struct vfsops *mnt_op)
1211 {
1212 	struct mount *mp, *nmp;
1213 	bool need_unbusy;
1214 
1215 	mtx_lock(&mountlist_mtx);
1216 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1217 		need_unbusy = false;
1218 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1219 			goto next;
1220 		if (mp->mnt_tmpfreevnodelistsize == 0)
1221 			goto next;
1222 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1223 			vnlru_return_batch(mp);
1224 			need_unbusy = true;
1225 			mtx_lock(&mountlist_mtx);
1226 		}
1227 next:
1228 		nmp = TAILQ_NEXT(mp, mnt_list);
1229 		if (need_unbusy)
1230 			vfs_unbusy(mp);
1231 	}
1232 	mtx_unlock(&mountlist_mtx);
1233 }
1234 
1235 /*
1236  * Attempt to recycle vnodes in a context that is always safe to block.
1237  * Calling vlrurecycle() from the bowels of filesystem code has some
1238  * interesting deadlock problems.
1239  */
1240 static struct proc *vnlruproc;
1241 static int vnlruproc_sig;
1242 
1243 static void
1244 vnlru_proc(void)
1245 {
1246 	struct mount *mp, *nmp;
1247 	unsigned long onumvnodes;
1248 	int done, force, trigger, usevnodes;
1249 	bool reclaim_nc_src;
1250 
1251 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1252 	    SHUTDOWN_PRI_FIRST);
1253 
1254 	force = 0;
1255 	for (;;) {
1256 		kproc_suspend_check(vnlruproc);
1257 		mtx_lock(&vnode_free_list_mtx);
1258 		/*
1259 		 * If numvnodes is too large (due to desiredvnodes being
1260 		 * adjusted using its sysctl, or emergency growth), first
1261 		 * try to reduce it by discarding from the free list.
1262 		 */
1263 		if (numvnodes > desiredvnodes)
1264 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1265 		/*
1266 		 * Sleep if the vnode cache is in a good state.  This is
1267 		 * when it is not over-full and has space for about a 4%
1268 		 * or 9% expansion (by growing its size or inexcessively
1269 		 * reducing its free list).  Otherwise, try to reclaim
1270 		 * space for a 10% expansion.
1271 		 */
1272 		if (vstir && force == 0) {
1273 			force = 1;
1274 			vstir = 0;
1275 		}
1276 		if (vspace() >= vlowat && force == 0) {
1277 			vnlruproc_sig = 0;
1278 			wakeup(&vnlruproc_sig);
1279 			msleep(vnlruproc, &vnode_free_list_mtx,
1280 			    PVFS|PDROP, "vlruwt", hz);
1281 			continue;
1282 		}
1283 		mtx_unlock(&vnode_free_list_mtx);
1284 		done = 0;
1285 		onumvnodes = numvnodes;
1286 		/*
1287 		 * Calculate parameters for recycling.  These are the same
1288 		 * throughout the loop to give some semblance of fairness.
1289 		 * The trigger point is to avoid recycling vnodes with lots
1290 		 * of resident pages.  We aren't trying to free memory; we
1291 		 * are trying to recycle or at least free vnodes.
1292 		 */
1293 		if (numvnodes <= desiredvnodes)
1294 			usevnodes = numvnodes - freevnodes;
1295 		else
1296 			usevnodes = numvnodes;
1297 		if (usevnodes <= 0)
1298 			usevnodes = 1;
1299 		/*
1300 		 * The trigger value is is chosen to give a conservatively
1301 		 * large value to ensure that it alone doesn't prevent
1302 		 * making progress.  The value can easily be so large that
1303 		 * it is effectively infinite in some congested and
1304 		 * misconfigured cases, and this is necessary.  Normally
1305 		 * it is about 8 to 100 (pages), which is quite large.
1306 		 */
1307 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1308 		if (force < 2)
1309 			trigger = vsmalltrigger;
1310 		reclaim_nc_src = force >= 3;
1311 		mtx_lock(&mountlist_mtx);
1312 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1313 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1314 				nmp = TAILQ_NEXT(mp, mnt_list);
1315 				continue;
1316 			}
1317 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1318 			mtx_lock(&mountlist_mtx);
1319 			nmp = TAILQ_NEXT(mp, mnt_list);
1320 			vfs_unbusy(mp);
1321 		}
1322 		mtx_unlock(&mountlist_mtx);
1323 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1324 			uma_reclaim();
1325 		if (done == 0) {
1326 			if (force == 0 || force == 1) {
1327 				force = 2;
1328 				continue;
1329 			}
1330 			if (force == 2) {
1331 				force = 3;
1332 				continue;
1333 			}
1334 			force = 0;
1335 			vnlru_nowhere++;
1336 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1337 		} else
1338 			kern_yield(PRI_USER);
1339 		/*
1340 		 * After becoming active to expand above low water, keep
1341 		 * active until above high water.
1342 		 */
1343 		force = vspace() < vhiwat;
1344 	}
1345 }
1346 
1347 static struct kproc_desc vnlru_kp = {
1348 	"vnlru",
1349 	vnlru_proc,
1350 	&vnlruproc
1351 };
1352 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1353     &vnlru_kp);
1354 
1355 /*
1356  * Routines having to do with the management of the vnode table.
1357  */
1358 
1359 /*
1360  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1361  * before we actually vgone().  This function must be called with the vnode
1362  * held to prevent the vnode from being returned to the free list midway
1363  * through vgone().
1364  */
1365 static int
1366 vtryrecycle(struct vnode *vp)
1367 {
1368 	struct mount *vnmp;
1369 
1370 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1371 	VNASSERT(vp->v_holdcnt, vp,
1372 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1373 	/*
1374 	 * This vnode may found and locked via some other list, if so we
1375 	 * can't recycle it yet.
1376 	 */
1377 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1378 		CTR2(KTR_VFS,
1379 		    "%s: impossible to recycle, vp %p lock is already held",
1380 		    __func__, vp);
1381 		return (EWOULDBLOCK);
1382 	}
1383 	/*
1384 	 * Don't recycle if its filesystem is being suspended.
1385 	 */
1386 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1387 		VOP_UNLOCK(vp, 0);
1388 		CTR2(KTR_VFS,
1389 		    "%s: impossible to recycle, cannot start the write for %p",
1390 		    __func__, vp);
1391 		return (EBUSY);
1392 	}
1393 	/*
1394 	 * If we got this far, we need to acquire the interlock and see if
1395 	 * anyone picked up this vnode from another list.  If not, we will
1396 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1397 	 * will skip over it.
1398 	 */
1399 	VI_LOCK(vp);
1400 	if (vp->v_usecount) {
1401 		VOP_UNLOCK(vp, LK_INTERLOCK);
1402 		vn_finished_write(vnmp);
1403 		CTR2(KTR_VFS,
1404 		    "%s: impossible to recycle, %p is already referenced",
1405 		    __func__, vp);
1406 		return (EBUSY);
1407 	}
1408 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1409 		counter_u64_add(recycles_count, 1);
1410 		vgonel(vp);
1411 	}
1412 	VOP_UNLOCK(vp, LK_INTERLOCK);
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_NOINSMNTQ) != 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 static void
2634 v_incr_usecount_locked(struct vnode *vp)
2635 {
2636 
2637 	ASSERT_VI_LOCKED(vp, __func__);
2638 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2639 		VNASSERT(vp->v_usecount == 0, vp,
2640 		    ("vnode with usecount and VI_OWEINACT set"));
2641 		vp->v_iflag &= ~VI_OWEINACT;
2642 	}
2643 	refcount_acquire(&vp->v_usecount);
2644 	v_incr_devcount(vp);
2645 }
2646 
2647 /*
2648  * Increment the use count on the vnode, taking care to reference
2649  * the driver's usecount if this is a chardev.
2650  */
2651 static void
2652 v_incr_usecount(struct vnode *vp)
2653 {
2654 
2655 	ASSERT_VI_UNLOCKED(vp, __func__);
2656 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2657 
2658 	if (vp->v_type != VCHR &&
2659 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2660 		VNODE_REFCOUNT_FENCE_ACQ();
2661 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2662 		    ("vnode with usecount and VI_OWEINACT set"));
2663 	} else {
2664 		VI_LOCK(vp);
2665 		v_incr_usecount_locked(vp);
2666 		VI_UNLOCK(vp);
2667 	}
2668 }
2669 
2670 /*
2671  * Increment si_usecount of the associated device, if any.
2672  */
2673 static void
2674 v_incr_devcount(struct vnode *vp)
2675 {
2676 
2677 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2678 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2679 		dev_lock();
2680 		vp->v_rdev->si_usecount++;
2681 		dev_unlock();
2682 	}
2683 }
2684 
2685 /*
2686  * Decrement si_usecount of the associated device, if any.
2687  */
2688 static void
2689 v_decr_devcount(struct vnode *vp)
2690 {
2691 
2692 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2693 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2694 		dev_lock();
2695 		vp->v_rdev->si_usecount--;
2696 		dev_unlock();
2697 	}
2698 }
2699 
2700 /*
2701  * Grab a particular vnode from the free list, increment its
2702  * reference count and lock it.  VI_DOOMED is set if the vnode
2703  * is being destroyed.  Only callers who specify LK_RETRY will
2704  * see doomed vnodes.  If inactive processing was delayed in
2705  * vput try to do it here.
2706  *
2707  * Notes on lockless counter manipulation:
2708  * _vhold, vputx and other routines make various decisions based
2709  * on either holdcnt or usecount being 0. As long as either counter
2710  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2711  * with atomic operations. Otherwise the interlock is taken covering
2712  * both the atomic and additional actions.
2713  */
2714 int
2715 vget(struct vnode *vp, int flags, struct thread *td)
2716 {
2717 	int error, oweinact;
2718 
2719 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2720 	    ("vget: invalid lock operation"));
2721 
2722 	if ((flags & LK_INTERLOCK) != 0)
2723 		ASSERT_VI_LOCKED(vp, __func__);
2724 	else
2725 		ASSERT_VI_UNLOCKED(vp, __func__);
2726 	if ((flags & LK_VNHELD) != 0)
2727 		VNASSERT((vp->v_holdcnt > 0), vp,
2728 		    ("vget: LK_VNHELD passed but vnode not held"));
2729 
2730 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2731 
2732 	if ((flags & LK_VNHELD) == 0)
2733 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2734 
2735 	if ((error = vn_lock(vp, flags)) != 0) {
2736 		vdrop(vp);
2737 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2738 		    vp);
2739 		return (error);
2740 	}
2741 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2742 		panic("vget: vn_lock failed to return ENOENT\n");
2743 	/*
2744 	 * We don't guarantee that any particular close will
2745 	 * trigger inactive processing so just make a best effort
2746 	 * here at preventing a reference to a removed file.  If
2747 	 * we don't succeed no harm is done.
2748 	 *
2749 	 * Upgrade our holdcnt to a usecount.
2750 	 */
2751 	if (vp->v_type == VCHR ||
2752 	    !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2753 		VI_LOCK(vp);
2754 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2755 			oweinact = 0;
2756 		} else {
2757 			oweinact = 1;
2758 			vp->v_iflag &= ~VI_OWEINACT;
2759 			VNODE_REFCOUNT_FENCE_REL();
2760 		}
2761 		refcount_acquire(&vp->v_usecount);
2762 		v_incr_devcount(vp);
2763 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2764 		    (flags & LK_NOWAIT) == 0)
2765 			vinactive(vp, td);
2766 		VI_UNLOCK(vp);
2767 	}
2768 	return (0);
2769 }
2770 
2771 /*
2772  * Increase the reference (use) and hold count of a vnode.
2773  * This will also remove the vnode from the free list if it is presently free.
2774  */
2775 void
2776 vref(struct vnode *vp)
2777 {
2778 
2779 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2780 	_vhold(vp, false);
2781 	v_incr_usecount(vp);
2782 }
2783 
2784 void
2785 vrefl(struct vnode *vp)
2786 {
2787 
2788 	ASSERT_VI_LOCKED(vp, __func__);
2789 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2790 	_vhold(vp, true);
2791 	v_incr_usecount_locked(vp);
2792 }
2793 
2794 void
2795 vrefact(struct vnode *vp)
2796 {
2797 
2798 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2799 	if (__predict_false(vp->v_type == VCHR)) {
2800 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2801 		    ("%s: wrong ref counts", __func__));
2802 		vref(vp);
2803 		return;
2804 	}
2805 #ifdef INVARIANTS
2806 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2807 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2808 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2809 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2810 #else
2811 	refcount_acquire(&vp->v_holdcnt);
2812 	refcount_acquire(&vp->v_usecount);
2813 #endif
2814 }
2815 
2816 /*
2817  * Return reference count of a vnode.
2818  *
2819  * The results of this call are only guaranteed when some mechanism is used to
2820  * stop other processes from gaining references to the vnode.  This may be the
2821  * case if the caller holds the only reference.  This is also useful when stale
2822  * data is acceptable as race conditions may be accounted for by some other
2823  * means.
2824  */
2825 int
2826 vrefcnt(struct vnode *vp)
2827 {
2828 
2829 	return (vp->v_usecount);
2830 }
2831 
2832 #define	VPUTX_VRELE	1
2833 #define	VPUTX_VPUT	2
2834 #define	VPUTX_VUNREF	3
2835 
2836 /*
2837  * Decrement the use and hold counts for a vnode.
2838  *
2839  * See an explanation near vget() as to why atomic operation is safe.
2840  */
2841 static void
2842 vputx(struct vnode *vp, int func)
2843 {
2844 	int error;
2845 
2846 	KASSERT(vp != NULL, ("vputx: null vp"));
2847 	if (func == VPUTX_VUNREF)
2848 		ASSERT_VOP_LOCKED(vp, "vunref");
2849 	else if (func == VPUTX_VPUT)
2850 		ASSERT_VOP_LOCKED(vp, "vput");
2851 	else
2852 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2853 	ASSERT_VI_UNLOCKED(vp, __func__);
2854 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2855 
2856 	if (vp->v_type != VCHR &&
2857 	    refcount_release_if_not_last(&vp->v_usecount)) {
2858 		if (func == VPUTX_VPUT)
2859 			VOP_UNLOCK(vp, 0);
2860 		vdrop(vp);
2861 		return;
2862 	}
2863 
2864 	VI_LOCK(vp);
2865 
2866 	/*
2867 	 * We want to hold the vnode until the inactive finishes to
2868 	 * prevent vgone() races.  We drop the use count here and the
2869 	 * hold count below when we're done.
2870 	 */
2871 	if (!refcount_release(&vp->v_usecount) ||
2872 	    (vp->v_iflag & VI_DOINGINACT)) {
2873 		if (func == VPUTX_VPUT)
2874 			VOP_UNLOCK(vp, 0);
2875 		v_decr_devcount(vp);
2876 		vdropl(vp);
2877 		return;
2878 	}
2879 
2880 	v_decr_devcount(vp);
2881 
2882 	error = 0;
2883 
2884 	if (vp->v_usecount != 0) {
2885 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2886 		panic("vputx: usecount not zero");
2887 	}
2888 
2889 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2890 
2891 	/*
2892 	 * We must call VOP_INACTIVE with the node locked. Mark
2893 	 * as VI_DOINGINACT to avoid recursion.
2894 	 */
2895 	vp->v_iflag |= VI_OWEINACT;
2896 	switch (func) {
2897 	case VPUTX_VRELE:
2898 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2899 		VI_LOCK(vp);
2900 		break;
2901 	case VPUTX_VPUT:
2902 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2903 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2904 			    LK_NOWAIT);
2905 			VI_LOCK(vp);
2906 		}
2907 		break;
2908 	case VPUTX_VUNREF:
2909 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2910 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2911 			VI_LOCK(vp);
2912 		}
2913 		break;
2914 	}
2915 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2916 	    ("vnode with usecount and VI_OWEINACT set"));
2917 	if (error == 0) {
2918 		if (vp->v_iflag & VI_OWEINACT)
2919 			vinactive(vp, curthread);
2920 		if (func != VPUTX_VUNREF)
2921 			VOP_UNLOCK(vp, 0);
2922 	}
2923 	vdropl(vp);
2924 }
2925 
2926 /*
2927  * Vnode put/release.
2928  * If count drops to zero, call inactive routine and return to freelist.
2929  */
2930 void
2931 vrele(struct vnode *vp)
2932 {
2933 
2934 	vputx(vp, VPUTX_VRELE);
2935 }
2936 
2937 /*
2938  * Release an already locked vnode.  This give the same effects as
2939  * unlock+vrele(), but takes less time and avoids releasing and
2940  * re-aquiring the lock (as vrele() acquires the lock internally.)
2941  */
2942 void
2943 vput(struct vnode *vp)
2944 {
2945 
2946 	vputx(vp, VPUTX_VPUT);
2947 }
2948 
2949 /*
2950  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2951  */
2952 void
2953 vunref(struct vnode *vp)
2954 {
2955 
2956 	vputx(vp, VPUTX_VUNREF);
2957 }
2958 
2959 /*
2960  * Increase the hold count and activate if this is the first reference.
2961  */
2962 void
2963 _vhold(struct vnode *vp, bool locked)
2964 {
2965 	struct mount *mp;
2966 
2967 	if (locked)
2968 		ASSERT_VI_LOCKED(vp, __func__);
2969 	else
2970 		ASSERT_VI_UNLOCKED(vp, __func__);
2971 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2972 	if (!locked) {
2973 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2974 			VNODE_REFCOUNT_FENCE_ACQ();
2975 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2976 			    ("_vhold: vnode with holdcnt is free"));
2977 			return;
2978 		}
2979 		VI_LOCK(vp);
2980 	}
2981 	if ((vp->v_iflag & VI_FREE) == 0) {
2982 		refcount_acquire(&vp->v_holdcnt);
2983 		if (!locked)
2984 			VI_UNLOCK(vp);
2985 		return;
2986 	}
2987 	VNASSERT(vp->v_holdcnt == 0, vp,
2988 	    ("%s: wrong hold count", __func__));
2989 	VNASSERT(vp->v_op != NULL, vp,
2990 	    ("%s: vnode already reclaimed.", __func__));
2991 	/*
2992 	 * Remove a vnode from the free list, mark it as in use,
2993 	 * and put it on the active list.
2994 	 */
2995 	VNASSERT(vp->v_mount != NULL, vp,
2996 	    ("_vhold: vnode not on per mount vnode list"));
2997 	mp = vp->v_mount;
2998 	mtx_lock(&mp->mnt_listmtx);
2999 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
3000 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3001 		mp->mnt_tmpfreevnodelistsize--;
3002 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
3003 	} else {
3004 		mtx_lock(&vnode_free_list_mtx);
3005 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3006 		freevnodes--;
3007 		mtx_unlock(&vnode_free_list_mtx);
3008 	}
3009 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3010 	    ("Activating already active vnode"));
3011 	vp->v_iflag &= ~VI_FREE;
3012 	vp->v_iflag |= VI_ACTIVE;
3013 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3014 	mp->mnt_activevnodelistsize++;
3015 	mtx_unlock(&mp->mnt_listmtx);
3016 	refcount_acquire(&vp->v_holdcnt);
3017 	if (!locked)
3018 		VI_UNLOCK(vp);
3019 }
3020 
3021 /*
3022  * Drop the hold count of the vnode.  If this is the last reference to
3023  * the vnode we place it on the free list unless it has been vgone'd
3024  * (marked VI_DOOMED) in which case we will free it.
3025  *
3026  * Because the vnode vm object keeps a hold reference on the vnode if
3027  * there is at least one resident non-cached page, the vnode cannot
3028  * leave the active list without the page cleanup done.
3029  */
3030 void
3031 _vdrop(struct vnode *vp, bool locked)
3032 {
3033 	struct bufobj *bo;
3034 	struct mount *mp;
3035 	int active;
3036 
3037 	if (locked)
3038 		ASSERT_VI_LOCKED(vp, __func__);
3039 	else
3040 		ASSERT_VI_UNLOCKED(vp, __func__);
3041 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3042 	if ((int)vp->v_holdcnt <= 0)
3043 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
3044 	if (!locked) {
3045 		if (refcount_release_if_not_last(&vp->v_holdcnt))
3046 			return;
3047 		VI_LOCK(vp);
3048 	}
3049 	if (refcount_release(&vp->v_holdcnt) == 0) {
3050 		VI_UNLOCK(vp);
3051 		return;
3052 	}
3053 	if ((vp->v_iflag & VI_DOOMED) == 0) {
3054 		/*
3055 		 * Mark a vnode as free: remove it from its active list
3056 		 * and put it up for recycling on the freelist.
3057 		 */
3058 		VNASSERT(vp->v_op != NULL, vp,
3059 		    ("vdropl: vnode already reclaimed."));
3060 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3061 		    ("vnode already free"));
3062 		VNASSERT(vp->v_holdcnt == 0, vp,
3063 		    ("vdropl: freeing when we shouldn't"));
3064 		active = vp->v_iflag & VI_ACTIVE;
3065 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
3066 			vp->v_iflag &= ~VI_ACTIVE;
3067 			mp = vp->v_mount;
3068 			if (mp != NULL) {
3069 				mtx_lock(&mp->mnt_listmtx);
3070 				if (active) {
3071 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
3072 					    vp, v_actfreelist);
3073 					mp->mnt_activevnodelistsize--;
3074 				}
3075 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3076 				    vp, v_actfreelist);
3077 				mp->mnt_tmpfreevnodelistsize++;
3078 				vp->v_iflag |= VI_FREE;
3079 				vp->v_mflag |= VMP_TMPMNTFREELIST;
3080 				VI_UNLOCK(vp);
3081 				if (mp->mnt_tmpfreevnodelistsize >=
3082 				    mnt_free_list_batch)
3083 					vnlru_return_batch_locked(mp);
3084 				mtx_unlock(&mp->mnt_listmtx);
3085 			} else {
3086 				VNASSERT(active == 0, vp,
3087 				    ("vdropl: active vnode not on per mount "
3088 				    "vnode list"));
3089 				mtx_lock(&vnode_free_list_mtx);
3090 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3091 				    v_actfreelist);
3092 				freevnodes++;
3093 				vp->v_iflag |= VI_FREE;
3094 				VI_UNLOCK(vp);
3095 				mtx_unlock(&vnode_free_list_mtx);
3096 			}
3097 		} else {
3098 			VI_UNLOCK(vp);
3099 			counter_u64_add(free_owe_inact, 1);
3100 		}
3101 		return;
3102 	}
3103 	/*
3104 	 * The vnode has been marked for destruction, so free it.
3105 	 *
3106 	 * The vnode will be returned to the zone where it will
3107 	 * normally remain until it is needed for another vnode. We
3108 	 * need to cleanup (or verify that the cleanup has already
3109 	 * been done) any residual data left from its current use
3110 	 * so as not to contaminate the freshly allocated vnode.
3111 	 */
3112 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
3113 	atomic_subtract_long(&numvnodes, 1);
3114 	bo = &vp->v_bufobj;
3115 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3116 	    ("cleaned vnode still on the free list."));
3117 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
3118 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
3119 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
3120 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
3121 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
3122 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
3123 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
3124 	    ("clean blk trie not empty"));
3125 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3126 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3127 	    ("dirty blk trie not empty"));
3128 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3129 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3130 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3131 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3132 	    ("Dangling rangelock waiters"));
3133 	VI_UNLOCK(vp);
3134 #ifdef MAC
3135 	mac_vnode_destroy(vp);
3136 #endif
3137 	if (vp->v_pollinfo != NULL) {
3138 		destroy_vpollinfo(vp->v_pollinfo);
3139 		vp->v_pollinfo = NULL;
3140 	}
3141 #ifdef INVARIANTS
3142 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3143 	vp->v_op = NULL;
3144 #endif
3145 	vp->v_mountedhere = NULL;
3146 	vp->v_unpcb = NULL;
3147 	vp->v_rdev = NULL;
3148 	vp->v_fifoinfo = NULL;
3149 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3150 	vp->v_iflag = 0;
3151 	vp->v_vflag = 0;
3152 	bo->bo_flag = 0;
3153 	uma_zfree(vnode_zone, vp);
3154 }
3155 
3156 /*
3157  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3158  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3159  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3160  * failed lock upgrade.
3161  */
3162 void
3163 vinactive(struct vnode *vp, struct thread *td)
3164 {
3165 	struct vm_object *obj;
3166 
3167 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3168 	ASSERT_VI_LOCKED(vp, "vinactive");
3169 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3170 	    ("vinactive: recursed on VI_DOINGINACT"));
3171 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3172 	vp->v_iflag |= VI_DOINGINACT;
3173 	vp->v_iflag &= ~VI_OWEINACT;
3174 	VI_UNLOCK(vp);
3175 	/*
3176 	 * Before moving off the active list, we must be sure that any
3177 	 * modified pages are converted into the vnode's dirty
3178 	 * buffers, since these will no longer be checked once the
3179 	 * vnode is on the inactive list.
3180 	 *
3181 	 * The write-out of the dirty pages is asynchronous.  At the
3182 	 * point that VOP_INACTIVE() is called, there could still be
3183 	 * pending I/O and dirty pages in the object.
3184 	 */
3185 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3186 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3187 		VM_OBJECT_WLOCK(obj);
3188 		vm_object_page_clean(obj, 0, 0, 0);
3189 		VM_OBJECT_WUNLOCK(obj);
3190 	}
3191 	VOP_INACTIVE(vp, td);
3192 	VI_LOCK(vp);
3193 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3194 	    ("vinactive: lost VI_DOINGINACT"));
3195 	vp->v_iflag &= ~VI_DOINGINACT;
3196 }
3197 
3198 /*
3199  * Remove any vnodes in the vnode table belonging to mount point mp.
3200  *
3201  * If FORCECLOSE is not specified, there should not be any active ones,
3202  * return error if any are found (nb: this is a user error, not a
3203  * system error). If FORCECLOSE is specified, detach any active vnodes
3204  * that are found.
3205  *
3206  * If WRITECLOSE is set, only flush out regular file vnodes open for
3207  * writing.
3208  *
3209  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3210  *
3211  * `rootrefs' specifies the base reference count for the root vnode
3212  * of this filesystem. The root vnode is considered busy if its
3213  * v_usecount exceeds this value. On a successful return, vflush(, td)
3214  * will call vrele() on the root vnode exactly rootrefs times.
3215  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3216  * be zero.
3217  */
3218 #ifdef DIAGNOSTIC
3219 static int busyprt = 0;		/* print out busy vnodes */
3220 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3221 #endif
3222 
3223 int
3224 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3225 {
3226 	struct vnode *vp, *mvp, *rootvp = NULL;
3227 	struct vattr vattr;
3228 	int busy = 0, error;
3229 
3230 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3231 	    rootrefs, flags);
3232 	if (rootrefs > 0) {
3233 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3234 		    ("vflush: bad args"));
3235 		/*
3236 		 * Get the filesystem root vnode. We can vput() it
3237 		 * immediately, since with rootrefs > 0, it won't go away.
3238 		 */
3239 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3240 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3241 			    __func__, error);
3242 			return (error);
3243 		}
3244 		vput(rootvp);
3245 	}
3246 loop:
3247 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3248 		vholdl(vp);
3249 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3250 		if (error) {
3251 			vdrop(vp);
3252 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3253 			goto loop;
3254 		}
3255 		/*
3256 		 * Skip over a vnodes marked VV_SYSTEM.
3257 		 */
3258 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3259 			VOP_UNLOCK(vp, 0);
3260 			vdrop(vp);
3261 			continue;
3262 		}
3263 		/*
3264 		 * If WRITECLOSE is set, flush out unlinked but still open
3265 		 * files (even if open only for reading) and regular file
3266 		 * vnodes open for writing.
3267 		 */
3268 		if (flags & WRITECLOSE) {
3269 			if (vp->v_object != NULL) {
3270 				VM_OBJECT_WLOCK(vp->v_object);
3271 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3272 				VM_OBJECT_WUNLOCK(vp->v_object);
3273 			}
3274 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3275 			if (error != 0) {
3276 				VOP_UNLOCK(vp, 0);
3277 				vdrop(vp);
3278 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3279 				return (error);
3280 			}
3281 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3282 			VI_LOCK(vp);
3283 
3284 			if ((vp->v_type == VNON ||
3285 			    (error == 0 && vattr.va_nlink > 0)) &&
3286 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3287 				VOP_UNLOCK(vp, 0);
3288 				vdropl(vp);
3289 				continue;
3290 			}
3291 		} else
3292 			VI_LOCK(vp);
3293 		/*
3294 		 * With v_usecount == 0, all we need to do is clear out the
3295 		 * vnode data structures and we are done.
3296 		 *
3297 		 * If FORCECLOSE is set, forcibly close the vnode.
3298 		 */
3299 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3300 			vgonel(vp);
3301 		} else {
3302 			busy++;
3303 #ifdef DIAGNOSTIC
3304 			if (busyprt)
3305 				vn_printf(vp, "vflush: busy vnode ");
3306 #endif
3307 		}
3308 		VOP_UNLOCK(vp, 0);
3309 		vdropl(vp);
3310 	}
3311 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3312 		/*
3313 		 * If just the root vnode is busy, and if its refcount
3314 		 * is equal to `rootrefs', then go ahead and kill it.
3315 		 */
3316 		VI_LOCK(rootvp);
3317 		KASSERT(busy > 0, ("vflush: not busy"));
3318 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3319 		    ("vflush: usecount %d < rootrefs %d",
3320 		     rootvp->v_usecount, rootrefs));
3321 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3322 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3323 			vgone(rootvp);
3324 			VOP_UNLOCK(rootvp, 0);
3325 			busy = 0;
3326 		} else
3327 			VI_UNLOCK(rootvp);
3328 	}
3329 	if (busy) {
3330 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3331 		    busy);
3332 		return (EBUSY);
3333 	}
3334 	for (; rootrefs > 0; rootrefs--)
3335 		vrele(rootvp);
3336 	return (0);
3337 }
3338 
3339 /*
3340  * Recycle an unused vnode to the front of the free list.
3341  */
3342 int
3343 vrecycle(struct vnode *vp)
3344 {
3345 	int recycled;
3346 
3347 	VI_LOCK(vp);
3348 	recycled = vrecyclel(vp);
3349 	VI_UNLOCK(vp);
3350 	return (recycled);
3351 }
3352 
3353 /*
3354  * vrecycle, with the vp interlock held.
3355  */
3356 int
3357 vrecyclel(struct vnode *vp)
3358 {
3359 	int recycled;
3360 
3361 	ASSERT_VOP_ELOCKED(vp, __func__);
3362 	ASSERT_VI_LOCKED(vp, __func__);
3363 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3364 	recycled = 0;
3365 	if (vp->v_usecount == 0) {
3366 		recycled = 1;
3367 		vgonel(vp);
3368 	}
3369 	return (recycled);
3370 }
3371 
3372 /*
3373  * Eliminate all activity associated with a vnode
3374  * in preparation for reuse.
3375  */
3376 void
3377 vgone(struct vnode *vp)
3378 {
3379 	VI_LOCK(vp);
3380 	vgonel(vp);
3381 	VI_UNLOCK(vp);
3382 }
3383 
3384 static void
3385 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3386     struct vnode *lowervp __unused)
3387 {
3388 }
3389 
3390 /*
3391  * Notify upper mounts about reclaimed or unlinked vnode.
3392  */
3393 void
3394 vfs_notify_upper(struct vnode *vp, int event)
3395 {
3396 	static struct vfsops vgonel_vfsops = {
3397 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3398 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3399 	};
3400 	struct mount *mp, *ump, *mmp;
3401 
3402 	mp = vp->v_mount;
3403 	if (mp == NULL)
3404 		return;
3405 
3406 	MNT_ILOCK(mp);
3407 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3408 		goto unlock;
3409 	MNT_IUNLOCK(mp);
3410 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3411 	mmp->mnt_op = &vgonel_vfsops;
3412 	mmp->mnt_kern_flag |= MNTK_MARKER;
3413 	MNT_ILOCK(mp);
3414 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3415 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3416 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3417 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3418 			continue;
3419 		}
3420 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3421 		MNT_IUNLOCK(mp);
3422 		switch (event) {
3423 		case VFS_NOTIFY_UPPER_RECLAIM:
3424 			VFS_RECLAIM_LOWERVP(ump, vp);
3425 			break;
3426 		case VFS_NOTIFY_UPPER_UNLINK:
3427 			VFS_UNLINK_LOWERVP(ump, vp);
3428 			break;
3429 		default:
3430 			KASSERT(0, ("invalid event %d", event));
3431 			break;
3432 		}
3433 		MNT_ILOCK(mp);
3434 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3435 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3436 	}
3437 	free(mmp, M_TEMP);
3438 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3439 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3440 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3441 		wakeup(&mp->mnt_uppers);
3442 	}
3443 unlock:
3444 	MNT_IUNLOCK(mp);
3445 }
3446 
3447 /*
3448  * vgone, with the vp interlock held.
3449  */
3450 static void
3451 vgonel(struct vnode *vp)
3452 {
3453 	struct thread *td;
3454 	int oweinact;
3455 	int active;
3456 	struct mount *mp;
3457 
3458 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3459 	ASSERT_VI_LOCKED(vp, "vgonel");
3460 	VNASSERT(vp->v_holdcnt, vp,
3461 	    ("vgonel: vp %p has no reference.", vp));
3462 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3463 	td = curthread;
3464 
3465 	/*
3466 	 * Don't vgonel if we're already doomed.
3467 	 */
3468 	if (vp->v_iflag & VI_DOOMED)
3469 		return;
3470 	vp->v_iflag |= VI_DOOMED;
3471 
3472 	/*
3473 	 * Check to see if the vnode is in use.  If so, we have to call
3474 	 * VOP_CLOSE() and VOP_INACTIVE().
3475 	 */
3476 	active = vp->v_usecount;
3477 	oweinact = (vp->v_iflag & VI_OWEINACT);
3478 	VI_UNLOCK(vp);
3479 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3480 
3481 	/*
3482 	 * If purging an active vnode, it must be closed and
3483 	 * deactivated before being reclaimed.
3484 	 */
3485 	if (active)
3486 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3487 	if (oweinact || active) {
3488 		VI_LOCK(vp);
3489 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3490 			vinactive(vp, td);
3491 		VI_UNLOCK(vp);
3492 	}
3493 	if (vp->v_type == VSOCK)
3494 		vfs_unp_reclaim(vp);
3495 
3496 	/*
3497 	 * Clean out any buffers associated with the vnode.
3498 	 * If the flush fails, just toss the buffers.
3499 	 */
3500 	mp = NULL;
3501 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3502 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3503 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3504 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3505 			;
3506 	}
3507 
3508 	BO_LOCK(&vp->v_bufobj);
3509 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3510 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3511 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3512 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3513 	    ("vp %p bufobj not invalidated", vp));
3514 
3515 	/*
3516 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3517 	 * after the object's page queue is flushed.
3518 	 */
3519 	if (vp->v_bufobj.bo_object == NULL)
3520 		vp->v_bufobj.bo_flag |= BO_DEAD;
3521 	BO_UNLOCK(&vp->v_bufobj);
3522 
3523 	/*
3524 	 * Reclaim the vnode.
3525 	 */
3526 	if (VOP_RECLAIM(vp, td))
3527 		panic("vgone: cannot reclaim");
3528 	if (mp != NULL)
3529 		vn_finished_secondary_write(mp);
3530 	VNASSERT(vp->v_object == NULL, vp,
3531 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3532 	/*
3533 	 * Clear the advisory locks and wake up waiting threads.
3534 	 */
3535 	(void)VOP_ADVLOCKPURGE(vp);
3536 	vp->v_lockf = NULL;
3537 	/*
3538 	 * Delete from old mount point vnode list.
3539 	 */
3540 	delmntque(vp);
3541 	cache_purge(vp);
3542 	/*
3543 	 * Done with purge, reset to the standard lock and invalidate
3544 	 * the vnode.
3545 	 */
3546 	VI_LOCK(vp);
3547 	vp->v_vnlock = &vp->v_lock;
3548 	vp->v_op = &dead_vnodeops;
3549 	vp->v_tag = "none";
3550 	vp->v_type = VBAD;
3551 }
3552 
3553 /*
3554  * Calculate the total number of references to a special device.
3555  */
3556 int
3557 vcount(struct vnode *vp)
3558 {
3559 	int count;
3560 
3561 	dev_lock();
3562 	count = vp->v_rdev->si_usecount;
3563 	dev_unlock();
3564 	return (count);
3565 }
3566 
3567 /*
3568  * Same as above, but using the struct cdev *as argument
3569  */
3570 int
3571 count_dev(struct cdev *dev)
3572 {
3573 	int count;
3574 
3575 	dev_lock();
3576 	count = dev->si_usecount;
3577 	dev_unlock();
3578 	return(count);
3579 }
3580 
3581 /*
3582  * Print out a description of a vnode.
3583  */
3584 static char *typename[] =
3585 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3586  "VMARKER"};
3587 
3588 void
3589 vn_printf(struct vnode *vp, const char *fmt, ...)
3590 {
3591 	va_list ap;
3592 	char buf[256], buf2[16];
3593 	u_long flags;
3594 
3595 	va_start(ap, fmt);
3596 	vprintf(fmt, ap);
3597 	va_end(ap);
3598 	printf("%p: ", (void *)vp);
3599 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3600 	printf("    usecount %d, writecount %d, refcount %d",
3601 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3602 	switch (vp->v_type) {
3603 	case VDIR:
3604 		printf(" mountedhere %p\n", vp->v_mountedhere);
3605 		break;
3606 	case VCHR:
3607 		printf(" rdev %p\n", vp->v_rdev);
3608 		break;
3609 	case VSOCK:
3610 		printf(" socket %p\n", vp->v_unpcb);
3611 		break;
3612 	case VFIFO:
3613 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3614 		break;
3615 	default:
3616 		printf("\n");
3617 		break;
3618 	}
3619 	buf[0] = '\0';
3620 	buf[1] = '\0';
3621 	if (vp->v_vflag & VV_ROOT)
3622 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3623 	if (vp->v_vflag & VV_ISTTY)
3624 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3625 	if (vp->v_vflag & VV_NOSYNC)
3626 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3627 	if (vp->v_vflag & VV_ETERNALDEV)
3628 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3629 	if (vp->v_vflag & VV_CACHEDLABEL)
3630 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3631 	if (vp->v_vflag & VV_COPYONWRITE)
3632 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3633 	if (vp->v_vflag & VV_SYSTEM)
3634 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3635 	if (vp->v_vflag & VV_PROCDEP)
3636 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3637 	if (vp->v_vflag & VV_NOKNOTE)
3638 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3639 	if (vp->v_vflag & VV_DELETED)
3640 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3641 	if (vp->v_vflag & VV_MD)
3642 		strlcat(buf, "|VV_MD", sizeof(buf));
3643 	if (vp->v_vflag & VV_FORCEINSMQ)
3644 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3645 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3646 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3647 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3648 	if (flags != 0) {
3649 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3650 		strlcat(buf, buf2, sizeof(buf));
3651 	}
3652 	if (vp->v_iflag & VI_MOUNT)
3653 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3654 	if (vp->v_iflag & VI_DOOMED)
3655 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3656 	if (vp->v_iflag & VI_FREE)
3657 		strlcat(buf, "|VI_FREE", sizeof(buf));
3658 	if (vp->v_iflag & VI_ACTIVE)
3659 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3660 	if (vp->v_iflag & VI_DOINGINACT)
3661 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3662 	if (vp->v_iflag & VI_OWEINACT)
3663 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3664 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3665 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3666 	if (flags != 0) {
3667 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3668 		strlcat(buf, buf2, sizeof(buf));
3669 	}
3670 	printf("    flags (%s)\n", buf + 1);
3671 	if (mtx_owned(VI_MTX(vp)))
3672 		printf(" VI_LOCKed");
3673 	if (vp->v_object != NULL)
3674 		printf("    v_object %p ref %d pages %d "
3675 		    "cleanbuf %d dirtybuf %d\n",
3676 		    vp->v_object, vp->v_object->ref_count,
3677 		    vp->v_object->resident_page_count,
3678 		    vp->v_bufobj.bo_clean.bv_cnt,
3679 		    vp->v_bufobj.bo_dirty.bv_cnt);
3680 	printf("    ");
3681 	lockmgr_printinfo(vp->v_vnlock);
3682 	if (vp->v_data != NULL)
3683 		VOP_PRINT(vp);
3684 }
3685 
3686 #ifdef DDB
3687 /*
3688  * List all of the locked vnodes in the system.
3689  * Called when debugging the kernel.
3690  */
3691 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3692 {
3693 	struct mount *mp;
3694 	struct vnode *vp;
3695 
3696 	/*
3697 	 * Note: because this is DDB, we can't obey the locking semantics
3698 	 * for these structures, which means we could catch an inconsistent
3699 	 * state and dereference a nasty pointer.  Not much to be done
3700 	 * about that.
3701 	 */
3702 	db_printf("Locked vnodes\n");
3703 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3704 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3705 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3706 				vn_printf(vp, "vnode ");
3707 		}
3708 	}
3709 }
3710 
3711 /*
3712  * Show details about the given vnode.
3713  */
3714 DB_SHOW_COMMAND(vnode, db_show_vnode)
3715 {
3716 	struct vnode *vp;
3717 
3718 	if (!have_addr)
3719 		return;
3720 	vp = (struct vnode *)addr;
3721 	vn_printf(vp, "vnode ");
3722 }
3723 
3724 /*
3725  * Show details about the given mount point.
3726  */
3727 DB_SHOW_COMMAND(mount, db_show_mount)
3728 {
3729 	struct mount *mp;
3730 	struct vfsopt *opt;
3731 	struct statfs *sp;
3732 	struct vnode *vp;
3733 	char buf[512];
3734 	uint64_t mflags;
3735 	u_int flags;
3736 
3737 	if (!have_addr) {
3738 		/* No address given, print short info about all mount points. */
3739 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3740 			db_printf("%p %s on %s (%s)\n", mp,
3741 			    mp->mnt_stat.f_mntfromname,
3742 			    mp->mnt_stat.f_mntonname,
3743 			    mp->mnt_stat.f_fstypename);
3744 			if (db_pager_quit)
3745 				break;
3746 		}
3747 		db_printf("\nMore info: show mount <addr>\n");
3748 		return;
3749 	}
3750 
3751 	mp = (struct mount *)addr;
3752 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3753 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3754 
3755 	buf[0] = '\0';
3756 	mflags = mp->mnt_flag;
3757 #define	MNT_FLAG(flag)	do {						\
3758 	if (mflags & (flag)) {						\
3759 		if (buf[0] != '\0')					\
3760 			strlcat(buf, ", ", sizeof(buf));		\
3761 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3762 		mflags &= ~(flag);					\
3763 	}								\
3764 } while (0)
3765 	MNT_FLAG(MNT_RDONLY);
3766 	MNT_FLAG(MNT_SYNCHRONOUS);
3767 	MNT_FLAG(MNT_NOEXEC);
3768 	MNT_FLAG(MNT_NOSUID);
3769 	MNT_FLAG(MNT_NFS4ACLS);
3770 	MNT_FLAG(MNT_UNION);
3771 	MNT_FLAG(MNT_ASYNC);
3772 	MNT_FLAG(MNT_SUIDDIR);
3773 	MNT_FLAG(MNT_SOFTDEP);
3774 	MNT_FLAG(MNT_NOSYMFOLLOW);
3775 	MNT_FLAG(MNT_GJOURNAL);
3776 	MNT_FLAG(MNT_MULTILABEL);
3777 	MNT_FLAG(MNT_ACLS);
3778 	MNT_FLAG(MNT_NOATIME);
3779 	MNT_FLAG(MNT_NOCLUSTERR);
3780 	MNT_FLAG(MNT_NOCLUSTERW);
3781 	MNT_FLAG(MNT_SUJ);
3782 	MNT_FLAG(MNT_EXRDONLY);
3783 	MNT_FLAG(MNT_EXPORTED);
3784 	MNT_FLAG(MNT_DEFEXPORTED);
3785 	MNT_FLAG(MNT_EXPORTANON);
3786 	MNT_FLAG(MNT_EXKERB);
3787 	MNT_FLAG(MNT_EXPUBLIC);
3788 	MNT_FLAG(MNT_LOCAL);
3789 	MNT_FLAG(MNT_QUOTA);
3790 	MNT_FLAG(MNT_ROOTFS);
3791 	MNT_FLAG(MNT_USER);
3792 	MNT_FLAG(MNT_IGNORE);
3793 	MNT_FLAG(MNT_UPDATE);
3794 	MNT_FLAG(MNT_DELEXPORT);
3795 	MNT_FLAG(MNT_RELOAD);
3796 	MNT_FLAG(MNT_FORCE);
3797 	MNT_FLAG(MNT_SNAPSHOT);
3798 	MNT_FLAG(MNT_BYFSID);
3799 #undef MNT_FLAG
3800 	if (mflags != 0) {
3801 		if (buf[0] != '\0')
3802 			strlcat(buf, ", ", sizeof(buf));
3803 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3804 		    "0x%016jx", mflags);
3805 	}
3806 	db_printf("    mnt_flag = %s\n", buf);
3807 
3808 	buf[0] = '\0';
3809 	flags = mp->mnt_kern_flag;
3810 #define	MNT_KERN_FLAG(flag)	do {					\
3811 	if (flags & (flag)) {						\
3812 		if (buf[0] != '\0')					\
3813 			strlcat(buf, ", ", sizeof(buf));		\
3814 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3815 		flags &= ~(flag);					\
3816 	}								\
3817 } while (0)
3818 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3819 	MNT_KERN_FLAG(MNTK_ASYNC);
3820 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3821 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3822 	MNT_KERN_FLAG(MNTK_DRAINING);
3823 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3824 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3825 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3826 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3827 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3828 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3829 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3830 	MNT_KERN_FLAG(MNTK_MARKER);
3831 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3832 	MNT_KERN_FLAG(MNTK_NOASYNC);
3833 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3834 	MNT_KERN_FLAG(MNTK_MWAIT);
3835 	MNT_KERN_FLAG(MNTK_SUSPEND);
3836 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3837 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3838 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3839 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3840 #undef MNT_KERN_FLAG
3841 	if (flags != 0) {
3842 		if (buf[0] != '\0')
3843 			strlcat(buf, ", ", sizeof(buf));
3844 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3845 		    "0x%08x", flags);
3846 	}
3847 	db_printf("    mnt_kern_flag = %s\n", buf);
3848 
3849 	db_printf("    mnt_opt = ");
3850 	opt = TAILQ_FIRST(mp->mnt_opt);
3851 	if (opt != NULL) {
3852 		db_printf("%s", opt->name);
3853 		opt = TAILQ_NEXT(opt, link);
3854 		while (opt != NULL) {
3855 			db_printf(", %s", opt->name);
3856 			opt = TAILQ_NEXT(opt, link);
3857 		}
3858 	}
3859 	db_printf("\n");
3860 
3861 	sp = &mp->mnt_stat;
3862 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3863 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3864 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3865 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3866 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3867 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3868 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3869 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3870 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3871 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3872 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3873 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3874 
3875 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3876 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3877 	if (jailed(mp->mnt_cred))
3878 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3879 	db_printf(" }\n");
3880 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3881 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3882 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3883 	db_printf("    mnt_activevnodelistsize = %d\n",
3884 	    mp->mnt_activevnodelistsize);
3885 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3886 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3887 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3888 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3889 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3890 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3891 	db_printf("    mnt_secondary_accwrites = %d\n",
3892 	    mp->mnt_secondary_accwrites);
3893 	db_printf("    mnt_gjprovider = %s\n",
3894 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3895 
3896 	db_printf("\n\nList of active vnodes\n");
3897 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3898 		if (vp->v_type != VMARKER) {
3899 			vn_printf(vp, "vnode ");
3900 			if (db_pager_quit)
3901 				break;
3902 		}
3903 	}
3904 	db_printf("\n\nList of inactive vnodes\n");
3905 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3906 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3907 			vn_printf(vp, "vnode ");
3908 			if (db_pager_quit)
3909 				break;
3910 		}
3911 	}
3912 }
3913 #endif	/* DDB */
3914 
3915 /*
3916  * Fill in a struct xvfsconf based on a struct vfsconf.
3917  */
3918 static int
3919 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3920 {
3921 	struct xvfsconf xvfsp;
3922 
3923 	bzero(&xvfsp, sizeof(xvfsp));
3924 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3925 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3926 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3927 	xvfsp.vfc_flags = vfsp->vfc_flags;
3928 	/*
3929 	 * These are unused in userland, we keep them
3930 	 * to not break binary compatibility.
3931 	 */
3932 	xvfsp.vfc_vfsops = NULL;
3933 	xvfsp.vfc_next = NULL;
3934 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3935 }
3936 
3937 #ifdef COMPAT_FREEBSD32
3938 struct xvfsconf32 {
3939 	uint32_t	vfc_vfsops;
3940 	char		vfc_name[MFSNAMELEN];
3941 	int32_t		vfc_typenum;
3942 	int32_t		vfc_refcount;
3943 	int32_t		vfc_flags;
3944 	uint32_t	vfc_next;
3945 };
3946 
3947 static int
3948 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3949 {
3950 	struct xvfsconf32 xvfsp;
3951 
3952 	bzero(&xvfsp, sizeof(xvfsp));
3953 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3954 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3955 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3956 	xvfsp.vfc_flags = vfsp->vfc_flags;
3957 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3958 }
3959 #endif
3960 
3961 /*
3962  * Top level filesystem related information gathering.
3963  */
3964 static int
3965 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3966 {
3967 	struct vfsconf *vfsp;
3968 	int error;
3969 
3970 	error = 0;
3971 	vfsconf_slock();
3972 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3973 #ifdef COMPAT_FREEBSD32
3974 		if (req->flags & SCTL_MASK32)
3975 			error = vfsconf2x32(req, vfsp);
3976 		else
3977 #endif
3978 			error = vfsconf2x(req, vfsp);
3979 		if (error)
3980 			break;
3981 	}
3982 	vfsconf_sunlock();
3983 	return (error);
3984 }
3985 
3986 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3987     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3988     "S,xvfsconf", "List of all configured filesystems");
3989 
3990 #ifndef BURN_BRIDGES
3991 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3992 
3993 static int
3994 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3995 {
3996 	int *name = (int *)arg1 - 1;	/* XXX */
3997 	u_int namelen = arg2 + 1;	/* XXX */
3998 	struct vfsconf *vfsp;
3999 
4000 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4001 	    "please rebuild world\n");
4002 
4003 #if 1 || defined(COMPAT_PRELITE2)
4004 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4005 	if (namelen == 1)
4006 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4007 #endif
4008 
4009 	switch (name[1]) {
4010 	case VFS_MAXTYPENUM:
4011 		if (namelen != 2)
4012 			return (ENOTDIR);
4013 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4014 	case VFS_CONF:
4015 		if (namelen != 3)
4016 			return (ENOTDIR);	/* overloaded */
4017 		vfsconf_slock();
4018 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4019 			if (vfsp->vfc_typenum == name[2])
4020 				break;
4021 		}
4022 		vfsconf_sunlock();
4023 		if (vfsp == NULL)
4024 			return (EOPNOTSUPP);
4025 #ifdef COMPAT_FREEBSD32
4026 		if (req->flags & SCTL_MASK32)
4027 			return (vfsconf2x32(req, vfsp));
4028 		else
4029 #endif
4030 			return (vfsconf2x(req, vfsp));
4031 	}
4032 	return (EOPNOTSUPP);
4033 }
4034 
4035 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4036     CTLFLAG_MPSAFE, vfs_sysctl,
4037     "Generic filesystem");
4038 
4039 #if 1 || defined(COMPAT_PRELITE2)
4040 
4041 static int
4042 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4043 {
4044 	int error;
4045 	struct vfsconf *vfsp;
4046 	struct ovfsconf ovfs;
4047 
4048 	vfsconf_slock();
4049 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4050 		bzero(&ovfs, sizeof(ovfs));
4051 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4052 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4053 		ovfs.vfc_index = vfsp->vfc_typenum;
4054 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4055 		ovfs.vfc_flags = vfsp->vfc_flags;
4056 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4057 		if (error != 0) {
4058 			vfsconf_sunlock();
4059 			return (error);
4060 		}
4061 	}
4062 	vfsconf_sunlock();
4063 	return (0);
4064 }
4065 
4066 #endif /* 1 || COMPAT_PRELITE2 */
4067 #endif /* !BURN_BRIDGES */
4068 
4069 #define KINFO_VNODESLOP		10
4070 #ifdef notyet
4071 /*
4072  * Dump vnode list (via sysctl).
4073  */
4074 /* ARGSUSED */
4075 static int
4076 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4077 {
4078 	struct xvnode *xvn;
4079 	struct mount *mp;
4080 	struct vnode *vp;
4081 	int error, len, n;
4082 
4083 	/*
4084 	 * Stale numvnodes access is not fatal here.
4085 	 */
4086 	req->lock = 0;
4087 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4088 	if (!req->oldptr)
4089 		/* Make an estimate */
4090 		return (SYSCTL_OUT(req, 0, len));
4091 
4092 	error = sysctl_wire_old_buffer(req, 0);
4093 	if (error != 0)
4094 		return (error);
4095 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4096 	n = 0;
4097 	mtx_lock(&mountlist_mtx);
4098 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4099 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4100 			continue;
4101 		MNT_ILOCK(mp);
4102 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4103 			if (n == len)
4104 				break;
4105 			vref(vp);
4106 			xvn[n].xv_size = sizeof *xvn;
4107 			xvn[n].xv_vnode = vp;
4108 			xvn[n].xv_id = 0;	/* XXX compat */
4109 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4110 			XV_COPY(usecount);
4111 			XV_COPY(writecount);
4112 			XV_COPY(holdcnt);
4113 			XV_COPY(mount);
4114 			XV_COPY(numoutput);
4115 			XV_COPY(type);
4116 #undef XV_COPY
4117 			xvn[n].xv_flag = vp->v_vflag;
4118 
4119 			switch (vp->v_type) {
4120 			case VREG:
4121 			case VDIR:
4122 			case VLNK:
4123 				break;
4124 			case VBLK:
4125 			case VCHR:
4126 				if (vp->v_rdev == NULL) {
4127 					vrele(vp);
4128 					continue;
4129 				}
4130 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4131 				break;
4132 			case VSOCK:
4133 				xvn[n].xv_socket = vp->v_socket;
4134 				break;
4135 			case VFIFO:
4136 				xvn[n].xv_fifo = vp->v_fifoinfo;
4137 				break;
4138 			case VNON:
4139 			case VBAD:
4140 			default:
4141 				/* shouldn't happen? */
4142 				vrele(vp);
4143 				continue;
4144 			}
4145 			vrele(vp);
4146 			++n;
4147 		}
4148 		MNT_IUNLOCK(mp);
4149 		mtx_lock(&mountlist_mtx);
4150 		vfs_unbusy(mp);
4151 		if (n == len)
4152 			break;
4153 	}
4154 	mtx_unlock(&mountlist_mtx);
4155 
4156 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4157 	free(xvn, M_TEMP);
4158 	return (error);
4159 }
4160 
4161 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4162     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4163     "");
4164 #endif
4165 
4166 static void
4167 unmount_or_warn(struct mount *mp)
4168 {
4169 	int error;
4170 
4171 	error = dounmount(mp, MNT_FORCE, curthread);
4172 	if (error != 0) {
4173 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4174 		if (error == EBUSY)
4175 			printf("BUSY)\n");
4176 		else
4177 			printf("%d)\n", error);
4178 	}
4179 }
4180 
4181 /*
4182  * Unmount all filesystems. The list is traversed in reverse order
4183  * of mounting to avoid dependencies.
4184  */
4185 void
4186 vfs_unmountall(void)
4187 {
4188 	struct mount *mp, *tmp;
4189 
4190 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4191 
4192 	/*
4193 	 * Since this only runs when rebooting, it is not interlocked.
4194 	 */
4195 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4196 		vfs_ref(mp);
4197 
4198 		/*
4199 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4200 		 * unmount of the latter.
4201 		 */
4202 		if (mp == rootdevmp)
4203 			continue;
4204 
4205 		unmount_or_warn(mp);
4206 	}
4207 
4208 	if (rootdevmp != NULL)
4209 		unmount_or_warn(rootdevmp);
4210 }
4211 
4212 /*
4213  * perform msync on all vnodes under a mount point
4214  * the mount point must be locked.
4215  */
4216 void
4217 vfs_msync(struct mount *mp, int flags)
4218 {
4219 	struct vnode *vp, *mvp;
4220 	struct vm_object *obj;
4221 
4222 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4223 
4224 	vnlru_return_batch(mp);
4225 
4226 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4227 		obj = vp->v_object;
4228 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4229 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4230 			if (!vget(vp,
4231 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4232 			    curthread)) {
4233 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4234 					vput(vp);
4235 					continue;
4236 				}
4237 
4238 				obj = vp->v_object;
4239 				if (obj != NULL) {
4240 					VM_OBJECT_WLOCK(obj);
4241 					vm_object_page_clean(obj, 0, 0,
4242 					    flags == MNT_WAIT ?
4243 					    OBJPC_SYNC : OBJPC_NOSYNC);
4244 					VM_OBJECT_WUNLOCK(obj);
4245 				}
4246 				vput(vp);
4247 			}
4248 		} else
4249 			VI_UNLOCK(vp);
4250 	}
4251 }
4252 
4253 static void
4254 destroy_vpollinfo_free(struct vpollinfo *vi)
4255 {
4256 
4257 	knlist_destroy(&vi->vpi_selinfo.si_note);
4258 	mtx_destroy(&vi->vpi_lock);
4259 	uma_zfree(vnodepoll_zone, vi);
4260 }
4261 
4262 static void
4263 destroy_vpollinfo(struct vpollinfo *vi)
4264 {
4265 
4266 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4267 	seldrain(&vi->vpi_selinfo);
4268 	destroy_vpollinfo_free(vi);
4269 }
4270 
4271 /*
4272  * Initialize per-vnode helper structure to hold poll-related state.
4273  */
4274 void
4275 v_addpollinfo(struct vnode *vp)
4276 {
4277 	struct vpollinfo *vi;
4278 
4279 	if (vp->v_pollinfo != NULL)
4280 		return;
4281 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4282 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4283 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4284 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4285 	VI_LOCK(vp);
4286 	if (vp->v_pollinfo != NULL) {
4287 		VI_UNLOCK(vp);
4288 		destroy_vpollinfo_free(vi);
4289 		return;
4290 	}
4291 	vp->v_pollinfo = vi;
4292 	VI_UNLOCK(vp);
4293 }
4294 
4295 /*
4296  * Record a process's interest in events which might happen to
4297  * a vnode.  Because poll uses the historic select-style interface
4298  * internally, this routine serves as both the ``check for any
4299  * pending events'' and the ``record my interest in future events''
4300  * functions.  (These are done together, while the lock is held,
4301  * to avoid race conditions.)
4302  */
4303 int
4304 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4305 {
4306 
4307 	v_addpollinfo(vp);
4308 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4309 	if (vp->v_pollinfo->vpi_revents & events) {
4310 		/*
4311 		 * This leaves events we are not interested
4312 		 * in available for the other process which
4313 		 * which presumably had requested them
4314 		 * (otherwise they would never have been
4315 		 * recorded).
4316 		 */
4317 		events &= vp->v_pollinfo->vpi_revents;
4318 		vp->v_pollinfo->vpi_revents &= ~events;
4319 
4320 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4321 		return (events);
4322 	}
4323 	vp->v_pollinfo->vpi_events |= events;
4324 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4325 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4326 	return (0);
4327 }
4328 
4329 /*
4330  * Routine to create and manage a filesystem syncer vnode.
4331  */
4332 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4333 static int	sync_fsync(struct  vop_fsync_args *);
4334 static int	sync_inactive(struct  vop_inactive_args *);
4335 static int	sync_reclaim(struct  vop_reclaim_args *);
4336 
4337 static struct vop_vector sync_vnodeops = {
4338 	.vop_bypass =	VOP_EOPNOTSUPP,
4339 	.vop_close =	sync_close,		/* close */
4340 	.vop_fsync =	sync_fsync,		/* fsync */
4341 	.vop_inactive =	sync_inactive,	/* inactive */
4342 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4343 	.vop_lock1 =	vop_stdlock,	/* lock */
4344 	.vop_unlock =	vop_stdunlock,	/* unlock */
4345 	.vop_islocked =	vop_stdislocked,	/* islocked */
4346 };
4347 
4348 /*
4349  * Create a new filesystem syncer vnode for the specified mount point.
4350  */
4351 void
4352 vfs_allocate_syncvnode(struct mount *mp)
4353 {
4354 	struct vnode *vp;
4355 	struct bufobj *bo;
4356 	static long start, incr, next;
4357 	int error;
4358 
4359 	/* Allocate a new vnode */
4360 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4361 	if (error != 0)
4362 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4363 	vp->v_type = VNON;
4364 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4365 	vp->v_vflag |= VV_FORCEINSMQ;
4366 	error = insmntque(vp, mp);
4367 	if (error != 0)
4368 		panic("vfs_allocate_syncvnode: insmntque() failed");
4369 	vp->v_vflag &= ~VV_FORCEINSMQ;
4370 	VOP_UNLOCK(vp, 0);
4371 	/*
4372 	 * Place the vnode onto the syncer worklist. We attempt to
4373 	 * scatter them about on the list so that they will go off
4374 	 * at evenly distributed times even if all the filesystems
4375 	 * are mounted at once.
4376 	 */
4377 	next += incr;
4378 	if (next == 0 || next > syncer_maxdelay) {
4379 		start /= 2;
4380 		incr /= 2;
4381 		if (start == 0) {
4382 			start = syncer_maxdelay / 2;
4383 			incr = syncer_maxdelay;
4384 		}
4385 		next = start;
4386 	}
4387 	bo = &vp->v_bufobj;
4388 	BO_LOCK(bo);
4389 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4390 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4391 	mtx_lock(&sync_mtx);
4392 	sync_vnode_count++;
4393 	if (mp->mnt_syncer == NULL) {
4394 		mp->mnt_syncer = vp;
4395 		vp = NULL;
4396 	}
4397 	mtx_unlock(&sync_mtx);
4398 	BO_UNLOCK(bo);
4399 	if (vp != NULL) {
4400 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4401 		vgone(vp);
4402 		vput(vp);
4403 	}
4404 }
4405 
4406 void
4407 vfs_deallocate_syncvnode(struct mount *mp)
4408 {
4409 	struct vnode *vp;
4410 
4411 	mtx_lock(&sync_mtx);
4412 	vp = mp->mnt_syncer;
4413 	if (vp != NULL)
4414 		mp->mnt_syncer = NULL;
4415 	mtx_unlock(&sync_mtx);
4416 	if (vp != NULL)
4417 		vrele(vp);
4418 }
4419 
4420 /*
4421  * Do a lazy sync of the filesystem.
4422  */
4423 static int
4424 sync_fsync(struct vop_fsync_args *ap)
4425 {
4426 	struct vnode *syncvp = ap->a_vp;
4427 	struct mount *mp = syncvp->v_mount;
4428 	int error, save;
4429 	struct bufobj *bo;
4430 
4431 	/*
4432 	 * We only need to do something if this is a lazy evaluation.
4433 	 */
4434 	if (ap->a_waitfor != MNT_LAZY)
4435 		return (0);
4436 
4437 	/*
4438 	 * Move ourselves to the back of the sync list.
4439 	 */
4440 	bo = &syncvp->v_bufobj;
4441 	BO_LOCK(bo);
4442 	vn_syncer_add_to_worklist(bo, syncdelay);
4443 	BO_UNLOCK(bo);
4444 
4445 	/*
4446 	 * Walk the list of vnodes pushing all that are dirty and
4447 	 * not already on the sync list.
4448 	 */
4449 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4450 		return (0);
4451 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4452 		vfs_unbusy(mp);
4453 		return (0);
4454 	}
4455 	save = curthread_pflags_set(TDP_SYNCIO);
4456 	vfs_msync(mp, MNT_NOWAIT);
4457 	error = VFS_SYNC(mp, MNT_LAZY);
4458 	curthread_pflags_restore(save);
4459 	vn_finished_write(mp);
4460 	vfs_unbusy(mp);
4461 	return (error);
4462 }
4463 
4464 /*
4465  * The syncer vnode is no referenced.
4466  */
4467 static int
4468 sync_inactive(struct vop_inactive_args *ap)
4469 {
4470 
4471 	vgone(ap->a_vp);
4472 	return (0);
4473 }
4474 
4475 /*
4476  * The syncer vnode is no longer needed and is being decommissioned.
4477  *
4478  * Modifications to the worklist must be protected by sync_mtx.
4479  */
4480 static int
4481 sync_reclaim(struct vop_reclaim_args *ap)
4482 {
4483 	struct vnode *vp = ap->a_vp;
4484 	struct bufobj *bo;
4485 
4486 	bo = &vp->v_bufobj;
4487 	BO_LOCK(bo);
4488 	mtx_lock(&sync_mtx);
4489 	if (vp->v_mount->mnt_syncer == vp)
4490 		vp->v_mount->mnt_syncer = NULL;
4491 	if (bo->bo_flag & BO_ONWORKLST) {
4492 		LIST_REMOVE(bo, bo_synclist);
4493 		syncer_worklist_len--;
4494 		sync_vnode_count--;
4495 		bo->bo_flag &= ~BO_ONWORKLST;
4496 	}
4497 	mtx_unlock(&sync_mtx);
4498 	BO_UNLOCK(bo);
4499 
4500 	return (0);
4501 }
4502 
4503 /*
4504  * Check if vnode represents a disk device
4505  */
4506 int
4507 vn_isdisk(struct vnode *vp, int *errp)
4508 {
4509 	int error;
4510 
4511 	if (vp->v_type != VCHR) {
4512 		error = ENOTBLK;
4513 		goto out;
4514 	}
4515 	error = 0;
4516 	dev_lock();
4517 	if (vp->v_rdev == NULL)
4518 		error = ENXIO;
4519 	else if (vp->v_rdev->si_devsw == NULL)
4520 		error = ENXIO;
4521 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4522 		error = ENOTBLK;
4523 	dev_unlock();
4524 out:
4525 	if (errp != NULL)
4526 		*errp = error;
4527 	return (error == 0);
4528 }
4529 
4530 /*
4531  * Common filesystem object access control check routine.  Accepts a
4532  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4533  * and optional call-by-reference privused argument allowing vaccess()
4534  * to indicate to the caller whether privilege was used to satisfy the
4535  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4536  */
4537 int
4538 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4539     accmode_t accmode, struct ucred *cred, int *privused)
4540 {
4541 	accmode_t dac_granted;
4542 	accmode_t priv_granted;
4543 
4544 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4545 	    ("invalid bit in accmode"));
4546 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4547 	    ("VAPPEND without VWRITE"));
4548 
4549 	/*
4550 	 * Look for a normal, non-privileged way to access the file/directory
4551 	 * as requested.  If it exists, go with that.
4552 	 */
4553 
4554 	if (privused != NULL)
4555 		*privused = 0;
4556 
4557 	dac_granted = 0;
4558 
4559 	/* Check the owner. */
4560 	if (cred->cr_uid == file_uid) {
4561 		dac_granted |= VADMIN;
4562 		if (file_mode & S_IXUSR)
4563 			dac_granted |= VEXEC;
4564 		if (file_mode & S_IRUSR)
4565 			dac_granted |= VREAD;
4566 		if (file_mode & S_IWUSR)
4567 			dac_granted |= (VWRITE | VAPPEND);
4568 
4569 		if ((accmode & dac_granted) == accmode)
4570 			return (0);
4571 
4572 		goto privcheck;
4573 	}
4574 
4575 	/* Otherwise, check the groups (first match) */
4576 	if (groupmember(file_gid, cred)) {
4577 		if (file_mode & S_IXGRP)
4578 			dac_granted |= VEXEC;
4579 		if (file_mode & S_IRGRP)
4580 			dac_granted |= VREAD;
4581 		if (file_mode & S_IWGRP)
4582 			dac_granted |= (VWRITE | VAPPEND);
4583 
4584 		if ((accmode & dac_granted) == accmode)
4585 			return (0);
4586 
4587 		goto privcheck;
4588 	}
4589 
4590 	/* Otherwise, check everyone else. */
4591 	if (file_mode & S_IXOTH)
4592 		dac_granted |= VEXEC;
4593 	if (file_mode & S_IROTH)
4594 		dac_granted |= VREAD;
4595 	if (file_mode & S_IWOTH)
4596 		dac_granted |= (VWRITE | VAPPEND);
4597 	if ((accmode & dac_granted) == accmode)
4598 		return (0);
4599 
4600 privcheck:
4601 	/*
4602 	 * Build a privilege mask to determine if the set of privileges
4603 	 * satisfies the requirements when combined with the granted mask
4604 	 * from above.  For each privilege, if the privilege is required,
4605 	 * bitwise or the request type onto the priv_granted mask.
4606 	 */
4607 	priv_granted = 0;
4608 
4609 	if (type == VDIR) {
4610 		/*
4611 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4612 		 * requests, instead of PRIV_VFS_EXEC.
4613 		 */
4614 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4615 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4616 			priv_granted |= VEXEC;
4617 	} else {
4618 		/*
4619 		 * Ensure that at least one execute bit is on. Otherwise,
4620 		 * a privileged user will always succeed, and we don't want
4621 		 * this to happen unless the file really is executable.
4622 		 */
4623 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4624 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4625 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4626 			priv_granted |= VEXEC;
4627 	}
4628 
4629 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4630 	    !priv_check_cred(cred, PRIV_VFS_READ))
4631 		priv_granted |= VREAD;
4632 
4633 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4634 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4635 		priv_granted |= (VWRITE | VAPPEND);
4636 
4637 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4638 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4639 		priv_granted |= VADMIN;
4640 
4641 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4642 		/* XXX audit: privilege used */
4643 		if (privused != NULL)
4644 			*privused = 1;
4645 		return (0);
4646 	}
4647 
4648 	return ((accmode & VADMIN) ? EPERM : EACCES);
4649 }
4650 
4651 /*
4652  * Credential check based on process requesting service, and per-attribute
4653  * permissions.
4654  */
4655 int
4656 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4657     struct thread *td, accmode_t accmode)
4658 {
4659 
4660 	/*
4661 	 * Kernel-invoked always succeeds.
4662 	 */
4663 	if (cred == NOCRED)
4664 		return (0);
4665 
4666 	/*
4667 	 * Do not allow privileged processes in jail to directly manipulate
4668 	 * system attributes.
4669 	 */
4670 	switch (attrnamespace) {
4671 	case EXTATTR_NAMESPACE_SYSTEM:
4672 		/* Potentially should be: return (EPERM); */
4673 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4674 	case EXTATTR_NAMESPACE_USER:
4675 		return (VOP_ACCESS(vp, accmode, cred, td));
4676 	default:
4677 		return (EPERM);
4678 	}
4679 }
4680 
4681 #ifdef DEBUG_VFS_LOCKS
4682 /*
4683  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4684  * no longer ok to have an unlocked VFS.
4685  */
4686 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4687 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4688 
4689 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4690 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4691     "Drop into debugger on lock violation");
4692 
4693 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4694 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4695     0, "Check for interlock across VOPs");
4696 
4697 int vfs_badlock_print = 1;	/* Print lock violations. */
4698 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4699     0, "Print lock violations");
4700 
4701 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4702 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4703     0, "Print vnode details on lock violations");
4704 
4705 #ifdef KDB
4706 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4707 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4708     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4709 #endif
4710 
4711 static void
4712 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4713 {
4714 
4715 #ifdef KDB
4716 	if (vfs_badlock_backtrace)
4717 		kdb_backtrace();
4718 #endif
4719 	if (vfs_badlock_vnode)
4720 		vn_printf(vp, "vnode ");
4721 	if (vfs_badlock_print)
4722 		printf("%s: %p %s\n", str, (void *)vp, msg);
4723 	if (vfs_badlock_ddb)
4724 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4725 }
4726 
4727 void
4728 assert_vi_locked(struct vnode *vp, const char *str)
4729 {
4730 
4731 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4732 		vfs_badlock("interlock is not locked but should be", str, vp);
4733 }
4734 
4735 void
4736 assert_vi_unlocked(struct vnode *vp, const char *str)
4737 {
4738 
4739 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4740 		vfs_badlock("interlock is locked but should not be", str, vp);
4741 }
4742 
4743 void
4744 assert_vop_locked(struct vnode *vp, const char *str)
4745 {
4746 	int locked;
4747 
4748 	if (!IGNORE_LOCK(vp)) {
4749 		locked = VOP_ISLOCKED(vp);
4750 		if (locked == 0 || locked == LK_EXCLOTHER)
4751 			vfs_badlock("is not locked but should be", str, vp);
4752 	}
4753 }
4754 
4755 void
4756 assert_vop_unlocked(struct vnode *vp, const char *str)
4757 {
4758 
4759 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4760 		vfs_badlock("is locked but should not be", str, vp);
4761 }
4762 
4763 void
4764 assert_vop_elocked(struct vnode *vp, const char *str)
4765 {
4766 
4767 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4768 		vfs_badlock("is not exclusive locked but should be", str, vp);
4769 }
4770 #endif /* DEBUG_VFS_LOCKS */
4771 
4772 void
4773 vop_rename_fail(struct vop_rename_args *ap)
4774 {
4775 
4776 	if (ap->a_tvp != NULL)
4777 		vput(ap->a_tvp);
4778 	if (ap->a_tdvp == ap->a_tvp)
4779 		vrele(ap->a_tdvp);
4780 	else
4781 		vput(ap->a_tdvp);
4782 	vrele(ap->a_fdvp);
4783 	vrele(ap->a_fvp);
4784 }
4785 
4786 void
4787 vop_rename_pre(void *ap)
4788 {
4789 	struct vop_rename_args *a = ap;
4790 
4791 #ifdef DEBUG_VFS_LOCKS
4792 	if (a->a_tvp)
4793 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4794 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4795 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4796 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4797 
4798 	/* Check the source (from). */
4799 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4800 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4801 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4802 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4803 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4804 
4805 	/* Check the target. */
4806 	if (a->a_tvp)
4807 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4808 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4809 #endif
4810 	if (a->a_tdvp != a->a_fdvp)
4811 		vhold(a->a_fdvp);
4812 	if (a->a_tvp != a->a_fvp)
4813 		vhold(a->a_fvp);
4814 	vhold(a->a_tdvp);
4815 	if (a->a_tvp)
4816 		vhold(a->a_tvp);
4817 }
4818 
4819 #ifdef DEBUG_VFS_LOCKS
4820 void
4821 vop_strategy_pre(void *ap)
4822 {
4823 	struct vop_strategy_args *a;
4824 	struct buf *bp;
4825 
4826 	a = ap;
4827 	bp = a->a_bp;
4828 
4829 	/*
4830 	 * Cluster ops lock their component buffers but not the IO container.
4831 	 */
4832 	if ((bp->b_flags & B_CLUSTER) != 0)
4833 		return;
4834 
4835 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4836 		if (vfs_badlock_print)
4837 			printf(
4838 			    "VOP_STRATEGY: bp is not locked but should be\n");
4839 		if (vfs_badlock_ddb)
4840 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4841 	}
4842 }
4843 
4844 void
4845 vop_lock_pre(void *ap)
4846 {
4847 	struct vop_lock1_args *a = ap;
4848 
4849 	if ((a->a_flags & LK_INTERLOCK) == 0)
4850 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4851 	else
4852 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4853 }
4854 
4855 void
4856 vop_lock_post(void *ap, int rc)
4857 {
4858 	struct vop_lock1_args *a = ap;
4859 
4860 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4861 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4862 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4863 }
4864 
4865 void
4866 vop_unlock_pre(void *ap)
4867 {
4868 	struct vop_unlock_args *a = ap;
4869 
4870 	if (a->a_flags & LK_INTERLOCK)
4871 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4872 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4873 }
4874 
4875 void
4876 vop_unlock_post(void *ap, int rc)
4877 {
4878 	struct vop_unlock_args *a = ap;
4879 
4880 	if (a->a_flags & LK_INTERLOCK)
4881 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4882 }
4883 #endif
4884 
4885 void
4886 vop_create_post(void *ap, int rc)
4887 {
4888 	struct vop_create_args *a = ap;
4889 
4890 	if (!rc)
4891 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4892 }
4893 
4894 void
4895 vop_deleteextattr_post(void *ap, int rc)
4896 {
4897 	struct vop_deleteextattr_args *a = ap;
4898 
4899 	if (!rc)
4900 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4901 }
4902 
4903 void
4904 vop_link_post(void *ap, int rc)
4905 {
4906 	struct vop_link_args *a = ap;
4907 
4908 	if (!rc) {
4909 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4910 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4911 	}
4912 }
4913 
4914 void
4915 vop_mkdir_post(void *ap, int rc)
4916 {
4917 	struct vop_mkdir_args *a = ap;
4918 
4919 	if (!rc)
4920 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4921 }
4922 
4923 void
4924 vop_mknod_post(void *ap, int rc)
4925 {
4926 	struct vop_mknod_args *a = ap;
4927 
4928 	if (!rc)
4929 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4930 }
4931 
4932 void
4933 vop_reclaim_post(void *ap, int rc)
4934 {
4935 	struct vop_reclaim_args *a = ap;
4936 
4937 	if (!rc)
4938 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4939 }
4940 
4941 void
4942 vop_remove_post(void *ap, int rc)
4943 {
4944 	struct vop_remove_args *a = ap;
4945 
4946 	if (!rc) {
4947 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4948 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4949 	}
4950 }
4951 
4952 void
4953 vop_rename_post(void *ap, int rc)
4954 {
4955 	struct vop_rename_args *a = ap;
4956 	long hint;
4957 
4958 	if (!rc) {
4959 		hint = NOTE_WRITE;
4960 		if (a->a_fdvp == a->a_tdvp) {
4961 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4962 				hint |= NOTE_LINK;
4963 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4964 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4965 		} else {
4966 			hint |= NOTE_EXTEND;
4967 			if (a->a_fvp->v_type == VDIR)
4968 				hint |= NOTE_LINK;
4969 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4970 
4971 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4972 			    a->a_tvp->v_type == VDIR)
4973 				hint &= ~NOTE_LINK;
4974 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4975 		}
4976 
4977 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4978 		if (a->a_tvp)
4979 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4980 	}
4981 	if (a->a_tdvp != a->a_fdvp)
4982 		vdrop(a->a_fdvp);
4983 	if (a->a_tvp != a->a_fvp)
4984 		vdrop(a->a_fvp);
4985 	vdrop(a->a_tdvp);
4986 	if (a->a_tvp)
4987 		vdrop(a->a_tvp);
4988 }
4989 
4990 void
4991 vop_rmdir_post(void *ap, int rc)
4992 {
4993 	struct vop_rmdir_args *a = ap;
4994 
4995 	if (!rc) {
4996 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4997 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4998 	}
4999 }
5000 
5001 void
5002 vop_setattr_post(void *ap, int rc)
5003 {
5004 	struct vop_setattr_args *a = ap;
5005 
5006 	if (!rc)
5007 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5008 }
5009 
5010 void
5011 vop_setextattr_post(void *ap, int rc)
5012 {
5013 	struct vop_setextattr_args *a = ap;
5014 
5015 	if (!rc)
5016 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5017 }
5018 
5019 void
5020 vop_symlink_post(void *ap, int rc)
5021 {
5022 	struct vop_symlink_args *a = ap;
5023 
5024 	if (!rc)
5025 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5026 }
5027 
5028 void
5029 vop_open_post(void *ap, int rc)
5030 {
5031 	struct vop_open_args *a = ap;
5032 
5033 	if (!rc)
5034 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5035 }
5036 
5037 void
5038 vop_close_post(void *ap, int rc)
5039 {
5040 	struct vop_close_args *a = ap;
5041 
5042 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5043 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
5044 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5045 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5046 	}
5047 }
5048 
5049 void
5050 vop_read_post(void *ap, int rc)
5051 {
5052 	struct vop_read_args *a = ap;
5053 
5054 	if (!rc)
5055 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5056 }
5057 
5058 void
5059 vop_readdir_post(void *ap, int rc)
5060 {
5061 	struct vop_readdir_args *a = ap;
5062 
5063 	if (!rc)
5064 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5065 }
5066 
5067 static struct knlist fs_knlist;
5068 
5069 static void
5070 vfs_event_init(void *arg)
5071 {
5072 	knlist_init_mtx(&fs_knlist, NULL);
5073 }
5074 /* XXX - correct order? */
5075 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5076 
5077 void
5078 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5079 {
5080 
5081 	KNOTE_UNLOCKED(&fs_knlist, event);
5082 }
5083 
5084 static int	filt_fsattach(struct knote *kn);
5085 static void	filt_fsdetach(struct knote *kn);
5086 static int	filt_fsevent(struct knote *kn, long hint);
5087 
5088 struct filterops fs_filtops = {
5089 	.f_isfd = 0,
5090 	.f_attach = filt_fsattach,
5091 	.f_detach = filt_fsdetach,
5092 	.f_event = filt_fsevent
5093 };
5094 
5095 static int
5096 filt_fsattach(struct knote *kn)
5097 {
5098 
5099 	kn->kn_flags |= EV_CLEAR;
5100 	knlist_add(&fs_knlist, kn, 0);
5101 	return (0);
5102 }
5103 
5104 static void
5105 filt_fsdetach(struct knote *kn)
5106 {
5107 
5108 	knlist_remove(&fs_knlist, kn, 0);
5109 }
5110 
5111 static int
5112 filt_fsevent(struct knote *kn, long hint)
5113 {
5114 
5115 	kn->kn_fflags |= hint;
5116 	return (kn->kn_fflags != 0);
5117 }
5118 
5119 static int
5120 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5121 {
5122 	struct vfsidctl vc;
5123 	int error;
5124 	struct mount *mp;
5125 
5126 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5127 	if (error)
5128 		return (error);
5129 	if (vc.vc_vers != VFS_CTL_VERS1)
5130 		return (EINVAL);
5131 	mp = vfs_getvfs(&vc.vc_fsid);
5132 	if (mp == NULL)
5133 		return (ENOENT);
5134 	/* ensure that a specific sysctl goes to the right filesystem. */
5135 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5136 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5137 		vfs_rel(mp);
5138 		return (EINVAL);
5139 	}
5140 	VCTLTOREQ(&vc, req);
5141 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5142 	vfs_rel(mp);
5143 	return (error);
5144 }
5145 
5146 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5147     NULL, 0, sysctl_vfs_ctl, "",
5148     "Sysctl by fsid");
5149 
5150 /*
5151  * Function to initialize a va_filerev field sensibly.
5152  * XXX: Wouldn't a random number make a lot more sense ??
5153  */
5154 u_quad_t
5155 init_va_filerev(void)
5156 {
5157 	struct bintime bt;
5158 
5159 	getbinuptime(&bt);
5160 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5161 }
5162 
5163 static int	filt_vfsread(struct knote *kn, long hint);
5164 static int	filt_vfswrite(struct knote *kn, long hint);
5165 static int	filt_vfsvnode(struct knote *kn, long hint);
5166 static void	filt_vfsdetach(struct knote *kn);
5167 static struct filterops vfsread_filtops = {
5168 	.f_isfd = 1,
5169 	.f_detach = filt_vfsdetach,
5170 	.f_event = filt_vfsread
5171 };
5172 static struct filterops vfswrite_filtops = {
5173 	.f_isfd = 1,
5174 	.f_detach = filt_vfsdetach,
5175 	.f_event = filt_vfswrite
5176 };
5177 static struct filterops vfsvnode_filtops = {
5178 	.f_isfd = 1,
5179 	.f_detach = filt_vfsdetach,
5180 	.f_event = filt_vfsvnode
5181 };
5182 
5183 static void
5184 vfs_knllock(void *arg)
5185 {
5186 	struct vnode *vp = arg;
5187 
5188 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5189 }
5190 
5191 static void
5192 vfs_knlunlock(void *arg)
5193 {
5194 	struct vnode *vp = arg;
5195 
5196 	VOP_UNLOCK(vp, 0);
5197 }
5198 
5199 static void
5200 vfs_knl_assert_locked(void *arg)
5201 {
5202 #ifdef DEBUG_VFS_LOCKS
5203 	struct vnode *vp = arg;
5204 
5205 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5206 #endif
5207 }
5208 
5209 static void
5210 vfs_knl_assert_unlocked(void *arg)
5211 {
5212 #ifdef DEBUG_VFS_LOCKS
5213 	struct vnode *vp = arg;
5214 
5215 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5216 #endif
5217 }
5218 
5219 int
5220 vfs_kqfilter(struct vop_kqfilter_args *ap)
5221 {
5222 	struct vnode *vp = ap->a_vp;
5223 	struct knote *kn = ap->a_kn;
5224 	struct knlist *knl;
5225 
5226 	switch (kn->kn_filter) {
5227 	case EVFILT_READ:
5228 		kn->kn_fop = &vfsread_filtops;
5229 		break;
5230 	case EVFILT_WRITE:
5231 		kn->kn_fop = &vfswrite_filtops;
5232 		break;
5233 	case EVFILT_VNODE:
5234 		kn->kn_fop = &vfsvnode_filtops;
5235 		break;
5236 	default:
5237 		return (EINVAL);
5238 	}
5239 
5240 	kn->kn_hook = (caddr_t)vp;
5241 
5242 	v_addpollinfo(vp);
5243 	if (vp->v_pollinfo == NULL)
5244 		return (ENOMEM);
5245 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5246 	vhold(vp);
5247 	knlist_add(knl, kn, 0);
5248 
5249 	return (0);
5250 }
5251 
5252 /*
5253  * Detach knote from vnode
5254  */
5255 static void
5256 filt_vfsdetach(struct knote *kn)
5257 {
5258 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5259 
5260 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5261 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5262 	vdrop(vp);
5263 }
5264 
5265 /*ARGSUSED*/
5266 static int
5267 filt_vfsread(struct knote *kn, long hint)
5268 {
5269 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5270 	struct vattr va;
5271 	int res;
5272 
5273 	/*
5274 	 * filesystem is gone, so set the EOF flag and schedule
5275 	 * the knote for deletion.
5276 	 */
5277 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5278 		VI_LOCK(vp);
5279 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5280 		VI_UNLOCK(vp);
5281 		return (1);
5282 	}
5283 
5284 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5285 		return (0);
5286 
5287 	VI_LOCK(vp);
5288 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5289 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5290 	VI_UNLOCK(vp);
5291 	return (res);
5292 }
5293 
5294 /*ARGSUSED*/
5295 static int
5296 filt_vfswrite(struct knote *kn, long hint)
5297 {
5298 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5299 
5300 	VI_LOCK(vp);
5301 
5302 	/*
5303 	 * filesystem is gone, so set the EOF flag and schedule
5304 	 * the knote for deletion.
5305 	 */
5306 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5307 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5308 
5309 	kn->kn_data = 0;
5310 	VI_UNLOCK(vp);
5311 	return (1);
5312 }
5313 
5314 static int
5315 filt_vfsvnode(struct knote *kn, long hint)
5316 {
5317 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5318 	int res;
5319 
5320 	VI_LOCK(vp);
5321 	if (kn->kn_sfflags & hint)
5322 		kn->kn_fflags |= hint;
5323 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5324 		kn->kn_flags |= EV_EOF;
5325 		VI_UNLOCK(vp);
5326 		return (1);
5327 	}
5328 	res = (kn->kn_fflags != 0);
5329 	VI_UNLOCK(vp);
5330 	return (res);
5331 }
5332 
5333 int
5334 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5335 {
5336 	int error;
5337 
5338 	if (dp->d_reclen > ap->a_uio->uio_resid)
5339 		return (ENAMETOOLONG);
5340 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5341 	if (error) {
5342 		if (ap->a_ncookies != NULL) {
5343 			if (ap->a_cookies != NULL)
5344 				free(ap->a_cookies, M_TEMP);
5345 			ap->a_cookies = NULL;
5346 			*ap->a_ncookies = 0;
5347 		}
5348 		return (error);
5349 	}
5350 	if (ap->a_ncookies == NULL)
5351 		return (0);
5352 
5353 	KASSERT(ap->a_cookies,
5354 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5355 
5356 	*ap->a_cookies = realloc(*ap->a_cookies,
5357 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5358 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5359 	*ap->a_ncookies += 1;
5360 	return (0);
5361 }
5362 
5363 /*
5364  * Mark for update the access time of the file if the filesystem
5365  * supports VOP_MARKATIME.  This functionality is used by execve and
5366  * mmap, so we want to avoid the I/O implied by directly setting
5367  * va_atime for the sake of efficiency.
5368  */
5369 void
5370 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5371 {
5372 	struct mount *mp;
5373 
5374 	mp = vp->v_mount;
5375 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5376 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5377 		(void)VOP_MARKATIME(vp);
5378 }
5379 
5380 /*
5381  * The purpose of this routine is to remove granularity from accmode_t,
5382  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5383  * VADMIN and VAPPEND.
5384  *
5385  * If it returns 0, the caller is supposed to continue with the usual
5386  * access checks using 'accmode' as modified by this routine.  If it
5387  * returns nonzero value, the caller is supposed to return that value
5388  * as errno.
5389  *
5390  * Note that after this routine runs, accmode may be zero.
5391  */
5392 int
5393 vfs_unixify_accmode(accmode_t *accmode)
5394 {
5395 	/*
5396 	 * There is no way to specify explicit "deny" rule using
5397 	 * file mode or POSIX.1e ACLs.
5398 	 */
5399 	if (*accmode & VEXPLICIT_DENY) {
5400 		*accmode = 0;
5401 		return (0);
5402 	}
5403 
5404 	/*
5405 	 * None of these can be translated into usual access bits.
5406 	 * Also, the common case for NFSv4 ACLs is to not contain
5407 	 * either of these bits. Caller should check for VWRITE
5408 	 * on the containing directory instead.
5409 	 */
5410 	if (*accmode & (VDELETE_CHILD | VDELETE))
5411 		return (EPERM);
5412 
5413 	if (*accmode & VADMIN_PERMS) {
5414 		*accmode &= ~VADMIN_PERMS;
5415 		*accmode |= VADMIN;
5416 	}
5417 
5418 	/*
5419 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5420 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5421 	 */
5422 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5423 
5424 	return (0);
5425 }
5426 
5427 /*
5428  * These are helper functions for filesystems to traverse all
5429  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5430  *
5431  * This interface replaces MNT_VNODE_FOREACH.
5432  */
5433 
5434 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5435 
5436 struct vnode *
5437 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5438 {
5439 	struct vnode *vp;
5440 
5441 	if (should_yield())
5442 		kern_yield(PRI_USER);
5443 	MNT_ILOCK(mp);
5444 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5445 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5446 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5447 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5448 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5449 			continue;
5450 		VI_LOCK(vp);
5451 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5452 			VI_UNLOCK(vp);
5453 			continue;
5454 		}
5455 		break;
5456 	}
5457 	if (vp == NULL) {
5458 		__mnt_vnode_markerfree_all(mvp, mp);
5459 		/* MNT_IUNLOCK(mp); -- done in above function */
5460 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5461 		return (NULL);
5462 	}
5463 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5464 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5465 	MNT_IUNLOCK(mp);
5466 	return (vp);
5467 }
5468 
5469 struct vnode *
5470 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5471 {
5472 	struct vnode *vp;
5473 
5474 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5475 	MNT_ILOCK(mp);
5476 	MNT_REF(mp);
5477 	(*mvp)->v_mount = mp;
5478 	(*mvp)->v_type = VMARKER;
5479 
5480 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5481 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5482 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5483 			continue;
5484 		VI_LOCK(vp);
5485 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5486 			VI_UNLOCK(vp);
5487 			continue;
5488 		}
5489 		break;
5490 	}
5491 	if (vp == NULL) {
5492 		MNT_REL(mp);
5493 		MNT_IUNLOCK(mp);
5494 		free(*mvp, M_VNODE_MARKER);
5495 		*mvp = NULL;
5496 		return (NULL);
5497 	}
5498 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5499 	MNT_IUNLOCK(mp);
5500 	return (vp);
5501 }
5502 
5503 void
5504 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5505 {
5506 
5507 	if (*mvp == NULL) {
5508 		MNT_IUNLOCK(mp);
5509 		return;
5510 	}
5511 
5512 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5513 
5514 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5515 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5516 	MNT_REL(mp);
5517 	MNT_IUNLOCK(mp);
5518 	free(*mvp, M_VNODE_MARKER);
5519 	*mvp = NULL;
5520 }
5521 
5522 /*
5523  * These are helper functions for filesystems to traverse their
5524  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5525  */
5526 static void
5527 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5528 {
5529 
5530 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5531 
5532 	MNT_ILOCK(mp);
5533 	MNT_REL(mp);
5534 	MNT_IUNLOCK(mp);
5535 	free(*mvp, M_VNODE_MARKER);
5536 	*mvp = NULL;
5537 }
5538 
5539 /*
5540  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5541  * conventional lock order during mnt_vnode_next_active iteration.
5542  *
5543  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5544  * The list lock is dropped and reacquired.  On success, both locks are held.
5545  * On failure, the mount vnode list lock is held but the vnode interlock is
5546  * not, and the procedure may have yielded.
5547  */
5548 static bool
5549 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5550     struct vnode *vp)
5551 {
5552 	const struct vnode *tmp;
5553 	bool held, ret;
5554 
5555 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5556 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5557 	    ("%s: bad marker", __func__));
5558 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5559 	    ("%s: inappropriate vnode", __func__));
5560 	ASSERT_VI_UNLOCKED(vp, __func__);
5561 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5562 
5563 	ret = false;
5564 
5565 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5566 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5567 
5568 	/*
5569 	 * Use a hold to prevent vp from disappearing while the mount vnode
5570 	 * list lock is dropped and reacquired.  Normally a hold would be
5571 	 * acquired with vhold(), but that might try to acquire the vnode
5572 	 * interlock, which would be a LOR with the mount vnode list lock.
5573 	 */
5574 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5575 	mtx_unlock(&mp->mnt_listmtx);
5576 	if (!held)
5577 		goto abort;
5578 	VI_LOCK(vp);
5579 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5580 		vdropl(vp);
5581 		goto abort;
5582 	}
5583 	mtx_lock(&mp->mnt_listmtx);
5584 
5585 	/*
5586 	 * Determine whether the vnode is still the next one after the marker,
5587 	 * excepting any other markers.  If the vnode has not been doomed by
5588 	 * vgone() then the hold should have ensured that it remained on the
5589 	 * active list.  If it has been doomed but is still on the active list,
5590 	 * don't abort, but rather skip over it (avoid spinning on doomed
5591 	 * vnodes).
5592 	 */
5593 	tmp = mvp;
5594 	do {
5595 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5596 	} while (tmp != NULL && tmp->v_type == VMARKER);
5597 	if (tmp != vp) {
5598 		mtx_unlock(&mp->mnt_listmtx);
5599 		VI_UNLOCK(vp);
5600 		goto abort;
5601 	}
5602 
5603 	ret = true;
5604 	goto out;
5605 abort:
5606 	maybe_yield();
5607 	mtx_lock(&mp->mnt_listmtx);
5608 out:
5609 	if (ret)
5610 		ASSERT_VI_LOCKED(vp, __func__);
5611 	else
5612 		ASSERT_VI_UNLOCKED(vp, __func__);
5613 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5614 	return (ret);
5615 }
5616 
5617 static struct vnode *
5618 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5619 {
5620 	struct vnode *vp, *nvp;
5621 
5622 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5623 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5624 restart:
5625 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5626 	while (vp != NULL) {
5627 		if (vp->v_type == VMARKER) {
5628 			vp = TAILQ_NEXT(vp, v_actfreelist);
5629 			continue;
5630 		}
5631 		/*
5632 		 * Try-lock because this is the wrong lock order.  If that does
5633 		 * not succeed, drop the mount vnode list lock and try to
5634 		 * reacquire it and the vnode interlock in the right order.
5635 		 */
5636 		if (!VI_TRYLOCK(vp) &&
5637 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5638 			goto restart;
5639 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5640 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5641 		    ("alien vnode on the active list %p %p", vp, mp));
5642 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5643 			break;
5644 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5645 		VI_UNLOCK(vp);
5646 		vp = nvp;
5647 	}
5648 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5649 
5650 	/* Check if we are done */
5651 	if (vp == NULL) {
5652 		mtx_unlock(&mp->mnt_listmtx);
5653 		mnt_vnode_markerfree_active(mvp, mp);
5654 		return (NULL);
5655 	}
5656 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5657 	mtx_unlock(&mp->mnt_listmtx);
5658 	ASSERT_VI_LOCKED(vp, "active iter");
5659 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5660 	return (vp);
5661 }
5662 
5663 struct vnode *
5664 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5665 {
5666 
5667 	if (should_yield())
5668 		kern_yield(PRI_USER);
5669 	mtx_lock(&mp->mnt_listmtx);
5670 	return (mnt_vnode_next_active(mvp, mp));
5671 }
5672 
5673 struct vnode *
5674 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5675 {
5676 	struct vnode *vp;
5677 
5678 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5679 	MNT_ILOCK(mp);
5680 	MNT_REF(mp);
5681 	MNT_IUNLOCK(mp);
5682 	(*mvp)->v_type = VMARKER;
5683 	(*mvp)->v_mount = mp;
5684 
5685 	mtx_lock(&mp->mnt_listmtx);
5686 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5687 	if (vp == NULL) {
5688 		mtx_unlock(&mp->mnt_listmtx);
5689 		mnt_vnode_markerfree_active(mvp, mp);
5690 		return (NULL);
5691 	}
5692 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5693 	return (mnt_vnode_next_active(mvp, mp));
5694 }
5695 
5696 void
5697 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5698 {
5699 
5700 	if (*mvp == NULL)
5701 		return;
5702 
5703 	mtx_lock(&mp->mnt_listmtx);
5704 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5705 	mtx_unlock(&mp->mnt_listmtx);
5706 	mnt_vnode_markerfree_active(mvp, mp);
5707 }
5708