xref: /freebsd/sys/kern/vfs_subr.c (revision 16fa3dcba027d13dcda9ee78e6057e3e5a79f80c)
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/asan.h>
52 #include <sys/bio.h>
53 #include <sys/buf.h>
54 #include <sys/capsicum.h>
55 #include <sys/condvar.h>
56 #include <sys/conf.h>
57 #include <sys/counter.h>
58 #include <sys/dirent.h>
59 #include <sys/event.h>
60 #include <sys/eventhandler.h>
61 #include <sys/extattr.h>
62 #include <sys/file.h>
63 #include <sys/fcntl.h>
64 #include <sys/jail.h>
65 #include <sys/kdb.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/ktr.h>
69 #include <sys/lockf.h>
70 #include <sys/malloc.h>
71 #include <sys/mount.h>
72 #include <sys/namei.h>
73 #include <sys/pctrie.h>
74 #include <sys/priv.h>
75 #include <sys/reboot.h>
76 #include <sys/refcount.h>
77 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/smr.h>
81 #include <sys/smp.h>
82 #include <sys/stat.h>
83 #include <sys/sysctl.h>
84 #include <sys/syslog.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/watchdog.h>
88 
89 #include <machine/stdarg.h>
90 
91 #include <security/mac/mac_framework.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_kern.h>
100 #include <vm/uma.h>
101 
102 #ifdef DDB
103 #include <ddb/ddb.h>
104 #endif
105 
106 static void	delmntque(struct vnode *vp);
107 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
108 		    int slpflag, int slptimeo);
109 static void	syncer_shutdown(void *arg, int howto);
110 static int	vtryrecycle(struct vnode *vp);
111 static void	v_init_counters(struct vnode *);
112 static void	vn_seqc_init(struct vnode *);
113 static void	vn_seqc_write_end_free(struct vnode *vp);
114 static void	vgonel(struct vnode *);
115 static bool	vhold_recycle_free(struct vnode *);
116 static void	vfs_knllock(void *arg);
117 static void	vfs_knlunlock(void *arg);
118 static void	vfs_knl_assert_lock(void *arg, int what);
119 static void	destroy_vpollinfo(struct vpollinfo *vi);
120 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
121 		    daddr_t startlbn, daddr_t endlbn);
122 static void	vnlru_recalc(void);
123 
124 /*
125  * These fences are intended for cases where some synchronization is
126  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
127  * and v_usecount) updates.  Access to v_iflags is generally synchronized
128  * by the interlock, but we have some internal assertions that check vnode
129  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
130  * for now.
131  */
132 #ifdef INVARIANTS
133 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
134 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
135 #else
136 #define	VNODE_REFCOUNT_FENCE_ACQ()
137 #define	VNODE_REFCOUNT_FENCE_REL()
138 #endif
139 
140 /*
141  * Number of vnodes in existence.  Increased whenever getnewvnode()
142  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
143  */
144 static u_long __exclusive_cache_line numvnodes;
145 
146 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
147     "Number of vnodes in existence");
148 
149 static counter_u64_t vnodes_created;
150 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
151     "Number of vnodes created by getnewvnode");
152 
153 /*
154  * Conversion tables for conversion from vnode types to inode formats
155  * and back.
156  */
157 enum vtype iftovt_tab[16] = {
158 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
159 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
160 };
161 int vttoif_tab[10] = {
162 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
163 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
164 };
165 
166 /*
167  * List of allocates vnodes in the system.
168  */
169 static TAILQ_HEAD(freelst, vnode) vnode_list;
170 static struct vnode *vnode_list_free_marker;
171 static struct vnode *vnode_list_reclaim_marker;
172 
173 /*
174  * "Free" vnode target.  Free vnodes are rarely completely free, but are
175  * just ones that are cheap to recycle.  Usually they are for files which
176  * have been stat'd but not read; these usually have inode and namecache
177  * data attached to them.  This target is the preferred minimum size of a
178  * sub-cache consisting mostly of such files. The system balances the size
179  * of this sub-cache with its complement to try to prevent either from
180  * thrashing while the other is relatively inactive.  The targets express
181  * a preference for the best balance.
182  *
183  * "Above" this target there are 2 further targets (watermarks) related
184  * to recyling of free vnodes.  In the best-operating case, the cache is
185  * exactly full, the free list has size between vlowat and vhiwat above the
186  * free target, and recycling from it and normal use maintains this state.
187  * Sometimes the free list is below vlowat or even empty, but this state
188  * is even better for immediate use provided the cache is not full.
189  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
190  * ones) to reach one of these states.  The watermarks are currently hard-
191  * coded as 4% and 9% of the available space higher.  These and the default
192  * of 25% for wantfreevnodes are too large if the memory size is large.
193  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
194  * whenever vnlru_proc() becomes active.
195  */
196 static long wantfreevnodes;
197 static long __exclusive_cache_line freevnodes;
198 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
199     &freevnodes, 0, "Number of \"free\" vnodes");
200 static long freevnodes_old;
201 
202 static counter_u64_t recycles_count;
203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
204     "Number of vnodes recycled to meet vnode cache targets");
205 
206 static counter_u64_t recycles_free_count;
207 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
208     "Number of free vnodes recycled to meet vnode cache targets");
209 
210 static counter_u64_t deferred_inact;
211 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
212     "Number of times inactive processing was deferred");
213 
214 /* To keep more than one thread at a time from running vfs_getnewfsid */
215 static struct mtx mntid_mtx;
216 
217 /*
218  * Lock for any access to the following:
219  *	vnode_list
220  *	numvnodes
221  *	freevnodes
222  */
223 static struct mtx __exclusive_cache_line vnode_list_mtx;
224 
225 /* Publicly exported FS */
226 struct nfs_public nfs_pub;
227 
228 static uma_zone_t buf_trie_zone;
229 static smr_t buf_trie_smr;
230 
231 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
232 static uma_zone_t vnode_zone;
233 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
234 
235 __read_frequently smr_t vfs_smr;
236 
237 /*
238  * The workitem queue.
239  *
240  * It is useful to delay writes of file data and filesystem metadata
241  * for tens of seconds so that quickly created and deleted files need
242  * not waste disk bandwidth being created and removed. To realize this,
243  * we append vnodes to a "workitem" queue. When running with a soft
244  * updates implementation, most pending metadata dependencies should
245  * not wait for more than a few seconds. Thus, mounted on block devices
246  * are delayed only about a half the time that file data is delayed.
247  * Similarly, directory updates are more critical, so are only delayed
248  * about a third the time that file data is delayed. Thus, there are
249  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
250  * one each second (driven off the filesystem syncer process). The
251  * syncer_delayno variable indicates the next queue that is to be processed.
252  * Items that need to be processed soon are placed in this queue:
253  *
254  *	syncer_workitem_pending[syncer_delayno]
255  *
256  * A delay of fifteen seconds is done by placing the request fifteen
257  * entries later in the queue:
258  *
259  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
260  *
261  */
262 static int syncer_delayno;
263 static long syncer_mask;
264 LIST_HEAD(synclist, bufobj);
265 static struct synclist *syncer_workitem_pending;
266 /*
267  * The sync_mtx protects:
268  *	bo->bo_synclist
269  *	sync_vnode_count
270  *	syncer_delayno
271  *	syncer_state
272  *	syncer_workitem_pending
273  *	syncer_worklist_len
274  *	rushjob
275  */
276 static struct mtx sync_mtx;
277 static struct cv sync_wakeup;
278 
279 #define SYNCER_MAXDELAY		32
280 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
281 static int syncdelay = 30;		/* max time to delay syncing data */
282 static int filedelay = 30;		/* time to delay syncing files */
283 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
284     "Time to delay syncing files (in seconds)");
285 static int dirdelay = 29;		/* time to delay syncing directories */
286 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
287     "Time to delay syncing directories (in seconds)");
288 static int metadelay = 28;		/* time to delay syncing metadata */
289 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
290     "Time to delay syncing metadata (in seconds)");
291 static int rushjob;		/* number of slots to run ASAP */
292 static int stat_rush_requests;	/* number of times I/O speeded up */
293 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
294     "Number of times I/O speeded up (rush requests)");
295 
296 #define	VDBATCH_SIZE 8
297 struct vdbatch {
298 	u_int index;
299 	long freevnodes;
300 	struct mtx lock;
301 	struct vnode *tab[VDBATCH_SIZE];
302 };
303 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
304 
305 static void	vdbatch_dequeue(struct vnode *vp);
306 
307 /*
308  * When shutting down the syncer, run it at four times normal speed.
309  */
310 #define SYNCER_SHUTDOWN_SPEEDUP		4
311 static int sync_vnode_count;
312 static int syncer_worklist_len;
313 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
314     syncer_state;
315 
316 /* Target for maximum number of vnodes. */
317 u_long desiredvnodes;
318 static u_long gapvnodes;		/* gap between wanted and desired */
319 static u_long vhiwat;		/* enough extras after expansion */
320 static u_long vlowat;		/* minimal extras before expansion */
321 static u_long vstir;		/* nonzero to stir non-free vnodes */
322 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
323 
324 static u_long vnlru_read_freevnodes(void);
325 
326 /*
327  * Note that no attempt is made to sanitize these parameters.
328  */
329 static int
330 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
331 {
332 	u_long val;
333 	int error;
334 
335 	val = desiredvnodes;
336 	error = sysctl_handle_long(oidp, &val, 0, req);
337 	if (error != 0 || req->newptr == NULL)
338 		return (error);
339 
340 	if (val == desiredvnodes)
341 		return (0);
342 	mtx_lock(&vnode_list_mtx);
343 	desiredvnodes = val;
344 	wantfreevnodes = desiredvnodes / 4;
345 	vnlru_recalc();
346 	mtx_unlock(&vnode_list_mtx);
347 	/*
348 	 * XXX There is no protection against multiple threads changing
349 	 * desiredvnodes at the same time. Locking above only helps vnlru and
350 	 * getnewvnode.
351 	 */
352 	vfs_hash_changesize(desiredvnodes);
353 	cache_changesize(desiredvnodes);
354 	return (0);
355 }
356 
357 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
358     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
359     "LU", "Target for maximum number of vnodes");
360 
361 static int
362 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
363 {
364 	u_long val;
365 	int error;
366 
367 	val = wantfreevnodes;
368 	error = sysctl_handle_long(oidp, &val, 0, req);
369 	if (error != 0 || req->newptr == NULL)
370 		return (error);
371 
372 	if (val == wantfreevnodes)
373 		return (0);
374 	mtx_lock(&vnode_list_mtx);
375 	wantfreevnodes = val;
376 	vnlru_recalc();
377 	mtx_unlock(&vnode_list_mtx);
378 	return (0);
379 }
380 
381 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
382     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
383     "LU", "Target for minimum number of \"free\" vnodes");
384 
385 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
386     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
387 static int vnlru_nowhere;
388 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
389     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
390 
391 static int
392 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
393 {
394 	struct vnode *vp;
395 	struct nameidata nd;
396 	char *buf;
397 	unsigned long ndflags;
398 	int error;
399 
400 	if (req->newptr == NULL)
401 		return (EINVAL);
402 	if (req->newlen >= PATH_MAX)
403 		return (E2BIG);
404 
405 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
406 	error = SYSCTL_IN(req, buf, req->newlen);
407 	if (error != 0)
408 		goto out;
409 
410 	buf[req->newlen] = '\0';
411 
412 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
413 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
414 	if ((error = namei(&nd)) != 0)
415 		goto out;
416 	vp = nd.ni_vp;
417 
418 	if (VN_IS_DOOMED(vp)) {
419 		/*
420 		 * This vnode is being recycled.  Return != 0 to let the caller
421 		 * know that the sysctl had no effect.  Return EAGAIN because a
422 		 * subsequent call will likely succeed (since namei will create
423 		 * a new vnode if necessary)
424 		 */
425 		error = EAGAIN;
426 		goto putvnode;
427 	}
428 
429 	counter_u64_add(recycles_count, 1);
430 	vgone(vp);
431 putvnode:
432 	NDFREE(&nd, 0);
433 out:
434 	free(buf, M_TEMP);
435 	return (error);
436 }
437 
438 static int
439 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
440 {
441 	struct thread *td = curthread;
442 	struct vnode *vp;
443 	struct file *fp;
444 	int error;
445 	int fd;
446 
447 	if (req->newptr == NULL)
448 		return (EBADF);
449 
450         error = sysctl_handle_int(oidp, &fd, 0, req);
451         if (error != 0)
452                 return (error);
453 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
454 	if (error != 0)
455 		return (error);
456 	vp = fp->f_vnode;
457 
458 	error = vn_lock(vp, LK_EXCLUSIVE);
459 	if (error != 0)
460 		goto drop;
461 
462 	counter_u64_add(recycles_count, 1);
463 	vgone(vp);
464 	VOP_UNLOCK(vp);
465 drop:
466 	fdrop(fp, td);
467 	return (error);
468 }
469 
470 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
471     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
472     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
473 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
474     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
475     sysctl_ftry_reclaim_vnode, "I",
476     "Try to reclaim a vnode by its file descriptor");
477 
478 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
479 static int vnsz2log;
480 
481 /*
482  * Support for the bufobj clean & dirty pctrie.
483  */
484 static void *
485 buf_trie_alloc(struct pctrie *ptree)
486 {
487 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
488 }
489 
490 static void
491 buf_trie_free(struct pctrie *ptree, void *node)
492 {
493 	uma_zfree_smr(buf_trie_zone, node);
494 }
495 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
496     buf_trie_smr);
497 
498 /*
499  * Initialize the vnode management data structures.
500  *
501  * Reevaluate the following cap on the number of vnodes after the physical
502  * memory size exceeds 512GB.  In the limit, as the physical memory size
503  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
504  */
505 #ifndef	MAXVNODES_MAX
506 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
507 #endif
508 
509 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
510 
511 static struct vnode *
512 vn_alloc_marker(struct mount *mp)
513 {
514 	struct vnode *vp;
515 
516 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
517 	vp->v_type = VMARKER;
518 	vp->v_mount = mp;
519 
520 	return (vp);
521 }
522 
523 static void
524 vn_free_marker(struct vnode *vp)
525 {
526 
527 	MPASS(vp->v_type == VMARKER);
528 	free(vp, M_VNODE_MARKER);
529 }
530 
531 #ifdef KASAN
532 static int
533 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
534 {
535 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
536 	return (0);
537 }
538 
539 static void
540 vnode_dtor(void *mem, int size, void *arg __unused)
541 {
542 	size_t end1, end2, off1, off2;
543 
544 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
545 	    offsetof(struct vnode, v_dbatchcpu),
546 	    "KASAN marks require updating");
547 
548 	off1 = offsetof(struct vnode, v_vnodelist);
549 	off2 = offsetof(struct vnode, v_dbatchcpu);
550 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
551 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
552 
553 	/*
554 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
555 	 * after the vnode has been freed.  Try to get some KASAN coverage by
556 	 * marking everything except those two fields as invalid.  Because
557 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
558 	 * the same 8-byte aligned word must also be marked valid.
559 	 */
560 
561 	/* Handle the area from the start until v_vnodelist... */
562 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
563 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
564 
565 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
566 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
567 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
568 	if (off2 > off1)
569 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
570 		    off2 - off1, KASAN_UMA_FREED);
571 
572 	/* ... and finally the area from v_dbatchcpu to the end. */
573 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
574 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
575 	    KASAN_UMA_FREED);
576 }
577 #endif /* KASAN */
578 
579 /*
580  * Initialize a vnode as it first enters the zone.
581  */
582 static int
583 vnode_init(void *mem, int size, int flags)
584 {
585 	struct vnode *vp;
586 
587 	vp = mem;
588 	bzero(vp, size);
589 	/*
590 	 * Setup locks.
591 	 */
592 	vp->v_vnlock = &vp->v_lock;
593 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
594 	/*
595 	 * By default, don't allow shared locks unless filesystems opt-in.
596 	 */
597 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
598 	    LK_NOSHARE | LK_IS_VNODE);
599 	/*
600 	 * Initialize bufobj.
601 	 */
602 	bufobj_init(&vp->v_bufobj, vp);
603 	/*
604 	 * Initialize namecache.
605 	 */
606 	cache_vnode_init(vp);
607 	/*
608 	 * Initialize rangelocks.
609 	 */
610 	rangelock_init(&vp->v_rl);
611 
612 	vp->v_dbatchcpu = NOCPU;
613 
614 	/*
615 	 * Check vhold_recycle_free for an explanation.
616 	 */
617 	vp->v_holdcnt = VHOLD_NO_SMR;
618 	vp->v_type = VNON;
619 	mtx_lock(&vnode_list_mtx);
620 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
621 	mtx_unlock(&vnode_list_mtx);
622 	return (0);
623 }
624 
625 /*
626  * Free a vnode when it is cleared from the zone.
627  */
628 static void
629 vnode_fini(void *mem, int size)
630 {
631 	struct vnode *vp;
632 	struct bufobj *bo;
633 
634 	vp = mem;
635 	vdbatch_dequeue(vp);
636 	mtx_lock(&vnode_list_mtx);
637 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
638 	mtx_unlock(&vnode_list_mtx);
639 	rangelock_destroy(&vp->v_rl);
640 	lockdestroy(vp->v_vnlock);
641 	mtx_destroy(&vp->v_interlock);
642 	bo = &vp->v_bufobj;
643 	rw_destroy(BO_LOCKPTR(bo));
644 
645 	kasan_mark(mem, size, size, 0);
646 }
647 
648 /*
649  * Provide the size of NFS nclnode and NFS fh for calculation of the
650  * vnode memory consumption.  The size is specified directly to
651  * eliminate dependency on NFS-private header.
652  *
653  * Other filesystems may use bigger or smaller (like UFS and ZFS)
654  * private inode data, but the NFS-based estimation is ample enough.
655  * Still, we care about differences in the size between 64- and 32-bit
656  * platforms.
657  *
658  * Namecache structure size is heuristically
659  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
660  */
661 #ifdef _LP64
662 #define	NFS_NCLNODE_SZ	(528 + 64)
663 #define	NC_SZ		148
664 #else
665 #define	NFS_NCLNODE_SZ	(360 + 32)
666 #define	NC_SZ		92
667 #endif
668 
669 static void
670 vntblinit(void *dummy __unused)
671 {
672 	struct vdbatch *vd;
673 	uma_ctor ctor;
674 	uma_dtor dtor;
675 	int cpu, physvnodes, virtvnodes;
676 	u_int i;
677 
678 	/*
679 	 * Desiredvnodes is a function of the physical memory size and the
680 	 * kernel's heap size.  Generally speaking, it scales with the
681 	 * physical memory size.  The ratio of desiredvnodes to the physical
682 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
683 	 * Thereafter, the
684 	 * marginal ratio of desiredvnodes to the physical memory size is
685 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
686 	 * size.  The memory required by desiredvnodes vnodes and vm objects
687 	 * must not exceed 1/10th of the kernel's heap size.
688 	 */
689 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
690 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
691 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
692 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
693 	desiredvnodes = min(physvnodes, virtvnodes);
694 	if (desiredvnodes > MAXVNODES_MAX) {
695 		if (bootverbose)
696 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
697 			    desiredvnodes, MAXVNODES_MAX);
698 		desiredvnodes = MAXVNODES_MAX;
699 	}
700 	wantfreevnodes = desiredvnodes / 4;
701 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
702 	TAILQ_INIT(&vnode_list);
703 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
704 	/*
705 	 * The lock is taken to appease WITNESS.
706 	 */
707 	mtx_lock(&vnode_list_mtx);
708 	vnlru_recalc();
709 	mtx_unlock(&vnode_list_mtx);
710 	vnode_list_free_marker = vn_alloc_marker(NULL);
711 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
712 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
713 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
714 
715 #ifdef KASAN
716 	ctor = vnode_ctor;
717 	dtor = vnode_dtor;
718 #else
719 	ctor = NULL;
720 	dtor = NULL;
721 #endif
722 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
723 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
724 	uma_zone_set_smr(vnode_zone, vfs_smr);
725 
726 	/*
727 	 * Preallocate enough nodes to support one-per buf so that
728 	 * we can not fail an insert.  reassignbuf() callers can not
729 	 * tolerate the insertion failure.
730 	 */
731 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
732 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
733 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
734 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
735 	uma_prealloc(buf_trie_zone, nbuf);
736 
737 	vnodes_created = counter_u64_alloc(M_WAITOK);
738 	recycles_count = counter_u64_alloc(M_WAITOK);
739 	recycles_free_count = counter_u64_alloc(M_WAITOK);
740 	deferred_inact = counter_u64_alloc(M_WAITOK);
741 
742 	/*
743 	 * Initialize the filesystem syncer.
744 	 */
745 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
746 	    &syncer_mask);
747 	syncer_maxdelay = syncer_mask + 1;
748 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
749 	cv_init(&sync_wakeup, "syncer");
750 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
751 		vnsz2log++;
752 	vnsz2log--;
753 
754 	CPU_FOREACH(cpu) {
755 		vd = DPCPU_ID_PTR((cpu), vd);
756 		bzero(vd, sizeof(*vd));
757 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
758 	}
759 }
760 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
761 
762 /*
763  * Mark a mount point as busy. Used to synchronize access and to delay
764  * unmounting. Eventually, mountlist_mtx is not released on failure.
765  *
766  * vfs_busy() is a custom lock, it can block the caller.
767  * vfs_busy() only sleeps if the unmount is active on the mount point.
768  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
769  * vnode belonging to mp.
770  *
771  * Lookup uses vfs_busy() to traverse mount points.
772  * root fs			var fs
773  * / vnode lock		A	/ vnode lock (/var)		D
774  * /var vnode lock	B	/log vnode lock(/var/log)	E
775  * vfs_busy lock	C	vfs_busy lock			F
776  *
777  * Within each file system, the lock order is C->A->B and F->D->E.
778  *
779  * When traversing across mounts, the system follows that lock order:
780  *
781  *        C->A->B
782  *              |
783  *              +->F->D->E
784  *
785  * The lookup() process for namei("/var") illustrates the process:
786  *  VOP_LOOKUP() obtains B while A is held
787  *  vfs_busy() obtains a shared lock on F while A and B are held
788  *  vput() releases lock on B
789  *  vput() releases lock on A
790  *  VFS_ROOT() obtains lock on D while shared lock on F is held
791  *  vfs_unbusy() releases shared lock on F
792  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
793  *    Attempt to lock A (instead of vp_crossmp) while D is held would
794  *    violate the global order, causing deadlocks.
795  *
796  * dounmount() locks B while F is drained.
797  */
798 int
799 vfs_busy(struct mount *mp, int flags)
800 {
801 	struct mount_pcpu *mpcpu;
802 
803 	MPASS((flags & ~MBF_MASK) == 0);
804 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
805 
806 	if (vfs_op_thread_enter(mp, mpcpu)) {
807 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
808 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
809 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
810 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
811 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
812 		vfs_op_thread_exit(mp, mpcpu);
813 		if (flags & MBF_MNTLSTLOCK)
814 			mtx_unlock(&mountlist_mtx);
815 		return (0);
816 	}
817 
818 	MNT_ILOCK(mp);
819 	vfs_assert_mount_counters(mp);
820 	MNT_REF(mp);
821 	/*
822 	 * If mount point is currently being unmounted, sleep until the
823 	 * mount point fate is decided.  If thread doing the unmounting fails,
824 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
825 	 * that this mount point has survived the unmount attempt and vfs_busy
826 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
827 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
828 	 * about to be really destroyed.  vfs_busy needs to release its
829 	 * reference on the mount point in this case and return with ENOENT,
830 	 * telling the caller that mount mount it tried to busy is no longer
831 	 * valid.
832 	 */
833 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
834 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
835 			MNT_REL(mp);
836 			MNT_IUNLOCK(mp);
837 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
838 			    __func__);
839 			return (ENOENT);
840 		}
841 		if (flags & MBF_MNTLSTLOCK)
842 			mtx_unlock(&mountlist_mtx);
843 		mp->mnt_kern_flag |= MNTK_MWAIT;
844 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
845 		if (flags & MBF_MNTLSTLOCK)
846 			mtx_lock(&mountlist_mtx);
847 		MNT_ILOCK(mp);
848 	}
849 	if (flags & MBF_MNTLSTLOCK)
850 		mtx_unlock(&mountlist_mtx);
851 	mp->mnt_lockref++;
852 	MNT_IUNLOCK(mp);
853 	return (0);
854 }
855 
856 /*
857  * Free a busy filesystem.
858  */
859 void
860 vfs_unbusy(struct mount *mp)
861 {
862 	struct mount_pcpu *mpcpu;
863 	int c;
864 
865 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
866 
867 	if (vfs_op_thread_enter(mp, mpcpu)) {
868 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
869 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
870 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
871 		vfs_op_thread_exit(mp, mpcpu);
872 		return;
873 	}
874 
875 	MNT_ILOCK(mp);
876 	vfs_assert_mount_counters(mp);
877 	MNT_REL(mp);
878 	c = --mp->mnt_lockref;
879 	if (mp->mnt_vfs_ops == 0) {
880 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
881 		MNT_IUNLOCK(mp);
882 		return;
883 	}
884 	if (c < 0)
885 		vfs_dump_mount_counters(mp);
886 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
887 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
888 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
889 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
890 		wakeup(&mp->mnt_lockref);
891 	}
892 	MNT_IUNLOCK(mp);
893 }
894 
895 /*
896  * Lookup a mount point by filesystem identifier.
897  */
898 struct mount *
899 vfs_getvfs(fsid_t *fsid)
900 {
901 	struct mount *mp;
902 
903 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
904 	mtx_lock(&mountlist_mtx);
905 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
906 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
907 			vfs_ref(mp);
908 			mtx_unlock(&mountlist_mtx);
909 			return (mp);
910 		}
911 	}
912 	mtx_unlock(&mountlist_mtx);
913 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
914 	return ((struct mount *) 0);
915 }
916 
917 /*
918  * Lookup a mount point by filesystem identifier, busying it before
919  * returning.
920  *
921  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
922  * cache for popular filesystem identifiers.  The cache is lockess, using
923  * the fact that struct mount's are never freed.  In worst case we may
924  * get pointer to unmounted or even different filesystem, so we have to
925  * check what we got, and go slow way if so.
926  */
927 struct mount *
928 vfs_busyfs(fsid_t *fsid)
929 {
930 #define	FSID_CACHE_SIZE	256
931 	typedef struct mount * volatile vmp_t;
932 	static vmp_t cache[FSID_CACHE_SIZE];
933 	struct mount *mp;
934 	int error;
935 	uint32_t hash;
936 
937 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
938 	hash = fsid->val[0] ^ fsid->val[1];
939 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
940 	mp = cache[hash];
941 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
942 		goto slow;
943 	if (vfs_busy(mp, 0) != 0) {
944 		cache[hash] = NULL;
945 		goto slow;
946 	}
947 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
948 		return (mp);
949 	else
950 	    vfs_unbusy(mp);
951 
952 slow:
953 	mtx_lock(&mountlist_mtx);
954 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
955 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
956 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
957 			if (error) {
958 				cache[hash] = NULL;
959 				mtx_unlock(&mountlist_mtx);
960 				return (NULL);
961 			}
962 			cache[hash] = mp;
963 			return (mp);
964 		}
965 	}
966 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
967 	mtx_unlock(&mountlist_mtx);
968 	return ((struct mount *) 0);
969 }
970 
971 /*
972  * Check if a user can access privileged mount options.
973  */
974 int
975 vfs_suser(struct mount *mp, struct thread *td)
976 {
977 	int error;
978 
979 	if (jailed(td->td_ucred)) {
980 		/*
981 		 * If the jail of the calling thread lacks permission for
982 		 * this type of file system, deny immediately.
983 		 */
984 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
985 			return (EPERM);
986 
987 		/*
988 		 * If the file system was mounted outside the jail of the
989 		 * calling thread, deny immediately.
990 		 */
991 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
992 			return (EPERM);
993 	}
994 
995 	/*
996 	 * If file system supports delegated administration, we don't check
997 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
998 	 * by the file system itself.
999 	 * If this is not the user that did original mount, we check for
1000 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1001 	 */
1002 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1003 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1004 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1005 			return (error);
1006 	}
1007 	return (0);
1008 }
1009 
1010 /*
1011  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1012  * will be used to create fake device numbers for stat().  Also try (but
1013  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1014  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1015  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1016  *
1017  * Keep in mind that several mounts may be running in parallel.  Starting
1018  * the search one past where the previous search terminated is both a
1019  * micro-optimization and a defense against returning the same fsid to
1020  * different mounts.
1021  */
1022 void
1023 vfs_getnewfsid(struct mount *mp)
1024 {
1025 	static uint16_t mntid_base;
1026 	struct mount *nmp;
1027 	fsid_t tfsid;
1028 	int mtype;
1029 
1030 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1031 	mtx_lock(&mntid_mtx);
1032 	mtype = mp->mnt_vfc->vfc_typenum;
1033 	tfsid.val[1] = mtype;
1034 	mtype = (mtype & 0xFF) << 24;
1035 	for (;;) {
1036 		tfsid.val[0] = makedev(255,
1037 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1038 		mntid_base++;
1039 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1040 			break;
1041 		vfs_rel(nmp);
1042 	}
1043 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1044 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1045 	mtx_unlock(&mntid_mtx);
1046 }
1047 
1048 /*
1049  * Knob to control the precision of file timestamps:
1050  *
1051  *   0 = seconds only; nanoseconds zeroed.
1052  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1053  *   2 = seconds and nanoseconds, truncated to microseconds.
1054  * >=3 = seconds and nanoseconds, maximum precision.
1055  */
1056 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1057 
1058 static int timestamp_precision = TSP_USEC;
1059 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1060     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1061     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1062     "3+: sec + ns (max. precision))");
1063 
1064 /*
1065  * Get a current timestamp.
1066  */
1067 void
1068 vfs_timestamp(struct timespec *tsp)
1069 {
1070 	struct timeval tv;
1071 
1072 	switch (timestamp_precision) {
1073 	case TSP_SEC:
1074 		tsp->tv_sec = time_second;
1075 		tsp->tv_nsec = 0;
1076 		break;
1077 	case TSP_HZ:
1078 		getnanotime(tsp);
1079 		break;
1080 	case TSP_USEC:
1081 		microtime(&tv);
1082 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1083 		break;
1084 	case TSP_NSEC:
1085 	default:
1086 		nanotime(tsp);
1087 		break;
1088 	}
1089 }
1090 
1091 /*
1092  * Set vnode attributes to VNOVAL
1093  */
1094 void
1095 vattr_null(struct vattr *vap)
1096 {
1097 
1098 	vap->va_type = VNON;
1099 	vap->va_size = VNOVAL;
1100 	vap->va_bytes = VNOVAL;
1101 	vap->va_mode = VNOVAL;
1102 	vap->va_nlink = VNOVAL;
1103 	vap->va_uid = VNOVAL;
1104 	vap->va_gid = VNOVAL;
1105 	vap->va_fsid = VNOVAL;
1106 	vap->va_fileid = VNOVAL;
1107 	vap->va_blocksize = VNOVAL;
1108 	vap->va_rdev = VNOVAL;
1109 	vap->va_atime.tv_sec = VNOVAL;
1110 	vap->va_atime.tv_nsec = VNOVAL;
1111 	vap->va_mtime.tv_sec = VNOVAL;
1112 	vap->va_mtime.tv_nsec = VNOVAL;
1113 	vap->va_ctime.tv_sec = VNOVAL;
1114 	vap->va_ctime.tv_nsec = VNOVAL;
1115 	vap->va_birthtime.tv_sec = VNOVAL;
1116 	vap->va_birthtime.tv_nsec = VNOVAL;
1117 	vap->va_flags = VNOVAL;
1118 	vap->va_gen = VNOVAL;
1119 	vap->va_vaflags = 0;
1120 }
1121 
1122 /*
1123  * Try to reduce the total number of vnodes.
1124  *
1125  * This routine (and its user) are buggy in at least the following ways:
1126  * - all parameters were picked years ago when RAM sizes were significantly
1127  *   smaller
1128  * - it can pick vnodes based on pages used by the vm object, but filesystems
1129  *   like ZFS don't use it making the pick broken
1130  * - since ZFS has its own aging policy it gets partially combated by this one
1131  * - a dedicated method should be provided for filesystems to let them decide
1132  *   whether the vnode should be recycled
1133  *
1134  * This routine is called when we have too many vnodes.  It attempts
1135  * to free <count> vnodes and will potentially free vnodes that still
1136  * have VM backing store (VM backing store is typically the cause
1137  * of a vnode blowout so we want to do this).  Therefore, this operation
1138  * is not considered cheap.
1139  *
1140  * A number of conditions may prevent a vnode from being reclaimed.
1141  * the buffer cache may have references on the vnode, a directory
1142  * vnode may still have references due to the namei cache representing
1143  * underlying files, or the vnode may be in active use.   It is not
1144  * desirable to reuse such vnodes.  These conditions may cause the
1145  * number of vnodes to reach some minimum value regardless of what
1146  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1147  *
1148  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1149  * 			 entries if this argument is strue
1150  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1151  *			 pages.
1152  * @param target	 How many vnodes to reclaim.
1153  * @return		 The number of vnodes that were reclaimed.
1154  */
1155 static int
1156 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1157 {
1158 	struct vnode *vp, *mvp;
1159 	struct mount *mp;
1160 	struct vm_object *object;
1161 	u_long done;
1162 	bool retried;
1163 
1164 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1165 
1166 	retried = false;
1167 	done = 0;
1168 
1169 	mvp = vnode_list_reclaim_marker;
1170 restart:
1171 	vp = mvp;
1172 	while (done < target) {
1173 		vp = TAILQ_NEXT(vp, v_vnodelist);
1174 		if (__predict_false(vp == NULL))
1175 			break;
1176 
1177 		if (__predict_false(vp->v_type == VMARKER))
1178 			continue;
1179 
1180 		/*
1181 		 * If it's been deconstructed already, it's still
1182 		 * referenced, or it exceeds the trigger, skip it.
1183 		 * Also skip free vnodes.  We are trying to make space
1184 		 * to expand the free list, not reduce it.
1185 		 */
1186 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1187 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1188 			goto next_iter;
1189 
1190 		if (vp->v_type == VBAD || vp->v_type == VNON)
1191 			goto next_iter;
1192 
1193 		object = atomic_load_ptr(&vp->v_object);
1194 		if (object == NULL || object->resident_page_count > trigger) {
1195 			goto next_iter;
1196 		}
1197 
1198 		/*
1199 		 * Handle races against vnode allocation. Filesystems lock the
1200 		 * vnode some time after it gets returned from getnewvnode,
1201 		 * despite type and hold count being manipulated earlier.
1202 		 * Resorting to checking v_mount restores guarantees present
1203 		 * before the global list was reworked to contain all vnodes.
1204 		 */
1205 		if (!VI_TRYLOCK(vp))
1206 			goto next_iter;
1207 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1208 			VI_UNLOCK(vp);
1209 			goto next_iter;
1210 		}
1211 		if (vp->v_mount == NULL) {
1212 			VI_UNLOCK(vp);
1213 			goto next_iter;
1214 		}
1215 		vholdl(vp);
1216 		VI_UNLOCK(vp);
1217 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1218 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1219 		mtx_unlock(&vnode_list_mtx);
1220 
1221 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1222 			vdrop(vp);
1223 			goto next_iter_unlocked;
1224 		}
1225 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1226 			vdrop(vp);
1227 			vn_finished_write(mp);
1228 			goto next_iter_unlocked;
1229 		}
1230 
1231 		VI_LOCK(vp);
1232 		if (vp->v_usecount > 0 ||
1233 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1234 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1235 		    vp->v_object->resident_page_count > trigger)) {
1236 			VOP_UNLOCK(vp);
1237 			vdropl(vp);
1238 			vn_finished_write(mp);
1239 			goto next_iter_unlocked;
1240 		}
1241 		counter_u64_add(recycles_count, 1);
1242 		vgonel(vp);
1243 		VOP_UNLOCK(vp);
1244 		vdropl(vp);
1245 		vn_finished_write(mp);
1246 		done++;
1247 next_iter_unlocked:
1248 		if (should_yield())
1249 			kern_yield(PRI_USER);
1250 		mtx_lock(&vnode_list_mtx);
1251 		goto restart;
1252 next_iter:
1253 		MPASS(vp->v_type != VMARKER);
1254 		if (!should_yield())
1255 			continue;
1256 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1257 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1258 		mtx_unlock(&vnode_list_mtx);
1259 		kern_yield(PRI_USER);
1260 		mtx_lock(&vnode_list_mtx);
1261 		goto restart;
1262 	}
1263 	if (done == 0 && !retried) {
1264 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1265 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1266 		retried = true;
1267 		goto restart;
1268 	}
1269 	return (done);
1270 }
1271 
1272 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1273 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1274     0,
1275     "limit on vnode free requests per call to the vnlru_free routine");
1276 
1277 /*
1278  * Attempt to reduce the free list by the requested amount.
1279  */
1280 static int
1281 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1282 {
1283 	struct vnode *vp;
1284 	struct mount *mp;
1285 	int ocount;
1286 
1287 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1288 	if (count > max_vnlru_free)
1289 		count = max_vnlru_free;
1290 	ocount = count;
1291 	vp = mvp;
1292 	for (;;) {
1293 		if (count == 0) {
1294 			break;
1295 		}
1296 		vp = TAILQ_NEXT(vp, v_vnodelist);
1297 		if (__predict_false(vp == NULL)) {
1298 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1299 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1300 			break;
1301 		}
1302 		if (__predict_false(vp->v_type == VMARKER))
1303 			continue;
1304 		if (vp->v_holdcnt > 0)
1305 			continue;
1306 		/*
1307 		 * Don't recycle if our vnode is from different type
1308 		 * of mount point.  Note that mp is type-safe, the
1309 		 * check does not reach unmapped address even if
1310 		 * vnode is reclaimed.
1311 		 */
1312 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1313 		    mp->mnt_op != mnt_op) {
1314 			continue;
1315 		}
1316 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1317 			continue;
1318 		}
1319 		if (!vhold_recycle_free(vp))
1320 			continue;
1321 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1322 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1323 		mtx_unlock(&vnode_list_mtx);
1324 		if (vtryrecycle(vp) == 0)
1325 			count--;
1326 		mtx_lock(&vnode_list_mtx);
1327 		vp = mvp;
1328 	}
1329 	return (ocount - count);
1330 }
1331 
1332 static int
1333 vnlru_free_locked(int count)
1334 {
1335 
1336 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1337 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1338 }
1339 
1340 void
1341 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1342 {
1343 
1344 	MPASS(mnt_op != NULL);
1345 	MPASS(mvp != NULL);
1346 	VNPASS(mvp->v_type == VMARKER, mvp);
1347 	mtx_lock(&vnode_list_mtx);
1348 	vnlru_free_impl(count, mnt_op, mvp);
1349 	mtx_unlock(&vnode_list_mtx);
1350 }
1351 
1352 struct vnode *
1353 vnlru_alloc_marker(void)
1354 {
1355 	struct vnode *mvp;
1356 
1357 	mvp = vn_alloc_marker(NULL);
1358 	mtx_lock(&vnode_list_mtx);
1359 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1360 	mtx_unlock(&vnode_list_mtx);
1361 	return (mvp);
1362 }
1363 
1364 void
1365 vnlru_free_marker(struct vnode *mvp)
1366 {
1367 	mtx_lock(&vnode_list_mtx);
1368 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1369 	mtx_unlock(&vnode_list_mtx);
1370 	vn_free_marker(mvp);
1371 }
1372 
1373 static void
1374 vnlru_recalc(void)
1375 {
1376 
1377 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1378 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1379 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1380 	vlowat = vhiwat / 2;
1381 }
1382 
1383 /*
1384  * Attempt to recycle vnodes in a context that is always safe to block.
1385  * Calling vlrurecycle() from the bowels of filesystem code has some
1386  * interesting deadlock problems.
1387  */
1388 static struct proc *vnlruproc;
1389 static int vnlruproc_sig;
1390 
1391 /*
1392  * The main freevnodes counter is only updated when threads requeue their vnode
1393  * batches. CPUs are conditionally walked to compute a more accurate total.
1394  *
1395  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1396  * at any given moment can still exceed slop, but it should not be by significant
1397  * margin in practice.
1398  */
1399 #define VNLRU_FREEVNODES_SLOP 128
1400 
1401 static __inline void
1402 vfs_freevnodes_inc(void)
1403 {
1404 	struct vdbatch *vd;
1405 
1406 	critical_enter();
1407 	vd = DPCPU_PTR(vd);
1408 	vd->freevnodes++;
1409 	critical_exit();
1410 }
1411 
1412 static __inline void
1413 vfs_freevnodes_dec(void)
1414 {
1415 	struct vdbatch *vd;
1416 
1417 	critical_enter();
1418 	vd = DPCPU_PTR(vd);
1419 	vd->freevnodes--;
1420 	critical_exit();
1421 }
1422 
1423 static u_long
1424 vnlru_read_freevnodes(void)
1425 {
1426 	struct vdbatch *vd;
1427 	long slop;
1428 	int cpu;
1429 
1430 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1431 	if (freevnodes > freevnodes_old)
1432 		slop = freevnodes - freevnodes_old;
1433 	else
1434 		slop = freevnodes_old - freevnodes;
1435 	if (slop < VNLRU_FREEVNODES_SLOP)
1436 		return (freevnodes >= 0 ? freevnodes : 0);
1437 	freevnodes_old = freevnodes;
1438 	CPU_FOREACH(cpu) {
1439 		vd = DPCPU_ID_PTR((cpu), vd);
1440 		freevnodes_old += vd->freevnodes;
1441 	}
1442 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1443 }
1444 
1445 static bool
1446 vnlru_under(u_long rnumvnodes, u_long limit)
1447 {
1448 	u_long rfreevnodes, space;
1449 
1450 	if (__predict_false(rnumvnodes > desiredvnodes))
1451 		return (true);
1452 
1453 	space = desiredvnodes - rnumvnodes;
1454 	if (space < limit) {
1455 		rfreevnodes = vnlru_read_freevnodes();
1456 		if (rfreevnodes > wantfreevnodes)
1457 			space += rfreevnodes - wantfreevnodes;
1458 	}
1459 	return (space < limit);
1460 }
1461 
1462 static bool
1463 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1464 {
1465 	long rfreevnodes, space;
1466 
1467 	if (__predict_false(rnumvnodes > desiredvnodes))
1468 		return (true);
1469 
1470 	space = desiredvnodes - rnumvnodes;
1471 	if (space < limit) {
1472 		rfreevnodes = atomic_load_long(&freevnodes);
1473 		if (rfreevnodes > wantfreevnodes)
1474 			space += rfreevnodes - wantfreevnodes;
1475 	}
1476 	return (space < limit);
1477 }
1478 
1479 static void
1480 vnlru_kick(void)
1481 {
1482 
1483 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1484 	if (vnlruproc_sig == 0) {
1485 		vnlruproc_sig = 1;
1486 		wakeup(vnlruproc);
1487 	}
1488 }
1489 
1490 static void
1491 vnlru_proc(void)
1492 {
1493 	u_long rnumvnodes, rfreevnodes, target;
1494 	unsigned long onumvnodes;
1495 	int done, force, trigger, usevnodes;
1496 	bool reclaim_nc_src, want_reread;
1497 
1498 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1499 	    SHUTDOWN_PRI_FIRST);
1500 
1501 	force = 0;
1502 	want_reread = false;
1503 	for (;;) {
1504 		kproc_suspend_check(vnlruproc);
1505 		mtx_lock(&vnode_list_mtx);
1506 		rnumvnodes = atomic_load_long(&numvnodes);
1507 
1508 		if (want_reread) {
1509 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1510 			want_reread = false;
1511 		}
1512 
1513 		/*
1514 		 * If numvnodes is too large (due to desiredvnodes being
1515 		 * adjusted using its sysctl, or emergency growth), first
1516 		 * try to reduce it by discarding from the free list.
1517 		 */
1518 		if (rnumvnodes > desiredvnodes) {
1519 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1520 			rnumvnodes = atomic_load_long(&numvnodes);
1521 		}
1522 		/*
1523 		 * Sleep if the vnode cache is in a good state.  This is
1524 		 * when it is not over-full and has space for about a 4%
1525 		 * or 9% expansion (by growing its size or inexcessively
1526 		 * reducing its free list).  Otherwise, try to reclaim
1527 		 * space for a 10% expansion.
1528 		 */
1529 		if (vstir && force == 0) {
1530 			force = 1;
1531 			vstir = 0;
1532 		}
1533 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1534 			vnlruproc_sig = 0;
1535 			wakeup(&vnlruproc_sig);
1536 			msleep(vnlruproc, &vnode_list_mtx,
1537 			    PVFS|PDROP, "vlruwt", hz);
1538 			continue;
1539 		}
1540 		rfreevnodes = vnlru_read_freevnodes();
1541 
1542 		onumvnodes = rnumvnodes;
1543 		/*
1544 		 * Calculate parameters for recycling.  These are the same
1545 		 * throughout the loop to give some semblance of fairness.
1546 		 * The trigger point is to avoid recycling vnodes with lots
1547 		 * of resident pages.  We aren't trying to free memory; we
1548 		 * are trying to recycle or at least free vnodes.
1549 		 */
1550 		if (rnumvnodes <= desiredvnodes)
1551 			usevnodes = rnumvnodes - rfreevnodes;
1552 		else
1553 			usevnodes = rnumvnodes;
1554 		if (usevnodes <= 0)
1555 			usevnodes = 1;
1556 		/*
1557 		 * The trigger value is is chosen to give a conservatively
1558 		 * large value to ensure that it alone doesn't prevent
1559 		 * making progress.  The value can easily be so large that
1560 		 * it is effectively infinite in some congested and
1561 		 * misconfigured cases, and this is necessary.  Normally
1562 		 * it is about 8 to 100 (pages), which is quite large.
1563 		 */
1564 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1565 		if (force < 2)
1566 			trigger = vsmalltrigger;
1567 		reclaim_nc_src = force >= 3;
1568 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1569 		target = target / 10 + 1;
1570 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1571 		mtx_unlock(&vnode_list_mtx);
1572 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1573 			uma_reclaim(UMA_RECLAIM_DRAIN);
1574 		if (done == 0) {
1575 			if (force == 0 || force == 1) {
1576 				force = 2;
1577 				continue;
1578 			}
1579 			if (force == 2) {
1580 				force = 3;
1581 				continue;
1582 			}
1583 			want_reread = true;
1584 			force = 0;
1585 			vnlru_nowhere++;
1586 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1587 		} else {
1588 			want_reread = true;
1589 			kern_yield(PRI_USER);
1590 		}
1591 	}
1592 }
1593 
1594 static struct kproc_desc vnlru_kp = {
1595 	"vnlru",
1596 	vnlru_proc,
1597 	&vnlruproc
1598 };
1599 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1600     &vnlru_kp);
1601 
1602 /*
1603  * Routines having to do with the management of the vnode table.
1604  */
1605 
1606 /*
1607  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1608  * before we actually vgone().  This function must be called with the vnode
1609  * held to prevent the vnode from being returned to the free list midway
1610  * through vgone().
1611  */
1612 static int
1613 vtryrecycle(struct vnode *vp)
1614 {
1615 	struct mount *vnmp;
1616 
1617 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1618 	VNASSERT(vp->v_holdcnt, vp,
1619 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1620 	/*
1621 	 * This vnode may found and locked via some other list, if so we
1622 	 * can't recycle it yet.
1623 	 */
1624 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1625 		CTR2(KTR_VFS,
1626 		    "%s: impossible to recycle, vp %p lock is already held",
1627 		    __func__, vp);
1628 		vdrop(vp);
1629 		return (EWOULDBLOCK);
1630 	}
1631 	/*
1632 	 * Don't recycle if its filesystem is being suspended.
1633 	 */
1634 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1635 		VOP_UNLOCK(vp);
1636 		CTR2(KTR_VFS,
1637 		    "%s: impossible to recycle, cannot start the write for %p",
1638 		    __func__, vp);
1639 		vdrop(vp);
1640 		return (EBUSY);
1641 	}
1642 	/*
1643 	 * If we got this far, we need to acquire the interlock and see if
1644 	 * anyone picked up this vnode from another list.  If not, we will
1645 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1646 	 * will skip over it.
1647 	 */
1648 	VI_LOCK(vp);
1649 	if (vp->v_usecount) {
1650 		VOP_UNLOCK(vp);
1651 		vdropl(vp);
1652 		vn_finished_write(vnmp);
1653 		CTR2(KTR_VFS,
1654 		    "%s: impossible to recycle, %p is already referenced",
1655 		    __func__, vp);
1656 		return (EBUSY);
1657 	}
1658 	if (!VN_IS_DOOMED(vp)) {
1659 		counter_u64_add(recycles_free_count, 1);
1660 		vgonel(vp);
1661 	}
1662 	VOP_UNLOCK(vp);
1663 	vdropl(vp);
1664 	vn_finished_write(vnmp);
1665 	return (0);
1666 }
1667 
1668 /*
1669  * Allocate a new vnode.
1670  *
1671  * The operation never returns an error. Returning an error was disabled
1672  * in r145385 (dated 2005) with the following comment:
1673  *
1674  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1675  *
1676  * Given the age of this commit (almost 15 years at the time of writing this
1677  * comment) restoring the ability to fail requires a significant audit of
1678  * all codepaths.
1679  *
1680  * The routine can try to free a vnode or stall for up to 1 second waiting for
1681  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1682  */
1683 static u_long vn_alloc_cyclecount;
1684 
1685 static struct vnode * __noinline
1686 vn_alloc_hard(struct mount *mp)
1687 {
1688 	u_long rnumvnodes, rfreevnodes;
1689 
1690 	mtx_lock(&vnode_list_mtx);
1691 	rnumvnodes = atomic_load_long(&numvnodes);
1692 	if (rnumvnodes + 1 < desiredvnodes) {
1693 		vn_alloc_cyclecount = 0;
1694 		goto alloc;
1695 	}
1696 	rfreevnodes = vnlru_read_freevnodes();
1697 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1698 		vn_alloc_cyclecount = 0;
1699 		vstir = 1;
1700 	}
1701 	/*
1702 	 * Grow the vnode cache if it will not be above its target max
1703 	 * after growing.  Otherwise, if the free list is nonempty, try
1704 	 * to reclaim 1 item from it before growing the cache (possibly
1705 	 * above its target max if the reclamation failed or is delayed).
1706 	 * Otherwise, wait for some space.  In all cases, schedule
1707 	 * vnlru_proc() if we are getting short of space.  The watermarks
1708 	 * should be chosen so that we never wait or even reclaim from
1709 	 * the free list to below its target minimum.
1710 	 */
1711 	if (vnlru_free_locked(1) > 0)
1712 		goto alloc;
1713 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1714 		/*
1715 		 * Wait for space for a new vnode.
1716 		 */
1717 		vnlru_kick();
1718 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1719 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1720 		    vnlru_read_freevnodes() > 1)
1721 			vnlru_free_locked(1);
1722 	}
1723 alloc:
1724 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1725 	if (vnlru_under(rnumvnodes, vlowat))
1726 		vnlru_kick();
1727 	mtx_unlock(&vnode_list_mtx);
1728 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1729 }
1730 
1731 static struct vnode *
1732 vn_alloc(struct mount *mp)
1733 {
1734 	u_long rnumvnodes;
1735 
1736 	if (__predict_false(vn_alloc_cyclecount != 0))
1737 		return (vn_alloc_hard(mp));
1738 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1739 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1740 		atomic_subtract_long(&numvnodes, 1);
1741 		return (vn_alloc_hard(mp));
1742 	}
1743 
1744 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1745 }
1746 
1747 static void
1748 vn_free(struct vnode *vp)
1749 {
1750 
1751 	atomic_subtract_long(&numvnodes, 1);
1752 	uma_zfree_smr(vnode_zone, vp);
1753 }
1754 
1755 /*
1756  * Return the next vnode from the free list.
1757  */
1758 int
1759 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1760     struct vnode **vpp)
1761 {
1762 	struct vnode *vp;
1763 	struct thread *td;
1764 	struct lock_object *lo;
1765 
1766 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1767 
1768 	KASSERT(vops->registered,
1769 	    ("%s: not registered vector op %p\n", __func__, vops));
1770 
1771 	td = curthread;
1772 	if (td->td_vp_reserved != NULL) {
1773 		vp = td->td_vp_reserved;
1774 		td->td_vp_reserved = NULL;
1775 	} else {
1776 		vp = vn_alloc(mp);
1777 	}
1778 	counter_u64_add(vnodes_created, 1);
1779 	/*
1780 	 * Locks are given the generic name "vnode" when created.
1781 	 * Follow the historic practice of using the filesystem
1782 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1783 	 *
1784 	 * Locks live in a witness group keyed on their name. Thus,
1785 	 * when a lock is renamed, it must also move from the witness
1786 	 * group of its old name to the witness group of its new name.
1787 	 *
1788 	 * The change only needs to be made when the vnode moves
1789 	 * from one filesystem type to another. We ensure that each
1790 	 * filesystem use a single static name pointer for its tag so
1791 	 * that we can compare pointers rather than doing a strcmp().
1792 	 */
1793 	lo = &vp->v_vnlock->lock_object;
1794 #ifdef WITNESS
1795 	if (lo->lo_name != tag) {
1796 #endif
1797 		lo->lo_name = tag;
1798 #ifdef WITNESS
1799 		WITNESS_DESTROY(lo);
1800 		WITNESS_INIT(lo, tag);
1801 	}
1802 #endif
1803 	/*
1804 	 * By default, don't allow shared locks unless filesystems opt-in.
1805 	 */
1806 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1807 	/*
1808 	 * Finalize various vnode identity bits.
1809 	 */
1810 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1811 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1812 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1813 	vp->v_type = VNON;
1814 	vp->v_op = vops;
1815 	vp->v_irflag = 0;
1816 	v_init_counters(vp);
1817 	vn_seqc_init(vp);
1818 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1819 #ifdef DIAGNOSTIC
1820 	if (mp == NULL && vops != &dead_vnodeops)
1821 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1822 #endif
1823 #ifdef MAC
1824 	mac_vnode_init(vp);
1825 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1826 		mac_vnode_associate_singlelabel(mp, vp);
1827 #endif
1828 	if (mp != NULL) {
1829 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1830 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1831 			vp->v_vflag |= VV_NOKNOTE;
1832 	}
1833 
1834 	/*
1835 	 * For the filesystems which do not use vfs_hash_insert(),
1836 	 * still initialize v_hash to have vfs_hash_index() useful.
1837 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1838 	 * its own hashing.
1839 	 */
1840 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1841 
1842 	*vpp = vp;
1843 	return (0);
1844 }
1845 
1846 void
1847 getnewvnode_reserve(void)
1848 {
1849 	struct thread *td;
1850 
1851 	td = curthread;
1852 	MPASS(td->td_vp_reserved == NULL);
1853 	td->td_vp_reserved = vn_alloc(NULL);
1854 }
1855 
1856 void
1857 getnewvnode_drop_reserve(void)
1858 {
1859 	struct thread *td;
1860 
1861 	td = curthread;
1862 	if (td->td_vp_reserved != NULL) {
1863 		vn_free(td->td_vp_reserved);
1864 		td->td_vp_reserved = NULL;
1865 	}
1866 }
1867 
1868 static void __noinline
1869 freevnode(struct vnode *vp)
1870 {
1871 	struct bufobj *bo;
1872 
1873 	/*
1874 	 * The vnode has been marked for destruction, so free it.
1875 	 *
1876 	 * The vnode will be returned to the zone where it will
1877 	 * normally remain until it is needed for another vnode. We
1878 	 * need to cleanup (or verify that the cleanup has already
1879 	 * been done) any residual data left from its current use
1880 	 * so as not to contaminate the freshly allocated vnode.
1881 	 */
1882 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1883 	/*
1884 	 * Paired with vgone.
1885 	 */
1886 	vn_seqc_write_end_free(vp);
1887 
1888 	bo = &vp->v_bufobj;
1889 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1890 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1891 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1892 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1893 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1894 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1895 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1896 	    ("clean blk trie not empty"));
1897 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1898 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1899 	    ("dirty blk trie not empty"));
1900 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1901 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1902 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1903 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1904 	    ("Dangling rangelock waiters"));
1905 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1906 	    ("Leaked inactivation"));
1907 	VI_UNLOCK(vp);
1908 #ifdef MAC
1909 	mac_vnode_destroy(vp);
1910 #endif
1911 	if (vp->v_pollinfo != NULL) {
1912 		destroy_vpollinfo(vp->v_pollinfo);
1913 		vp->v_pollinfo = NULL;
1914 	}
1915 	vp->v_mountedhere = NULL;
1916 	vp->v_unpcb = NULL;
1917 	vp->v_rdev = NULL;
1918 	vp->v_fifoinfo = NULL;
1919 	vp->v_iflag = 0;
1920 	vp->v_vflag = 0;
1921 	bo->bo_flag = 0;
1922 	vn_free(vp);
1923 }
1924 
1925 /*
1926  * Delete from old mount point vnode list, if on one.
1927  */
1928 static void
1929 delmntque(struct vnode *vp)
1930 {
1931 	struct mount *mp;
1932 
1933 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1934 
1935 	mp = vp->v_mount;
1936 	if (mp == NULL)
1937 		return;
1938 	MNT_ILOCK(mp);
1939 	VI_LOCK(vp);
1940 	vp->v_mount = NULL;
1941 	VI_UNLOCK(vp);
1942 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1943 		("bad mount point vnode list size"));
1944 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1945 	mp->mnt_nvnodelistsize--;
1946 	MNT_REL(mp);
1947 	MNT_IUNLOCK(mp);
1948 }
1949 
1950 static void
1951 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1952 {
1953 
1954 	vp->v_data = NULL;
1955 	vp->v_op = &dead_vnodeops;
1956 	vgone(vp);
1957 	vput(vp);
1958 }
1959 
1960 /*
1961  * Insert into list of vnodes for the new mount point, if available.
1962  */
1963 int
1964 insmntque1(struct vnode *vp, struct mount *mp,
1965 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1966 {
1967 
1968 	KASSERT(vp->v_mount == NULL,
1969 		("insmntque: vnode already on per mount vnode list"));
1970 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1971 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1972 
1973 	/*
1974 	 * We acquire the vnode interlock early to ensure that the
1975 	 * vnode cannot be recycled by another process releasing a
1976 	 * holdcnt on it before we get it on both the vnode list
1977 	 * and the active vnode list. The mount mutex protects only
1978 	 * manipulation of the vnode list and the vnode freelist
1979 	 * mutex protects only manipulation of the active vnode list.
1980 	 * Hence the need to hold the vnode interlock throughout.
1981 	 */
1982 	MNT_ILOCK(mp);
1983 	VI_LOCK(vp);
1984 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1985 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1986 	    mp->mnt_nvnodelistsize == 0)) &&
1987 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1988 		VI_UNLOCK(vp);
1989 		MNT_IUNLOCK(mp);
1990 		if (dtr != NULL)
1991 			dtr(vp, dtr_arg);
1992 		return (EBUSY);
1993 	}
1994 	vp->v_mount = mp;
1995 	MNT_REF(mp);
1996 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1997 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1998 		("neg mount point vnode list size"));
1999 	mp->mnt_nvnodelistsize++;
2000 	VI_UNLOCK(vp);
2001 	MNT_IUNLOCK(mp);
2002 	return (0);
2003 }
2004 
2005 int
2006 insmntque(struct vnode *vp, struct mount *mp)
2007 {
2008 
2009 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
2010 }
2011 
2012 /*
2013  * Flush out and invalidate all buffers associated with a bufobj
2014  * Called with the underlying object locked.
2015  */
2016 int
2017 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2018 {
2019 	int error;
2020 
2021 	BO_LOCK(bo);
2022 	if (flags & V_SAVE) {
2023 		error = bufobj_wwait(bo, slpflag, slptimeo);
2024 		if (error) {
2025 			BO_UNLOCK(bo);
2026 			return (error);
2027 		}
2028 		if (bo->bo_dirty.bv_cnt > 0) {
2029 			BO_UNLOCK(bo);
2030 			do {
2031 				error = BO_SYNC(bo, MNT_WAIT);
2032 			} while (error == ERELOOKUP);
2033 			if (error != 0)
2034 				return (error);
2035 			/*
2036 			 * XXX We could save a lock/unlock if this was only
2037 			 * enabled under INVARIANTS
2038 			 */
2039 			BO_LOCK(bo);
2040 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
2041 				panic("vinvalbuf: dirty bufs");
2042 		}
2043 	}
2044 	/*
2045 	 * If you alter this loop please notice that interlock is dropped and
2046 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2047 	 * no race conditions occur from this.
2048 	 */
2049 	do {
2050 		error = flushbuflist(&bo->bo_clean,
2051 		    flags, bo, slpflag, slptimeo);
2052 		if (error == 0 && !(flags & V_CLEANONLY))
2053 			error = flushbuflist(&bo->bo_dirty,
2054 			    flags, bo, slpflag, slptimeo);
2055 		if (error != 0 && error != EAGAIN) {
2056 			BO_UNLOCK(bo);
2057 			return (error);
2058 		}
2059 	} while (error != 0);
2060 
2061 	/*
2062 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2063 	 * have write I/O in-progress but if there is a VM object then the
2064 	 * VM object can also have read-I/O in-progress.
2065 	 */
2066 	do {
2067 		bufobj_wwait(bo, 0, 0);
2068 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2069 			BO_UNLOCK(bo);
2070 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2071 			BO_LOCK(bo);
2072 		}
2073 	} while (bo->bo_numoutput > 0);
2074 	BO_UNLOCK(bo);
2075 
2076 	/*
2077 	 * Destroy the copy in the VM cache, too.
2078 	 */
2079 	if (bo->bo_object != NULL &&
2080 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2081 		VM_OBJECT_WLOCK(bo->bo_object);
2082 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2083 		    OBJPR_CLEANONLY : 0);
2084 		VM_OBJECT_WUNLOCK(bo->bo_object);
2085 	}
2086 
2087 #ifdef INVARIANTS
2088 	BO_LOCK(bo);
2089 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2090 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2091 	    bo->bo_clean.bv_cnt > 0))
2092 		panic("vinvalbuf: flush failed");
2093 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2094 	    bo->bo_dirty.bv_cnt > 0)
2095 		panic("vinvalbuf: flush dirty failed");
2096 	BO_UNLOCK(bo);
2097 #endif
2098 	return (0);
2099 }
2100 
2101 /*
2102  * Flush out and invalidate all buffers associated with a vnode.
2103  * Called with the underlying object locked.
2104  */
2105 int
2106 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2107 {
2108 
2109 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2110 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2111 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2112 		return (0);
2113 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2114 }
2115 
2116 /*
2117  * Flush out buffers on the specified list.
2118  *
2119  */
2120 static int
2121 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2122     int slptimeo)
2123 {
2124 	struct buf *bp, *nbp;
2125 	int retval, error;
2126 	daddr_t lblkno;
2127 	b_xflags_t xflags;
2128 
2129 	ASSERT_BO_WLOCKED(bo);
2130 
2131 	retval = 0;
2132 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2133 		/*
2134 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2135 		 * do not skip any buffers. If we are flushing only V_NORMAL
2136 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2137 		 * flushing only V_ALT buffers then skip buffers not marked
2138 		 * as BX_ALTDATA.
2139 		 */
2140 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2141 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2142 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2143 			continue;
2144 		}
2145 		if (nbp != NULL) {
2146 			lblkno = nbp->b_lblkno;
2147 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2148 		}
2149 		retval = EAGAIN;
2150 		error = BUF_TIMELOCK(bp,
2151 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2152 		    "flushbuf", slpflag, slptimeo);
2153 		if (error) {
2154 			BO_LOCK(bo);
2155 			return (error != ENOLCK ? error : EAGAIN);
2156 		}
2157 		KASSERT(bp->b_bufobj == bo,
2158 		    ("bp %p wrong b_bufobj %p should be %p",
2159 		    bp, bp->b_bufobj, bo));
2160 		/*
2161 		 * XXX Since there are no node locks for NFS, I
2162 		 * believe there is a slight chance that a delayed
2163 		 * write will occur while sleeping just above, so
2164 		 * check for it.
2165 		 */
2166 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2167 		    (flags & V_SAVE)) {
2168 			bremfree(bp);
2169 			bp->b_flags |= B_ASYNC;
2170 			bwrite(bp);
2171 			BO_LOCK(bo);
2172 			return (EAGAIN);	/* XXX: why not loop ? */
2173 		}
2174 		bremfree(bp);
2175 		bp->b_flags |= (B_INVAL | B_RELBUF);
2176 		bp->b_flags &= ~B_ASYNC;
2177 		brelse(bp);
2178 		BO_LOCK(bo);
2179 		if (nbp == NULL)
2180 			break;
2181 		nbp = gbincore(bo, lblkno);
2182 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2183 		    != xflags)
2184 			break;			/* nbp invalid */
2185 	}
2186 	return (retval);
2187 }
2188 
2189 int
2190 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2191 {
2192 	struct buf *bp;
2193 	int error;
2194 	daddr_t lblkno;
2195 
2196 	ASSERT_BO_LOCKED(bo);
2197 
2198 	for (lblkno = startn;;) {
2199 again:
2200 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2201 		if (bp == NULL || bp->b_lblkno >= endn ||
2202 		    bp->b_lblkno < startn)
2203 			break;
2204 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2205 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2206 		if (error != 0) {
2207 			BO_RLOCK(bo);
2208 			if (error == ENOLCK)
2209 				goto again;
2210 			return (error);
2211 		}
2212 		KASSERT(bp->b_bufobj == bo,
2213 		    ("bp %p wrong b_bufobj %p should be %p",
2214 		    bp, bp->b_bufobj, bo));
2215 		lblkno = bp->b_lblkno + 1;
2216 		if ((bp->b_flags & B_MANAGED) == 0)
2217 			bremfree(bp);
2218 		bp->b_flags |= B_RELBUF;
2219 		/*
2220 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2221 		 * pages backing each buffer in the range are unlikely to be
2222 		 * reused.  Dirty buffers will have the hint applied once
2223 		 * they've been written.
2224 		 */
2225 		if ((bp->b_flags & B_VMIO) != 0)
2226 			bp->b_flags |= B_NOREUSE;
2227 		brelse(bp);
2228 		BO_RLOCK(bo);
2229 	}
2230 	return (0);
2231 }
2232 
2233 /*
2234  * Truncate a file's buffer and pages to a specified length.  This
2235  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2236  * sync activity.
2237  */
2238 int
2239 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2240 {
2241 	struct buf *bp, *nbp;
2242 	struct bufobj *bo;
2243 	daddr_t startlbn;
2244 
2245 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2246 	    vp, blksize, (uintmax_t)length);
2247 
2248 	/*
2249 	 * Round up to the *next* lbn.
2250 	 */
2251 	startlbn = howmany(length, blksize);
2252 
2253 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2254 
2255 	bo = &vp->v_bufobj;
2256 restart_unlocked:
2257 	BO_LOCK(bo);
2258 
2259 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2260 		;
2261 
2262 	if (length > 0) {
2263 restartsync:
2264 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2265 			if (bp->b_lblkno > 0)
2266 				continue;
2267 			/*
2268 			 * Since we hold the vnode lock this should only
2269 			 * fail if we're racing with the buf daemon.
2270 			 */
2271 			if (BUF_LOCK(bp,
2272 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2273 			    BO_LOCKPTR(bo)) == ENOLCK)
2274 				goto restart_unlocked;
2275 
2276 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2277 			    ("buf(%p) on dirty queue without DELWRI", bp));
2278 
2279 			bremfree(bp);
2280 			bawrite(bp);
2281 			BO_LOCK(bo);
2282 			goto restartsync;
2283 		}
2284 	}
2285 
2286 	bufobj_wwait(bo, 0, 0);
2287 	BO_UNLOCK(bo);
2288 	vnode_pager_setsize(vp, length);
2289 
2290 	return (0);
2291 }
2292 
2293 /*
2294  * Invalidate the cached pages of a file's buffer within the range of block
2295  * numbers [startlbn, endlbn).
2296  */
2297 void
2298 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2299     int blksize)
2300 {
2301 	struct bufobj *bo;
2302 	off_t start, end;
2303 
2304 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2305 
2306 	start = blksize * startlbn;
2307 	end = blksize * endlbn;
2308 
2309 	bo = &vp->v_bufobj;
2310 	BO_LOCK(bo);
2311 	MPASS(blksize == bo->bo_bsize);
2312 
2313 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2314 		;
2315 
2316 	BO_UNLOCK(bo);
2317 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2318 }
2319 
2320 static int
2321 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2322     daddr_t startlbn, daddr_t endlbn)
2323 {
2324 	struct buf *bp, *nbp;
2325 	bool anyfreed;
2326 
2327 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2328 	ASSERT_BO_LOCKED(bo);
2329 
2330 	do {
2331 		anyfreed = false;
2332 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2333 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2334 				continue;
2335 			if (BUF_LOCK(bp,
2336 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2337 			    BO_LOCKPTR(bo)) == ENOLCK) {
2338 				BO_LOCK(bo);
2339 				return (EAGAIN);
2340 			}
2341 
2342 			bremfree(bp);
2343 			bp->b_flags |= B_INVAL | B_RELBUF;
2344 			bp->b_flags &= ~B_ASYNC;
2345 			brelse(bp);
2346 			anyfreed = true;
2347 
2348 			BO_LOCK(bo);
2349 			if (nbp != NULL &&
2350 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2351 			    nbp->b_vp != vp ||
2352 			    (nbp->b_flags & B_DELWRI) != 0))
2353 				return (EAGAIN);
2354 		}
2355 
2356 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2357 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2358 				continue;
2359 			if (BUF_LOCK(bp,
2360 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2361 			    BO_LOCKPTR(bo)) == ENOLCK) {
2362 				BO_LOCK(bo);
2363 				return (EAGAIN);
2364 			}
2365 			bremfree(bp);
2366 			bp->b_flags |= B_INVAL | B_RELBUF;
2367 			bp->b_flags &= ~B_ASYNC;
2368 			brelse(bp);
2369 			anyfreed = true;
2370 
2371 			BO_LOCK(bo);
2372 			if (nbp != NULL &&
2373 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2374 			    (nbp->b_vp != vp) ||
2375 			    (nbp->b_flags & B_DELWRI) == 0))
2376 				return (EAGAIN);
2377 		}
2378 	} while (anyfreed);
2379 	return (0);
2380 }
2381 
2382 static void
2383 buf_vlist_remove(struct buf *bp)
2384 {
2385 	struct bufv *bv;
2386 	b_xflags_t flags;
2387 
2388 	flags = bp->b_xflags;
2389 
2390 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2391 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2392 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2393 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2394 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2395 
2396 	if ((flags & BX_VNDIRTY) != 0)
2397 		bv = &bp->b_bufobj->bo_dirty;
2398 	else
2399 		bv = &bp->b_bufobj->bo_clean;
2400 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2401 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2402 	bv->bv_cnt--;
2403 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2404 }
2405 
2406 /*
2407  * Add the buffer to the sorted clean or dirty block list.
2408  *
2409  * NOTE: xflags is passed as a constant, optimizing this inline function!
2410  */
2411 static void
2412 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2413 {
2414 	struct bufv *bv;
2415 	struct buf *n;
2416 	int error;
2417 
2418 	ASSERT_BO_WLOCKED(bo);
2419 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2420 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2421 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2422 	    ("dead bo %p", bo));
2423 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2424 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2425 	bp->b_xflags |= xflags;
2426 	if (xflags & BX_VNDIRTY)
2427 		bv = &bo->bo_dirty;
2428 	else
2429 		bv = &bo->bo_clean;
2430 
2431 	/*
2432 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2433 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2434 	 * than _ge.
2435 	 */
2436 	if (bv->bv_cnt == 0 ||
2437 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2438 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2439 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2440 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2441 	else
2442 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2443 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2444 	if (error)
2445 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2446 	bv->bv_cnt++;
2447 }
2448 
2449 /*
2450  * Look up a buffer using the buffer tries.
2451  */
2452 struct buf *
2453 gbincore(struct bufobj *bo, daddr_t lblkno)
2454 {
2455 	struct buf *bp;
2456 
2457 	ASSERT_BO_LOCKED(bo);
2458 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2459 	if (bp != NULL)
2460 		return (bp);
2461 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2462 }
2463 
2464 /*
2465  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2466  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2467  * stability of the result.  Like other lockless lookups, the found buf may
2468  * already be invalid by the time this function returns.
2469  */
2470 struct buf *
2471 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2472 {
2473 	struct buf *bp;
2474 
2475 	ASSERT_BO_UNLOCKED(bo);
2476 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2477 	if (bp != NULL)
2478 		return (bp);
2479 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2480 }
2481 
2482 /*
2483  * Associate a buffer with a vnode.
2484  */
2485 void
2486 bgetvp(struct vnode *vp, struct buf *bp)
2487 {
2488 	struct bufobj *bo;
2489 
2490 	bo = &vp->v_bufobj;
2491 	ASSERT_BO_WLOCKED(bo);
2492 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2493 
2494 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2495 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2496 	    ("bgetvp: bp already attached! %p", bp));
2497 
2498 	vhold(vp);
2499 	bp->b_vp = vp;
2500 	bp->b_bufobj = bo;
2501 	/*
2502 	 * Insert onto list for new vnode.
2503 	 */
2504 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2505 }
2506 
2507 /*
2508  * Disassociate a buffer from a vnode.
2509  */
2510 void
2511 brelvp(struct buf *bp)
2512 {
2513 	struct bufobj *bo;
2514 	struct vnode *vp;
2515 
2516 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2517 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2518 
2519 	/*
2520 	 * Delete from old vnode list, if on one.
2521 	 */
2522 	vp = bp->b_vp;		/* XXX */
2523 	bo = bp->b_bufobj;
2524 	BO_LOCK(bo);
2525 	buf_vlist_remove(bp);
2526 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2527 		bo->bo_flag &= ~BO_ONWORKLST;
2528 		mtx_lock(&sync_mtx);
2529 		LIST_REMOVE(bo, bo_synclist);
2530 		syncer_worklist_len--;
2531 		mtx_unlock(&sync_mtx);
2532 	}
2533 	bp->b_vp = NULL;
2534 	bp->b_bufobj = NULL;
2535 	BO_UNLOCK(bo);
2536 	vdrop(vp);
2537 }
2538 
2539 /*
2540  * Add an item to the syncer work queue.
2541  */
2542 static void
2543 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2544 {
2545 	int slot;
2546 
2547 	ASSERT_BO_WLOCKED(bo);
2548 
2549 	mtx_lock(&sync_mtx);
2550 	if (bo->bo_flag & BO_ONWORKLST)
2551 		LIST_REMOVE(bo, bo_synclist);
2552 	else {
2553 		bo->bo_flag |= BO_ONWORKLST;
2554 		syncer_worklist_len++;
2555 	}
2556 
2557 	if (delay > syncer_maxdelay - 2)
2558 		delay = syncer_maxdelay - 2;
2559 	slot = (syncer_delayno + delay) & syncer_mask;
2560 
2561 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2562 	mtx_unlock(&sync_mtx);
2563 }
2564 
2565 static int
2566 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2567 {
2568 	int error, len;
2569 
2570 	mtx_lock(&sync_mtx);
2571 	len = syncer_worklist_len - sync_vnode_count;
2572 	mtx_unlock(&sync_mtx);
2573 	error = SYSCTL_OUT(req, &len, sizeof(len));
2574 	return (error);
2575 }
2576 
2577 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2578     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2579     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2580 
2581 static struct proc *updateproc;
2582 static void sched_sync(void);
2583 static struct kproc_desc up_kp = {
2584 	"syncer",
2585 	sched_sync,
2586 	&updateproc
2587 };
2588 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2589 
2590 static int
2591 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2592 {
2593 	struct vnode *vp;
2594 	struct mount *mp;
2595 
2596 	*bo = LIST_FIRST(slp);
2597 	if (*bo == NULL)
2598 		return (0);
2599 	vp = bo2vnode(*bo);
2600 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2601 		return (1);
2602 	/*
2603 	 * We use vhold in case the vnode does not
2604 	 * successfully sync.  vhold prevents the vnode from
2605 	 * going away when we unlock the sync_mtx so that
2606 	 * we can acquire the vnode interlock.
2607 	 */
2608 	vholdl(vp);
2609 	mtx_unlock(&sync_mtx);
2610 	VI_UNLOCK(vp);
2611 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2612 		vdrop(vp);
2613 		mtx_lock(&sync_mtx);
2614 		return (*bo == LIST_FIRST(slp));
2615 	}
2616 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2617 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2618 	VOP_UNLOCK(vp);
2619 	vn_finished_write(mp);
2620 	BO_LOCK(*bo);
2621 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2622 		/*
2623 		 * Put us back on the worklist.  The worklist
2624 		 * routine will remove us from our current
2625 		 * position and then add us back in at a later
2626 		 * position.
2627 		 */
2628 		vn_syncer_add_to_worklist(*bo, syncdelay);
2629 	}
2630 	BO_UNLOCK(*bo);
2631 	vdrop(vp);
2632 	mtx_lock(&sync_mtx);
2633 	return (0);
2634 }
2635 
2636 static int first_printf = 1;
2637 
2638 /*
2639  * System filesystem synchronizer daemon.
2640  */
2641 static void
2642 sched_sync(void)
2643 {
2644 	struct synclist *next, *slp;
2645 	struct bufobj *bo;
2646 	long starttime;
2647 	struct thread *td = curthread;
2648 	int last_work_seen;
2649 	int net_worklist_len;
2650 	int syncer_final_iter;
2651 	int error;
2652 
2653 	last_work_seen = 0;
2654 	syncer_final_iter = 0;
2655 	syncer_state = SYNCER_RUNNING;
2656 	starttime = time_uptime;
2657 	td->td_pflags |= TDP_NORUNNINGBUF;
2658 
2659 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2660 	    SHUTDOWN_PRI_LAST);
2661 
2662 	mtx_lock(&sync_mtx);
2663 	for (;;) {
2664 		if (syncer_state == SYNCER_FINAL_DELAY &&
2665 		    syncer_final_iter == 0) {
2666 			mtx_unlock(&sync_mtx);
2667 			kproc_suspend_check(td->td_proc);
2668 			mtx_lock(&sync_mtx);
2669 		}
2670 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2671 		if (syncer_state != SYNCER_RUNNING &&
2672 		    starttime != time_uptime) {
2673 			if (first_printf) {
2674 				printf("\nSyncing disks, vnodes remaining... ");
2675 				first_printf = 0;
2676 			}
2677 			printf("%d ", net_worklist_len);
2678 		}
2679 		starttime = time_uptime;
2680 
2681 		/*
2682 		 * Push files whose dirty time has expired.  Be careful
2683 		 * of interrupt race on slp queue.
2684 		 *
2685 		 * Skip over empty worklist slots when shutting down.
2686 		 */
2687 		do {
2688 			slp = &syncer_workitem_pending[syncer_delayno];
2689 			syncer_delayno += 1;
2690 			if (syncer_delayno == syncer_maxdelay)
2691 				syncer_delayno = 0;
2692 			next = &syncer_workitem_pending[syncer_delayno];
2693 			/*
2694 			 * If the worklist has wrapped since the
2695 			 * it was emptied of all but syncer vnodes,
2696 			 * switch to the FINAL_DELAY state and run
2697 			 * for one more second.
2698 			 */
2699 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2700 			    net_worklist_len == 0 &&
2701 			    last_work_seen == syncer_delayno) {
2702 				syncer_state = SYNCER_FINAL_DELAY;
2703 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2704 			}
2705 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2706 		    syncer_worklist_len > 0);
2707 
2708 		/*
2709 		 * Keep track of the last time there was anything
2710 		 * on the worklist other than syncer vnodes.
2711 		 * Return to the SHUTTING_DOWN state if any
2712 		 * new work appears.
2713 		 */
2714 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2715 			last_work_seen = syncer_delayno;
2716 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2717 			syncer_state = SYNCER_SHUTTING_DOWN;
2718 		while (!LIST_EMPTY(slp)) {
2719 			error = sync_vnode(slp, &bo, td);
2720 			if (error == 1) {
2721 				LIST_REMOVE(bo, bo_synclist);
2722 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2723 				continue;
2724 			}
2725 
2726 			if (first_printf == 0) {
2727 				/*
2728 				 * Drop the sync mutex, because some watchdog
2729 				 * drivers need to sleep while patting
2730 				 */
2731 				mtx_unlock(&sync_mtx);
2732 				wdog_kern_pat(WD_LASTVAL);
2733 				mtx_lock(&sync_mtx);
2734 			}
2735 		}
2736 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2737 			syncer_final_iter--;
2738 		/*
2739 		 * The variable rushjob allows the kernel to speed up the
2740 		 * processing of the filesystem syncer process. A rushjob
2741 		 * value of N tells the filesystem syncer to process the next
2742 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2743 		 * is used by the soft update code to speed up the filesystem
2744 		 * syncer process when the incore state is getting so far
2745 		 * ahead of the disk that the kernel memory pool is being
2746 		 * threatened with exhaustion.
2747 		 */
2748 		if (rushjob > 0) {
2749 			rushjob -= 1;
2750 			continue;
2751 		}
2752 		/*
2753 		 * Just sleep for a short period of time between
2754 		 * iterations when shutting down to allow some I/O
2755 		 * to happen.
2756 		 *
2757 		 * If it has taken us less than a second to process the
2758 		 * current work, then wait. Otherwise start right over
2759 		 * again. We can still lose time if any single round
2760 		 * takes more than two seconds, but it does not really
2761 		 * matter as we are just trying to generally pace the
2762 		 * filesystem activity.
2763 		 */
2764 		if (syncer_state != SYNCER_RUNNING ||
2765 		    time_uptime == starttime) {
2766 			thread_lock(td);
2767 			sched_prio(td, PPAUSE);
2768 			thread_unlock(td);
2769 		}
2770 		if (syncer_state != SYNCER_RUNNING)
2771 			cv_timedwait(&sync_wakeup, &sync_mtx,
2772 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2773 		else if (time_uptime == starttime)
2774 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2775 	}
2776 }
2777 
2778 /*
2779  * Request the syncer daemon to speed up its work.
2780  * We never push it to speed up more than half of its
2781  * normal turn time, otherwise it could take over the cpu.
2782  */
2783 int
2784 speedup_syncer(void)
2785 {
2786 	int ret = 0;
2787 
2788 	mtx_lock(&sync_mtx);
2789 	if (rushjob < syncdelay / 2) {
2790 		rushjob += 1;
2791 		stat_rush_requests += 1;
2792 		ret = 1;
2793 	}
2794 	mtx_unlock(&sync_mtx);
2795 	cv_broadcast(&sync_wakeup);
2796 	return (ret);
2797 }
2798 
2799 /*
2800  * Tell the syncer to speed up its work and run though its work
2801  * list several times, then tell it to shut down.
2802  */
2803 static void
2804 syncer_shutdown(void *arg, int howto)
2805 {
2806 
2807 	if (howto & RB_NOSYNC)
2808 		return;
2809 	mtx_lock(&sync_mtx);
2810 	syncer_state = SYNCER_SHUTTING_DOWN;
2811 	rushjob = 0;
2812 	mtx_unlock(&sync_mtx);
2813 	cv_broadcast(&sync_wakeup);
2814 	kproc_shutdown(arg, howto);
2815 }
2816 
2817 void
2818 syncer_suspend(void)
2819 {
2820 
2821 	syncer_shutdown(updateproc, 0);
2822 }
2823 
2824 void
2825 syncer_resume(void)
2826 {
2827 
2828 	mtx_lock(&sync_mtx);
2829 	first_printf = 1;
2830 	syncer_state = SYNCER_RUNNING;
2831 	mtx_unlock(&sync_mtx);
2832 	cv_broadcast(&sync_wakeup);
2833 	kproc_resume(updateproc);
2834 }
2835 
2836 /*
2837  * Move the buffer between the clean and dirty lists of its vnode.
2838  */
2839 void
2840 reassignbuf(struct buf *bp)
2841 {
2842 	struct vnode *vp;
2843 	struct bufobj *bo;
2844 	int delay;
2845 #ifdef INVARIANTS
2846 	struct bufv *bv;
2847 #endif
2848 
2849 	vp = bp->b_vp;
2850 	bo = bp->b_bufobj;
2851 
2852 	KASSERT((bp->b_flags & B_PAGING) == 0,
2853 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2854 
2855 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2856 	    bp, bp->b_vp, bp->b_flags);
2857 
2858 	BO_LOCK(bo);
2859 	buf_vlist_remove(bp);
2860 
2861 	/*
2862 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2863 	 * of clean buffers.
2864 	 */
2865 	if (bp->b_flags & B_DELWRI) {
2866 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2867 			switch (vp->v_type) {
2868 			case VDIR:
2869 				delay = dirdelay;
2870 				break;
2871 			case VCHR:
2872 				delay = metadelay;
2873 				break;
2874 			default:
2875 				delay = filedelay;
2876 			}
2877 			vn_syncer_add_to_worklist(bo, delay);
2878 		}
2879 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2880 	} else {
2881 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2882 
2883 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2884 			mtx_lock(&sync_mtx);
2885 			LIST_REMOVE(bo, bo_synclist);
2886 			syncer_worklist_len--;
2887 			mtx_unlock(&sync_mtx);
2888 			bo->bo_flag &= ~BO_ONWORKLST;
2889 		}
2890 	}
2891 #ifdef INVARIANTS
2892 	bv = &bo->bo_clean;
2893 	bp = TAILQ_FIRST(&bv->bv_hd);
2894 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2895 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2896 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2897 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2898 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2899 	bv = &bo->bo_dirty;
2900 	bp = TAILQ_FIRST(&bv->bv_hd);
2901 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2902 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2903 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2904 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2905 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2906 #endif
2907 	BO_UNLOCK(bo);
2908 }
2909 
2910 static void
2911 v_init_counters(struct vnode *vp)
2912 {
2913 
2914 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2915 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2916 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2917 
2918 	refcount_init(&vp->v_holdcnt, 1);
2919 	refcount_init(&vp->v_usecount, 1);
2920 }
2921 
2922 /*
2923  * Grab a particular vnode from the free list, increment its
2924  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2925  * is being destroyed.  Only callers who specify LK_RETRY will
2926  * see doomed vnodes.  If inactive processing was delayed in
2927  * vput try to do it here.
2928  *
2929  * usecount is manipulated using atomics without holding any locks.
2930  *
2931  * holdcnt can be manipulated using atomics without holding any locks,
2932  * except when transitioning 1<->0, in which case the interlock is held.
2933  *
2934  * Consumers which don't guarantee liveness of the vnode can use SMR to
2935  * try to get a reference. Note this operation can fail since the vnode
2936  * may be awaiting getting freed by the time they get to it.
2937  */
2938 enum vgetstate
2939 vget_prep_smr(struct vnode *vp)
2940 {
2941 	enum vgetstate vs;
2942 
2943 	VFS_SMR_ASSERT_ENTERED();
2944 
2945 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2946 		vs = VGET_USECOUNT;
2947 	} else {
2948 		if (vhold_smr(vp))
2949 			vs = VGET_HOLDCNT;
2950 		else
2951 			vs = VGET_NONE;
2952 	}
2953 	return (vs);
2954 }
2955 
2956 enum vgetstate
2957 vget_prep(struct vnode *vp)
2958 {
2959 	enum vgetstate vs;
2960 
2961 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2962 		vs = VGET_USECOUNT;
2963 	} else {
2964 		vhold(vp);
2965 		vs = VGET_HOLDCNT;
2966 	}
2967 	return (vs);
2968 }
2969 
2970 void
2971 vget_abort(struct vnode *vp, enum vgetstate vs)
2972 {
2973 
2974 	switch (vs) {
2975 	case VGET_USECOUNT:
2976 		vrele(vp);
2977 		break;
2978 	case VGET_HOLDCNT:
2979 		vdrop(vp);
2980 		break;
2981 	default:
2982 		__assert_unreachable();
2983 	}
2984 }
2985 
2986 int
2987 vget(struct vnode *vp, int flags)
2988 {
2989 	enum vgetstate vs;
2990 
2991 	vs = vget_prep(vp);
2992 	return (vget_finish(vp, flags, vs));
2993 }
2994 
2995 int
2996 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2997 {
2998 	int error;
2999 
3000 	if ((flags & LK_INTERLOCK) != 0)
3001 		ASSERT_VI_LOCKED(vp, __func__);
3002 	else
3003 		ASSERT_VI_UNLOCKED(vp, __func__);
3004 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3005 	VNPASS(vp->v_holdcnt > 0, vp);
3006 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3007 
3008 	error = vn_lock(vp, flags);
3009 	if (__predict_false(error != 0)) {
3010 		vget_abort(vp, vs);
3011 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3012 		    vp);
3013 		return (error);
3014 	}
3015 
3016 	vget_finish_ref(vp, vs);
3017 	return (0);
3018 }
3019 
3020 void
3021 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3022 {
3023 	int old;
3024 
3025 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3026 	VNPASS(vp->v_holdcnt > 0, vp);
3027 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3028 
3029 	if (vs == VGET_USECOUNT)
3030 		return;
3031 
3032 	/*
3033 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3034 	 * the vnode around. Otherwise someone else lended their hold count and
3035 	 * we have to drop ours.
3036 	 */
3037 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3038 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3039 	if (old != 0) {
3040 #ifdef INVARIANTS
3041 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3042 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3043 #else
3044 		refcount_release(&vp->v_holdcnt);
3045 #endif
3046 	}
3047 }
3048 
3049 void
3050 vref(struct vnode *vp)
3051 {
3052 	enum vgetstate vs;
3053 
3054 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3055 	vs = vget_prep(vp);
3056 	vget_finish_ref(vp, vs);
3057 }
3058 
3059 void
3060 vrefact(struct vnode *vp)
3061 {
3062 
3063 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3064 #ifdef INVARIANTS
3065 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3066 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3067 #else
3068 	refcount_acquire(&vp->v_usecount);
3069 #endif
3070 }
3071 
3072 void
3073 vlazy(struct vnode *vp)
3074 {
3075 	struct mount *mp;
3076 
3077 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3078 
3079 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3080 		return;
3081 	/*
3082 	 * We may get here for inactive routines after the vnode got doomed.
3083 	 */
3084 	if (VN_IS_DOOMED(vp))
3085 		return;
3086 	mp = vp->v_mount;
3087 	mtx_lock(&mp->mnt_listmtx);
3088 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3089 		vp->v_mflag |= VMP_LAZYLIST;
3090 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3091 		mp->mnt_lazyvnodelistsize++;
3092 	}
3093 	mtx_unlock(&mp->mnt_listmtx);
3094 }
3095 
3096 static void
3097 vunlazy(struct vnode *vp)
3098 {
3099 	struct mount *mp;
3100 
3101 	ASSERT_VI_LOCKED(vp, __func__);
3102 	VNPASS(!VN_IS_DOOMED(vp), vp);
3103 
3104 	mp = vp->v_mount;
3105 	mtx_lock(&mp->mnt_listmtx);
3106 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3107 	/*
3108 	 * Don't remove the vnode from the lazy list if another thread
3109 	 * has increased the hold count. It may have re-enqueued the
3110 	 * vnode to the lazy list and is now responsible for its
3111 	 * removal.
3112 	 */
3113 	if (vp->v_holdcnt == 0) {
3114 		vp->v_mflag &= ~VMP_LAZYLIST;
3115 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3116 		mp->mnt_lazyvnodelistsize--;
3117 	}
3118 	mtx_unlock(&mp->mnt_listmtx);
3119 }
3120 
3121 /*
3122  * This routine is only meant to be called from vgonel prior to dooming
3123  * the vnode.
3124  */
3125 static void
3126 vunlazy_gone(struct vnode *vp)
3127 {
3128 	struct mount *mp;
3129 
3130 	ASSERT_VOP_ELOCKED(vp, __func__);
3131 	ASSERT_VI_LOCKED(vp, __func__);
3132 	VNPASS(!VN_IS_DOOMED(vp), vp);
3133 
3134 	if (vp->v_mflag & VMP_LAZYLIST) {
3135 		mp = vp->v_mount;
3136 		mtx_lock(&mp->mnt_listmtx);
3137 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3138 		vp->v_mflag &= ~VMP_LAZYLIST;
3139 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3140 		mp->mnt_lazyvnodelistsize--;
3141 		mtx_unlock(&mp->mnt_listmtx);
3142 	}
3143 }
3144 
3145 static void
3146 vdefer_inactive(struct vnode *vp)
3147 {
3148 
3149 	ASSERT_VI_LOCKED(vp, __func__);
3150 	VNASSERT(vp->v_holdcnt > 0, vp,
3151 	    ("%s: vnode without hold count", __func__));
3152 	if (VN_IS_DOOMED(vp)) {
3153 		vdropl(vp);
3154 		return;
3155 	}
3156 	if (vp->v_iflag & VI_DEFINACT) {
3157 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3158 		vdropl(vp);
3159 		return;
3160 	}
3161 	if (vp->v_usecount > 0) {
3162 		vp->v_iflag &= ~VI_OWEINACT;
3163 		vdropl(vp);
3164 		return;
3165 	}
3166 	vlazy(vp);
3167 	vp->v_iflag |= VI_DEFINACT;
3168 	VI_UNLOCK(vp);
3169 	counter_u64_add(deferred_inact, 1);
3170 }
3171 
3172 static void
3173 vdefer_inactive_unlocked(struct vnode *vp)
3174 {
3175 
3176 	VI_LOCK(vp);
3177 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3178 		vdropl(vp);
3179 		return;
3180 	}
3181 	vdefer_inactive(vp);
3182 }
3183 
3184 enum vput_op { VRELE, VPUT, VUNREF };
3185 
3186 /*
3187  * Handle ->v_usecount transitioning to 0.
3188  *
3189  * By releasing the last usecount we take ownership of the hold count which
3190  * provides liveness of the vnode, meaning we have to vdrop.
3191  *
3192  * For all vnodes we may need to perform inactive processing. It requires an
3193  * exclusive lock on the vnode, while it is legal to call here with only a
3194  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3195  * inactive processing gets deferred to the syncer.
3196  *
3197  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3198  * on the lock being held all the way until VOP_INACTIVE. This in particular
3199  * happens with UFS which adds half-constructed vnodes to the hash, where they
3200  * can be found by other code.
3201  */
3202 static void
3203 vput_final(struct vnode *vp, enum vput_op func)
3204 {
3205 	int error;
3206 	bool want_unlock;
3207 
3208 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3209 	VNPASS(vp->v_holdcnt > 0, vp);
3210 
3211 	VI_LOCK(vp);
3212 
3213 	/*
3214 	 * By the time we got here someone else might have transitioned
3215 	 * the count back to > 0.
3216 	 */
3217 	if (vp->v_usecount > 0)
3218 		goto out;
3219 
3220 	/*
3221 	 * If the vnode is doomed vgone already performed inactive processing
3222 	 * (if needed).
3223 	 */
3224 	if (VN_IS_DOOMED(vp))
3225 		goto out;
3226 
3227 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3228 		goto out;
3229 
3230 	if (vp->v_iflag & VI_DOINGINACT)
3231 		goto out;
3232 
3233 	/*
3234 	 * Locking operations here will drop the interlock and possibly the
3235 	 * vnode lock, opening a window where the vnode can get doomed all the
3236 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3237 	 * perform inactive.
3238 	 */
3239 	vp->v_iflag |= VI_OWEINACT;
3240 	want_unlock = false;
3241 	error = 0;
3242 	switch (func) {
3243 	case VRELE:
3244 		switch (VOP_ISLOCKED(vp)) {
3245 		case LK_EXCLUSIVE:
3246 			break;
3247 		case LK_EXCLOTHER:
3248 		case 0:
3249 			want_unlock = true;
3250 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3251 			VI_LOCK(vp);
3252 			break;
3253 		default:
3254 			/*
3255 			 * The lock has at least one sharer, but we have no way
3256 			 * to conclude whether this is us. Play it safe and
3257 			 * defer processing.
3258 			 */
3259 			error = EAGAIN;
3260 			break;
3261 		}
3262 		break;
3263 	case VPUT:
3264 		want_unlock = true;
3265 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3266 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3267 			    LK_NOWAIT);
3268 			VI_LOCK(vp);
3269 		}
3270 		break;
3271 	case VUNREF:
3272 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3273 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3274 			VI_LOCK(vp);
3275 		}
3276 		break;
3277 	}
3278 	if (error == 0) {
3279 		if (func == VUNREF) {
3280 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3281 			    ("recursive vunref"));
3282 			vp->v_vflag |= VV_UNREF;
3283 		}
3284 		for (;;) {
3285 			error = vinactive(vp);
3286 			if (want_unlock)
3287 				VOP_UNLOCK(vp);
3288 			if (error != ERELOOKUP || !want_unlock)
3289 				break;
3290 			VOP_LOCK(vp, LK_EXCLUSIVE);
3291 		}
3292 		if (func == VUNREF)
3293 			vp->v_vflag &= ~VV_UNREF;
3294 		vdropl(vp);
3295 	} else {
3296 		vdefer_inactive(vp);
3297 	}
3298 	return;
3299 out:
3300 	if (func == VPUT)
3301 		VOP_UNLOCK(vp);
3302 	vdropl(vp);
3303 }
3304 
3305 /*
3306  * Decrement ->v_usecount for a vnode.
3307  *
3308  * Releasing the last use count requires additional processing, see vput_final
3309  * above for details.
3310  *
3311  * Comment above each variant denotes lock state on entry and exit.
3312  */
3313 
3314 /*
3315  * in: any
3316  * out: same as passed in
3317  */
3318 void
3319 vrele(struct vnode *vp)
3320 {
3321 
3322 	ASSERT_VI_UNLOCKED(vp, __func__);
3323 	if (!refcount_release(&vp->v_usecount))
3324 		return;
3325 	vput_final(vp, VRELE);
3326 }
3327 
3328 /*
3329  * in: locked
3330  * out: unlocked
3331  */
3332 void
3333 vput(struct vnode *vp)
3334 {
3335 
3336 	ASSERT_VOP_LOCKED(vp, __func__);
3337 	ASSERT_VI_UNLOCKED(vp, __func__);
3338 	if (!refcount_release(&vp->v_usecount)) {
3339 		VOP_UNLOCK(vp);
3340 		return;
3341 	}
3342 	vput_final(vp, VPUT);
3343 }
3344 
3345 /*
3346  * in: locked
3347  * out: locked
3348  */
3349 void
3350 vunref(struct vnode *vp)
3351 {
3352 
3353 	ASSERT_VOP_LOCKED(vp, __func__);
3354 	ASSERT_VI_UNLOCKED(vp, __func__);
3355 	if (!refcount_release(&vp->v_usecount))
3356 		return;
3357 	vput_final(vp, VUNREF);
3358 }
3359 
3360 void
3361 vhold(struct vnode *vp)
3362 {
3363 	int old;
3364 
3365 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3366 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3367 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3368 	    ("%s: wrong hold count %d", __func__, old));
3369 	if (old == 0)
3370 		vfs_freevnodes_dec();
3371 }
3372 
3373 void
3374 vholdnz(struct vnode *vp)
3375 {
3376 
3377 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3378 #ifdef INVARIANTS
3379 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3380 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3381 	    ("%s: wrong hold count %d", __func__, old));
3382 #else
3383 	atomic_add_int(&vp->v_holdcnt, 1);
3384 #endif
3385 }
3386 
3387 /*
3388  * Grab a hold count unless the vnode is freed.
3389  *
3390  * Only use this routine if vfs smr is the only protection you have against
3391  * freeing the vnode.
3392  *
3393  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3394  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3395  * the thread which managed to set the flag.
3396  *
3397  * It may be tempting to replace the loop with:
3398  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3399  * if (count & VHOLD_NO_SMR) {
3400  *     backpedal and error out;
3401  * }
3402  *
3403  * However, while this is more performant, it hinders debugging by eliminating
3404  * the previously mentioned invariant.
3405  */
3406 bool
3407 vhold_smr(struct vnode *vp)
3408 {
3409 	int count;
3410 
3411 	VFS_SMR_ASSERT_ENTERED();
3412 
3413 	count = atomic_load_int(&vp->v_holdcnt);
3414 	for (;;) {
3415 		if (count & VHOLD_NO_SMR) {
3416 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3417 			    ("non-zero hold count with flags %d\n", count));
3418 			return (false);
3419 		}
3420 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3421 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3422 			if (count == 0)
3423 				vfs_freevnodes_dec();
3424 			return (true);
3425 		}
3426 	}
3427 }
3428 
3429 /*
3430  * Hold a free vnode for recycling.
3431  *
3432  * Note: vnode_init references this comment.
3433  *
3434  * Attempts to recycle only need the global vnode list lock and have no use for
3435  * SMR.
3436  *
3437  * However, vnodes get inserted into the global list before they get fully
3438  * initialized and stay there until UMA decides to free the memory. This in
3439  * particular means the target can be found before it becomes usable and after
3440  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3441  * VHOLD_NO_SMR.
3442  *
3443  * Note: the vnode may gain more references after we transition the count 0->1.
3444  */
3445 static bool
3446 vhold_recycle_free(struct vnode *vp)
3447 {
3448 	int count;
3449 
3450 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3451 
3452 	count = atomic_load_int(&vp->v_holdcnt);
3453 	for (;;) {
3454 		if (count & VHOLD_NO_SMR) {
3455 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3456 			    ("non-zero hold count with flags %d\n", count));
3457 			return (false);
3458 		}
3459 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3460 		if (count > 0) {
3461 			return (false);
3462 		}
3463 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3464 			vfs_freevnodes_dec();
3465 			return (true);
3466 		}
3467 	}
3468 }
3469 
3470 static void __noinline
3471 vdbatch_process(struct vdbatch *vd)
3472 {
3473 	struct vnode *vp;
3474 	int i;
3475 
3476 	mtx_assert(&vd->lock, MA_OWNED);
3477 	MPASS(curthread->td_pinned > 0);
3478 	MPASS(vd->index == VDBATCH_SIZE);
3479 
3480 	mtx_lock(&vnode_list_mtx);
3481 	critical_enter();
3482 	freevnodes += vd->freevnodes;
3483 	for (i = 0; i < VDBATCH_SIZE; i++) {
3484 		vp = vd->tab[i];
3485 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3486 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3487 		MPASS(vp->v_dbatchcpu != NOCPU);
3488 		vp->v_dbatchcpu = NOCPU;
3489 	}
3490 	mtx_unlock(&vnode_list_mtx);
3491 	vd->freevnodes = 0;
3492 	bzero(vd->tab, sizeof(vd->tab));
3493 	vd->index = 0;
3494 	critical_exit();
3495 }
3496 
3497 static void
3498 vdbatch_enqueue(struct vnode *vp)
3499 {
3500 	struct vdbatch *vd;
3501 
3502 	ASSERT_VI_LOCKED(vp, __func__);
3503 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3504 	    ("%s: deferring requeue of a doomed vnode", __func__));
3505 
3506 	if (vp->v_dbatchcpu != NOCPU) {
3507 		VI_UNLOCK(vp);
3508 		return;
3509 	}
3510 
3511 	sched_pin();
3512 	vd = DPCPU_PTR(vd);
3513 	mtx_lock(&vd->lock);
3514 	MPASS(vd->index < VDBATCH_SIZE);
3515 	MPASS(vd->tab[vd->index] == NULL);
3516 	/*
3517 	 * A hack: we depend on being pinned so that we know what to put in
3518 	 * ->v_dbatchcpu.
3519 	 */
3520 	vp->v_dbatchcpu = curcpu;
3521 	vd->tab[vd->index] = vp;
3522 	vd->index++;
3523 	VI_UNLOCK(vp);
3524 	if (vd->index == VDBATCH_SIZE)
3525 		vdbatch_process(vd);
3526 	mtx_unlock(&vd->lock);
3527 	sched_unpin();
3528 }
3529 
3530 /*
3531  * This routine must only be called for vnodes which are about to be
3532  * deallocated. Supporting dequeue for arbitrary vndoes would require
3533  * validating that the locked batch matches.
3534  */
3535 static void
3536 vdbatch_dequeue(struct vnode *vp)
3537 {
3538 	struct vdbatch *vd;
3539 	int i;
3540 	short cpu;
3541 
3542 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3543 	    ("%s: called for a used vnode\n", __func__));
3544 
3545 	cpu = vp->v_dbatchcpu;
3546 	if (cpu == NOCPU)
3547 		return;
3548 
3549 	vd = DPCPU_ID_PTR(cpu, vd);
3550 	mtx_lock(&vd->lock);
3551 	for (i = 0; i < vd->index; i++) {
3552 		if (vd->tab[i] != vp)
3553 			continue;
3554 		vp->v_dbatchcpu = NOCPU;
3555 		vd->index--;
3556 		vd->tab[i] = vd->tab[vd->index];
3557 		vd->tab[vd->index] = NULL;
3558 		break;
3559 	}
3560 	mtx_unlock(&vd->lock);
3561 	/*
3562 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3563 	 */
3564 	MPASS(vp->v_dbatchcpu == NOCPU);
3565 }
3566 
3567 /*
3568  * Drop the hold count of the vnode.  If this is the last reference to
3569  * the vnode we place it on the free list unless it has been vgone'd
3570  * (marked VIRF_DOOMED) in which case we will free it.
3571  *
3572  * Because the vnode vm object keeps a hold reference on the vnode if
3573  * there is at least one resident non-cached page, the vnode cannot
3574  * leave the active list without the page cleanup done.
3575  */
3576 static void __noinline
3577 vdropl_final(struct vnode *vp)
3578 {
3579 
3580 	ASSERT_VI_LOCKED(vp, __func__);
3581 	VNPASS(VN_IS_DOOMED(vp), vp);
3582 	/*
3583 	 * Set the VHOLD_NO_SMR flag.
3584 	 *
3585 	 * We may be racing against vhold_smr. If they win we can just pretend
3586 	 * we never got this far, they will vdrop later.
3587 	 */
3588 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3589 		vfs_freevnodes_inc();
3590 		VI_UNLOCK(vp);
3591 		/*
3592 		 * We lost the aforementioned race. Any subsequent access is
3593 		 * invalid as they might have managed to vdropl on their own.
3594 		 */
3595 		return;
3596 	}
3597 	/*
3598 	 * Don't bump freevnodes as this one is going away.
3599 	 */
3600 	freevnode(vp);
3601 }
3602 
3603 void
3604 vdrop(struct vnode *vp)
3605 {
3606 
3607 	ASSERT_VI_UNLOCKED(vp, __func__);
3608 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3609 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3610 		return;
3611 	VI_LOCK(vp);
3612 	vdropl(vp);
3613 }
3614 
3615 void
3616 vdropl(struct vnode *vp)
3617 {
3618 
3619 	ASSERT_VI_LOCKED(vp, __func__);
3620 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3621 	if (!refcount_release(&vp->v_holdcnt)) {
3622 		VI_UNLOCK(vp);
3623 		return;
3624 	}
3625 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3626 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3627 	if (VN_IS_DOOMED(vp)) {
3628 		vdropl_final(vp);
3629 		return;
3630 	}
3631 
3632 	vfs_freevnodes_inc();
3633 	if (vp->v_mflag & VMP_LAZYLIST) {
3634 		vunlazy(vp);
3635 	}
3636 	/*
3637 	 * Also unlocks the interlock. We can't assert on it as we
3638 	 * released our hold and by now the vnode might have been
3639 	 * freed.
3640 	 */
3641 	vdbatch_enqueue(vp);
3642 }
3643 
3644 /*
3645  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3646  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3647  */
3648 static int
3649 vinactivef(struct vnode *vp)
3650 {
3651 	struct vm_object *obj;
3652 	int error;
3653 
3654 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3655 	ASSERT_VI_LOCKED(vp, "vinactive");
3656 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3657 	    ("vinactive: recursed on VI_DOINGINACT"));
3658 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3659 	vp->v_iflag |= VI_DOINGINACT;
3660 	vp->v_iflag &= ~VI_OWEINACT;
3661 	VI_UNLOCK(vp);
3662 	/*
3663 	 * Before moving off the active list, we must be sure that any
3664 	 * modified pages are converted into the vnode's dirty
3665 	 * buffers, since these will no longer be checked once the
3666 	 * vnode is on the inactive list.
3667 	 *
3668 	 * The write-out of the dirty pages is asynchronous.  At the
3669 	 * point that VOP_INACTIVE() is called, there could still be
3670 	 * pending I/O and dirty pages in the object.
3671 	 */
3672 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3673 	    vm_object_mightbedirty(obj)) {
3674 		VM_OBJECT_WLOCK(obj);
3675 		vm_object_page_clean(obj, 0, 0, 0);
3676 		VM_OBJECT_WUNLOCK(obj);
3677 	}
3678 	error = VOP_INACTIVE(vp);
3679 	VI_LOCK(vp);
3680 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3681 	    ("vinactive: lost VI_DOINGINACT"));
3682 	vp->v_iflag &= ~VI_DOINGINACT;
3683 	return (error);
3684 }
3685 
3686 int
3687 vinactive(struct vnode *vp)
3688 {
3689 
3690 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3691 	ASSERT_VI_LOCKED(vp, "vinactive");
3692 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3693 
3694 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3695 		return (0);
3696 	if (vp->v_iflag & VI_DOINGINACT)
3697 		return (0);
3698 	if (vp->v_usecount > 0) {
3699 		vp->v_iflag &= ~VI_OWEINACT;
3700 		return (0);
3701 	}
3702 	return (vinactivef(vp));
3703 }
3704 
3705 /*
3706  * Remove any vnodes in the vnode table belonging to mount point mp.
3707  *
3708  * If FORCECLOSE is not specified, there should not be any active ones,
3709  * return error if any are found (nb: this is a user error, not a
3710  * system error). If FORCECLOSE is specified, detach any active vnodes
3711  * that are found.
3712  *
3713  * If WRITECLOSE is set, only flush out regular file vnodes open for
3714  * writing.
3715  *
3716  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3717  *
3718  * `rootrefs' specifies the base reference count for the root vnode
3719  * of this filesystem. The root vnode is considered busy if its
3720  * v_usecount exceeds this value. On a successful return, vflush(, td)
3721  * will call vrele() on the root vnode exactly rootrefs times.
3722  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3723  * be zero.
3724  */
3725 #ifdef DIAGNOSTIC
3726 static int busyprt = 0;		/* print out busy vnodes */
3727 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3728 #endif
3729 
3730 int
3731 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3732 {
3733 	struct vnode *vp, *mvp, *rootvp = NULL;
3734 	struct vattr vattr;
3735 	int busy = 0, error;
3736 
3737 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3738 	    rootrefs, flags);
3739 	if (rootrefs > 0) {
3740 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3741 		    ("vflush: bad args"));
3742 		/*
3743 		 * Get the filesystem root vnode. We can vput() it
3744 		 * immediately, since with rootrefs > 0, it won't go away.
3745 		 */
3746 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3747 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3748 			    __func__, error);
3749 			return (error);
3750 		}
3751 		vput(rootvp);
3752 	}
3753 loop:
3754 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3755 		vholdl(vp);
3756 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3757 		if (error) {
3758 			vdrop(vp);
3759 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3760 			goto loop;
3761 		}
3762 		/*
3763 		 * Skip over a vnodes marked VV_SYSTEM.
3764 		 */
3765 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3766 			VOP_UNLOCK(vp);
3767 			vdrop(vp);
3768 			continue;
3769 		}
3770 		/*
3771 		 * If WRITECLOSE is set, flush out unlinked but still open
3772 		 * files (even if open only for reading) and regular file
3773 		 * vnodes open for writing.
3774 		 */
3775 		if (flags & WRITECLOSE) {
3776 			if (vp->v_object != NULL) {
3777 				VM_OBJECT_WLOCK(vp->v_object);
3778 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3779 				VM_OBJECT_WUNLOCK(vp->v_object);
3780 			}
3781 			do {
3782 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3783 			} while (error == ERELOOKUP);
3784 			if (error != 0) {
3785 				VOP_UNLOCK(vp);
3786 				vdrop(vp);
3787 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3788 				return (error);
3789 			}
3790 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3791 			VI_LOCK(vp);
3792 
3793 			if ((vp->v_type == VNON ||
3794 			    (error == 0 && vattr.va_nlink > 0)) &&
3795 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3796 				VOP_UNLOCK(vp);
3797 				vdropl(vp);
3798 				continue;
3799 			}
3800 		} else
3801 			VI_LOCK(vp);
3802 		/*
3803 		 * With v_usecount == 0, all we need to do is clear out the
3804 		 * vnode data structures and we are done.
3805 		 *
3806 		 * If FORCECLOSE is set, forcibly close the vnode.
3807 		 */
3808 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3809 			vgonel(vp);
3810 		} else {
3811 			busy++;
3812 #ifdef DIAGNOSTIC
3813 			if (busyprt)
3814 				vn_printf(vp, "vflush: busy vnode ");
3815 #endif
3816 		}
3817 		VOP_UNLOCK(vp);
3818 		vdropl(vp);
3819 	}
3820 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3821 		/*
3822 		 * If just the root vnode is busy, and if its refcount
3823 		 * is equal to `rootrefs', then go ahead and kill it.
3824 		 */
3825 		VI_LOCK(rootvp);
3826 		KASSERT(busy > 0, ("vflush: not busy"));
3827 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3828 		    ("vflush: usecount %d < rootrefs %d",
3829 		     rootvp->v_usecount, rootrefs));
3830 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3831 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3832 			vgone(rootvp);
3833 			VOP_UNLOCK(rootvp);
3834 			busy = 0;
3835 		} else
3836 			VI_UNLOCK(rootvp);
3837 	}
3838 	if (busy) {
3839 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3840 		    busy);
3841 		return (EBUSY);
3842 	}
3843 	for (; rootrefs > 0; rootrefs--)
3844 		vrele(rootvp);
3845 	return (0);
3846 }
3847 
3848 /*
3849  * Recycle an unused vnode to the front of the free list.
3850  */
3851 int
3852 vrecycle(struct vnode *vp)
3853 {
3854 	int recycled;
3855 
3856 	VI_LOCK(vp);
3857 	recycled = vrecyclel(vp);
3858 	VI_UNLOCK(vp);
3859 	return (recycled);
3860 }
3861 
3862 /*
3863  * vrecycle, with the vp interlock held.
3864  */
3865 int
3866 vrecyclel(struct vnode *vp)
3867 {
3868 	int recycled;
3869 
3870 	ASSERT_VOP_ELOCKED(vp, __func__);
3871 	ASSERT_VI_LOCKED(vp, __func__);
3872 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3873 	recycled = 0;
3874 	if (vp->v_usecount == 0) {
3875 		recycled = 1;
3876 		vgonel(vp);
3877 	}
3878 	return (recycled);
3879 }
3880 
3881 /*
3882  * Eliminate all activity associated with a vnode
3883  * in preparation for reuse.
3884  */
3885 void
3886 vgone(struct vnode *vp)
3887 {
3888 	VI_LOCK(vp);
3889 	vgonel(vp);
3890 	VI_UNLOCK(vp);
3891 }
3892 
3893 static void
3894 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3895     struct vnode *lowervp __unused)
3896 {
3897 }
3898 
3899 /*
3900  * Notify upper mounts about reclaimed or unlinked vnode.
3901  */
3902 void
3903 vfs_notify_upper(struct vnode *vp, int event)
3904 {
3905 	static struct vfsops vgonel_vfsops = {
3906 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3907 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3908 	};
3909 	struct mount *mp, *ump, *mmp;
3910 
3911 	mp = vp->v_mount;
3912 	if (mp == NULL)
3913 		return;
3914 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3915 		return;
3916 
3917 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3918 	mmp->mnt_op = &vgonel_vfsops;
3919 	mmp->mnt_kern_flag |= MNTK_MARKER;
3920 	MNT_ILOCK(mp);
3921 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3922 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3923 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3924 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3925 			continue;
3926 		}
3927 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3928 		MNT_IUNLOCK(mp);
3929 		switch (event) {
3930 		case VFS_NOTIFY_UPPER_RECLAIM:
3931 			VFS_RECLAIM_LOWERVP(ump, vp);
3932 			break;
3933 		case VFS_NOTIFY_UPPER_UNLINK:
3934 			VFS_UNLINK_LOWERVP(ump, vp);
3935 			break;
3936 		default:
3937 			KASSERT(0, ("invalid event %d", event));
3938 			break;
3939 		}
3940 		MNT_ILOCK(mp);
3941 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3942 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3943 	}
3944 	free(mmp, M_TEMP);
3945 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3946 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3947 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3948 		wakeup(&mp->mnt_uppers);
3949 	}
3950 	MNT_IUNLOCK(mp);
3951 }
3952 
3953 /*
3954  * vgone, with the vp interlock held.
3955  */
3956 static void
3957 vgonel(struct vnode *vp)
3958 {
3959 	struct thread *td;
3960 	struct mount *mp;
3961 	vm_object_t object;
3962 	bool active, doinginact, oweinact;
3963 
3964 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3965 	ASSERT_VI_LOCKED(vp, "vgonel");
3966 	VNASSERT(vp->v_holdcnt, vp,
3967 	    ("vgonel: vp %p has no reference.", vp));
3968 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3969 	td = curthread;
3970 
3971 	/*
3972 	 * Don't vgonel if we're already doomed.
3973 	 */
3974 	if (VN_IS_DOOMED(vp))
3975 		return;
3976 	/*
3977 	 * Paired with freevnode.
3978 	 */
3979 	vn_seqc_write_begin_locked(vp);
3980 	vunlazy_gone(vp);
3981 	vn_irflag_set_locked(vp, VIRF_DOOMED);
3982 
3983 	/*
3984 	 * Check to see if the vnode is in use.  If so, we have to
3985 	 * call VOP_CLOSE() and VOP_INACTIVE().
3986 	 *
3987 	 * It could be that VOP_INACTIVE() requested reclamation, in
3988 	 * which case we should avoid recursion, so check
3989 	 * VI_DOINGINACT.  This is not precise but good enough.
3990 	 */
3991 	active = vp->v_usecount > 0;
3992 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3993 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3994 
3995 	/*
3996 	 * If we need to do inactive VI_OWEINACT will be set.
3997 	 */
3998 	if (vp->v_iflag & VI_DEFINACT) {
3999 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4000 		vp->v_iflag &= ~VI_DEFINACT;
4001 		vdropl(vp);
4002 	} else {
4003 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4004 		VI_UNLOCK(vp);
4005 	}
4006 	cache_purge_vgone(vp);
4007 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4008 
4009 	/*
4010 	 * If purging an active vnode, it must be closed and
4011 	 * deactivated before being reclaimed.
4012 	 */
4013 	if (active)
4014 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4015 	if (!doinginact) {
4016 		do {
4017 			if (oweinact || active) {
4018 				VI_LOCK(vp);
4019 				vinactivef(vp);
4020 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4021 				VI_UNLOCK(vp);
4022 			}
4023 		} while (oweinact);
4024 	}
4025 	if (vp->v_type == VSOCK)
4026 		vfs_unp_reclaim(vp);
4027 
4028 	/*
4029 	 * Clean out any buffers associated with the vnode.
4030 	 * If the flush fails, just toss the buffers.
4031 	 */
4032 	mp = NULL;
4033 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4034 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4035 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4036 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4037 			;
4038 	}
4039 
4040 	BO_LOCK(&vp->v_bufobj);
4041 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4042 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4043 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4044 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4045 	    ("vp %p bufobj not invalidated", vp));
4046 
4047 	/*
4048 	 * For VMIO bufobj, BO_DEAD is set later, or in
4049 	 * vm_object_terminate() after the object's page queue is
4050 	 * flushed.
4051 	 */
4052 	object = vp->v_bufobj.bo_object;
4053 	if (object == NULL)
4054 		vp->v_bufobj.bo_flag |= BO_DEAD;
4055 	BO_UNLOCK(&vp->v_bufobj);
4056 
4057 	/*
4058 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4059 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4060 	 * should not touch the object borrowed from the lower vnode
4061 	 * (the handle check).
4062 	 */
4063 	if (object != NULL && object->type == OBJT_VNODE &&
4064 	    object->handle == vp)
4065 		vnode_destroy_vobject(vp);
4066 
4067 	/*
4068 	 * Reclaim the vnode.
4069 	 */
4070 	if (VOP_RECLAIM(vp))
4071 		panic("vgone: cannot reclaim");
4072 	if (mp != NULL)
4073 		vn_finished_secondary_write(mp);
4074 	VNASSERT(vp->v_object == NULL, vp,
4075 	    ("vop_reclaim left v_object vp=%p", vp));
4076 	/*
4077 	 * Clear the advisory locks and wake up waiting threads.
4078 	 */
4079 	(void)VOP_ADVLOCKPURGE(vp);
4080 	vp->v_lockf = NULL;
4081 	/*
4082 	 * Delete from old mount point vnode list.
4083 	 */
4084 	delmntque(vp);
4085 	/*
4086 	 * Done with purge, reset to the standard lock and invalidate
4087 	 * the vnode.
4088 	 */
4089 	VI_LOCK(vp);
4090 	vp->v_vnlock = &vp->v_lock;
4091 	vp->v_op = &dead_vnodeops;
4092 	vp->v_type = VBAD;
4093 }
4094 
4095 /*
4096  * Print out a description of a vnode.
4097  */
4098 static const char * const typename[] =
4099 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4100  "VMARKER"};
4101 
4102 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4103     "new hold count flag not added to vn_printf");
4104 
4105 void
4106 vn_printf(struct vnode *vp, const char *fmt, ...)
4107 {
4108 	va_list ap;
4109 	char buf[256], buf2[16];
4110 	u_long flags;
4111 	u_int holdcnt;
4112 	short irflag;
4113 
4114 	va_start(ap, fmt);
4115 	vprintf(fmt, ap);
4116 	va_end(ap);
4117 	printf("%p: ", (void *)vp);
4118 	printf("type %s\n", typename[vp->v_type]);
4119 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4120 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4121 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4122 	    vp->v_seqc_users);
4123 	switch (vp->v_type) {
4124 	case VDIR:
4125 		printf(" mountedhere %p\n", vp->v_mountedhere);
4126 		break;
4127 	case VCHR:
4128 		printf(" rdev %p\n", vp->v_rdev);
4129 		break;
4130 	case VSOCK:
4131 		printf(" socket %p\n", vp->v_unpcb);
4132 		break;
4133 	case VFIFO:
4134 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4135 		break;
4136 	default:
4137 		printf("\n");
4138 		break;
4139 	}
4140 	buf[0] = '\0';
4141 	buf[1] = '\0';
4142 	if (holdcnt & VHOLD_NO_SMR)
4143 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4144 	printf("    hold count flags (%s)\n", buf + 1);
4145 
4146 	buf[0] = '\0';
4147 	buf[1] = '\0';
4148 	irflag = vn_irflag_read(vp);
4149 	if (irflag & VIRF_DOOMED)
4150 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4151 	if (irflag & VIRF_PGREAD)
4152 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4153 	if (irflag & VIRF_MOUNTPOINT)
4154 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4155 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT);
4156 	if (flags != 0) {
4157 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4158 		strlcat(buf, buf2, sizeof(buf));
4159 	}
4160 	if (vp->v_vflag & VV_ROOT)
4161 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4162 	if (vp->v_vflag & VV_ISTTY)
4163 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4164 	if (vp->v_vflag & VV_NOSYNC)
4165 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4166 	if (vp->v_vflag & VV_ETERNALDEV)
4167 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4168 	if (vp->v_vflag & VV_CACHEDLABEL)
4169 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4170 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4171 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4172 	if (vp->v_vflag & VV_COPYONWRITE)
4173 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4174 	if (vp->v_vflag & VV_SYSTEM)
4175 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4176 	if (vp->v_vflag & VV_PROCDEP)
4177 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4178 	if (vp->v_vflag & VV_NOKNOTE)
4179 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4180 	if (vp->v_vflag & VV_DELETED)
4181 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4182 	if (vp->v_vflag & VV_MD)
4183 		strlcat(buf, "|VV_MD", sizeof(buf));
4184 	if (vp->v_vflag & VV_FORCEINSMQ)
4185 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4186 	if (vp->v_vflag & VV_READLINK)
4187 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4188 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4189 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4190 	    VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ |
4191 	    VV_READLINK);
4192 	if (flags != 0) {
4193 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4194 		strlcat(buf, buf2, sizeof(buf));
4195 	}
4196 	if (vp->v_iflag & VI_TEXT_REF)
4197 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4198 	if (vp->v_iflag & VI_MOUNT)
4199 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4200 	if (vp->v_iflag & VI_DOINGINACT)
4201 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4202 	if (vp->v_iflag & VI_OWEINACT)
4203 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4204 	if (vp->v_iflag & VI_DEFINACT)
4205 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4206 	if (vp->v_iflag & VI_FOPENING)
4207 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4208 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4209 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4210 	if (flags != 0) {
4211 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4212 		strlcat(buf, buf2, sizeof(buf));
4213 	}
4214 	if (vp->v_mflag & VMP_LAZYLIST)
4215 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4216 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4217 	if (flags != 0) {
4218 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4219 		strlcat(buf, buf2, sizeof(buf));
4220 	}
4221 	printf("    flags (%s)", buf + 1);
4222 	if (mtx_owned(VI_MTX(vp)))
4223 		printf(" VI_LOCKed");
4224 	printf("\n");
4225 	if (vp->v_object != NULL)
4226 		printf("    v_object %p ref %d pages %d "
4227 		    "cleanbuf %d dirtybuf %d\n",
4228 		    vp->v_object, vp->v_object->ref_count,
4229 		    vp->v_object->resident_page_count,
4230 		    vp->v_bufobj.bo_clean.bv_cnt,
4231 		    vp->v_bufobj.bo_dirty.bv_cnt);
4232 	printf("    ");
4233 	lockmgr_printinfo(vp->v_vnlock);
4234 	if (vp->v_data != NULL)
4235 		VOP_PRINT(vp);
4236 }
4237 
4238 #ifdef DDB
4239 /*
4240  * List all of the locked vnodes in the system.
4241  * Called when debugging the kernel.
4242  */
4243 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4244 {
4245 	struct mount *mp;
4246 	struct vnode *vp;
4247 
4248 	/*
4249 	 * Note: because this is DDB, we can't obey the locking semantics
4250 	 * for these structures, which means we could catch an inconsistent
4251 	 * state and dereference a nasty pointer.  Not much to be done
4252 	 * about that.
4253 	 */
4254 	db_printf("Locked vnodes\n");
4255 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4256 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4257 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4258 				vn_printf(vp, "vnode ");
4259 		}
4260 	}
4261 }
4262 
4263 /*
4264  * Show details about the given vnode.
4265  */
4266 DB_SHOW_COMMAND(vnode, db_show_vnode)
4267 {
4268 	struct vnode *vp;
4269 
4270 	if (!have_addr)
4271 		return;
4272 	vp = (struct vnode *)addr;
4273 	vn_printf(vp, "vnode ");
4274 }
4275 
4276 /*
4277  * Show details about the given mount point.
4278  */
4279 DB_SHOW_COMMAND(mount, db_show_mount)
4280 {
4281 	struct mount *mp;
4282 	struct vfsopt *opt;
4283 	struct statfs *sp;
4284 	struct vnode *vp;
4285 	char buf[512];
4286 	uint64_t mflags;
4287 	u_int flags;
4288 
4289 	if (!have_addr) {
4290 		/* No address given, print short info about all mount points. */
4291 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4292 			db_printf("%p %s on %s (%s)\n", mp,
4293 			    mp->mnt_stat.f_mntfromname,
4294 			    mp->mnt_stat.f_mntonname,
4295 			    mp->mnt_stat.f_fstypename);
4296 			if (db_pager_quit)
4297 				break;
4298 		}
4299 		db_printf("\nMore info: show mount <addr>\n");
4300 		return;
4301 	}
4302 
4303 	mp = (struct mount *)addr;
4304 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4305 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4306 
4307 	buf[0] = '\0';
4308 	mflags = mp->mnt_flag;
4309 #define	MNT_FLAG(flag)	do {						\
4310 	if (mflags & (flag)) {						\
4311 		if (buf[0] != '\0')					\
4312 			strlcat(buf, ", ", sizeof(buf));		\
4313 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4314 		mflags &= ~(flag);					\
4315 	}								\
4316 } while (0)
4317 	MNT_FLAG(MNT_RDONLY);
4318 	MNT_FLAG(MNT_SYNCHRONOUS);
4319 	MNT_FLAG(MNT_NOEXEC);
4320 	MNT_FLAG(MNT_NOSUID);
4321 	MNT_FLAG(MNT_NFS4ACLS);
4322 	MNT_FLAG(MNT_UNION);
4323 	MNT_FLAG(MNT_ASYNC);
4324 	MNT_FLAG(MNT_SUIDDIR);
4325 	MNT_FLAG(MNT_SOFTDEP);
4326 	MNT_FLAG(MNT_NOSYMFOLLOW);
4327 	MNT_FLAG(MNT_GJOURNAL);
4328 	MNT_FLAG(MNT_MULTILABEL);
4329 	MNT_FLAG(MNT_ACLS);
4330 	MNT_FLAG(MNT_NOATIME);
4331 	MNT_FLAG(MNT_NOCLUSTERR);
4332 	MNT_FLAG(MNT_NOCLUSTERW);
4333 	MNT_FLAG(MNT_SUJ);
4334 	MNT_FLAG(MNT_EXRDONLY);
4335 	MNT_FLAG(MNT_EXPORTED);
4336 	MNT_FLAG(MNT_DEFEXPORTED);
4337 	MNT_FLAG(MNT_EXPORTANON);
4338 	MNT_FLAG(MNT_EXKERB);
4339 	MNT_FLAG(MNT_EXPUBLIC);
4340 	MNT_FLAG(MNT_LOCAL);
4341 	MNT_FLAG(MNT_QUOTA);
4342 	MNT_FLAG(MNT_ROOTFS);
4343 	MNT_FLAG(MNT_USER);
4344 	MNT_FLAG(MNT_IGNORE);
4345 	MNT_FLAG(MNT_UPDATE);
4346 	MNT_FLAG(MNT_DELEXPORT);
4347 	MNT_FLAG(MNT_RELOAD);
4348 	MNT_FLAG(MNT_FORCE);
4349 	MNT_FLAG(MNT_SNAPSHOT);
4350 	MNT_FLAG(MNT_BYFSID);
4351 #undef MNT_FLAG
4352 	if (mflags != 0) {
4353 		if (buf[0] != '\0')
4354 			strlcat(buf, ", ", sizeof(buf));
4355 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4356 		    "0x%016jx", mflags);
4357 	}
4358 	db_printf("    mnt_flag = %s\n", buf);
4359 
4360 	buf[0] = '\0';
4361 	flags = mp->mnt_kern_flag;
4362 #define	MNT_KERN_FLAG(flag)	do {					\
4363 	if (flags & (flag)) {						\
4364 		if (buf[0] != '\0')					\
4365 			strlcat(buf, ", ", sizeof(buf));		\
4366 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4367 		flags &= ~(flag);					\
4368 	}								\
4369 } while (0)
4370 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4371 	MNT_KERN_FLAG(MNTK_ASYNC);
4372 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4373 	MNT_KERN_FLAG(MNTK_DRAINING);
4374 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4375 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4376 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4377 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4378 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4379 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4380 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4381 	MNT_KERN_FLAG(MNTK_MARKER);
4382 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4383 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4384 	MNT_KERN_FLAG(MNTK_NOASYNC);
4385 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4386 	MNT_KERN_FLAG(MNTK_MWAIT);
4387 	MNT_KERN_FLAG(MNTK_SUSPEND);
4388 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4389 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4390 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4391 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4392 #undef MNT_KERN_FLAG
4393 	if (flags != 0) {
4394 		if (buf[0] != '\0')
4395 			strlcat(buf, ", ", sizeof(buf));
4396 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4397 		    "0x%08x", flags);
4398 	}
4399 	db_printf("    mnt_kern_flag = %s\n", buf);
4400 
4401 	db_printf("    mnt_opt = ");
4402 	opt = TAILQ_FIRST(mp->mnt_opt);
4403 	if (opt != NULL) {
4404 		db_printf("%s", opt->name);
4405 		opt = TAILQ_NEXT(opt, link);
4406 		while (opt != NULL) {
4407 			db_printf(", %s", opt->name);
4408 			opt = TAILQ_NEXT(opt, link);
4409 		}
4410 	}
4411 	db_printf("\n");
4412 
4413 	sp = &mp->mnt_stat;
4414 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4415 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4416 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4417 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4418 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4419 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4420 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4421 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4422 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4423 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4424 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4425 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4426 
4427 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4428 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4429 	if (jailed(mp->mnt_cred))
4430 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4431 	db_printf(" }\n");
4432 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4433 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4434 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4435 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4436 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4437 	    mp->mnt_lazyvnodelistsize);
4438 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4439 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4440 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4441 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4442 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4443 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4444 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4445 	db_printf("    mnt_secondary_accwrites = %d\n",
4446 	    mp->mnt_secondary_accwrites);
4447 	db_printf("    mnt_gjprovider = %s\n",
4448 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4449 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4450 
4451 	db_printf("\n\nList of active vnodes\n");
4452 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4453 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4454 			vn_printf(vp, "vnode ");
4455 			if (db_pager_quit)
4456 				break;
4457 		}
4458 	}
4459 	db_printf("\n\nList of inactive vnodes\n");
4460 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4461 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4462 			vn_printf(vp, "vnode ");
4463 			if (db_pager_quit)
4464 				break;
4465 		}
4466 	}
4467 }
4468 #endif	/* DDB */
4469 
4470 /*
4471  * Fill in a struct xvfsconf based on a struct vfsconf.
4472  */
4473 static int
4474 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4475 {
4476 	struct xvfsconf xvfsp;
4477 
4478 	bzero(&xvfsp, sizeof(xvfsp));
4479 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4480 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4481 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4482 	xvfsp.vfc_flags = vfsp->vfc_flags;
4483 	/*
4484 	 * These are unused in userland, we keep them
4485 	 * to not break binary compatibility.
4486 	 */
4487 	xvfsp.vfc_vfsops = NULL;
4488 	xvfsp.vfc_next = NULL;
4489 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4490 }
4491 
4492 #ifdef COMPAT_FREEBSD32
4493 struct xvfsconf32 {
4494 	uint32_t	vfc_vfsops;
4495 	char		vfc_name[MFSNAMELEN];
4496 	int32_t		vfc_typenum;
4497 	int32_t		vfc_refcount;
4498 	int32_t		vfc_flags;
4499 	uint32_t	vfc_next;
4500 };
4501 
4502 static int
4503 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4504 {
4505 	struct xvfsconf32 xvfsp;
4506 
4507 	bzero(&xvfsp, sizeof(xvfsp));
4508 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4509 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4510 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4511 	xvfsp.vfc_flags = vfsp->vfc_flags;
4512 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4513 }
4514 #endif
4515 
4516 /*
4517  * Top level filesystem related information gathering.
4518  */
4519 static int
4520 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4521 {
4522 	struct vfsconf *vfsp;
4523 	int error;
4524 
4525 	error = 0;
4526 	vfsconf_slock();
4527 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4528 #ifdef COMPAT_FREEBSD32
4529 		if (req->flags & SCTL_MASK32)
4530 			error = vfsconf2x32(req, vfsp);
4531 		else
4532 #endif
4533 			error = vfsconf2x(req, vfsp);
4534 		if (error)
4535 			break;
4536 	}
4537 	vfsconf_sunlock();
4538 	return (error);
4539 }
4540 
4541 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4542     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4543     "S,xvfsconf", "List of all configured filesystems");
4544 
4545 #ifndef BURN_BRIDGES
4546 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4547 
4548 static int
4549 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4550 {
4551 	int *name = (int *)arg1 - 1;	/* XXX */
4552 	u_int namelen = arg2 + 1;	/* XXX */
4553 	struct vfsconf *vfsp;
4554 
4555 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4556 	    "please rebuild world\n");
4557 
4558 #if 1 || defined(COMPAT_PRELITE2)
4559 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4560 	if (namelen == 1)
4561 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4562 #endif
4563 
4564 	switch (name[1]) {
4565 	case VFS_MAXTYPENUM:
4566 		if (namelen != 2)
4567 			return (ENOTDIR);
4568 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4569 	case VFS_CONF:
4570 		if (namelen != 3)
4571 			return (ENOTDIR);	/* overloaded */
4572 		vfsconf_slock();
4573 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4574 			if (vfsp->vfc_typenum == name[2])
4575 				break;
4576 		}
4577 		vfsconf_sunlock();
4578 		if (vfsp == NULL)
4579 			return (EOPNOTSUPP);
4580 #ifdef COMPAT_FREEBSD32
4581 		if (req->flags & SCTL_MASK32)
4582 			return (vfsconf2x32(req, vfsp));
4583 		else
4584 #endif
4585 			return (vfsconf2x(req, vfsp));
4586 	}
4587 	return (EOPNOTSUPP);
4588 }
4589 
4590 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4591     CTLFLAG_MPSAFE, vfs_sysctl,
4592     "Generic filesystem");
4593 
4594 #if 1 || defined(COMPAT_PRELITE2)
4595 
4596 static int
4597 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4598 {
4599 	int error;
4600 	struct vfsconf *vfsp;
4601 	struct ovfsconf ovfs;
4602 
4603 	vfsconf_slock();
4604 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4605 		bzero(&ovfs, sizeof(ovfs));
4606 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4607 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4608 		ovfs.vfc_index = vfsp->vfc_typenum;
4609 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4610 		ovfs.vfc_flags = vfsp->vfc_flags;
4611 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4612 		if (error != 0) {
4613 			vfsconf_sunlock();
4614 			return (error);
4615 		}
4616 	}
4617 	vfsconf_sunlock();
4618 	return (0);
4619 }
4620 
4621 #endif /* 1 || COMPAT_PRELITE2 */
4622 #endif /* !BURN_BRIDGES */
4623 
4624 #define KINFO_VNODESLOP		10
4625 #ifdef notyet
4626 /*
4627  * Dump vnode list (via sysctl).
4628  */
4629 /* ARGSUSED */
4630 static int
4631 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4632 {
4633 	struct xvnode *xvn;
4634 	struct mount *mp;
4635 	struct vnode *vp;
4636 	int error, len, n;
4637 
4638 	/*
4639 	 * Stale numvnodes access is not fatal here.
4640 	 */
4641 	req->lock = 0;
4642 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4643 	if (!req->oldptr)
4644 		/* Make an estimate */
4645 		return (SYSCTL_OUT(req, 0, len));
4646 
4647 	error = sysctl_wire_old_buffer(req, 0);
4648 	if (error != 0)
4649 		return (error);
4650 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4651 	n = 0;
4652 	mtx_lock(&mountlist_mtx);
4653 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4654 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4655 			continue;
4656 		MNT_ILOCK(mp);
4657 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4658 			if (n == len)
4659 				break;
4660 			vref(vp);
4661 			xvn[n].xv_size = sizeof *xvn;
4662 			xvn[n].xv_vnode = vp;
4663 			xvn[n].xv_id = 0;	/* XXX compat */
4664 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4665 			XV_COPY(usecount);
4666 			XV_COPY(writecount);
4667 			XV_COPY(holdcnt);
4668 			XV_COPY(mount);
4669 			XV_COPY(numoutput);
4670 			XV_COPY(type);
4671 #undef XV_COPY
4672 			xvn[n].xv_flag = vp->v_vflag;
4673 
4674 			switch (vp->v_type) {
4675 			case VREG:
4676 			case VDIR:
4677 			case VLNK:
4678 				break;
4679 			case VBLK:
4680 			case VCHR:
4681 				if (vp->v_rdev == NULL) {
4682 					vrele(vp);
4683 					continue;
4684 				}
4685 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4686 				break;
4687 			case VSOCK:
4688 				xvn[n].xv_socket = vp->v_socket;
4689 				break;
4690 			case VFIFO:
4691 				xvn[n].xv_fifo = vp->v_fifoinfo;
4692 				break;
4693 			case VNON:
4694 			case VBAD:
4695 			default:
4696 				/* shouldn't happen? */
4697 				vrele(vp);
4698 				continue;
4699 			}
4700 			vrele(vp);
4701 			++n;
4702 		}
4703 		MNT_IUNLOCK(mp);
4704 		mtx_lock(&mountlist_mtx);
4705 		vfs_unbusy(mp);
4706 		if (n == len)
4707 			break;
4708 	}
4709 	mtx_unlock(&mountlist_mtx);
4710 
4711 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4712 	free(xvn, M_TEMP);
4713 	return (error);
4714 }
4715 
4716 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4717     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4718     "");
4719 #endif
4720 
4721 static void
4722 unmount_or_warn(struct mount *mp)
4723 {
4724 	int error;
4725 
4726 	error = dounmount(mp, MNT_FORCE, curthread);
4727 	if (error != 0) {
4728 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4729 		if (error == EBUSY)
4730 			printf("BUSY)\n");
4731 		else
4732 			printf("%d)\n", error);
4733 	}
4734 }
4735 
4736 /*
4737  * Unmount all filesystems. The list is traversed in reverse order
4738  * of mounting to avoid dependencies.
4739  */
4740 void
4741 vfs_unmountall(void)
4742 {
4743 	struct mount *mp, *tmp;
4744 
4745 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4746 
4747 	/*
4748 	 * Since this only runs when rebooting, it is not interlocked.
4749 	 */
4750 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4751 		vfs_ref(mp);
4752 
4753 		/*
4754 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4755 		 * unmount of the latter.
4756 		 */
4757 		if (mp == rootdevmp)
4758 			continue;
4759 
4760 		unmount_or_warn(mp);
4761 	}
4762 
4763 	if (rootdevmp != NULL)
4764 		unmount_or_warn(rootdevmp);
4765 }
4766 
4767 static void
4768 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4769 {
4770 
4771 	ASSERT_VI_LOCKED(vp, __func__);
4772 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4773 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4774 		vdropl(vp);
4775 		return;
4776 	}
4777 	if (vn_lock(vp, lkflags) == 0) {
4778 		VI_LOCK(vp);
4779 		vinactive(vp);
4780 		VOP_UNLOCK(vp);
4781 		vdropl(vp);
4782 		return;
4783 	}
4784 	vdefer_inactive_unlocked(vp);
4785 }
4786 
4787 static int
4788 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4789 {
4790 
4791 	return (vp->v_iflag & VI_DEFINACT);
4792 }
4793 
4794 static void __noinline
4795 vfs_periodic_inactive(struct mount *mp, int flags)
4796 {
4797 	struct vnode *vp, *mvp;
4798 	int lkflags;
4799 
4800 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4801 	if (flags != MNT_WAIT)
4802 		lkflags |= LK_NOWAIT;
4803 
4804 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4805 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4806 			VI_UNLOCK(vp);
4807 			continue;
4808 		}
4809 		vp->v_iflag &= ~VI_DEFINACT;
4810 		vfs_deferred_inactive(vp, lkflags);
4811 	}
4812 }
4813 
4814 static inline bool
4815 vfs_want_msync(struct vnode *vp)
4816 {
4817 	struct vm_object *obj;
4818 
4819 	/*
4820 	 * This test may be performed without any locks held.
4821 	 * We rely on vm_object's type stability.
4822 	 */
4823 	if (vp->v_vflag & VV_NOSYNC)
4824 		return (false);
4825 	obj = vp->v_object;
4826 	return (obj != NULL && vm_object_mightbedirty(obj));
4827 }
4828 
4829 static int
4830 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4831 {
4832 
4833 	if (vp->v_vflag & VV_NOSYNC)
4834 		return (false);
4835 	if (vp->v_iflag & VI_DEFINACT)
4836 		return (true);
4837 	return (vfs_want_msync(vp));
4838 }
4839 
4840 static void __noinline
4841 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4842 {
4843 	struct vnode *vp, *mvp;
4844 	struct vm_object *obj;
4845 	int lkflags, objflags;
4846 	bool seen_defer;
4847 
4848 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4849 	if (flags != MNT_WAIT) {
4850 		lkflags |= LK_NOWAIT;
4851 		objflags = OBJPC_NOSYNC;
4852 	} else {
4853 		objflags = OBJPC_SYNC;
4854 	}
4855 
4856 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4857 		seen_defer = false;
4858 		if (vp->v_iflag & VI_DEFINACT) {
4859 			vp->v_iflag &= ~VI_DEFINACT;
4860 			seen_defer = true;
4861 		}
4862 		if (!vfs_want_msync(vp)) {
4863 			if (seen_defer)
4864 				vfs_deferred_inactive(vp, lkflags);
4865 			else
4866 				VI_UNLOCK(vp);
4867 			continue;
4868 		}
4869 		if (vget(vp, lkflags) == 0) {
4870 			obj = vp->v_object;
4871 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4872 				VM_OBJECT_WLOCK(obj);
4873 				vm_object_page_clean(obj, 0, 0, objflags);
4874 				VM_OBJECT_WUNLOCK(obj);
4875 			}
4876 			vput(vp);
4877 			if (seen_defer)
4878 				vdrop(vp);
4879 		} else {
4880 			if (seen_defer)
4881 				vdefer_inactive_unlocked(vp);
4882 		}
4883 	}
4884 }
4885 
4886 void
4887 vfs_periodic(struct mount *mp, int flags)
4888 {
4889 
4890 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4891 
4892 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4893 		vfs_periodic_inactive(mp, flags);
4894 	else
4895 		vfs_periodic_msync_inactive(mp, flags);
4896 }
4897 
4898 static void
4899 destroy_vpollinfo_free(struct vpollinfo *vi)
4900 {
4901 
4902 	knlist_destroy(&vi->vpi_selinfo.si_note);
4903 	mtx_destroy(&vi->vpi_lock);
4904 	free(vi, M_VNODEPOLL);
4905 }
4906 
4907 static void
4908 destroy_vpollinfo(struct vpollinfo *vi)
4909 {
4910 
4911 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4912 	seldrain(&vi->vpi_selinfo);
4913 	destroy_vpollinfo_free(vi);
4914 }
4915 
4916 /*
4917  * Initialize per-vnode helper structure to hold poll-related state.
4918  */
4919 void
4920 v_addpollinfo(struct vnode *vp)
4921 {
4922 	struct vpollinfo *vi;
4923 
4924 	if (vp->v_pollinfo != NULL)
4925 		return;
4926 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4927 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4928 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4929 	    vfs_knlunlock, vfs_knl_assert_lock);
4930 	VI_LOCK(vp);
4931 	if (vp->v_pollinfo != NULL) {
4932 		VI_UNLOCK(vp);
4933 		destroy_vpollinfo_free(vi);
4934 		return;
4935 	}
4936 	vp->v_pollinfo = vi;
4937 	VI_UNLOCK(vp);
4938 }
4939 
4940 /*
4941  * Record a process's interest in events which might happen to
4942  * a vnode.  Because poll uses the historic select-style interface
4943  * internally, this routine serves as both the ``check for any
4944  * pending events'' and the ``record my interest in future events''
4945  * functions.  (These are done together, while the lock is held,
4946  * to avoid race conditions.)
4947  */
4948 int
4949 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4950 {
4951 
4952 	v_addpollinfo(vp);
4953 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4954 	if (vp->v_pollinfo->vpi_revents & events) {
4955 		/*
4956 		 * This leaves events we are not interested
4957 		 * in available for the other process which
4958 		 * which presumably had requested them
4959 		 * (otherwise they would never have been
4960 		 * recorded).
4961 		 */
4962 		events &= vp->v_pollinfo->vpi_revents;
4963 		vp->v_pollinfo->vpi_revents &= ~events;
4964 
4965 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4966 		return (events);
4967 	}
4968 	vp->v_pollinfo->vpi_events |= events;
4969 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4970 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4971 	return (0);
4972 }
4973 
4974 /*
4975  * Routine to create and manage a filesystem syncer vnode.
4976  */
4977 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4978 static int	sync_fsync(struct  vop_fsync_args *);
4979 static int	sync_inactive(struct  vop_inactive_args *);
4980 static int	sync_reclaim(struct  vop_reclaim_args *);
4981 
4982 static struct vop_vector sync_vnodeops = {
4983 	.vop_bypass =	VOP_EOPNOTSUPP,
4984 	.vop_close =	sync_close,		/* close */
4985 	.vop_fsync =	sync_fsync,		/* fsync */
4986 	.vop_inactive =	sync_inactive,	/* inactive */
4987 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4988 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4989 	.vop_lock1 =	vop_stdlock,	/* lock */
4990 	.vop_unlock =	vop_stdunlock,	/* unlock */
4991 	.vop_islocked =	vop_stdislocked,	/* islocked */
4992 };
4993 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4994 
4995 /*
4996  * Create a new filesystem syncer vnode for the specified mount point.
4997  */
4998 void
4999 vfs_allocate_syncvnode(struct mount *mp)
5000 {
5001 	struct vnode *vp;
5002 	struct bufobj *bo;
5003 	static long start, incr, next;
5004 	int error;
5005 
5006 	/* Allocate a new vnode */
5007 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5008 	if (error != 0)
5009 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5010 	vp->v_type = VNON;
5011 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5012 	vp->v_vflag |= VV_FORCEINSMQ;
5013 	error = insmntque(vp, mp);
5014 	if (error != 0)
5015 		panic("vfs_allocate_syncvnode: insmntque() failed");
5016 	vp->v_vflag &= ~VV_FORCEINSMQ;
5017 	VOP_UNLOCK(vp);
5018 	/*
5019 	 * Place the vnode onto the syncer worklist. We attempt to
5020 	 * scatter them about on the list so that they will go off
5021 	 * at evenly distributed times even if all the filesystems
5022 	 * are mounted at once.
5023 	 */
5024 	next += incr;
5025 	if (next == 0 || next > syncer_maxdelay) {
5026 		start /= 2;
5027 		incr /= 2;
5028 		if (start == 0) {
5029 			start = syncer_maxdelay / 2;
5030 			incr = syncer_maxdelay;
5031 		}
5032 		next = start;
5033 	}
5034 	bo = &vp->v_bufobj;
5035 	BO_LOCK(bo);
5036 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5037 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5038 	mtx_lock(&sync_mtx);
5039 	sync_vnode_count++;
5040 	if (mp->mnt_syncer == NULL) {
5041 		mp->mnt_syncer = vp;
5042 		vp = NULL;
5043 	}
5044 	mtx_unlock(&sync_mtx);
5045 	BO_UNLOCK(bo);
5046 	if (vp != NULL) {
5047 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5048 		vgone(vp);
5049 		vput(vp);
5050 	}
5051 }
5052 
5053 void
5054 vfs_deallocate_syncvnode(struct mount *mp)
5055 {
5056 	struct vnode *vp;
5057 
5058 	mtx_lock(&sync_mtx);
5059 	vp = mp->mnt_syncer;
5060 	if (vp != NULL)
5061 		mp->mnt_syncer = NULL;
5062 	mtx_unlock(&sync_mtx);
5063 	if (vp != NULL)
5064 		vrele(vp);
5065 }
5066 
5067 /*
5068  * Do a lazy sync of the filesystem.
5069  */
5070 static int
5071 sync_fsync(struct vop_fsync_args *ap)
5072 {
5073 	struct vnode *syncvp = ap->a_vp;
5074 	struct mount *mp = syncvp->v_mount;
5075 	int error, save;
5076 	struct bufobj *bo;
5077 
5078 	/*
5079 	 * We only need to do something if this is a lazy evaluation.
5080 	 */
5081 	if (ap->a_waitfor != MNT_LAZY)
5082 		return (0);
5083 
5084 	/*
5085 	 * Move ourselves to the back of the sync list.
5086 	 */
5087 	bo = &syncvp->v_bufobj;
5088 	BO_LOCK(bo);
5089 	vn_syncer_add_to_worklist(bo, syncdelay);
5090 	BO_UNLOCK(bo);
5091 
5092 	/*
5093 	 * Walk the list of vnodes pushing all that are dirty and
5094 	 * not already on the sync list.
5095 	 */
5096 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5097 		return (0);
5098 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5099 		vfs_unbusy(mp);
5100 		return (0);
5101 	}
5102 	save = curthread_pflags_set(TDP_SYNCIO);
5103 	/*
5104 	 * The filesystem at hand may be idle with free vnodes stored in the
5105 	 * batch.  Return them instead of letting them stay there indefinitely.
5106 	 */
5107 	vfs_periodic(mp, MNT_NOWAIT);
5108 	error = VFS_SYNC(mp, MNT_LAZY);
5109 	curthread_pflags_restore(save);
5110 	vn_finished_write(mp);
5111 	vfs_unbusy(mp);
5112 	return (error);
5113 }
5114 
5115 /*
5116  * The syncer vnode is no referenced.
5117  */
5118 static int
5119 sync_inactive(struct vop_inactive_args *ap)
5120 {
5121 
5122 	vgone(ap->a_vp);
5123 	return (0);
5124 }
5125 
5126 /*
5127  * The syncer vnode is no longer needed and is being decommissioned.
5128  *
5129  * Modifications to the worklist must be protected by sync_mtx.
5130  */
5131 static int
5132 sync_reclaim(struct vop_reclaim_args *ap)
5133 {
5134 	struct vnode *vp = ap->a_vp;
5135 	struct bufobj *bo;
5136 
5137 	bo = &vp->v_bufobj;
5138 	BO_LOCK(bo);
5139 	mtx_lock(&sync_mtx);
5140 	if (vp->v_mount->mnt_syncer == vp)
5141 		vp->v_mount->mnt_syncer = NULL;
5142 	if (bo->bo_flag & BO_ONWORKLST) {
5143 		LIST_REMOVE(bo, bo_synclist);
5144 		syncer_worklist_len--;
5145 		sync_vnode_count--;
5146 		bo->bo_flag &= ~BO_ONWORKLST;
5147 	}
5148 	mtx_unlock(&sync_mtx);
5149 	BO_UNLOCK(bo);
5150 
5151 	return (0);
5152 }
5153 
5154 int
5155 vn_need_pageq_flush(struct vnode *vp)
5156 {
5157 	struct vm_object *obj;
5158 
5159 	obj = vp->v_object;
5160 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5161 	    vm_object_mightbedirty(obj));
5162 }
5163 
5164 /*
5165  * Check if vnode represents a disk device
5166  */
5167 bool
5168 vn_isdisk_error(struct vnode *vp, int *errp)
5169 {
5170 	int error;
5171 
5172 	if (vp->v_type != VCHR) {
5173 		error = ENOTBLK;
5174 		goto out;
5175 	}
5176 	error = 0;
5177 	dev_lock();
5178 	if (vp->v_rdev == NULL)
5179 		error = ENXIO;
5180 	else if (vp->v_rdev->si_devsw == NULL)
5181 		error = ENXIO;
5182 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5183 		error = ENOTBLK;
5184 	dev_unlock();
5185 out:
5186 	*errp = error;
5187 	return (error == 0);
5188 }
5189 
5190 bool
5191 vn_isdisk(struct vnode *vp)
5192 {
5193 	int error;
5194 
5195 	return (vn_isdisk_error(vp, &error));
5196 }
5197 
5198 /*
5199  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5200  * the comment above cache_fplookup for details.
5201  */
5202 int
5203 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5204 {
5205 	int error;
5206 
5207 	VFS_SMR_ASSERT_ENTERED();
5208 
5209 	/* Check the owner. */
5210 	if (cred->cr_uid == file_uid) {
5211 		if (file_mode & S_IXUSR)
5212 			return (0);
5213 		goto out_error;
5214 	}
5215 
5216 	/* Otherwise, check the groups (first match) */
5217 	if (groupmember(file_gid, cred)) {
5218 		if (file_mode & S_IXGRP)
5219 			return (0);
5220 		goto out_error;
5221 	}
5222 
5223 	/* Otherwise, check everyone else. */
5224 	if (file_mode & S_IXOTH)
5225 		return (0);
5226 out_error:
5227 	/*
5228 	 * Permission check failed, but it is possible denial will get overwritten
5229 	 * (e.g., when root is traversing through a 700 directory owned by someone
5230 	 * else).
5231 	 *
5232 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5233 	 * modules overriding this result. It's quite unclear what semantics
5234 	 * are allowed for them to operate, thus for safety we don't call them
5235 	 * from within the SMR section. This also means if any such modules
5236 	 * are present, we have to let the regular lookup decide.
5237 	 */
5238 	error = priv_check_cred_vfs_lookup_nomac(cred);
5239 	switch (error) {
5240 	case 0:
5241 		return (0);
5242 	case EAGAIN:
5243 		/*
5244 		 * MAC modules present.
5245 		 */
5246 		return (EAGAIN);
5247 	case EPERM:
5248 		return (EACCES);
5249 	default:
5250 		return (error);
5251 	}
5252 }
5253 
5254 /*
5255  * Common filesystem object access control check routine.  Accepts a
5256  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5257  * Returns 0 on success, or an errno on failure.
5258  */
5259 int
5260 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5261     accmode_t accmode, struct ucred *cred)
5262 {
5263 	accmode_t dac_granted;
5264 	accmode_t priv_granted;
5265 
5266 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5267 	    ("invalid bit in accmode"));
5268 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5269 	    ("VAPPEND without VWRITE"));
5270 
5271 	/*
5272 	 * Look for a normal, non-privileged way to access the file/directory
5273 	 * as requested.  If it exists, go with that.
5274 	 */
5275 
5276 	dac_granted = 0;
5277 
5278 	/* Check the owner. */
5279 	if (cred->cr_uid == file_uid) {
5280 		dac_granted |= VADMIN;
5281 		if (file_mode & S_IXUSR)
5282 			dac_granted |= VEXEC;
5283 		if (file_mode & S_IRUSR)
5284 			dac_granted |= VREAD;
5285 		if (file_mode & S_IWUSR)
5286 			dac_granted |= (VWRITE | VAPPEND);
5287 
5288 		if ((accmode & dac_granted) == accmode)
5289 			return (0);
5290 
5291 		goto privcheck;
5292 	}
5293 
5294 	/* Otherwise, check the groups (first match) */
5295 	if (groupmember(file_gid, cred)) {
5296 		if (file_mode & S_IXGRP)
5297 			dac_granted |= VEXEC;
5298 		if (file_mode & S_IRGRP)
5299 			dac_granted |= VREAD;
5300 		if (file_mode & S_IWGRP)
5301 			dac_granted |= (VWRITE | VAPPEND);
5302 
5303 		if ((accmode & dac_granted) == accmode)
5304 			return (0);
5305 
5306 		goto privcheck;
5307 	}
5308 
5309 	/* Otherwise, check everyone else. */
5310 	if (file_mode & S_IXOTH)
5311 		dac_granted |= VEXEC;
5312 	if (file_mode & S_IROTH)
5313 		dac_granted |= VREAD;
5314 	if (file_mode & S_IWOTH)
5315 		dac_granted |= (VWRITE | VAPPEND);
5316 	if ((accmode & dac_granted) == accmode)
5317 		return (0);
5318 
5319 privcheck:
5320 	/*
5321 	 * Build a privilege mask to determine if the set of privileges
5322 	 * satisfies the requirements when combined with the granted mask
5323 	 * from above.  For each privilege, if the privilege is required,
5324 	 * bitwise or the request type onto the priv_granted mask.
5325 	 */
5326 	priv_granted = 0;
5327 
5328 	if (type == VDIR) {
5329 		/*
5330 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5331 		 * requests, instead of PRIV_VFS_EXEC.
5332 		 */
5333 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5334 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5335 			priv_granted |= VEXEC;
5336 	} else {
5337 		/*
5338 		 * Ensure that at least one execute bit is on. Otherwise,
5339 		 * a privileged user will always succeed, and we don't want
5340 		 * this to happen unless the file really is executable.
5341 		 */
5342 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5343 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5344 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5345 			priv_granted |= VEXEC;
5346 	}
5347 
5348 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5349 	    !priv_check_cred(cred, PRIV_VFS_READ))
5350 		priv_granted |= VREAD;
5351 
5352 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5353 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5354 		priv_granted |= (VWRITE | VAPPEND);
5355 
5356 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5357 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5358 		priv_granted |= VADMIN;
5359 
5360 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5361 		return (0);
5362 	}
5363 
5364 	return ((accmode & VADMIN) ? EPERM : EACCES);
5365 }
5366 
5367 /*
5368  * Credential check based on process requesting service, and per-attribute
5369  * permissions.
5370  */
5371 int
5372 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5373     struct thread *td, accmode_t accmode)
5374 {
5375 
5376 	/*
5377 	 * Kernel-invoked always succeeds.
5378 	 */
5379 	if (cred == NOCRED)
5380 		return (0);
5381 
5382 	/*
5383 	 * Do not allow privileged processes in jail to directly manipulate
5384 	 * system attributes.
5385 	 */
5386 	switch (attrnamespace) {
5387 	case EXTATTR_NAMESPACE_SYSTEM:
5388 		/* Potentially should be: return (EPERM); */
5389 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5390 	case EXTATTR_NAMESPACE_USER:
5391 		return (VOP_ACCESS(vp, accmode, cred, td));
5392 	default:
5393 		return (EPERM);
5394 	}
5395 }
5396 
5397 #ifdef DEBUG_VFS_LOCKS
5398 /*
5399  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5400  * no longer ok to have an unlocked VFS.
5401  */
5402 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5403 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5404 
5405 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5406 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5407     "Drop into debugger on lock violation");
5408 
5409 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5410 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5411     0, "Check for interlock across VOPs");
5412 
5413 int vfs_badlock_print = 1;	/* Print lock violations. */
5414 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5415     0, "Print lock violations");
5416 
5417 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5418 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5419     0, "Print vnode details on lock violations");
5420 
5421 #ifdef KDB
5422 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5423 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5424     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5425 #endif
5426 
5427 static void
5428 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5429 {
5430 
5431 #ifdef KDB
5432 	if (vfs_badlock_backtrace)
5433 		kdb_backtrace();
5434 #endif
5435 	if (vfs_badlock_vnode)
5436 		vn_printf(vp, "vnode ");
5437 	if (vfs_badlock_print)
5438 		printf("%s: %p %s\n", str, (void *)vp, msg);
5439 	if (vfs_badlock_ddb)
5440 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5441 }
5442 
5443 void
5444 assert_vi_locked(struct vnode *vp, const char *str)
5445 {
5446 
5447 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5448 		vfs_badlock("interlock is not locked but should be", str, vp);
5449 }
5450 
5451 void
5452 assert_vi_unlocked(struct vnode *vp, const char *str)
5453 {
5454 
5455 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5456 		vfs_badlock("interlock is locked but should not be", str, vp);
5457 }
5458 
5459 void
5460 assert_vop_locked(struct vnode *vp, const char *str)
5461 {
5462 	int locked;
5463 
5464 	if (!IGNORE_LOCK(vp)) {
5465 		locked = VOP_ISLOCKED(vp);
5466 		if (locked == 0 || locked == LK_EXCLOTHER)
5467 			vfs_badlock("is not locked but should be", str, vp);
5468 	}
5469 }
5470 
5471 void
5472 assert_vop_unlocked(struct vnode *vp, const char *str)
5473 {
5474 
5475 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5476 		vfs_badlock("is locked but should not be", str, vp);
5477 }
5478 
5479 void
5480 assert_vop_elocked(struct vnode *vp, const char *str)
5481 {
5482 
5483 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5484 		vfs_badlock("is not exclusive locked but should be", str, vp);
5485 }
5486 #endif /* DEBUG_VFS_LOCKS */
5487 
5488 void
5489 vop_rename_fail(struct vop_rename_args *ap)
5490 {
5491 
5492 	if (ap->a_tvp != NULL)
5493 		vput(ap->a_tvp);
5494 	if (ap->a_tdvp == ap->a_tvp)
5495 		vrele(ap->a_tdvp);
5496 	else
5497 		vput(ap->a_tdvp);
5498 	vrele(ap->a_fdvp);
5499 	vrele(ap->a_fvp);
5500 }
5501 
5502 void
5503 vop_rename_pre(void *ap)
5504 {
5505 	struct vop_rename_args *a = ap;
5506 
5507 #ifdef DEBUG_VFS_LOCKS
5508 	if (a->a_tvp)
5509 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5510 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5511 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5512 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5513 
5514 	/* Check the source (from). */
5515 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5516 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5517 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5518 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5519 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5520 
5521 	/* Check the target. */
5522 	if (a->a_tvp)
5523 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5524 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5525 #endif
5526 	/*
5527 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5528 	 * in vop_rename_post but that's not going to work out since some
5529 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5530 	 *
5531 	 * For now filesystems are expected to do the relevant calls after they
5532 	 * decide what vnodes to operate on.
5533 	 */
5534 	if (a->a_tdvp != a->a_fdvp)
5535 		vhold(a->a_fdvp);
5536 	if (a->a_tvp != a->a_fvp)
5537 		vhold(a->a_fvp);
5538 	vhold(a->a_tdvp);
5539 	if (a->a_tvp)
5540 		vhold(a->a_tvp);
5541 }
5542 
5543 #ifdef DEBUG_VFS_LOCKS
5544 void
5545 vop_fplookup_vexec_debugpre(void *ap __unused)
5546 {
5547 
5548 	VFS_SMR_ASSERT_ENTERED();
5549 }
5550 
5551 void
5552 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5553 {
5554 
5555 	VFS_SMR_ASSERT_ENTERED();
5556 }
5557 
5558 void
5559 vop_fplookup_symlink_debugpre(void *ap __unused)
5560 {
5561 
5562 	VFS_SMR_ASSERT_ENTERED();
5563 }
5564 
5565 void
5566 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5567 {
5568 
5569 	VFS_SMR_ASSERT_ENTERED();
5570 }
5571 void
5572 vop_strategy_debugpre(void *ap)
5573 {
5574 	struct vop_strategy_args *a;
5575 	struct buf *bp;
5576 
5577 	a = ap;
5578 	bp = a->a_bp;
5579 
5580 	/*
5581 	 * Cluster ops lock their component buffers but not the IO container.
5582 	 */
5583 	if ((bp->b_flags & B_CLUSTER) != 0)
5584 		return;
5585 
5586 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5587 		if (vfs_badlock_print)
5588 			printf(
5589 			    "VOP_STRATEGY: bp is not locked but should be\n");
5590 		if (vfs_badlock_ddb)
5591 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5592 	}
5593 }
5594 
5595 void
5596 vop_lock_debugpre(void *ap)
5597 {
5598 	struct vop_lock1_args *a = ap;
5599 
5600 	if ((a->a_flags & LK_INTERLOCK) == 0)
5601 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5602 	else
5603 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5604 }
5605 
5606 void
5607 vop_lock_debugpost(void *ap, int rc)
5608 {
5609 	struct vop_lock1_args *a = ap;
5610 
5611 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5612 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5613 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5614 }
5615 
5616 void
5617 vop_unlock_debugpre(void *ap)
5618 {
5619 	struct vop_unlock_args *a = ap;
5620 
5621 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5622 }
5623 
5624 void
5625 vop_need_inactive_debugpre(void *ap)
5626 {
5627 	struct vop_need_inactive_args *a = ap;
5628 
5629 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5630 }
5631 
5632 void
5633 vop_need_inactive_debugpost(void *ap, int rc)
5634 {
5635 	struct vop_need_inactive_args *a = ap;
5636 
5637 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5638 }
5639 #endif
5640 
5641 void
5642 vop_create_pre(void *ap)
5643 {
5644 	struct vop_create_args *a;
5645 	struct vnode *dvp;
5646 
5647 	a = ap;
5648 	dvp = a->a_dvp;
5649 	vn_seqc_write_begin(dvp);
5650 }
5651 
5652 void
5653 vop_create_post(void *ap, int rc)
5654 {
5655 	struct vop_create_args *a;
5656 	struct vnode *dvp;
5657 
5658 	a = ap;
5659 	dvp = a->a_dvp;
5660 	vn_seqc_write_end(dvp);
5661 	if (!rc)
5662 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5663 }
5664 
5665 void
5666 vop_whiteout_pre(void *ap)
5667 {
5668 	struct vop_whiteout_args *a;
5669 	struct vnode *dvp;
5670 
5671 	a = ap;
5672 	dvp = a->a_dvp;
5673 	vn_seqc_write_begin(dvp);
5674 }
5675 
5676 void
5677 vop_whiteout_post(void *ap, int rc)
5678 {
5679 	struct vop_whiteout_args *a;
5680 	struct vnode *dvp;
5681 
5682 	a = ap;
5683 	dvp = a->a_dvp;
5684 	vn_seqc_write_end(dvp);
5685 }
5686 
5687 void
5688 vop_deleteextattr_pre(void *ap)
5689 {
5690 	struct vop_deleteextattr_args *a;
5691 	struct vnode *vp;
5692 
5693 	a = ap;
5694 	vp = a->a_vp;
5695 	vn_seqc_write_begin(vp);
5696 }
5697 
5698 void
5699 vop_deleteextattr_post(void *ap, int rc)
5700 {
5701 	struct vop_deleteextattr_args *a;
5702 	struct vnode *vp;
5703 
5704 	a = ap;
5705 	vp = a->a_vp;
5706 	vn_seqc_write_end(vp);
5707 	if (!rc)
5708 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5709 }
5710 
5711 void
5712 vop_link_pre(void *ap)
5713 {
5714 	struct vop_link_args *a;
5715 	struct vnode *vp, *tdvp;
5716 
5717 	a = ap;
5718 	vp = a->a_vp;
5719 	tdvp = a->a_tdvp;
5720 	vn_seqc_write_begin(vp);
5721 	vn_seqc_write_begin(tdvp);
5722 }
5723 
5724 void
5725 vop_link_post(void *ap, int rc)
5726 {
5727 	struct vop_link_args *a;
5728 	struct vnode *vp, *tdvp;
5729 
5730 	a = ap;
5731 	vp = a->a_vp;
5732 	tdvp = a->a_tdvp;
5733 	vn_seqc_write_end(vp);
5734 	vn_seqc_write_end(tdvp);
5735 	if (!rc) {
5736 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5737 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5738 	}
5739 }
5740 
5741 void
5742 vop_mkdir_pre(void *ap)
5743 {
5744 	struct vop_mkdir_args *a;
5745 	struct vnode *dvp;
5746 
5747 	a = ap;
5748 	dvp = a->a_dvp;
5749 	vn_seqc_write_begin(dvp);
5750 }
5751 
5752 void
5753 vop_mkdir_post(void *ap, int rc)
5754 {
5755 	struct vop_mkdir_args *a;
5756 	struct vnode *dvp;
5757 
5758 	a = ap;
5759 	dvp = a->a_dvp;
5760 	vn_seqc_write_end(dvp);
5761 	if (!rc)
5762 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5763 }
5764 
5765 #ifdef DEBUG_VFS_LOCKS
5766 void
5767 vop_mkdir_debugpost(void *ap, int rc)
5768 {
5769 	struct vop_mkdir_args *a;
5770 
5771 	a = ap;
5772 	if (!rc)
5773 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5774 }
5775 #endif
5776 
5777 void
5778 vop_mknod_pre(void *ap)
5779 {
5780 	struct vop_mknod_args *a;
5781 	struct vnode *dvp;
5782 
5783 	a = ap;
5784 	dvp = a->a_dvp;
5785 	vn_seqc_write_begin(dvp);
5786 }
5787 
5788 void
5789 vop_mknod_post(void *ap, int rc)
5790 {
5791 	struct vop_mknod_args *a;
5792 	struct vnode *dvp;
5793 
5794 	a = ap;
5795 	dvp = a->a_dvp;
5796 	vn_seqc_write_end(dvp);
5797 	if (!rc)
5798 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5799 }
5800 
5801 void
5802 vop_reclaim_post(void *ap, int rc)
5803 {
5804 	struct vop_reclaim_args *a;
5805 	struct vnode *vp;
5806 
5807 	a = ap;
5808 	vp = a->a_vp;
5809 	ASSERT_VOP_IN_SEQC(vp);
5810 	if (!rc)
5811 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5812 }
5813 
5814 void
5815 vop_remove_pre(void *ap)
5816 {
5817 	struct vop_remove_args *a;
5818 	struct vnode *dvp, *vp;
5819 
5820 	a = ap;
5821 	dvp = a->a_dvp;
5822 	vp = a->a_vp;
5823 	vn_seqc_write_begin(dvp);
5824 	vn_seqc_write_begin(vp);
5825 }
5826 
5827 void
5828 vop_remove_post(void *ap, int rc)
5829 {
5830 	struct vop_remove_args *a;
5831 	struct vnode *dvp, *vp;
5832 
5833 	a = ap;
5834 	dvp = a->a_dvp;
5835 	vp = a->a_vp;
5836 	vn_seqc_write_end(dvp);
5837 	vn_seqc_write_end(vp);
5838 	if (!rc) {
5839 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5840 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5841 	}
5842 }
5843 
5844 void
5845 vop_rename_post(void *ap, int rc)
5846 {
5847 	struct vop_rename_args *a = ap;
5848 	long hint;
5849 
5850 	if (!rc) {
5851 		hint = NOTE_WRITE;
5852 		if (a->a_fdvp == a->a_tdvp) {
5853 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5854 				hint |= NOTE_LINK;
5855 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5856 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5857 		} else {
5858 			hint |= NOTE_EXTEND;
5859 			if (a->a_fvp->v_type == VDIR)
5860 				hint |= NOTE_LINK;
5861 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5862 
5863 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5864 			    a->a_tvp->v_type == VDIR)
5865 				hint &= ~NOTE_LINK;
5866 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5867 		}
5868 
5869 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5870 		if (a->a_tvp)
5871 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5872 	}
5873 	if (a->a_tdvp != a->a_fdvp)
5874 		vdrop(a->a_fdvp);
5875 	if (a->a_tvp != a->a_fvp)
5876 		vdrop(a->a_fvp);
5877 	vdrop(a->a_tdvp);
5878 	if (a->a_tvp)
5879 		vdrop(a->a_tvp);
5880 }
5881 
5882 void
5883 vop_rmdir_pre(void *ap)
5884 {
5885 	struct vop_rmdir_args *a;
5886 	struct vnode *dvp, *vp;
5887 
5888 	a = ap;
5889 	dvp = a->a_dvp;
5890 	vp = a->a_vp;
5891 	vn_seqc_write_begin(dvp);
5892 	vn_seqc_write_begin(vp);
5893 }
5894 
5895 void
5896 vop_rmdir_post(void *ap, int rc)
5897 {
5898 	struct vop_rmdir_args *a;
5899 	struct vnode *dvp, *vp;
5900 
5901 	a = ap;
5902 	dvp = a->a_dvp;
5903 	vp = a->a_vp;
5904 	vn_seqc_write_end(dvp);
5905 	vn_seqc_write_end(vp);
5906 	if (!rc) {
5907 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5908 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5909 	}
5910 }
5911 
5912 void
5913 vop_setattr_pre(void *ap)
5914 {
5915 	struct vop_setattr_args *a;
5916 	struct vnode *vp;
5917 
5918 	a = ap;
5919 	vp = a->a_vp;
5920 	vn_seqc_write_begin(vp);
5921 }
5922 
5923 void
5924 vop_setattr_post(void *ap, int rc)
5925 {
5926 	struct vop_setattr_args *a;
5927 	struct vnode *vp;
5928 
5929 	a = ap;
5930 	vp = a->a_vp;
5931 	vn_seqc_write_end(vp);
5932 	if (!rc)
5933 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5934 }
5935 
5936 void
5937 vop_setacl_pre(void *ap)
5938 {
5939 	struct vop_setacl_args *a;
5940 	struct vnode *vp;
5941 
5942 	a = ap;
5943 	vp = a->a_vp;
5944 	vn_seqc_write_begin(vp);
5945 }
5946 
5947 void
5948 vop_setacl_post(void *ap, int rc __unused)
5949 {
5950 	struct vop_setacl_args *a;
5951 	struct vnode *vp;
5952 
5953 	a = ap;
5954 	vp = a->a_vp;
5955 	vn_seqc_write_end(vp);
5956 }
5957 
5958 void
5959 vop_setextattr_pre(void *ap)
5960 {
5961 	struct vop_setextattr_args *a;
5962 	struct vnode *vp;
5963 
5964 	a = ap;
5965 	vp = a->a_vp;
5966 	vn_seqc_write_begin(vp);
5967 }
5968 
5969 void
5970 vop_setextattr_post(void *ap, int rc)
5971 {
5972 	struct vop_setextattr_args *a;
5973 	struct vnode *vp;
5974 
5975 	a = ap;
5976 	vp = a->a_vp;
5977 	vn_seqc_write_end(vp);
5978 	if (!rc)
5979 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5980 }
5981 
5982 void
5983 vop_symlink_pre(void *ap)
5984 {
5985 	struct vop_symlink_args *a;
5986 	struct vnode *dvp;
5987 
5988 	a = ap;
5989 	dvp = a->a_dvp;
5990 	vn_seqc_write_begin(dvp);
5991 }
5992 
5993 void
5994 vop_symlink_post(void *ap, int rc)
5995 {
5996 	struct vop_symlink_args *a;
5997 	struct vnode *dvp;
5998 
5999 	a = ap;
6000 	dvp = a->a_dvp;
6001 	vn_seqc_write_end(dvp);
6002 	if (!rc)
6003 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6004 }
6005 
6006 void
6007 vop_open_post(void *ap, int rc)
6008 {
6009 	struct vop_open_args *a = ap;
6010 
6011 	if (!rc)
6012 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6013 }
6014 
6015 void
6016 vop_close_post(void *ap, int rc)
6017 {
6018 	struct vop_close_args *a = ap;
6019 
6020 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6021 	    !VN_IS_DOOMED(a->a_vp))) {
6022 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6023 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6024 	}
6025 }
6026 
6027 void
6028 vop_read_post(void *ap, int rc)
6029 {
6030 	struct vop_read_args *a = ap;
6031 
6032 	if (!rc)
6033 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6034 }
6035 
6036 void
6037 vop_read_pgcache_post(void *ap, int rc)
6038 {
6039 	struct vop_read_pgcache_args *a = ap;
6040 
6041 	if (!rc)
6042 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6043 }
6044 
6045 void
6046 vop_readdir_post(void *ap, int rc)
6047 {
6048 	struct vop_readdir_args *a = ap;
6049 
6050 	if (!rc)
6051 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6052 }
6053 
6054 static struct knlist fs_knlist;
6055 
6056 static void
6057 vfs_event_init(void *arg)
6058 {
6059 	knlist_init_mtx(&fs_knlist, NULL);
6060 }
6061 /* XXX - correct order? */
6062 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6063 
6064 void
6065 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6066 {
6067 
6068 	KNOTE_UNLOCKED(&fs_knlist, event);
6069 }
6070 
6071 static int	filt_fsattach(struct knote *kn);
6072 static void	filt_fsdetach(struct knote *kn);
6073 static int	filt_fsevent(struct knote *kn, long hint);
6074 
6075 struct filterops fs_filtops = {
6076 	.f_isfd = 0,
6077 	.f_attach = filt_fsattach,
6078 	.f_detach = filt_fsdetach,
6079 	.f_event = filt_fsevent
6080 };
6081 
6082 static int
6083 filt_fsattach(struct knote *kn)
6084 {
6085 
6086 	kn->kn_flags |= EV_CLEAR;
6087 	knlist_add(&fs_knlist, kn, 0);
6088 	return (0);
6089 }
6090 
6091 static void
6092 filt_fsdetach(struct knote *kn)
6093 {
6094 
6095 	knlist_remove(&fs_knlist, kn, 0);
6096 }
6097 
6098 static int
6099 filt_fsevent(struct knote *kn, long hint)
6100 {
6101 
6102 	kn->kn_fflags |= kn->kn_sfflags & hint;
6103 
6104 	return (kn->kn_fflags != 0);
6105 }
6106 
6107 static int
6108 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6109 {
6110 	struct vfsidctl vc;
6111 	int error;
6112 	struct mount *mp;
6113 
6114 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6115 	if (error)
6116 		return (error);
6117 	if (vc.vc_vers != VFS_CTL_VERS1)
6118 		return (EINVAL);
6119 	mp = vfs_getvfs(&vc.vc_fsid);
6120 	if (mp == NULL)
6121 		return (ENOENT);
6122 	/* ensure that a specific sysctl goes to the right filesystem. */
6123 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6124 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6125 		vfs_rel(mp);
6126 		return (EINVAL);
6127 	}
6128 	VCTLTOREQ(&vc, req);
6129 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6130 	vfs_rel(mp);
6131 	return (error);
6132 }
6133 
6134 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6135     NULL, 0, sysctl_vfs_ctl, "",
6136     "Sysctl by fsid");
6137 
6138 /*
6139  * Function to initialize a va_filerev field sensibly.
6140  * XXX: Wouldn't a random number make a lot more sense ??
6141  */
6142 u_quad_t
6143 init_va_filerev(void)
6144 {
6145 	struct bintime bt;
6146 
6147 	getbinuptime(&bt);
6148 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6149 }
6150 
6151 static int	filt_vfsread(struct knote *kn, long hint);
6152 static int	filt_vfswrite(struct knote *kn, long hint);
6153 static int	filt_vfsvnode(struct knote *kn, long hint);
6154 static void	filt_vfsdetach(struct knote *kn);
6155 static struct filterops vfsread_filtops = {
6156 	.f_isfd = 1,
6157 	.f_detach = filt_vfsdetach,
6158 	.f_event = filt_vfsread
6159 };
6160 static struct filterops vfswrite_filtops = {
6161 	.f_isfd = 1,
6162 	.f_detach = filt_vfsdetach,
6163 	.f_event = filt_vfswrite
6164 };
6165 static struct filterops vfsvnode_filtops = {
6166 	.f_isfd = 1,
6167 	.f_detach = filt_vfsdetach,
6168 	.f_event = filt_vfsvnode
6169 };
6170 
6171 static void
6172 vfs_knllock(void *arg)
6173 {
6174 	struct vnode *vp = arg;
6175 
6176 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6177 }
6178 
6179 static void
6180 vfs_knlunlock(void *arg)
6181 {
6182 	struct vnode *vp = arg;
6183 
6184 	VOP_UNLOCK(vp);
6185 }
6186 
6187 static void
6188 vfs_knl_assert_lock(void *arg, int what)
6189 {
6190 #ifdef DEBUG_VFS_LOCKS
6191 	struct vnode *vp = arg;
6192 
6193 	if (what == LA_LOCKED)
6194 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6195 	else
6196 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6197 #endif
6198 }
6199 
6200 int
6201 vfs_kqfilter(struct vop_kqfilter_args *ap)
6202 {
6203 	struct vnode *vp = ap->a_vp;
6204 	struct knote *kn = ap->a_kn;
6205 	struct knlist *knl;
6206 
6207 	switch (kn->kn_filter) {
6208 	case EVFILT_READ:
6209 		kn->kn_fop = &vfsread_filtops;
6210 		break;
6211 	case EVFILT_WRITE:
6212 		kn->kn_fop = &vfswrite_filtops;
6213 		break;
6214 	case EVFILT_VNODE:
6215 		kn->kn_fop = &vfsvnode_filtops;
6216 		break;
6217 	default:
6218 		return (EINVAL);
6219 	}
6220 
6221 	kn->kn_hook = (caddr_t)vp;
6222 
6223 	v_addpollinfo(vp);
6224 	if (vp->v_pollinfo == NULL)
6225 		return (ENOMEM);
6226 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6227 	vhold(vp);
6228 	knlist_add(knl, kn, 0);
6229 
6230 	return (0);
6231 }
6232 
6233 /*
6234  * Detach knote from vnode
6235  */
6236 static void
6237 filt_vfsdetach(struct knote *kn)
6238 {
6239 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6240 
6241 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6242 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6243 	vdrop(vp);
6244 }
6245 
6246 /*ARGSUSED*/
6247 static int
6248 filt_vfsread(struct knote *kn, long hint)
6249 {
6250 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6251 	struct vattr va;
6252 	int res;
6253 
6254 	/*
6255 	 * filesystem is gone, so set the EOF flag and schedule
6256 	 * the knote for deletion.
6257 	 */
6258 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6259 		VI_LOCK(vp);
6260 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6261 		VI_UNLOCK(vp);
6262 		return (1);
6263 	}
6264 
6265 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6266 		return (0);
6267 
6268 	VI_LOCK(vp);
6269 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6270 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6271 	VI_UNLOCK(vp);
6272 	return (res);
6273 }
6274 
6275 /*ARGSUSED*/
6276 static int
6277 filt_vfswrite(struct knote *kn, long hint)
6278 {
6279 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6280 
6281 	VI_LOCK(vp);
6282 
6283 	/*
6284 	 * filesystem is gone, so set the EOF flag and schedule
6285 	 * the knote for deletion.
6286 	 */
6287 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6288 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6289 
6290 	kn->kn_data = 0;
6291 	VI_UNLOCK(vp);
6292 	return (1);
6293 }
6294 
6295 static int
6296 filt_vfsvnode(struct knote *kn, long hint)
6297 {
6298 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6299 	int res;
6300 
6301 	VI_LOCK(vp);
6302 	if (kn->kn_sfflags & hint)
6303 		kn->kn_fflags |= hint;
6304 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6305 		kn->kn_flags |= EV_EOF;
6306 		VI_UNLOCK(vp);
6307 		return (1);
6308 	}
6309 	res = (kn->kn_fflags != 0);
6310 	VI_UNLOCK(vp);
6311 	return (res);
6312 }
6313 
6314 /*
6315  * Returns whether the directory is empty or not.
6316  * If it is empty, the return value is 0; otherwise
6317  * the return value is an error value (which may
6318  * be ENOTEMPTY).
6319  */
6320 int
6321 vfs_emptydir(struct vnode *vp)
6322 {
6323 	struct uio uio;
6324 	struct iovec iov;
6325 	struct dirent *dirent, *dp, *endp;
6326 	int error, eof;
6327 
6328 	error = 0;
6329 	eof = 0;
6330 
6331 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6332 
6333 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6334 	iov.iov_base = dirent;
6335 	iov.iov_len = sizeof(struct dirent);
6336 
6337 	uio.uio_iov = &iov;
6338 	uio.uio_iovcnt = 1;
6339 	uio.uio_offset = 0;
6340 	uio.uio_resid = sizeof(struct dirent);
6341 	uio.uio_segflg = UIO_SYSSPACE;
6342 	uio.uio_rw = UIO_READ;
6343 	uio.uio_td = curthread;
6344 
6345 	while (eof == 0 && error == 0) {
6346 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6347 		    NULL, NULL);
6348 		if (error != 0)
6349 			break;
6350 		endp = (void *)((uint8_t *)dirent +
6351 		    sizeof(struct dirent) - uio.uio_resid);
6352 		for (dp = dirent; dp < endp;
6353 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6354 			if (dp->d_type == DT_WHT)
6355 				continue;
6356 			if (dp->d_namlen == 0)
6357 				continue;
6358 			if (dp->d_type != DT_DIR &&
6359 			    dp->d_type != DT_UNKNOWN) {
6360 				error = ENOTEMPTY;
6361 				break;
6362 			}
6363 			if (dp->d_namlen > 2) {
6364 				error = ENOTEMPTY;
6365 				break;
6366 			}
6367 			if (dp->d_namlen == 1 &&
6368 			    dp->d_name[0] != '.') {
6369 				error = ENOTEMPTY;
6370 				break;
6371 			}
6372 			if (dp->d_namlen == 2 &&
6373 			    dp->d_name[1] != '.') {
6374 				error = ENOTEMPTY;
6375 				break;
6376 			}
6377 			uio.uio_resid = sizeof(struct dirent);
6378 		}
6379 	}
6380 	free(dirent, M_TEMP);
6381 	return (error);
6382 }
6383 
6384 int
6385 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6386 {
6387 	int error;
6388 
6389 	if (dp->d_reclen > ap->a_uio->uio_resid)
6390 		return (ENAMETOOLONG);
6391 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6392 	if (error) {
6393 		if (ap->a_ncookies != NULL) {
6394 			if (ap->a_cookies != NULL)
6395 				free(ap->a_cookies, M_TEMP);
6396 			ap->a_cookies = NULL;
6397 			*ap->a_ncookies = 0;
6398 		}
6399 		return (error);
6400 	}
6401 	if (ap->a_ncookies == NULL)
6402 		return (0);
6403 
6404 	KASSERT(ap->a_cookies,
6405 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6406 
6407 	*ap->a_cookies = realloc(*ap->a_cookies,
6408 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6409 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6410 	*ap->a_ncookies += 1;
6411 	return (0);
6412 }
6413 
6414 /*
6415  * The purpose of this routine is to remove granularity from accmode_t,
6416  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6417  * VADMIN and VAPPEND.
6418  *
6419  * If it returns 0, the caller is supposed to continue with the usual
6420  * access checks using 'accmode' as modified by this routine.  If it
6421  * returns nonzero value, the caller is supposed to return that value
6422  * as errno.
6423  *
6424  * Note that after this routine runs, accmode may be zero.
6425  */
6426 int
6427 vfs_unixify_accmode(accmode_t *accmode)
6428 {
6429 	/*
6430 	 * There is no way to specify explicit "deny" rule using
6431 	 * file mode or POSIX.1e ACLs.
6432 	 */
6433 	if (*accmode & VEXPLICIT_DENY) {
6434 		*accmode = 0;
6435 		return (0);
6436 	}
6437 
6438 	/*
6439 	 * None of these can be translated into usual access bits.
6440 	 * Also, the common case for NFSv4 ACLs is to not contain
6441 	 * either of these bits. Caller should check for VWRITE
6442 	 * on the containing directory instead.
6443 	 */
6444 	if (*accmode & (VDELETE_CHILD | VDELETE))
6445 		return (EPERM);
6446 
6447 	if (*accmode & VADMIN_PERMS) {
6448 		*accmode &= ~VADMIN_PERMS;
6449 		*accmode |= VADMIN;
6450 	}
6451 
6452 	/*
6453 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6454 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6455 	 */
6456 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6457 
6458 	return (0);
6459 }
6460 
6461 /*
6462  * Clear out a doomed vnode (if any) and replace it with a new one as long
6463  * as the fs is not being unmounted. Return the root vnode to the caller.
6464  */
6465 static int __noinline
6466 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6467 {
6468 	struct vnode *vp;
6469 	int error;
6470 
6471 restart:
6472 	if (mp->mnt_rootvnode != NULL) {
6473 		MNT_ILOCK(mp);
6474 		vp = mp->mnt_rootvnode;
6475 		if (vp != NULL) {
6476 			if (!VN_IS_DOOMED(vp)) {
6477 				vrefact(vp);
6478 				MNT_IUNLOCK(mp);
6479 				error = vn_lock(vp, flags);
6480 				if (error == 0) {
6481 					*vpp = vp;
6482 					return (0);
6483 				}
6484 				vrele(vp);
6485 				goto restart;
6486 			}
6487 			/*
6488 			 * Clear the old one.
6489 			 */
6490 			mp->mnt_rootvnode = NULL;
6491 		}
6492 		MNT_IUNLOCK(mp);
6493 		if (vp != NULL) {
6494 			vfs_op_barrier_wait(mp);
6495 			vrele(vp);
6496 		}
6497 	}
6498 	error = VFS_CACHEDROOT(mp, flags, vpp);
6499 	if (error != 0)
6500 		return (error);
6501 	if (mp->mnt_vfs_ops == 0) {
6502 		MNT_ILOCK(mp);
6503 		if (mp->mnt_vfs_ops != 0) {
6504 			MNT_IUNLOCK(mp);
6505 			return (0);
6506 		}
6507 		if (mp->mnt_rootvnode == NULL) {
6508 			vrefact(*vpp);
6509 			mp->mnt_rootvnode = *vpp;
6510 		} else {
6511 			if (mp->mnt_rootvnode != *vpp) {
6512 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6513 					panic("%s: mismatch between vnode returned "
6514 					    " by VFS_CACHEDROOT and the one cached "
6515 					    " (%p != %p)",
6516 					    __func__, *vpp, mp->mnt_rootvnode);
6517 				}
6518 			}
6519 		}
6520 		MNT_IUNLOCK(mp);
6521 	}
6522 	return (0);
6523 }
6524 
6525 int
6526 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6527 {
6528 	struct mount_pcpu *mpcpu;
6529 	struct vnode *vp;
6530 	int error;
6531 
6532 	if (!vfs_op_thread_enter(mp, mpcpu))
6533 		return (vfs_cache_root_fallback(mp, flags, vpp));
6534 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6535 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6536 		vfs_op_thread_exit(mp, mpcpu);
6537 		return (vfs_cache_root_fallback(mp, flags, vpp));
6538 	}
6539 	vrefact(vp);
6540 	vfs_op_thread_exit(mp, mpcpu);
6541 	error = vn_lock(vp, flags);
6542 	if (error != 0) {
6543 		vrele(vp);
6544 		return (vfs_cache_root_fallback(mp, flags, vpp));
6545 	}
6546 	*vpp = vp;
6547 	return (0);
6548 }
6549 
6550 struct vnode *
6551 vfs_cache_root_clear(struct mount *mp)
6552 {
6553 	struct vnode *vp;
6554 
6555 	/*
6556 	 * ops > 0 guarantees there is nobody who can see this vnode
6557 	 */
6558 	MPASS(mp->mnt_vfs_ops > 0);
6559 	vp = mp->mnt_rootvnode;
6560 	if (vp != NULL)
6561 		vn_seqc_write_begin(vp);
6562 	mp->mnt_rootvnode = NULL;
6563 	return (vp);
6564 }
6565 
6566 void
6567 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6568 {
6569 
6570 	MPASS(mp->mnt_vfs_ops > 0);
6571 	vrefact(vp);
6572 	mp->mnt_rootvnode = vp;
6573 }
6574 
6575 /*
6576  * These are helper functions for filesystems to traverse all
6577  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6578  *
6579  * This interface replaces MNT_VNODE_FOREACH.
6580  */
6581 
6582 struct vnode *
6583 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6584 {
6585 	struct vnode *vp;
6586 
6587 	if (should_yield())
6588 		kern_yield(PRI_USER);
6589 	MNT_ILOCK(mp);
6590 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6591 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6592 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6593 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6594 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6595 			continue;
6596 		VI_LOCK(vp);
6597 		if (VN_IS_DOOMED(vp)) {
6598 			VI_UNLOCK(vp);
6599 			continue;
6600 		}
6601 		break;
6602 	}
6603 	if (vp == NULL) {
6604 		__mnt_vnode_markerfree_all(mvp, mp);
6605 		/* MNT_IUNLOCK(mp); -- done in above function */
6606 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6607 		return (NULL);
6608 	}
6609 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6610 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6611 	MNT_IUNLOCK(mp);
6612 	return (vp);
6613 }
6614 
6615 struct vnode *
6616 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6617 {
6618 	struct vnode *vp;
6619 
6620 	*mvp = vn_alloc_marker(mp);
6621 	MNT_ILOCK(mp);
6622 	MNT_REF(mp);
6623 
6624 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6625 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6626 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6627 			continue;
6628 		VI_LOCK(vp);
6629 		if (VN_IS_DOOMED(vp)) {
6630 			VI_UNLOCK(vp);
6631 			continue;
6632 		}
6633 		break;
6634 	}
6635 	if (vp == NULL) {
6636 		MNT_REL(mp);
6637 		MNT_IUNLOCK(mp);
6638 		vn_free_marker(*mvp);
6639 		*mvp = NULL;
6640 		return (NULL);
6641 	}
6642 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6643 	MNT_IUNLOCK(mp);
6644 	return (vp);
6645 }
6646 
6647 void
6648 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6649 {
6650 
6651 	if (*mvp == NULL) {
6652 		MNT_IUNLOCK(mp);
6653 		return;
6654 	}
6655 
6656 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6657 
6658 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6659 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6660 	MNT_REL(mp);
6661 	MNT_IUNLOCK(mp);
6662 	vn_free_marker(*mvp);
6663 	*mvp = NULL;
6664 }
6665 
6666 /*
6667  * These are helper functions for filesystems to traverse their
6668  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6669  */
6670 static void
6671 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6672 {
6673 
6674 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6675 
6676 	MNT_ILOCK(mp);
6677 	MNT_REL(mp);
6678 	MNT_IUNLOCK(mp);
6679 	vn_free_marker(*mvp);
6680 	*mvp = NULL;
6681 }
6682 
6683 /*
6684  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6685  * conventional lock order during mnt_vnode_next_lazy iteration.
6686  *
6687  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6688  * The list lock is dropped and reacquired.  On success, both locks are held.
6689  * On failure, the mount vnode list lock is held but the vnode interlock is
6690  * not, and the procedure may have yielded.
6691  */
6692 static bool
6693 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6694     struct vnode *vp)
6695 {
6696 
6697 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6698 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6699 	    ("%s: bad marker", __func__));
6700 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6701 	    ("%s: inappropriate vnode", __func__));
6702 	ASSERT_VI_UNLOCKED(vp, __func__);
6703 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6704 
6705 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6706 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6707 
6708 	/*
6709 	 * Note we may be racing against vdrop which transitioned the hold
6710 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6711 	 * if we are the only user after we get the interlock we will just
6712 	 * vdrop.
6713 	 */
6714 	vhold(vp);
6715 	mtx_unlock(&mp->mnt_listmtx);
6716 	VI_LOCK(vp);
6717 	if (VN_IS_DOOMED(vp)) {
6718 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6719 		goto out_lost;
6720 	}
6721 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6722 	/*
6723 	 * There is nothing to do if we are the last user.
6724 	 */
6725 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6726 		goto out_lost;
6727 	mtx_lock(&mp->mnt_listmtx);
6728 	return (true);
6729 out_lost:
6730 	vdropl(vp);
6731 	maybe_yield();
6732 	mtx_lock(&mp->mnt_listmtx);
6733 	return (false);
6734 }
6735 
6736 static struct vnode *
6737 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6738     void *cbarg)
6739 {
6740 	struct vnode *vp;
6741 
6742 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6743 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6744 restart:
6745 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6746 	while (vp != NULL) {
6747 		if (vp->v_type == VMARKER) {
6748 			vp = TAILQ_NEXT(vp, v_lazylist);
6749 			continue;
6750 		}
6751 		/*
6752 		 * See if we want to process the vnode. Note we may encounter a
6753 		 * long string of vnodes we don't care about and hog the list
6754 		 * as a result. Check for it and requeue the marker.
6755 		 */
6756 		VNPASS(!VN_IS_DOOMED(vp), vp);
6757 		if (!cb(vp, cbarg)) {
6758 			if (!should_yield()) {
6759 				vp = TAILQ_NEXT(vp, v_lazylist);
6760 				continue;
6761 			}
6762 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6763 			    v_lazylist);
6764 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6765 			    v_lazylist);
6766 			mtx_unlock(&mp->mnt_listmtx);
6767 			kern_yield(PRI_USER);
6768 			mtx_lock(&mp->mnt_listmtx);
6769 			goto restart;
6770 		}
6771 		/*
6772 		 * Try-lock because this is the wrong lock order.
6773 		 */
6774 		if (!VI_TRYLOCK(vp) &&
6775 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6776 			goto restart;
6777 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6778 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6779 		    ("alien vnode on the lazy list %p %p", vp, mp));
6780 		VNPASS(vp->v_mount == mp, vp);
6781 		VNPASS(!VN_IS_DOOMED(vp), vp);
6782 		break;
6783 	}
6784 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6785 
6786 	/* Check if we are done */
6787 	if (vp == NULL) {
6788 		mtx_unlock(&mp->mnt_listmtx);
6789 		mnt_vnode_markerfree_lazy(mvp, mp);
6790 		return (NULL);
6791 	}
6792 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6793 	mtx_unlock(&mp->mnt_listmtx);
6794 	ASSERT_VI_LOCKED(vp, "lazy iter");
6795 	return (vp);
6796 }
6797 
6798 struct vnode *
6799 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6800     void *cbarg)
6801 {
6802 
6803 	if (should_yield())
6804 		kern_yield(PRI_USER);
6805 	mtx_lock(&mp->mnt_listmtx);
6806 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6807 }
6808 
6809 struct vnode *
6810 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6811     void *cbarg)
6812 {
6813 	struct vnode *vp;
6814 
6815 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6816 		return (NULL);
6817 
6818 	*mvp = vn_alloc_marker(mp);
6819 	MNT_ILOCK(mp);
6820 	MNT_REF(mp);
6821 	MNT_IUNLOCK(mp);
6822 
6823 	mtx_lock(&mp->mnt_listmtx);
6824 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6825 	if (vp == NULL) {
6826 		mtx_unlock(&mp->mnt_listmtx);
6827 		mnt_vnode_markerfree_lazy(mvp, mp);
6828 		return (NULL);
6829 	}
6830 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6831 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6832 }
6833 
6834 void
6835 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6836 {
6837 
6838 	if (*mvp == NULL)
6839 		return;
6840 
6841 	mtx_lock(&mp->mnt_listmtx);
6842 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6843 	mtx_unlock(&mp->mnt_listmtx);
6844 	mnt_vnode_markerfree_lazy(mvp, mp);
6845 }
6846 
6847 int
6848 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6849 {
6850 
6851 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6852 		cnp->cn_flags &= ~NOEXECCHECK;
6853 		return (0);
6854 	}
6855 
6856 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6857 }
6858 
6859 /*
6860  * Do not use this variant unless you have means other than the hold count
6861  * to prevent the vnode from getting freed.
6862  */
6863 void
6864 vn_seqc_write_begin_unheld_locked(struct vnode *vp)
6865 {
6866 
6867 	ASSERT_VI_LOCKED(vp, __func__);
6868 	VNPASS(vp->v_seqc_users >= 0, vp);
6869 	vp->v_seqc_users++;
6870 	if (vp->v_seqc_users == 1)
6871 		seqc_sleepable_write_begin(&vp->v_seqc);
6872 }
6873 
6874 void
6875 vn_seqc_write_begin_locked(struct vnode *vp)
6876 {
6877 
6878 	ASSERT_VI_LOCKED(vp, __func__);
6879 	VNPASS(vp->v_holdcnt > 0, vp);
6880 	vn_seqc_write_begin_unheld_locked(vp);
6881 }
6882 
6883 void
6884 vn_seqc_write_begin(struct vnode *vp)
6885 {
6886 
6887 	VI_LOCK(vp);
6888 	vn_seqc_write_begin_locked(vp);
6889 	VI_UNLOCK(vp);
6890 }
6891 
6892 void
6893 vn_seqc_write_begin_unheld(struct vnode *vp)
6894 {
6895 
6896 	VI_LOCK(vp);
6897 	vn_seqc_write_begin_unheld_locked(vp);
6898 	VI_UNLOCK(vp);
6899 }
6900 
6901 void
6902 vn_seqc_write_end_locked(struct vnode *vp)
6903 {
6904 
6905 	ASSERT_VI_LOCKED(vp, __func__);
6906 	VNPASS(vp->v_seqc_users > 0, vp);
6907 	vp->v_seqc_users--;
6908 	if (vp->v_seqc_users == 0)
6909 		seqc_sleepable_write_end(&vp->v_seqc);
6910 }
6911 
6912 void
6913 vn_seqc_write_end(struct vnode *vp)
6914 {
6915 
6916 	VI_LOCK(vp);
6917 	vn_seqc_write_end_locked(vp);
6918 	VI_UNLOCK(vp);
6919 }
6920 
6921 /*
6922  * Special case handling for allocating and freeing vnodes.
6923  *
6924  * The counter remains unchanged on free so that a doomed vnode will
6925  * keep testing as in modify as long as it is accessible with SMR.
6926  */
6927 static void
6928 vn_seqc_init(struct vnode *vp)
6929 {
6930 
6931 	vp->v_seqc = 0;
6932 	vp->v_seqc_users = 0;
6933 }
6934 
6935 static void
6936 vn_seqc_write_end_free(struct vnode *vp)
6937 {
6938 
6939 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6940 	VNPASS(vp->v_seqc_users == 1, vp);
6941 }
6942 
6943 void
6944 vn_irflag_set_locked(struct vnode *vp, short toset)
6945 {
6946 	short flags;
6947 
6948 	ASSERT_VI_LOCKED(vp, __func__);
6949 	flags = vn_irflag_read(vp);
6950 	VNASSERT((flags & toset) == 0, vp,
6951 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
6952 	    __func__, flags, toset));
6953 	atomic_store_short(&vp->v_irflag, flags | toset);
6954 }
6955 
6956 void
6957 vn_irflag_set(struct vnode *vp, short toset)
6958 {
6959 
6960 	VI_LOCK(vp);
6961 	vn_irflag_set_locked(vp, toset);
6962 	VI_UNLOCK(vp);
6963 }
6964 
6965 void
6966 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6967 {
6968 	short flags;
6969 
6970 	ASSERT_VI_LOCKED(vp, __func__);
6971 	flags = vn_irflag_read(vp);
6972 	atomic_store_short(&vp->v_irflag, flags | toset);
6973 }
6974 
6975 void
6976 vn_irflag_set_cond(struct vnode *vp, short toset)
6977 {
6978 
6979 	VI_LOCK(vp);
6980 	vn_irflag_set_cond_locked(vp, toset);
6981 	VI_UNLOCK(vp);
6982 }
6983 
6984 void
6985 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6986 {
6987 	short flags;
6988 
6989 	ASSERT_VI_LOCKED(vp, __func__);
6990 	flags = vn_irflag_read(vp);
6991 	VNASSERT((flags & tounset) == tounset, vp,
6992 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
6993 	    __func__, flags, tounset));
6994 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
6995 }
6996 
6997 void
6998 vn_irflag_unset(struct vnode *vp, short tounset)
6999 {
7000 
7001 	VI_LOCK(vp);
7002 	vn_irflag_unset_locked(vp, tounset);
7003 	VI_UNLOCK(vp);
7004 }
7005