xref: /freebsd/sys/kern/vfs_subr.c (revision 840aca288042eaf625a23908e807abdfde0bc21d)
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, 0);
1038 			vdropl(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, 0);
1402 		VI_UNLOCK(vp);
1403 		vn_finished_write(vnmp);
1404 		CTR2(KTR_VFS,
1405 		    "%s: impossible to recycle, %p is already referenced",
1406 		    __func__, vp);
1407 		return (EBUSY);
1408 	}
1409 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1410 		counter_u64_add(recycles_count, 1);
1411 		vgonel(vp);
1412 	}
1413 	VOP_UNLOCK(vp, 0);
1414 	VI_UNLOCK(vp);
1415 	vn_finished_write(vnmp);
1416 	return (0);
1417 }
1418 
1419 static void
1420 vcheckspace(void)
1421 {
1422 
1423 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1424 		vnlruproc_sig = 1;
1425 		wakeup(vnlruproc);
1426 	}
1427 }
1428 
1429 /*
1430  * Wait if necessary for space for a new vnode.
1431  */
1432 static int
1433 getnewvnode_wait(int suspended)
1434 {
1435 
1436 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1437 	if (numvnodes >= desiredvnodes) {
1438 		if (suspended) {
1439 			/*
1440 			 * The file system is being suspended.  We cannot
1441 			 * risk a deadlock here, so allow allocation of
1442 			 * another vnode even if this would give too many.
1443 			 */
1444 			return (0);
1445 		}
1446 		if (vnlruproc_sig == 0) {
1447 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1448 			wakeup(vnlruproc);
1449 		}
1450 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1451 		    "vlruwk", hz);
1452 	}
1453 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1454 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1455 		vnlru_free_locked(1, NULL);
1456 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1457 }
1458 
1459 /*
1460  * This hack is fragile, and probably not needed any more now that the
1461  * watermark handling works.
1462  */
1463 void
1464 getnewvnode_reserve(u_int count)
1465 {
1466 	struct thread *td;
1467 
1468 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1469 	/* XXX no longer so quick, but this part is not racy. */
1470 	mtx_lock(&vnode_free_list_mtx);
1471 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1472 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1473 		    freevnodes - wantfreevnodes), NULL);
1474 	mtx_unlock(&vnode_free_list_mtx);
1475 
1476 	td = curthread;
1477 	/* First try to be quick and racy. */
1478 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1479 		td->td_vp_reserv += count;
1480 		vcheckspace();	/* XXX no longer so quick, but more racy */
1481 		return;
1482 	} else
1483 		atomic_subtract_long(&numvnodes, count);
1484 
1485 	mtx_lock(&vnode_free_list_mtx);
1486 	while (count > 0) {
1487 		if (getnewvnode_wait(0) == 0) {
1488 			count--;
1489 			td->td_vp_reserv++;
1490 			atomic_add_long(&numvnodes, 1);
1491 		}
1492 	}
1493 	vcheckspace();
1494 	mtx_unlock(&vnode_free_list_mtx);
1495 }
1496 
1497 /*
1498  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1499  * misconfgured or changed significantly.  Reducing desiredvnodes below
1500  * the reserved amount should cause bizarre behaviour like reducing it
1501  * below the number of active vnodes -- the system will try to reduce
1502  * numvnodes to match, but should fail, so the subtraction below should
1503  * not overflow.
1504  */
1505 void
1506 getnewvnode_drop_reserve(void)
1507 {
1508 	struct thread *td;
1509 
1510 	td = curthread;
1511 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1512 	td->td_vp_reserv = 0;
1513 }
1514 
1515 /*
1516  * Return the next vnode from the free list.
1517  */
1518 int
1519 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1520     struct vnode **vpp)
1521 {
1522 	struct vnode *vp;
1523 	struct thread *td;
1524 	struct lock_object *lo;
1525 	static int cyclecount;
1526 	int error __unused;
1527 
1528 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1529 	vp = NULL;
1530 	td = curthread;
1531 	if (td->td_vp_reserv > 0) {
1532 		td->td_vp_reserv -= 1;
1533 		goto alloc;
1534 	}
1535 	mtx_lock(&vnode_free_list_mtx);
1536 	if (numvnodes < desiredvnodes)
1537 		cyclecount = 0;
1538 	else if (cyclecount++ >= freevnodes) {
1539 		cyclecount = 0;
1540 		vstir = 1;
1541 	}
1542 	/*
1543 	 * Grow the vnode cache if it will not be above its target max
1544 	 * after growing.  Otherwise, if the free list is nonempty, try
1545 	 * to reclaim 1 item from it before growing the cache (possibly
1546 	 * above its target max if the reclamation failed or is delayed).
1547 	 * Otherwise, wait for some space.  In all cases, schedule
1548 	 * vnlru_proc() if we are getting short of space.  The watermarks
1549 	 * should be chosen so that we never wait or even reclaim from
1550 	 * the free list to below its target minimum.
1551 	 */
1552 	if (numvnodes + 1 <= desiredvnodes)
1553 		;
1554 	else if (freevnodes > 0)
1555 		vnlru_free_locked(1, NULL);
1556 	else {
1557 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1558 		    MNTK_SUSPEND));
1559 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1560 		if (error != 0) {
1561 			mtx_unlock(&vnode_free_list_mtx);
1562 			return (error);
1563 		}
1564 #endif
1565 	}
1566 	vcheckspace();
1567 	atomic_add_long(&numvnodes, 1);
1568 	mtx_unlock(&vnode_free_list_mtx);
1569 alloc:
1570 	counter_u64_add(vnodes_created, 1);
1571 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1572 	/*
1573 	 * Locks are given the generic name "vnode" when created.
1574 	 * Follow the historic practice of using the filesystem
1575 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1576 	 *
1577 	 * Locks live in a witness group keyed on their name. Thus,
1578 	 * when a lock is renamed, it must also move from the witness
1579 	 * group of its old name to the witness group of its new name.
1580 	 *
1581 	 * The change only needs to be made when the vnode moves
1582 	 * from one filesystem type to another. We ensure that each
1583 	 * filesystem use a single static name pointer for its tag so
1584 	 * that we can compare pointers rather than doing a strcmp().
1585 	 */
1586 	lo = &vp->v_vnlock->lock_object;
1587 	if (lo->lo_name != tag) {
1588 		lo->lo_name = tag;
1589 		WITNESS_DESTROY(lo);
1590 		WITNESS_INIT(lo, tag);
1591 	}
1592 	/*
1593 	 * By default, don't allow shared locks unless filesystems opt-in.
1594 	 */
1595 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1596 	/*
1597 	 * Finalize various vnode identity bits.
1598 	 */
1599 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1600 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1601 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1602 	vp->v_type = VNON;
1603 	vp->v_tag = tag;
1604 	vp->v_op = vops;
1605 	v_init_counters(vp);
1606 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1607 #ifdef DIAGNOSTIC
1608 	if (mp == NULL && vops != &dead_vnodeops)
1609 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1610 #endif
1611 #ifdef MAC
1612 	mac_vnode_init(vp);
1613 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1614 		mac_vnode_associate_singlelabel(mp, vp);
1615 #endif
1616 	if (mp != NULL) {
1617 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1618 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1619 			vp->v_vflag |= VV_NOKNOTE;
1620 	}
1621 
1622 	/*
1623 	 * For the filesystems which do not use vfs_hash_insert(),
1624 	 * still initialize v_hash to have vfs_hash_index() useful.
1625 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1626 	 * its own hashing.
1627 	 */
1628 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1629 
1630 	*vpp = vp;
1631 	return (0);
1632 }
1633 
1634 /*
1635  * Delete from old mount point vnode list, if on one.
1636  */
1637 static void
1638 delmntque(struct vnode *vp)
1639 {
1640 	struct mount *mp;
1641 	int active;
1642 
1643 	mp = vp->v_mount;
1644 	if (mp == NULL)
1645 		return;
1646 	MNT_ILOCK(mp);
1647 	VI_LOCK(vp);
1648 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1649 	    ("Active vnode list size %d > Vnode list size %d",
1650 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1651 	active = vp->v_iflag & VI_ACTIVE;
1652 	vp->v_iflag &= ~VI_ACTIVE;
1653 	if (active) {
1654 		mtx_lock(&mp->mnt_listmtx);
1655 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1656 		mp->mnt_activevnodelistsize--;
1657 		mtx_unlock(&mp->mnt_listmtx);
1658 	}
1659 	vp->v_mount = NULL;
1660 	VI_UNLOCK(vp);
1661 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1662 		("bad mount point vnode list size"));
1663 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1664 	mp->mnt_nvnodelistsize--;
1665 	MNT_REL(mp);
1666 	MNT_IUNLOCK(mp);
1667 }
1668 
1669 static void
1670 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1671 {
1672 
1673 	vp->v_data = NULL;
1674 	vp->v_op = &dead_vnodeops;
1675 	vgone(vp);
1676 	vput(vp);
1677 }
1678 
1679 /*
1680  * Insert into list of vnodes for the new mount point, if available.
1681  */
1682 int
1683 insmntque1(struct vnode *vp, struct mount *mp,
1684 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1685 {
1686 
1687 	KASSERT(vp->v_mount == NULL,
1688 		("insmntque: vnode already on per mount vnode list"));
1689 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1690 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1691 
1692 	/*
1693 	 * We acquire the vnode interlock early to ensure that the
1694 	 * vnode cannot be recycled by another process releasing a
1695 	 * holdcnt on it before we get it on both the vnode list
1696 	 * and the active vnode list. The mount mutex protects only
1697 	 * manipulation of the vnode list and the vnode freelist
1698 	 * mutex protects only manipulation of the active vnode list.
1699 	 * Hence the need to hold the vnode interlock throughout.
1700 	 */
1701 	MNT_ILOCK(mp);
1702 	VI_LOCK(vp);
1703 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1704 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1705 	    mp->mnt_nvnodelistsize == 0)) &&
1706 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1707 		VI_UNLOCK(vp);
1708 		MNT_IUNLOCK(mp);
1709 		if (dtr != NULL)
1710 			dtr(vp, dtr_arg);
1711 		return (EBUSY);
1712 	}
1713 	vp->v_mount = mp;
1714 	MNT_REF(mp);
1715 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1716 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1717 		("neg mount point vnode list size"));
1718 	mp->mnt_nvnodelistsize++;
1719 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1720 	    ("Activating already active vnode"));
1721 	vp->v_iflag |= VI_ACTIVE;
1722 	mtx_lock(&mp->mnt_listmtx);
1723 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1724 	mp->mnt_activevnodelistsize++;
1725 	mtx_unlock(&mp->mnt_listmtx);
1726 	VI_UNLOCK(vp);
1727 	MNT_IUNLOCK(mp);
1728 	return (0);
1729 }
1730 
1731 int
1732 insmntque(struct vnode *vp, struct mount *mp)
1733 {
1734 
1735 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1736 }
1737 
1738 /*
1739  * Flush out and invalidate all buffers associated with a bufobj
1740  * Called with the underlying object locked.
1741  */
1742 int
1743 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1744 {
1745 	int error;
1746 
1747 	BO_LOCK(bo);
1748 	if (flags & V_SAVE) {
1749 		error = bufobj_wwait(bo, slpflag, slptimeo);
1750 		if (error) {
1751 			BO_UNLOCK(bo);
1752 			return (error);
1753 		}
1754 		if (bo->bo_dirty.bv_cnt > 0) {
1755 			BO_UNLOCK(bo);
1756 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1757 				return (error);
1758 			/*
1759 			 * XXX We could save a lock/unlock if this was only
1760 			 * enabled under INVARIANTS
1761 			 */
1762 			BO_LOCK(bo);
1763 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1764 				panic("vinvalbuf: dirty bufs");
1765 		}
1766 	}
1767 	/*
1768 	 * If you alter this loop please notice that interlock is dropped and
1769 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1770 	 * no race conditions occur from this.
1771 	 */
1772 	do {
1773 		error = flushbuflist(&bo->bo_clean,
1774 		    flags, bo, slpflag, slptimeo);
1775 		if (error == 0 && !(flags & V_CLEANONLY))
1776 			error = flushbuflist(&bo->bo_dirty,
1777 			    flags, bo, slpflag, slptimeo);
1778 		if (error != 0 && error != EAGAIN) {
1779 			BO_UNLOCK(bo);
1780 			return (error);
1781 		}
1782 	} while (error != 0);
1783 
1784 	/*
1785 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1786 	 * have write I/O in-progress but if there is a VM object then the
1787 	 * VM object can also have read-I/O in-progress.
1788 	 */
1789 	do {
1790 		bufobj_wwait(bo, 0, 0);
1791 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1792 			BO_UNLOCK(bo);
1793 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1794 			BO_LOCK(bo);
1795 		}
1796 	} while (bo->bo_numoutput > 0);
1797 	BO_UNLOCK(bo);
1798 
1799 	/*
1800 	 * Destroy the copy in the VM cache, too.
1801 	 */
1802 	if (bo->bo_object != NULL &&
1803 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1804 		VM_OBJECT_WLOCK(bo->bo_object);
1805 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1806 		    OBJPR_CLEANONLY : 0);
1807 		VM_OBJECT_WUNLOCK(bo->bo_object);
1808 	}
1809 
1810 #ifdef INVARIANTS
1811 	BO_LOCK(bo);
1812 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1813 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1814 	    bo->bo_clean.bv_cnt > 0))
1815 		panic("vinvalbuf: flush failed");
1816 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1817 	    bo->bo_dirty.bv_cnt > 0)
1818 		panic("vinvalbuf: flush dirty failed");
1819 	BO_UNLOCK(bo);
1820 #endif
1821 	return (0);
1822 }
1823 
1824 /*
1825  * Flush out and invalidate all buffers associated with a vnode.
1826  * Called with the underlying object locked.
1827  */
1828 int
1829 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1830 {
1831 
1832 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1833 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1834 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1835 		return (0);
1836 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1837 }
1838 
1839 /*
1840  * Flush out buffers on the specified list.
1841  *
1842  */
1843 static int
1844 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1845     int slptimeo)
1846 {
1847 	struct buf *bp, *nbp;
1848 	int retval, error;
1849 	daddr_t lblkno;
1850 	b_xflags_t xflags;
1851 
1852 	ASSERT_BO_WLOCKED(bo);
1853 
1854 	retval = 0;
1855 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1856 		/*
1857 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1858 		 * do not skip any buffers. If we are flushing only V_NORMAL
1859 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1860 		 * flushing only V_ALT buffers then skip buffers not marked
1861 		 * as BX_ALTDATA.
1862 		 */
1863 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1864 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1865 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1866 			continue;
1867 		}
1868 		if (nbp != NULL) {
1869 			lblkno = nbp->b_lblkno;
1870 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1871 		}
1872 		retval = EAGAIN;
1873 		error = BUF_TIMELOCK(bp,
1874 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1875 		    "flushbuf", slpflag, slptimeo);
1876 		if (error) {
1877 			BO_LOCK(bo);
1878 			return (error != ENOLCK ? error : EAGAIN);
1879 		}
1880 		KASSERT(bp->b_bufobj == bo,
1881 		    ("bp %p wrong b_bufobj %p should be %p",
1882 		    bp, bp->b_bufobj, bo));
1883 		/*
1884 		 * XXX Since there are no node locks for NFS, I
1885 		 * believe there is a slight chance that a delayed
1886 		 * write will occur while sleeping just above, so
1887 		 * check for it.
1888 		 */
1889 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1890 		    (flags & V_SAVE)) {
1891 			bremfree(bp);
1892 			bp->b_flags |= B_ASYNC;
1893 			bwrite(bp);
1894 			BO_LOCK(bo);
1895 			return (EAGAIN);	/* XXX: why not loop ? */
1896 		}
1897 		bremfree(bp);
1898 		bp->b_flags |= (B_INVAL | B_RELBUF);
1899 		bp->b_flags &= ~B_ASYNC;
1900 		brelse(bp);
1901 		BO_LOCK(bo);
1902 		if (nbp == NULL)
1903 			break;
1904 		nbp = gbincore(bo, lblkno);
1905 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1906 		    != xflags)
1907 			break;			/* nbp invalid */
1908 	}
1909 	return (retval);
1910 }
1911 
1912 int
1913 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1914 {
1915 	struct buf *bp;
1916 	int error;
1917 	daddr_t lblkno;
1918 
1919 	ASSERT_BO_LOCKED(bo);
1920 
1921 	for (lblkno = startn;;) {
1922 again:
1923 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1924 		if (bp == NULL || bp->b_lblkno >= endn ||
1925 		    bp->b_lblkno < startn)
1926 			break;
1927 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1928 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1929 		if (error != 0) {
1930 			BO_RLOCK(bo);
1931 			if (error == ENOLCK)
1932 				goto again;
1933 			return (error);
1934 		}
1935 		KASSERT(bp->b_bufobj == bo,
1936 		    ("bp %p wrong b_bufobj %p should be %p",
1937 		    bp, bp->b_bufobj, bo));
1938 		lblkno = bp->b_lblkno + 1;
1939 		if ((bp->b_flags & B_MANAGED) == 0)
1940 			bremfree(bp);
1941 		bp->b_flags |= B_RELBUF;
1942 		/*
1943 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1944 		 * pages backing each buffer in the range are unlikely to be
1945 		 * reused.  Dirty buffers will have the hint applied once
1946 		 * they've been written.
1947 		 */
1948 		if ((bp->b_flags & B_VMIO) != 0)
1949 			bp->b_flags |= B_NOREUSE;
1950 		brelse(bp);
1951 		BO_RLOCK(bo);
1952 	}
1953 	return (0);
1954 }
1955 
1956 /*
1957  * Truncate a file's buffer and pages to a specified length.  This
1958  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1959  * sync activity.
1960  */
1961 int
1962 vtruncbuf(struct vnode *vp, off_t length, int blksize)
1963 {
1964 	struct buf *bp, *nbp;
1965 	struct bufobj *bo;
1966 	daddr_t startlbn;
1967 
1968 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
1969 	    vp, blksize, (uintmax_t)length);
1970 
1971 	/*
1972 	 * Round up to the *next* lbn.
1973 	 */
1974 	startlbn = howmany(length, blksize);
1975 
1976 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1977 
1978 	bo = &vp->v_bufobj;
1979 restart_unlocked:
1980 	BO_LOCK(bo);
1981 
1982 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
1983 		;
1984 
1985 	if (length > 0) {
1986 restartsync:
1987 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1988 			if (bp->b_lblkno > 0)
1989 				continue;
1990 			/*
1991 			 * Since we hold the vnode lock this should only
1992 			 * fail if we're racing with the buf daemon.
1993 			 */
1994 			if (BUF_LOCK(bp,
1995 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1996 			    BO_LOCKPTR(bo)) == ENOLCK)
1997 				goto restart_unlocked;
1998 
1999 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2000 			    ("buf(%p) on dirty queue without DELWRI", bp));
2001 
2002 			bremfree(bp);
2003 			bawrite(bp);
2004 			BO_LOCK(bo);
2005 			goto restartsync;
2006 		}
2007 	}
2008 
2009 	bufobj_wwait(bo, 0, 0);
2010 	BO_UNLOCK(bo);
2011 	vnode_pager_setsize(vp, length);
2012 
2013 	return (0);
2014 }
2015 
2016 /*
2017  * Invalidate the cached pages of a file's buffer within the range of block
2018  * numbers [startlbn, endlbn).
2019  */
2020 void
2021 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2022     int blksize)
2023 {
2024 	struct bufobj *bo;
2025 	off_t start, end;
2026 
2027 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2028 
2029 	start = blksize * startlbn;
2030 	end = blksize * endlbn;
2031 
2032 	bo = &vp->v_bufobj;
2033 	BO_LOCK(bo);
2034 	MPASS(blksize == bo->bo_bsize);
2035 
2036 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2037 		;
2038 
2039 	BO_UNLOCK(bo);
2040 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2041 }
2042 
2043 static int
2044 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2045     daddr_t startlbn, daddr_t endlbn)
2046 {
2047 	struct buf *bp, *nbp;
2048 	bool anyfreed;
2049 
2050 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2051 	ASSERT_BO_LOCKED(bo);
2052 
2053 	do {
2054 		anyfreed = false;
2055 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2056 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2057 				continue;
2058 			if (BUF_LOCK(bp,
2059 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2060 			    BO_LOCKPTR(bo)) == ENOLCK) {
2061 				BO_LOCK(bo);
2062 				return (EAGAIN);
2063 			}
2064 
2065 			bremfree(bp);
2066 			bp->b_flags |= B_INVAL | B_RELBUF;
2067 			bp->b_flags &= ~B_ASYNC;
2068 			brelse(bp);
2069 			anyfreed = true;
2070 
2071 			BO_LOCK(bo);
2072 			if (nbp != NULL &&
2073 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2074 			    nbp->b_vp != vp ||
2075 			    (nbp->b_flags & B_DELWRI) != 0))
2076 				return (EAGAIN);
2077 		}
2078 
2079 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2080 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2081 				continue;
2082 			if (BUF_LOCK(bp,
2083 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2084 			    BO_LOCKPTR(bo)) == ENOLCK) {
2085 				BO_LOCK(bo);
2086 				return (EAGAIN);
2087 			}
2088 			bremfree(bp);
2089 			bp->b_flags |= B_INVAL | B_RELBUF;
2090 			bp->b_flags &= ~B_ASYNC;
2091 			brelse(bp);
2092 			anyfreed = true;
2093 
2094 			BO_LOCK(bo);
2095 			if (nbp != NULL &&
2096 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2097 			    (nbp->b_vp != vp) ||
2098 			    (nbp->b_flags & B_DELWRI) == 0))
2099 				return (EAGAIN);
2100 		}
2101 	} while (anyfreed);
2102 	return (0);
2103 }
2104 
2105 static void
2106 buf_vlist_remove(struct buf *bp)
2107 {
2108 	struct bufv *bv;
2109 
2110 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2111 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2112 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2113 	    (BX_VNDIRTY|BX_VNCLEAN),
2114 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2115 	if (bp->b_xflags & BX_VNDIRTY)
2116 		bv = &bp->b_bufobj->bo_dirty;
2117 	else
2118 		bv = &bp->b_bufobj->bo_clean;
2119 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2120 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2121 	bv->bv_cnt--;
2122 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2123 }
2124 
2125 /*
2126  * Add the buffer to the sorted clean or dirty block list.
2127  *
2128  * NOTE: xflags is passed as a constant, optimizing this inline function!
2129  */
2130 static void
2131 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2132 {
2133 	struct bufv *bv;
2134 	struct buf *n;
2135 	int error;
2136 
2137 	ASSERT_BO_WLOCKED(bo);
2138 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2139 	    ("dead bo %p", bo));
2140 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2141 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2142 	bp->b_xflags |= xflags;
2143 	if (xflags & BX_VNDIRTY)
2144 		bv = &bo->bo_dirty;
2145 	else
2146 		bv = &bo->bo_clean;
2147 
2148 	/*
2149 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2150 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2151 	 * than _ge.
2152 	 */
2153 	if (bv->bv_cnt == 0 ||
2154 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2155 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2156 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2157 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2158 	else
2159 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2160 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2161 	if (error)
2162 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2163 	bv->bv_cnt++;
2164 }
2165 
2166 /*
2167  * Look up a buffer using the buffer tries.
2168  */
2169 struct buf *
2170 gbincore(struct bufobj *bo, daddr_t lblkno)
2171 {
2172 	struct buf *bp;
2173 
2174 	ASSERT_BO_LOCKED(bo);
2175 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2176 	if (bp != NULL)
2177 		return (bp);
2178 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2179 }
2180 
2181 /*
2182  * Associate a buffer with a vnode.
2183  */
2184 void
2185 bgetvp(struct vnode *vp, struct buf *bp)
2186 {
2187 	struct bufobj *bo;
2188 
2189 	bo = &vp->v_bufobj;
2190 	ASSERT_BO_WLOCKED(bo);
2191 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2192 
2193 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2194 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2195 	    ("bgetvp: bp already attached! %p", bp));
2196 
2197 	vhold(vp);
2198 	bp->b_vp = vp;
2199 	bp->b_bufobj = bo;
2200 	/*
2201 	 * Insert onto list for new vnode.
2202 	 */
2203 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2204 }
2205 
2206 /*
2207  * Disassociate a buffer from a vnode.
2208  */
2209 void
2210 brelvp(struct buf *bp)
2211 {
2212 	struct bufobj *bo;
2213 	struct vnode *vp;
2214 
2215 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2216 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2217 
2218 	/*
2219 	 * Delete from old vnode list, if on one.
2220 	 */
2221 	vp = bp->b_vp;		/* XXX */
2222 	bo = bp->b_bufobj;
2223 	BO_LOCK(bo);
2224 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2225 		buf_vlist_remove(bp);
2226 	else
2227 		panic("brelvp: Buffer %p not on queue.", bp);
2228 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2229 		bo->bo_flag &= ~BO_ONWORKLST;
2230 		mtx_lock(&sync_mtx);
2231 		LIST_REMOVE(bo, bo_synclist);
2232 		syncer_worklist_len--;
2233 		mtx_unlock(&sync_mtx);
2234 	}
2235 	bp->b_vp = NULL;
2236 	bp->b_bufobj = NULL;
2237 	BO_UNLOCK(bo);
2238 	vdrop(vp);
2239 }
2240 
2241 /*
2242  * Add an item to the syncer work queue.
2243  */
2244 static void
2245 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2246 {
2247 	int slot;
2248 
2249 	ASSERT_BO_WLOCKED(bo);
2250 
2251 	mtx_lock(&sync_mtx);
2252 	if (bo->bo_flag & BO_ONWORKLST)
2253 		LIST_REMOVE(bo, bo_synclist);
2254 	else {
2255 		bo->bo_flag |= BO_ONWORKLST;
2256 		syncer_worklist_len++;
2257 	}
2258 
2259 	if (delay > syncer_maxdelay - 2)
2260 		delay = syncer_maxdelay - 2;
2261 	slot = (syncer_delayno + delay) & syncer_mask;
2262 
2263 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2264 	mtx_unlock(&sync_mtx);
2265 }
2266 
2267 static int
2268 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2269 {
2270 	int error, len;
2271 
2272 	mtx_lock(&sync_mtx);
2273 	len = syncer_worklist_len - sync_vnode_count;
2274 	mtx_unlock(&sync_mtx);
2275 	error = SYSCTL_OUT(req, &len, sizeof(len));
2276 	return (error);
2277 }
2278 
2279 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2280     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2281 
2282 static struct proc *updateproc;
2283 static void sched_sync(void);
2284 static struct kproc_desc up_kp = {
2285 	"syncer",
2286 	sched_sync,
2287 	&updateproc
2288 };
2289 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2290 
2291 static int
2292 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2293 {
2294 	struct vnode *vp;
2295 	struct mount *mp;
2296 
2297 	*bo = LIST_FIRST(slp);
2298 	if (*bo == NULL)
2299 		return (0);
2300 	vp = bo2vnode(*bo);
2301 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2302 		return (1);
2303 	/*
2304 	 * We use vhold in case the vnode does not
2305 	 * successfully sync.  vhold prevents the vnode from
2306 	 * going away when we unlock the sync_mtx so that
2307 	 * we can acquire the vnode interlock.
2308 	 */
2309 	vholdl(vp);
2310 	mtx_unlock(&sync_mtx);
2311 	VI_UNLOCK(vp);
2312 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2313 		vdrop(vp);
2314 		mtx_lock(&sync_mtx);
2315 		return (*bo == LIST_FIRST(slp));
2316 	}
2317 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2318 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2319 	VOP_UNLOCK(vp, 0);
2320 	vn_finished_write(mp);
2321 	BO_LOCK(*bo);
2322 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2323 		/*
2324 		 * Put us back on the worklist.  The worklist
2325 		 * routine will remove us from our current
2326 		 * position and then add us back in at a later
2327 		 * position.
2328 		 */
2329 		vn_syncer_add_to_worklist(*bo, syncdelay);
2330 	}
2331 	BO_UNLOCK(*bo);
2332 	vdrop(vp);
2333 	mtx_lock(&sync_mtx);
2334 	return (0);
2335 }
2336 
2337 static int first_printf = 1;
2338 
2339 /*
2340  * System filesystem synchronizer daemon.
2341  */
2342 static void
2343 sched_sync(void)
2344 {
2345 	struct synclist *next, *slp;
2346 	struct bufobj *bo;
2347 	long starttime;
2348 	struct thread *td = curthread;
2349 	int last_work_seen;
2350 	int net_worklist_len;
2351 	int syncer_final_iter;
2352 	int error;
2353 
2354 	last_work_seen = 0;
2355 	syncer_final_iter = 0;
2356 	syncer_state = SYNCER_RUNNING;
2357 	starttime = time_uptime;
2358 	td->td_pflags |= TDP_NORUNNINGBUF;
2359 
2360 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2361 	    SHUTDOWN_PRI_LAST);
2362 
2363 	mtx_lock(&sync_mtx);
2364 	for (;;) {
2365 		if (syncer_state == SYNCER_FINAL_DELAY &&
2366 		    syncer_final_iter == 0) {
2367 			mtx_unlock(&sync_mtx);
2368 			kproc_suspend_check(td->td_proc);
2369 			mtx_lock(&sync_mtx);
2370 		}
2371 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2372 		if (syncer_state != SYNCER_RUNNING &&
2373 		    starttime != time_uptime) {
2374 			if (first_printf) {
2375 				printf("\nSyncing disks, vnodes remaining... ");
2376 				first_printf = 0;
2377 			}
2378 			printf("%d ", net_worklist_len);
2379 		}
2380 		starttime = time_uptime;
2381 
2382 		/*
2383 		 * Push files whose dirty time has expired.  Be careful
2384 		 * of interrupt race on slp queue.
2385 		 *
2386 		 * Skip over empty worklist slots when shutting down.
2387 		 */
2388 		do {
2389 			slp = &syncer_workitem_pending[syncer_delayno];
2390 			syncer_delayno += 1;
2391 			if (syncer_delayno == syncer_maxdelay)
2392 				syncer_delayno = 0;
2393 			next = &syncer_workitem_pending[syncer_delayno];
2394 			/*
2395 			 * If the worklist has wrapped since the
2396 			 * it was emptied of all but syncer vnodes,
2397 			 * switch to the FINAL_DELAY state and run
2398 			 * for one more second.
2399 			 */
2400 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2401 			    net_worklist_len == 0 &&
2402 			    last_work_seen == syncer_delayno) {
2403 				syncer_state = SYNCER_FINAL_DELAY;
2404 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2405 			}
2406 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2407 		    syncer_worklist_len > 0);
2408 
2409 		/*
2410 		 * Keep track of the last time there was anything
2411 		 * on the worklist other than syncer vnodes.
2412 		 * Return to the SHUTTING_DOWN state if any
2413 		 * new work appears.
2414 		 */
2415 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2416 			last_work_seen = syncer_delayno;
2417 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2418 			syncer_state = SYNCER_SHUTTING_DOWN;
2419 		while (!LIST_EMPTY(slp)) {
2420 			error = sync_vnode(slp, &bo, td);
2421 			if (error == 1) {
2422 				LIST_REMOVE(bo, bo_synclist);
2423 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2424 				continue;
2425 			}
2426 
2427 			if (first_printf == 0) {
2428 				/*
2429 				 * Drop the sync mutex, because some watchdog
2430 				 * drivers need to sleep while patting
2431 				 */
2432 				mtx_unlock(&sync_mtx);
2433 				wdog_kern_pat(WD_LASTVAL);
2434 				mtx_lock(&sync_mtx);
2435 			}
2436 
2437 		}
2438 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2439 			syncer_final_iter--;
2440 		/*
2441 		 * The variable rushjob allows the kernel to speed up the
2442 		 * processing of the filesystem syncer process. A rushjob
2443 		 * value of N tells the filesystem syncer to process the next
2444 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2445 		 * is used by the soft update code to speed up the filesystem
2446 		 * syncer process when the incore state is getting so far
2447 		 * ahead of the disk that the kernel memory pool is being
2448 		 * threatened with exhaustion.
2449 		 */
2450 		if (rushjob > 0) {
2451 			rushjob -= 1;
2452 			continue;
2453 		}
2454 		/*
2455 		 * Just sleep for a short period of time between
2456 		 * iterations when shutting down to allow some I/O
2457 		 * to happen.
2458 		 *
2459 		 * If it has taken us less than a second to process the
2460 		 * current work, then wait. Otherwise start right over
2461 		 * again. We can still lose time if any single round
2462 		 * takes more than two seconds, but it does not really
2463 		 * matter as we are just trying to generally pace the
2464 		 * filesystem activity.
2465 		 */
2466 		if (syncer_state != SYNCER_RUNNING ||
2467 		    time_uptime == starttime) {
2468 			thread_lock(td);
2469 			sched_prio(td, PPAUSE);
2470 			thread_unlock(td);
2471 		}
2472 		if (syncer_state != SYNCER_RUNNING)
2473 			cv_timedwait(&sync_wakeup, &sync_mtx,
2474 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2475 		else if (time_uptime == starttime)
2476 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2477 	}
2478 }
2479 
2480 /*
2481  * Request the syncer daemon to speed up its work.
2482  * We never push it to speed up more than half of its
2483  * normal turn time, otherwise it could take over the cpu.
2484  */
2485 int
2486 speedup_syncer(void)
2487 {
2488 	int ret = 0;
2489 
2490 	mtx_lock(&sync_mtx);
2491 	if (rushjob < syncdelay / 2) {
2492 		rushjob += 1;
2493 		stat_rush_requests += 1;
2494 		ret = 1;
2495 	}
2496 	mtx_unlock(&sync_mtx);
2497 	cv_broadcast(&sync_wakeup);
2498 	return (ret);
2499 }
2500 
2501 /*
2502  * Tell the syncer to speed up its work and run though its work
2503  * list several times, then tell it to shut down.
2504  */
2505 static void
2506 syncer_shutdown(void *arg, int howto)
2507 {
2508 
2509 	if (howto & RB_NOSYNC)
2510 		return;
2511 	mtx_lock(&sync_mtx);
2512 	syncer_state = SYNCER_SHUTTING_DOWN;
2513 	rushjob = 0;
2514 	mtx_unlock(&sync_mtx);
2515 	cv_broadcast(&sync_wakeup);
2516 	kproc_shutdown(arg, howto);
2517 }
2518 
2519 void
2520 syncer_suspend(void)
2521 {
2522 
2523 	syncer_shutdown(updateproc, 0);
2524 }
2525 
2526 void
2527 syncer_resume(void)
2528 {
2529 
2530 	mtx_lock(&sync_mtx);
2531 	first_printf = 1;
2532 	syncer_state = SYNCER_RUNNING;
2533 	mtx_unlock(&sync_mtx);
2534 	cv_broadcast(&sync_wakeup);
2535 	kproc_resume(updateproc);
2536 }
2537 
2538 /*
2539  * Reassign a buffer from one vnode to another.
2540  * Used to assign file specific control information
2541  * (indirect blocks) to the vnode to which they belong.
2542  */
2543 void
2544 reassignbuf(struct buf *bp)
2545 {
2546 	struct vnode *vp;
2547 	struct bufobj *bo;
2548 	int delay;
2549 #ifdef INVARIANTS
2550 	struct bufv *bv;
2551 #endif
2552 
2553 	vp = bp->b_vp;
2554 	bo = bp->b_bufobj;
2555 	++reassignbufcalls;
2556 
2557 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2558 	    bp, bp->b_vp, bp->b_flags);
2559 	/*
2560 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2561 	 * is not fully linked in.
2562 	 */
2563 	if (bp->b_flags & B_PAGING)
2564 		panic("cannot reassign paging buffer");
2565 
2566 	/*
2567 	 * Delete from old vnode list, if on one.
2568 	 */
2569 	BO_LOCK(bo);
2570 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2571 		buf_vlist_remove(bp);
2572 	else
2573 		panic("reassignbuf: Buffer %p not on queue.", bp);
2574 	/*
2575 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2576 	 * of clean buffers.
2577 	 */
2578 	if (bp->b_flags & B_DELWRI) {
2579 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2580 			switch (vp->v_type) {
2581 			case VDIR:
2582 				delay = dirdelay;
2583 				break;
2584 			case VCHR:
2585 				delay = metadelay;
2586 				break;
2587 			default:
2588 				delay = filedelay;
2589 			}
2590 			vn_syncer_add_to_worklist(bo, delay);
2591 		}
2592 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2593 	} else {
2594 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2595 
2596 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2597 			mtx_lock(&sync_mtx);
2598 			LIST_REMOVE(bo, bo_synclist);
2599 			syncer_worklist_len--;
2600 			mtx_unlock(&sync_mtx);
2601 			bo->bo_flag &= ~BO_ONWORKLST;
2602 		}
2603 	}
2604 #ifdef INVARIANTS
2605 	bv = &bo->bo_clean;
2606 	bp = TAILQ_FIRST(&bv->bv_hd);
2607 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2608 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2609 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2610 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2611 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2612 	bv = &bo->bo_dirty;
2613 	bp = TAILQ_FIRST(&bv->bv_hd);
2614 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2615 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2616 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2617 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2618 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2619 #endif
2620 	BO_UNLOCK(bo);
2621 }
2622 
2623 static void
2624 v_init_counters(struct vnode *vp)
2625 {
2626 
2627 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2628 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2629 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2630 
2631 	refcount_init(&vp->v_holdcnt, 1);
2632 	refcount_init(&vp->v_usecount, 1);
2633 }
2634 
2635 static void
2636 v_incr_usecount_locked(struct vnode *vp)
2637 {
2638 
2639 	ASSERT_VI_LOCKED(vp, __func__);
2640 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2641 		VNASSERT(vp->v_usecount == 0, vp,
2642 		    ("vnode with usecount and VI_OWEINACT set"));
2643 		vp->v_iflag &= ~VI_OWEINACT;
2644 		VNODE_REFCOUNT_FENCE_REL();
2645 	}
2646 	refcount_acquire(&vp->v_usecount);
2647 	v_incr_devcount(vp);
2648 }
2649 
2650 /*
2651  * Increment the use count on the vnode, taking care to reference
2652  * the driver's usecount if this is a chardev.
2653  */
2654 static void
2655 v_incr_usecount(struct vnode *vp)
2656 {
2657 
2658 	ASSERT_VI_UNLOCKED(vp, __func__);
2659 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2660 
2661 	if (vp->v_type != VCHR &&
2662 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2663 		VNODE_REFCOUNT_FENCE_ACQ();
2664 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2665 		    ("vnode with usecount and VI_OWEINACT set"));
2666 	} else {
2667 		VI_LOCK(vp);
2668 		v_incr_usecount_locked(vp);
2669 		VI_UNLOCK(vp);
2670 	}
2671 }
2672 
2673 /*
2674  * Increment si_usecount of the associated device, if any.
2675  */
2676 static void
2677 v_incr_devcount(struct vnode *vp)
2678 {
2679 
2680 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2681 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2682 		dev_lock();
2683 		vp->v_rdev->si_usecount++;
2684 		dev_unlock();
2685 	}
2686 }
2687 
2688 /*
2689  * Decrement si_usecount of the associated device, if any.
2690  */
2691 static void
2692 v_decr_devcount(struct vnode *vp)
2693 {
2694 
2695 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2696 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2697 		dev_lock();
2698 		vp->v_rdev->si_usecount--;
2699 		dev_unlock();
2700 	}
2701 }
2702 
2703 /*
2704  * Grab a particular vnode from the free list, increment its
2705  * reference count and lock it.  VI_DOOMED is set if the vnode
2706  * is being destroyed.  Only callers who specify LK_RETRY will
2707  * see doomed vnodes.  If inactive processing was delayed in
2708  * vput try to do it here.
2709  *
2710  * Notes on lockless counter manipulation:
2711  * _vhold, vputx and other routines make various decisions based
2712  * on either holdcnt or usecount being 0. As long as either counter
2713  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2714  * with atomic operations. Otherwise the interlock is taken covering
2715  * both the atomic and additional actions.
2716  */
2717 int
2718 vget(struct vnode *vp, int flags, struct thread *td)
2719 {
2720 	int error, oweinact;
2721 
2722 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2723 	    ("vget: invalid lock operation"));
2724 
2725 	if ((flags & LK_INTERLOCK) != 0)
2726 		ASSERT_VI_LOCKED(vp, __func__);
2727 	else
2728 		ASSERT_VI_UNLOCKED(vp, __func__);
2729 	if ((flags & LK_VNHELD) != 0)
2730 		VNASSERT((vp->v_holdcnt > 0), vp,
2731 		    ("vget: LK_VNHELD passed but vnode not held"));
2732 
2733 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2734 
2735 	if ((flags & LK_VNHELD) == 0)
2736 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2737 
2738 	if ((error = vn_lock(vp, flags)) != 0) {
2739 		vdrop(vp);
2740 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2741 		    vp);
2742 		return (error);
2743 	}
2744 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2745 		panic("vget: vn_lock failed to return ENOENT\n");
2746 	/*
2747 	 * We don't guarantee that any particular close will
2748 	 * trigger inactive processing so just make a best effort
2749 	 * here at preventing a reference to a removed file.  If
2750 	 * we don't succeed no harm is done.
2751 	 *
2752 	 * Upgrade our holdcnt to a usecount.
2753 	 */
2754 	if (vp->v_type == VCHR ||
2755 	    !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2756 		VI_LOCK(vp);
2757 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2758 			oweinact = 0;
2759 		} else {
2760 			oweinact = 1;
2761 			vp->v_iflag &= ~VI_OWEINACT;
2762 			VNODE_REFCOUNT_FENCE_REL();
2763 		}
2764 		refcount_acquire(&vp->v_usecount);
2765 		v_incr_devcount(vp);
2766 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2767 		    (flags & LK_NOWAIT) == 0)
2768 			vinactive(vp, td);
2769 		VI_UNLOCK(vp);
2770 	}
2771 	return (0);
2772 }
2773 
2774 /*
2775  * Increase the reference (use) and hold count of a vnode.
2776  * This will also remove the vnode from the free list if it is presently free.
2777  */
2778 void
2779 vref(struct vnode *vp)
2780 {
2781 
2782 	ASSERT_VI_UNLOCKED(vp, __func__);
2783 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2784 	_vhold(vp, false);
2785 	v_incr_usecount(vp);
2786 }
2787 
2788 void
2789 vrefl(struct vnode *vp)
2790 {
2791 
2792 	ASSERT_VI_LOCKED(vp, __func__);
2793 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2794 	_vhold(vp, true);
2795 	v_incr_usecount_locked(vp);
2796 }
2797 
2798 void
2799 vrefact(struct vnode *vp)
2800 {
2801 
2802 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2803 	if (__predict_false(vp->v_type == VCHR)) {
2804 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2805 		    ("%s: wrong ref counts", __func__));
2806 		vref(vp);
2807 		return;
2808 	}
2809 #ifdef INVARIANTS
2810 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2811 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2812 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2813 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2814 #else
2815 	refcount_acquire(&vp->v_holdcnt);
2816 	refcount_acquire(&vp->v_usecount);
2817 #endif
2818 }
2819 
2820 /*
2821  * Return reference count of a vnode.
2822  *
2823  * The results of this call are only guaranteed when some mechanism is used to
2824  * stop other processes from gaining references to the vnode.  This may be the
2825  * case if the caller holds the only reference.  This is also useful when stale
2826  * data is acceptable as race conditions may be accounted for by some other
2827  * means.
2828  */
2829 int
2830 vrefcnt(struct vnode *vp)
2831 {
2832 
2833 	return (vp->v_usecount);
2834 }
2835 
2836 #define	VPUTX_VRELE	1
2837 #define	VPUTX_VPUT	2
2838 #define	VPUTX_VUNREF	3
2839 
2840 /*
2841  * Decrement the use and hold counts for a vnode.
2842  *
2843  * See an explanation near vget() as to why atomic operation is safe.
2844  */
2845 static void
2846 vputx(struct vnode *vp, int func)
2847 {
2848 	int error;
2849 
2850 	KASSERT(vp != NULL, ("vputx: null vp"));
2851 	if (func == VPUTX_VUNREF)
2852 		ASSERT_VOP_LOCKED(vp, "vunref");
2853 	else if (func == VPUTX_VPUT)
2854 		ASSERT_VOP_LOCKED(vp, "vput");
2855 	else
2856 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2857 	ASSERT_VI_UNLOCKED(vp, __func__);
2858 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2859 	    ("%s: wrong ref counts", __func__));
2860 
2861 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2862 
2863 	if (vp->v_type != VCHR &&
2864 	    refcount_release_if_not_last(&vp->v_usecount)) {
2865 		if (func == VPUTX_VPUT)
2866 			VOP_UNLOCK(vp, 0);
2867 		vdrop(vp);
2868 		return;
2869 	}
2870 
2871 	VI_LOCK(vp);
2872 
2873 	/*
2874 	 * We want to hold the vnode until the inactive finishes to
2875 	 * prevent vgone() races.  We drop the use count here and the
2876 	 * hold count below when we're done.
2877 	 */
2878 	if (!refcount_release(&vp->v_usecount) ||
2879 	    (vp->v_iflag & VI_DOINGINACT)) {
2880 		if (func == VPUTX_VPUT)
2881 			VOP_UNLOCK(vp, 0);
2882 		v_decr_devcount(vp);
2883 		vdropl(vp);
2884 		return;
2885 	}
2886 
2887 	v_decr_devcount(vp);
2888 
2889 	error = 0;
2890 
2891 	if (vp->v_usecount != 0) {
2892 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2893 		panic("vputx: usecount not zero");
2894 	}
2895 
2896 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2897 
2898 	/*
2899 	 * Check if the fs wants to perform inactive processing. Note we
2900 	 * may be only holding the interlock, in which case it is possible
2901 	 * someone else called vgone on the vnode and ->v_data is now NULL.
2902 	 * Since vgone performs inactive on its own there is nothing to do
2903 	 * here but to drop our hold count.
2904 	 */
2905 	if (__predict_false(vp->v_iflag & VI_DOOMED) ||
2906 	    VOP_NEED_INACTIVE(vp) == 0) {
2907 		if (func == VPUTX_VPUT)
2908 			VOP_UNLOCK(vp, 0);
2909 		vdropl(vp);
2910 		return;
2911 	}
2912 
2913 	/*
2914 	 * We must call VOP_INACTIVE with the node locked. Mark
2915 	 * as VI_DOINGINACT to avoid recursion.
2916 	 */
2917 	vp->v_iflag |= VI_OWEINACT;
2918 	switch (func) {
2919 	case VPUTX_VRELE:
2920 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2921 		VI_LOCK(vp);
2922 		break;
2923 	case VPUTX_VPUT:
2924 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2925 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2926 			    LK_NOWAIT);
2927 			VI_LOCK(vp);
2928 		}
2929 		break;
2930 	case VPUTX_VUNREF:
2931 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2932 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2933 			VI_LOCK(vp);
2934 		}
2935 		break;
2936 	}
2937 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2938 	    ("vnode with usecount and VI_OWEINACT set"));
2939 	if (error == 0) {
2940 		if (vp->v_iflag & VI_OWEINACT)
2941 			vinactive(vp, curthread);
2942 		if (func != VPUTX_VUNREF)
2943 			VOP_UNLOCK(vp, 0);
2944 	}
2945 	vdropl(vp);
2946 }
2947 
2948 /*
2949  * Vnode put/release.
2950  * If count drops to zero, call inactive routine and return to freelist.
2951  */
2952 void
2953 vrele(struct vnode *vp)
2954 {
2955 
2956 	vputx(vp, VPUTX_VRELE);
2957 }
2958 
2959 /*
2960  * Release an already locked vnode.  This give the same effects as
2961  * unlock+vrele(), but takes less time and avoids releasing and
2962  * re-aquiring the lock (as vrele() acquires the lock internally.)
2963  */
2964 void
2965 vput(struct vnode *vp)
2966 {
2967 
2968 	vputx(vp, VPUTX_VPUT);
2969 }
2970 
2971 /*
2972  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2973  */
2974 void
2975 vunref(struct vnode *vp)
2976 {
2977 
2978 	vputx(vp, VPUTX_VUNREF);
2979 }
2980 
2981 /*
2982  * Increase the hold count and activate if this is the first reference.
2983  */
2984 void
2985 _vhold(struct vnode *vp, bool locked)
2986 {
2987 	struct mount *mp;
2988 
2989 	if (locked)
2990 		ASSERT_VI_LOCKED(vp, __func__);
2991 	else
2992 		ASSERT_VI_UNLOCKED(vp, __func__);
2993 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2994 	if (!locked) {
2995 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2996 			VNODE_REFCOUNT_FENCE_ACQ();
2997 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2998 			    ("_vhold: vnode with holdcnt is free"));
2999 			return;
3000 		}
3001 		VI_LOCK(vp);
3002 	}
3003 	if ((vp->v_iflag & VI_FREE) == 0) {
3004 		refcount_acquire(&vp->v_holdcnt);
3005 		if (!locked)
3006 			VI_UNLOCK(vp);
3007 		return;
3008 	}
3009 	VNASSERT(vp->v_holdcnt == 0, vp,
3010 	    ("%s: wrong hold count", __func__));
3011 	VNASSERT(vp->v_op != NULL, vp,
3012 	    ("%s: vnode already reclaimed.", __func__));
3013 	/*
3014 	 * Remove a vnode from the free list, mark it as in use,
3015 	 * and put it on the active list.
3016 	 */
3017 	VNASSERT(vp->v_mount != NULL, vp,
3018 	    ("_vhold: vnode not on per mount vnode list"));
3019 	mp = vp->v_mount;
3020 	mtx_lock(&mp->mnt_listmtx);
3021 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
3022 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3023 		mp->mnt_tmpfreevnodelistsize--;
3024 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
3025 	} else {
3026 		mtx_lock(&vnode_free_list_mtx);
3027 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3028 		freevnodes--;
3029 		mtx_unlock(&vnode_free_list_mtx);
3030 	}
3031 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3032 	    ("Activating already active vnode"));
3033 	vp->v_iflag &= ~VI_FREE;
3034 	vp->v_iflag |= VI_ACTIVE;
3035 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3036 	mp->mnt_activevnodelistsize++;
3037 	mtx_unlock(&mp->mnt_listmtx);
3038 	refcount_acquire(&vp->v_holdcnt);
3039 	if (!locked)
3040 		VI_UNLOCK(vp);
3041 }
3042 
3043 void
3044 vholdnz(struct vnode *vp)
3045 {
3046 
3047 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3048 #ifdef INVARIANTS
3049 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3050 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
3051 #else
3052 	atomic_add_int(&vp->v_holdcnt, 1);
3053 #endif
3054 }
3055 
3056 /*
3057  * Drop the hold count of the vnode.  If this is the last reference to
3058  * the vnode we place it on the free list unless it has been vgone'd
3059  * (marked VI_DOOMED) in which case we will free it.
3060  *
3061  * Because the vnode vm object keeps a hold reference on the vnode if
3062  * there is at least one resident non-cached page, the vnode cannot
3063  * leave the active list without the page cleanup done.
3064  */
3065 void
3066 _vdrop(struct vnode *vp, bool locked)
3067 {
3068 	struct bufobj *bo;
3069 	struct mount *mp;
3070 	int active;
3071 
3072 	if (locked)
3073 		ASSERT_VI_LOCKED(vp, __func__);
3074 	else
3075 		ASSERT_VI_UNLOCKED(vp, __func__);
3076 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3077 	if (__predict_false((int)vp->v_holdcnt <= 0)) {
3078 		vn_printf(vp, "vdrop: holdcnt %d", vp->v_holdcnt);
3079 		panic("vdrop: wrong holdcnt");
3080 	}
3081 	if (!locked) {
3082 		if (refcount_release_if_not_last(&vp->v_holdcnt))
3083 			return;
3084 		VI_LOCK(vp);
3085 	}
3086 	if (refcount_release(&vp->v_holdcnt) == 0) {
3087 		VI_UNLOCK(vp);
3088 		return;
3089 	}
3090 	if ((vp->v_iflag & VI_DOOMED) == 0) {
3091 		/*
3092 		 * Mark a vnode as free: remove it from its active list
3093 		 * and put it up for recycling on the freelist.
3094 		 */
3095 		VNASSERT(vp->v_op != NULL, vp,
3096 		    ("vdropl: vnode already reclaimed."));
3097 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3098 		    ("vnode already free"));
3099 		VNASSERT(vp->v_holdcnt == 0, vp,
3100 		    ("vdropl: freeing when we shouldn't"));
3101 		active = vp->v_iflag & VI_ACTIVE;
3102 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
3103 			vp->v_iflag &= ~VI_ACTIVE;
3104 			mp = vp->v_mount;
3105 			if (mp != NULL) {
3106 				mtx_lock(&mp->mnt_listmtx);
3107 				if (active) {
3108 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
3109 					    vp, v_actfreelist);
3110 					mp->mnt_activevnodelistsize--;
3111 				}
3112 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3113 				    vp, v_actfreelist);
3114 				mp->mnt_tmpfreevnodelistsize++;
3115 				vp->v_iflag |= VI_FREE;
3116 				vp->v_mflag |= VMP_TMPMNTFREELIST;
3117 				VI_UNLOCK(vp);
3118 				if (mp->mnt_tmpfreevnodelistsize >=
3119 				    mnt_free_list_batch)
3120 					vnlru_return_batch_locked(mp);
3121 				mtx_unlock(&mp->mnt_listmtx);
3122 			} else {
3123 				VNASSERT(active == 0, vp,
3124 				    ("vdropl: active vnode not on per mount "
3125 				    "vnode list"));
3126 				mtx_lock(&vnode_free_list_mtx);
3127 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3128 				    v_actfreelist);
3129 				freevnodes++;
3130 				vp->v_iflag |= VI_FREE;
3131 				VI_UNLOCK(vp);
3132 				mtx_unlock(&vnode_free_list_mtx);
3133 			}
3134 		} else {
3135 			VI_UNLOCK(vp);
3136 			counter_u64_add(free_owe_inact, 1);
3137 		}
3138 		return;
3139 	}
3140 	/*
3141 	 * The vnode has been marked for destruction, so free it.
3142 	 *
3143 	 * The vnode will be returned to the zone where it will
3144 	 * normally remain until it is needed for another vnode. We
3145 	 * need to cleanup (or verify that the cleanup has already
3146 	 * been done) any residual data left from its current use
3147 	 * so as not to contaminate the freshly allocated vnode.
3148 	 */
3149 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
3150 	atomic_subtract_long(&numvnodes, 1);
3151 	bo = &vp->v_bufobj;
3152 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3153 	    ("cleaned vnode still on the free list."));
3154 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
3155 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
3156 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
3157 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
3158 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
3159 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
3160 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
3161 	    ("clean blk trie not empty"));
3162 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3163 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3164 	    ("dirty blk trie not empty"));
3165 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3166 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3167 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3168 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3169 	    ("Dangling rangelock waiters"));
3170 	VI_UNLOCK(vp);
3171 #ifdef MAC
3172 	mac_vnode_destroy(vp);
3173 #endif
3174 	if (vp->v_pollinfo != NULL) {
3175 		destroy_vpollinfo(vp->v_pollinfo);
3176 		vp->v_pollinfo = NULL;
3177 	}
3178 #ifdef INVARIANTS
3179 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3180 	vp->v_op = NULL;
3181 #endif
3182 	vp->v_mountedhere = NULL;
3183 	vp->v_unpcb = NULL;
3184 	vp->v_rdev = NULL;
3185 	vp->v_fifoinfo = NULL;
3186 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3187 	vp->v_iflag = 0;
3188 	vp->v_vflag = 0;
3189 	bo->bo_flag = 0;
3190 	uma_zfree(vnode_zone, vp);
3191 }
3192 
3193 /*
3194  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3195  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3196  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3197  * failed lock upgrade.
3198  */
3199 void
3200 vinactive(struct vnode *vp, struct thread *td)
3201 {
3202 	struct vm_object *obj;
3203 
3204 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3205 	ASSERT_VI_LOCKED(vp, "vinactive");
3206 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3207 	    ("vinactive: recursed on VI_DOINGINACT"));
3208 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3209 	vp->v_iflag |= VI_DOINGINACT;
3210 	vp->v_iflag &= ~VI_OWEINACT;
3211 	VI_UNLOCK(vp);
3212 	/*
3213 	 * Before moving off the active list, we must be sure that any
3214 	 * modified pages are converted into the vnode's dirty
3215 	 * buffers, since these will no longer be checked once the
3216 	 * vnode is on the inactive list.
3217 	 *
3218 	 * The write-out of the dirty pages is asynchronous.  At the
3219 	 * point that VOP_INACTIVE() is called, there could still be
3220 	 * pending I/O and dirty pages in the object.
3221 	 */
3222 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3223 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3224 		VM_OBJECT_WLOCK(obj);
3225 		vm_object_page_clean(obj, 0, 0, 0);
3226 		VM_OBJECT_WUNLOCK(obj);
3227 	}
3228 	VOP_INACTIVE(vp, td);
3229 	VI_LOCK(vp);
3230 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3231 	    ("vinactive: lost VI_DOINGINACT"));
3232 	vp->v_iflag &= ~VI_DOINGINACT;
3233 }
3234 
3235 /*
3236  * Remove any vnodes in the vnode table belonging to mount point mp.
3237  *
3238  * If FORCECLOSE is not specified, there should not be any active ones,
3239  * return error if any are found (nb: this is a user error, not a
3240  * system error). If FORCECLOSE is specified, detach any active vnodes
3241  * that are found.
3242  *
3243  * If WRITECLOSE is set, only flush out regular file vnodes open for
3244  * writing.
3245  *
3246  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3247  *
3248  * `rootrefs' specifies the base reference count for the root vnode
3249  * of this filesystem. The root vnode is considered busy if its
3250  * v_usecount exceeds this value. On a successful return, vflush(, td)
3251  * will call vrele() on the root vnode exactly rootrefs times.
3252  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3253  * be zero.
3254  */
3255 #ifdef DIAGNOSTIC
3256 static int busyprt = 0;		/* print out busy vnodes */
3257 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3258 #endif
3259 
3260 int
3261 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3262 {
3263 	struct vnode *vp, *mvp, *rootvp = NULL;
3264 	struct vattr vattr;
3265 	int busy = 0, error;
3266 
3267 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3268 	    rootrefs, flags);
3269 	if (rootrefs > 0) {
3270 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3271 		    ("vflush: bad args"));
3272 		/*
3273 		 * Get the filesystem root vnode. We can vput() it
3274 		 * immediately, since with rootrefs > 0, it won't go away.
3275 		 */
3276 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3277 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3278 			    __func__, error);
3279 			return (error);
3280 		}
3281 		vput(rootvp);
3282 	}
3283 loop:
3284 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3285 		vholdl(vp);
3286 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3287 		if (error) {
3288 			vdrop(vp);
3289 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3290 			goto loop;
3291 		}
3292 		/*
3293 		 * Skip over a vnodes marked VV_SYSTEM.
3294 		 */
3295 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3296 			VOP_UNLOCK(vp, 0);
3297 			vdrop(vp);
3298 			continue;
3299 		}
3300 		/*
3301 		 * If WRITECLOSE is set, flush out unlinked but still open
3302 		 * files (even if open only for reading) and regular file
3303 		 * vnodes open for writing.
3304 		 */
3305 		if (flags & WRITECLOSE) {
3306 			if (vp->v_object != NULL) {
3307 				VM_OBJECT_WLOCK(vp->v_object);
3308 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3309 				VM_OBJECT_WUNLOCK(vp->v_object);
3310 			}
3311 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3312 			if (error != 0) {
3313 				VOP_UNLOCK(vp, 0);
3314 				vdrop(vp);
3315 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3316 				return (error);
3317 			}
3318 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3319 			VI_LOCK(vp);
3320 
3321 			if ((vp->v_type == VNON ||
3322 			    (error == 0 && vattr.va_nlink > 0)) &&
3323 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3324 				VOP_UNLOCK(vp, 0);
3325 				vdropl(vp);
3326 				continue;
3327 			}
3328 		} else
3329 			VI_LOCK(vp);
3330 		/*
3331 		 * With v_usecount == 0, all we need to do is clear out the
3332 		 * vnode data structures and we are done.
3333 		 *
3334 		 * If FORCECLOSE is set, forcibly close the vnode.
3335 		 */
3336 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3337 			vgonel(vp);
3338 		} else {
3339 			busy++;
3340 #ifdef DIAGNOSTIC
3341 			if (busyprt)
3342 				vn_printf(vp, "vflush: busy vnode ");
3343 #endif
3344 		}
3345 		VOP_UNLOCK(vp, 0);
3346 		vdropl(vp);
3347 	}
3348 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3349 		/*
3350 		 * If just the root vnode is busy, and if its refcount
3351 		 * is equal to `rootrefs', then go ahead and kill it.
3352 		 */
3353 		VI_LOCK(rootvp);
3354 		KASSERT(busy > 0, ("vflush: not busy"));
3355 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3356 		    ("vflush: usecount %d < rootrefs %d",
3357 		     rootvp->v_usecount, rootrefs));
3358 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3359 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3360 			vgone(rootvp);
3361 			VOP_UNLOCK(rootvp, 0);
3362 			busy = 0;
3363 		} else
3364 			VI_UNLOCK(rootvp);
3365 	}
3366 	if (busy) {
3367 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3368 		    busy);
3369 		return (EBUSY);
3370 	}
3371 	for (; rootrefs > 0; rootrefs--)
3372 		vrele(rootvp);
3373 	return (0);
3374 }
3375 
3376 /*
3377  * Recycle an unused vnode to the front of the free list.
3378  */
3379 int
3380 vrecycle(struct vnode *vp)
3381 {
3382 	int recycled;
3383 
3384 	VI_LOCK(vp);
3385 	recycled = vrecyclel(vp);
3386 	VI_UNLOCK(vp);
3387 	return (recycled);
3388 }
3389 
3390 /*
3391  * vrecycle, with the vp interlock held.
3392  */
3393 int
3394 vrecyclel(struct vnode *vp)
3395 {
3396 	int recycled;
3397 
3398 	ASSERT_VOP_ELOCKED(vp, __func__);
3399 	ASSERT_VI_LOCKED(vp, __func__);
3400 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3401 	recycled = 0;
3402 	if (vp->v_usecount == 0) {
3403 		recycled = 1;
3404 		vgonel(vp);
3405 	}
3406 	return (recycled);
3407 }
3408 
3409 /*
3410  * Eliminate all activity associated with a vnode
3411  * in preparation for reuse.
3412  */
3413 void
3414 vgone(struct vnode *vp)
3415 {
3416 	VI_LOCK(vp);
3417 	vgonel(vp);
3418 	VI_UNLOCK(vp);
3419 }
3420 
3421 static void
3422 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3423     struct vnode *lowervp __unused)
3424 {
3425 }
3426 
3427 /*
3428  * Notify upper mounts about reclaimed or unlinked vnode.
3429  */
3430 void
3431 vfs_notify_upper(struct vnode *vp, int event)
3432 {
3433 	static struct vfsops vgonel_vfsops = {
3434 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3435 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3436 	};
3437 	struct mount *mp, *ump, *mmp;
3438 
3439 	mp = vp->v_mount;
3440 	if (mp == NULL)
3441 		return;
3442 
3443 	MNT_ILOCK(mp);
3444 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3445 		goto unlock;
3446 	MNT_IUNLOCK(mp);
3447 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3448 	mmp->mnt_op = &vgonel_vfsops;
3449 	mmp->mnt_kern_flag |= MNTK_MARKER;
3450 	MNT_ILOCK(mp);
3451 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3452 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3453 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3454 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3455 			continue;
3456 		}
3457 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3458 		MNT_IUNLOCK(mp);
3459 		switch (event) {
3460 		case VFS_NOTIFY_UPPER_RECLAIM:
3461 			VFS_RECLAIM_LOWERVP(ump, vp);
3462 			break;
3463 		case VFS_NOTIFY_UPPER_UNLINK:
3464 			VFS_UNLINK_LOWERVP(ump, vp);
3465 			break;
3466 		default:
3467 			KASSERT(0, ("invalid event %d", event));
3468 			break;
3469 		}
3470 		MNT_ILOCK(mp);
3471 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3472 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3473 	}
3474 	free(mmp, M_TEMP);
3475 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3476 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3477 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3478 		wakeup(&mp->mnt_uppers);
3479 	}
3480 unlock:
3481 	MNT_IUNLOCK(mp);
3482 }
3483 
3484 /*
3485  * vgone, with the vp interlock held.
3486  */
3487 static void
3488 vgonel(struct vnode *vp)
3489 {
3490 	struct thread *td;
3491 	struct mount *mp;
3492 	vm_object_t object;
3493 	bool active, oweinact;
3494 
3495 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3496 	ASSERT_VI_LOCKED(vp, "vgonel");
3497 	VNASSERT(vp->v_holdcnt, vp,
3498 	    ("vgonel: vp %p has no reference.", vp));
3499 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3500 	td = curthread;
3501 
3502 	/*
3503 	 * Don't vgonel if we're already doomed.
3504 	 */
3505 	if (vp->v_iflag & VI_DOOMED)
3506 		return;
3507 	vp->v_iflag |= VI_DOOMED;
3508 
3509 	/*
3510 	 * Check to see if the vnode is in use.  If so, we have to call
3511 	 * VOP_CLOSE() and VOP_INACTIVE().
3512 	 */
3513 	active = vp->v_usecount > 0;
3514 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3515 	VI_UNLOCK(vp);
3516 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3517 
3518 	/*
3519 	 * If purging an active vnode, it must be closed and
3520 	 * deactivated before being reclaimed.
3521 	 */
3522 	if (active)
3523 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3524 	if (oweinact || active) {
3525 		VI_LOCK(vp);
3526 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3527 			vinactive(vp, td);
3528 		VI_UNLOCK(vp);
3529 	}
3530 	if (vp->v_type == VSOCK)
3531 		vfs_unp_reclaim(vp);
3532 
3533 	/*
3534 	 * Clean out any buffers associated with the vnode.
3535 	 * If the flush fails, just toss the buffers.
3536 	 */
3537 	mp = NULL;
3538 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3539 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3540 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3541 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3542 			;
3543 	}
3544 
3545 	BO_LOCK(&vp->v_bufobj);
3546 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3547 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3548 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3549 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3550 	    ("vp %p bufobj not invalidated", vp));
3551 
3552 	/*
3553 	 * For VMIO bufobj, BO_DEAD is set later, or in
3554 	 * vm_object_terminate() after the object's page queue is
3555 	 * flushed.
3556 	 */
3557 	object = vp->v_bufobj.bo_object;
3558 	if (object == NULL)
3559 		vp->v_bufobj.bo_flag |= BO_DEAD;
3560 	BO_UNLOCK(&vp->v_bufobj);
3561 
3562 	/*
3563 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3564 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3565 	 * should not touch the object borrowed from the lower vnode
3566 	 * (the handle check).
3567 	 */
3568 	if (object != NULL && object->type == OBJT_VNODE &&
3569 	    object->handle == vp)
3570 		vnode_destroy_vobject(vp);
3571 
3572 	/*
3573 	 * Reclaim the vnode.
3574 	 */
3575 	if (VOP_RECLAIM(vp, td))
3576 		panic("vgone: cannot reclaim");
3577 	if (mp != NULL)
3578 		vn_finished_secondary_write(mp);
3579 	VNASSERT(vp->v_object == NULL, vp,
3580 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3581 	/*
3582 	 * Clear the advisory locks and wake up waiting threads.
3583 	 */
3584 	(void)VOP_ADVLOCKPURGE(vp);
3585 	vp->v_lockf = NULL;
3586 	/*
3587 	 * Delete from old mount point vnode list.
3588 	 */
3589 	delmntque(vp);
3590 	cache_purge(vp);
3591 	/*
3592 	 * Done with purge, reset to the standard lock and invalidate
3593 	 * the vnode.
3594 	 */
3595 	VI_LOCK(vp);
3596 	vp->v_vnlock = &vp->v_lock;
3597 	vp->v_op = &dead_vnodeops;
3598 	vp->v_tag = "none";
3599 	vp->v_type = VBAD;
3600 }
3601 
3602 /*
3603  * Calculate the total number of references to a special device.
3604  */
3605 int
3606 vcount(struct vnode *vp)
3607 {
3608 	int count;
3609 
3610 	dev_lock();
3611 	count = vp->v_rdev->si_usecount;
3612 	dev_unlock();
3613 	return (count);
3614 }
3615 
3616 /*
3617  * Same as above, but using the struct cdev *as argument
3618  */
3619 int
3620 count_dev(struct cdev *dev)
3621 {
3622 	int count;
3623 
3624 	dev_lock();
3625 	count = dev->si_usecount;
3626 	dev_unlock();
3627 	return(count);
3628 }
3629 
3630 /*
3631  * Print out a description of a vnode.
3632  */
3633 static char *typename[] =
3634 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3635  "VMARKER"};
3636 
3637 void
3638 vn_printf(struct vnode *vp, const char *fmt, ...)
3639 {
3640 	va_list ap;
3641 	char buf[256], buf2[16];
3642 	u_long flags;
3643 
3644 	va_start(ap, fmt);
3645 	vprintf(fmt, ap);
3646 	va_end(ap);
3647 	printf("%p: ", (void *)vp);
3648 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3649 	printf("    usecount %d, writecount %d, refcount %d",
3650 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3651 	switch (vp->v_type) {
3652 	case VDIR:
3653 		printf(" mountedhere %p\n", vp->v_mountedhere);
3654 		break;
3655 	case VCHR:
3656 		printf(" rdev %p\n", vp->v_rdev);
3657 		break;
3658 	case VSOCK:
3659 		printf(" socket %p\n", vp->v_unpcb);
3660 		break;
3661 	case VFIFO:
3662 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3663 		break;
3664 	default:
3665 		printf("\n");
3666 		break;
3667 	}
3668 	buf[0] = '\0';
3669 	buf[1] = '\0';
3670 	if (vp->v_vflag & VV_ROOT)
3671 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3672 	if (vp->v_vflag & VV_ISTTY)
3673 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3674 	if (vp->v_vflag & VV_NOSYNC)
3675 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3676 	if (vp->v_vflag & VV_ETERNALDEV)
3677 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3678 	if (vp->v_vflag & VV_CACHEDLABEL)
3679 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3680 	if (vp->v_vflag & VV_COPYONWRITE)
3681 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3682 	if (vp->v_vflag & VV_SYSTEM)
3683 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3684 	if (vp->v_vflag & VV_PROCDEP)
3685 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3686 	if (vp->v_vflag & VV_NOKNOTE)
3687 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3688 	if (vp->v_vflag & VV_DELETED)
3689 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3690 	if (vp->v_vflag & VV_MD)
3691 		strlcat(buf, "|VV_MD", sizeof(buf));
3692 	if (vp->v_vflag & VV_FORCEINSMQ)
3693 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3694 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3695 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3696 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3697 	if (flags != 0) {
3698 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3699 		strlcat(buf, buf2, sizeof(buf));
3700 	}
3701 	if (vp->v_iflag & VI_MOUNT)
3702 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3703 	if (vp->v_iflag & VI_DOOMED)
3704 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3705 	if (vp->v_iflag & VI_FREE)
3706 		strlcat(buf, "|VI_FREE", sizeof(buf));
3707 	if (vp->v_iflag & VI_ACTIVE)
3708 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3709 	if (vp->v_iflag & VI_DOINGINACT)
3710 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3711 	if (vp->v_iflag & VI_OWEINACT)
3712 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3713 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3714 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3715 	if (flags != 0) {
3716 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3717 		strlcat(buf, buf2, sizeof(buf));
3718 	}
3719 	printf("    flags (%s)\n", buf + 1);
3720 	if (mtx_owned(VI_MTX(vp)))
3721 		printf(" VI_LOCKed");
3722 	if (vp->v_object != NULL)
3723 		printf("    v_object %p ref %d pages %d "
3724 		    "cleanbuf %d dirtybuf %d\n",
3725 		    vp->v_object, vp->v_object->ref_count,
3726 		    vp->v_object->resident_page_count,
3727 		    vp->v_bufobj.bo_clean.bv_cnt,
3728 		    vp->v_bufobj.bo_dirty.bv_cnt);
3729 	printf("    ");
3730 	lockmgr_printinfo(vp->v_vnlock);
3731 	if (vp->v_data != NULL)
3732 		VOP_PRINT(vp);
3733 }
3734 
3735 #ifdef DDB
3736 /*
3737  * List all of the locked vnodes in the system.
3738  * Called when debugging the kernel.
3739  */
3740 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3741 {
3742 	struct mount *mp;
3743 	struct vnode *vp;
3744 
3745 	/*
3746 	 * Note: because this is DDB, we can't obey the locking semantics
3747 	 * for these structures, which means we could catch an inconsistent
3748 	 * state and dereference a nasty pointer.  Not much to be done
3749 	 * about that.
3750 	 */
3751 	db_printf("Locked vnodes\n");
3752 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3753 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3754 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3755 				vn_printf(vp, "vnode ");
3756 		}
3757 	}
3758 }
3759 
3760 /*
3761  * Show details about the given vnode.
3762  */
3763 DB_SHOW_COMMAND(vnode, db_show_vnode)
3764 {
3765 	struct vnode *vp;
3766 
3767 	if (!have_addr)
3768 		return;
3769 	vp = (struct vnode *)addr;
3770 	vn_printf(vp, "vnode ");
3771 }
3772 
3773 /*
3774  * Show details about the given mount point.
3775  */
3776 DB_SHOW_COMMAND(mount, db_show_mount)
3777 {
3778 	struct mount *mp;
3779 	struct vfsopt *opt;
3780 	struct statfs *sp;
3781 	struct vnode *vp;
3782 	char buf[512];
3783 	uint64_t mflags;
3784 	u_int flags;
3785 
3786 	if (!have_addr) {
3787 		/* No address given, print short info about all mount points. */
3788 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3789 			db_printf("%p %s on %s (%s)\n", mp,
3790 			    mp->mnt_stat.f_mntfromname,
3791 			    mp->mnt_stat.f_mntonname,
3792 			    mp->mnt_stat.f_fstypename);
3793 			if (db_pager_quit)
3794 				break;
3795 		}
3796 		db_printf("\nMore info: show mount <addr>\n");
3797 		return;
3798 	}
3799 
3800 	mp = (struct mount *)addr;
3801 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3802 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3803 
3804 	buf[0] = '\0';
3805 	mflags = mp->mnt_flag;
3806 #define	MNT_FLAG(flag)	do {						\
3807 	if (mflags & (flag)) {						\
3808 		if (buf[0] != '\0')					\
3809 			strlcat(buf, ", ", sizeof(buf));		\
3810 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3811 		mflags &= ~(flag);					\
3812 	}								\
3813 } while (0)
3814 	MNT_FLAG(MNT_RDONLY);
3815 	MNT_FLAG(MNT_SYNCHRONOUS);
3816 	MNT_FLAG(MNT_NOEXEC);
3817 	MNT_FLAG(MNT_NOSUID);
3818 	MNT_FLAG(MNT_NFS4ACLS);
3819 	MNT_FLAG(MNT_UNION);
3820 	MNT_FLAG(MNT_ASYNC);
3821 	MNT_FLAG(MNT_SUIDDIR);
3822 	MNT_FLAG(MNT_SOFTDEP);
3823 	MNT_FLAG(MNT_NOSYMFOLLOW);
3824 	MNT_FLAG(MNT_GJOURNAL);
3825 	MNT_FLAG(MNT_MULTILABEL);
3826 	MNT_FLAG(MNT_ACLS);
3827 	MNT_FLAG(MNT_NOATIME);
3828 	MNT_FLAG(MNT_NOCLUSTERR);
3829 	MNT_FLAG(MNT_NOCLUSTERW);
3830 	MNT_FLAG(MNT_SUJ);
3831 	MNT_FLAG(MNT_EXRDONLY);
3832 	MNT_FLAG(MNT_EXPORTED);
3833 	MNT_FLAG(MNT_DEFEXPORTED);
3834 	MNT_FLAG(MNT_EXPORTANON);
3835 	MNT_FLAG(MNT_EXKERB);
3836 	MNT_FLAG(MNT_EXPUBLIC);
3837 	MNT_FLAG(MNT_LOCAL);
3838 	MNT_FLAG(MNT_QUOTA);
3839 	MNT_FLAG(MNT_ROOTFS);
3840 	MNT_FLAG(MNT_USER);
3841 	MNT_FLAG(MNT_IGNORE);
3842 	MNT_FLAG(MNT_UPDATE);
3843 	MNT_FLAG(MNT_DELEXPORT);
3844 	MNT_FLAG(MNT_RELOAD);
3845 	MNT_FLAG(MNT_FORCE);
3846 	MNT_FLAG(MNT_SNAPSHOT);
3847 	MNT_FLAG(MNT_BYFSID);
3848 #undef MNT_FLAG
3849 	if (mflags != 0) {
3850 		if (buf[0] != '\0')
3851 			strlcat(buf, ", ", sizeof(buf));
3852 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3853 		    "0x%016jx", mflags);
3854 	}
3855 	db_printf("    mnt_flag = %s\n", buf);
3856 
3857 	buf[0] = '\0';
3858 	flags = mp->mnt_kern_flag;
3859 #define	MNT_KERN_FLAG(flag)	do {					\
3860 	if (flags & (flag)) {						\
3861 		if (buf[0] != '\0')					\
3862 			strlcat(buf, ", ", sizeof(buf));		\
3863 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3864 		flags &= ~(flag);					\
3865 	}								\
3866 } while (0)
3867 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3868 	MNT_KERN_FLAG(MNTK_ASYNC);
3869 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3870 	MNT_KERN_FLAG(MNTK_DRAINING);
3871 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3872 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3873 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3874 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3875 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3876 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3877 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3878 	MNT_KERN_FLAG(MNTK_MARKER);
3879 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3880 	MNT_KERN_FLAG(MNTK_NOASYNC);
3881 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3882 	MNT_KERN_FLAG(MNTK_MWAIT);
3883 	MNT_KERN_FLAG(MNTK_SUSPEND);
3884 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3885 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3886 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3887 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3888 #undef MNT_KERN_FLAG
3889 	if (flags != 0) {
3890 		if (buf[0] != '\0')
3891 			strlcat(buf, ", ", sizeof(buf));
3892 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3893 		    "0x%08x", flags);
3894 	}
3895 	db_printf("    mnt_kern_flag = %s\n", buf);
3896 
3897 	db_printf("    mnt_opt = ");
3898 	opt = TAILQ_FIRST(mp->mnt_opt);
3899 	if (opt != NULL) {
3900 		db_printf("%s", opt->name);
3901 		opt = TAILQ_NEXT(opt, link);
3902 		while (opt != NULL) {
3903 			db_printf(", %s", opt->name);
3904 			opt = TAILQ_NEXT(opt, link);
3905 		}
3906 	}
3907 	db_printf("\n");
3908 
3909 	sp = &mp->mnt_stat;
3910 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3911 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3912 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3913 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3914 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3915 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3916 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3917 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3918 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3919 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3920 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3921 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3922 
3923 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3924 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3925 	if (jailed(mp->mnt_cred))
3926 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3927 	db_printf(" }\n");
3928 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3929 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3930 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3931 	db_printf("    mnt_activevnodelistsize = %d\n",
3932 	    mp->mnt_activevnodelistsize);
3933 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3934 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3935 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3936 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3937 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3938 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3939 	db_printf("    mnt_secondary_accwrites = %d\n",
3940 	    mp->mnt_secondary_accwrites);
3941 	db_printf("    mnt_gjprovider = %s\n",
3942 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3943 
3944 	db_printf("\n\nList of active vnodes\n");
3945 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3946 		if (vp->v_type != VMARKER) {
3947 			vn_printf(vp, "vnode ");
3948 			if (db_pager_quit)
3949 				break;
3950 		}
3951 	}
3952 	db_printf("\n\nList of inactive vnodes\n");
3953 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3954 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3955 			vn_printf(vp, "vnode ");
3956 			if (db_pager_quit)
3957 				break;
3958 		}
3959 	}
3960 }
3961 #endif	/* DDB */
3962 
3963 /*
3964  * Fill in a struct xvfsconf based on a struct vfsconf.
3965  */
3966 static int
3967 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3968 {
3969 	struct xvfsconf xvfsp;
3970 
3971 	bzero(&xvfsp, sizeof(xvfsp));
3972 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3973 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3974 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3975 	xvfsp.vfc_flags = vfsp->vfc_flags;
3976 	/*
3977 	 * These are unused in userland, we keep them
3978 	 * to not break binary compatibility.
3979 	 */
3980 	xvfsp.vfc_vfsops = NULL;
3981 	xvfsp.vfc_next = NULL;
3982 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3983 }
3984 
3985 #ifdef COMPAT_FREEBSD32
3986 struct xvfsconf32 {
3987 	uint32_t	vfc_vfsops;
3988 	char		vfc_name[MFSNAMELEN];
3989 	int32_t		vfc_typenum;
3990 	int32_t		vfc_refcount;
3991 	int32_t		vfc_flags;
3992 	uint32_t	vfc_next;
3993 };
3994 
3995 static int
3996 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3997 {
3998 	struct xvfsconf32 xvfsp;
3999 
4000 	bzero(&xvfsp, sizeof(xvfsp));
4001 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4002 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4003 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4004 	xvfsp.vfc_flags = vfsp->vfc_flags;
4005 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4006 }
4007 #endif
4008 
4009 /*
4010  * Top level filesystem related information gathering.
4011  */
4012 static int
4013 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4014 {
4015 	struct vfsconf *vfsp;
4016 	int error;
4017 
4018 	error = 0;
4019 	vfsconf_slock();
4020 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4021 #ifdef COMPAT_FREEBSD32
4022 		if (req->flags & SCTL_MASK32)
4023 			error = vfsconf2x32(req, vfsp);
4024 		else
4025 #endif
4026 			error = vfsconf2x(req, vfsp);
4027 		if (error)
4028 			break;
4029 	}
4030 	vfsconf_sunlock();
4031 	return (error);
4032 }
4033 
4034 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4035     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4036     "S,xvfsconf", "List of all configured filesystems");
4037 
4038 #ifndef BURN_BRIDGES
4039 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4040 
4041 static int
4042 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4043 {
4044 	int *name = (int *)arg1 - 1;	/* XXX */
4045 	u_int namelen = arg2 + 1;	/* XXX */
4046 	struct vfsconf *vfsp;
4047 
4048 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4049 	    "please rebuild world\n");
4050 
4051 #if 1 || defined(COMPAT_PRELITE2)
4052 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4053 	if (namelen == 1)
4054 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4055 #endif
4056 
4057 	switch (name[1]) {
4058 	case VFS_MAXTYPENUM:
4059 		if (namelen != 2)
4060 			return (ENOTDIR);
4061 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4062 	case VFS_CONF:
4063 		if (namelen != 3)
4064 			return (ENOTDIR);	/* overloaded */
4065 		vfsconf_slock();
4066 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4067 			if (vfsp->vfc_typenum == name[2])
4068 				break;
4069 		}
4070 		vfsconf_sunlock();
4071 		if (vfsp == NULL)
4072 			return (EOPNOTSUPP);
4073 #ifdef COMPAT_FREEBSD32
4074 		if (req->flags & SCTL_MASK32)
4075 			return (vfsconf2x32(req, vfsp));
4076 		else
4077 #endif
4078 			return (vfsconf2x(req, vfsp));
4079 	}
4080 	return (EOPNOTSUPP);
4081 }
4082 
4083 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4084     CTLFLAG_MPSAFE, vfs_sysctl,
4085     "Generic filesystem");
4086 
4087 #if 1 || defined(COMPAT_PRELITE2)
4088 
4089 static int
4090 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4091 {
4092 	int error;
4093 	struct vfsconf *vfsp;
4094 	struct ovfsconf ovfs;
4095 
4096 	vfsconf_slock();
4097 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4098 		bzero(&ovfs, sizeof(ovfs));
4099 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4100 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4101 		ovfs.vfc_index = vfsp->vfc_typenum;
4102 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4103 		ovfs.vfc_flags = vfsp->vfc_flags;
4104 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4105 		if (error != 0) {
4106 			vfsconf_sunlock();
4107 			return (error);
4108 		}
4109 	}
4110 	vfsconf_sunlock();
4111 	return (0);
4112 }
4113 
4114 #endif /* 1 || COMPAT_PRELITE2 */
4115 #endif /* !BURN_BRIDGES */
4116 
4117 #define KINFO_VNODESLOP		10
4118 #ifdef notyet
4119 /*
4120  * Dump vnode list (via sysctl).
4121  */
4122 /* ARGSUSED */
4123 static int
4124 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4125 {
4126 	struct xvnode *xvn;
4127 	struct mount *mp;
4128 	struct vnode *vp;
4129 	int error, len, n;
4130 
4131 	/*
4132 	 * Stale numvnodes access is not fatal here.
4133 	 */
4134 	req->lock = 0;
4135 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4136 	if (!req->oldptr)
4137 		/* Make an estimate */
4138 		return (SYSCTL_OUT(req, 0, len));
4139 
4140 	error = sysctl_wire_old_buffer(req, 0);
4141 	if (error != 0)
4142 		return (error);
4143 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4144 	n = 0;
4145 	mtx_lock(&mountlist_mtx);
4146 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4147 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4148 			continue;
4149 		MNT_ILOCK(mp);
4150 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4151 			if (n == len)
4152 				break;
4153 			vref(vp);
4154 			xvn[n].xv_size = sizeof *xvn;
4155 			xvn[n].xv_vnode = vp;
4156 			xvn[n].xv_id = 0;	/* XXX compat */
4157 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4158 			XV_COPY(usecount);
4159 			XV_COPY(writecount);
4160 			XV_COPY(holdcnt);
4161 			XV_COPY(mount);
4162 			XV_COPY(numoutput);
4163 			XV_COPY(type);
4164 #undef XV_COPY
4165 			xvn[n].xv_flag = vp->v_vflag;
4166 
4167 			switch (vp->v_type) {
4168 			case VREG:
4169 			case VDIR:
4170 			case VLNK:
4171 				break;
4172 			case VBLK:
4173 			case VCHR:
4174 				if (vp->v_rdev == NULL) {
4175 					vrele(vp);
4176 					continue;
4177 				}
4178 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4179 				break;
4180 			case VSOCK:
4181 				xvn[n].xv_socket = vp->v_socket;
4182 				break;
4183 			case VFIFO:
4184 				xvn[n].xv_fifo = vp->v_fifoinfo;
4185 				break;
4186 			case VNON:
4187 			case VBAD:
4188 			default:
4189 				/* shouldn't happen? */
4190 				vrele(vp);
4191 				continue;
4192 			}
4193 			vrele(vp);
4194 			++n;
4195 		}
4196 		MNT_IUNLOCK(mp);
4197 		mtx_lock(&mountlist_mtx);
4198 		vfs_unbusy(mp);
4199 		if (n == len)
4200 			break;
4201 	}
4202 	mtx_unlock(&mountlist_mtx);
4203 
4204 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4205 	free(xvn, M_TEMP);
4206 	return (error);
4207 }
4208 
4209 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4210     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4211     "");
4212 #endif
4213 
4214 static void
4215 unmount_or_warn(struct mount *mp)
4216 {
4217 	int error;
4218 
4219 	error = dounmount(mp, MNT_FORCE, curthread);
4220 	if (error != 0) {
4221 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4222 		if (error == EBUSY)
4223 			printf("BUSY)\n");
4224 		else
4225 			printf("%d)\n", error);
4226 	}
4227 }
4228 
4229 /*
4230  * Unmount all filesystems. The list is traversed in reverse order
4231  * of mounting to avoid dependencies.
4232  */
4233 void
4234 vfs_unmountall(void)
4235 {
4236 	struct mount *mp, *tmp;
4237 
4238 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4239 
4240 	/*
4241 	 * Since this only runs when rebooting, it is not interlocked.
4242 	 */
4243 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4244 		vfs_ref(mp);
4245 
4246 		/*
4247 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4248 		 * unmount of the latter.
4249 		 */
4250 		if (mp == rootdevmp)
4251 			continue;
4252 
4253 		unmount_or_warn(mp);
4254 	}
4255 
4256 	if (rootdevmp != NULL)
4257 		unmount_or_warn(rootdevmp);
4258 }
4259 
4260 /*
4261  * perform msync on all vnodes under a mount point
4262  * the mount point must be locked.
4263  */
4264 void
4265 vfs_msync(struct mount *mp, int flags)
4266 {
4267 	struct vnode *vp, *mvp;
4268 	struct vm_object *obj;
4269 
4270 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4271 
4272 	vnlru_return_batch(mp);
4273 
4274 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4275 		obj = vp->v_object;
4276 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4277 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4278 			if (!vget(vp,
4279 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4280 			    curthread)) {
4281 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4282 					vput(vp);
4283 					continue;
4284 				}
4285 
4286 				obj = vp->v_object;
4287 				if (obj != NULL) {
4288 					VM_OBJECT_WLOCK(obj);
4289 					vm_object_page_clean(obj, 0, 0,
4290 					    flags == MNT_WAIT ?
4291 					    OBJPC_SYNC : OBJPC_NOSYNC);
4292 					VM_OBJECT_WUNLOCK(obj);
4293 				}
4294 				vput(vp);
4295 			}
4296 		} else
4297 			VI_UNLOCK(vp);
4298 	}
4299 }
4300 
4301 static void
4302 destroy_vpollinfo_free(struct vpollinfo *vi)
4303 {
4304 
4305 	knlist_destroy(&vi->vpi_selinfo.si_note);
4306 	mtx_destroy(&vi->vpi_lock);
4307 	uma_zfree(vnodepoll_zone, vi);
4308 }
4309 
4310 static void
4311 destroy_vpollinfo(struct vpollinfo *vi)
4312 {
4313 
4314 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4315 	seldrain(&vi->vpi_selinfo);
4316 	destroy_vpollinfo_free(vi);
4317 }
4318 
4319 /*
4320  * Initialize per-vnode helper structure to hold poll-related state.
4321  */
4322 void
4323 v_addpollinfo(struct vnode *vp)
4324 {
4325 	struct vpollinfo *vi;
4326 
4327 	if (vp->v_pollinfo != NULL)
4328 		return;
4329 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4330 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4331 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4332 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4333 	VI_LOCK(vp);
4334 	if (vp->v_pollinfo != NULL) {
4335 		VI_UNLOCK(vp);
4336 		destroy_vpollinfo_free(vi);
4337 		return;
4338 	}
4339 	vp->v_pollinfo = vi;
4340 	VI_UNLOCK(vp);
4341 }
4342 
4343 /*
4344  * Record a process's interest in events which might happen to
4345  * a vnode.  Because poll uses the historic select-style interface
4346  * internally, this routine serves as both the ``check for any
4347  * pending events'' and the ``record my interest in future events''
4348  * functions.  (These are done together, while the lock is held,
4349  * to avoid race conditions.)
4350  */
4351 int
4352 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4353 {
4354 
4355 	v_addpollinfo(vp);
4356 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4357 	if (vp->v_pollinfo->vpi_revents & events) {
4358 		/*
4359 		 * This leaves events we are not interested
4360 		 * in available for the other process which
4361 		 * which presumably had requested them
4362 		 * (otherwise they would never have been
4363 		 * recorded).
4364 		 */
4365 		events &= vp->v_pollinfo->vpi_revents;
4366 		vp->v_pollinfo->vpi_revents &= ~events;
4367 
4368 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4369 		return (events);
4370 	}
4371 	vp->v_pollinfo->vpi_events |= events;
4372 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4373 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4374 	return (0);
4375 }
4376 
4377 /*
4378  * Routine to create and manage a filesystem syncer vnode.
4379  */
4380 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4381 static int	sync_fsync(struct  vop_fsync_args *);
4382 static int	sync_inactive(struct  vop_inactive_args *);
4383 static int	sync_reclaim(struct  vop_reclaim_args *);
4384 
4385 static struct vop_vector sync_vnodeops = {
4386 	.vop_bypass =	VOP_EOPNOTSUPP,
4387 	.vop_close =	sync_close,		/* close */
4388 	.vop_fsync =	sync_fsync,		/* fsync */
4389 	.vop_inactive =	sync_inactive,	/* inactive */
4390 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4391 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4392 	.vop_lock1 =	vop_stdlock,	/* lock */
4393 	.vop_unlock =	vop_stdunlock,	/* unlock */
4394 	.vop_islocked =	vop_stdislocked,	/* islocked */
4395 };
4396 
4397 /*
4398  * Create a new filesystem syncer vnode for the specified mount point.
4399  */
4400 void
4401 vfs_allocate_syncvnode(struct mount *mp)
4402 {
4403 	struct vnode *vp;
4404 	struct bufobj *bo;
4405 	static long start, incr, next;
4406 	int error;
4407 
4408 	/* Allocate a new vnode */
4409 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4410 	if (error != 0)
4411 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4412 	vp->v_type = VNON;
4413 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4414 	vp->v_vflag |= VV_FORCEINSMQ;
4415 	error = insmntque(vp, mp);
4416 	if (error != 0)
4417 		panic("vfs_allocate_syncvnode: insmntque() failed");
4418 	vp->v_vflag &= ~VV_FORCEINSMQ;
4419 	VOP_UNLOCK(vp, 0);
4420 	/*
4421 	 * Place the vnode onto the syncer worklist. We attempt to
4422 	 * scatter them about on the list so that they will go off
4423 	 * at evenly distributed times even if all the filesystems
4424 	 * are mounted at once.
4425 	 */
4426 	next += incr;
4427 	if (next == 0 || next > syncer_maxdelay) {
4428 		start /= 2;
4429 		incr /= 2;
4430 		if (start == 0) {
4431 			start = syncer_maxdelay / 2;
4432 			incr = syncer_maxdelay;
4433 		}
4434 		next = start;
4435 	}
4436 	bo = &vp->v_bufobj;
4437 	BO_LOCK(bo);
4438 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4439 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4440 	mtx_lock(&sync_mtx);
4441 	sync_vnode_count++;
4442 	if (mp->mnt_syncer == NULL) {
4443 		mp->mnt_syncer = vp;
4444 		vp = NULL;
4445 	}
4446 	mtx_unlock(&sync_mtx);
4447 	BO_UNLOCK(bo);
4448 	if (vp != NULL) {
4449 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4450 		vgone(vp);
4451 		vput(vp);
4452 	}
4453 }
4454 
4455 void
4456 vfs_deallocate_syncvnode(struct mount *mp)
4457 {
4458 	struct vnode *vp;
4459 
4460 	mtx_lock(&sync_mtx);
4461 	vp = mp->mnt_syncer;
4462 	if (vp != NULL)
4463 		mp->mnt_syncer = NULL;
4464 	mtx_unlock(&sync_mtx);
4465 	if (vp != NULL)
4466 		vrele(vp);
4467 }
4468 
4469 /*
4470  * Do a lazy sync of the filesystem.
4471  */
4472 static int
4473 sync_fsync(struct vop_fsync_args *ap)
4474 {
4475 	struct vnode *syncvp = ap->a_vp;
4476 	struct mount *mp = syncvp->v_mount;
4477 	int error, save;
4478 	struct bufobj *bo;
4479 
4480 	/*
4481 	 * We only need to do something if this is a lazy evaluation.
4482 	 */
4483 	if (ap->a_waitfor != MNT_LAZY)
4484 		return (0);
4485 
4486 	/*
4487 	 * Move ourselves to the back of the sync list.
4488 	 */
4489 	bo = &syncvp->v_bufobj;
4490 	BO_LOCK(bo);
4491 	vn_syncer_add_to_worklist(bo, syncdelay);
4492 	BO_UNLOCK(bo);
4493 
4494 	/*
4495 	 * Walk the list of vnodes pushing all that are dirty and
4496 	 * not already on the sync list.
4497 	 */
4498 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4499 		return (0);
4500 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4501 		vfs_unbusy(mp);
4502 		return (0);
4503 	}
4504 	save = curthread_pflags_set(TDP_SYNCIO);
4505 	vfs_msync(mp, MNT_NOWAIT);
4506 	error = VFS_SYNC(mp, MNT_LAZY);
4507 	curthread_pflags_restore(save);
4508 	vn_finished_write(mp);
4509 	vfs_unbusy(mp);
4510 	return (error);
4511 }
4512 
4513 /*
4514  * The syncer vnode is no referenced.
4515  */
4516 static int
4517 sync_inactive(struct vop_inactive_args *ap)
4518 {
4519 
4520 	vgone(ap->a_vp);
4521 	return (0);
4522 }
4523 
4524 /*
4525  * The syncer vnode is no longer needed and is being decommissioned.
4526  *
4527  * Modifications to the worklist must be protected by sync_mtx.
4528  */
4529 static int
4530 sync_reclaim(struct vop_reclaim_args *ap)
4531 {
4532 	struct vnode *vp = ap->a_vp;
4533 	struct bufobj *bo;
4534 
4535 	bo = &vp->v_bufobj;
4536 	BO_LOCK(bo);
4537 	mtx_lock(&sync_mtx);
4538 	if (vp->v_mount->mnt_syncer == vp)
4539 		vp->v_mount->mnt_syncer = NULL;
4540 	if (bo->bo_flag & BO_ONWORKLST) {
4541 		LIST_REMOVE(bo, bo_synclist);
4542 		syncer_worklist_len--;
4543 		sync_vnode_count--;
4544 		bo->bo_flag &= ~BO_ONWORKLST;
4545 	}
4546 	mtx_unlock(&sync_mtx);
4547 	BO_UNLOCK(bo);
4548 
4549 	return (0);
4550 }
4551 
4552 int
4553 vn_need_pageq_flush(struct vnode *vp)
4554 {
4555 	struct vm_object *obj;
4556 	int need;
4557 
4558 	MPASS(mtx_owned(VI_MTX(vp)));
4559 	need = 0;
4560 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4561 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0)
4562 		need = 1;
4563 	return (need);
4564 }
4565 
4566 /*
4567  * Check if vnode represents a disk device
4568  */
4569 int
4570 vn_isdisk(struct vnode *vp, int *errp)
4571 {
4572 	int error;
4573 
4574 	if (vp->v_type != VCHR) {
4575 		error = ENOTBLK;
4576 		goto out;
4577 	}
4578 	error = 0;
4579 	dev_lock();
4580 	if (vp->v_rdev == NULL)
4581 		error = ENXIO;
4582 	else if (vp->v_rdev->si_devsw == NULL)
4583 		error = ENXIO;
4584 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4585 		error = ENOTBLK;
4586 	dev_unlock();
4587 out:
4588 	if (errp != NULL)
4589 		*errp = error;
4590 	return (error == 0);
4591 }
4592 
4593 /*
4594  * Common filesystem object access control check routine.  Accepts a
4595  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4596  * and optional call-by-reference privused argument allowing vaccess()
4597  * to indicate to the caller whether privilege was used to satisfy the
4598  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4599  */
4600 int
4601 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4602     accmode_t accmode, struct ucred *cred, int *privused)
4603 {
4604 	accmode_t dac_granted;
4605 	accmode_t priv_granted;
4606 
4607 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4608 	    ("invalid bit in accmode"));
4609 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4610 	    ("VAPPEND without VWRITE"));
4611 
4612 	/*
4613 	 * Look for a normal, non-privileged way to access the file/directory
4614 	 * as requested.  If it exists, go with that.
4615 	 */
4616 
4617 	if (privused != NULL)
4618 		*privused = 0;
4619 
4620 	dac_granted = 0;
4621 
4622 	/* Check the owner. */
4623 	if (cred->cr_uid == file_uid) {
4624 		dac_granted |= VADMIN;
4625 		if (file_mode & S_IXUSR)
4626 			dac_granted |= VEXEC;
4627 		if (file_mode & S_IRUSR)
4628 			dac_granted |= VREAD;
4629 		if (file_mode & S_IWUSR)
4630 			dac_granted |= (VWRITE | VAPPEND);
4631 
4632 		if ((accmode & dac_granted) == accmode)
4633 			return (0);
4634 
4635 		goto privcheck;
4636 	}
4637 
4638 	/* Otherwise, check the groups (first match) */
4639 	if (groupmember(file_gid, cred)) {
4640 		if (file_mode & S_IXGRP)
4641 			dac_granted |= VEXEC;
4642 		if (file_mode & S_IRGRP)
4643 			dac_granted |= VREAD;
4644 		if (file_mode & S_IWGRP)
4645 			dac_granted |= (VWRITE | VAPPEND);
4646 
4647 		if ((accmode & dac_granted) == accmode)
4648 			return (0);
4649 
4650 		goto privcheck;
4651 	}
4652 
4653 	/* Otherwise, check everyone else. */
4654 	if (file_mode & S_IXOTH)
4655 		dac_granted |= VEXEC;
4656 	if (file_mode & S_IROTH)
4657 		dac_granted |= VREAD;
4658 	if (file_mode & S_IWOTH)
4659 		dac_granted |= (VWRITE | VAPPEND);
4660 	if ((accmode & dac_granted) == accmode)
4661 		return (0);
4662 
4663 privcheck:
4664 	/*
4665 	 * Build a privilege mask to determine if the set of privileges
4666 	 * satisfies the requirements when combined with the granted mask
4667 	 * from above.  For each privilege, if the privilege is required,
4668 	 * bitwise or the request type onto the priv_granted mask.
4669 	 */
4670 	priv_granted = 0;
4671 
4672 	if (type == VDIR) {
4673 		/*
4674 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4675 		 * requests, instead of PRIV_VFS_EXEC.
4676 		 */
4677 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4678 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4679 			priv_granted |= VEXEC;
4680 	} else {
4681 		/*
4682 		 * Ensure that at least one execute bit is on. Otherwise,
4683 		 * a privileged user will always succeed, and we don't want
4684 		 * this to happen unless the file really is executable.
4685 		 */
4686 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4687 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4688 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4689 			priv_granted |= VEXEC;
4690 	}
4691 
4692 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4693 	    !priv_check_cred(cred, PRIV_VFS_READ))
4694 		priv_granted |= VREAD;
4695 
4696 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4697 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4698 		priv_granted |= (VWRITE | VAPPEND);
4699 
4700 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4701 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4702 		priv_granted |= VADMIN;
4703 
4704 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4705 		/* XXX audit: privilege used */
4706 		if (privused != NULL)
4707 			*privused = 1;
4708 		return (0);
4709 	}
4710 
4711 	return ((accmode & VADMIN) ? EPERM : EACCES);
4712 }
4713 
4714 /*
4715  * Credential check based on process requesting service, and per-attribute
4716  * permissions.
4717  */
4718 int
4719 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4720     struct thread *td, accmode_t accmode)
4721 {
4722 
4723 	/*
4724 	 * Kernel-invoked always succeeds.
4725 	 */
4726 	if (cred == NOCRED)
4727 		return (0);
4728 
4729 	/*
4730 	 * Do not allow privileged processes in jail to directly manipulate
4731 	 * system attributes.
4732 	 */
4733 	switch (attrnamespace) {
4734 	case EXTATTR_NAMESPACE_SYSTEM:
4735 		/* Potentially should be: return (EPERM); */
4736 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4737 	case EXTATTR_NAMESPACE_USER:
4738 		return (VOP_ACCESS(vp, accmode, cred, td));
4739 	default:
4740 		return (EPERM);
4741 	}
4742 }
4743 
4744 #ifdef DEBUG_VFS_LOCKS
4745 /*
4746  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4747  * no longer ok to have an unlocked VFS.
4748  */
4749 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4750 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4751 
4752 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4753 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4754     "Drop into debugger on lock violation");
4755 
4756 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4757 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4758     0, "Check for interlock across VOPs");
4759 
4760 int vfs_badlock_print = 1;	/* Print lock violations. */
4761 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4762     0, "Print lock violations");
4763 
4764 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4765 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4766     0, "Print vnode details on lock violations");
4767 
4768 #ifdef KDB
4769 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4770 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4771     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4772 #endif
4773 
4774 static void
4775 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4776 {
4777 
4778 #ifdef KDB
4779 	if (vfs_badlock_backtrace)
4780 		kdb_backtrace();
4781 #endif
4782 	if (vfs_badlock_vnode)
4783 		vn_printf(vp, "vnode ");
4784 	if (vfs_badlock_print)
4785 		printf("%s: %p %s\n", str, (void *)vp, msg);
4786 	if (vfs_badlock_ddb)
4787 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4788 }
4789 
4790 void
4791 assert_vi_locked(struct vnode *vp, const char *str)
4792 {
4793 
4794 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4795 		vfs_badlock("interlock is not locked but should be", str, vp);
4796 }
4797 
4798 void
4799 assert_vi_unlocked(struct vnode *vp, const char *str)
4800 {
4801 
4802 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4803 		vfs_badlock("interlock is locked but should not be", str, vp);
4804 }
4805 
4806 void
4807 assert_vop_locked(struct vnode *vp, const char *str)
4808 {
4809 	int locked;
4810 
4811 	if (!IGNORE_LOCK(vp)) {
4812 		locked = VOP_ISLOCKED(vp);
4813 		if (locked == 0 || locked == LK_EXCLOTHER)
4814 			vfs_badlock("is not locked but should be", str, vp);
4815 	}
4816 }
4817 
4818 void
4819 assert_vop_unlocked(struct vnode *vp, const char *str)
4820 {
4821 
4822 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4823 		vfs_badlock("is locked but should not be", str, vp);
4824 }
4825 
4826 void
4827 assert_vop_elocked(struct vnode *vp, const char *str)
4828 {
4829 
4830 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4831 		vfs_badlock("is not exclusive locked but should be", str, vp);
4832 }
4833 #endif /* DEBUG_VFS_LOCKS */
4834 
4835 void
4836 vop_rename_fail(struct vop_rename_args *ap)
4837 {
4838 
4839 	if (ap->a_tvp != NULL)
4840 		vput(ap->a_tvp);
4841 	if (ap->a_tdvp == ap->a_tvp)
4842 		vrele(ap->a_tdvp);
4843 	else
4844 		vput(ap->a_tdvp);
4845 	vrele(ap->a_fdvp);
4846 	vrele(ap->a_fvp);
4847 }
4848 
4849 void
4850 vop_rename_pre(void *ap)
4851 {
4852 	struct vop_rename_args *a = ap;
4853 
4854 #ifdef DEBUG_VFS_LOCKS
4855 	if (a->a_tvp)
4856 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4857 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4858 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4859 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4860 
4861 	/* Check the source (from). */
4862 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4863 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4864 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4865 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4866 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4867 
4868 	/* Check the target. */
4869 	if (a->a_tvp)
4870 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4871 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4872 #endif
4873 	if (a->a_tdvp != a->a_fdvp)
4874 		vhold(a->a_fdvp);
4875 	if (a->a_tvp != a->a_fvp)
4876 		vhold(a->a_fvp);
4877 	vhold(a->a_tdvp);
4878 	if (a->a_tvp)
4879 		vhold(a->a_tvp);
4880 }
4881 
4882 #ifdef DEBUG_VFS_LOCKS
4883 void
4884 vop_strategy_pre(void *ap)
4885 {
4886 	struct vop_strategy_args *a;
4887 	struct buf *bp;
4888 
4889 	a = ap;
4890 	bp = a->a_bp;
4891 
4892 	/*
4893 	 * Cluster ops lock their component buffers but not the IO container.
4894 	 */
4895 	if ((bp->b_flags & B_CLUSTER) != 0)
4896 		return;
4897 
4898 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4899 		if (vfs_badlock_print)
4900 			printf(
4901 			    "VOP_STRATEGY: bp is not locked but should be\n");
4902 		if (vfs_badlock_ddb)
4903 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4904 	}
4905 }
4906 
4907 void
4908 vop_lock_pre(void *ap)
4909 {
4910 	struct vop_lock1_args *a = ap;
4911 
4912 	if ((a->a_flags & LK_INTERLOCK) == 0)
4913 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4914 	else
4915 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4916 }
4917 
4918 void
4919 vop_lock_post(void *ap, int rc)
4920 {
4921 	struct vop_lock1_args *a = ap;
4922 
4923 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4924 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4925 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4926 }
4927 
4928 void
4929 vop_unlock_pre(void *ap)
4930 {
4931 	struct vop_unlock_args *a = ap;
4932 
4933 	if (a->a_flags & LK_INTERLOCK)
4934 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4935 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4936 }
4937 
4938 void
4939 vop_unlock_post(void *ap, int rc)
4940 {
4941 	struct vop_unlock_args *a = ap;
4942 
4943 	if (a->a_flags & LK_INTERLOCK)
4944 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4945 }
4946 
4947 void
4948 vop_need_inactive_pre(void *ap)
4949 {
4950 	struct vop_need_inactive_args *a = ap;
4951 
4952 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
4953 }
4954 
4955 void
4956 vop_need_inactive_post(void *ap, int rc)
4957 {
4958 	struct vop_need_inactive_args *a = ap;
4959 
4960 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
4961 }
4962 #endif
4963 
4964 void
4965 vop_create_post(void *ap, int rc)
4966 {
4967 	struct vop_create_args *a = ap;
4968 
4969 	if (!rc)
4970 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4971 }
4972 
4973 void
4974 vop_deleteextattr_post(void *ap, int rc)
4975 {
4976 	struct vop_deleteextattr_args *a = ap;
4977 
4978 	if (!rc)
4979 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4980 }
4981 
4982 void
4983 vop_link_post(void *ap, int rc)
4984 {
4985 	struct vop_link_args *a = ap;
4986 
4987 	if (!rc) {
4988 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4989 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4990 	}
4991 }
4992 
4993 void
4994 vop_mkdir_post(void *ap, int rc)
4995 {
4996 	struct vop_mkdir_args *a = ap;
4997 
4998 	if (!rc)
4999 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5000 }
5001 
5002 void
5003 vop_mknod_post(void *ap, int rc)
5004 {
5005 	struct vop_mknod_args *a = ap;
5006 
5007 	if (!rc)
5008 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5009 }
5010 
5011 void
5012 vop_reclaim_post(void *ap, int rc)
5013 {
5014 	struct vop_reclaim_args *a = ap;
5015 
5016 	if (!rc)
5017 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5018 }
5019 
5020 void
5021 vop_remove_post(void *ap, int rc)
5022 {
5023 	struct vop_remove_args *a = ap;
5024 
5025 	if (!rc) {
5026 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5027 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5028 	}
5029 }
5030 
5031 void
5032 vop_rename_post(void *ap, int rc)
5033 {
5034 	struct vop_rename_args *a = ap;
5035 	long hint;
5036 
5037 	if (!rc) {
5038 		hint = NOTE_WRITE;
5039 		if (a->a_fdvp == a->a_tdvp) {
5040 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5041 				hint |= NOTE_LINK;
5042 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5043 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5044 		} else {
5045 			hint |= NOTE_EXTEND;
5046 			if (a->a_fvp->v_type == VDIR)
5047 				hint |= NOTE_LINK;
5048 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5049 
5050 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5051 			    a->a_tvp->v_type == VDIR)
5052 				hint &= ~NOTE_LINK;
5053 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5054 		}
5055 
5056 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5057 		if (a->a_tvp)
5058 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5059 	}
5060 	if (a->a_tdvp != a->a_fdvp)
5061 		vdrop(a->a_fdvp);
5062 	if (a->a_tvp != a->a_fvp)
5063 		vdrop(a->a_fvp);
5064 	vdrop(a->a_tdvp);
5065 	if (a->a_tvp)
5066 		vdrop(a->a_tvp);
5067 }
5068 
5069 void
5070 vop_rmdir_post(void *ap, int rc)
5071 {
5072 	struct vop_rmdir_args *a = ap;
5073 
5074 	if (!rc) {
5075 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5076 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5077 	}
5078 }
5079 
5080 void
5081 vop_setattr_post(void *ap, int rc)
5082 {
5083 	struct vop_setattr_args *a = ap;
5084 
5085 	if (!rc)
5086 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5087 }
5088 
5089 void
5090 vop_setextattr_post(void *ap, int rc)
5091 {
5092 	struct vop_setextattr_args *a = ap;
5093 
5094 	if (!rc)
5095 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5096 }
5097 
5098 void
5099 vop_symlink_post(void *ap, int rc)
5100 {
5101 	struct vop_symlink_args *a = ap;
5102 
5103 	if (!rc)
5104 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5105 }
5106 
5107 void
5108 vop_open_post(void *ap, int rc)
5109 {
5110 	struct vop_open_args *a = ap;
5111 
5112 	if (!rc)
5113 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5114 }
5115 
5116 void
5117 vop_close_post(void *ap, int rc)
5118 {
5119 	struct vop_close_args *a = ap;
5120 
5121 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5122 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
5123 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5124 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5125 	}
5126 }
5127 
5128 void
5129 vop_read_post(void *ap, int rc)
5130 {
5131 	struct vop_read_args *a = ap;
5132 
5133 	if (!rc)
5134 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5135 }
5136 
5137 void
5138 vop_readdir_post(void *ap, int rc)
5139 {
5140 	struct vop_readdir_args *a = ap;
5141 
5142 	if (!rc)
5143 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5144 }
5145 
5146 static struct knlist fs_knlist;
5147 
5148 static void
5149 vfs_event_init(void *arg)
5150 {
5151 	knlist_init_mtx(&fs_knlist, NULL);
5152 }
5153 /* XXX - correct order? */
5154 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5155 
5156 void
5157 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5158 {
5159 
5160 	KNOTE_UNLOCKED(&fs_knlist, event);
5161 }
5162 
5163 static int	filt_fsattach(struct knote *kn);
5164 static void	filt_fsdetach(struct knote *kn);
5165 static int	filt_fsevent(struct knote *kn, long hint);
5166 
5167 struct filterops fs_filtops = {
5168 	.f_isfd = 0,
5169 	.f_attach = filt_fsattach,
5170 	.f_detach = filt_fsdetach,
5171 	.f_event = filt_fsevent
5172 };
5173 
5174 static int
5175 filt_fsattach(struct knote *kn)
5176 {
5177 
5178 	kn->kn_flags |= EV_CLEAR;
5179 	knlist_add(&fs_knlist, kn, 0);
5180 	return (0);
5181 }
5182 
5183 static void
5184 filt_fsdetach(struct knote *kn)
5185 {
5186 
5187 	knlist_remove(&fs_knlist, kn, 0);
5188 }
5189 
5190 static int
5191 filt_fsevent(struct knote *kn, long hint)
5192 {
5193 
5194 	kn->kn_fflags |= hint;
5195 	return (kn->kn_fflags != 0);
5196 }
5197 
5198 static int
5199 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5200 {
5201 	struct vfsidctl vc;
5202 	int error;
5203 	struct mount *mp;
5204 
5205 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5206 	if (error)
5207 		return (error);
5208 	if (vc.vc_vers != VFS_CTL_VERS1)
5209 		return (EINVAL);
5210 	mp = vfs_getvfs(&vc.vc_fsid);
5211 	if (mp == NULL)
5212 		return (ENOENT);
5213 	/* ensure that a specific sysctl goes to the right filesystem. */
5214 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5215 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5216 		vfs_rel(mp);
5217 		return (EINVAL);
5218 	}
5219 	VCTLTOREQ(&vc, req);
5220 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5221 	vfs_rel(mp);
5222 	return (error);
5223 }
5224 
5225 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5226     NULL, 0, sysctl_vfs_ctl, "",
5227     "Sysctl by fsid");
5228 
5229 /*
5230  * Function to initialize a va_filerev field sensibly.
5231  * XXX: Wouldn't a random number make a lot more sense ??
5232  */
5233 u_quad_t
5234 init_va_filerev(void)
5235 {
5236 	struct bintime bt;
5237 
5238 	getbinuptime(&bt);
5239 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5240 }
5241 
5242 static int	filt_vfsread(struct knote *kn, long hint);
5243 static int	filt_vfswrite(struct knote *kn, long hint);
5244 static int	filt_vfsvnode(struct knote *kn, long hint);
5245 static void	filt_vfsdetach(struct knote *kn);
5246 static struct filterops vfsread_filtops = {
5247 	.f_isfd = 1,
5248 	.f_detach = filt_vfsdetach,
5249 	.f_event = filt_vfsread
5250 };
5251 static struct filterops vfswrite_filtops = {
5252 	.f_isfd = 1,
5253 	.f_detach = filt_vfsdetach,
5254 	.f_event = filt_vfswrite
5255 };
5256 static struct filterops vfsvnode_filtops = {
5257 	.f_isfd = 1,
5258 	.f_detach = filt_vfsdetach,
5259 	.f_event = filt_vfsvnode
5260 };
5261 
5262 static void
5263 vfs_knllock(void *arg)
5264 {
5265 	struct vnode *vp = arg;
5266 
5267 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5268 }
5269 
5270 static void
5271 vfs_knlunlock(void *arg)
5272 {
5273 	struct vnode *vp = arg;
5274 
5275 	VOP_UNLOCK(vp, 0);
5276 }
5277 
5278 static void
5279 vfs_knl_assert_locked(void *arg)
5280 {
5281 #ifdef DEBUG_VFS_LOCKS
5282 	struct vnode *vp = arg;
5283 
5284 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5285 #endif
5286 }
5287 
5288 static void
5289 vfs_knl_assert_unlocked(void *arg)
5290 {
5291 #ifdef DEBUG_VFS_LOCKS
5292 	struct vnode *vp = arg;
5293 
5294 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5295 #endif
5296 }
5297 
5298 int
5299 vfs_kqfilter(struct vop_kqfilter_args *ap)
5300 {
5301 	struct vnode *vp = ap->a_vp;
5302 	struct knote *kn = ap->a_kn;
5303 	struct knlist *knl;
5304 
5305 	switch (kn->kn_filter) {
5306 	case EVFILT_READ:
5307 		kn->kn_fop = &vfsread_filtops;
5308 		break;
5309 	case EVFILT_WRITE:
5310 		kn->kn_fop = &vfswrite_filtops;
5311 		break;
5312 	case EVFILT_VNODE:
5313 		kn->kn_fop = &vfsvnode_filtops;
5314 		break;
5315 	default:
5316 		return (EINVAL);
5317 	}
5318 
5319 	kn->kn_hook = (caddr_t)vp;
5320 
5321 	v_addpollinfo(vp);
5322 	if (vp->v_pollinfo == NULL)
5323 		return (ENOMEM);
5324 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5325 	vhold(vp);
5326 	knlist_add(knl, kn, 0);
5327 
5328 	return (0);
5329 }
5330 
5331 /*
5332  * Detach knote from vnode
5333  */
5334 static void
5335 filt_vfsdetach(struct knote *kn)
5336 {
5337 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5338 
5339 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5340 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5341 	vdrop(vp);
5342 }
5343 
5344 /*ARGSUSED*/
5345 static int
5346 filt_vfsread(struct knote *kn, long hint)
5347 {
5348 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5349 	struct vattr va;
5350 	int res;
5351 
5352 	/*
5353 	 * filesystem is gone, so set the EOF flag and schedule
5354 	 * the knote for deletion.
5355 	 */
5356 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5357 		VI_LOCK(vp);
5358 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5359 		VI_UNLOCK(vp);
5360 		return (1);
5361 	}
5362 
5363 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5364 		return (0);
5365 
5366 	VI_LOCK(vp);
5367 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5368 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5369 	VI_UNLOCK(vp);
5370 	return (res);
5371 }
5372 
5373 /*ARGSUSED*/
5374 static int
5375 filt_vfswrite(struct knote *kn, long hint)
5376 {
5377 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5378 
5379 	VI_LOCK(vp);
5380 
5381 	/*
5382 	 * filesystem is gone, so set the EOF flag and schedule
5383 	 * the knote for deletion.
5384 	 */
5385 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5386 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5387 
5388 	kn->kn_data = 0;
5389 	VI_UNLOCK(vp);
5390 	return (1);
5391 }
5392 
5393 static int
5394 filt_vfsvnode(struct knote *kn, long hint)
5395 {
5396 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5397 	int res;
5398 
5399 	VI_LOCK(vp);
5400 	if (kn->kn_sfflags & hint)
5401 		kn->kn_fflags |= hint;
5402 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5403 		kn->kn_flags |= EV_EOF;
5404 		VI_UNLOCK(vp);
5405 		return (1);
5406 	}
5407 	res = (kn->kn_fflags != 0);
5408 	VI_UNLOCK(vp);
5409 	return (res);
5410 }
5411 
5412 int
5413 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5414 {
5415 	int error;
5416 
5417 	if (dp->d_reclen > ap->a_uio->uio_resid)
5418 		return (ENAMETOOLONG);
5419 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5420 	if (error) {
5421 		if (ap->a_ncookies != NULL) {
5422 			if (ap->a_cookies != NULL)
5423 				free(ap->a_cookies, M_TEMP);
5424 			ap->a_cookies = NULL;
5425 			*ap->a_ncookies = 0;
5426 		}
5427 		return (error);
5428 	}
5429 	if (ap->a_ncookies == NULL)
5430 		return (0);
5431 
5432 	KASSERT(ap->a_cookies,
5433 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5434 
5435 	*ap->a_cookies = realloc(*ap->a_cookies,
5436 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5437 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5438 	*ap->a_ncookies += 1;
5439 	return (0);
5440 }
5441 
5442 /*
5443  * Mark for update the access time of the file if the filesystem
5444  * supports VOP_MARKATIME.  This functionality is used by execve and
5445  * mmap, so we want to avoid the I/O implied by directly setting
5446  * va_atime for the sake of efficiency.
5447  */
5448 void
5449 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5450 {
5451 	struct mount *mp;
5452 
5453 	mp = vp->v_mount;
5454 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5455 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5456 		(void)VOP_MARKATIME(vp);
5457 }
5458 
5459 /*
5460  * The purpose of this routine is to remove granularity from accmode_t,
5461  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5462  * VADMIN and VAPPEND.
5463  *
5464  * If it returns 0, the caller is supposed to continue with the usual
5465  * access checks using 'accmode' as modified by this routine.  If it
5466  * returns nonzero value, the caller is supposed to return that value
5467  * as errno.
5468  *
5469  * Note that after this routine runs, accmode may be zero.
5470  */
5471 int
5472 vfs_unixify_accmode(accmode_t *accmode)
5473 {
5474 	/*
5475 	 * There is no way to specify explicit "deny" rule using
5476 	 * file mode or POSIX.1e ACLs.
5477 	 */
5478 	if (*accmode & VEXPLICIT_DENY) {
5479 		*accmode = 0;
5480 		return (0);
5481 	}
5482 
5483 	/*
5484 	 * None of these can be translated into usual access bits.
5485 	 * Also, the common case for NFSv4 ACLs is to not contain
5486 	 * either of these bits. Caller should check for VWRITE
5487 	 * on the containing directory instead.
5488 	 */
5489 	if (*accmode & (VDELETE_CHILD | VDELETE))
5490 		return (EPERM);
5491 
5492 	if (*accmode & VADMIN_PERMS) {
5493 		*accmode &= ~VADMIN_PERMS;
5494 		*accmode |= VADMIN;
5495 	}
5496 
5497 	/*
5498 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5499 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5500 	 */
5501 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5502 
5503 	return (0);
5504 }
5505 
5506 /*
5507  * These are helper functions for filesystems to traverse all
5508  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5509  *
5510  * This interface replaces MNT_VNODE_FOREACH.
5511  */
5512 
5513 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5514 
5515 struct vnode *
5516 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5517 {
5518 	struct vnode *vp;
5519 
5520 	if (should_yield())
5521 		kern_yield(PRI_USER);
5522 	MNT_ILOCK(mp);
5523 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5524 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5525 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5526 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5527 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5528 			continue;
5529 		VI_LOCK(vp);
5530 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5531 			VI_UNLOCK(vp);
5532 			continue;
5533 		}
5534 		break;
5535 	}
5536 	if (vp == NULL) {
5537 		__mnt_vnode_markerfree_all(mvp, mp);
5538 		/* MNT_IUNLOCK(mp); -- done in above function */
5539 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5540 		return (NULL);
5541 	}
5542 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5543 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5544 	MNT_IUNLOCK(mp);
5545 	return (vp);
5546 }
5547 
5548 struct vnode *
5549 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5550 {
5551 	struct vnode *vp;
5552 
5553 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5554 	MNT_ILOCK(mp);
5555 	MNT_REF(mp);
5556 	(*mvp)->v_mount = mp;
5557 	(*mvp)->v_type = VMARKER;
5558 
5559 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5560 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5561 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5562 			continue;
5563 		VI_LOCK(vp);
5564 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5565 			VI_UNLOCK(vp);
5566 			continue;
5567 		}
5568 		break;
5569 	}
5570 	if (vp == NULL) {
5571 		MNT_REL(mp);
5572 		MNT_IUNLOCK(mp);
5573 		free(*mvp, M_VNODE_MARKER);
5574 		*mvp = NULL;
5575 		return (NULL);
5576 	}
5577 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5578 	MNT_IUNLOCK(mp);
5579 	return (vp);
5580 }
5581 
5582 void
5583 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5584 {
5585 
5586 	if (*mvp == NULL) {
5587 		MNT_IUNLOCK(mp);
5588 		return;
5589 	}
5590 
5591 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5592 
5593 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5594 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5595 	MNT_REL(mp);
5596 	MNT_IUNLOCK(mp);
5597 	free(*mvp, M_VNODE_MARKER);
5598 	*mvp = NULL;
5599 }
5600 
5601 /*
5602  * These are helper functions for filesystems to traverse their
5603  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5604  */
5605 static void
5606 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5607 {
5608 
5609 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5610 
5611 	MNT_ILOCK(mp);
5612 	MNT_REL(mp);
5613 	MNT_IUNLOCK(mp);
5614 	free(*mvp, M_VNODE_MARKER);
5615 	*mvp = NULL;
5616 }
5617 
5618 /*
5619  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5620  * conventional lock order during mnt_vnode_next_active iteration.
5621  *
5622  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5623  * The list lock is dropped and reacquired.  On success, both locks are held.
5624  * On failure, the mount vnode list lock is held but the vnode interlock is
5625  * not, and the procedure may have yielded.
5626  */
5627 static bool
5628 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5629     struct vnode *vp)
5630 {
5631 	const struct vnode *tmp;
5632 	bool held, ret;
5633 
5634 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5635 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5636 	    ("%s: bad marker", __func__));
5637 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5638 	    ("%s: inappropriate vnode", __func__));
5639 	ASSERT_VI_UNLOCKED(vp, __func__);
5640 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5641 
5642 	ret = false;
5643 
5644 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5645 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5646 
5647 	/*
5648 	 * Use a hold to prevent vp from disappearing while the mount vnode
5649 	 * list lock is dropped and reacquired.  Normally a hold would be
5650 	 * acquired with vhold(), but that might try to acquire the vnode
5651 	 * interlock, which would be a LOR with the mount vnode list lock.
5652 	 */
5653 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5654 	mtx_unlock(&mp->mnt_listmtx);
5655 	if (!held)
5656 		goto abort;
5657 	VI_LOCK(vp);
5658 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5659 		vdropl(vp);
5660 		goto abort;
5661 	}
5662 	mtx_lock(&mp->mnt_listmtx);
5663 
5664 	/*
5665 	 * Determine whether the vnode is still the next one after the marker,
5666 	 * excepting any other markers.  If the vnode has not been doomed by
5667 	 * vgone() then the hold should have ensured that it remained on the
5668 	 * active list.  If it has been doomed but is still on the active list,
5669 	 * don't abort, but rather skip over it (avoid spinning on doomed
5670 	 * vnodes).
5671 	 */
5672 	tmp = mvp;
5673 	do {
5674 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5675 	} while (tmp != NULL && tmp->v_type == VMARKER);
5676 	if (tmp != vp) {
5677 		mtx_unlock(&mp->mnt_listmtx);
5678 		VI_UNLOCK(vp);
5679 		goto abort;
5680 	}
5681 
5682 	ret = true;
5683 	goto out;
5684 abort:
5685 	maybe_yield();
5686 	mtx_lock(&mp->mnt_listmtx);
5687 out:
5688 	if (ret)
5689 		ASSERT_VI_LOCKED(vp, __func__);
5690 	else
5691 		ASSERT_VI_UNLOCKED(vp, __func__);
5692 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5693 	return (ret);
5694 }
5695 
5696 static struct vnode *
5697 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5698 {
5699 	struct vnode *vp, *nvp;
5700 
5701 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5702 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5703 restart:
5704 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5705 	while (vp != NULL) {
5706 		if (vp->v_type == VMARKER) {
5707 			vp = TAILQ_NEXT(vp, v_actfreelist);
5708 			continue;
5709 		}
5710 		/*
5711 		 * Try-lock because this is the wrong lock order.  If that does
5712 		 * not succeed, drop the mount vnode list lock and try to
5713 		 * reacquire it and the vnode interlock in the right order.
5714 		 */
5715 		if (!VI_TRYLOCK(vp) &&
5716 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5717 			goto restart;
5718 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5719 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5720 		    ("alien vnode on the active list %p %p", vp, mp));
5721 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5722 			break;
5723 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5724 		VI_UNLOCK(vp);
5725 		vp = nvp;
5726 	}
5727 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5728 
5729 	/* Check if we are done */
5730 	if (vp == NULL) {
5731 		mtx_unlock(&mp->mnt_listmtx);
5732 		mnt_vnode_markerfree_active(mvp, mp);
5733 		return (NULL);
5734 	}
5735 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5736 	mtx_unlock(&mp->mnt_listmtx);
5737 	ASSERT_VI_LOCKED(vp, "active iter");
5738 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5739 	return (vp);
5740 }
5741 
5742 struct vnode *
5743 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5744 {
5745 
5746 	if (should_yield())
5747 		kern_yield(PRI_USER);
5748 	mtx_lock(&mp->mnt_listmtx);
5749 	return (mnt_vnode_next_active(mvp, mp));
5750 }
5751 
5752 struct vnode *
5753 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5754 {
5755 	struct vnode *vp;
5756 
5757 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5758 	MNT_ILOCK(mp);
5759 	MNT_REF(mp);
5760 	MNT_IUNLOCK(mp);
5761 	(*mvp)->v_type = VMARKER;
5762 	(*mvp)->v_mount = mp;
5763 
5764 	mtx_lock(&mp->mnt_listmtx);
5765 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5766 	if (vp == NULL) {
5767 		mtx_unlock(&mp->mnt_listmtx);
5768 		mnt_vnode_markerfree_active(mvp, mp);
5769 		return (NULL);
5770 	}
5771 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5772 	return (mnt_vnode_next_active(mvp, mp));
5773 }
5774 
5775 void
5776 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5777 {
5778 
5779 	if (*mvp == NULL)
5780 		return;
5781 
5782 	mtx_lock(&mp->mnt_listmtx);
5783 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5784 	mtx_unlock(&mp->mnt_listmtx);
5785 	mnt_vnode_markerfree_active(mvp, mp);
5786 }
5787