xref: /freebsd/sys/kern/vfs_subr.c (revision 162ae9c834f6d9f9cb443bd62cceb23e0b5fef48)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/stat.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/uma.h>
99 
100 #ifdef DDB
101 #include <ddb/ddb.h>
102 #endif
103 
104 static void	delmntque(struct vnode *vp);
105 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 		    int slpflag, int slptimeo);
107 static void	syncer_shutdown(void *arg, int howto);
108 static int	vtryrecycle(struct vnode *vp);
109 static void	v_init_counters(struct vnode *);
110 static void	v_incr_devcount(struct vnode *);
111 static void	v_decr_devcount(struct vnode *);
112 static void	vgonel(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_locked(void *arg);
116 static void	vfs_knl_assert_unlocked(void *arg);
117 static void	destroy_vpollinfo(struct vpollinfo *vi);
118 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
119 		    daddr_t startlbn, daddr_t endlbn);
120 static void	vnlru_recalc(void);
121 
122 /*
123  * These fences are intended for cases where some synchronization is
124  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
125  * and v_usecount) updates.  Access to v_iflags is generally synchronized
126  * by the interlock, but we have some internal assertions that check vnode
127  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
128  * for now.
129  */
130 #ifdef INVARIANTS
131 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
132 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
133 #else
134 #define	VNODE_REFCOUNT_FENCE_ACQ()
135 #define	VNODE_REFCOUNT_FENCE_REL()
136 #endif
137 
138 /*
139  * Number of vnodes in existence.  Increased whenever getnewvnode()
140  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141  */
142 static u_long __exclusive_cache_line numvnodes;
143 
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145     "Number of vnodes in existence");
146 
147 static counter_u64_t vnodes_created;
148 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
149     "Number of vnodes created by getnewvnode");
150 
151 /*
152  * Conversion tables for conversion from vnode types to inode formats
153  * and back.
154  */
155 enum vtype iftovt_tab[16] = {
156 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
157 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
158 };
159 int vttoif_tab[10] = {
160 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
161 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
162 };
163 
164 /*
165  * List of allocates vnodes in the system.
166  */
167 static TAILQ_HEAD(freelst, vnode) vnode_list;
168 static struct vnode *vnode_list_free_marker;
169 static struct vnode *vnode_list_reclaim_marker;
170 
171 /*
172  * "Free" vnode target.  Free vnodes are rarely completely free, but are
173  * just ones that are cheap to recycle.  Usually they are for files which
174  * have been stat'd but not read; these usually have inode and namecache
175  * data attached to them.  This target is the preferred minimum size of a
176  * sub-cache consisting mostly of such files. The system balances the size
177  * of this sub-cache with its complement to try to prevent either from
178  * thrashing while the other is relatively inactive.  The targets express
179  * a preference for the best balance.
180  *
181  * "Above" this target there are 2 further targets (watermarks) related
182  * to recyling of free vnodes.  In the best-operating case, the cache is
183  * exactly full, the free list has size between vlowat and vhiwat above the
184  * free target, and recycling from it and normal use maintains this state.
185  * Sometimes the free list is below vlowat or even empty, but this state
186  * is even better for immediate use provided the cache is not full.
187  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
188  * ones) to reach one of these states.  The watermarks are currently hard-
189  * coded as 4% and 9% of the available space higher.  These and the default
190  * of 25% for wantfreevnodes are too large if the memory size is large.
191  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
192  * whenever vnlru_proc() becomes active.
193  */
194 static long wantfreevnodes;
195 static long __exclusive_cache_line freevnodes;
196 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
197     &freevnodes, 0, "Number of \"free\" vnodes");
198 static long freevnodes_old;
199 
200 static counter_u64_t recycles_count;
201 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
202     "Number of vnodes recycled to meet vnode cache targets");
203 
204 static counter_u64_t recycles_free_count;
205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
206     "Number of free vnodes recycled to meet vnode cache targets");
207 
208 /*
209  * Various variables used for debugging the new implementation of
210  * reassignbuf().
211  * XXX these are probably of (very) limited utility now.
212  */
213 static int reassignbufcalls;
214 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
215     &reassignbufcalls, 0, "Number of calls to reassignbuf");
216 
217 static counter_u64_t deferred_inact;
218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
219     "Number of times inactive processing was deferred");
220 
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
223 
224 /*
225  * Lock for any access to the following:
226  *	vnode_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx __exclusive_cache_line vnode_list_mtx;
231 
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
234 
235 static uma_zone_t buf_trie_zone;
236 
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
240 
241 /*
242  * The workitem queue.
243  *
244  * It is useful to delay writes of file data and filesystem metadata
245  * for tens of seconds so that quickly created and deleted files need
246  * not waste disk bandwidth being created and removed. To realize this,
247  * we append vnodes to a "workitem" queue. When running with a soft
248  * updates implementation, most pending metadata dependencies should
249  * not wait for more than a few seconds. Thus, mounted on block devices
250  * are delayed only about a half the time that file data is delayed.
251  * Similarly, directory updates are more critical, so are only delayed
252  * about a third the time that file data is delayed. Thus, there are
253  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254  * one each second (driven off the filesystem syncer process). The
255  * syncer_delayno variable indicates the next queue that is to be processed.
256  * Items that need to be processed soon are placed in this queue:
257  *
258  *	syncer_workitem_pending[syncer_delayno]
259  *
260  * A delay of fifteen seconds is done by placing the request fifteen
261  * entries later in the queue:
262  *
263  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264  *
265  */
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
270 /*
271  * The sync_mtx protects:
272  *	bo->bo_synclist
273  *	sync_vnode_count
274  *	syncer_delayno
275  *	syncer_state
276  *	syncer_workitem_pending
277  *	syncer_worklist_len
278  *	rushjob
279  */
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
282 
283 #define SYNCER_MAXDELAY		32
284 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285 static int syncdelay = 30;		/* max time to delay syncing data */
286 static int filedelay = 30;		/* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288     "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29;		/* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291     "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28;		/* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294     "Time to delay syncing metadata (in seconds)");
295 static int rushjob;		/* number of slots to run ASAP */
296 static int stat_rush_requests;	/* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298     "Number of times I/O speeded up (rush requests)");
299 
300 #define	VDBATCH_SIZE 8
301 struct vdbatch {
302 	u_int index;
303 	long freevnodes;
304 	struct mtx lock;
305 	struct vnode *tab[VDBATCH_SIZE];
306 };
307 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
308 
309 static void	vdbatch_dequeue(struct vnode *vp);
310 
311 /*
312  * When shutting down the syncer, run it at four times normal speed.
313  */
314 #define SYNCER_SHUTDOWN_SPEEDUP		4
315 static int sync_vnode_count;
316 static int syncer_worklist_len;
317 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
318     syncer_state;
319 
320 /* Target for maximum number of vnodes. */
321 u_long desiredvnodes;
322 static u_long gapvnodes;		/* gap between wanted and desired */
323 static u_long vhiwat;		/* enough extras after expansion */
324 static u_long vlowat;		/* minimal extras before expansion */
325 static u_long vstir;		/* nonzero to stir non-free vnodes */
326 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
327 
328 static u_long vnlru_read_freevnodes(void);
329 
330 /*
331  * Note that no attempt is made to sanitize these parameters.
332  */
333 static int
334 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
335 {
336 	u_long val;
337 	int error;
338 
339 	val = desiredvnodes;
340 	error = sysctl_handle_long(oidp, &val, 0, req);
341 	if (error != 0 || req->newptr == NULL)
342 		return (error);
343 
344 	if (val == desiredvnodes)
345 		return (0);
346 	mtx_lock(&vnode_list_mtx);
347 	desiredvnodes = val;
348 	wantfreevnodes = desiredvnodes / 4;
349 	vnlru_recalc();
350 	mtx_unlock(&vnode_list_mtx);
351 	/*
352 	 * XXX There is no protection against multiple threads changing
353 	 * desiredvnodes at the same time. Locking above only helps vnlru and
354 	 * getnewvnode.
355 	 */
356 	vfs_hash_changesize(desiredvnodes);
357 	cache_changesize(desiredvnodes);
358 	return (0);
359 }
360 
361 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
362     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
363     "UL", "Target for maximum number of vnodes");
364 
365 static int
366 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
367 {
368 	u_long val;
369 	int error;
370 
371 	val = wantfreevnodes;
372 	error = sysctl_handle_long(oidp, &val, 0, req);
373 	if (error != 0 || req->newptr == NULL)
374 		return (error);
375 
376 	if (val == wantfreevnodes)
377 		return (0);
378 	mtx_lock(&vnode_list_mtx);
379 	wantfreevnodes = val;
380 	vnlru_recalc();
381 	mtx_unlock(&vnode_list_mtx);
382 	return (0);
383 }
384 
385 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
386     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
387     "UL", "Target for minimum number of \"free\" vnodes");
388 
389 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
390     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
391 static int vnlru_nowhere;
392 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
393     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
394 
395 static int
396 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
397 {
398 	struct vnode *vp;
399 	struct nameidata nd;
400 	char *buf;
401 	unsigned long ndflags;
402 	int error;
403 
404 	if (req->newptr == NULL)
405 		return (EINVAL);
406 	if (req->newlen >= PATH_MAX)
407 		return (E2BIG);
408 
409 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
410 	error = SYSCTL_IN(req, buf, req->newlen);
411 	if (error != 0)
412 		goto out;
413 
414 	buf[req->newlen] = '\0';
415 
416 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
417 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
418 	if ((error = namei(&nd)) != 0)
419 		goto out;
420 	vp = nd.ni_vp;
421 
422 	if (VN_IS_DOOMED(vp)) {
423 		/*
424 		 * This vnode is being recycled.  Return != 0 to let the caller
425 		 * know that the sysctl had no effect.  Return EAGAIN because a
426 		 * subsequent call will likely succeed (since namei will create
427 		 * a new vnode if necessary)
428 		 */
429 		error = EAGAIN;
430 		goto putvnode;
431 	}
432 
433 	counter_u64_add(recycles_count, 1);
434 	vgone(vp);
435 putvnode:
436 	NDFREE(&nd, 0);
437 out:
438 	free(buf, M_TEMP);
439 	return (error);
440 }
441 
442 static int
443 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
444 {
445 	struct thread *td = curthread;
446 	struct vnode *vp;
447 	struct file *fp;
448 	int error;
449 	int fd;
450 
451 	if (req->newptr == NULL)
452 		return (EBADF);
453 
454         error = sysctl_handle_int(oidp, &fd, 0, req);
455         if (error != 0)
456                 return (error);
457 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
458 	if (error != 0)
459 		return (error);
460 	vp = fp->f_vnode;
461 
462 	error = vn_lock(vp, LK_EXCLUSIVE);
463 	if (error != 0)
464 		goto drop;
465 
466 	counter_u64_add(recycles_count, 1);
467 	vgone(vp);
468 	VOP_UNLOCK(vp);
469 drop:
470 	fdrop(fp, td);
471 	return (error);
472 }
473 
474 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
475     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
476     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
477 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
478     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
479     sysctl_ftry_reclaim_vnode, "I",
480     "Try to reclaim a vnode by its file descriptor");
481 
482 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
483 static int vnsz2log;
484 
485 /*
486  * Support for the bufobj clean & dirty pctrie.
487  */
488 static void *
489 buf_trie_alloc(struct pctrie *ptree)
490 {
491 
492 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
493 }
494 
495 static void
496 buf_trie_free(struct pctrie *ptree, void *node)
497 {
498 
499 	uma_zfree(buf_trie_zone, node);
500 }
501 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
502 
503 /*
504  * Initialize the vnode management data structures.
505  *
506  * Reevaluate the following cap on the number of vnodes after the physical
507  * memory size exceeds 512GB.  In the limit, as the physical memory size
508  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
509  */
510 #ifndef	MAXVNODES_MAX
511 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
512 #endif
513 
514 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
515 
516 static struct vnode *
517 vn_alloc_marker(struct mount *mp)
518 {
519 	struct vnode *vp;
520 
521 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
522 	vp->v_type = VMARKER;
523 	vp->v_mount = mp;
524 
525 	return (vp);
526 }
527 
528 static void
529 vn_free_marker(struct vnode *vp)
530 {
531 
532 	MPASS(vp->v_type == VMARKER);
533 	free(vp, M_VNODE_MARKER);
534 }
535 
536 /*
537  * Initialize a vnode as it first enters the zone.
538  */
539 static int
540 vnode_init(void *mem, int size, int flags)
541 {
542 	struct vnode *vp;
543 
544 	vp = mem;
545 	bzero(vp, size);
546 	/*
547 	 * Setup locks.
548 	 */
549 	vp->v_vnlock = &vp->v_lock;
550 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
551 	/*
552 	 * By default, don't allow shared locks unless filesystems opt-in.
553 	 */
554 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
555 	    LK_NOSHARE | LK_IS_VNODE);
556 	/*
557 	 * Initialize bufobj.
558 	 */
559 	bufobj_init(&vp->v_bufobj, vp);
560 	/*
561 	 * Initialize namecache.
562 	 */
563 	LIST_INIT(&vp->v_cache_src);
564 	TAILQ_INIT(&vp->v_cache_dst);
565 	/*
566 	 * Initialize rangelocks.
567 	 */
568 	rangelock_init(&vp->v_rl);
569 
570 	vp->v_dbatchcpu = NOCPU;
571 
572 	mtx_lock(&vnode_list_mtx);
573 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
574 	mtx_unlock(&vnode_list_mtx);
575 	return (0);
576 }
577 
578 /*
579  * Free a vnode when it is cleared from the zone.
580  */
581 static void
582 vnode_fini(void *mem, int size)
583 {
584 	struct vnode *vp;
585 	struct bufobj *bo;
586 
587 	vp = mem;
588 	vdbatch_dequeue(vp);
589 	mtx_lock(&vnode_list_mtx);
590 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
591 	mtx_unlock(&vnode_list_mtx);
592 	rangelock_destroy(&vp->v_rl);
593 	lockdestroy(vp->v_vnlock);
594 	mtx_destroy(&vp->v_interlock);
595 	bo = &vp->v_bufobj;
596 	rw_destroy(BO_LOCKPTR(bo));
597 }
598 
599 /*
600  * Provide the size of NFS nclnode and NFS fh for calculation of the
601  * vnode memory consumption.  The size is specified directly to
602  * eliminate dependency on NFS-private header.
603  *
604  * Other filesystems may use bigger or smaller (like UFS and ZFS)
605  * private inode data, but the NFS-based estimation is ample enough.
606  * Still, we care about differences in the size between 64- and 32-bit
607  * platforms.
608  *
609  * Namecache structure size is heuristically
610  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
611  */
612 #ifdef _LP64
613 #define	NFS_NCLNODE_SZ	(528 + 64)
614 #define	NC_SZ		148
615 #else
616 #define	NFS_NCLNODE_SZ	(360 + 32)
617 #define	NC_SZ		92
618 #endif
619 
620 static void
621 vntblinit(void *dummy __unused)
622 {
623 	struct vdbatch *vd;
624 	int cpu, physvnodes, virtvnodes;
625 	u_int i;
626 
627 	/*
628 	 * Desiredvnodes is a function of the physical memory size and the
629 	 * kernel's heap size.  Generally speaking, it scales with the
630 	 * physical memory size.  The ratio of desiredvnodes to the physical
631 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
632 	 * Thereafter, the
633 	 * marginal ratio of desiredvnodes to the physical memory size is
634 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
635 	 * size.  The memory required by desiredvnodes vnodes and vm objects
636 	 * must not exceed 1/10th of the kernel's heap size.
637 	 */
638 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
639 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
640 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
641 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
642 	desiredvnodes = min(physvnodes, virtvnodes);
643 	if (desiredvnodes > MAXVNODES_MAX) {
644 		if (bootverbose)
645 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
646 			    desiredvnodes, MAXVNODES_MAX);
647 		desiredvnodes = MAXVNODES_MAX;
648 	}
649 	wantfreevnodes = desiredvnodes / 4;
650 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
651 	TAILQ_INIT(&vnode_list);
652 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
653 	/*
654 	 * The lock is taken to appease WITNESS.
655 	 */
656 	mtx_lock(&vnode_list_mtx);
657 	vnlru_recalc();
658 	mtx_unlock(&vnode_list_mtx);
659 	vnode_list_free_marker = vn_alloc_marker(NULL);
660 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
661 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
662 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
663 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
664 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
665 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
666 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
667 	/*
668 	 * Preallocate enough nodes to support one-per buf so that
669 	 * we can not fail an insert.  reassignbuf() callers can not
670 	 * tolerate the insertion failure.
671 	 */
672 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
673 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
674 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
675 	uma_prealloc(buf_trie_zone, nbuf);
676 
677 	vnodes_created = counter_u64_alloc(M_WAITOK);
678 	recycles_count = counter_u64_alloc(M_WAITOK);
679 	recycles_free_count = counter_u64_alloc(M_WAITOK);
680 	deferred_inact = counter_u64_alloc(M_WAITOK);
681 
682 	/*
683 	 * Initialize the filesystem syncer.
684 	 */
685 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
686 	    &syncer_mask);
687 	syncer_maxdelay = syncer_mask + 1;
688 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
689 	cv_init(&sync_wakeup, "syncer");
690 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
691 		vnsz2log++;
692 	vnsz2log--;
693 
694 	CPU_FOREACH(cpu) {
695 		vd = DPCPU_ID_PTR((cpu), vd);
696 		bzero(vd, sizeof(*vd));
697 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
698 	}
699 }
700 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
701 
702 
703 /*
704  * Mark a mount point as busy. Used to synchronize access and to delay
705  * unmounting. Eventually, mountlist_mtx is not released on failure.
706  *
707  * vfs_busy() is a custom lock, it can block the caller.
708  * vfs_busy() only sleeps if the unmount is active on the mount point.
709  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
710  * vnode belonging to mp.
711  *
712  * Lookup uses vfs_busy() to traverse mount points.
713  * root fs			var fs
714  * / vnode lock		A	/ vnode lock (/var)		D
715  * /var vnode lock	B	/log vnode lock(/var/log)	E
716  * vfs_busy lock	C	vfs_busy lock			F
717  *
718  * Within each file system, the lock order is C->A->B and F->D->E.
719  *
720  * When traversing across mounts, the system follows that lock order:
721  *
722  *        C->A->B
723  *              |
724  *              +->F->D->E
725  *
726  * The lookup() process for namei("/var") illustrates the process:
727  *  VOP_LOOKUP() obtains B while A is held
728  *  vfs_busy() obtains a shared lock on F while A and B are held
729  *  vput() releases lock on B
730  *  vput() releases lock on A
731  *  VFS_ROOT() obtains lock on D while shared lock on F is held
732  *  vfs_unbusy() releases shared lock on F
733  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
734  *    Attempt to lock A (instead of vp_crossmp) while D is held would
735  *    violate the global order, causing deadlocks.
736  *
737  * dounmount() locks B while F is drained.
738  */
739 int
740 vfs_busy(struct mount *mp, int flags)
741 {
742 
743 	MPASS((flags & ~MBF_MASK) == 0);
744 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
745 
746 	if (vfs_op_thread_enter(mp)) {
747 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
748 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
749 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
750 		vfs_mp_count_add_pcpu(mp, ref, 1);
751 		vfs_mp_count_add_pcpu(mp, lockref, 1);
752 		vfs_op_thread_exit(mp);
753 		if (flags & MBF_MNTLSTLOCK)
754 			mtx_unlock(&mountlist_mtx);
755 		return (0);
756 	}
757 
758 	MNT_ILOCK(mp);
759 	vfs_assert_mount_counters(mp);
760 	MNT_REF(mp);
761 	/*
762 	 * If mount point is currently being unmounted, sleep until the
763 	 * mount point fate is decided.  If thread doing the unmounting fails,
764 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
765 	 * that this mount point has survived the unmount attempt and vfs_busy
766 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
767 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
768 	 * about to be really destroyed.  vfs_busy needs to release its
769 	 * reference on the mount point in this case and return with ENOENT,
770 	 * telling the caller that mount mount it tried to busy is no longer
771 	 * valid.
772 	 */
773 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
774 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
775 			MNT_REL(mp);
776 			MNT_IUNLOCK(mp);
777 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
778 			    __func__);
779 			return (ENOENT);
780 		}
781 		if (flags & MBF_MNTLSTLOCK)
782 			mtx_unlock(&mountlist_mtx);
783 		mp->mnt_kern_flag |= MNTK_MWAIT;
784 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
785 		if (flags & MBF_MNTLSTLOCK)
786 			mtx_lock(&mountlist_mtx);
787 		MNT_ILOCK(mp);
788 	}
789 	if (flags & MBF_MNTLSTLOCK)
790 		mtx_unlock(&mountlist_mtx);
791 	mp->mnt_lockref++;
792 	MNT_IUNLOCK(mp);
793 	return (0);
794 }
795 
796 /*
797  * Free a busy filesystem.
798  */
799 void
800 vfs_unbusy(struct mount *mp)
801 {
802 	int c;
803 
804 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
805 
806 	if (vfs_op_thread_enter(mp)) {
807 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
808 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
809 		vfs_mp_count_sub_pcpu(mp, ref, 1);
810 		vfs_op_thread_exit(mp);
811 		return;
812 	}
813 
814 	MNT_ILOCK(mp);
815 	vfs_assert_mount_counters(mp);
816 	MNT_REL(mp);
817 	c = --mp->mnt_lockref;
818 	if (mp->mnt_vfs_ops == 0) {
819 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
820 		MNT_IUNLOCK(mp);
821 		return;
822 	}
823 	if (c < 0)
824 		vfs_dump_mount_counters(mp);
825 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
826 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
827 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
828 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
829 		wakeup(&mp->mnt_lockref);
830 	}
831 	MNT_IUNLOCK(mp);
832 }
833 
834 /*
835  * Lookup a mount point by filesystem identifier.
836  */
837 struct mount *
838 vfs_getvfs(fsid_t *fsid)
839 {
840 	struct mount *mp;
841 
842 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
843 	mtx_lock(&mountlist_mtx);
844 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
845 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
846 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
847 			vfs_ref(mp);
848 			mtx_unlock(&mountlist_mtx);
849 			return (mp);
850 		}
851 	}
852 	mtx_unlock(&mountlist_mtx);
853 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
854 	return ((struct mount *) 0);
855 }
856 
857 /*
858  * Lookup a mount point by filesystem identifier, busying it before
859  * returning.
860  *
861  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
862  * cache for popular filesystem identifiers.  The cache is lockess, using
863  * the fact that struct mount's are never freed.  In worst case we may
864  * get pointer to unmounted or even different filesystem, so we have to
865  * check what we got, and go slow way if so.
866  */
867 struct mount *
868 vfs_busyfs(fsid_t *fsid)
869 {
870 #define	FSID_CACHE_SIZE	256
871 	typedef struct mount * volatile vmp_t;
872 	static vmp_t cache[FSID_CACHE_SIZE];
873 	struct mount *mp;
874 	int error;
875 	uint32_t hash;
876 
877 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
878 	hash = fsid->val[0] ^ fsid->val[1];
879 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
880 	mp = cache[hash];
881 	if (mp == NULL ||
882 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
883 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
884 		goto slow;
885 	if (vfs_busy(mp, 0) != 0) {
886 		cache[hash] = NULL;
887 		goto slow;
888 	}
889 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
890 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
891 		return (mp);
892 	else
893 	    vfs_unbusy(mp);
894 
895 slow:
896 	mtx_lock(&mountlist_mtx);
897 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
898 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
899 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
900 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
901 			if (error) {
902 				cache[hash] = NULL;
903 				mtx_unlock(&mountlist_mtx);
904 				return (NULL);
905 			}
906 			cache[hash] = mp;
907 			return (mp);
908 		}
909 	}
910 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
911 	mtx_unlock(&mountlist_mtx);
912 	return ((struct mount *) 0);
913 }
914 
915 /*
916  * Check if a user can access privileged mount options.
917  */
918 int
919 vfs_suser(struct mount *mp, struct thread *td)
920 {
921 	int error;
922 
923 	if (jailed(td->td_ucred)) {
924 		/*
925 		 * If the jail of the calling thread lacks permission for
926 		 * this type of file system, deny immediately.
927 		 */
928 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
929 			return (EPERM);
930 
931 		/*
932 		 * If the file system was mounted outside the jail of the
933 		 * calling thread, deny immediately.
934 		 */
935 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
936 			return (EPERM);
937 	}
938 
939 	/*
940 	 * If file system supports delegated administration, we don't check
941 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
942 	 * by the file system itself.
943 	 * If this is not the user that did original mount, we check for
944 	 * the PRIV_VFS_MOUNT_OWNER privilege.
945 	 */
946 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
947 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
948 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
949 			return (error);
950 	}
951 	return (0);
952 }
953 
954 /*
955  * Get a new unique fsid.  Try to make its val[0] unique, since this value
956  * will be used to create fake device numbers for stat().  Also try (but
957  * not so hard) make its val[0] unique mod 2^16, since some emulators only
958  * support 16-bit device numbers.  We end up with unique val[0]'s for the
959  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
960  *
961  * Keep in mind that several mounts may be running in parallel.  Starting
962  * the search one past where the previous search terminated is both a
963  * micro-optimization and a defense against returning the same fsid to
964  * different mounts.
965  */
966 void
967 vfs_getnewfsid(struct mount *mp)
968 {
969 	static uint16_t mntid_base;
970 	struct mount *nmp;
971 	fsid_t tfsid;
972 	int mtype;
973 
974 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
975 	mtx_lock(&mntid_mtx);
976 	mtype = mp->mnt_vfc->vfc_typenum;
977 	tfsid.val[1] = mtype;
978 	mtype = (mtype & 0xFF) << 24;
979 	for (;;) {
980 		tfsid.val[0] = makedev(255,
981 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
982 		mntid_base++;
983 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
984 			break;
985 		vfs_rel(nmp);
986 	}
987 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
988 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
989 	mtx_unlock(&mntid_mtx);
990 }
991 
992 /*
993  * Knob to control the precision of file timestamps:
994  *
995  *   0 = seconds only; nanoseconds zeroed.
996  *   1 = seconds and nanoseconds, accurate within 1/HZ.
997  *   2 = seconds and nanoseconds, truncated to microseconds.
998  * >=3 = seconds and nanoseconds, maximum precision.
999  */
1000 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1001 
1002 static int timestamp_precision = TSP_USEC;
1003 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1004     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1005     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1006     "3+: sec + ns (max. precision))");
1007 
1008 /*
1009  * Get a current timestamp.
1010  */
1011 void
1012 vfs_timestamp(struct timespec *tsp)
1013 {
1014 	struct timeval tv;
1015 
1016 	switch (timestamp_precision) {
1017 	case TSP_SEC:
1018 		tsp->tv_sec = time_second;
1019 		tsp->tv_nsec = 0;
1020 		break;
1021 	case TSP_HZ:
1022 		getnanotime(tsp);
1023 		break;
1024 	case TSP_USEC:
1025 		microtime(&tv);
1026 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1027 		break;
1028 	case TSP_NSEC:
1029 	default:
1030 		nanotime(tsp);
1031 		break;
1032 	}
1033 }
1034 
1035 /*
1036  * Set vnode attributes to VNOVAL
1037  */
1038 void
1039 vattr_null(struct vattr *vap)
1040 {
1041 
1042 	vap->va_type = VNON;
1043 	vap->va_size = VNOVAL;
1044 	vap->va_bytes = VNOVAL;
1045 	vap->va_mode = VNOVAL;
1046 	vap->va_nlink = VNOVAL;
1047 	vap->va_uid = VNOVAL;
1048 	vap->va_gid = VNOVAL;
1049 	vap->va_fsid = VNOVAL;
1050 	vap->va_fileid = VNOVAL;
1051 	vap->va_blocksize = VNOVAL;
1052 	vap->va_rdev = VNOVAL;
1053 	vap->va_atime.tv_sec = VNOVAL;
1054 	vap->va_atime.tv_nsec = VNOVAL;
1055 	vap->va_mtime.tv_sec = VNOVAL;
1056 	vap->va_mtime.tv_nsec = VNOVAL;
1057 	vap->va_ctime.tv_sec = VNOVAL;
1058 	vap->va_ctime.tv_nsec = VNOVAL;
1059 	vap->va_birthtime.tv_sec = VNOVAL;
1060 	vap->va_birthtime.tv_nsec = VNOVAL;
1061 	vap->va_flags = VNOVAL;
1062 	vap->va_gen = VNOVAL;
1063 	vap->va_vaflags = 0;
1064 }
1065 
1066 /*
1067  * Try to reduce the total number of vnodes.
1068  *
1069  * This routine (and its user) are buggy in at least the following ways:
1070  * - all parameters were picked years ago when RAM sizes were significantly
1071  *   smaller
1072  * - it can pick vnodes based on pages used by the vm object, but filesystems
1073  *   like ZFS don't use it making the pick broken
1074  * - since ZFS has its own aging policy it gets partially combated by this one
1075  * - a dedicated method should be provided for filesystems to let them decide
1076  *   whether the vnode should be recycled
1077  *
1078  * This routine is called when we have too many vnodes.  It attempts
1079  * to free <count> vnodes and will potentially free vnodes that still
1080  * have VM backing store (VM backing store is typically the cause
1081  * of a vnode blowout so we want to do this).  Therefore, this operation
1082  * is not considered cheap.
1083  *
1084  * A number of conditions may prevent a vnode from being reclaimed.
1085  * the buffer cache may have references on the vnode, a directory
1086  * vnode may still have references due to the namei cache representing
1087  * underlying files, or the vnode may be in active use.   It is not
1088  * desirable to reuse such vnodes.  These conditions may cause the
1089  * number of vnodes to reach some minimum value regardless of what
1090  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1091  *
1092  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1093  * 			 entries if this argument is strue
1094  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1095  *			 pages.
1096  * @param target	 How many vnodes to reclaim.
1097  * @return		 The number of vnodes that were reclaimed.
1098  */
1099 static int
1100 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1101 {
1102 	struct vnode *vp, *mvp;
1103 	struct mount *mp;
1104 	u_long done;
1105 	bool retried;
1106 
1107 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1108 
1109 	retried = false;
1110 	done = 0;
1111 
1112 	mvp = vnode_list_reclaim_marker;
1113 restart:
1114 	vp = mvp;
1115 	while (done < target) {
1116 		vp = TAILQ_NEXT(vp, v_vnodelist);
1117 		if (__predict_false(vp == NULL))
1118 			break;
1119 
1120 		if (__predict_false(vp->v_type == VMARKER))
1121 			continue;
1122 
1123 		/*
1124 		 * If it's been deconstructed already, it's still
1125 		 * referenced, or it exceeds the trigger, skip it.
1126 		 * Also skip free vnodes.  We are trying to make space
1127 		 * to expand the free list, not reduce it.
1128 		 */
1129 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1130 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1131 			goto next_iter;
1132 
1133 		if (vp->v_type == VBAD || vp->v_type == VNON)
1134 			goto next_iter;
1135 
1136 		if (!VI_TRYLOCK(vp))
1137 			goto next_iter;
1138 
1139 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1140 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1141 		    vp->v_type == VBAD || vp->v_type == VNON ||
1142 		    (vp->v_object != NULL &&
1143 		    vp->v_object->resident_page_count > trigger)) {
1144 			VI_UNLOCK(vp);
1145 			goto next_iter;
1146 		}
1147 		vholdl(vp);
1148 		VI_UNLOCK(vp);
1149 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1150 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1151 		mtx_unlock(&vnode_list_mtx);
1152 
1153 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1154 			vdrop(vp);
1155 			goto next_iter_unlocked;
1156 		}
1157 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1158 			vdrop(vp);
1159 			vn_finished_write(mp);
1160 			goto next_iter_unlocked;
1161 		}
1162 
1163 		VI_LOCK(vp);
1164 		if (vp->v_usecount > 0 ||
1165 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1166 		    (vp->v_object != NULL &&
1167 		    vp->v_object->resident_page_count > trigger)) {
1168 			VOP_UNLOCK(vp);
1169 			vdropl(vp);
1170 			vn_finished_write(mp);
1171 			goto next_iter_unlocked;
1172 		}
1173 		counter_u64_add(recycles_count, 1);
1174 		vgonel(vp);
1175 		VOP_UNLOCK(vp);
1176 		vdropl(vp);
1177 		vn_finished_write(mp);
1178 		done++;
1179 next_iter_unlocked:
1180 		if (should_yield())
1181 			kern_yield(PRI_USER);
1182 		mtx_lock(&vnode_list_mtx);
1183 		goto restart;
1184 next_iter:
1185 		MPASS(vp->v_type != VMARKER);
1186 		if (!should_yield())
1187 			continue;
1188 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1189 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1190 		mtx_unlock(&vnode_list_mtx);
1191 		kern_yield(PRI_USER);
1192 		mtx_lock(&vnode_list_mtx);
1193 		goto restart;
1194 	}
1195 	if (done == 0 && !retried) {
1196 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1197 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1198 		retried = true;
1199 		goto restart;
1200 	}
1201 	return (done);
1202 }
1203 
1204 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1205 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1206     0,
1207     "limit on vnode free requests per call to the vnlru_free routine");
1208 
1209 /*
1210  * Attempt to reduce the free list by the requested amount.
1211  */
1212 static int
1213 vnlru_free_locked(int count, struct vfsops *mnt_op)
1214 {
1215 	struct vnode *vp, *mvp;
1216 	struct mount *mp;
1217 	int ocount;
1218 
1219 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1220 	if (count > max_vnlru_free)
1221 		count = max_vnlru_free;
1222 	ocount = count;
1223 	mvp = vnode_list_free_marker;
1224 restart:
1225 	vp = mvp;
1226 	while (count > 0) {
1227 		vp = TAILQ_NEXT(vp, v_vnodelist);
1228 		if (__predict_false(vp == NULL)) {
1229 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1230 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1231 			break;
1232 		}
1233 		if (__predict_false(vp->v_type == VMARKER))
1234 			continue;
1235 
1236 		/*
1237 		 * Don't recycle if our vnode is from different type
1238 		 * of mount point.  Note that mp is type-safe, the
1239 		 * check does not reach unmapped address even if
1240 		 * vnode is reclaimed.
1241 		 * Don't recycle if we can't get the interlock without
1242 		 * blocking.
1243 		 */
1244 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1245 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1246 			continue;
1247 		}
1248 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1249 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1250 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1251 			VI_UNLOCK(vp);
1252 			continue;
1253 		}
1254 		vholdl(vp);
1255 		count--;
1256 		mtx_unlock(&vnode_list_mtx);
1257 		VI_UNLOCK(vp);
1258 		vtryrecycle(vp);
1259 		vdrop(vp);
1260 		mtx_lock(&vnode_list_mtx);
1261 		goto restart;
1262 	}
1263 	return (ocount - count);
1264 }
1265 
1266 void
1267 vnlru_free(int count, struct vfsops *mnt_op)
1268 {
1269 
1270 	mtx_lock(&vnode_list_mtx);
1271 	vnlru_free_locked(count, mnt_op);
1272 	mtx_unlock(&vnode_list_mtx);
1273 }
1274 
1275 static void
1276 vnlru_recalc(void)
1277 {
1278 
1279 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1280 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1281 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1282 	vlowat = vhiwat / 2;
1283 }
1284 
1285 /*
1286  * Attempt to recycle vnodes in a context that is always safe to block.
1287  * Calling vlrurecycle() from the bowels of filesystem code has some
1288  * interesting deadlock problems.
1289  */
1290 static struct proc *vnlruproc;
1291 static int vnlruproc_sig;
1292 
1293 /*
1294  * The main freevnodes counter is only updated when threads requeue their vnode
1295  * batches. CPUs are conditionally walked to compute a more accurate total.
1296  *
1297  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1298  * at any given moment can still exceed slop, but it should not be by significant
1299  * margin in practice.
1300  */
1301 #define VNLRU_FREEVNODES_SLOP 128
1302 
1303 static u_long
1304 vnlru_read_freevnodes(void)
1305 {
1306 	struct vdbatch *vd;
1307 	long slop;
1308 	int cpu;
1309 
1310 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1311 	if (freevnodes > freevnodes_old)
1312 		slop = freevnodes - freevnodes_old;
1313 	else
1314 		slop = freevnodes_old - freevnodes;
1315 	if (slop < VNLRU_FREEVNODES_SLOP)
1316 		return (freevnodes >= 0 ? freevnodes : 0);
1317 	freevnodes_old = freevnodes;
1318 	CPU_FOREACH(cpu) {
1319 		vd = DPCPU_ID_PTR((cpu), vd);
1320 		freevnodes_old += vd->freevnodes;
1321 	}
1322 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1323 }
1324 
1325 static bool
1326 vnlru_under(u_long rnumvnodes, u_long limit)
1327 {
1328 	u_long rfreevnodes, space;
1329 
1330 	if (__predict_false(rnumvnodes > desiredvnodes))
1331 		return (true);
1332 
1333 	space = desiredvnodes - rnumvnodes;
1334 	if (space < limit) {
1335 		rfreevnodes = vnlru_read_freevnodes();
1336 		if (rfreevnodes > wantfreevnodes)
1337 			space += rfreevnodes - wantfreevnodes;
1338 	}
1339 	return (space < limit);
1340 }
1341 
1342 static bool
1343 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1344 {
1345 	long rfreevnodes, space;
1346 
1347 	if (__predict_false(rnumvnodes > desiredvnodes))
1348 		return (true);
1349 
1350 	space = desiredvnodes - rnumvnodes;
1351 	if (space < limit) {
1352 		rfreevnodes = atomic_load_long(&freevnodes);
1353 		if (rfreevnodes > wantfreevnodes)
1354 			space += rfreevnodes - wantfreevnodes;
1355 	}
1356 	return (space < limit);
1357 }
1358 
1359 static void
1360 vnlru_kick(void)
1361 {
1362 
1363 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1364 	if (vnlruproc_sig == 0) {
1365 		vnlruproc_sig = 1;
1366 		wakeup(vnlruproc);
1367 	}
1368 }
1369 
1370 static void
1371 vnlru_proc(void)
1372 {
1373 	u_long rnumvnodes, rfreevnodes, target;
1374 	unsigned long onumvnodes;
1375 	int done, force, trigger, usevnodes;
1376 	bool reclaim_nc_src, want_reread;
1377 
1378 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1379 	    SHUTDOWN_PRI_FIRST);
1380 
1381 	force = 0;
1382 	want_reread = false;
1383 	for (;;) {
1384 		kproc_suspend_check(vnlruproc);
1385 		mtx_lock(&vnode_list_mtx);
1386 		rnumvnodes = atomic_load_long(&numvnodes);
1387 
1388 		if (want_reread) {
1389 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1390 			want_reread = false;
1391 		}
1392 
1393 		/*
1394 		 * If numvnodes is too large (due to desiredvnodes being
1395 		 * adjusted using its sysctl, or emergency growth), first
1396 		 * try to reduce it by discarding from the free list.
1397 		 */
1398 		if (rnumvnodes > desiredvnodes) {
1399 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1400 			rnumvnodes = atomic_load_long(&numvnodes);
1401 		}
1402 		/*
1403 		 * Sleep if the vnode cache is in a good state.  This is
1404 		 * when it is not over-full and has space for about a 4%
1405 		 * or 9% expansion (by growing its size or inexcessively
1406 		 * reducing its free list).  Otherwise, try to reclaim
1407 		 * space for a 10% expansion.
1408 		 */
1409 		if (vstir && force == 0) {
1410 			force = 1;
1411 			vstir = 0;
1412 		}
1413 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1414 			vnlruproc_sig = 0;
1415 			wakeup(&vnlruproc_sig);
1416 			msleep(vnlruproc, &vnode_list_mtx,
1417 			    PVFS|PDROP, "vlruwt", hz);
1418 			continue;
1419 		}
1420 		rfreevnodes = vnlru_read_freevnodes();
1421 
1422 		onumvnodes = rnumvnodes;
1423 		/*
1424 		 * Calculate parameters for recycling.  These are the same
1425 		 * throughout the loop to give some semblance of fairness.
1426 		 * The trigger point is to avoid recycling vnodes with lots
1427 		 * of resident pages.  We aren't trying to free memory; we
1428 		 * are trying to recycle or at least free vnodes.
1429 		 */
1430 		if (rnumvnodes <= desiredvnodes)
1431 			usevnodes = rnumvnodes - rfreevnodes;
1432 		else
1433 			usevnodes = rnumvnodes;
1434 		if (usevnodes <= 0)
1435 			usevnodes = 1;
1436 		/*
1437 		 * The trigger value is is chosen to give a conservatively
1438 		 * large value to ensure that it alone doesn't prevent
1439 		 * making progress.  The value can easily be so large that
1440 		 * it is effectively infinite in some congested and
1441 		 * misconfigured cases, and this is necessary.  Normally
1442 		 * it is about 8 to 100 (pages), which is quite large.
1443 		 */
1444 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1445 		if (force < 2)
1446 			trigger = vsmalltrigger;
1447 		reclaim_nc_src = force >= 3;
1448 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1449 		target = target / 10 + 1;
1450 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1451 		mtx_unlock(&vnode_list_mtx);
1452 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1453 			uma_reclaim(UMA_RECLAIM_DRAIN);
1454 		if (done == 0) {
1455 			if (force == 0 || force == 1) {
1456 				force = 2;
1457 				continue;
1458 			}
1459 			if (force == 2) {
1460 				force = 3;
1461 				continue;
1462 			}
1463 			want_reread = true;
1464 			force = 0;
1465 			vnlru_nowhere++;
1466 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1467 		} else {
1468 			want_reread = true;
1469 			kern_yield(PRI_USER);
1470 		}
1471 	}
1472 }
1473 
1474 static struct kproc_desc vnlru_kp = {
1475 	"vnlru",
1476 	vnlru_proc,
1477 	&vnlruproc
1478 };
1479 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1480     &vnlru_kp);
1481 
1482 /*
1483  * Routines having to do with the management of the vnode table.
1484  */
1485 
1486 /*
1487  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1488  * before we actually vgone().  This function must be called with the vnode
1489  * held to prevent the vnode from being returned to the free list midway
1490  * through vgone().
1491  */
1492 static int
1493 vtryrecycle(struct vnode *vp)
1494 {
1495 	struct mount *vnmp;
1496 
1497 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1498 	VNASSERT(vp->v_holdcnt, vp,
1499 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1500 	/*
1501 	 * This vnode may found and locked via some other list, if so we
1502 	 * can't recycle it yet.
1503 	 */
1504 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1505 		CTR2(KTR_VFS,
1506 		    "%s: impossible to recycle, vp %p lock is already held",
1507 		    __func__, vp);
1508 		return (EWOULDBLOCK);
1509 	}
1510 	/*
1511 	 * Don't recycle if its filesystem is being suspended.
1512 	 */
1513 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1514 		VOP_UNLOCK(vp);
1515 		CTR2(KTR_VFS,
1516 		    "%s: impossible to recycle, cannot start the write for %p",
1517 		    __func__, vp);
1518 		return (EBUSY);
1519 	}
1520 	/*
1521 	 * If we got this far, we need to acquire the interlock and see if
1522 	 * anyone picked up this vnode from another list.  If not, we will
1523 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1524 	 * will skip over it.
1525 	 */
1526 	VI_LOCK(vp);
1527 	if (vp->v_usecount) {
1528 		VOP_UNLOCK(vp);
1529 		VI_UNLOCK(vp);
1530 		vn_finished_write(vnmp);
1531 		CTR2(KTR_VFS,
1532 		    "%s: impossible to recycle, %p is already referenced",
1533 		    __func__, vp);
1534 		return (EBUSY);
1535 	}
1536 	if (!VN_IS_DOOMED(vp)) {
1537 		counter_u64_add(recycles_free_count, 1);
1538 		vgonel(vp);
1539 	}
1540 	VOP_UNLOCK(vp);
1541 	VI_UNLOCK(vp);
1542 	vn_finished_write(vnmp);
1543 	return (0);
1544 }
1545 
1546 /*
1547  * Allocate a new vnode.
1548  *
1549  * The operation never returns an error. Returning an error was disabled
1550  * in r145385 (dated 2005) with the following comment:
1551  *
1552  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1553  *
1554  * Given the age of this commit (almost 15 years at the time of writing this
1555  * comment) restoring the ability to fail requires a significant audit of
1556  * all codepaths.
1557  *
1558  * The routine can try to free a vnode or stall for up to 1 second waiting for
1559  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1560  */
1561 static u_long vn_alloc_cyclecount;
1562 
1563 static struct vnode * __noinline
1564 vn_alloc_hard(struct mount *mp)
1565 {
1566 	u_long rnumvnodes, rfreevnodes;
1567 
1568 	mtx_lock(&vnode_list_mtx);
1569 	rnumvnodes = atomic_load_long(&numvnodes);
1570 	if (rnumvnodes + 1 < desiredvnodes) {
1571 		vn_alloc_cyclecount = 0;
1572 		goto alloc;
1573 	}
1574 	rfreevnodes = vnlru_read_freevnodes();
1575 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1576 		vn_alloc_cyclecount = 0;
1577 		vstir = 1;
1578 	}
1579 	/*
1580 	 * Grow the vnode cache if it will not be above its target max
1581 	 * after growing.  Otherwise, if the free list is nonempty, try
1582 	 * to reclaim 1 item from it before growing the cache (possibly
1583 	 * above its target max if the reclamation failed or is delayed).
1584 	 * Otherwise, wait for some space.  In all cases, schedule
1585 	 * vnlru_proc() if we are getting short of space.  The watermarks
1586 	 * should be chosen so that we never wait or even reclaim from
1587 	 * the free list to below its target minimum.
1588 	 */
1589 	if (vnlru_free_locked(1, NULL) > 0)
1590 		goto alloc;
1591 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1592 		/*
1593 		 * Wait for space for a new vnode.
1594 		 */
1595 		vnlru_kick();
1596 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1597 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1598 		    vnlru_read_freevnodes() > 1)
1599 			vnlru_free_locked(1, NULL);
1600 	}
1601 alloc:
1602 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1603 	if (vnlru_under(rnumvnodes, vlowat))
1604 		vnlru_kick();
1605 	mtx_unlock(&vnode_list_mtx);
1606 	return (uma_zalloc(vnode_zone, M_WAITOK));
1607 }
1608 
1609 static struct vnode *
1610 vn_alloc(struct mount *mp)
1611 {
1612 	u_long rnumvnodes;
1613 
1614 	if (__predict_false(vn_alloc_cyclecount != 0))
1615 		return (vn_alloc_hard(mp));
1616 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1617 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1618 		atomic_subtract_long(&numvnodes, 1);
1619 		return (vn_alloc_hard(mp));
1620 	}
1621 
1622 	return (uma_zalloc(vnode_zone, M_WAITOK));
1623 }
1624 
1625 static void
1626 vn_free(struct vnode *vp)
1627 {
1628 
1629 	atomic_subtract_long(&numvnodes, 1);
1630 	uma_zfree(vnode_zone, vp);
1631 }
1632 
1633 /*
1634  * Return the next vnode from the free list.
1635  */
1636 int
1637 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1638     struct vnode **vpp)
1639 {
1640 	struct vnode *vp;
1641 	struct thread *td;
1642 	struct lock_object *lo;
1643 
1644 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1645 
1646 	KASSERT(vops->registered,
1647 	    ("%s: not registered vector op %p\n", __func__, vops));
1648 
1649 	td = curthread;
1650 	if (td->td_vp_reserved != NULL) {
1651 		vp = td->td_vp_reserved;
1652 		td->td_vp_reserved = NULL;
1653 	} else {
1654 		vp = vn_alloc(mp);
1655 	}
1656 	counter_u64_add(vnodes_created, 1);
1657 	/*
1658 	 * Locks are given the generic name "vnode" when created.
1659 	 * Follow the historic practice of using the filesystem
1660 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1661 	 *
1662 	 * Locks live in a witness group keyed on their name. Thus,
1663 	 * when a lock is renamed, it must also move from the witness
1664 	 * group of its old name to the witness group of its new name.
1665 	 *
1666 	 * The change only needs to be made when the vnode moves
1667 	 * from one filesystem type to another. We ensure that each
1668 	 * filesystem use a single static name pointer for its tag so
1669 	 * that we can compare pointers rather than doing a strcmp().
1670 	 */
1671 	lo = &vp->v_vnlock->lock_object;
1672 #ifdef WITNESS
1673 	if (lo->lo_name != tag) {
1674 #endif
1675 		lo->lo_name = tag;
1676 #ifdef WITNESS
1677 		WITNESS_DESTROY(lo);
1678 		WITNESS_INIT(lo, tag);
1679 	}
1680 #endif
1681 	/*
1682 	 * By default, don't allow shared locks unless filesystems opt-in.
1683 	 */
1684 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1685 	/*
1686 	 * Finalize various vnode identity bits.
1687 	 */
1688 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1689 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1690 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1691 	vp->v_type = VNON;
1692 	vp->v_op = vops;
1693 	v_init_counters(vp);
1694 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1695 #ifdef DIAGNOSTIC
1696 	if (mp == NULL && vops != &dead_vnodeops)
1697 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1698 #endif
1699 #ifdef MAC
1700 	mac_vnode_init(vp);
1701 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1702 		mac_vnode_associate_singlelabel(mp, vp);
1703 #endif
1704 	if (mp != NULL) {
1705 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1706 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1707 			vp->v_vflag |= VV_NOKNOTE;
1708 	}
1709 
1710 	/*
1711 	 * For the filesystems which do not use vfs_hash_insert(),
1712 	 * still initialize v_hash to have vfs_hash_index() useful.
1713 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1714 	 * its own hashing.
1715 	 */
1716 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1717 
1718 	*vpp = vp;
1719 	return (0);
1720 }
1721 
1722 void
1723 getnewvnode_reserve(void)
1724 {
1725 	struct thread *td;
1726 
1727 	td = curthread;
1728 	MPASS(td->td_vp_reserved == NULL);
1729 	td->td_vp_reserved = vn_alloc(NULL);
1730 }
1731 
1732 void
1733 getnewvnode_drop_reserve(void)
1734 {
1735 	struct thread *td;
1736 
1737 	td = curthread;
1738 	if (td->td_vp_reserved != NULL) {
1739 		vn_free(td->td_vp_reserved);
1740 		td->td_vp_reserved = NULL;
1741 	}
1742 }
1743 
1744 static void
1745 freevnode(struct vnode *vp)
1746 {
1747 	struct bufobj *bo;
1748 
1749 	/*
1750 	 * The vnode has been marked for destruction, so free it.
1751 	 *
1752 	 * The vnode will be returned to the zone where it will
1753 	 * normally remain until it is needed for another vnode. We
1754 	 * need to cleanup (or verify that the cleanup has already
1755 	 * been done) any residual data left from its current use
1756 	 * so as not to contaminate the freshly allocated vnode.
1757 	 */
1758 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1759 	bo = &vp->v_bufobj;
1760 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1761 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1762 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1763 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1764 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1765 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1766 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1767 	    ("clean blk trie not empty"));
1768 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1769 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1770 	    ("dirty blk trie not empty"));
1771 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1772 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1773 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1774 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1775 	    ("Dangling rangelock waiters"));
1776 	VI_UNLOCK(vp);
1777 #ifdef MAC
1778 	mac_vnode_destroy(vp);
1779 #endif
1780 	if (vp->v_pollinfo != NULL) {
1781 		destroy_vpollinfo(vp->v_pollinfo);
1782 		vp->v_pollinfo = NULL;
1783 	}
1784 #ifdef INVARIANTS
1785 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1786 	vp->v_op = NULL;
1787 #endif
1788 	vp->v_mountedhere = NULL;
1789 	vp->v_unpcb = NULL;
1790 	vp->v_rdev = NULL;
1791 	vp->v_fifoinfo = NULL;
1792 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1793 	vp->v_irflag = 0;
1794 	vp->v_iflag = 0;
1795 	vp->v_vflag = 0;
1796 	bo->bo_flag = 0;
1797 	vn_free(vp);
1798 }
1799 
1800 /*
1801  * Delete from old mount point vnode list, if on one.
1802  */
1803 static void
1804 delmntque(struct vnode *vp)
1805 {
1806 	struct mount *mp;
1807 
1808 	mp = vp->v_mount;
1809 	if (mp == NULL)
1810 		return;
1811 	MNT_ILOCK(mp);
1812 	VI_LOCK(vp);
1813 	if (vp->v_mflag & VMP_LAZYLIST) {
1814 		mtx_lock(&mp->mnt_listmtx);
1815 		if (vp->v_mflag & VMP_LAZYLIST) {
1816 			vp->v_mflag &= ~VMP_LAZYLIST;
1817 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
1818 			mp->mnt_lazyvnodelistsize--;
1819 		}
1820 		mtx_unlock(&mp->mnt_listmtx);
1821 	}
1822 	vp->v_mount = NULL;
1823 	VI_UNLOCK(vp);
1824 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1825 		("bad mount point vnode list size"));
1826 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1827 	mp->mnt_nvnodelistsize--;
1828 	MNT_REL(mp);
1829 	MNT_IUNLOCK(mp);
1830 }
1831 
1832 static void
1833 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1834 {
1835 
1836 	vp->v_data = NULL;
1837 	vp->v_op = &dead_vnodeops;
1838 	vgone(vp);
1839 	vput(vp);
1840 }
1841 
1842 /*
1843  * Insert into list of vnodes for the new mount point, if available.
1844  */
1845 int
1846 insmntque1(struct vnode *vp, struct mount *mp,
1847 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1848 {
1849 
1850 	KASSERT(vp->v_mount == NULL,
1851 		("insmntque: vnode already on per mount vnode list"));
1852 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1853 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1854 
1855 	/*
1856 	 * We acquire the vnode interlock early to ensure that the
1857 	 * vnode cannot be recycled by another process releasing a
1858 	 * holdcnt on it before we get it on both the vnode list
1859 	 * and the active vnode list. The mount mutex protects only
1860 	 * manipulation of the vnode list and the vnode freelist
1861 	 * mutex protects only manipulation of the active vnode list.
1862 	 * Hence the need to hold the vnode interlock throughout.
1863 	 */
1864 	MNT_ILOCK(mp);
1865 	VI_LOCK(vp);
1866 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1867 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1868 	    mp->mnt_nvnodelistsize == 0)) &&
1869 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1870 		VI_UNLOCK(vp);
1871 		MNT_IUNLOCK(mp);
1872 		if (dtr != NULL)
1873 			dtr(vp, dtr_arg);
1874 		return (EBUSY);
1875 	}
1876 	vp->v_mount = mp;
1877 	MNT_REF(mp);
1878 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1879 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1880 		("neg mount point vnode list size"));
1881 	mp->mnt_nvnodelistsize++;
1882 	VI_UNLOCK(vp);
1883 	MNT_IUNLOCK(mp);
1884 	return (0);
1885 }
1886 
1887 int
1888 insmntque(struct vnode *vp, struct mount *mp)
1889 {
1890 
1891 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1892 }
1893 
1894 /*
1895  * Flush out and invalidate all buffers associated with a bufobj
1896  * Called with the underlying object locked.
1897  */
1898 int
1899 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1900 {
1901 	int error;
1902 
1903 	BO_LOCK(bo);
1904 	if (flags & V_SAVE) {
1905 		error = bufobj_wwait(bo, slpflag, slptimeo);
1906 		if (error) {
1907 			BO_UNLOCK(bo);
1908 			return (error);
1909 		}
1910 		if (bo->bo_dirty.bv_cnt > 0) {
1911 			BO_UNLOCK(bo);
1912 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1913 				return (error);
1914 			/*
1915 			 * XXX We could save a lock/unlock if this was only
1916 			 * enabled under INVARIANTS
1917 			 */
1918 			BO_LOCK(bo);
1919 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1920 				panic("vinvalbuf: dirty bufs");
1921 		}
1922 	}
1923 	/*
1924 	 * If you alter this loop please notice that interlock is dropped and
1925 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1926 	 * no race conditions occur from this.
1927 	 */
1928 	do {
1929 		error = flushbuflist(&bo->bo_clean,
1930 		    flags, bo, slpflag, slptimeo);
1931 		if (error == 0 && !(flags & V_CLEANONLY))
1932 			error = flushbuflist(&bo->bo_dirty,
1933 			    flags, bo, slpflag, slptimeo);
1934 		if (error != 0 && error != EAGAIN) {
1935 			BO_UNLOCK(bo);
1936 			return (error);
1937 		}
1938 	} while (error != 0);
1939 
1940 	/*
1941 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1942 	 * have write I/O in-progress but if there is a VM object then the
1943 	 * VM object can also have read-I/O in-progress.
1944 	 */
1945 	do {
1946 		bufobj_wwait(bo, 0, 0);
1947 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1948 			BO_UNLOCK(bo);
1949 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1950 			BO_LOCK(bo);
1951 		}
1952 	} while (bo->bo_numoutput > 0);
1953 	BO_UNLOCK(bo);
1954 
1955 	/*
1956 	 * Destroy the copy in the VM cache, too.
1957 	 */
1958 	if (bo->bo_object != NULL &&
1959 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1960 		VM_OBJECT_WLOCK(bo->bo_object);
1961 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1962 		    OBJPR_CLEANONLY : 0);
1963 		VM_OBJECT_WUNLOCK(bo->bo_object);
1964 	}
1965 
1966 #ifdef INVARIANTS
1967 	BO_LOCK(bo);
1968 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1969 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1970 	    bo->bo_clean.bv_cnt > 0))
1971 		panic("vinvalbuf: flush failed");
1972 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1973 	    bo->bo_dirty.bv_cnt > 0)
1974 		panic("vinvalbuf: flush dirty failed");
1975 	BO_UNLOCK(bo);
1976 #endif
1977 	return (0);
1978 }
1979 
1980 /*
1981  * Flush out and invalidate all buffers associated with a vnode.
1982  * Called with the underlying object locked.
1983  */
1984 int
1985 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1986 {
1987 
1988 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1989 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1990 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1991 		return (0);
1992 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1993 }
1994 
1995 /*
1996  * Flush out buffers on the specified list.
1997  *
1998  */
1999 static int
2000 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2001     int slptimeo)
2002 {
2003 	struct buf *bp, *nbp;
2004 	int retval, error;
2005 	daddr_t lblkno;
2006 	b_xflags_t xflags;
2007 
2008 	ASSERT_BO_WLOCKED(bo);
2009 
2010 	retval = 0;
2011 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2012 		/*
2013 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2014 		 * do not skip any buffers. If we are flushing only V_NORMAL
2015 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2016 		 * flushing only V_ALT buffers then skip buffers not marked
2017 		 * as BX_ALTDATA.
2018 		 */
2019 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2020 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2021 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2022 			continue;
2023 		}
2024 		if (nbp != NULL) {
2025 			lblkno = nbp->b_lblkno;
2026 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2027 		}
2028 		retval = EAGAIN;
2029 		error = BUF_TIMELOCK(bp,
2030 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2031 		    "flushbuf", slpflag, slptimeo);
2032 		if (error) {
2033 			BO_LOCK(bo);
2034 			return (error != ENOLCK ? error : EAGAIN);
2035 		}
2036 		KASSERT(bp->b_bufobj == bo,
2037 		    ("bp %p wrong b_bufobj %p should be %p",
2038 		    bp, bp->b_bufobj, bo));
2039 		/*
2040 		 * XXX Since there are no node locks for NFS, I
2041 		 * believe there is a slight chance that a delayed
2042 		 * write will occur while sleeping just above, so
2043 		 * check for it.
2044 		 */
2045 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2046 		    (flags & V_SAVE)) {
2047 			bremfree(bp);
2048 			bp->b_flags |= B_ASYNC;
2049 			bwrite(bp);
2050 			BO_LOCK(bo);
2051 			return (EAGAIN);	/* XXX: why not loop ? */
2052 		}
2053 		bremfree(bp);
2054 		bp->b_flags |= (B_INVAL | B_RELBUF);
2055 		bp->b_flags &= ~B_ASYNC;
2056 		brelse(bp);
2057 		BO_LOCK(bo);
2058 		if (nbp == NULL)
2059 			break;
2060 		nbp = gbincore(bo, lblkno);
2061 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2062 		    != xflags)
2063 			break;			/* nbp invalid */
2064 	}
2065 	return (retval);
2066 }
2067 
2068 int
2069 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2070 {
2071 	struct buf *bp;
2072 	int error;
2073 	daddr_t lblkno;
2074 
2075 	ASSERT_BO_LOCKED(bo);
2076 
2077 	for (lblkno = startn;;) {
2078 again:
2079 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2080 		if (bp == NULL || bp->b_lblkno >= endn ||
2081 		    bp->b_lblkno < startn)
2082 			break;
2083 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2084 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2085 		if (error != 0) {
2086 			BO_RLOCK(bo);
2087 			if (error == ENOLCK)
2088 				goto again;
2089 			return (error);
2090 		}
2091 		KASSERT(bp->b_bufobj == bo,
2092 		    ("bp %p wrong b_bufobj %p should be %p",
2093 		    bp, bp->b_bufobj, bo));
2094 		lblkno = bp->b_lblkno + 1;
2095 		if ((bp->b_flags & B_MANAGED) == 0)
2096 			bremfree(bp);
2097 		bp->b_flags |= B_RELBUF;
2098 		/*
2099 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2100 		 * pages backing each buffer in the range are unlikely to be
2101 		 * reused.  Dirty buffers will have the hint applied once
2102 		 * they've been written.
2103 		 */
2104 		if ((bp->b_flags & B_VMIO) != 0)
2105 			bp->b_flags |= B_NOREUSE;
2106 		brelse(bp);
2107 		BO_RLOCK(bo);
2108 	}
2109 	return (0);
2110 }
2111 
2112 /*
2113  * Truncate a file's buffer and pages to a specified length.  This
2114  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2115  * sync activity.
2116  */
2117 int
2118 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2119 {
2120 	struct buf *bp, *nbp;
2121 	struct bufobj *bo;
2122 	daddr_t startlbn;
2123 
2124 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2125 	    vp, blksize, (uintmax_t)length);
2126 
2127 	/*
2128 	 * Round up to the *next* lbn.
2129 	 */
2130 	startlbn = howmany(length, blksize);
2131 
2132 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2133 
2134 	bo = &vp->v_bufobj;
2135 restart_unlocked:
2136 	BO_LOCK(bo);
2137 
2138 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2139 		;
2140 
2141 	if (length > 0) {
2142 restartsync:
2143 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2144 			if (bp->b_lblkno > 0)
2145 				continue;
2146 			/*
2147 			 * Since we hold the vnode lock this should only
2148 			 * fail if we're racing with the buf daemon.
2149 			 */
2150 			if (BUF_LOCK(bp,
2151 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2152 			    BO_LOCKPTR(bo)) == ENOLCK)
2153 				goto restart_unlocked;
2154 
2155 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2156 			    ("buf(%p) on dirty queue without DELWRI", bp));
2157 
2158 			bremfree(bp);
2159 			bawrite(bp);
2160 			BO_LOCK(bo);
2161 			goto restartsync;
2162 		}
2163 	}
2164 
2165 	bufobj_wwait(bo, 0, 0);
2166 	BO_UNLOCK(bo);
2167 	vnode_pager_setsize(vp, length);
2168 
2169 	return (0);
2170 }
2171 
2172 /*
2173  * Invalidate the cached pages of a file's buffer within the range of block
2174  * numbers [startlbn, endlbn).
2175  */
2176 void
2177 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2178     int blksize)
2179 {
2180 	struct bufobj *bo;
2181 	off_t start, end;
2182 
2183 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2184 
2185 	start = blksize * startlbn;
2186 	end = blksize * endlbn;
2187 
2188 	bo = &vp->v_bufobj;
2189 	BO_LOCK(bo);
2190 	MPASS(blksize == bo->bo_bsize);
2191 
2192 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2193 		;
2194 
2195 	BO_UNLOCK(bo);
2196 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2197 }
2198 
2199 static int
2200 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2201     daddr_t startlbn, daddr_t endlbn)
2202 {
2203 	struct buf *bp, *nbp;
2204 	bool anyfreed;
2205 
2206 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2207 	ASSERT_BO_LOCKED(bo);
2208 
2209 	do {
2210 		anyfreed = false;
2211 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2212 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2213 				continue;
2214 			if (BUF_LOCK(bp,
2215 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2216 			    BO_LOCKPTR(bo)) == ENOLCK) {
2217 				BO_LOCK(bo);
2218 				return (EAGAIN);
2219 			}
2220 
2221 			bremfree(bp);
2222 			bp->b_flags |= B_INVAL | B_RELBUF;
2223 			bp->b_flags &= ~B_ASYNC;
2224 			brelse(bp);
2225 			anyfreed = true;
2226 
2227 			BO_LOCK(bo);
2228 			if (nbp != NULL &&
2229 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2230 			    nbp->b_vp != vp ||
2231 			    (nbp->b_flags & B_DELWRI) != 0))
2232 				return (EAGAIN);
2233 		}
2234 
2235 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2236 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2237 				continue;
2238 			if (BUF_LOCK(bp,
2239 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2240 			    BO_LOCKPTR(bo)) == ENOLCK) {
2241 				BO_LOCK(bo);
2242 				return (EAGAIN);
2243 			}
2244 			bremfree(bp);
2245 			bp->b_flags |= B_INVAL | B_RELBUF;
2246 			bp->b_flags &= ~B_ASYNC;
2247 			brelse(bp);
2248 			anyfreed = true;
2249 
2250 			BO_LOCK(bo);
2251 			if (nbp != NULL &&
2252 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2253 			    (nbp->b_vp != vp) ||
2254 			    (nbp->b_flags & B_DELWRI) == 0))
2255 				return (EAGAIN);
2256 		}
2257 	} while (anyfreed);
2258 	return (0);
2259 }
2260 
2261 static void
2262 buf_vlist_remove(struct buf *bp)
2263 {
2264 	struct bufv *bv;
2265 
2266 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2267 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2268 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2269 	    (BX_VNDIRTY|BX_VNCLEAN),
2270 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2271 	if (bp->b_xflags & BX_VNDIRTY)
2272 		bv = &bp->b_bufobj->bo_dirty;
2273 	else
2274 		bv = &bp->b_bufobj->bo_clean;
2275 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2276 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2277 	bv->bv_cnt--;
2278 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2279 }
2280 
2281 /*
2282  * Add the buffer to the sorted clean or dirty block list.
2283  *
2284  * NOTE: xflags is passed as a constant, optimizing this inline function!
2285  */
2286 static void
2287 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2288 {
2289 	struct bufv *bv;
2290 	struct buf *n;
2291 	int error;
2292 
2293 	ASSERT_BO_WLOCKED(bo);
2294 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2295 	    ("dead bo %p", bo));
2296 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2297 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2298 	bp->b_xflags |= xflags;
2299 	if (xflags & BX_VNDIRTY)
2300 		bv = &bo->bo_dirty;
2301 	else
2302 		bv = &bo->bo_clean;
2303 
2304 	/*
2305 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2306 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2307 	 * than _ge.
2308 	 */
2309 	if (bv->bv_cnt == 0 ||
2310 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2311 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2312 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2313 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2314 	else
2315 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2316 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2317 	if (error)
2318 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2319 	bv->bv_cnt++;
2320 }
2321 
2322 /*
2323  * Look up a buffer using the buffer tries.
2324  */
2325 struct buf *
2326 gbincore(struct bufobj *bo, daddr_t lblkno)
2327 {
2328 	struct buf *bp;
2329 
2330 	ASSERT_BO_LOCKED(bo);
2331 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2332 	if (bp != NULL)
2333 		return (bp);
2334 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2335 }
2336 
2337 /*
2338  * Associate a buffer with a vnode.
2339  */
2340 void
2341 bgetvp(struct vnode *vp, struct buf *bp)
2342 {
2343 	struct bufobj *bo;
2344 
2345 	bo = &vp->v_bufobj;
2346 	ASSERT_BO_WLOCKED(bo);
2347 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2348 
2349 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2350 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2351 	    ("bgetvp: bp already attached! %p", bp));
2352 
2353 	vhold(vp);
2354 	bp->b_vp = vp;
2355 	bp->b_bufobj = bo;
2356 	/*
2357 	 * Insert onto list for new vnode.
2358 	 */
2359 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2360 }
2361 
2362 /*
2363  * Disassociate a buffer from a vnode.
2364  */
2365 void
2366 brelvp(struct buf *bp)
2367 {
2368 	struct bufobj *bo;
2369 	struct vnode *vp;
2370 
2371 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2372 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2373 
2374 	/*
2375 	 * Delete from old vnode list, if on one.
2376 	 */
2377 	vp = bp->b_vp;		/* XXX */
2378 	bo = bp->b_bufobj;
2379 	BO_LOCK(bo);
2380 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2381 		buf_vlist_remove(bp);
2382 	else
2383 		panic("brelvp: Buffer %p not on queue.", bp);
2384 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2385 		bo->bo_flag &= ~BO_ONWORKLST;
2386 		mtx_lock(&sync_mtx);
2387 		LIST_REMOVE(bo, bo_synclist);
2388 		syncer_worklist_len--;
2389 		mtx_unlock(&sync_mtx);
2390 	}
2391 	bp->b_vp = NULL;
2392 	bp->b_bufobj = NULL;
2393 	BO_UNLOCK(bo);
2394 	vdrop(vp);
2395 }
2396 
2397 /*
2398  * Add an item to the syncer work queue.
2399  */
2400 static void
2401 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2402 {
2403 	int slot;
2404 
2405 	ASSERT_BO_WLOCKED(bo);
2406 
2407 	mtx_lock(&sync_mtx);
2408 	if (bo->bo_flag & BO_ONWORKLST)
2409 		LIST_REMOVE(bo, bo_synclist);
2410 	else {
2411 		bo->bo_flag |= BO_ONWORKLST;
2412 		syncer_worklist_len++;
2413 	}
2414 
2415 	if (delay > syncer_maxdelay - 2)
2416 		delay = syncer_maxdelay - 2;
2417 	slot = (syncer_delayno + delay) & syncer_mask;
2418 
2419 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2420 	mtx_unlock(&sync_mtx);
2421 }
2422 
2423 static int
2424 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2425 {
2426 	int error, len;
2427 
2428 	mtx_lock(&sync_mtx);
2429 	len = syncer_worklist_len - sync_vnode_count;
2430 	mtx_unlock(&sync_mtx);
2431 	error = SYSCTL_OUT(req, &len, sizeof(len));
2432 	return (error);
2433 }
2434 
2435 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2436     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2437     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2438 
2439 static struct proc *updateproc;
2440 static void sched_sync(void);
2441 static struct kproc_desc up_kp = {
2442 	"syncer",
2443 	sched_sync,
2444 	&updateproc
2445 };
2446 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2447 
2448 static int
2449 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2450 {
2451 	struct vnode *vp;
2452 	struct mount *mp;
2453 
2454 	*bo = LIST_FIRST(slp);
2455 	if (*bo == NULL)
2456 		return (0);
2457 	vp = bo2vnode(*bo);
2458 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2459 		return (1);
2460 	/*
2461 	 * We use vhold in case the vnode does not
2462 	 * successfully sync.  vhold prevents the vnode from
2463 	 * going away when we unlock the sync_mtx so that
2464 	 * we can acquire the vnode interlock.
2465 	 */
2466 	vholdl(vp);
2467 	mtx_unlock(&sync_mtx);
2468 	VI_UNLOCK(vp);
2469 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2470 		vdrop(vp);
2471 		mtx_lock(&sync_mtx);
2472 		return (*bo == LIST_FIRST(slp));
2473 	}
2474 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2475 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2476 	VOP_UNLOCK(vp);
2477 	vn_finished_write(mp);
2478 	BO_LOCK(*bo);
2479 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2480 		/*
2481 		 * Put us back on the worklist.  The worklist
2482 		 * routine will remove us from our current
2483 		 * position and then add us back in at a later
2484 		 * position.
2485 		 */
2486 		vn_syncer_add_to_worklist(*bo, syncdelay);
2487 	}
2488 	BO_UNLOCK(*bo);
2489 	vdrop(vp);
2490 	mtx_lock(&sync_mtx);
2491 	return (0);
2492 }
2493 
2494 static int first_printf = 1;
2495 
2496 /*
2497  * System filesystem synchronizer daemon.
2498  */
2499 static void
2500 sched_sync(void)
2501 {
2502 	struct synclist *next, *slp;
2503 	struct bufobj *bo;
2504 	long starttime;
2505 	struct thread *td = curthread;
2506 	int last_work_seen;
2507 	int net_worklist_len;
2508 	int syncer_final_iter;
2509 	int error;
2510 
2511 	last_work_seen = 0;
2512 	syncer_final_iter = 0;
2513 	syncer_state = SYNCER_RUNNING;
2514 	starttime = time_uptime;
2515 	td->td_pflags |= TDP_NORUNNINGBUF;
2516 
2517 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2518 	    SHUTDOWN_PRI_LAST);
2519 
2520 	mtx_lock(&sync_mtx);
2521 	for (;;) {
2522 		if (syncer_state == SYNCER_FINAL_DELAY &&
2523 		    syncer_final_iter == 0) {
2524 			mtx_unlock(&sync_mtx);
2525 			kproc_suspend_check(td->td_proc);
2526 			mtx_lock(&sync_mtx);
2527 		}
2528 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2529 		if (syncer_state != SYNCER_RUNNING &&
2530 		    starttime != time_uptime) {
2531 			if (first_printf) {
2532 				printf("\nSyncing disks, vnodes remaining... ");
2533 				first_printf = 0;
2534 			}
2535 			printf("%d ", net_worklist_len);
2536 		}
2537 		starttime = time_uptime;
2538 
2539 		/*
2540 		 * Push files whose dirty time has expired.  Be careful
2541 		 * of interrupt race on slp queue.
2542 		 *
2543 		 * Skip over empty worklist slots when shutting down.
2544 		 */
2545 		do {
2546 			slp = &syncer_workitem_pending[syncer_delayno];
2547 			syncer_delayno += 1;
2548 			if (syncer_delayno == syncer_maxdelay)
2549 				syncer_delayno = 0;
2550 			next = &syncer_workitem_pending[syncer_delayno];
2551 			/*
2552 			 * If the worklist has wrapped since the
2553 			 * it was emptied of all but syncer vnodes,
2554 			 * switch to the FINAL_DELAY state and run
2555 			 * for one more second.
2556 			 */
2557 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2558 			    net_worklist_len == 0 &&
2559 			    last_work_seen == syncer_delayno) {
2560 				syncer_state = SYNCER_FINAL_DELAY;
2561 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2562 			}
2563 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2564 		    syncer_worklist_len > 0);
2565 
2566 		/*
2567 		 * Keep track of the last time there was anything
2568 		 * on the worklist other than syncer vnodes.
2569 		 * Return to the SHUTTING_DOWN state if any
2570 		 * new work appears.
2571 		 */
2572 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2573 			last_work_seen = syncer_delayno;
2574 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2575 			syncer_state = SYNCER_SHUTTING_DOWN;
2576 		while (!LIST_EMPTY(slp)) {
2577 			error = sync_vnode(slp, &bo, td);
2578 			if (error == 1) {
2579 				LIST_REMOVE(bo, bo_synclist);
2580 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2581 				continue;
2582 			}
2583 
2584 			if (first_printf == 0) {
2585 				/*
2586 				 * Drop the sync mutex, because some watchdog
2587 				 * drivers need to sleep while patting
2588 				 */
2589 				mtx_unlock(&sync_mtx);
2590 				wdog_kern_pat(WD_LASTVAL);
2591 				mtx_lock(&sync_mtx);
2592 			}
2593 
2594 		}
2595 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2596 			syncer_final_iter--;
2597 		/*
2598 		 * The variable rushjob allows the kernel to speed up the
2599 		 * processing of the filesystem syncer process. A rushjob
2600 		 * value of N tells the filesystem syncer to process the next
2601 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2602 		 * is used by the soft update code to speed up the filesystem
2603 		 * syncer process when the incore state is getting so far
2604 		 * ahead of the disk that the kernel memory pool is being
2605 		 * threatened with exhaustion.
2606 		 */
2607 		if (rushjob > 0) {
2608 			rushjob -= 1;
2609 			continue;
2610 		}
2611 		/*
2612 		 * Just sleep for a short period of time between
2613 		 * iterations when shutting down to allow some I/O
2614 		 * to happen.
2615 		 *
2616 		 * If it has taken us less than a second to process the
2617 		 * current work, then wait. Otherwise start right over
2618 		 * again. We can still lose time if any single round
2619 		 * takes more than two seconds, but it does not really
2620 		 * matter as we are just trying to generally pace the
2621 		 * filesystem activity.
2622 		 */
2623 		if (syncer_state != SYNCER_RUNNING ||
2624 		    time_uptime == starttime) {
2625 			thread_lock(td);
2626 			sched_prio(td, PPAUSE);
2627 			thread_unlock(td);
2628 		}
2629 		if (syncer_state != SYNCER_RUNNING)
2630 			cv_timedwait(&sync_wakeup, &sync_mtx,
2631 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2632 		else if (time_uptime == starttime)
2633 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2634 	}
2635 }
2636 
2637 /*
2638  * Request the syncer daemon to speed up its work.
2639  * We never push it to speed up more than half of its
2640  * normal turn time, otherwise it could take over the cpu.
2641  */
2642 int
2643 speedup_syncer(void)
2644 {
2645 	int ret = 0;
2646 
2647 	mtx_lock(&sync_mtx);
2648 	if (rushjob < syncdelay / 2) {
2649 		rushjob += 1;
2650 		stat_rush_requests += 1;
2651 		ret = 1;
2652 	}
2653 	mtx_unlock(&sync_mtx);
2654 	cv_broadcast(&sync_wakeup);
2655 	return (ret);
2656 }
2657 
2658 /*
2659  * Tell the syncer to speed up its work and run though its work
2660  * list several times, then tell it to shut down.
2661  */
2662 static void
2663 syncer_shutdown(void *arg, int howto)
2664 {
2665 
2666 	if (howto & RB_NOSYNC)
2667 		return;
2668 	mtx_lock(&sync_mtx);
2669 	syncer_state = SYNCER_SHUTTING_DOWN;
2670 	rushjob = 0;
2671 	mtx_unlock(&sync_mtx);
2672 	cv_broadcast(&sync_wakeup);
2673 	kproc_shutdown(arg, howto);
2674 }
2675 
2676 void
2677 syncer_suspend(void)
2678 {
2679 
2680 	syncer_shutdown(updateproc, 0);
2681 }
2682 
2683 void
2684 syncer_resume(void)
2685 {
2686 
2687 	mtx_lock(&sync_mtx);
2688 	first_printf = 1;
2689 	syncer_state = SYNCER_RUNNING;
2690 	mtx_unlock(&sync_mtx);
2691 	cv_broadcast(&sync_wakeup);
2692 	kproc_resume(updateproc);
2693 }
2694 
2695 /*
2696  * Reassign a buffer from one vnode to another.
2697  * Used to assign file specific control information
2698  * (indirect blocks) to the vnode to which they belong.
2699  */
2700 void
2701 reassignbuf(struct buf *bp)
2702 {
2703 	struct vnode *vp;
2704 	struct bufobj *bo;
2705 	int delay;
2706 #ifdef INVARIANTS
2707 	struct bufv *bv;
2708 #endif
2709 
2710 	vp = bp->b_vp;
2711 	bo = bp->b_bufobj;
2712 	++reassignbufcalls;
2713 
2714 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2715 	    bp, bp->b_vp, bp->b_flags);
2716 	/*
2717 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2718 	 * is not fully linked in.
2719 	 */
2720 	if (bp->b_flags & B_PAGING)
2721 		panic("cannot reassign paging buffer");
2722 
2723 	/*
2724 	 * Delete from old vnode list, if on one.
2725 	 */
2726 	BO_LOCK(bo);
2727 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2728 		buf_vlist_remove(bp);
2729 	else
2730 		panic("reassignbuf: Buffer %p not on queue.", bp);
2731 	/*
2732 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2733 	 * of clean buffers.
2734 	 */
2735 	if (bp->b_flags & B_DELWRI) {
2736 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2737 			switch (vp->v_type) {
2738 			case VDIR:
2739 				delay = dirdelay;
2740 				break;
2741 			case VCHR:
2742 				delay = metadelay;
2743 				break;
2744 			default:
2745 				delay = filedelay;
2746 			}
2747 			vn_syncer_add_to_worklist(bo, delay);
2748 		}
2749 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2750 	} else {
2751 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2752 
2753 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2754 			mtx_lock(&sync_mtx);
2755 			LIST_REMOVE(bo, bo_synclist);
2756 			syncer_worklist_len--;
2757 			mtx_unlock(&sync_mtx);
2758 			bo->bo_flag &= ~BO_ONWORKLST;
2759 		}
2760 	}
2761 #ifdef INVARIANTS
2762 	bv = &bo->bo_clean;
2763 	bp = TAILQ_FIRST(&bv->bv_hd);
2764 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2765 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2766 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2767 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2768 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2769 	bv = &bo->bo_dirty;
2770 	bp = TAILQ_FIRST(&bv->bv_hd);
2771 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2772 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2773 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2774 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2775 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2776 #endif
2777 	BO_UNLOCK(bo);
2778 }
2779 
2780 static void
2781 v_init_counters(struct vnode *vp)
2782 {
2783 
2784 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2785 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2786 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2787 
2788 	refcount_init(&vp->v_holdcnt, 1);
2789 	refcount_init(&vp->v_usecount, 1);
2790 }
2791 
2792 /*
2793  * Increment si_usecount of the associated device, if any.
2794  */
2795 static void
2796 v_incr_devcount(struct vnode *vp)
2797 {
2798 
2799 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2800 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2801 		dev_lock();
2802 		vp->v_rdev->si_usecount++;
2803 		dev_unlock();
2804 	}
2805 }
2806 
2807 /*
2808  * Decrement si_usecount of the associated device, if any.
2809  */
2810 static void
2811 v_decr_devcount(struct vnode *vp)
2812 {
2813 
2814 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2815 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2816 		dev_lock();
2817 		vp->v_rdev->si_usecount--;
2818 		dev_unlock();
2819 	}
2820 }
2821 
2822 /*
2823  * Grab a particular vnode from the free list, increment its
2824  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2825  * is being destroyed.  Only callers who specify LK_RETRY will
2826  * see doomed vnodes.  If inactive processing was delayed in
2827  * vput try to do it here.
2828  *
2829  * usecount is manipulated using atomics without holding any locks.
2830  *
2831  * holdcnt can be manipulated using atomics without holding any locks,
2832  * except when transitioning 1<->0, in which case the interlock is held.
2833  */
2834 enum vgetstate
2835 vget_prep(struct vnode *vp)
2836 {
2837 	enum vgetstate vs;
2838 
2839 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2840 		vs = VGET_USECOUNT;
2841 	} else {
2842 		vhold(vp);
2843 		vs = VGET_HOLDCNT;
2844 	}
2845 	return (vs);
2846 }
2847 
2848 int
2849 vget(struct vnode *vp, int flags, struct thread *td)
2850 {
2851 	enum vgetstate vs;
2852 
2853 	MPASS(td == curthread);
2854 
2855 	vs = vget_prep(vp);
2856 	return (vget_finish(vp, flags, vs));
2857 }
2858 
2859 static int __noinline
2860 vget_finish_vchr(struct vnode *vp)
2861 {
2862 
2863 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2864 
2865 	/*
2866 	 * See the comment in vget_finish before usecount bump.
2867 	 */
2868 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2869 #ifdef INVARIANTS
2870 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2871 		VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2872 #else
2873 		refcount_release(&vp->v_holdcnt);
2874 #endif
2875 		return (0);
2876 	}
2877 
2878 	VI_LOCK(vp);
2879 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2880 #ifdef INVARIANTS
2881 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2882 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2883 #else
2884 		refcount_release(&vp->v_holdcnt);
2885 #endif
2886 		VI_UNLOCK(vp);
2887 		return (0);
2888 	}
2889 	v_incr_devcount(vp);
2890 	refcount_acquire(&vp->v_usecount);
2891 	VI_UNLOCK(vp);
2892 	return (0);
2893 }
2894 
2895 int
2896 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2897 {
2898 	int error, old;
2899 
2900 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2901 	    ("%s: invalid lock operation", __func__));
2902 
2903 	if ((flags & LK_INTERLOCK) != 0)
2904 		ASSERT_VI_LOCKED(vp, __func__);
2905 	else
2906 		ASSERT_VI_UNLOCKED(vp, __func__);
2907 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2908 	if (vs == VGET_USECOUNT) {
2909 		VNASSERT(vp->v_usecount > 0, vp,
2910 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2911 		    __func__));
2912 	}
2913 
2914 	error = vn_lock(vp, flags);
2915 	if (__predict_false(error != 0)) {
2916 		if (vs == VGET_USECOUNT)
2917 			vrele(vp);
2918 		else
2919 			vdrop(vp);
2920 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2921 		    vp);
2922 		return (error);
2923 	}
2924 
2925 	if (vs == VGET_USECOUNT) {
2926 		return (0);
2927 	}
2928 
2929 	if (__predict_false(vp->v_type == VCHR))
2930 		return (vget_finish_vchr(vp));
2931 
2932 	/*
2933 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2934 	 * the vnode around. Otherwise someone else lended their hold count and
2935 	 * we have to drop ours.
2936 	 */
2937 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2938 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2939 	if (old != 0) {
2940 #ifdef INVARIANTS
2941 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2942 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2943 #else
2944 		refcount_release(&vp->v_holdcnt);
2945 #endif
2946 	}
2947 	return (0);
2948 }
2949 
2950 /*
2951  * Increase the reference (use) and hold count of a vnode.
2952  * This will also remove the vnode from the free list if it is presently free.
2953  */
2954 static void __noinline
2955 vref_vchr(struct vnode *vp, bool interlock)
2956 {
2957 
2958 	/*
2959 	 * See the comment in vget_finish before usecount bump.
2960 	 */
2961 	if (!interlock) {
2962 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2963 			VNODE_REFCOUNT_FENCE_ACQ();
2964 			VNASSERT(vp->v_holdcnt > 0, vp,
2965 			    ("%s: active vnode not held", __func__));
2966 			return;
2967 		}
2968 		VI_LOCK(vp);
2969 		/*
2970 		 * By the time we get here the vnode might have been doomed, at
2971 		 * which point the 0->1 use count transition is no longer
2972 		 * protected by the interlock. Since it can't bounce back to
2973 		 * VCHR and requires vref semantics, punt it back
2974 		 */
2975 		if (__predict_false(vp->v_type == VBAD)) {
2976 			VI_UNLOCK(vp);
2977 			vref(vp);
2978 			return;
2979 		}
2980 	}
2981 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2982 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2983 		VNODE_REFCOUNT_FENCE_ACQ();
2984 		VNASSERT(vp->v_holdcnt > 0, vp,
2985 		    ("%s: active vnode not held", __func__));
2986 		if (!interlock)
2987 			VI_UNLOCK(vp);
2988 		return;
2989 	}
2990 	vhold(vp);
2991 	v_incr_devcount(vp);
2992 	refcount_acquire(&vp->v_usecount);
2993 	if (!interlock)
2994 		VI_UNLOCK(vp);
2995 	return;
2996 }
2997 
2998 void
2999 vref(struct vnode *vp)
3000 {
3001 	int old;
3002 
3003 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3004 	if (__predict_false(vp->v_type == VCHR)) {
3005 		 vref_vchr(vp, false);
3006 		 return;
3007 	}
3008 
3009 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3010 		VNODE_REFCOUNT_FENCE_ACQ();
3011 		VNASSERT(vp->v_holdcnt > 0, vp,
3012 		    ("%s: active vnode not held", __func__));
3013 		return;
3014 	}
3015 	vhold(vp);
3016 	/*
3017 	 * See the comment in vget_finish.
3018 	 */
3019 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3020 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3021 	if (old != 0) {
3022 #ifdef INVARIANTS
3023 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3024 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3025 #else
3026 		refcount_release(&vp->v_holdcnt);
3027 #endif
3028 	}
3029 }
3030 
3031 void
3032 vrefl(struct vnode *vp)
3033 {
3034 
3035 	ASSERT_VI_LOCKED(vp, __func__);
3036 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3037 	if (__predict_false(vp->v_type == VCHR)) {
3038 		vref_vchr(vp, true);
3039 		return;
3040 	}
3041 	vref(vp);
3042 }
3043 
3044 void
3045 vrefact(struct vnode *vp)
3046 {
3047 
3048 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3049 #ifdef INVARIANTS
3050 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3051 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3052 #else
3053 	refcount_acquire(&vp->v_usecount);
3054 #endif
3055 }
3056 
3057 /*
3058  * Return reference count of a vnode.
3059  *
3060  * The results of this call are only guaranteed when some mechanism is used to
3061  * stop other processes from gaining references to the vnode.  This may be the
3062  * case if the caller holds the only reference.  This is also useful when stale
3063  * data is acceptable as race conditions may be accounted for by some other
3064  * means.
3065  */
3066 int
3067 vrefcnt(struct vnode *vp)
3068 {
3069 
3070 	return (vp->v_usecount);
3071 }
3072 
3073 void
3074 vlazy(struct vnode *vp)
3075 {
3076 	struct mount *mp;
3077 
3078 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3079 
3080 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3081 		return;
3082 	mp = vp->v_mount;
3083 	mtx_lock(&mp->mnt_listmtx);
3084 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3085 		vp->v_mflag |= VMP_LAZYLIST;
3086 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3087 		mp->mnt_lazyvnodelistsize++;
3088 	}
3089 	mtx_unlock(&mp->mnt_listmtx);
3090 }
3091 
3092 static void
3093 vdefer_inactive(struct vnode *vp)
3094 {
3095 
3096 	ASSERT_VI_LOCKED(vp, __func__);
3097 	VNASSERT(vp->v_holdcnt > 0, vp,
3098 	    ("%s: vnode without hold count", __func__));
3099 	if (VN_IS_DOOMED(vp)) {
3100 		vdropl(vp);
3101 		return;
3102 	}
3103 	if (vp->v_iflag & VI_DEFINACT) {
3104 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3105 		vdropl(vp);
3106 		return;
3107 	}
3108 	if (vp->v_usecount > 0) {
3109 		vp->v_iflag &= ~VI_OWEINACT;
3110 		vdropl(vp);
3111 		return;
3112 	}
3113 	vlazy(vp);
3114 	vp->v_iflag |= VI_DEFINACT;
3115 	VI_UNLOCK(vp);
3116 	counter_u64_add(deferred_inact, 1);
3117 }
3118 
3119 static void
3120 vdefer_inactive_unlocked(struct vnode *vp)
3121 {
3122 
3123 	VI_LOCK(vp);
3124 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3125 		vdropl(vp);
3126 		return;
3127 	}
3128 	vdefer_inactive(vp);
3129 }
3130 
3131 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
3132 
3133 /*
3134  * Decrement the use and hold counts for a vnode.
3135  *
3136  * See an explanation near vget() as to why atomic operation is safe.
3137  *
3138  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3139  * on the lock being held all the way until VOP_INACTIVE. This in particular
3140  * happens with UFS which adds half-constructed vnodes to the hash, where they
3141  * can be found by other code.
3142  */
3143 static void
3144 vputx(struct vnode *vp, enum vputx_op func)
3145 {
3146 	int error;
3147 
3148 	KASSERT(vp != NULL, ("vputx: null vp"));
3149 	if (func == VPUTX_VUNREF)
3150 		ASSERT_VOP_LOCKED(vp, "vunref");
3151 	else if (func == VPUTX_VPUT)
3152 		ASSERT_VOP_LOCKED(vp, "vput");
3153 	ASSERT_VI_UNLOCKED(vp, __func__);
3154 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
3155 	    ("%s: wrong ref counts", __func__));
3156 
3157 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3158 
3159 	/*
3160 	 * We want to hold the vnode until the inactive finishes to
3161 	 * prevent vgone() races.  We drop the use count here and the
3162 	 * hold count below when we're done.
3163 	 *
3164 	 * If we release the last usecount we take ownership of the hold
3165 	 * count which provides liveness of the vnode, in which case we
3166 	 * have to vdrop.
3167 	 */
3168 	if (!refcount_release(&vp->v_usecount)) {
3169 		if (func == VPUTX_VPUT)
3170 			VOP_UNLOCK(vp);
3171 		return;
3172 	}
3173 	VI_LOCK(vp);
3174 	v_decr_devcount(vp);
3175 	/*
3176 	 * By the time we got here someone else might have transitioned
3177 	 * the count back to > 0.
3178 	 */
3179 	if (vp->v_usecount > 0 || vp->v_iflag & VI_DOINGINACT)
3180 		goto out;
3181 
3182 	/*
3183 	 * Check if the fs wants to perform inactive processing. Note we
3184 	 * may be only holding the interlock, in which case it is possible
3185 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3186 	 * Since vgone performs inactive on its own there is nothing to do
3187 	 * here but to drop our hold count.
3188 	 */
3189 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3190 	    VOP_NEED_INACTIVE(vp) == 0)
3191 		goto out;
3192 
3193 	/*
3194 	 * We must call VOP_INACTIVE with the node locked. Mark
3195 	 * as VI_DOINGINACT to avoid recursion.
3196 	 */
3197 	vp->v_iflag |= VI_OWEINACT;
3198 	switch (func) {
3199 	case VPUTX_VRELE:
3200 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3201 		VI_LOCK(vp);
3202 		break;
3203 	case VPUTX_VPUT:
3204 		error = 0;
3205 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3206 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3207 			    LK_NOWAIT);
3208 			VI_LOCK(vp);
3209 		}
3210 		break;
3211 	case VPUTX_VUNREF:
3212 		error = 0;
3213 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3214 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3215 			VI_LOCK(vp);
3216 		}
3217 		break;
3218 	}
3219 	if (error == 0) {
3220 		vinactive(vp);
3221 		if (func != VPUTX_VUNREF)
3222 			VOP_UNLOCK(vp);
3223 		vdropl(vp);
3224 	} else {
3225 		vdefer_inactive(vp);
3226 	}
3227 	return;
3228 out:
3229 	if (func == VPUTX_VPUT)
3230 		VOP_UNLOCK(vp);
3231 	vdropl(vp);
3232 }
3233 
3234 /*
3235  * Vnode put/release.
3236  * If count drops to zero, call inactive routine and return to freelist.
3237  */
3238 void
3239 vrele(struct vnode *vp)
3240 {
3241 
3242 	vputx(vp, VPUTX_VRELE);
3243 }
3244 
3245 /*
3246  * Release an already locked vnode.  This give the same effects as
3247  * unlock+vrele(), but takes less time and avoids releasing and
3248  * re-aquiring the lock (as vrele() acquires the lock internally.)
3249  */
3250 void
3251 vput(struct vnode *vp)
3252 {
3253 
3254 	vputx(vp, VPUTX_VPUT);
3255 }
3256 
3257 /*
3258  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3259  */
3260 void
3261 vunref(struct vnode *vp)
3262 {
3263 
3264 	vputx(vp, VPUTX_VUNREF);
3265 }
3266 
3267 void
3268 vhold(struct vnode *vp)
3269 {
3270 	struct vdbatch *vd;
3271 	int old;
3272 
3273 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3274 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3275 	VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3276 	if (old != 0)
3277 		return;
3278 	critical_enter();
3279 	vd = DPCPU_PTR(vd);
3280 	vd->freevnodes--;
3281 	critical_exit();
3282 }
3283 
3284 void
3285 vholdl(struct vnode *vp)
3286 {
3287 
3288 	ASSERT_VI_LOCKED(vp, __func__);
3289 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3290 	vhold(vp);
3291 }
3292 
3293 void
3294 vholdnz(struct vnode *vp)
3295 {
3296 
3297 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3298 #ifdef INVARIANTS
3299 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3300 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3301 #else
3302 	atomic_add_int(&vp->v_holdcnt, 1);
3303 #endif
3304 }
3305 
3306 static void __noinline
3307 vdbatch_process(struct vdbatch *vd)
3308 {
3309 	struct vnode *vp;
3310 	int i;
3311 
3312 	mtx_assert(&vd->lock, MA_OWNED);
3313 	MPASS(curthread->td_pinned > 0);
3314 	MPASS(vd->index == VDBATCH_SIZE);
3315 
3316 	mtx_lock(&vnode_list_mtx);
3317 	critical_enter();
3318 	freevnodes += vd->freevnodes;
3319 	for (i = 0; i < VDBATCH_SIZE; i++) {
3320 		vp = vd->tab[i];
3321 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3322 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3323 		MPASS(vp->v_dbatchcpu != NOCPU);
3324 		vp->v_dbatchcpu = NOCPU;
3325 	}
3326 	mtx_unlock(&vnode_list_mtx);
3327 	vd->freevnodes = 0;
3328 	bzero(vd->tab, sizeof(vd->tab));
3329 	vd->index = 0;
3330 	critical_exit();
3331 }
3332 
3333 static void
3334 vdbatch_enqueue(struct vnode *vp)
3335 {
3336 	struct vdbatch *vd;
3337 
3338 	ASSERT_VI_LOCKED(vp, __func__);
3339 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3340 	    ("%s: deferring requeue of a doomed vnode", __func__));
3341 
3342 	critical_enter();
3343 	vd = DPCPU_PTR(vd);
3344 	vd->freevnodes++;
3345 	if (vp->v_dbatchcpu != NOCPU) {
3346 		VI_UNLOCK(vp);
3347 		critical_exit();
3348 		return;
3349 	}
3350 
3351 	sched_pin();
3352 	critical_exit();
3353 	mtx_lock(&vd->lock);
3354 	MPASS(vd->index < VDBATCH_SIZE);
3355 	MPASS(vd->tab[vd->index] == NULL);
3356 	/*
3357 	 * A hack: we depend on being pinned so that we know what to put in
3358 	 * ->v_dbatchcpu.
3359 	 */
3360 	vp->v_dbatchcpu = curcpu;
3361 	vd->tab[vd->index] = vp;
3362 	vd->index++;
3363 	VI_UNLOCK(vp);
3364 	if (vd->index == VDBATCH_SIZE)
3365 		vdbatch_process(vd);
3366 	mtx_unlock(&vd->lock);
3367 	sched_unpin();
3368 }
3369 
3370 /*
3371  * This routine must only be called for vnodes which are about to be
3372  * deallocated. Supporting dequeue for arbitrary vndoes would require
3373  * validating that the locked batch matches.
3374  */
3375 static void
3376 vdbatch_dequeue(struct vnode *vp)
3377 {
3378 	struct vdbatch *vd;
3379 	int i;
3380 	short cpu;
3381 
3382 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3383 	    ("%s: called for a used vnode\n", __func__));
3384 
3385 	cpu = vp->v_dbatchcpu;
3386 	if (cpu == NOCPU)
3387 		return;
3388 
3389 	vd = DPCPU_ID_PTR(cpu, vd);
3390 	mtx_lock(&vd->lock);
3391 	for (i = 0; i < vd->index; i++) {
3392 		if (vd->tab[i] != vp)
3393 			continue;
3394 		vp->v_dbatchcpu = NOCPU;
3395 		vd->index--;
3396 		vd->tab[i] = vd->tab[vd->index];
3397 		vd->tab[vd->index] = NULL;
3398 		break;
3399 	}
3400 	mtx_unlock(&vd->lock);
3401 	/*
3402 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3403 	 */
3404 	MPASS(vp->v_dbatchcpu == NOCPU);
3405 }
3406 
3407 /*
3408  * Drop the hold count of the vnode.  If this is the last reference to
3409  * the vnode we place it on the free list unless it has been vgone'd
3410  * (marked VIRF_DOOMED) in which case we will free it.
3411  *
3412  * Because the vnode vm object keeps a hold reference on the vnode if
3413  * there is at least one resident non-cached page, the vnode cannot
3414  * leave the active list without the page cleanup done.
3415  */
3416 static void
3417 vdrop_deactivate(struct vnode *vp)
3418 {
3419 	struct mount *mp;
3420 
3421 	ASSERT_VI_LOCKED(vp, __func__);
3422 	/*
3423 	 * Mark a vnode as free: remove it from its active list
3424 	 * and put it up for recycling on the freelist.
3425 	 */
3426 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3427 	    ("vdrop: returning doomed vnode"));
3428 	VNASSERT(vp->v_op != NULL, vp,
3429 	    ("vdrop: vnode already reclaimed."));
3430 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3431 	    ("vnode with VI_OWEINACT set"));
3432 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3433 	    ("vnode with VI_DEFINACT set"));
3434 	if (vp->v_mflag & VMP_LAZYLIST) {
3435 		mp = vp->v_mount;
3436 		mtx_lock(&mp->mnt_listmtx);
3437 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3438 		/*
3439 		 * Don't remove the vnode from the lazy list if another thread
3440 		 * has increased the hold count. It may have re-enqueued the
3441 		 * vnode to the lazy list and is now responsible for its
3442 		 * removal.
3443 		 */
3444 		if (vp->v_holdcnt == 0) {
3445 			vp->v_mflag &= ~VMP_LAZYLIST;
3446 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3447 			mp->mnt_lazyvnodelistsize--;
3448 		}
3449 		mtx_unlock(&mp->mnt_listmtx);
3450 	}
3451 	vdbatch_enqueue(vp);
3452 }
3453 
3454 void
3455 vdrop(struct vnode *vp)
3456 {
3457 
3458 	ASSERT_VI_UNLOCKED(vp, __func__);
3459 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3460 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3461 		return;
3462 	VI_LOCK(vp);
3463 	vdropl(vp);
3464 }
3465 
3466 void
3467 vdropl(struct vnode *vp)
3468 {
3469 
3470 	ASSERT_VI_LOCKED(vp, __func__);
3471 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3472 	if (!refcount_release(&vp->v_holdcnt)) {
3473 		VI_UNLOCK(vp);
3474 		return;
3475 	}
3476 	if (VN_IS_DOOMED(vp)) {
3477 		freevnode(vp);
3478 		return;
3479 	}
3480 	vdrop_deactivate(vp);
3481 }
3482 
3483 /*
3484  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3485  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3486  */
3487 static void
3488 vinactivef(struct vnode *vp)
3489 {
3490 	struct vm_object *obj;
3491 
3492 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3493 	ASSERT_VI_LOCKED(vp, "vinactive");
3494 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3495 	    ("vinactive: recursed on VI_DOINGINACT"));
3496 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3497 	vp->v_iflag |= VI_DOINGINACT;
3498 	vp->v_iflag &= ~VI_OWEINACT;
3499 	VI_UNLOCK(vp);
3500 	/*
3501 	 * Before moving off the active list, we must be sure that any
3502 	 * modified pages are converted into the vnode's dirty
3503 	 * buffers, since these will no longer be checked once the
3504 	 * vnode is on the inactive list.
3505 	 *
3506 	 * The write-out of the dirty pages is asynchronous.  At the
3507 	 * point that VOP_INACTIVE() is called, there could still be
3508 	 * pending I/O and dirty pages in the object.
3509 	 */
3510 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3511 	    vm_object_mightbedirty(obj)) {
3512 		VM_OBJECT_WLOCK(obj);
3513 		vm_object_page_clean(obj, 0, 0, 0);
3514 		VM_OBJECT_WUNLOCK(obj);
3515 	}
3516 	VOP_INACTIVE(vp, curthread);
3517 	VI_LOCK(vp);
3518 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3519 	    ("vinactive: lost VI_DOINGINACT"));
3520 	vp->v_iflag &= ~VI_DOINGINACT;
3521 }
3522 
3523 void
3524 vinactive(struct vnode *vp)
3525 {
3526 
3527 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3528 	ASSERT_VI_LOCKED(vp, "vinactive");
3529 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3530 
3531 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3532 		return;
3533 	if (vp->v_iflag & VI_DOINGINACT)
3534 		return;
3535 	if (vp->v_usecount > 0) {
3536 		vp->v_iflag &= ~VI_OWEINACT;
3537 		return;
3538 	}
3539 	vinactivef(vp);
3540 }
3541 
3542 /*
3543  * Remove any vnodes in the vnode table belonging to mount point mp.
3544  *
3545  * If FORCECLOSE is not specified, there should not be any active ones,
3546  * return error if any are found (nb: this is a user error, not a
3547  * system error). If FORCECLOSE is specified, detach any active vnodes
3548  * that are found.
3549  *
3550  * If WRITECLOSE is set, only flush out regular file vnodes open for
3551  * writing.
3552  *
3553  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3554  *
3555  * `rootrefs' specifies the base reference count for the root vnode
3556  * of this filesystem. The root vnode is considered busy if its
3557  * v_usecount exceeds this value. On a successful return, vflush(, td)
3558  * will call vrele() on the root vnode exactly rootrefs times.
3559  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3560  * be zero.
3561  */
3562 #ifdef DIAGNOSTIC
3563 static int busyprt = 0;		/* print out busy vnodes */
3564 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3565 #endif
3566 
3567 int
3568 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3569 {
3570 	struct vnode *vp, *mvp, *rootvp = NULL;
3571 	struct vattr vattr;
3572 	int busy = 0, error;
3573 
3574 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3575 	    rootrefs, flags);
3576 	if (rootrefs > 0) {
3577 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3578 		    ("vflush: bad args"));
3579 		/*
3580 		 * Get the filesystem root vnode. We can vput() it
3581 		 * immediately, since with rootrefs > 0, it won't go away.
3582 		 */
3583 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3584 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3585 			    __func__, error);
3586 			return (error);
3587 		}
3588 		vput(rootvp);
3589 	}
3590 loop:
3591 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3592 		vholdl(vp);
3593 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3594 		if (error) {
3595 			vdrop(vp);
3596 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3597 			goto loop;
3598 		}
3599 		/*
3600 		 * Skip over a vnodes marked VV_SYSTEM.
3601 		 */
3602 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3603 			VOP_UNLOCK(vp);
3604 			vdrop(vp);
3605 			continue;
3606 		}
3607 		/*
3608 		 * If WRITECLOSE is set, flush out unlinked but still open
3609 		 * files (even if open only for reading) and regular file
3610 		 * vnodes open for writing.
3611 		 */
3612 		if (flags & WRITECLOSE) {
3613 			if (vp->v_object != NULL) {
3614 				VM_OBJECT_WLOCK(vp->v_object);
3615 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3616 				VM_OBJECT_WUNLOCK(vp->v_object);
3617 			}
3618 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3619 			if (error != 0) {
3620 				VOP_UNLOCK(vp);
3621 				vdrop(vp);
3622 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3623 				return (error);
3624 			}
3625 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3626 			VI_LOCK(vp);
3627 
3628 			if ((vp->v_type == VNON ||
3629 			    (error == 0 && vattr.va_nlink > 0)) &&
3630 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3631 				VOP_UNLOCK(vp);
3632 				vdropl(vp);
3633 				continue;
3634 			}
3635 		} else
3636 			VI_LOCK(vp);
3637 		/*
3638 		 * With v_usecount == 0, all we need to do is clear out the
3639 		 * vnode data structures and we are done.
3640 		 *
3641 		 * If FORCECLOSE is set, forcibly close the vnode.
3642 		 */
3643 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3644 			vgonel(vp);
3645 		} else {
3646 			busy++;
3647 #ifdef DIAGNOSTIC
3648 			if (busyprt)
3649 				vn_printf(vp, "vflush: busy vnode ");
3650 #endif
3651 		}
3652 		VOP_UNLOCK(vp);
3653 		vdropl(vp);
3654 	}
3655 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3656 		/*
3657 		 * If just the root vnode is busy, and if its refcount
3658 		 * is equal to `rootrefs', then go ahead and kill it.
3659 		 */
3660 		VI_LOCK(rootvp);
3661 		KASSERT(busy > 0, ("vflush: not busy"));
3662 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3663 		    ("vflush: usecount %d < rootrefs %d",
3664 		     rootvp->v_usecount, rootrefs));
3665 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3666 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3667 			vgone(rootvp);
3668 			VOP_UNLOCK(rootvp);
3669 			busy = 0;
3670 		} else
3671 			VI_UNLOCK(rootvp);
3672 	}
3673 	if (busy) {
3674 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3675 		    busy);
3676 		return (EBUSY);
3677 	}
3678 	for (; rootrefs > 0; rootrefs--)
3679 		vrele(rootvp);
3680 	return (0);
3681 }
3682 
3683 /*
3684  * Recycle an unused vnode to the front of the free list.
3685  */
3686 int
3687 vrecycle(struct vnode *vp)
3688 {
3689 	int recycled;
3690 
3691 	VI_LOCK(vp);
3692 	recycled = vrecyclel(vp);
3693 	VI_UNLOCK(vp);
3694 	return (recycled);
3695 }
3696 
3697 /*
3698  * vrecycle, with the vp interlock held.
3699  */
3700 int
3701 vrecyclel(struct vnode *vp)
3702 {
3703 	int recycled;
3704 
3705 	ASSERT_VOP_ELOCKED(vp, __func__);
3706 	ASSERT_VI_LOCKED(vp, __func__);
3707 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3708 	recycled = 0;
3709 	if (vp->v_usecount == 0) {
3710 		recycled = 1;
3711 		vgonel(vp);
3712 	}
3713 	return (recycled);
3714 }
3715 
3716 /*
3717  * Eliminate all activity associated with a vnode
3718  * in preparation for reuse.
3719  */
3720 void
3721 vgone(struct vnode *vp)
3722 {
3723 	VI_LOCK(vp);
3724 	vgonel(vp);
3725 	VI_UNLOCK(vp);
3726 }
3727 
3728 static void
3729 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3730     struct vnode *lowervp __unused)
3731 {
3732 }
3733 
3734 /*
3735  * Notify upper mounts about reclaimed or unlinked vnode.
3736  */
3737 void
3738 vfs_notify_upper(struct vnode *vp, int event)
3739 {
3740 	static struct vfsops vgonel_vfsops = {
3741 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3742 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3743 	};
3744 	struct mount *mp, *ump, *mmp;
3745 
3746 	mp = vp->v_mount;
3747 	if (mp == NULL)
3748 		return;
3749 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3750 		return;
3751 
3752 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3753 	mmp->mnt_op = &vgonel_vfsops;
3754 	mmp->mnt_kern_flag |= MNTK_MARKER;
3755 	MNT_ILOCK(mp);
3756 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3757 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3758 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3759 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3760 			continue;
3761 		}
3762 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3763 		MNT_IUNLOCK(mp);
3764 		switch (event) {
3765 		case VFS_NOTIFY_UPPER_RECLAIM:
3766 			VFS_RECLAIM_LOWERVP(ump, vp);
3767 			break;
3768 		case VFS_NOTIFY_UPPER_UNLINK:
3769 			VFS_UNLINK_LOWERVP(ump, vp);
3770 			break;
3771 		default:
3772 			KASSERT(0, ("invalid event %d", event));
3773 			break;
3774 		}
3775 		MNT_ILOCK(mp);
3776 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3777 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3778 	}
3779 	free(mmp, M_TEMP);
3780 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3781 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3782 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3783 		wakeup(&mp->mnt_uppers);
3784 	}
3785 	MNT_IUNLOCK(mp);
3786 }
3787 
3788 /*
3789  * vgone, with the vp interlock held.
3790  */
3791 static void
3792 vgonel(struct vnode *vp)
3793 {
3794 	struct thread *td;
3795 	struct mount *mp;
3796 	vm_object_t object;
3797 	bool active, oweinact;
3798 
3799 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3800 	ASSERT_VI_LOCKED(vp, "vgonel");
3801 	VNASSERT(vp->v_holdcnt, vp,
3802 	    ("vgonel: vp %p has no reference.", vp));
3803 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3804 	td = curthread;
3805 
3806 	/*
3807 	 * Don't vgonel if we're already doomed.
3808 	 */
3809 	if (vp->v_irflag & VIRF_DOOMED)
3810 		return;
3811 	vp->v_irflag |= VIRF_DOOMED;
3812 
3813 	/*
3814 	 * Check to see if the vnode is in use.  If so, we have to call
3815 	 * VOP_CLOSE() and VOP_INACTIVE().
3816 	 */
3817 	active = vp->v_usecount > 0;
3818 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3819 	/*
3820 	 * If we need to do inactive VI_OWEINACT will be set.
3821 	 */
3822 	if (vp->v_iflag & VI_DEFINACT) {
3823 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3824 		vp->v_iflag &= ~VI_DEFINACT;
3825 		vdropl(vp);
3826 	} else {
3827 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3828 		VI_UNLOCK(vp);
3829 	}
3830 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3831 
3832 	/*
3833 	 * If purging an active vnode, it must be closed and
3834 	 * deactivated before being reclaimed.
3835 	 */
3836 	if (active)
3837 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3838 	if (oweinact || active) {
3839 		VI_LOCK(vp);
3840 		vinactivef(vp);
3841 		VI_UNLOCK(vp);
3842 	}
3843 	if (vp->v_type == VSOCK)
3844 		vfs_unp_reclaim(vp);
3845 
3846 	/*
3847 	 * Clean out any buffers associated with the vnode.
3848 	 * If the flush fails, just toss the buffers.
3849 	 */
3850 	mp = NULL;
3851 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3852 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3853 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3854 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3855 			;
3856 	}
3857 
3858 	BO_LOCK(&vp->v_bufobj);
3859 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3860 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3861 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3862 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3863 	    ("vp %p bufobj not invalidated", vp));
3864 
3865 	/*
3866 	 * For VMIO bufobj, BO_DEAD is set later, or in
3867 	 * vm_object_terminate() after the object's page queue is
3868 	 * flushed.
3869 	 */
3870 	object = vp->v_bufobj.bo_object;
3871 	if (object == NULL)
3872 		vp->v_bufobj.bo_flag |= BO_DEAD;
3873 	BO_UNLOCK(&vp->v_bufobj);
3874 
3875 	/*
3876 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3877 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3878 	 * should not touch the object borrowed from the lower vnode
3879 	 * (the handle check).
3880 	 */
3881 	if (object != NULL && object->type == OBJT_VNODE &&
3882 	    object->handle == vp)
3883 		vnode_destroy_vobject(vp);
3884 
3885 	/*
3886 	 * Reclaim the vnode.
3887 	 */
3888 	if (VOP_RECLAIM(vp, td))
3889 		panic("vgone: cannot reclaim");
3890 	if (mp != NULL)
3891 		vn_finished_secondary_write(mp);
3892 	VNASSERT(vp->v_object == NULL, vp,
3893 	    ("vop_reclaim left v_object vp=%p", vp));
3894 	/*
3895 	 * Clear the advisory locks and wake up waiting threads.
3896 	 */
3897 	(void)VOP_ADVLOCKPURGE(vp);
3898 	vp->v_lockf = NULL;
3899 	/*
3900 	 * Delete from old mount point vnode list.
3901 	 */
3902 	delmntque(vp);
3903 	cache_purge(vp);
3904 	/*
3905 	 * Done with purge, reset to the standard lock and invalidate
3906 	 * the vnode.
3907 	 */
3908 	VI_LOCK(vp);
3909 	vp->v_vnlock = &vp->v_lock;
3910 	vp->v_op = &dead_vnodeops;
3911 	vp->v_type = VBAD;
3912 }
3913 
3914 /*
3915  * Calculate the total number of references to a special device.
3916  */
3917 int
3918 vcount(struct vnode *vp)
3919 {
3920 	int count;
3921 
3922 	dev_lock();
3923 	count = vp->v_rdev->si_usecount;
3924 	dev_unlock();
3925 	return (count);
3926 }
3927 
3928 /*
3929  * Print out a description of a vnode.
3930  */
3931 static char *typename[] =
3932 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3933  "VMARKER"};
3934 
3935 void
3936 vn_printf(struct vnode *vp, const char *fmt, ...)
3937 {
3938 	va_list ap;
3939 	char buf[256], buf2[16];
3940 	u_long flags;
3941 
3942 	va_start(ap, fmt);
3943 	vprintf(fmt, ap);
3944 	va_end(ap);
3945 	printf("%p: ", (void *)vp);
3946 	printf("type %s\n", typename[vp->v_type]);
3947 	printf("    usecount %d, writecount %d, refcount %d",
3948 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3949 	switch (vp->v_type) {
3950 	case VDIR:
3951 		printf(" mountedhere %p\n", vp->v_mountedhere);
3952 		break;
3953 	case VCHR:
3954 		printf(" rdev %p\n", vp->v_rdev);
3955 		break;
3956 	case VSOCK:
3957 		printf(" socket %p\n", vp->v_unpcb);
3958 		break;
3959 	case VFIFO:
3960 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3961 		break;
3962 	default:
3963 		printf("\n");
3964 		break;
3965 	}
3966 	buf[0] = '\0';
3967 	buf[1] = '\0';
3968 	if (vp->v_irflag & VIRF_DOOMED)
3969 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3970 	flags = vp->v_irflag & ~(VIRF_DOOMED);
3971 	if (flags != 0) {
3972 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3973 		strlcat(buf, buf2, sizeof(buf));
3974 	}
3975 	if (vp->v_vflag & VV_ROOT)
3976 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3977 	if (vp->v_vflag & VV_ISTTY)
3978 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3979 	if (vp->v_vflag & VV_NOSYNC)
3980 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3981 	if (vp->v_vflag & VV_ETERNALDEV)
3982 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3983 	if (vp->v_vflag & VV_CACHEDLABEL)
3984 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3985 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3986 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3987 	if (vp->v_vflag & VV_COPYONWRITE)
3988 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3989 	if (vp->v_vflag & VV_SYSTEM)
3990 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3991 	if (vp->v_vflag & VV_PROCDEP)
3992 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3993 	if (vp->v_vflag & VV_NOKNOTE)
3994 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3995 	if (vp->v_vflag & VV_DELETED)
3996 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3997 	if (vp->v_vflag & VV_MD)
3998 		strlcat(buf, "|VV_MD", sizeof(buf));
3999 	if (vp->v_vflag & VV_FORCEINSMQ)
4000 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4001 	if (vp->v_vflag & VV_READLINK)
4002 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4003 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4004 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4005 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4006 	if (flags != 0) {
4007 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4008 		strlcat(buf, buf2, sizeof(buf));
4009 	}
4010 	if (vp->v_iflag & VI_TEXT_REF)
4011 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4012 	if (vp->v_iflag & VI_MOUNT)
4013 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4014 	if (vp->v_iflag & VI_DOINGINACT)
4015 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4016 	if (vp->v_iflag & VI_OWEINACT)
4017 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4018 	if (vp->v_iflag & VI_DEFINACT)
4019 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4020 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4021 	    VI_OWEINACT | VI_DEFINACT);
4022 	if (flags != 0) {
4023 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4024 		strlcat(buf, buf2, sizeof(buf));
4025 	}
4026 	if (vp->v_mflag & VMP_LAZYLIST)
4027 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4028 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4029 	if (flags != 0) {
4030 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4031 		strlcat(buf, buf2, sizeof(buf));
4032 	}
4033 	printf("    flags (%s)\n", buf + 1);
4034 	if (mtx_owned(VI_MTX(vp)))
4035 		printf(" VI_LOCKed");
4036 	if (vp->v_object != NULL)
4037 		printf("    v_object %p ref %d pages %d "
4038 		    "cleanbuf %d dirtybuf %d\n",
4039 		    vp->v_object, vp->v_object->ref_count,
4040 		    vp->v_object->resident_page_count,
4041 		    vp->v_bufobj.bo_clean.bv_cnt,
4042 		    vp->v_bufobj.bo_dirty.bv_cnt);
4043 	printf("    ");
4044 	lockmgr_printinfo(vp->v_vnlock);
4045 	if (vp->v_data != NULL)
4046 		VOP_PRINT(vp);
4047 }
4048 
4049 #ifdef DDB
4050 /*
4051  * List all of the locked vnodes in the system.
4052  * Called when debugging the kernel.
4053  */
4054 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4055 {
4056 	struct mount *mp;
4057 	struct vnode *vp;
4058 
4059 	/*
4060 	 * Note: because this is DDB, we can't obey the locking semantics
4061 	 * for these structures, which means we could catch an inconsistent
4062 	 * state and dereference a nasty pointer.  Not much to be done
4063 	 * about that.
4064 	 */
4065 	db_printf("Locked vnodes\n");
4066 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4067 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4068 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4069 				vn_printf(vp, "vnode ");
4070 		}
4071 	}
4072 }
4073 
4074 /*
4075  * Show details about the given vnode.
4076  */
4077 DB_SHOW_COMMAND(vnode, db_show_vnode)
4078 {
4079 	struct vnode *vp;
4080 
4081 	if (!have_addr)
4082 		return;
4083 	vp = (struct vnode *)addr;
4084 	vn_printf(vp, "vnode ");
4085 }
4086 
4087 /*
4088  * Show details about the given mount point.
4089  */
4090 DB_SHOW_COMMAND(mount, db_show_mount)
4091 {
4092 	struct mount *mp;
4093 	struct vfsopt *opt;
4094 	struct statfs *sp;
4095 	struct vnode *vp;
4096 	char buf[512];
4097 	uint64_t mflags;
4098 	u_int flags;
4099 
4100 	if (!have_addr) {
4101 		/* No address given, print short info about all mount points. */
4102 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4103 			db_printf("%p %s on %s (%s)\n", mp,
4104 			    mp->mnt_stat.f_mntfromname,
4105 			    mp->mnt_stat.f_mntonname,
4106 			    mp->mnt_stat.f_fstypename);
4107 			if (db_pager_quit)
4108 				break;
4109 		}
4110 		db_printf("\nMore info: show mount <addr>\n");
4111 		return;
4112 	}
4113 
4114 	mp = (struct mount *)addr;
4115 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4116 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4117 
4118 	buf[0] = '\0';
4119 	mflags = mp->mnt_flag;
4120 #define	MNT_FLAG(flag)	do {						\
4121 	if (mflags & (flag)) {						\
4122 		if (buf[0] != '\0')					\
4123 			strlcat(buf, ", ", sizeof(buf));		\
4124 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4125 		mflags &= ~(flag);					\
4126 	}								\
4127 } while (0)
4128 	MNT_FLAG(MNT_RDONLY);
4129 	MNT_FLAG(MNT_SYNCHRONOUS);
4130 	MNT_FLAG(MNT_NOEXEC);
4131 	MNT_FLAG(MNT_NOSUID);
4132 	MNT_FLAG(MNT_NFS4ACLS);
4133 	MNT_FLAG(MNT_UNION);
4134 	MNT_FLAG(MNT_ASYNC);
4135 	MNT_FLAG(MNT_SUIDDIR);
4136 	MNT_FLAG(MNT_SOFTDEP);
4137 	MNT_FLAG(MNT_NOSYMFOLLOW);
4138 	MNT_FLAG(MNT_GJOURNAL);
4139 	MNT_FLAG(MNT_MULTILABEL);
4140 	MNT_FLAG(MNT_ACLS);
4141 	MNT_FLAG(MNT_NOATIME);
4142 	MNT_FLAG(MNT_NOCLUSTERR);
4143 	MNT_FLAG(MNT_NOCLUSTERW);
4144 	MNT_FLAG(MNT_SUJ);
4145 	MNT_FLAG(MNT_EXRDONLY);
4146 	MNT_FLAG(MNT_EXPORTED);
4147 	MNT_FLAG(MNT_DEFEXPORTED);
4148 	MNT_FLAG(MNT_EXPORTANON);
4149 	MNT_FLAG(MNT_EXKERB);
4150 	MNT_FLAG(MNT_EXPUBLIC);
4151 	MNT_FLAG(MNT_LOCAL);
4152 	MNT_FLAG(MNT_QUOTA);
4153 	MNT_FLAG(MNT_ROOTFS);
4154 	MNT_FLAG(MNT_USER);
4155 	MNT_FLAG(MNT_IGNORE);
4156 	MNT_FLAG(MNT_UPDATE);
4157 	MNT_FLAG(MNT_DELEXPORT);
4158 	MNT_FLAG(MNT_RELOAD);
4159 	MNT_FLAG(MNT_FORCE);
4160 	MNT_FLAG(MNT_SNAPSHOT);
4161 	MNT_FLAG(MNT_BYFSID);
4162 #undef MNT_FLAG
4163 	if (mflags != 0) {
4164 		if (buf[0] != '\0')
4165 			strlcat(buf, ", ", sizeof(buf));
4166 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4167 		    "0x%016jx", mflags);
4168 	}
4169 	db_printf("    mnt_flag = %s\n", buf);
4170 
4171 	buf[0] = '\0';
4172 	flags = mp->mnt_kern_flag;
4173 #define	MNT_KERN_FLAG(flag)	do {					\
4174 	if (flags & (flag)) {						\
4175 		if (buf[0] != '\0')					\
4176 			strlcat(buf, ", ", sizeof(buf));		\
4177 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4178 		flags &= ~(flag);					\
4179 	}								\
4180 } while (0)
4181 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4182 	MNT_KERN_FLAG(MNTK_ASYNC);
4183 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4184 	MNT_KERN_FLAG(MNTK_DRAINING);
4185 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4186 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4187 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4188 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4189 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4190 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4191 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4192 	MNT_KERN_FLAG(MNTK_MARKER);
4193 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4194 	MNT_KERN_FLAG(MNTK_NOASYNC);
4195 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4196 	MNT_KERN_FLAG(MNTK_MWAIT);
4197 	MNT_KERN_FLAG(MNTK_SUSPEND);
4198 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4199 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4200 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4201 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4202 #undef MNT_KERN_FLAG
4203 	if (flags != 0) {
4204 		if (buf[0] != '\0')
4205 			strlcat(buf, ", ", sizeof(buf));
4206 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4207 		    "0x%08x", flags);
4208 	}
4209 	db_printf("    mnt_kern_flag = %s\n", buf);
4210 
4211 	db_printf("    mnt_opt = ");
4212 	opt = TAILQ_FIRST(mp->mnt_opt);
4213 	if (opt != NULL) {
4214 		db_printf("%s", opt->name);
4215 		opt = TAILQ_NEXT(opt, link);
4216 		while (opt != NULL) {
4217 			db_printf(", %s", opt->name);
4218 			opt = TAILQ_NEXT(opt, link);
4219 		}
4220 	}
4221 	db_printf("\n");
4222 
4223 	sp = &mp->mnt_stat;
4224 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4225 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4226 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4227 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4228 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4229 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4230 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4231 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4232 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4233 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4234 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4235 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4236 
4237 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4238 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4239 	if (jailed(mp->mnt_cred))
4240 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4241 	db_printf(" }\n");
4242 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4243 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4244 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4245 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4246 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4247 	    mp->mnt_lazyvnodelistsize);
4248 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4249 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4250 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4251 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4252 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4253 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4254 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4255 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4256 	db_printf("    mnt_secondary_accwrites = %d\n",
4257 	    mp->mnt_secondary_accwrites);
4258 	db_printf("    mnt_gjprovider = %s\n",
4259 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4260 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4261 
4262 	db_printf("\n\nList of active vnodes\n");
4263 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4264 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4265 			vn_printf(vp, "vnode ");
4266 			if (db_pager_quit)
4267 				break;
4268 		}
4269 	}
4270 	db_printf("\n\nList of inactive vnodes\n");
4271 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4272 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4273 			vn_printf(vp, "vnode ");
4274 			if (db_pager_quit)
4275 				break;
4276 		}
4277 	}
4278 }
4279 #endif	/* DDB */
4280 
4281 /*
4282  * Fill in a struct xvfsconf based on a struct vfsconf.
4283  */
4284 static int
4285 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4286 {
4287 	struct xvfsconf xvfsp;
4288 
4289 	bzero(&xvfsp, sizeof(xvfsp));
4290 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4291 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4292 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4293 	xvfsp.vfc_flags = vfsp->vfc_flags;
4294 	/*
4295 	 * These are unused in userland, we keep them
4296 	 * to not break binary compatibility.
4297 	 */
4298 	xvfsp.vfc_vfsops = NULL;
4299 	xvfsp.vfc_next = NULL;
4300 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4301 }
4302 
4303 #ifdef COMPAT_FREEBSD32
4304 struct xvfsconf32 {
4305 	uint32_t	vfc_vfsops;
4306 	char		vfc_name[MFSNAMELEN];
4307 	int32_t		vfc_typenum;
4308 	int32_t		vfc_refcount;
4309 	int32_t		vfc_flags;
4310 	uint32_t	vfc_next;
4311 };
4312 
4313 static int
4314 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4315 {
4316 	struct xvfsconf32 xvfsp;
4317 
4318 	bzero(&xvfsp, sizeof(xvfsp));
4319 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4320 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4321 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4322 	xvfsp.vfc_flags = vfsp->vfc_flags;
4323 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4324 }
4325 #endif
4326 
4327 /*
4328  * Top level filesystem related information gathering.
4329  */
4330 static int
4331 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4332 {
4333 	struct vfsconf *vfsp;
4334 	int error;
4335 
4336 	error = 0;
4337 	vfsconf_slock();
4338 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4339 #ifdef COMPAT_FREEBSD32
4340 		if (req->flags & SCTL_MASK32)
4341 			error = vfsconf2x32(req, vfsp);
4342 		else
4343 #endif
4344 			error = vfsconf2x(req, vfsp);
4345 		if (error)
4346 			break;
4347 	}
4348 	vfsconf_sunlock();
4349 	return (error);
4350 }
4351 
4352 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4353     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4354     "S,xvfsconf", "List of all configured filesystems");
4355 
4356 #ifndef BURN_BRIDGES
4357 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4358 
4359 static int
4360 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4361 {
4362 	int *name = (int *)arg1 - 1;	/* XXX */
4363 	u_int namelen = arg2 + 1;	/* XXX */
4364 	struct vfsconf *vfsp;
4365 
4366 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4367 	    "please rebuild world\n");
4368 
4369 #if 1 || defined(COMPAT_PRELITE2)
4370 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4371 	if (namelen == 1)
4372 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4373 #endif
4374 
4375 	switch (name[1]) {
4376 	case VFS_MAXTYPENUM:
4377 		if (namelen != 2)
4378 			return (ENOTDIR);
4379 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4380 	case VFS_CONF:
4381 		if (namelen != 3)
4382 			return (ENOTDIR);	/* overloaded */
4383 		vfsconf_slock();
4384 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4385 			if (vfsp->vfc_typenum == name[2])
4386 				break;
4387 		}
4388 		vfsconf_sunlock();
4389 		if (vfsp == NULL)
4390 			return (EOPNOTSUPP);
4391 #ifdef COMPAT_FREEBSD32
4392 		if (req->flags & SCTL_MASK32)
4393 			return (vfsconf2x32(req, vfsp));
4394 		else
4395 #endif
4396 			return (vfsconf2x(req, vfsp));
4397 	}
4398 	return (EOPNOTSUPP);
4399 }
4400 
4401 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4402     CTLFLAG_MPSAFE, vfs_sysctl,
4403     "Generic filesystem");
4404 
4405 #if 1 || defined(COMPAT_PRELITE2)
4406 
4407 static int
4408 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4409 {
4410 	int error;
4411 	struct vfsconf *vfsp;
4412 	struct ovfsconf ovfs;
4413 
4414 	vfsconf_slock();
4415 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4416 		bzero(&ovfs, sizeof(ovfs));
4417 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4418 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4419 		ovfs.vfc_index = vfsp->vfc_typenum;
4420 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4421 		ovfs.vfc_flags = vfsp->vfc_flags;
4422 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4423 		if (error != 0) {
4424 			vfsconf_sunlock();
4425 			return (error);
4426 		}
4427 	}
4428 	vfsconf_sunlock();
4429 	return (0);
4430 }
4431 
4432 #endif /* 1 || COMPAT_PRELITE2 */
4433 #endif /* !BURN_BRIDGES */
4434 
4435 #define KINFO_VNODESLOP		10
4436 #ifdef notyet
4437 /*
4438  * Dump vnode list (via sysctl).
4439  */
4440 /* ARGSUSED */
4441 static int
4442 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4443 {
4444 	struct xvnode *xvn;
4445 	struct mount *mp;
4446 	struct vnode *vp;
4447 	int error, len, n;
4448 
4449 	/*
4450 	 * Stale numvnodes access is not fatal here.
4451 	 */
4452 	req->lock = 0;
4453 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4454 	if (!req->oldptr)
4455 		/* Make an estimate */
4456 		return (SYSCTL_OUT(req, 0, len));
4457 
4458 	error = sysctl_wire_old_buffer(req, 0);
4459 	if (error != 0)
4460 		return (error);
4461 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4462 	n = 0;
4463 	mtx_lock(&mountlist_mtx);
4464 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4465 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4466 			continue;
4467 		MNT_ILOCK(mp);
4468 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4469 			if (n == len)
4470 				break;
4471 			vref(vp);
4472 			xvn[n].xv_size = sizeof *xvn;
4473 			xvn[n].xv_vnode = vp;
4474 			xvn[n].xv_id = 0;	/* XXX compat */
4475 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4476 			XV_COPY(usecount);
4477 			XV_COPY(writecount);
4478 			XV_COPY(holdcnt);
4479 			XV_COPY(mount);
4480 			XV_COPY(numoutput);
4481 			XV_COPY(type);
4482 #undef XV_COPY
4483 			xvn[n].xv_flag = vp->v_vflag;
4484 
4485 			switch (vp->v_type) {
4486 			case VREG:
4487 			case VDIR:
4488 			case VLNK:
4489 				break;
4490 			case VBLK:
4491 			case VCHR:
4492 				if (vp->v_rdev == NULL) {
4493 					vrele(vp);
4494 					continue;
4495 				}
4496 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4497 				break;
4498 			case VSOCK:
4499 				xvn[n].xv_socket = vp->v_socket;
4500 				break;
4501 			case VFIFO:
4502 				xvn[n].xv_fifo = vp->v_fifoinfo;
4503 				break;
4504 			case VNON:
4505 			case VBAD:
4506 			default:
4507 				/* shouldn't happen? */
4508 				vrele(vp);
4509 				continue;
4510 			}
4511 			vrele(vp);
4512 			++n;
4513 		}
4514 		MNT_IUNLOCK(mp);
4515 		mtx_lock(&mountlist_mtx);
4516 		vfs_unbusy(mp);
4517 		if (n == len)
4518 			break;
4519 	}
4520 	mtx_unlock(&mountlist_mtx);
4521 
4522 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4523 	free(xvn, M_TEMP);
4524 	return (error);
4525 }
4526 
4527 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4528     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4529     "");
4530 #endif
4531 
4532 static void
4533 unmount_or_warn(struct mount *mp)
4534 {
4535 	int error;
4536 
4537 	error = dounmount(mp, MNT_FORCE, curthread);
4538 	if (error != 0) {
4539 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4540 		if (error == EBUSY)
4541 			printf("BUSY)\n");
4542 		else
4543 			printf("%d)\n", error);
4544 	}
4545 }
4546 
4547 /*
4548  * Unmount all filesystems. The list is traversed in reverse order
4549  * of mounting to avoid dependencies.
4550  */
4551 void
4552 vfs_unmountall(void)
4553 {
4554 	struct mount *mp, *tmp;
4555 
4556 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4557 
4558 	/*
4559 	 * Since this only runs when rebooting, it is not interlocked.
4560 	 */
4561 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4562 		vfs_ref(mp);
4563 
4564 		/*
4565 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4566 		 * unmount of the latter.
4567 		 */
4568 		if (mp == rootdevmp)
4569 			continue;
4570 
4571 		unmount_or_warn(mp);
4572 	}
4573 
4574 	if (rootdevmp != NULL)
4575 		unmount_or_warn(rootdevmp);
4576 }
4577 
4578 static void
4579 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4580 {
4581 
4582 	ASSERT_VI_LOCKED(vp, __func__);
4583 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4584 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4585 		vdropl(vp);
4586 		return;
4587 	}
4588 	if (vn_lock(vp, lkflags) == 0) {
4589 		VI_LOCK(vp);
4590 		vinactive(vp);
4591 		VOP_UNLOCK(vp);
4592 		vdropl(vp);
4593 		return;
4594 	}
4595 	vdefer_inactive_unlocked(vp);
4596 }
4597 
4598 static int
4599 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4600 {
4601 
4602 	return (vp->v_iflag & VI_DEFINACT);
4603 }
4604 
4605 static void __noinline
4606 vfs_periodic_inactive(struct mount *mp, int flags)
4607 {
4608 	struct vnode *vp, *mvp;
4609 	int lkflags;
4610 
4611 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4612 	if (flags != MNT_WAIT)
4613 		lkflags |= LK_NOWAIT;
4614 
4615 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4616 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4617 			VI_UNLOCK(vp);
4618 			continue;
4619 		}
4620 		vp->v_iflag &= ~VI_DEFINACT;
4621 		vfs_deferred_inactive(vp, lkflags);
4622 	}
4623 }
4624 
4625 static inline bool
4626 vfs_want_msync(struct vnode *vp)
4627 {
4628 	struct vm_object *obj;
4629 
4630 	/*
4631 	 * This test may be performed without any locks held.
4632 	 * We rely on vm_object's type stability.
4633 	 */
4634 	if (vp->v_vflag & VV_NOSYNC)
4635 		return (false);
4636 	obj = vp->v_object;
4637 	return (obj != NULL && vm_object_mightbedirty(obj));
4638 }
4639 
4640 static int
4641 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4642 {
4643 
4644 	if (vp->v_vflag & VV_NOSYNC)
4645 		return (false);
4646 	if (vp->v_iflag & VI_DEFINACT)
4647 		return (true);
4648 	return (vfs_want_msync(vp));
4649 }
4650 
4651 static void __noinline
4652 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4653 {
4654 	struct vnode *vp, *mvp;
4655 	struct vm_object *obj;
4656 	struct thread *td;
4657 	int lkflags, objflags;
4658 	bool seen_defer;
4659 
4660 	td = curthread;
4661 
4662 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4663 	if (flags != MNT_WAIT) {
4664 		lkflags |= LK_NOWAIT;
4665 		objflags = OBJPC_NOSYNC;
4666 	} else {
4667 		objflags = OBJPC_SYNC;
4668 	}
4669 
4670 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4671 		seen_defer = false;
4672 		if (vp->v_iflag & VI_DEFINACT) {
4673 			vp->v_iflag &= ~VI_DEFINACT;
4674 			seen_defer = true;
4675 		}
4676 		if (!vfs_want_msync(vp)) {
4677 			if (seen_defer)
4678 				vfs_deferred_inactive(vp, lkflags);
4679 			else
4680 				VI_UNLOCK(vp);
4681 			continue;
4682 		}
4683 		if (vget(vp, lkflags, td) == 0) {
4684 			obj = vp->v_object;
4685 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4686 				VM_OBJECT_WLOCK(obj);
4687 				vm_object_page_clean(obj, 0, 0, objflags);
4688 				VM_OBJECT_WUNLOCK(obj);
4689 			}
4690 			vput(vp);
4691 			if (seen_defer)
4692 				vdrop(vp);
4693 		} else {
4694 			if (seen_defer)
4695 				vdefer_inactive_unlocked(vp);
4696 		}
4697 	}
4698 }
4699 
4700 void
4701 vfs_periodic(struct mount *mp, int flags)
4702 {
4703 
4704 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4705 
4706 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4707 		vfs_periodic_inactive(mp, flags);
4708 	else
4709 		vfs_periodic_msync_inactive(mp, flags);
4710 }
4711 
4712 static void
4713 destroy_vpollinfo_free(struct vpollinfo *vi)
4714 {
4715 
4716 	knlist_destroy(&vi->vpi_selinfo.si_note);
4717 	mtx_destroy(&vi->vpi_lock);
4718 	uma_zfree(vnodepoll_zone, vi);
4719 }
4720 
4721 static void
4722 destroy_vpollinfo(struct vpollinfo *vi)
4723 {
4724 
4725 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4726 	seldrain(&vi->vpi_selinfo);
4727 	destroy_vpollinfo_free(vi);
4728 }
4729 
4730 /*
4731  * Initialize per-vnode helper structure to hold poll-related state.
4732  */
4733 void
4734 v_addpollinfo(struct vnode *vp)
4735 {
4736 	struct vpollinfo *vi;
4737 
4738 	if (vp->v_pollinfo != NULL)
4739 		return;
4740 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4741 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4742 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4743 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4744 	VI_LOCK(vp);
4745 	if (vp->v_pollinfo != NULL) {
4746 		VI_UNLOCK(vp);
4747 		destroy_vpollinfo_free(vi);
4748 		return;
4749 	}
4750 	vp->v_pollinfo = vi;
4751 	VI_UNLOCK(vp);
4752 }
4753 
4754 /*
4755  * Record a process's interest in events which might happen to
4756  * a vnode.  Because poll uses the historic select-style interface
4757  * internally, this routine serves as both the ``check for any
4758  * pending events'' and the ``record my interest in future events''
4759  * functions.  (These are done together, while the lock is held,
4760  * to avoid race conditions.)
4761  */
4762 int
4763 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4764 {
4765 
4766 	v_addpollinfo(vp);
4767 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4768 	if (vp->v_pollinfo->vpi_revents & events) {
4769 		/*
4770 		 * This leaves events we are not interested
4771 		 * in available for the other process which
4772 		 * which presumably had requested them
4773 		 * (otherwise they would never have been
4774 		 * recorded).
4775 		 */
4776 		events &= vp->v_pollinfo->vpi_revents;
4777 		vp->v_pollinfo->vpi_revents &= ~events;
4778 
4779 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4780 		return (events);
4781 	}
4782 	vp->v_pollinfo->vpi_events |= events;
4783 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4784 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4785 	return (0);
4786 }
4787 
4788 /*
4789  * Routine to create and manage a filesystem syncer vnode.
4790  */
4791 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4792 static int	sync_fsync(struct  vop_fsync_args *);
4793 static int	sync_inactive(struct  vop_inactive_args *);
4794 static int	sync_reclaim(struct  vop_reclaim_args *);
4795 
4796 static struct vop_vector sync_vnodeops = {
4797 	.vop_bypass =	VOP_EOPNOTSUPP,
4798 	.vop_close =	sync_close,		/* close */
4799 	.vop_fsync =	sync_fsync,		/* fsync */
4800 	.vop_inactive =	sync_inactive,	/* inactive */
4801 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4802 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4803 	.vop_lock1 =	vop_stdlock,	/* lock */
4804 	.vop_unlock =	vop_stdunlock,	/* unlock */
4805 	.vop_islocked =	vop_stdislocked,	/* islocked */
4806 };
4807 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4808 
4809 /*
4810  * Create a new filesystem syncer vnode for the specified mount point.
4811  */
4812 void
4813 vfs_allocate_syncvnode(struct mount *mp)
4814 {
4815 	struct vnode *vp;
4816 	struct bufobj *bo;
4817 	static long start, incr, next;
4818 	int error;
4819 
4820 	/* Allocate a new vnode */
4821 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4822 	if (error != 0)
4823 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4824 	vp->v_type = VNON;
4825 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4826 	vp->v_vflag |= VV_FORCEINSMQ;
4827 	error = insmntque(vp, mp);
4828 	if (error != 0)
4829 		panic("vfs_allocate_syncvnode: insmntque() failed");
4830 	vp->v_vflag &= ~VV_FORCEINSMQ;
4831 	VOP_UNLOCK(vp);
4832 	/*
4833 	 * Place the vnode onto the syncer worklist. We attempt to
4834 	 * scatter them about on the list so that they will go off
4835 	 * at evenly distributed times even if all the filesystems
4836 	 * are mounted at once.
4837 	 */
4838 	next += incr;
4839 	if (next == 0 || next > syncer_maxdelay) {
4840 		start /= 2;
4841 		incr /= 2;
4842 		if (start == 0) {
4843 			start = syncer_maxdelay / 2;
4844 			incr = syncer_maxdelay;
4845 		}
4846 		next = start;
4847 	}
4848 	bo = &vp->v_bufobj;
4849 	BO_LOCK(bo);
4850 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4851 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4852 	mtx_lock(&sync_mtx);
4853 	sync_vnode_count++;
4854 	if (mp->mnt_syncer == NULL) {
4855 		mp->mnt_syncer = vp;
4856 		vp = NULL;
4857 	}
4858 	mtx_unlock(&sync_mtx);
4859 	BO_UNLOCK(bo);
4860 	if (vp != NULL) {
4861 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4862 		vgone(vp);
4863 		vput(vp);
4864 	}
4865 }
4866 
4867 void
4868 vfs_deallocate_syncvnode(struct mount *mp)
4869 {
4870 	struct vnode *vp;
4871 
4872 	mtx_lock(&sync_mtx);
4873 	vp = mp->mnt_syncer;
4874 	if (vp != NULL)
4875 		mp->mnt_syncer = NULL;
4876 	mtx_unlock(&sync_mtx);
4877 	if (vp != NULL)
4878 		vrele(vp);
4879 }
4880 
4881 /*
4882  * Do a lazy sync of the filesystem.
4883  */
4884 static int
4885 sync_fsync(struct vop_fsync_args *ap)
4886 {
4887 	struct vnode *syncvp = ap->a_vp;
4888 	struct mount *mp = syncvp->v_mount;
4889 	int error, save;
4890 	struct bufobj *bo;
4891 
4892 	/*
4893 	 * We only need to do something if this is a lazy evaluation.
4894 	 */
4895 	if (ap->a_waitfor != MNT_LAZY)
4896 		return (0);
4897 
4898 	/*
4899 	 * Move ourselves to the back of the sync list.
4900 	 */
4901 	bo = &syncvp->v_bufobj;
4902 	BO_LOCK(bo);
4903 	vn_syncer_add_to_worklist(bo, syncdelay);
4904 	BO_UNLOCK(bo);
4905 
4906 	/*
4907 	 * Walk the list of vnodes pushing all that are dirty and
4908 	 * not already on the sync list.
4909 	 */
4910 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4911 		return (0);
4912 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4913 		vfs_unbusy(mp);
4914 		return (0);
4915 	}
4916 	save = curthread_pflags_set(TDP_SYNCIO);
4917 	/*
4918 	 * The filesystem at hand may be idle with free vnodes stored in the
4919 	 * batch.  Return them instead of letting them stay there indefinitely.
4920 	 */
4921 	vfs_periodic(mp, MNT_NOWAIT);
4922 	error = VFS_SYNC(mp, MNT_LAZY);
4923 	curthread_pflags_restore(save);
4924 	vn_finished_write(mp);
4925 	vfs_unbusy(mp);
4926 	return (error);
4927 }
4928 
4929 /*
4930  * The syncer vnode is no referenced.
4931  */
4932 static int
4933 sync_inactive(struct vop_inactive_args *ap)
4934 {
4935 
4936 	vgone(ap->a_vp);
4937 	return (0);
4938 }
4939 
4940 /*
4941  * The syncer vnode is no longer needed and is being decommissioned.
4942  *
4943  * Modifications to the worklist must be protected by sync_mtx.
4944  */
4945 static int
4946 sync_reclaim(struct vop_reclaim_args *ap)
4947 {
4948 	struct vnode *vp = ap->a_vp;
4949 	struct bufobj *bo;
4950 
4951 	bo = &vp->v_bufobj;
4952 	BO_LOCK(bo);
4953 	mtx_lock(&sync_mtx);
4954 	if (vp->v_mount->mnt_syncer == vp)
4955 		vp->v_mount->mnt_syncer = NULL;
4956 	if (bo->bo_flag & BO_ONWORKLST) {
4957 		LIST_REMOVE(bo, bo_synclist);
4958 		syncer_worklist_len--;
4959 		sync_vnode_count--;
4960 		bo->bo_flag &= ~BO_ONWORKLST;
4961 	}
4962 	mtx_unlock(&sync_mtx);
4963 	BO_UNLOCK(bo);
4964 
4965 	return (0);
4966 }
4967 
4968 int
4969 vn_need_pageq_flush(struct vnode *vp)
4970 {
4971 	struct vm_object *obj;
4972 	int need;
4973 
4974 	MPASS(mtx_owned(VI_MTX(vp)));
4975 	need = 0;
4976 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4977 	    vm_object_mightbedirty(obj))
4978 		need = 1;
4979 	return (need);
4980 }
4981 
4982 /*
4983  * Check if vnode represents a disk device
4984  */
4985 int
4986 vn_isdisk(struct vnode *vp, int *errp)
4987 {
4988 	int error;
4989 
4990 	if (vp->v_type != VCHR) {
4991 		error = ENOTBLK;
4992 		goto out;
4993 	}
4994 	error = 0;
4995 	dev_lock();
4996 	if (vp->v_rdev == NULL)
4997 		error = ENXIO;
4998 	else if (vp->v_rdev->si_devsw == NULL)
4999 		error = ENXIO;
5000 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5001 		error = ENOTBLK;
5002 	dev_unlock();
5003 out:
5004 	if (errp != NULL)
5005 		*errp = error;
5006 	return (error == 0);
5007 }
5008 
5009 /*
5010  * Common filesystem object access control check routine.  Accepts a
5011  * vnode's type, "mode", uid and gid, requested access mode, credentials,
5012  * and optional call-by-reference privused argument allowing vaccess()
5013  * to indicate to the caller whether privilege was used to satisfy the
5014  * request (obsoleted).  Returns 0 on success, or an errno on failure.
5015  */
5016 int
5017 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5018     accmode_t accmode, struct ucred *cred, int *privused)
5019 {
5020 	accmode_t dac_granted;
5021 	accmode_t priv_granted;
5022 
5023 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5024 	    ("invalid bit in accmode"));
5025 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5026 	    ("VAPPEND without VWRITE"));
5027 
5028 	/*
5029 	 * Look for a normal, non-privileged way to access the file/directory
5030 	 * as requested.  If it exists, go with that.
5031 	 */
5032 
5033 	if (privused != NULL)
5034 		*privused = 0;
5035 
5036 	dac_granted = 0;
5037 
5038 	/* Check the owner. */
5039 	if (cred->cr_uid == file_uid) {
5040 		dac_granted |= VADMIN;
5041 		if (file_mode & S_IXUSR)
5042 			dac_granted |= VEXEC;
5043 		if (file_mode & S_IRUSR)
5044 			dac_granted |= VREAD;
5045 		if (file_mode & S_IWUSR)
5046 			dac_granted |= (VWRITE | VAPPEND);
5047 
5048 		if ((accmode & dac_granted) == accmode)
5049 			return (0);
5050 
5051 		goto privcheck;
5052 	}
5053 
5054 	/* Otherwise, check the groups (first match) */
5055 	if (groupmember(file_gid, cred)) {
5056 		if (file_mode & S_IXGRP)
5057 			dac_granted |= VEXEC;
5058 		if (file_mode & S_IRGRP)
5059 			dac_granted |= VREAD;
5060 		if (file_mode & S_IWGRP)
5061 			dac_granted |= (VWRITE | VAPPEND);
5062 
5063 		if ((accmode & dac_granted) == accmode)
5064 			return (0);
5065 
5066 		goto privcheck;
5067 	}
5068 
5069 	/* Otherwise, check everyone else. */
5070 	if (file_mode & S_IXOTH)
5071 		dac_granted |= VEXEC;
5072 	if (file_mode & S_IROTH)
5073 		dac_granted |= VREAD;
5074 	if (file_mode & S_IWOTH)
5075 		dac_granted |= (VWRITE | VAPPEND);
5076 	if ((accmode & dac_granted) == accmode)
5077 		return (0);
5078 
5079 privcheck:
5080 	/*
5081 	 * Build a privilege mask to determine if the set of privileges
5082 	 * satisfies the requirements when combined with the granted mask
5083 	 * from above.  For each privilege, if the privilege is required,
5084 	 * bitwise or the request type onto the priv_granted mask.
5085 	 */
5086 	priv_granted = 0;
5087 
5088 	if (type == VDIR) {
5089 		/*
5090 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5091 		 * requests, instead of PRIV_VFS_EXEC.
5092 		 */
5093 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5094 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5095 			priv_granted |= VEXEC;
5096 	} else {
5097 		/*
5098 		 * Ensure that at least one execute bit is on. Otherwise,
5099 		 * a privileged user will always succeed, and we don't want
5100 		 * this to happen unless the file really is executable.
5101 		 */
5102 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5103 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5104 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5105 			priv_granted |= VEXEC;
5106 	}
5107 
5108 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5109 	    !priv_check_cred(cred, PRIV_VFS_READ))
5110 		priv_granted |= VREAD;
5111 
5112 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5113 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5114 		priv_granted |= (VWRITE | VAPPEND);
5115 
5116 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5117 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5118 		priv_granted |= VADMIN;
5119 
5120 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5121 		/* XXX audit: privilege used */
5122 		if (privused != NULL)
5123 			*privused = 1;
5124 		return (0);
5125 	}
5126 
5127 	return ((accmode & VADMIN) ? EPERM : EACCES);
5128 }
5129 
5130 /*
5131  * Credential check based on process requesting service, and per-attribute
5132  * permissions.
5133  */
5134 int
5135 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5136     struct thread *td, accmode_t accmode)
5137 {
5138 
5139 	/*
5140 	 * Kernel-invoked always succeeds.
5141 	 */
5142 	if (cred == NOCRED)
5143 		return (0);
5144 
5145 	/*
5146 	 * Do not allow privileged processes in jail to directly manipulate
5147 	 * system attributes.
5148 	 */
5149 	switch (attrnamespace) {
5150 	case EXTATTR_NAMESPACE_SYSTEM:
5151 		/* Potentially should be: return (EPERM); */
5152 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5153 	case EXTATTR_NAMESPACE_USER:
5154 		return (VOP_ACCESS(vp, accmode, cred, td));
5155 	default:
5156 		return (EPERM);
5157 	}
5158 }
5159 
5160 #ifdef DEBUG_VFS_LOCKS
5161 /*
5162  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5163  * no longer ok to have an unlocked VFS.
5164  */
5165 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5166 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5167 
5168 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5169 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5170     "Drop into debugger on lock violation");
5171 
5172 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5173 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5174     0, "Check for interlock across VOPs");
5175 
5176 int vfs_badlock_print = 1;	/* Print lock violations. */
5177 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5178     0, "Print lock violations");
5179 
5180 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5181 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5182     0, "Print vnode details on lock violations");
5183 
5184 #ifdef KDB
5185 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5186 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5187     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5188 #endif
5189 
5190 static void
5191 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5192 {
5193 
5194 #ifdef KDB
5195 	if (vfs_badlock_backtrace)
5196 		kdb_backtrace();
5197 #endif
5198 	if (vfs_badlock_vnode)
5199 		vn_printf(vp, "vnode ");
5200 	if (vfs_badlock_print)
5201 		printf("%s: %p %s\n", str, (void *)vp, msg);
5202 	if (vfs_badlock_ddb)
5203 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5204 }
5205 
5206 void
5207 assert_vi_locked(struct vnode *vp, const char *str)
5208 {
5209 
5210 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5211 		vfs_badlock("interlock is not locked but should be", str, vp);
5212 }
5213 
5214 void
5215 assert_vi_unlocked(struct vnode *vp, const char *str)
5216 {
5217 
5218 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5219 		vfs_badlock("interlock is locked but should not be", str, vp);
5220 }
5221 
5222 void
5223 assert_vop_locked(struct vnode *vp, const char *str)
5224 {
5225 	int locked;
5226 
5227 	if (!IGNORE_LOCK(vp)) {
5228 		locked = VOP_ISLOCKED(vp);
5229 		if (locked == 0 || locked == LK_EXCLOTHER)
5230 			vfs_badlock("is not locked but should be", str, vp);
5231 	}
5232 }
5233 
5234 void
5235 assert_vop_unlocked(struct vnode *vp, const char *str)
5236 {
5237 
5238 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5239 		vfs_badlock("is locked but should not be", str, vp);
5240 }
5241 
5242 void
5243 assert_vop_elocked(struct vnode *vp, const char *str)
5244 {
5245 
5246 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5247 		vfs_badlock("is not exclusive locked but should be", str, vp);
5248 }
5249 #endif /* DEBUG_VFS_LOCKS */
5250 
5251 void
5252 vop_rename_fail(struct vop_rename_args *ap)
5253 {
5254 
5255 	if (ap->a_tvp != NULL)
5256 		vput(ap->a_tvp);
5257 	if (ap->a_tdvp == ap->a_tvp)
5258 		vrele(ap->a_tdvp);
5259 	else
5260 		vput(ap->a_tdvp);
5261 	vrele(ap->a_fdvp);
5262 	vrele(ap->a_fvp);
5263 }
5264 
5265 void
5266 vop_rename_pre(void *ap)
5267 {
5268 	struct vop_rename_args *a = ap;
5269 
5270 #ifdef DEBUG_VFS_LOCKS
5271 	if (a->a_tvp)
5272 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5273 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5274 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5275 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5276 
5277 	/* Check the source (from). */
5278 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5279 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5280 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5281 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5282 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5283 
5284 	/* Check the target. */
5285 	if (a->a_tvp)
5286 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5287 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5288 #endif
5289 	if (a->a_tdvp != a->a_fdvp)
5290 		vhold(a->a_fdvp);
5291 	if (a->a_tvp != a->a_fvp)
5292 		vhold(a->a_fvp);
5293 	vhold(a->a_tdvp);
5294 	if (a->a_tvp)
5295 		vhold(a->a_tvp);
5296 }
5297 
5298 #ifdef DEBUG_VFS_LOCKS
5299 void
5300 vop_strategy_pre(void *ap)
5301 {
5302 	struct vop_strategy_args *a;
5303 	struct buf *bp;
5304 
5305 	a = ap;
5306 	bp = a->a_bp;
5307 
5308 	/*
5309 	 * Cluster ops lock their component buffers but not the IO container.
5310 	 */
5311 	if ((bp->b_flags & B_CLUSTER) != 0)
5312 		return;
5313 
5314 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5315 		if (vfs_badlock_print)
5316 			printf(
5317 			    "VOP_STRATEGY: bp is not locked but should be\n");
5318 		if (vfs_badlock_ddb)
5319 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5320 	}
5321 }
5322 
5323 void
5324 vop_lock_pre(void *ap)
5325 {
5326 	struct vop_lock1_args *a = ap;
5327 
5328 	if ((a->a_flags & LK_INTERLOCK) == 0)
5329 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5330 	else
5331 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5332 }
5333 
5334 void
5335 vop_lock_post(void *ap, int rc)
5336 {
5337 	struct vop_lock1_args *a = ap;
5338 
5339 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5340 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5341 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5342 }
5343 
5344 void
5345 vop_unlock_pre(void *ap)
5346 {
5347 	struct vop_unlock_args *a = ap;
5348 
5349 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5350 }
5351 
5352 void
5353 vop_unlock_post(void *ap, int rc)
5354 {
5355 	return;
5356 }
5357 
5358 void
5359 vop_need_inactive_pre(void *ap)
5360 {
5361 	struct vop_need_inactive_args *a = ap;
5362 
5363 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5364 }
5365 
5366 void
5367 vop_need_inactive_post(void *ap, int rc)
5368 {
5369 	struct vop_need_inactive_args *a = ap;
5370 
5371 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5372 }
5373 #endif
5374 
5375 void
5376 vop_create_post(void *ap, int rc)
5377 {
5378 	struct vop_create_args *a = ap;
5379 
5380 	if (!rc)
5381 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5382 }
5383 
5384 void
5385 vop_deleteextattr_post(void *ap, int rc)
5386 {
5387 	struct vop_deleteextattr_args *a = ap;
5388 
5389 	if (!rc)
5390 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5391 }
5392 
5393 void
5394 vop_link_post(void *ap, int rc)
5395 {
5396 	struct vop_link_args *a = ap;
5397 
5398 	if (!rc) {
5399 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5400 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5401 	}
5402 }
5403 
5404 void
5405 vop_mkdir_post(void *ap, int rc)
5406 {
5407 	struct vop_mkdir_args *a = ap;
5408 
5409 	if (!rc)
5410 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5411 }
5412 
5413 void
5414 vop_mknod_post(void *ap, int rc)
5415 {
5416 	struct vop_mknod_args *a = ap;
5417 
5418 	if (!rc)
5419 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5420 }
5421 
5422 void
5423 vop_reclaim_post(void *ap, int rc)
5424 {
5425 	struct vop_reclaim_args *a = ap;
5426 
5427 	if (!rc)
5428 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5429 }
5430 
5431 void
5432 vop_remove_post(void *ap, int rc)
5433 {
5434 	struct vop_remove_args *a = ap;
5435 
5436 	if (!rc) {
5437 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5438 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5439 	}
5440 }
5441 
5442 void
5443 vop_rename_post(void *ap, int rc)
5444 {
5445 	struct vop_rename_args *a = ap;
5446 	long hint;
5447 
5448 	if (!rc) {
5449 		hint = NOTE_WRITE;
5450 		if (a->a_fdvp == a->a_tdvp) {
5451 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5452 				hint |= NOTE_LINK;
5453 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5454 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5455 		} else {
5456 			hint |= NOTE_EXTEND;
5457 			if (a->a_fvp->v_type == VDIR)
5458 				hint |= NOTE_LINK;
5459 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5460 
5461 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5462 			    a->a_tvp->v_type == VDIR)
5463 				hint &= ~NOTE_LINK;
5464 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5465 		}
5466 
5467 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5468 		if (a->a_tvp)
5469 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5470 	}
5471 	if (a->a_tdvp != a->a_fdvp)
5472 		vdrop(a->a_fdvp);
5473 	if (a->a_tvp != a->a_fvp)
5474 		vdrop(a->a_fvp);
5475 	vdrop(a->a_tdvp);
5476 	if (a->a_tvp)
5477 		vdrop(a->a_tvp);
5478 }
5479 
5480 void
5481 vop_rmdir_post(void *ap, int rc)
5482 {
5483 	struct vop_rmdir_args *a = ap;
5484 
5485 	if (!rc) {
5486 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5487 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5488 	}
5489 }
5490 
5491 void
5492 vop_setattr_post(void *ap, int rc)
5493 {
5494 	struct vop_setattr_args *a = ap;
5495 
5496 	if (!rc)
5497 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5498 }
5499 
5500 void
5501 vop_setextattr_post(void *ap, int rc)
5502 {
5503 	struct vop_setextattr_args *a = ap;
5504 
5505 	if (!rc)
5506 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5507 }
5508 
5509 void
5510 vop_symlink_post(void *ap, int rc)
5511 {
5512 	struct vop_symlink_args *a = ap;
5513 
5514 	if (!rc)
5515 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5516 }
5517 
5518 void
5519 vop_open_post(void *ap, int rc)
5520 {
5521 	struct vop_open_args *a = ap;
5522 
5523 	if (!rc)
5524 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5525 }
5526 
5527 void
5528 vop_close_post(void *ap, int rc)
5529 {
5530 	struct vop_close_args *a = ap;
5531 
5532 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5533 	    !VN_IS_DOOMED(a->a_vp))) {
5534 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5535 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5536 	}
5537 }
5538 
5539 void
5540 vop_read_post(void *ap, int rc)
5541 {
5542 	struct vop_read_args *a = ap;
5543 
5544 	if (!rc)
5545 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5546 }
5547 
5548 void
5549 vop_readdir_post(void *ap, int rc)
5550 {
5551 	struct vop_readdir_args *a = ap;
5552 
5553 	if (!rc)
5554 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5555 }
5556 
5557 static struct knlist fs_knlist;
5558 
5559 static void
5560 vfs_event_init(void *arg)
5561 {
5562 	knlist_init_mtx(&fs_knlist, NULL);
5563 }
5564 /* XXX - correct order? */
5565 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5566 
5567 void
5568 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5569 {
5570 
5571 	KNOTE_UNLOCKED(&fs_knlist, event);
5572 }
5573 
5574 static int	filt_fsattach(struct knote *kn);
5575 static void	filt_fsdetach(struct knote *kn);
5576 static int	filt_fsevent(struct knote *kn, long hint);
5577 
5578 struct filterops fs_filtops = {
5579 	.f_isfd = 0,
5580 	.f_attach = filt_fsattach,
5581 	.f_detach = filt_fsdetach,
5582 	.f_event = filt_fsevent
5583 };
5584 
5585 static int
5586 filt_fsattach(struct knote *kn)
5587 {
5588 
5589 	kn->kn_flags |= EV_CLEAR;
5590 	knlist_add(&fs_knlist, kn, 0);
5591 	return (0);
5592 }
5593 
5594 static void
5595 filt_fsdetach(struct knote *kn)
5596 {
5597 
5598 	knlist_remove(&fs_knlist, kn, 0);
5599 }
5600 
5601 static int
5602 filt_fsevent(struct knote *kn, long hint)
5603 {
5604 
5605 	kn->kn_fflags |= hint;
5606 	return (kn->kn_fflags != 0);
5607 }
5608 
5609 static int
5610 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5611 {
5612 	struct vfsidctl vc;
5613 	int error;
5614 	struct mount *mp;
5615 
5616 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5617 	if (error)
5618 		return (error);
5619 	if (vc.vc_vers != VFS_CTL_VERS1)
5620 		return (EINVAL);
5621 	mp = vfs_getvfs(&vc.vc_fsid);
5622 	if (mp == NULL)
5623 		return (ENOENT);
5624 	/* ensure that a specific sysctl goes to the right filesystem. */
5625 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5626 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5627 		vfs_rel(mp);
5628 		return (EINVAL);
5629 	}
5630 	VCTLTOREQ(&vc, req);
5631 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5632 	vfs_rel(mp);
5633 	return (error);
5634 }
5635 
5636 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5637     NULL, 0, sysctl_vfs_ctl, "",
5638     "Sysctl by fsid");
5639 
5640 /*
5641  * Function to initialize a va_filerev field sensibly.
5642  * XXX: Wouldn't a random number make a lot more sense ??
5643  */
5644 u_quad_t
5645 init_va_filerev(void)
5646 {
5647 	struct bintime bt;
5648 
5649 	getbinuptime(&bt);
5650 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5651 }
5652 
5653 static int	filt_vfsread(struct knote *kn, long hint);
5654 static int	filt_vfswrite(struct knote *kn, long hint);
5655 static int	filt_vfsvnode(struct knote *kn, long hint);
5656 static void	filt_vfsdetach(struct knote *kn);
5657 static struct filterops vfsread_filtops = {
5658 	.f_isfd = 1,
5659 	.f_detach = filt_vfsdetach,
5660 	.f_event = filt_vfsread
5661 };
5662 static struct filterops vfswrite_filtops = {
5663 	.f_isfd = 1,
5664 	.f_detach = filt_vfsdetach,
5665 	.f_event = filt_vfswrite
5666 };
5667 static struct filterops vfsvnode_filtops = {
5668 	.f_isfd = 1,
5669 	.f_detach = filt_vfsdetach,
5670 	.f_event = filt_vfsvnode
5671 };
5672 
5673 static void
5674 vfs_knllock(void *arg)
5675 {
5676 	struct vnode *vp = arg;
5677 
5678 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5679 }
5680 
5681 static void
5682 vfs_knlunlock(void *arg)
5683 {
5684 	struct vnode *vp = arg;
5685 
5686 	VOP_UNLOCK(vp);
5687 }
5688 
5689 static void
5690 vfs_knl_assert_locked(void *arg)
5691 {
5692 #ifdef DEBUG_VFS_LOCKS
5693 	struct vnode *vp = arg;
5694 
5695 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5696 #endif
5697 }
5698 
5699 static void
5700 vfs_knl_assert_unlocked(void *arg)
5701 {
5702 #ifdef DEBUG_VFS_LOCKS
5703 	struct vnode *vp = arg;
5704 
5705 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5706 #endif
5707 }
5708 
5709 int
5710 vfs_kqfilter(struct vop_kqfilter_args *ap)
5711 {
5712 	struct vnode *vp = ap->a_vp;
5713 	struct knote *kn = ap->a_kn;
5714 	struct knlist *knl;
5715 
5716 	switch (kn->kn_filter) {
5717 	case EVFILT_READ:
5718 		kn->kn_fop = &vfsread_filtops;
5719 		break;
5720 	case EVFILT_WRITE:
5721 		kn->kn_fop = &vfswrite_filtops;
5722 		break;
5723 	case EVFILT_VNODE:
5724 		kn->kn_fop = &vfsvnode_filtops;
5725 		break;
5726 	default:
5727 		return (EINVAL);
5728 	}
5729 
5730 	kn->kn_hook = (caddr_t)vp;
5731 
5732 	v_addpollinfo(vp);
5733 	if (vp->v_pollinfo == NULL)
5734 		return (ENOMEM);
5735 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5736 	vhold(vp);
5737 	knlist_add(knl, kn, 0);
5738 
5739 	return (0);
5740 }
5741 
5742 /*
5743  * Detach knote from vnode
5744  */
5745 static void
5746 filt_vfsdetach(struct knote *kn)
5747 {
5748 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5749 
5750 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5751 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5752 	vdrop(vp);
5753 }
5754 
5755 /*ARGSUSED*/
5756 static int
5757 filt_vfsread(struct knote *kn, long hint)
5758 {
5759 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5760 	struct vattr va;
5761 	int res;
5762 
5763 	/*
5764 	 * filesystem is gone, so set the EOF flag and schedule
5765 	 * the knote for deletion.
5766 	 */
5767 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5768 		VI_LOCK(vp);
5769 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5770 		VI_UNLOCK(vp);
5771 		return (1);
5772 	}
5773 
5774 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5775 		return (0);
5776 
5777 	VI_LOCK(vp);
5778 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5779 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5780 	VI_UNLOCK(vp);
5781 	return (res);
5782 }
5783 
5784 /*ARGSUSED*/
5785 static int
5786 filt_vfswrite(struct knote *kn, long hint)
5787 {
5788 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5789 
5790 	VI_LOCK(vp);
5791 
5792 	/*
5793 	 * filesystem is gone, so set the EOF flag and schedule
5794 	 * the knote for deletion.
5795 	 */
5796 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5797 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5798 
5799 	kn->kn_data = 0;
5800 	VI_UNLOCK(vp);
5801 	return (1);
5802 }
5803 
5804 static int
5805 filt_vfsvnode(struct knote *kn, long hint)
5806 {
5807 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5808 	int res;
5809 
5810 	VI_LOCK(vp);
5811 	if (kn->kn_sfflags & hint)
5812 		kn->kn_fflags |= hint;
5813 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5814 		kn->kn_flags |= EV_EOF;
5815 		VI_UNLOCK(vp);
5816 		return (1);
5817 	}
5818 	res = (kn->kn_fflags != 0);
5819 	VI_UNLOCK(vp);
5820 	return (res);
5821 }
5822 
5823 /*
5824  * Returns whether the directory is empty or not.
5825  * If it is empty, the return value is 0; otherwise
5826  * the return value is an error value (which may
5827  * be ENOTEMPTY).
5828  */
5829 int
5830 vfs_emptydir(struct vnode *vp)
5831 {
5832 	struct uio uio;
5833 	struct iovec iov;
5834 	struct dirent *dirent, *dp, *endp;
5835 	int error, eof;
5836 
5837 	error = 0;
5838 	eof = 0;
5839 
5840 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5841 
5842 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5843 	iov.iov_base = dirent;
5844 	iov.iov_len = sizeof(struct dirent);
5845 
5846 	uio.uio_iov = &iov;
5847 	uio.uio_iovcnt = 1;
5848 	uio.uio_offset = 0;
5849 	uio.uio_resid = sizeof(struct dirent);
5850 	uio.uio_segflg = UIO_SYSSPACE;
5851 	uio.uio_rw = UIO_READ;
5852 	uio.uio_td = curthread;
5853 
5854 	while (eof == 0 && error == 0) {
5855 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5856 		    NULL, NULL);
5857 		if (error != 0)
5858 			break;
5859 		endp = (void *)((uint8_t *)dirent +
5860 		    sizeof(struct dirent) - uio.uio_resid);
5861 		for (dp = dirent; dp < endp;
5862 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5863 			if (dp->d_type == DT_WHT)
5864 				continue;
5865 			if (dp->d_namlen == 0)
5866 				continue;
5867 			if (dp->d_type != DT_DIR &&
5868 			    dp->d_type != DT_UNKNOWN) {
5869 				error = ENOTEMPTY;
5870 				break;
5871 			}
5872 			if (dp->d_namlen > 2) {
5873 				error = ENOTEMPTY;
5874 				break;
5875 			}
5876 			if (dp->d_namlen == 1 &&
5877 			    dp->d_name[0] != '.') {
5878 				error = ENOTEMPTY;
5879 				break;
5880 			}
5881 			if (dp->d_namlen == 2 &&
5882 			    dp->d_name[1] != '.') {
5883 				error = ENOTEMPTY;
5884 				break;
5885 			}
5886 			uio.uio_resid = sizeof(struct dirent);
5887 		}
5888 	}
5889 	free(dirent, M_TEMP);
5890 	return (error);
5891 }
5892 
5893 int
5894 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5895 {
5896 	int error;
5897 
5898 	if (dp->d_reclen > ap->a_uio->uio_resid)
5899 		return (ENAMETOOLONG);
5900 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5901 	if (error) {
5902 		if (ap->a_ncookies != NULL) {
5903 			if (ap->a_cookies != NULL)
5904 				free(ap->a_cookies, M_TEMP);
5905 			ap->a_cookies = NULL;
5906 			*ap->a_ncookies = 0;
5907 		}
5908 		return (error);
5909 	}
5910 	if (ap->a_ncookies == NULL)
5911 		return (0);
5912 
5913 	KASSERT(ap->a_cookies,
5914 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5915 
5916 	*ap->a_cookies = realloc(*ap->a_cookies,
5917 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5918 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5919 	*ap->a_ncookies += 1;
5920 	return (0);
5921 }
5922 
5923 /*
5924  * Mark for update the access time of the file if the filesystem
5925  * supports VOP_MARKATIME.  This functionality is used by execve and
5926  * mmap, so we want to avoid the I/O implied by directly setting
5927  * va_atime for the sake of efficiency.
5928  */
5929 void
5930 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5931 {
5932 	struct mount *mp;
5933 
5934 	mp = vp->v_mount;
5935 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5936 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5937 		(void)VOP_MARKATIME(vp);
5938 }
5939 
5940 /*
5941  * The purpose of this routine is to remove granularity from accmode_t,
5942  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5943  * VADMIN and VAPPEND.
5944  *
5945  * If it returns 0, the caller is supposed to continue with the usual
5946  * access checks using 'accmode' as modified by this routine.  If it
5947  * returns nonzero value, the caller is supposed to return that value
5948  * as errno.
5949  *
5950  * Note that after this routine runs, accmode may be zero.
5951  */
5952 int
5953 vfs_unixify_accmode(accmode_t *accmode)
5954 {
5955 	/*
5956 	 * There is no way to specify explicit "deny" rule using
5957 	 * file mode or POSIX.1e ACLs.
5958 	 */
5959 	if (*accmode & VEXPLICIT_DENY) {
5960 		*accmode = 0;
5961 		return (0);
5962 	}
5963 
5964 	/*
5965 	 * None of these can be translated into usual access bits.
5966 	 * Also, the common case for NFSv4 ACLs is to not contain
5967 	 * either of these bits. Caller should check for VWRITE
5968 	 * on the containing directory instead.
5969 	 */
5970 	if (*accmode & (VDELETE_CHILD | VDELETE))
5971 		return (EPERM);
5972 
5973 	if (*accmode & VADMIN_PERMS) {
5974 		*accmode &= ~VADMIN_PERMS;
5975 		*accmode |= VADMIN;
5976 	}
5977 
5978 	/*
5979 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5980 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5981 	 */
5982 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5983 
5984 	return (0);
5985 }
5986 
5987 /*
5988  * Clear out a doomed vnode (if any) and replace it with a new one as long
5989  * as the fs is not being unmounted. Return the root vnode to the caller.
5990  */
5991 static int __noinline
5992 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
5993 {
5994 	struct vnode *vp;
5995 	int error;
5996 
5997 restart:
5998 	if (mp->mnt_rootvnode != NULL) {
5999 		MNT_ILOCK(mp);
6000 		vp = mp->mnt_rootvnode;
6001 		if (vp != NULL) {
6002 			if (!VN_IS_DOOMED(vp)) {
6003 				vrefact(vp);
6004 				MNT_IUNLOCK(mp);
6005 				error = vn_lock(vp, flags);
6006 				if (error == 0) {
6007 					*vpp = vp;
6008 					return (0);
6009 				}
6010 				vrele(vp);
6011 				goto restart;
6012 			}
6013 			/*
6014 			 * Clear the old one.
6015 			 */
6016 			mp->mnt_rootvnode = NULL;
6017 		}
6018 		MNT_IUNLOCK(mp);
6019 		if (vp != NULL) {
6020 			/*
6021 			 * Paired with a fence in vfs_op_thread_exit().
6022 			 */
6023 			atomic_thread_fence_acq();
6024 			vfs_op_barrier_wait(mp);
6025 			vrele(vp);
6026 		}
6027 	}
6028 	error = VFS_CACHEDROOT(mp, flags, vpp);
6029 	if (error != 0)
6030 		return (error);
6031 	if (mp->mnt_vfs_ops == 0) {
6032 		MNT_ILOCK(mp);
6033 		if (mp->mnt_vfs_ops != 0) {
6034 			MNT_IUNLOCK(mp);
6035 			return (0);
6036 		}
6037 		if (mp->mnt_rootvnode == NULL) {
6038 			vrefact(*vpp);
6039 			mp->mnt_rootvnode = *vpp;
6040 		} else {
6041 			if (mp->mnt_rootvnode != *vpp) {
6042 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6043 					panic("%s: mismatch between vnode returned "
6044 					    " by VFS_CACHEDROOT and the one cached "
6045 					    " (%p != %p)",
6046 					    __func__, *vpp, mp->mnt_rootvnode);
6047 				}
6048 			}
6049 		}
6050 		MNT_IUNLOCK(mp);
6051 	}
6052 	return (0);
6053 }
6054 
6055 int
6056 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6057 {
6058 	struct vnode *vp;
6059 	int error;
6060 
6061 	if (!vfs_op_thread_enter(mp))
6062 		return (vfs_cache_root_fallback(mp, flags, vpp));
6063 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
6064 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6065 		vfs_op_thread_exit(mp);
6066 		return (vfs_cache_root_fallback(mp, flags, vpp));
6067 	}
6068 	vrefact(vp);
6069 	vfs_op_thread_exit(mp);
6070 	error = vn_lock(vp, flags);
6071 	if (error != 0) {
6072 		vrele(vp);
6073 		return (vfs_cache_root_fallback(mp, flags, vpp));
6074 	}
6075 	*vpp = vp;
6076 	return (0);
6077 }
6078 
6079 struct vnode *
6080 vfs_cache_root_clear(struct mount *mp)
6081 {
6082 	struct vnode *vp;
6083 
6084 	/*
6085 	 * ops > 0 guarantees there is nobody who can see this vnode
6086 	 */
6087 	MPASS(mp->mnt_vfs_ops > 0);
6088 	vp = mp->mnt_rootvnode;
6089 	mp->mnt_rootvnode = NULL;
6090 	return (vp);
6091 }
6092 
6093 void
6094 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6095 {
6096 
6097 	MPASS(mp->mnt_vfs_ops > 0);
6098 	vrefact(vp);
6099 	mp->mnt_rootvnode = vp;
6100 }
6101 
6102 /*
6103  * These are helper functions for filesystems to traverse all
6104  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6105  *
6106  * This interface replaces MNT_VNODE_FOREACH.
6107  */
6108 
6109 
6110 struct vnode *
6111 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6112 {
6113 	struct vnode *vp;
6114 
6115 	if (should_yield())
6116 		kern_yield(PRI_USER);
6117 	MNT_ILOCK(mp);
6118 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6119 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6120 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6121 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6122 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6123 			continue;
6124 		VI_LOCK(vp);
6125 		if (VN_IS_DOOMED(vp)) {
6126 			VI_UNLOCK(vp);
6127 			continue;
6128 		}
6129 		break;
6130 	}
6131 	if (vp == NULL) {
6132 		__mnt_vnode_markerfree_all(mvp, mp);
6133 		/* MNT_IUNLOCK(mp); -- done in above function */
6134 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6135 		return (NULL);
6136 	}
6137 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6138 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6139 	MNT_IUNLOCK(mp);
6140 	return (vp);
6141 }
6142 
6143 struct vnode *
6144 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6145 {
6146 	struct vnode *vp;
6147 
6148 	*mvp = vn_alloc_marker(mp);
6149 	MNT_ILOCK(mp);
6150 	MNT_REF(mp);
6151 
6152 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6153 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6154 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6155 			continue;
6156 		VI_LOCK(vp);
6157 		if (VN_IS_DOOMED(vp)) {
6158 			VI_UNLOCK(vp);
6159 			continue;
6160 		}
6161 		break;
6162 	}
6163 	if (vp == NULL) {
6164 		MNT_REL(mp);
6165 		MNT_IUNLOCK(mp);
6166 		vn_free_marker(*mvp);
6167 		*mvp = NULL;
6168 		return (NULL);
6169 	}
6170 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6171 	MNT_IUNLOCK(mp);
6172 	return (vp);
6173 }
6174 
6175 void
6176 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6177 {
6178 
6179 	if (*mvp == NULL) {
6180 		MNT_IUNLOCK(mp);
6181 		return;
6182 	}
6183 
6184 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6185 
6186 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6187 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6188 	MNT_REL(mp);
6189 	MNT_IUNLOCK(mp);
6190 	vn_free_marker(*mvp);
6191 	*mvp = NULL;
6192 }
6193 
6194 /*
6195  * These are helper functions for filesystems to traverse their
6196  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6197  */
6198 static void
6199 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6200 {
6201 
6202 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6203 
6204 	MNT_ILOCK(mp);
6205 	MNT_REL(mp);
6206 	MNT_IUNLOCK(mp);
6207 	vn_free_marker(*mvp);
6208 	*mvp = NULL;
6209 }
6210 
6211 /*
6212  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6213  * conventional lock order during mnt_vnode_next_lazy iteration.
6214  *
6215  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6216  * The list lock is dropped and reacquired.  On success, both locks are held.
6217  * On failure, the mount vnode list lock is held but the vnode interlock is
6218  * not, and the procedure may have yielded.
6219  */
6220 static bool
6221 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6222     struct vnode *vp)
6223 {
6224 	const struct vnode *tmp;
6225 	bool held, ret;
6226 
6227 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6228 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6229 	    ("%s: bad marker", __func__));
6230 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6231 	    ("%s: inappropriate vnode", __func__));
6232 	ASSERT_VI_UNLOCKED(vp, __func__);
6233 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6234 
6235 	ret = false;
6236 
6237 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6238 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6239 
6240 	/*
6241 	 * Use a hold to prevent vp from disappearing while the mount vnode
6242 	 * list lock is dropped and reacquired.  Normally a hold would be
6243 	 * acquired with vhold(), but that might try to acquire the vnode
6244 	 * interlock, which would be a LOR with the mount vnode list lock.
6245 	 */
6246 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
6247 	mtx_unlock(&mp->mnt_listmtx);
6248 	if (!held)
6249 		goto abort;
6250 	VI_LOCK(vp);
6251 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
6252 		vdropl(vp);
6253 		goto abort;
6254 	}
6255 	mtx_lock(&mp->mnt_listmtx);
6256 
6257 	/*
6258 	 * Determine whether the vnode is still the next one after the marker,
6259 	 * excepting any other markers.  If the vnode has not been doomed by
6260 	 * vgone() then the hold should have ensured that it remained on the
6261 	 * lazy list.  If it has been doomed but is still on the lazy list,
6262 	 * don't abort, but rather skip over it (avoid spinning on doomed
6263 	 * vnodes).
6264 	 */
6265 	tmp = mvp;
6266 	do {
6267 		tmp = TAILQ_NEXT(tmp, v_lazylist);
6268 	} while (tmp != NULL && tmp->v_type == VMARKER);
6269 	if (tmp != vp) {
6270 		mtx_unlock(&mp->mnt_listmtx);
6271 		VI_UNLOCK(vp);
6272 		goto abort;
6273 	}
6274 
6275 	ret = true;
6276 	goto out;
6277 abort:
6278 	maybe_yield();
6279 	mtx_lock(&mp->mnt_listmtx);
6280 out:
6281 	if (ret)
6282 		ASSERT_VI_LOCKED(vp, __func__);
6283 	else
6284 		ASSERT_VI_UNLOCKED(vp, __func__);
6285 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6286 	return (ret);
6287 }
6288 
6289 static struct vnode *
6290 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6291     void *cbarg)
6292 {
6293 	struct vnode *vp, *nvp;
6294 
6295 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6296 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6297 restart:
6298 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6299 	while (vp != NULL) {
6300 		if (vp->v_type == VMARKER) {
6301 			vp = TAILQ_NEXT(vp, v_lazylist);
6302 			continue;
6303 		}
6304 		/*
6305 		 * See if we want to process the vnode. Note we may encounter a
6306 		 * long string of vnodes we don't care about and hog the list
6307 		 * as a result. Check for it and requeue the marker.
6308 		 */
6309 		if (VN_IS_DOOMED(vp) || !cb(vp, cbarg)) {
6310 			if (!should_yield()) {
6311 				vp = TAILQ_NEXT(vp, v_lazylist);
6312 				continue;
6313 			}
6314 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6315 			    v_lazylist);
6316 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6317 			    v_lazylist);
6318 			mtx_unlock(&mp->mnt_listmtx);
6319 			kern_yield(PRI_USER);
6320 			mtx_lock(&mp->mnt_listmtx);
6321 			goto restart;
6322 		}
6323 		/*
6324 		 * Try-lock because this is the wrong lock order.  If that does
6325 		 * not succeed, drop the mount vnode list lock and try to
6326 		 * reacquire it and the vnode interlock in the right order.
6327 		 */
6328 		if (!VI_TRYLOCK(vp) &&
6329 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6330 			goto restart;
6331 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6332 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6333 		    ("alien vnode on the lazy list %p %p", vp, mp));
6334 		if (vp->v_mount == mp && !VN_IS_DOOMED(vp))
6335 			break;
6336 		nvp = TAILQ_NEXT(vp, v_lazylist);
6337 		VI_UNLOCK(vp);
6338 		vp = nvp;
6339 	}
6340 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6341 
6342 	/* Check if we are done */
6343 	if (vp == NULL) {
6344 		mtx_unlock(&mp->mnt_listmtx);
6345 		mnt_vnode_markerfree_lazy(mvp, mp);
6346 		return (NULL);
6347 	}
6348 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6349 	mtx_unlock(&mp->mnt_listmtx);
6350 	ASSERT_VI_LOCKED(vp, "lazy iter");
6351 	return (vp);
6352 }
6353 
6354 struct vnode *
6355 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6356     void *cbarg)
6357 {
6358 
6359 	if (should_yield())
6360 		kern_yield(PRI_USER);
6361 	mtx_lock(&mp->mnt_listmtx);
6362 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6363 }
6364 
6365 struct vnode *
6366 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6367     void *cbarg)
6368 {
6369 	struct vnode *vp;
6370 
6371 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6372 		return (NULL);
6373 
6374 	*mvp = vn_alloc_marker(mp);
6375 	MNT_ILOCK(mp);
6376 	MNT_REF(mp);
6377 	MNT_IUNLOCK(mp);
6378 
6379 	mtx_lock(&mp->mnt_listmtx);
6380 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6381 	if (vp == NULL) {
6382 		mtx_unlock(&mp->mnt_listmtx);
6383 		mnt_vnode_markerfree_lazy(mvp, mp);
6384 		return (NULL);
6385 	}
6386 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6387 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6388 }
6389 
6390 void
6391 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6392 {
6393 
6394 	if (*mvp == NULL)
6395 		return;
6396 
6397 	mtx_lock(&mp->mnt_listmtx);
6398 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6399 	mtx_unlock(&mp->mnt_listmtx);
6400 	mnt_vnode_markerfree_lazy(mvp, mp);
6401 }
6402