xref: /freebsd/sys/kern/vfs_subr.c (revision 0bf48626aaa33768078f5872b922b1487b3a9296)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/stat.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/uma.h>
99 
100 #ifdef DDB
101 #include <ddb/ddb.h>
102 #endif
103 
104 static void	delmntque(struct vnode *vp);
105 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 		    int slpflag, int slptimeo);
107 static void	syncer_shutdown(void *arg, int howto);
108 static int	vtryrecycle(struct vnode *vp);
109 static void	v_init_counters(struct vnode *);
110 static void	v_incr_usecount(struct vnode *);
111 static void	v_incr_usecount_locked(struct vnode *);
112 static void	v_incr_devcount(struct vnode *);
113 static void	v_decr_devcount(struct vnode *);
114 static void	vgonel(struct vnode *);
115 static void	vfs_knllock(void *arg);
116 static void	vfs_knlunlock(void *arg);
117 static void	vfs_knl_assert_locked(void *arg);
118 static void	vfs_knl_assert_unlocked(void *arg);
119 static void	vnlru_return_batches(struct vfsops *mnt_op);
120 static void	destroy_vpollinfo(struct vpollinfo *vi);
121 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
122 		    daddr_t startlbn, daddr_t endlbn);
123 
124 /*
125  * These fences are intended for cases where some synchronization is
126  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
127  * and v_usecount) updates.  Access to v_iflags is generally synchronized
128  * by the interlock, but we have some internal assertions that check vnode
129  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
130  * for now.
131  */
132 #ifdef INVARIANTS
133 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
134 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
135 #else
136 #define	VNODE_REFCOUNT_FENCE_ACQ()
137 #define	VNODE_REFCOUNT_FENCE_REL()
138 #endif
139 
140 /*
141  * Number of vnodes in existence.  Increased whenever getnewvnode()
142  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
143  */
144 static unsigned long	numvnodes;
145 
146 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
147     "Number of vnodes in existence");
148 
149 static counter_u64_t vnodes_created;
150 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
151     "Number of vnodes created by getnewvnode");
152 
153 static u_long mnt_free_list_batch = 128;
154 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
155     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
156 
157 /*
158  * Conversion tables for conversion from vnode types to inode formats
159  * and back.
160  */
161 enum vtype iftovt_tab[16] = {
162 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
163 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
164 };
165 int vttoif_tab[10] = {
166 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
167 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
168 };
169 
170 /*
171  * List of vnodes that are ready for recycling.
172  */
173 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
174 
175 /*
176  * "Free" vnode target.  Free vnodes are rarely completely free, but are
177  * just ones that are cheap to recycle.  Usually they are for files which
178  * have been stat'd but not read; these usually have inode and namecache
179  * data attached to them.  This target is the preferred minimum size of a
180  * sub-cache consisting mostly of such files. The system balances the size
181  * of this sub-cache with its complement to try to prevent either from
182  * thrashing while the other is relatively inactive.  The targets express
183  * a preference for the best balance.
184  *
185  * "Above" this target there are 2 further targets (watermarks) related
186  * to recyling of free vnodes.  In the best-operating case, the cache is
187  * exactly full, the free list has size between vlowat and vhiwat above the
188  * free target, and recycling from it and normal use maintains this state.
189  * Sometimes the free list is below vlowat or even empty, but this state
190  * is even better for immediate use provided the cache is not full.
191  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
192  * ones) to reach one of these states.  The watermarks are currently hard-
193  * coded as 4% and 9% of the available space higher.  These and the default
194  * of 25% for wantfreevnodes are too large if the memory size is large.
195  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
196  * whenever vnlru_proc() becomes active.
197  */
198 static u_long wantfreevnodes;
199 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
200     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
201 static u_long freevnodes;
202 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
203     &freevnodes, 0, "Number of \"free\" vnodes");
204 
205 static counter_u64_t recycles_count;
206 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
207     "Number of vnodes recycled to meet vnode cache targets");
208 
209 /*
210  * Various variables used for debugging the new implementation of
211  * reassignbuf().
212  * XXX these are probably of (very) limited utility now.
213  */
214 static int reassignbufcalls;
215 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
216     "Number of calls to reassignbuf");
217 
218 static counter_u64_t free_owe_inact;
219 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
220     "Number of times free vnodes kept on active list due to VFS "
221     "owing inactivation");
222 
223 /* To keep more than one thread at a time from running vfs_getnewfsid */
224 static struct mtx mntid_mtx;
225 
226 /*
227  * Lock for any access to the following:
228  *	vnode_free_list
229  *	numvnodes
230  *	freevnodes
231  */
232 static struct mtx vnode_free_list_mtx;
233 
234 /* Publicly exported FS */
235 struct nfs_public nfs_pub;
236 
237 static uma_zone_t buf_trie_zone;
238 
239 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
240 static uma_zone_t vnode_zone;
241 static uma_zone_t vnodepoll_zone;
242 
243 /*
244  * The workitem queue.
245  *
246  * It is useful to delay writes of file data and filesystem metadata
247  * for tens of seconds so that quickly created and deleted files need
248  * not waste disk bandwidth being created and removed. To realize this,
249  * we append vnodes to a "workitem" queue. When running with a soft
250  * updates implementation, most pending metadata dependencies should
251  * not wait for more than a few seconds. Thus, mounted on block devices
252  * are delayed only about a half the time that file data is delayed.
253  * Similarly, directory updates are more critical, so are only delayed
254  * about a third the time that file data is delayed. Thus, there are
255  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
256  * one each second (driven off the filesystem syncer process). The
257  * syncer_delayno variable indicates the next queue that is to be processed.
258  * Items that need to be processed soon are placed in this queue:
259  *
260  *	syncer_workitem_pending[syncer_delayno]
261  *
262  * A delay of fifteen seconds is done by placing the request fifteen
263  * entries later in the queue:
264  *
265  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
266  *
267  */
268 static int syncer_delayno;
269 static long syncer_mask;
270 LIST_HEAD(synclist, bufobj);
271 static struct synclist *syncer_workitem_pending;
272 /*
273  * The sync_mtx protects:
274  *	bo->bo_synclist
275  *	sync_vnode_count
276  *	syncer_delayno
277  *	syncer_state
278  *	syncer_workitem_pending
279  *	syncer_worklist_len
280  *	rushjob
281  */
282 static struct mtx sync_mtx;
283 static struct cv sync_wakeup;
284 
285 #define SYNCER_MAXDELAY		32
286 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
287 static int syncdelay = 30;		/* max time to delay syncing data */
288 static int filedelay = 30;		/* time to delay syncing files */
289 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
290     "Time to delay syncing files (in seconds)");
291 static int dirdelay = 29;		/* time to delay syncing directories */
292 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
293     "Time to delay syncing directories (in seconds)");
294 static int metadelay = 28;		/* time to delay syncing metadata */
295 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
296     "Time to delay syncing metadata (in seconds)");
297 static int rushjob;		/* number of slots to run ASAP */
298 static int stat_rush_requests;	/* number of times I/O speeded up */
299 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
300     "Number of times I/O speeded up (rush requests)");
301 
302 /*
303  * When shutting down the syncer, run it at four times normal speed.
304  */
305 #define SYNCER_SHUTDOWN_SPEEDUP		4
306 static int sync_vnode_count;
307 static int syncer_worklist_len;
308 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
309     syncer_state;
310 
311 /* Target for maximum number of vnodes. */
312 int desiredvnodes;
313 static int gapvnodes;		/* gap between wanted and desired */
314 static int vhiwat;		/* enough extras after expansion */
315 static int vlowat;		/* minimal extras before expansion */
316 static int vstir;		/* nonzero to stir non-free vnodes */
317 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
318 
319 static int
320 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
321 {
322 	int error, old_desiredvnodes;
323 
324 	old_desiredvnodes = desiredvnodes;
325 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
326 		return (error);
327 	if (old_desiredvnodes != desiredvnodes) {
328 		wantfreevnodes = desiredvnodes / 4;
329 		/* XXX locking seems to be incomplete. */
330 		vfs_hash_changesize(desiredvnodes);
331 		cache_changesize(desiredvnodes);
332 	}
333 	return (0);
334 }
335 
336 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
337     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
338     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
339 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
340     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
341 static int vnlru_nowhere;
342 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
343     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
344 
345 static int
346 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
347 {
348 	struct vnode *vp;
349 	struct nameidata nd;
350 	char *buf;
351 	unsigned long ndflags;
352 	int error;
353 
354 	if (req->newptr == NULL)
355 		return (EINVAL);
356 	if (req->newlen > PATH_MAX)
357 		return (E2BIG);
358 
359 	buf = malloc(PATH_MAX + 1, M_TEMP, M_WAITOK);
360 	error = SYSCTL_IN(req, buf, req->newlen);
361 	if (error != 0)
362 		goto out;
363 
364 	buf[req->newlen] = '\0';
365 
366 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
367 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
368 	if ((error = namei(&nd)) != 0)
369 		goto out;
370 	vp = nd.ni_vp;
371 
372 	if ((vp->v_iflag & VI_DOOMED) != 0) {
373 		/*
374 		 * This vnode is being recycled.  Return != 0 to let the caller
375 		 * know that the sysctl had no effect.  Return EAGAIN because a
376 		 * subsequent call will likely succeed (since namei will create
377 		 * a new vnode if necessary)
378 		 */
379 		error = EAGAIN;
380 		goto putvnode;
381 	}
382 
383 	counter_u64_add(recycles_count, 1);
384 	vgone(vp);
385 putvnode:
386 	NDFREE(&nd, 0);
387 out:
388 	free(buf, M_TEMP);
389 	return (error);
390 }
391 
392 static int
393 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
394 {
395 	struct thread *td = curthread;
396 	struct vnode *vp;
397 	struct file *fp;
398 	int error;
399 	int fd;
400 
401 	if (req->newptr == NULL)
402 		return (EBADF);
403 
404         error = sysctl_handle_int(oidp, &fd, 0, req);
405         if (error != 0)
406                 return (error);
407 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
408 	if (error != 0)
409 		return (error);
410 	vp = fp->f_vnode;
411 
412 	error = vn_lock(vp, LK_EXCLUSIVE);
413 	if (error != 0)
414 		goto drop;
415 
416 	counter_u64_add(recycles_count, 1);
417 	vgone(vp);
418 	VOP_UNLOCK(vp, 0);
419 drop:
420 	fdrop(fp, td);
421 	return (error);
422 }
423 
424 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
425     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
426     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
427 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
428     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
429     sysctl_ftry_reclaim_vnode, "I",
430     "Try to reclaim a vnode by its file descriptor");
431 
432 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
433 static int vnsz2log;
434 
435 /*
436  * Support for the bufobj clean & dirty pctrie.
437  */
438 static void *
439 buf_trie_alloc(struct pctrie *ptree)
440 {
441 
442 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
443 }
444 
445 static void
446 buf_trie_free(struct pctrie *ptree, void *node)
447 {
448 
449 	uma_zfree(buf_trie_zone, node);
450 }
451 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
452 
453 /*
454  * Initialize the vnode management data structures.
455  *
456  * Reevaluate the following cap on the number of vnodes after the physical
457  * memory size exceeds 512GB.  In the limit, as the physical memory size
458  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
459  */
460 #ifndef	MAXVNODES_MAX
461 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
462 #endif
463 
464 /*
465  * Initialize a vnode as it first enters the zone.
466  */
467 static int
468 vnode_init(void *mem, int size, int flags)
469 {
470 	struct vnode *vp;
471 
472 	vp = mem;
473 	bzero(vp, size);
474 	/*
475 	 * Setup locks.
476 	 */
477 	vp->v_vnlock = &vp->v_lock;
478 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
479 	/*
480 	 * By default, don't allow shared locks unless filesystems opt-in.
481 	 */
482 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
483 	    LK_NOSHARE | LK_IS_VNODE);
484 	/*
485 	 * Initialize bufobj.
486 	 */
487 	bufobj_init(&vp->v_bufobj, vp);
488 	/*
489 	 * Initialize namecache.
490 	 */
491 	LIST_INIT(&vp->v_cache_src);
492 	TAILQ_INIT(&vp->v_cache_dst);
493 	/*
494 	 * Initialize rangelocks.
495 	 */
496 	rangelock_init(&vp->v_rl);
497 	return (0);
498 }
499 
500 /*
501  * Free a vnode when it is cleared from the zone.
502  */
503 static void
504 vnode_fini(void *mem, int size)
505 {
506 	struct vnode *vp;
507 	struct bufobj *bo;
508 
509 	vp = mem;
510 	rangelock_destroy(&vp->v_rl);
511 	lockdestroy(vp->v_vnlock);
512 	mtx_destroy(&vp->v_interlock);
513 	bo = &vp->v_bufobj;
514 	rw_destroy(BO_LOCKPTR(bo));
515 }
516 
517 /*
518  * Provide the size of NFS nclnode and NFS fh for calculation of the
519  * vnode memory consumption.  The size is specified directly to
520  * eliminate dependency on NFS-private header.
521  *
522  * Other filesystems may use bigger or smaller (like UFS and ZFS)
523  * private inode data, but the NFS-based estimation is ample enough.
524  * Still, we care about differences in the size between 64- and 32-bit
525  * platforms.
526  *
527  * Namecache structure size is heuristically
528  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
529  */
530 #ifdef _LP64
531 #define	NFS_NCLNODE_SZ	(528 + 64)
532 #define	NC_SZ		148
533 #else
534 #define	NFS_NCLNODE_SZ	(360 + 32)
535 #define	NC_SZ		92
536 #endif
537 
538 static void
539 vntblinit(void *dummy __unused)
540 {
541 	u_int i;
542 	int physvnodes, virtvnodes;
543 
544 	/*
545 	 * Desiredvnodes is a function of the physical memory size and the
546 	 * kernel's heap size.  Generally speaking, it scales with the
547 	 * physical memory size.  The ratio of desiredvnodes to the physical
548 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
549 	 * Thereafter, the
550 	 * marginal ratio of desiredvnodes to the physical memory size is
551 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
552 	 * size.  The memory required by desiredvnodes vnodes and vm objects
553 	 * must not exceed 1/10th of the kernel's heap size.
554 	 */
555 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
556 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
557 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
558 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
559 	desiredvnodes = min(physvnodes, virtvnodes);
560 	if (desiredvnodes > MAXVNODES_MAX) {
561 		if (bootverbose)
562 			printf("Reducing kern.maxvnodes %d -> %d\n",
563 			    desiredvnodes, MAXVNODES_MAX);
564 		desiredvnodes = MAXVNODES_MAX;
565 	}
566 	wantfreevnodes = desiredvnodes / 4;
567 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
568 	TAILQ_INIT(&vnode_free_list);
569 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
570 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
571 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
572 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
573 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
574 	/*
575 	 * Preallocate enough nodes to support one-per buf so that
576 	 * we can not fail an insert.  reassignbuf() callers can not
577 	 * tolerate the insertion failure.
578 	 */
579 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
580 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
581 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
582 	uma_prealloc(buf_trie_zone, nbuf);
583 
584 	vnodes_created = counter_u64_alloc(M_WAITOK);
585 	recycles_count = counter_u64_alloc(M_WAITOK);
586 	free_owe_inact = counter_u64_alloc(M_WAITOK);
587 
588 	/*
589 	 * Initialize the filesystem syncer.
590 	 */
591 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
592 	    &syncer_mask);
593 	syncer_maxdelay = syncer_mask + 1;
594 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
595 	cv_init(&sync_wakeup, "syncer");
596 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
597 		vnsz2log++;
598 	vnsz2log--;
599 }
600 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
601 
602 
603 /*
604  * Mark a mount point as busy. Used to synchronize access and to delay
605  * unmounting. Eventually, mountlist_mtx is not released on failure.
606  *
607  * vfs_busy() is a custom lock, it can block the caller.
608  * vfs_busy() only sleeps if the unmount is active on the mount point.
609  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
610  * vnode belonging to mp.
611  *
612  * Lookup uses vfs_busy() to traverse mount points.
613  * root fs			var fs
614  * / vnode lock		A	/ vnode lock (/var)		D
615  * /var vnode lock	B	/log vnode lock(/var/log)	E
616  * vfs_busy lock	C	vfs_busy lock			F
617  *
618  * Within each file system, the lock order is C->A->B and F->D->E.
619  *
620  * When traversing across mounts, the system follows that lock order:
621  *
622  *        C->A->B
623  *              |
624  *              +->F->D->E
625  *
626  * The lookup() process for namei("/var") illustrates the process:
627  *  VOP_LOOKUP() obtains B while A is held
628  *  vfs_busy() obtains a shared lock on F while A and B are held
629  *  vput() releases lock on B
630  *  vput() releases lock on A
631  *  VFS_ROOT() obtains lock on D while shared lock on F is held
632  *  vfs_unbusy() releases shared lock on F
633  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
634  *    Attempt to lock A (instead of vp_crossmp) while D is held would
635  *    violate the global order, causing deadlocks.
636  *
637  * dounmount() locks B while F is drained.
638  */
639 int
640 vfs_busy(struct mount *mp, int flags)
641 {
642 
643 	MPASS((flags & ~MBF_MASK) == 0);
644 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
645 
646 	MNT_ILOCK(mp);
647 	MNT_REF(mp);
648 	/*
649 	 * If mount point is currently being unmounted, sleep until the
650 	 * mount point fate is decided.  If thread doing the unmounting fails,
651 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
652 	 * that this mount point has survived the unmount attempt and vfs_busy
653 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
654 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
655 	 * about to be really destroyed.  vfs_busy needs to release its
656 	 * reference on the mount point in this case and return with ENOENT,
657 	 * telling the caller that mount mount it tried to busy is no longer
658 	 * valid.
659 	 */
660 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
661 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
662 			MNT_REL(mp);
663 			MNT_IUNLOCK(mp);
664 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
665 			    __func__);
666 			return (ENOENT);
667 		}
668 		if (flags & MBF_MNTLSTLOCK)
669 			mtx_unlock(&mountlist_mtx);
670 		mp->mnt_kern_flag |= MNTK_MWAIT;
671 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
672 		if (flags & MBF_MNTLSTLOCK)
673 			mtx_lock(&mountlist_mtx);
674 		MNT_ILOCK(mp);
675 	}
676 	if (flags & MBF_MNTLSTLOCK)
677 		mtx_unlock(&mountlist_mtx);
678 	mp->mnt_lockref++;
679 	MNT_IUNLOCK(mp);
680 	return (0);
681 }
682 
683 /*
684  * Free a busy filesystem.
685  */
686 void
687 vfs_unbusy(struct mount *mp)
688 {
689 
690 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
691 	MNT_ILOCK(mp);
692 	MNT_REL(mp);
693 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
694 	mp->mnt_lockref--;
695 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
696 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
697 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
698 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
699 		wakeup(&mp->mnt_lockref);
700 	}
701 	MNT_IUNLOCK(mp);
702 }
703 
704 /*
705  * Lookup a mount point by filesystem identifier.
706  */
707 struct mount *
708 vfs_getvfs(fsid_t *fsid)
709 {
710 	struct mount *mp;
711 
712 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
713 	mtx_lock(&mountlist_mtx);
714 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
715 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
716 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
717 			vfs_ref(mp);
718 			mtx_unlock(&mountlist_mtx);
719 			return (mp);
720 		}
721 	}
722 	mtx_unlock(&mountlist_mtx);
723 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
724 	return ((struct mount *) 0);
725 }
726 
727 /*
728  * Lookup a mount point by filesystem identifier, busying it before
729  * returning.
730  *
731  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
732  * cache for popular filesystem identifiers.  The cache is lockess, using
733  * the fact that struct mount's are never freed.  In worst case we may
734  * get pointer to unmounted or even different filesystem, so we have to
735  * check what we got, and go slow way if so.
736  */
737 struct mount *
738 vfs_busyfs(fsid_t *fsid)
739 {
740 #define	FSID_CACHE_SIZE	256
741 	typedef struct mount * volatile vmp_t;
742 	static vmp_t cache[FSID_CACHE_SIZE];
743 	struct mount *mp;
744 	int error;
745 	uint32_t hash;
746 
747 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
748 	hash = fsid->val[0] ^ fsid->val[1];
749 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
750 	mp = cache[hash];
751 	if (mp == NULL ||
752 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
753 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
754 		goto slow;
755 	if (vfs_busy(mp, 0) != 0) {
756 		cache[hash] = NULL;
757 		goto slow;
758 	}
759 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
760 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
761 		return (mp);
762 	else
763 	    vfs_unbusy(mp);
764 
765 slow:
766 	mtx_lock(&mountlist_mtx);
767 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
768 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
769 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
770 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
771 			if (error) {
772 				cache[hash] = NULL;
773 				mtx_unlock(&mountlist_mtx);
774 				return (NULL);
775 			}
776 			cache[hash] = mp;
777 			return (mp);
778 		}
779 	}
780 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
781 	mtx_unlock(&mountlist_mtx);
782 	return ((struct mount *) 0);
783 }
784 
785 /*
786  * Check if a user can access privileged mount options.
787  */
788 int
789 vfs_suser(struct mount *mp, struct thread *td)
790 {
791 	int error;
792 
793 	if (jailed(td->td_ucred)) {
794 		/*
795 		 * If the jail of the calling thread lacks permission for
796 		 * this type of file system, deny immediately.
797 		 */
798 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
799 			return (EPERM);
800 
801 		/*
802 		 * If the file system was mounted outside the jail of the
803 		 * calling thread, deny immediately.
804 		 */
805 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
806 			return (EPERM);
807 	}
808 
809 	/*
810 	 * If file system supports delegated administration, we don't check
811 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
812 	 * by the file system itself.
813 	 * If this is not the user that did original mount, we check for
814 	 * the PRIV_VFS_MOUNT_OWNER privilege.
815 	 */
816 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
817 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
818 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
819 			return (error);
820 	}
821 	return (0);
822 }
823 
824 /*
825  * Get a new unique fsid.  Try to make its val[0] unique, since this value
826  * will be used to create fake device numbers for stat().  Also try (but
827  * not so hard) make its val[0] unique mod 2^16, since some emulators only
828  * support 16-bit device numbers.  We end up with unique val[0]'s for the
829  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
830  *
831  * Keep in mind that several mounts may be running in parallel.  Starting
832  * the search one past where the previous search terminated is both a
833  * micro-optimization and a defense against returning the same fsid to
834  * different mounts.
835  */
836 void
837 vfs_getnewfsid(struct mount *mp)
838 {
839 	static uint16_t mntid_base;
840 	struct mount *nmp;
841 	fsid_t tfsid;
842 	int mtype;
843 
844 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
845 	mtx_lock(&mntid_mtx);
846 	mtype = mp->mnt_vfc->vfc_typenum;
847 	tfsid.val[1] = mtype;
848 	mtype = (mtype & 0xFF) << 24;
849 	for (;;) {
850 		tfsid.val[0] = makedev(255,
851 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
852 		mntid_base++;
853 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
854 			break;
855 		vfs_rel(nmp);
856 	}
857 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
858 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
859 	mtx_unlock(&mntid_mtx);
860 }
861 
862 /*
863  * Knob to control the precision of file timestamps:
864  *
865  *   0 = seconds only; nanoseconds zeroed.
866  *   1 = seconds and nanoseconds, accurate within 1/HZ.
867  *   2 = seconds and nanoseconds, truncated to microseconds.
868  * >=3 = seconds and nanoseconds, maximum precision.
869  */
870 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
871 
872 static int timestamp_precision = TSP_USEC;
873 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
874     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
875     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
876     "3+: sec + ns (max. precision))");
877 
878 /*
879  * Get a current timestamp.
880  */
881 void
882 vfs_timestamp(struct timespec *tsp)
883 {
884 	struct timeval tv;
885 
886 	switch (timestamp_precision) {
887 	case TSP_SEC:
888 		tsp->tv_sec = time_second;
889 		tsp->tv_nsec = 0;
890 		break;
891 	case TSP_HZ:
892 		getnanotime(tsp);
893 		break;
894 	case TSP_USEC:
895 		microtime(&tv);
896 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
897 		break;
898 	case TSP_NSEC:
899 	default:
900 		nanotime(tsp);
901 		break;
902 	}
903 }
904 
905 /*
906  * Set vnode attributes to VNOVAL
907  */
908 void
909 vattr_null(struct vattr *vap)
910 {
911 
912 	vap->va_type = VNON;
913 	vap->va_size = VNOVAL;
914 	vap->va_bytes = VNOVAL;
915 	vap->va_mode = VNOVAL;
916 	vap->va_nlink = VNOVAL;
917 	vap->va_uid = VNOVAL;
918 	vap->va_gid = VNOVAL;
919 	vap->va_fsid = VNOVAL;
920 	vap->va_fileid = VNOVAL;
921 	vap->va_blocksize = VNOVAL;
922 	vap->va_rdev = VNOVAL;
923 	vap->va_atime.tv_sec = VNOVAL;
924 	vap->va_atime.tv_nsec = VNOVAL;
925 	vap->va_mtime.tv_sec = VNOVAL;
926 	vap->va_mtime.tv_nsec = VNOVAL;
927 	vap->va_ctime.tv_sec = VNOVAL;
928 	vap->va_ctime.tv_nsec = VNOVAL;
929 	vap->va_birthtime.tv_sec = VNOVAL;
930 	vap->va_birthtime.tv_nsec = VNOVAL;
931 	vap->va_flags = VNOVAL;
932 	vap->va_gen = VNOVAL;
933 	vap->va_vaflags = 0;
934 }
935 
936 /*
937  * This routine is called when we have too many vnodes.  It attempts
938  * to free <count> vnodes and will potentially free vnodes that still
939  * have VM backing store (VM backing store is typically the cause
940  * of a vnode blowout so we want to do this).  Therefore, this operation
941  * is not considered cheap.
942  *
943  * A number of conditions may prevent a vnode from being reclaimed.
944  * the buffer cache may have references on the vnode, a directory
945  * vnode may still have references due to the namei cache representing
946  * underlying files, or the vnode may be in active use.   It is not
947  * desirable to reuse such vnodes.  These conditions may cause the
948  * number of vnodes to reach some minimum value regardless of what
949  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
950  *
951  * @param mp		 Try to reclaim vnodes from this mountpoint
952  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
953  * 			 entries if this argument is strue
954  * @param trigger	 Only reclaim vnodes with fewer than this many resident
955  *			 pages.
956  * @return		 The number of vnodes that were reclaimed.
957  */
958 static int
959 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
960 {
961 	struct vnode *vp;
962 	int count, done, target;
963 
964 	done = 0;
965 	vn_start_write(NULL, &mp, V_WAIT);
966 	MNT_ILOCK(mp);
967 	count = mp->mnt_nvnodelistsize;
968 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
969 	target = target / 10 + 1;
970 	while (count != 0 && done < target) {
971 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
972 		while (vp != NULL && vp->v_type == VMARKER)
973 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
974 		if (vp == NULL)
975 			break;
976 		/*
977 		 * XXX LRU is completely broken for non-free vnodes.  First
978 		 * by calling here in mountpoint order, then by moving
979 		 * unselected vnodes to the end here, and most grossly by
980 		 * removing the vlruvp() function that was supposed to
981 		 * maintain the order.  (This function was born broken
982 		 * since syncer problems prevented it doing anything.)  The
983 		 * order is closer to LRC (C = Created).
984 		 *
985 		 * LRU reclaiming of vnodes seems to have last worked in
986 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
987 		 * Then there was no hold count, and inactive vnodes were
988 		 * simply put on the free list in LRU order.  The separate
989 		 * lists also break LRU.  We prefer to reclaim from the
990 		 * free list for technical reasons.  This tends to thrash
991 		 * the free list to keep very unrecently used held vnodes.
992 		 * The problem is mitigated by keeping the free list large.
993 		 */
994 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
995 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
996 		--count;
997 		if (!VI_TRYLOCK(vp))
998 			goto next_iter;
999 		/*
1000 		 * If it's been deconstructed already, it's still
1001 		 * referenced, or it exceeds the trigger, skip it.
1002 		 * Also skip free vnodes.  We are trying to make space
1003 		 * to expand the free list, not reduce it.
1004 		 */
1005 		if (vp->v_usecount ||
1006 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1007 		    ((vp->v_iflag & VI_FREE) != 0) ||
1008 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
1009 		    vp->v_object->resident_page_count > trigger)) {
1010 			VI_UNLOCK(vp);
1011 			goto next_iter;
1012 		}
1013 		MNT_IUNLOCK(mp);
1014 		vholdl(vp);
1015 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1016 			vdrop(vp);
1017 			goto next_iter_mntunlocked;
1018 		}
1019 		VI_LOCK(vp);
1020 		/*
1021 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1022 		 * the vnode interlock and before it was locked again.
1023 		 *
1024 		 * It is not necessary to recheck VI_DOOMED because it can
1025 		 * only be set by another thread that holds both the vnode
1026 		 * lock and vnode interlock.  If another thread has the
1027 		 * vnode lock before we get to VOP_LOCK() and obtains the
1028 		 * vnode interlock after VOP_LOCK() drops the vnode
1029 		 * interlock, the other thread will be unable to drop the
1030 		 * vnode lock before our VOP_LOCK() call fails.
1031 		 */
1032 		if (vp->v_usecount ||
1033 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1034 		    (vp->v_iflag & VI_FREE) != 0 ||
1035 		    (vp->v_object != NULL &&
1036 		    vp->v_object->resident_page_count > trigger)) {
1037 			VOP_UNLOCK(vp, LK_INTERLOCK);
1038 			vdrop(vp);
1039 			goto next_iter_mntunlocked;
1040 		}
1041 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
1042 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
1043 		counter_u64_add(recycles_count, 1);
1044 		vgonel(vp);
1045 		VOP_UNLOCK(vp, 0);
1046 		vdropl(vp);
1047 		done++;
1048 next_iter_mntunlocked:
1049 		if (!should_yield())
1050 			goto relock_mnt;
1051 		goto yield;
1052 next_iter:
1053 		if (!should_yield())
1054 			continue;
1055 		MNT_IUNLOCK(mp);
1056 yield:
1057 		kern_yield(PRI_USER);
1058 relock_mnt:
1059 		MNT_ILOCK(mp);
1060 	}
1061 	MNT_IUNLOCK(mp);
1062 	vn_finished_write(mp);
1063 	return done;
1064 }
1065 
1066 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1067 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1068     0,
1069     "limit on vnode free requests per call to the vnlru_free routine");
1070 
1071 /*
1072  * Attempt to reduce the free list by the requested amount.
1073  */
1074 static void
1075 vnlru_free_locked(int count, struct vfsops *mnt_op)
1076 {
1077 	struct vnode *vp;
1078 	struct mount *mp;
1079 	bool tried_batches;
1080 
1081 	tried_batches = false;
1082 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1083 	if (count > max_vnlru_free)
1084 		count = max_vnlru_free;
1085 	for (; count > 0; count--) {
1086 		vp = TAILQ_FIRST(&vnode_free_list);
1087 		/*
1088 		 * The list can be modified while the free_list_mtx
1089 		 * has been dropped and vp could be NULL here.
1090 		 */
1091 		if (vp == NULL) {
1092 			if (tried_batches)
1093 				break;
1094 			mtx_unlock(&vnode_free_list_mtx);
1095 			vnlru_return_batches(mnt_op);
1096 			tried_batches = true;
1097 			mtx_lock(&vnode_free_list_mtx);
1098 			continue;
1099 		}
1100 
1101 		VNASSERT(vp->v_op != NULL, vp,
1102 		    ("vnlru_free: vnode already reclaimed."));
1103 		KASSERT((vp->v_iflag & VI_FREE) != 0,
1104 		    ("Removing vnode not on freelist"));
1105 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1106 		    ("Mangling active vnode"));
1107 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1108 
1109 		/*
1110 		 * Don't recycle if our vnode is from different type
1111 		 * of mount point.  Note that mp is type-safe, the
1112 		 * check does not reach unmapped address even if
1113 		 * vnode is reclaimed.
1114 		 * Don't recycle if we can't get the interlock without
1115 		 * blocking.
1116 		 */
1117 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1118 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1119 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1120 			continue;
1121 		}
1122 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1123 		    vp, ("vp inconsistent on freelist"));
1124 
1125 		/*
1126 		 * The clear of VI_FREE prevents activation of the
1127 		 * vnode.  There is no sense in putting the vnode on
1128 		 * the mount point active list, only to remove it
1129 		 * later during recycling.  Inline the relevant part
1130 		 * of vholdl(), to avoid triggering assertions or
1131 		 * activating.
1132 		 */
1133 		freevnodes--;
1134 		vp->v_iflag &= ~VI_FREE;
1135 		VNODE_REFCOUNT_FENCE_REL();
1136 		refcount_acquire(&vp->v_holdcnt);
1137 
1138 		mtx_unlock(&vnode_free_list_mtx);
1139 		VI_UNLOCK(vp);
1140 		vtryrecycle(vp);
1141 		/*
1142 		 * If the recycled succeeded this vdrop will actually free
1143 		 * the vnode.  If not it will simply place it back on
1144 		 * the free list.
1145 		 */
1146 		vdrop(vp);
1147 		mtx_lock(&vnode_free_list_mtx);
1148 	}
1149 }
1150 
1151 void
1152 vnlru_free(int count, struct vfsops *mnt_op)
1153 {
1154 
1155 	mtx_lock(&vnode_free_list_mtx);
1156 	vnlru_free_locked(count, mnt_op);
1157 	mtx_unlock(&vnode_free_list_mtx);
1158 }
1159 
1160 
1161 /* XXX some names and initialization are bad for limits and watermarks. */
1162 static int
1163 vspace(void)
1164 {
1165 	int space;
1166 
1167 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1168 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1169 	vlowat = vhiwat / 2;
1170 	if (numvnodes > desiredvnodes)
1171 		return (0);
1172 	space = desiredvnodes - numvnodes;
1173 	if (freevnodes > wantfreevnodes)
1174 		space += freevnodes - wantfreevnodes;
1175 	return (space);
1176 }
1177 
1178 static void
1179 vnlru_return_batch_locked(struct mount *mp)
1180 {
1181 	struct vnode *vp;
1182 
1183 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1184 
1185 	if (mp->mnt_tmpfreevnodelistsize == 0)
1186 		return;
1187 
1188 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1189 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1190 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1191 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1192 	}
1193 	mtx_lock(&vnode_free_list_mtx);
1194 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1195 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1196 	mtx_unlock(&vnode_free_list_mtx);
1197 	mp->mnt_tmpfreevnodelistsize = 0;
1198 }
1199 
1200 static void
1201 vnlru_return_batch(struct mount *mp)
1202 {
1203 
1204 	mtx_lock(&mp->mnt_listmtx);
1205 	vnlru_return_batch_locked(mp);
1206 	mtx_unlock(&mp->mnt_listmtx);
1207 }
1208 
1209 static void
1210 vnlru_return_batches(struct vfsops *mnt_op)
1211 {
1212 	struct mount *mp, *nmp;
1213 	bool need_unbusy;
1214 
1215 	mtx_lock(&mountlist_mtx);
1216 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1217 		need_unbusy = false;
1218 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1219 			goto next;
1220 		if (mp->mnt_tmpfreevnodelistsize == 0)
1221 			goto next;
1222 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1223 			vnlru_return_batch(mp);
1224 			need_unbusy = true;
1225 			mtx_lock(&mountlist_mtx);
1226 		}
1227 next:
1228 		nmp = TAILQ_NEXT(mp, mnt_list);
1229 		if (need_unbusy)
1230 			vfs_unbusy(mp);
1231 	}
1232 	mtx_unlock(&mountlist_mtx);
1233 }
1234 
1235 /*
1236  * Attempt to recycle vnodes in a context that is always safe to block.
1237  * Calling vlrurecycle() from the bowels of filesystem code has some
1238  * interesting deadlock problems.
1239  */
1240 static struct proc *vnlruproc;
1241 static int vnlruproc_sig;
1242 
1243 static void
1244 vnlru_proc(void)
1245 {
1246 	struct mount *mp, *nmp;
1247 	unsigned long onumvnodes;
1248 	int done, force, trigger, usevnodes;
1249 	bool reclaim_nc_src;
1250 
1251 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1252 	    SHUTDOWN_PRI_FIRST);
1253 
1254 	force = 0;
1255 	for (;;) {
1256 		kproc_suspend_check(vnlruproc);
1257 		mtx_lock(&vnode_free_list_mtx);
1258 		/*
1259 		 * If numvnodes is too large (due to desiredvnodes being
1260 		 * adjusted using its sysctl, or emergency growth), first
1261 		 * try to reduce it by discarding from the free list.
1262 		 */
1263 		if (numvnodes > desiredvnodes)
1264 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1265 		/*
1266 		 * Sleep if the vnode cache is in a good state.  This is
1267 		 * when it is not over-full and has space for about a 4%
1268 		 * or 9% expansion (by growing its size or inexcessively
1269 		 * reducing its free list).  Otherwise, try to reclaim
1270 		 * space for a 10% expansion.
1271 		 */
1272 		if (vstir && force == 0) {
1273 			force = 1;
1274 			vstir = 0;
1275 		}
1276 		if (vspace() >= vlowat && force == 0) {
1277 			vnlruproc_sig = 0;
1278 			wakeup(&vnlruproc_sig);
1279 			msleep(vnlruproc, &vnode_free_list_mtx,
1280 			    PVFS|PDROP, "vlruwt", hz);
1281 			continue;
1282 		}
1283 		mtx_unlock(&vnode_free_list_mtx);
1284 		done = 0;
1285 		onumvnodes = numvnodes;
1286 		/*
1287 		 * Calculate parameters for recycling.  These are the same
1288 		 * throughout the loop to give some semblance of fairness.
1289 		 * The trigger point is to avoid recycling vnodes with lots
1290 		 * of resident pages.  We aren't trying to free memory; we
1291 		 * are trying to recycle or at least free vnodes.
1292 		 */
1293 		if (numvnodes <= desiredvnodes)
1294 			usevnodes = numvnodes - freevnodes;
1295 		else
1296 			usevnodes = numvnodes;
1297 		if (usevnodes <= 0)
1298 			usevnodes = 1;
1299 		/*
1300 		 * The trigger value is is chosen to give a conservatively
1301 		 * large value to ensure that it alone doesn't prevent
1302 		 * making progress.  The value can easily be so large that
1303 		 * it is effectively infinite in some congested and
1304 		 * misconfigured cases, and this is necessary.  Normally
1305 		 * it is about 8 to 100 (pages), which is quite large.
1306 		 */
1307 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1308 		if (force < 2)
1309 			trigger = vsmalltrigger;
1310 		reclaim_nc_src = force >= 3;
1311 		mtx_lock(&mountlist_mtx);
1312 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1313 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1314 				nmp = TAILQ_NEXT(mp, mnt_list);
1315 				continue;
1316 			}
1317 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1318 			mtx_lock(&mountlist_mtx);
1319 			nmp = TAILQ_NEXT(mp, mnt_list);
1320 			vfs_unbusy(mp);
1321 		}
1322 		mtx_unlock(&mountlist_mtx);
1323 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1324 			uma_reclaim();
1325 		if (done == 0) {
1326 			if (force == 0 || force == 1) {
1327 				force = 2;
1328 				continue;
1329 			}
1330 			if (force == 2) {
1331 				force = 3;
1332 				continue;
1333 			}
1334 			force = 0;
1335 			vnlru_nowhere++;
1336 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1337 		} else
1338 			kern_yield(PRI_USER);
1339 		/*
1340 		 * After becoming active to expand above low water, keep
1341 		 * active until above high water.
1342 		 */
1343 		force = vspace() < vhiwat;
1344 	}
1345 }
1346 
1347 static struct kproc_desc vnlru_kp = {
1348 	"vnlru",
1349 	vnlru_proc,
1350 	&vnlruproc
1351 };
1352 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1353     &vnlru_kp);
1354 
1355 /*
1356  * Routines having to do with the management of the vnode table.
1357  */
1358 
1359 /*
1360  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1361  * before we actually vgone().  This function must be called with the vnode
1362  * held to prevent the vnode from being returned to the free list midway
1363  * through vgone().
1364  */
1365 static int
1366 vtryrecycle(struct vnode *vp)
1367 {
1368 	struct mount *vnmp;
1369 
1370 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1371 	VNASSERT(vp->v_holdcnt, vp,
1372 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1373 	/*
1374 	 * This vnode may found and locked via some other list, if so we
1375 	 * can't recycle it yet.
1376 	 */
1377 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1378 		CTR2(KTR_VFS,
1379 		    "%s: impossible to recycle, vp %p lock is already held",
1380 		    __func__, vp);
1381 		return (EWOULDBLOCK);
1382 	}
1383 	/*
1384 	 * Don't recycle if its filesystem is being suspended.
1385 	 */
1386 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1387 		VOP_UNLOCK(vp, 0);
1388 		CTR2(KTR_VFS,
1389 		    "%s: impossible to recycle, cannot start the write for %p",
1390 		    __func__, vp);
1391 		return (EBUSY);
1392 	}
1393 	/*
1394 	 * If we got this far, we need to acquire the interlock and see if
1395 	 * anyone picked up this vnode from another list.  If not, we will
1396 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1397 	 * will skip over it.
1398 	 */
1399 	VI_LOCK(vp);
1400 	if (vp->v_usecount) {
1401 		VOP_UNLOCK(vp, LK_INTERLOCK);
1402 		vn_finished_write(vnmp);
1403 		CTR2(KTR_VFS,
1404 		    "%s: impossible to recycle, %p is already referenced",
1405 		    __func__, vp);
1406 		return (EBUSY);
1407 	}
1408 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1409 		counter_u64_add(recycles_count, 1);
1410 		vgonel(vp);
1411 	}
1412 	VOP_UNLOCK(vp, LK_INTERLOCK);
1413 	vn_finished_write(vnmp);
1414 	return (0);
1415 }
1416 
1417 static void
1418 vcheckspace(void)
1419 {
1420 
1421 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1422 		vnlruproc_sig = 1;
1423 		wakeup(vnlruproc);
1424 	}
1425 }
1426 
1427 /*
1428  * Wait if necessary for space for a new vnode.
1429  */
1430 static int
1431 getnewvnode_wait(int suspended)
1432 {
1433 
1434 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1435 	if (numvnodes >= desiredvnodes) {
1436 		if (suspended) {
1437 			/*
1438 			 * The file system is being suspended.  We cannot
1439 			 * risk a deadlock here, so allow allocation of
1440 			 * another vnode even if this would give too many.
1441 			 */
1442 			return (0);
1443 		}
1444 		if (vnlruproc_sig == 0) {
1445 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1446 			wakeup(vnlruproc);
1447 		}
1448 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1449 		    "vlruwk", hz);
1450 	}
1451 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1452 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1453 		vnlru_free_locked(1, NULL);
1454 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1455 }
1456 
1457 /*
1458  * This hack is fragile, and probably not needed any more now that the
1459  * watermark handling works.
1460  */
1461 void
1462 getnewvnode_reserve(u_int count)
1463 {
1464 	struct thread *td;
1465 
1466 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1467 	/* XXX no longer so quick, but this part is not racy. */
1468 	mtx_lock(&vnode_free_list_mtx);
1469 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1470 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1471 		    freevnodes - wantfreevnodes), NULL);
1472 	mtx_unlock(&vnode_free_list_mtx);
1473 
1474 	td = curthread;
1475 	/* First try to be quick and racy. */
1476 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1477 		td->td_vp_reserv += count;
1478 		vcheckspace();	/* XXX no longer so quick, but more racy */
1479 		return;
1480 	} else
1481 		atomic_subtract_long(&numvnodes, count);
1482 
1483 	mtx_lock(&vnode_free_list_mtx);
1484 	while (count > 0) {
1485 		if (getnewvnode_wait(0) == 0) {
1486 			count--;
1487 			td->td_vp_reserv++;
1488 			atomic_add_long(&numvnodes, 1);
1489 		}
1490 	}
1491 	vcheckspace();
1492 	mtx_unlock(&vnode_free_list_mtx);
1493 }
1494 
1495 /*
1496  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1497  * misconfgured or changed significantly.  Reducing desiredvnodes below
1498  * the reserved amount should cause bizarre behaviour like reducing it
1499  * below the number of active vnodes -- the system will try to reduce
1500  * numvnodes to match, but should fail, so the subtraction below should
1501  * not overflow.
1502  */
1503 void
1504 getnewvnode_drop_reserve(void)
1505 {
1506 	struct thread *td;
1507 
1508 	td = curthread;
1509 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1510 	td->td_vp_reserv = 0;
1511 }
1512 
1513 /*
1514  * Return the next vnode from the free list.
1515  */
1516 int
1517 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1518     struct vnode **vpp)
1519 {
1520 	struct vnode *vp;
1521 	struct thread *td;
1522 	struct lock_object *lo;
1523 	static int cyclecount;
1524 	int error __unused;
1525 
1526 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1527 	vp = NULL;
1528 	td = curthread;
1529 	if (td->td_vp_reserv > 0) {
1530 		td->td_vp_reserv -= 1;
1531 		goto alloc;
1532 	}
1533 	mtx_lock(&vnode_free_list_mtx);
1534 	if (numvnodes < desiredvnodes)
1535 		cyclecount = 0;
1536 	else if (cyclecount++ >= freevnodes) {
1537 		cyclecount = 0;
1538 		vstir = 1;
1539 	}
1540 	/*
1541 	 * Grow the vnode cache if it will not be above its target max
1542 	 * after growing.  Otherwise, if the free list is nonempty, try
1543 	 * to reclaim 1 item from it before growing the cache (possibly
1544 	 * above its target max if the reclamation failed or is delayed).
1545 	 * Otherwise, wait for some space.  In all cases, schedule
1546 	 * vnlru_proc() if we are getting short of space.  The watermarks
1547 	 * should be chosen so that we never wait or even reclaim from
1548 	 * the free list to below its target minimum.
1549 	 */
1550 	if (numvnodes + 1 <= desiredvnodes)
1551 		;
1552 	else if (freevnodes > 0)
1553 		vnlru_free_locked(1, NULL);
1554 	else {
1555 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1556 		    MNTK_SUSPEND));
1557 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1558 		if (error != 0) {
1559 			mtx_unlock(&vnode_free_list_mtx);
1560 			return (error);
1561 		}
1562 #endif
1563 	}
1564 	vcheckspace();
1565 	atomic_add_long(&numvnodes, 1);
1566 	mtx_unlock(&vnode_free_list_mtx);
1567 alloc:
1568 	counter_u64_add(vnodes_created, 1);
1569 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1570 	/*
1571 	 * Locks are given the generic name "vnode" when created.
1572 	 * Follow the historic practice of using the filesystem
1573 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1574 	 *
1575 	 * Locks live in a witness group keyed on their name. Thus,
1576 	 * when a lock is renamed, it must also move from the witness
1577 	 * group of its old name to the witness group of its new name.
1578 	 *
1579 	 * The change only needs to be made when the vnode moves
1580 	 * from one filesystem type to another. We ensure that each
1581 	 * filesystem use a single static name pointer for its tag so
1582 	 * that we can compare pointers rather than doing a strcmp().
1583 	 */
1584 	lo = &vp->v_vnlock->lock_object;
1585 	if (lo->lo_name != tag) {
1586 		lo->lo_name = tag;
1587 		WITNESS_DESTROY(lo);
1588 		WITNESS_INIT(lo, tag);
1589 	}
1590 	/*
1591 	 * By default, don't allow shared locks unless filesystems opt-in.
1592 	 */
1593 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1594 	/*
1595 	 * Finalize various vnode identity bits.
1596 	 */
1597 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1598 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1599 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1600 	vp->v_type = VNON;
1601 	vp->v_tag = tag;
1602 	vp->v_op = vops;
1603 	v_init_counters(vp);
1604 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1605 #ifdef DIAGNOSTIC
1606 	if (mp == NULL && vops != &dead_vnodeops)
1607 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1608 #endif
1609 #ifdef MAC
1610 	mac_vnode_init(vp);
1611 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1612 		mac_vnode_associate_singlelabel(mp, vp);
1613 #endif
1614 	if (mp != NULL) {
1615 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1616 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1617 			vp->v_vflag |= VV_NOKNOTE;
1618 	}
1619 
1620 	/*
1621 	 * For the filesystems which do not use vfs_hash_insert(),
1622 	 * still initialize v_hash to have vfs_hash_index() useful.
1623 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1624 	 * its own hashing.
1625 	 */
1626 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1627 
1628 	*vpp = vp;
1629 	return (0);
1630 }
1631 
1632 /*
1633  * Delete from old mount point vnode list, if on one.
1634  */
1635 static void
1636 delmntque(struct vnode *vp)
1637 {
1638 	struct mount *mp;
1639 	int active;
1640 
1641 	mp = vp->v_mount;
1642 	if (mp == NULL)
1643 		return;
1644 	MNT_ILOCK(mp);
1645 	VI_LOCK(vp);
1646 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1647 	    ("Active vnode list size %d > Vnode list size %d",
1648 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1649 	active = vp->v_iflag & VI_ACTIVE;
1650 	vp->v_iflag &= ~VI_ACTIVE;
1651 	if (active) {
1652 		mtx_lock(&mp->mnt_listmtx);
1653 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1654 		mp->mnt_activevnodelistsize--;
1655 		mtx_unlock(&mp->mnt_listmtx);
1656 	}
1657 	vp->v_mount = NULL;
1658 	VI_UNLOCK(vp);
1659 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1660 		("bad mount point vnode list size"));
1661 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1662 	mp->mnt_nvnodelistsize--;
1663 	MNT_REL(mp);
1664 	MNT_IUNLOCK(mp);
1665 }
1666 
1667 static void
1668 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1669 {
1670 
1671 	vp->v_data = NULL;
1672 	vp->v_op = &dead_vnodeops;
1673 	vgone(vp);
1674 	vput(vp);
1675 }
1676 
1677 /*
1678  * Insert into list of vnodes for the new mount point, if available.
1679  */
1680 int
1681 insmntque1(struct vnode *vp, struct mount *mp,
1682 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1683 {
1684 
1685 	KASSERT(vp->v_mount == NULL,
1686 		("insmntque: vnode already on per mount vnode list"));
1687 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1688 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1689 
1690 	/*
1691 	 * We acquire the vnode interlock early to ensure that the
1692 	 * vnode cannot be recycled by another process releasing a
1693 	 * holdcnt on it before we get it on both the vnode list
1694 	 * and the active vnode list. The mount mutex protects only
1695 	 * manipulation of the vnode list and the vnode freelist
1696 	 * mutex protects only manipulation of the active vnode list.
1697 	 * Hence the need to hold the vnode interlock throughout.
1698 	 */
1699 	MNT_ILOCK(mp);
1700 	VI_LOCK(vp);
1701 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1702 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1703 	    mp->mnt_nvnodelistsize == 0)) &&
1704 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1705 		VI_UNLOCK(vp);
1706 		MNT_IUNLOCK(mp);
1707 		if (dtr != NULL)
1708 			dtr(vp, dtr_arg);
1709 		return (EBUSY);
1710 	}
1711 	vp->v_mount = mp;
1712 	MNT_REF(mp);
1713 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1714 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1715 		("neg mount point vnode list size"));
1716 	mp->mnt_nvnodelistsize++;
1717 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1718 	    ("Activating already active vnode"));
1719 	vp->v_iflag |= VI_ACTIVE;
1720 	mtx_lock(&mp->mnt_listmtx);
1721 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1722 	mp->mnt_activevnodelistsize++;
1723 	mtx_unlock(&mp->mnt_listmtx);
1724 	VI_UNLOCK(vp);
1725 	MNT_IUNLOCK(mp);
1726 	return (0);
1727 }
1728 
1729 int
1730 insmntque(struct vnode *vp, struct mount *mp)
1731 {
1732 
1733 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1734 }
1735 
1736 /*
1737  * Flush out and invalidate all buffers associated with a bufobj
1738  * Called with the underlying object locked.
1739  */
1740 int
1741 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1742 {
1743 	int error;
1744 
1745 	BO_LOCK(bo);
1746 	if (flags & V_SAVE) {
1747 		error = bufobj_wwait(bo, slpflag, slptimeo);
1748 		if (error) {
1749 			BO_UNLOCK(bo);
1750 			return (error);
1751 		}
1752 		if (bo->bo_dirty.bv_cnt > 0) {
1753 			BO_UNLOCK(bo);
1754 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1755 				return (error);
1756 			/*
1757 			 * XXX We could save a lock/unlock if this was only
1758 			 * enabled under INVARIANTS
1759 			 */
1760 			BO_LOCK(bo);
1761 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1762 				panic("vinvalbuf: dirty bufs");
1763 		}
1764 	}
1765 	/*
1766 	 * If you alter this loop please notice that interlock is dropped and
1767 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1768 	 * no race conditions occur from this.
1769 	 */
1770 	do {
1771 		error = flushbuflist(&bo->bo_clean,
1772 		    flags, bo, slpflag, slptimeo);
1773 		if (error == 0 && !(flags & V_CLEANONLY))
1774 			error = flushbuflist(&bo->bo_dirty,
1775 			    flags, bo, slpflag, slptimeo);
1776 		if (error != 0 && error != EAGAIN) {
1777 			BO_UNLOCK(bo);
1778 			return (error);
1779 		}
1780 	} while (error != 0);
1781 
1782 	/*
1783 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1784 	 * have write I/O in-progress but if there is a VM object then the
1785 	 * VM object can also have read-I/O in-progress.
1786 	 */
1787 	do {
1788 		bufobj_wwait(bo, 0, 0);
1789 		if ((flags & V_VMIO) == 0) {
1790 			BO_UNLOCK(bo);
1791 			if (bo->bo_object != NULL) {
1792 				VM_OBJECT_WLOCK(bo->bo_object);
1793 				vm_object_pip_wait(bo->bo_object, "bovlbx");
1794 				VM_OBJECT_WUNLOCK(bo->bo_object);
1795 			}
1796 			BO_LOCK(bo);
1797 		}
1798 	} while (bo->bo_numoutput > 0);
1799 	BO_UNLOCK(bo);
1800 
1801 	/*
1802 	 * Destroy the copy in the VM cache, too.
1803 	 */
1804 	if (bo->bo_object != NULL &&
1805 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1806 		VM_OBJECT_WLOCK(bo->bo_object);
1807 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1808 		    OBJPR_CLEANONLY : 0);
1809 		VM_OBJECT_WUNLOCK(bo->bo_object);
1810 	}
1811 
1812 #ifdef INVARIANTS
1813 	BO_LOCK(bo);
1814 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1815 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1816 	    bo->bo_clean.bv_cnt > 0))
1817 		panic("vinvalbuf: flush failed");
1818 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1819 	    bo->bo_dirty.bv_cnt > 0)
1820 		panic("vinvalbuf: flush dirty failed");
1821 	BO_UNLOCK(bo);
1822 #endif
1823 	return (0);
1824 }
1825 
1826 /*
1827  * Flush out and invalidate all buffers associated with a vnode.
1828  * Called with the underlying object locked.
1829  */
1830 int
1831 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1832 {
1833 
1834 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1835 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1836 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1837 		return (0);
1838 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1839 }
1840 
1841 /*
1842  * Flush out buffers on the specified list.
1843  *
1844  */
1845 static int
1846 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1847     int slptimeo)
1848 {
1849 	struct buf *bp, *nbp;
1850 	int retval, error;
1851 	daddr_t lblkno;
1852 	b_xflags_t xflags;
1853 
1854 	ASSERT_BO_WLOCKED(bo);
1855 
1856 	retval = 0;
1857 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1858 		/*
1859 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1860 		 * do not skip any buffers. If we are flushing only V_NORMAL
1861 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1862 		 * flushing only V_ALT buffers then skip buffers not marked
1863 		 * as BX_ALTDATA.
1864 		 */
1865 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1866 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1867 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1868 			continue;
1869 		}
1870 		if (nbp != NULL) {
1871 			lblkno = nbp->b_lblkno;
1872 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1873 		}
1874 		retval = EAGAIN;
1875 		error = BUF_TIMELOCK(bp,
1876 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1877 		    "flushbuf", slpflag, slptimeo);
1878 		if (error) {
1879 			BO_LOCK(bo);
1880 			return (error != ENOLCK ? error : EAGAIN);
1881 		}
1882 		KASSERT(bp->b_bufobj == bo,
1883 		    ("bp %p wrong b_bufobj %p should be %p",
1884 		    bp, bp->b_bufobj, bo));
1885 		/*
1886 		 * XXX Since there are no node locks for NFS, I
1887 		 * believe there is a slight chance that a delayed
1888 		 * write will occur while sleeping just above, so
1889 		 * check for it.
1890 		 */
1891 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1892 		    (flags & V_SAVE)) {
1893 			bremfree(bp);
1894 			bp->b_flags |= B_ASYNC;
1895 			bwrite(bp);
1896 			BO_LOCK(bo);
1897 			return (EAGAIN);	/* XXX: why not loop ? */
1898 		}
1899 		bremfree(bp);
1900 		bp->b_flags |= (B_INVAL | B_RELBUF);
1901 		bp->b_flags &= ~B_ASYNC;
1902 		brelse(bp);
1903 		BO_LOCK(bo);
1904 		if (nbp == NULL)
1905 			break;
1906 		nbp = gbincore(bo, lblkno);
1907 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1908 		    != xflags)
1909 			break;			/* nbp invalid */
1910 	}
1911 	return (retval);
1912 }
1913 
1914 int
1915 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1916 {
1917 	struct buf *bp;
1918 	int error;
1919 	daddr_t lblkno;
1920 
1921 	ASSERT_BO_LOCKED(bo);
1922 
1923 	for (lblkno = startn;;) {
1924 again:
1925 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1926 		if (bp == NULL || bp->b_lblkno >= endn ||
1927 		    bp->b_lblkno < startn)
1928 			break;
1929 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1930 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1931 		if (error != 0) {
1932 			BO_RLOCK(bo);
1933 			if (error == ENOLCK)
1934 				goto again;
1935 			return (error);
1936 		}
1937 		KASSERT(bp->b_bufobj == bo,
1938 		    ("bp %p wrong b_bufobj %p should be %p",
1939 		    bp, bp->b_bufobj, bo));
1940 		lblkno = bp->b_lblkno + 1;
1941 		if ((bp->b_flags & B_MANAGED) == 0)
1942 			bremfree(bp);
1943 		bp->b_flags |= B_RELBUF;
1944 		/*
1945 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1946 		 * pages backing each buffer in the range are unlikely to be
1947 		 * reused.  Dirty buffers will have the hint applied once
1948 		 * they've been written.
1949 		 */
1950 		if ((bp->b_flags & B_VMIO) != 0)
1951 			bp->b_flags |= B_NOREUSE;
1952 		brelse(bp);
1953 		BO_RLOCK(bo);
1954 	}
1955 	return (0);
1956 }
1957 
1958 /*
1959  * Truncate a file's buffer and pages to a specified length.  This
1960  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1961  * sync activity.
1962  */
1963 int
1964 vtruncbuf(struct vnode *vp, off_t length, int blksize)
1965 {
1966 	struct buf *bp, *nbp;
1967 	struct bufobj *bo;
1968 	daddr_t startlbn;
1969 
1970 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
1971 	    vp, blksize, (uintmax_t)length);
1972 
1973 	/*
1974 	 * Round up to the *next* lbn.
1975 	 */
1976 	startlbn = howmany(length, blksize);
1977 
1978 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1979 
1980 	bo = &vp->v_bufobj;
1981 restart_unlocked:
1982 	BO_LOCK(bo);
1983 
1984 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
1985 		;
1986 
1987 	if (length > 0) {
1988 restartsync:
1989 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1990 			if (bp->b_lblkno > 0)
1991 				continue;
1992 			/*
1993 			 * Since we hold the vnode lock this should only
1994 			 * fail if we're racing with the buf daemon.
1995 			 */
1996 			if (BUF_LOCK(bp,
1997 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1998 			    BO_LOCKPTR(bo)) == ENOLCK)
1999 				goto restart_unlocked;
2000 
2001 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2002 			    ("buf(%p) on dirty queue without DELWRI", bp));
2003 
2004 			bremfree(bp);
2005 			bawrite(bp);
2006 			BO_LOCK(bo);
2007 			goto restartsync;
2008 		}
2009 	}
2010 
2011 	bufobj_wwait(bo, 0, 0);
2012 	BO_UNLOCK(bo);
2013 	vnode_pager_setsize(vp, length);
2014 
2015 	return (0);
2016 }
2017 
2018 /*
2019  * Invalidate the cached pages of a file's buffer within the range of block
2020  * numbers [startlbn, endlbn).
2021  */
2022 void
2023 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2024     int blksize)
2025 {
2026 	struct bufobj *bo;
2027 	off_t start, end;
2028 
2029 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2030 
2031 	start = blksize * startlbn;
2032 	end = blksize * endlbn;
2033 
2034 	bo = &vp->v_bufobj;
2035 	BO_LOCK(bo);
2036 	MPASS(blksize == bo->bo_bsize);
2037 
2038 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2039 		;
2040 
2041 	BO_UNLOCK(bo);
2042 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2043 }
2044 
2045 static int
2046 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2047     daddr_t startlbn, daddr_t endlbn)
2048 {
2049 	struct buf *bp, *nbp;
2050 	bool anyfreed;
2051 
2052 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2053 	ASSERT_BO_LOCKED(bo);
2054 
2055 	do {
2056 		anyfreed = false;
2057 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2058 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2059 				continue;
2060 			if (BUF_LOCK(bp,
2061 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2062 			    BO_LOCKPTR(bo)) == ENOLCK) {
2063 				BO_LOCK(bo);
2064 				return (EAGAIN);
2065 			}
2066 
2067 			bremfree(bp);
2068 			bp->b_flags |= B_INVAL | B_RELBUF;
2069 			bp->b_flags &= ~B_ASYNC;
2070 			brelse(bp);
2071 			anyfreed = true;
2072 
2073 			BO_LOCK(bo);
2074 			if (nbp != NULL &&
2075 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2076 			    nbp->b_vp != vp ||
2077 			    (nbp->b_flags & B_DELWRI) != 0))
2078 				return (EAGAIN);
2079 		}
2080 
2081 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2082 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2083 				continue;
2084 			if (BUF_LOCK(bp,
2085 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2086 			    BO_LOCKPTR(bo)) == ENOLCK) {
2087 				BO_LOCK(bo);
2088 				return (EAGAIN);
2089 			}
2090 			bremfree(bp);
2091 			bp->b_flags |= B_INVAL | B_RELBUF;
2092 			bp->b_flags &= ~B_ASYNC;
2093 			brelse(bp);
2094 			anyfreed = true;
2095 
2096 			BO_LOCK(bo);
2097 			if (nbp != NULL &&
2098 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2099 			    (nbp->b_vp != vp) ||
2100 			    (nbp->b_flags & B_DELWRI) == 0))
2101 				return (EAGAIN);
2102 		}
2103 	} while (anyfreed);
2104 	return (0);
2105 }
2106 
2107 static void
2108 buf_vlist_remove(struct buf *bp)
2109 {
2110 	struct bufv *bv;
2111 
2112 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2113 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2114 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2115 	    (BX_VNDIRTY|BX_VNCLEAN),
2116 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2117 	if (bp->b_xflags & BX_VNDIRTY)
2118 		bv = &bp->b_bufobj->bo_dirty;
2119 	else
2120 		bv = &bp->b_bufobj->bo_clean;
2121 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2122 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2123 	bv->bv_cnt--;
2124 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2125 }
2126 
2127 /*
2128  * Add the buffer to the sorted clean or dirty block list.
2129  *
2130  * NOTE: xflags is passed as a constant, optimizing this inline function!
2131  */
2132 static void
2133 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2134 {
2135 	struct bufv *bv;
2136 	struct buf *n;
2137 	int error;
2138 
2139 	ASSERT_BO_WLOCKED(bo);
2140 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2141 	    ("dead bo %p", bo));
2142 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2143 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2144 	bp->b_xflags |= xflags;
2145 	if (xflags & BX_VNDIRTY)
2146 		bv = &bo->bo_dirty;
2147 	else
2148 		bv = &bo->bo_clean;
2149 
2150 	/*
2151 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2152 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2153 	 * than _ge.
2154 	 */
2155 	if (bv->bv_cnt == 0 ||
2156 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2157 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2158 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2159 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2160 	else
2161 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2162 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2163 	if (error)
2164 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2165 	bv->bv_cnt++;
2166 }
2167 
2168 /*
2169  * Look up a buffer using the buffer tries.
2170  */
2171 struct buf *
2172 gbincore(struct bufobj *bo, daddr_t lblkno)
2173 {
2174 	struct buf *bp;
2175 
2176 	ASSERT_BO_LOCKED(bo);
2177 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2178 	if (bp != NULL)
2179 		return (bp);
2180 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2181 }
2182 
2183 /*
2184  * Associate a buffer with a vnode.
2185  */
2186 void
2187 bgetvp(struct vnode *vp, struct buf *bp)
2188 {
2189 	struct bufobj *bo;
2190 
2191 	bo = &vp->v_bufobj;
2192 	ASSERT_BO_WLOCKED(bo);
2193 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2194 
2195 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2196 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2197 	    ("bgetvp: bp already attached! %p", bp));
2198 
2199 	vhold(vp);
2200 	bp->b_vp = vp;
2201 	bp->b_bufobj = bo;
2202 	/*
2203 	 * Insert onto list for new vnode.
2204 	 */
2205 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2206 }
2207 
2208 /*
2209  * Disassociate a buffer from a vnode.
2210  */
2211 void
2212 brelvp(struct buf *bp)
2213 {
2214 	struct bufobj *bo;
2215 	struct vnode *vp;
2216 
2217 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2218 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2219 
2220 	/*
2221 	 * Delete from old vnode list, if on one.
2222 	 */
2223 	vp = bp->b_vp;		/* XXX */
2224 	bo = bp->b_bufobj;
2225 	BO_LOCK(bo);
2226 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2227 		buf_vlist_remove(bp);
2228 	else
2229 		panic("brelvp: Buffer %p not on queue.", bp);
2230 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2231 		bo->bo_flag &= ~BO_ONWORKLST;
2232 		mtx_lock(&sync_mtx);
2233 		LIST_REMOVE(bo, bo_synclist);
2234 		syncer_worklist_len--;
2235 		mtx_unlock(&sync_mtx);
2236 	}
2237 	bp->b_vp = NULL;
2238 	bp->b_bufobj = NULL;
2239 	BO_UNLOCK(bo);
2240 	vdrop(vp);
2241 }
2242 
2243 /*
2244  * Add an item to the syncer work queue.
2245  */
2246 static void
2247 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2248 {
2249 	int slot;
2250 
2251 	ASSERT_BO_WLOCKED(bo);
2252 
2253 	mtx_lock(&sync_mtx);
2254 	if (bo->bo_flag & BO_ONWORKLST)
2255 		LIST_REMOVE(bo, bo_synclist);
2256 	else {
2257 		bo->bo_flag |= BO_ONWORKLST;
2258 		syncer_worklist_len++;
2259 	}
2260 
2261 	if (delay > syncer_maxdelay - 2)
2262 		delay = syncer_maxdelay - 2;
2263 	slot = (syncer_delayno + delay) & syncer_mask;
2264 
2265 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2266 	mtx_unlock(&sync_mtx);
2267 }
2268 
2269 static int
2270 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2271 {
2272 	int error, len;
2273 
2274 	mtx_lock(&sync_mtx);
2275 	len = syncer_worklist_len - sync_vnode_count;
2276 	mtx_unlock(&sync_mtx);
2277 	error = SYSCTL_OUT(req, &len, sizeof(len));
2278 	return (error);
2279 }
2280 
2281 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2282     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2283 
2284 static struct proc *updateproc;
2285 static void sched_sync(void);
2286 static struct kproc_desc up_kp = {
2287 	"syncer",
2288 	sched_sync,
2289 	&updateproc
2290 };
2291 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2292 
2293 static int
2294 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2295 {
2296 	struct vnode *vp;
2297 	struct mount *mp;
2298 
2299 	*bo = LIST_FIRST(slp);
2300 	if (*bo == NULL)
2301 		return (0);
2302 	vp = bo2vnode(*bo);
2303 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2304 		return (1);
2305 	/*
2306 	 * We use vhold in case the vnode does not
2307 	 * successfully sync.  vhold prevents the vnode from
2308 	 * going away when we unlock the sync_mtx so that
2309 	 * we can acquire the vnode interlock.
2310 	 */
2311 	vholdl(vp);
2312 	mtx_unlock(&sync_mtx);
2313 	VI_UNLOCK(vp);
2314 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2315 		vdrop(vp);
2316 		mtx_lock(&sync_mtx);
2317 		return (*bo == LIST_FIRST(slp));
2318 	}
2319 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2320 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2321 	VOP_UNLOCK(vp, 0);
2322 	vn_finished_write(mp);
2323 	BO_LOCK(*bo);
2324 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2325 		/*
2326 		 * Put us back on the worklist.  The worklist
2327 		 * routine will remove us from our current
2328 		 * position and then add us back in at a later
2329 		 * position.
2330 		 */
2331 		vn_syncer_add_to_worklist(*bo, syncdelay);
2332 	}
2333 	BO_UNLOCK(*bo);
2334 	vdrop(vp);
2335 	mtx_lock(&sync_mtx);
2336 	return (0);
2337 }
2338 
2339 static int first_printf = 1;
2340 
2341 /*
2342  * System filesystem synchronizer daemon.
2343  */
2344 static void
2345 sched_sync(void)
2346 {
2347 	struct synclist *next, *slp;
2348 	struct bufobj *bo;
2349 	long starttime;
2350 	struct thread *td = curthread;
2351 	int last_work_seen;
2352 	int net_worklist_len;
2353 	int syncer_final_iter;
2354 	int error;
2355 
2356 	last_work_seen = 0;
2357 	syncer_final_iter = 0;
2358 	syncer_state = SYNCER_RUNNING;
2359 	starttime = time_uptime;
2360 	td->td_pflags |= TDP_NORUNNINGBUF;
2361 
2362 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2363 	    SHUTDOWN_PRI_LAST);
2364 
2365 	mtx_lock(&sync_mtx);
2366 	for (;;) {
2367 		if (syncer_state == SYNCER_FINAL_DELAY &&
2368 		    syncer_final_iter == 0) {
2369 			mtx_unlock(&sync_mtx);
2370 			kproc_suspend_check(td->td_proc);
2371 			mtx_lock(&sync_mtx);
2372 		}
2373 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2374 		if (syncer_state != SYNCER_RUNNING &&
2375 		    starttime != time_uptime) {
2376 			if (first_printf) {
2377 				printf("\nSyncing disks, vnodes remaining... ");
2378 				first_printf = 0;
2379 			}
2380 			printf("%d ", net_worklist_len);
2381 		}
2382 		starttime = time_uptime;
2383 
2384 		/*
2385 		 * Push files whose dirty time has expired.  Be careful
2386 		 * of interrupt race on slp queue.
2387 		 *
2388 		 * Skip over empty worklist slots when shutting down.
2389 		 */
2390 		do {
2391 			slp = &syncer_workitem_pending[syncer_delayno];
2392 			syncer_delayno += 1;
2393 			if (syncer_delayno == syncer_maxdelay)
2394 				syncer_delayno = 0;
2395 			next = &syncer_workitem_pending[syncer_delayno];
2396 			/*
2397 			 * If the worklist has wrapped since the
2398 			 * it was emptied of all but syncer vnodes,
2399 			 * switch to the FINAL_DELAY state and run
2400 			 * for one more second.
2401 			 */
2402 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2403 			    net_worklist_len == 0 &&
2404 			    last_work_seen == syncer_delayno) {
2405 				syncer_state = SYNCER_FINAL_DELAY;
2406 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2407 			}
2408 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2409 		    syncer_worklist_len > 0);
2410 
2411 		/*
2412 		 * Keep track of the last time there was anything
2413 		 * on the worklist other than syncer vnodes.
2414 		 * Return to the SHUTTING_DOWN state if any
2415 		 * new work appears.
2416 		 */
2417 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2418 			last_work_seen = syncer_delayno;
2419 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2420 			syncer_state = SYNCER_SHUTTING_DOWN;
2421 		while (!LIST_EMPTY(slp)) {
2422 			error = sync_vnode(slp, &bo, td);
2423 			if (error == 1) {
2424 				LIST_REMOVE(bo, bo_synclist);
2425 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2426 				continue;
2427 			}
2428 
2429 			if (first_printf == 0) {
2430 				/*
2431 				 * Drop the sync mutex, because some watchdog
2432 				 * drivers need to sleep while patting
2433 				 */
2434 				mtx_unlock(&sync_mtx);
2435 				wdog_kern_pat(WD_LASTVAL);
2436 				mtx_lock(&sync_mtx);
2437 			}
2438 
2439 		}
2440 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2441 			syncer_final_iter--;
2442 		/*
2443 		 * The variable rushjob allows the kernel to speed up the
2444 		 * processing of the filesystem syncer process. A rushjob
2445 		 * value of N tells the filesystem syncer to process the next
2446 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2447 		 * is used by the soft update code to speed up the filesystem
2448 		 * syncer process when the incore state is getting so far
2449 		 * ahead of the disk that the kernel memory pool is being
2450 		 * threatened with exhaustion.
2451 		 */
2452 		if (rushjob > 0) {
2453 			rushjob -= 1;
2454 			continue;
2455 		}
2456 		/*
2457 		 * Just sleep for a short period of time between
2458 		 * iterations when shutting down to allow some I/O
2459 		 * to happen.
2460 		 *
2461 		 * If it has taken us less than a second to process the
2462 		 * current work, then wait. Otherwise start right over
2463 		 * again. We can still lose time if any single round
2464 		 * takes more than two seconds, but it does not really
2465 		 * matter as we are just trying to generally pace the
2466 		 * filesystem activity.
2467 		 */
2468 		if (syncer_state != SYNCER_RUNNING ||
2469 		    time_uptime == starttime) {
2470 			thread_lock(td);
2471 			sched_prio(td, PPAUSE);
2472 			thread_unlock(td);
2473 		}
2474 		if (syncer_state != SYNCER_RUNNING)
2475 			cv_timedwait(&sync_wakeup, &sync_mtx,
2476 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2477 		else if (time_uptime == starttime)
2478 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2479 	}
2480 }
2481 
2482 /*
2483  * Request the syncer daemon to speed up its work.
2484  * We never push it to speed up more than half of its
2485  * normal turn time, otherwise it could take over the cpu.
2486  */
2487 int
2488 speedup_syncer(void)
2489 {
2490 	int ret = 0;
2491 
2492 	mtx_lock(&sync_mtx);
2493 	if (rushjob < syncdelay / 2) {
2494 		rushjob += 1;
2495 		stat_rush_requests += 1;
2496 		ret = 1;
2497 	}
2498 	mtx_unlock(&sync_mtx);
2499 	cv_broadcast(&sync_wakeup);
2500 	return (ret);
2501 }
2502 
2503 /*
2504  * Tell the syncer to speed up its work and run though its work
2505  * list several times, then tell it to shut down.
2506  */
2507 static void
2508 syncer_shutdown(void *arg, int howto)
2509 {
2510 
2511 	if (howto & RB_NOSYNC)
2512 		return;
2513 	mtx_lock(&sync_mtx);
2514 	syncer_state = SYNCER_SHUTTING_DOWN;
2515 	rushjob = 0;
2516 	mtx_unlock(&sync_mtx);
2517 	cv_broadcast(&sync_wakeup);
2518 	kproc_shutdown(arg, howto);
2519 }
2520 
2521 void
2522 syncer_suspend(void)
2523 {
2524 
2525 	syncer_shutdown(updateproc, 0);
2526 }
2527 
2528 void
2529 syncer_resume(void)
2530 {
2531 
2532 	mtx_lock(&sync_mtx);
2533 	first_printf = 1;
2534 	syncer_state = SYNCER_RUNNING;
2535 	mtx_unlock(&sync_mtx);
2536 	cv_broadcast(&sync_wakeup);
2537 	kproc_resume(updateproc);
2538 }
2539 
2540 /*
2541  * Reassign a buffer from one vnode to another.
2542  * Used to assign file specific control information
2543  * (indirect blocks) to the vnode to which they belong.
2544  */
2545 void
2546 reassignbuf(struct buf *bp)
2547 {
2548 	struct vnode *vp;
2549 	struct bufobj *bo;
2550 	int delay;
2551 #ifdef INVARIANTS
2552 	struct bufv *bv;
2553 #endif
2554 
2555 	vp = bp->b_vp;
2556 	bo = bp->b_bufobj;
2557 	++reassignbufcalls;
2558 
2559 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2560 	    bp, bp->b_vp, bp->b_flags);
2561 	/*
2562 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2563 	 * is not fully linked in.
2564 	 */
2565 	if (bp->b_flags & B_PAGING)
2566 		panic("cannot reassign paging buffer");
2567 
2568 	/*
2569 	 * Delete from old vnode list, if on one.
2570 	 */
2571 	BO_LOCK(bo);
2572 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2573 		buf_vlist_remove(bp);
2574 	else
2575 		panic("reassignbuf: Buffer %p not on queue.", bp);
2576 	/*
2577 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2578 	 * of clean buffers.
2579 	 */
2580 	if (bp->b_flags & B_DELWRI) {
2581 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2582 			switch (vp->v_type) {
2583 			case VDIR:
2584 				delay = dirdelay;
2585 				break;
2586 			case VCHR:
2587 				delay = metadelay;
2588 				break;
2589 			default:
2590 				delay = filedelay;
2591 			}
2592 			vn_syncer_add_to_worklist(bo, delay);
2593 		}
2594 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2595 	} else {
2596 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2597 
2598 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2599 			mtx_lock(&sync_mtx);
2600 			LIST_REMOVE(bo, bo_synclist);
2601 			syncer_worklist_len--;
2602 			mtx_unlock(&sync_mtx);
2603 			bo->bo_flag &= ~BO_ONWORKLST;
2604 		}
2605 	}
2606 #ifdef INVARIANTS
2607 	bv = &bo->bo_clean;
2608 	bp = TAILQ_FIRST(&bv->bv_hd);
2609 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2610 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2611 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2612 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2613 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2614 	bv = &bo->bo_dirty;
2615 	bp = TAILQ_FIRST(&bv->bv_hd);
2616 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2617 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2618 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2619 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2620 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2621 #endif
2622 	BO_UNLOCK(bo);
2623 }
2624 
2625 static void
2626 v_init_counters(struct vnode *vp)
2627 {
2628 
2629 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2630 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2631 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2632 
2633 	refcount_init(&vp->v_holdcnt, 1);
2634 	refcount_init(&vp->v_usecount, 1);
2635 }
2636 
2637 static void
2638 v_incr_usecount_locked(struct vnode *vp)
2639 {
2640 
2641 	ASSERT_VI_LOCKED(vp, __func__);
2642 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2643 		VNASSERT(vp->v_usecount == 0, vp,
2644 		    ("vnode with usecount and VI_OWEINACT set"));
2645 		vp->v_iflag &= ~VI_OWEINACT;
2646 	}
2647 	refcount_acquire(&vp->v_usecount);
2648 	v_incr_devcount(vp);
2649 }
2650 
2651 /*
2652  * Increment the use count on the vnode, taking care to reference
2653  * the driver's usecount if this is a chardev.
2654  */
2655 static void
2656 v_incr_usecount(struct vnode *vp)
2657 {
2658 
2659 	ASSERT_VI_UNLOCKED(vp, __func__);
2660 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2661 
2662 	if (vp->v_type != VCHR &&
2663 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2664 		VNODE_REFCOUNT_FENCE_ACQ();
2665 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2666 		    ("vnode with usecount and VI_OWEINACT set"));
2667 	} else {
2668 		VI_LOCK(vp);
2669 		v_incr_usecount_locked(vp);
2670 		VI_UNLOCK(vp);
2671 	}
2672 }
2673 
2674 /*
2675  * Increment si_usecount of the associated device, if any.
2676  */
2677 static void
2678 v_incr_devcount(struct vnode *vp)
2679 {
2680 
2681 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2682 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2683 		dev_lock();
2684 		vp->v_rdev->si_usecount++;
2685 		dev_unlock();
2686 	}
2687 }
2688 
2689 /*
2690  * Decrement si_usecount of the associated device, if any.
2691  */
2692 static void
2693 v_decr_devcount(struct vnode *vp)
2694 {
2695 
2696 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2697 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2698 		dev_lock();
2699 		vp->v_rdev->si_usecount--;
2700 		dev_unlock();
2701 	}
2702 }
2703 
2704 /*
2705  * Grab a particular vnode from the free list, increment its
2706  * reference count and lock it.  VI_DOOMED is set if the vnode
2707  * is being destroyed.  Only callers who specify LK_RETRY will
2708  * see doomed vnodes.  If inactive processing was delayed in
2709  * vput try to do it here.
2710  *
2711  * Notes on lockless counter manipulation:
2712  * _vhold, vputx and other routines make various decisions based
2713  * on either holdcnt or usecount being 0. As long as either counter
2714  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2715  * with atomic operations. Otherwise the interlock is taken covering
2716  * both the atomic and additional actions.
2717  */
2718 int
2719 vget(struct vnode *vp, int flags, struct thread *td)
2720 {
2721 	int error, oweinact;
2722 
2723 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2724 	    ("vget: invalid lock operation"));
2725 
2726 	if ((flags & LK_INTERLOCK) != 0)
2727 		ASSERT_VI_LOCKED(vp, __func__);
2728 	else
2729 		ASSERT_VI_UNLOCKED(vp, __func__);
2730 	if ((flags & LK_VNHELD) != 0)
2731 		VNASSERT((vp->v_holdcnt > 0), vp,
2732 		    ("vget: LK_VNHELD passed but vnode not held"));
2733 
2734 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2735 
2736 	if ((flags & LK_VNHELD) == 0)
2737 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2738 
2739 	if ((error = vn_lock(vp, flags)) != 0) {
2740 		vdrop(vp);
2741 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2742 		    vp);
2743 		return (error);
2744 	}
2745 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2746 		panic("vget: vn_lock failed to return ENOENT\n");
2747 	/*
2748 	 * We don't guarantee that any particular close will
2749 	 * trigger inactive processing so just make a best effort
2750 	 * here at preventing a reference to a removed file.  If
2751 	 * we don't succeed no harm is done.
2752 	 *
2753 	 * Upgrade our holdcnt to a usecount.
2754 	 */
2755 	if (vp->v_type == VCHR ||
2756 	    !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2757 		VI_LOCK(vp);
2758 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2759 			oweinact = 0;
2760 		} else {
2761 			oweinact = 1;
2762 			vp->v_iflag &= ~VI_OWEINACT;
2763 			VNODE_REFCOUNT_FENCE_REL();
2764 		}
2765 		refcount_acquire(&vp->v_usecount);
2766 		v_incr_devcount(vp);
2767 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2768 		    (flags & LK_NOWAIT) == 0)
2769 			vinactive(vp, td);
2770 		VI_UNLOCK(vp);
2771 	}
2772 	return (0);
2773 }
2774 
2775 /*
2776  * Increase the reference (use) and hold count of a vnode.
2777  * This will also remove the vnode from the free list if it is presently free.
2778  */
2779 void
2780 vref(struct vnode *vp)
2781 {
2782 
2783 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2784 	_vhold(vp, false);
2785 	v_incr_usecount(vp);
2786 }
2787 
2788 void
2789 vrefl(struct vnode *vp)
2790 {
2791 
2792 	ASSERT_VI_LOCKED(vp, __func__);
2793 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2794 	_vhold(vp, true);
2795 	v_incr_usecount_locked(vp);
2796 }
2797 
2798 void
2799 vrefact(struct vnode *vp)
2800 {
2801 
2802 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2803 	if (__predict_false(vp->v_type == VCHR)) {
2804 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2805 		    ("%s: wrong ref counts", __func__));
2806 		vref(vp);
2807 		return;
2808 	}
2809 #ifdef INVARIANTS
2810 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2811 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2812 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2813 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2814 #else
2815 	refcount_acquire(&vp->v_holdcnt);
2816 	refcount_acquire(&vp->v_usecount);
2817 #endif
2818 }
2819 
2820 /*
2821  * Return reference count of a vnode.
2822  *
2823  * The results of this call are only guaranteed when some mechanism is used to
2824  * stop other processes from gaining references to the vnode.  This may be the
2825  * case if the caller holds the only reference.  This is also useful when stale
2826  * data is acceptable as race conditions may be accounted for by some other
2827  * means.
2828  */
2829 int
2830 vrefcnt(struct vnode *vp)
2831 {
2832 
2833 	return (vp->v_usecount);
2834 }
2835 
2836 #define	VPUTX_VRELE	1
2837 #define	VPUTX_VPUT	2
2838 #define	VPUTX_VUNREF	3
2839 
2840 /*
2841  * Decrement the use and hold counts for a vnode.
2842  *
2843  * See an explanation near vget() as to why atomic operation is safe.
2844  */
2845 static void
2846 vputx(struct vnode *vp, int func)
2847 {
2848 	int error;
2849 
2850 	KASSERT(vp != NULL, ("vputx: null vp"));
2851 	if (func == VPUTX_VUNREF)
2852 		ASSERT_VOP_LOCKED(vp, "vunref");
2853 	else if (func == VPUTX_VPUT)
2854 		ASSERT_VOP_LOCKED(vp, "vput");
2855 	else
2856 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2857 	ASSERT_VI_UNLOCKED(vp, __func__);
2858 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2859 
2860 	if (vp->v_type != VCHR &&
2861 	    refcount_release_if_not_last(&vp->v_usecount)) {
2862 		if (func == VPUTX_VPUT)
2863 			VOP_UNLOCK(vp, 0);
2864 		vdrop(vp);
2865 		return;
2866 	}
2867 
2868 	VI_LOCK(vp);
2869 
2870 	/*
2871 	 * We want to hold the vnode until the inactive finishes to
2872 	 * prevent vgone() races.  We drop the use count here and the
2873 	 * hold count below when we're done.
2874 	 */
2875 	if (!refcount_release(&vp->v_usecount) ||
2876 	    (vp->v_iflag & VI_DOINGINACT)) {
2877 		if (func == VPUTX_VPUT)
2878 			VOP_UNLOCK(vp, 0);
2879 		v_decr_devcount(vp);
2880 		vdropl(vp);
2881 		return;
2882 	}
2883 
2884 	v_decr_devcount(vp);
2885 
2886 	error = 0;
2887 
2888 	if (vp->v_usecount != 0) {
2889 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2890 		panic("vputx: usecount not zero");
2891 	}
2892 
2893 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2894 
2895 	/*
2896 	 * We must call VOP_INACTIVE with the node locked. Mark
2897 	 * as VI_DOINGINACT to avoid recursion.
2898 	 */
2899 	vp->v_iflag |= VI_OWEINACT;
2900 	switch (func) {
2901 	case VPUTX_VRELE:
2902 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2903 		VI_LOCK(vp);
2904 		break;
2905 	case VPUTX_VPUT:
2906 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2907 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2908 			    LK_NOWAIT);
2909 			VI_LOCK(vp);
2910 		}
2911 		break;
2912 	case VPUTX_VUNREF:
2913 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2914 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2915 			VI_LOCK(vp);
2916 		}
2917 		break;
2918 	}
2919 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2920 	    ("vnode with usecount and VI_OWEINACT set"));
2921 	if (error == 0) {
2922 		if (vp->v_iflag & VI_OWEINACT)
2923 			vinactive(vp, curthread);
2924 		if (func != VPUTX_VUNREF)
2925 			VOP_UNLOCK(vp, 0);
2926 	}
2927 	vdropl(vp);
2928 }
2929 
2930 /*
2931  * Vnode put/release.
2932  * If count drops to zero, call inactive routine and return to freelist.
2933  */
2934 void
2935 vrele(struct vnode *vp)
2936 {
2937 
2938 	vputx(vp, VPUTX_VRELE);
2939 }
2940 
2941 /*
2942  * Release an already locked vnode.  This give the same effects as
2943  * unlock+vrele(), but takes less time and avoids releasing and
2944  * re-aquiring the lock (as vrele() acquires the lock internally.)
2945  */
2946 void
2947 vput(struct vnode *vp)
2948 {
2949 
2950 	vputx(vp, VPUTX_VPUT);
2951 }
2952 
2953 /*
2954  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2955  */
2956 void
2957 vunref(struct vnode *vp)
2958 {
2959 
2960 	vputx(vp, VPUTX_VUNREF);
2961 }
2962 
2963 /*
2964  * Increase the hold count and activate if this is the first reference.
2965  */
2966 void
2967 _vhold(struct vnode *vp, bool locked)
2968 {
2969 	struct mount *mp;
2970 
2971 	if (locked)
2972 		ASSERT_VI_LOCKED(vp, __func__);
2973 	else
2974 		ASSERT_VI_UNLOCKED(vp, __func__);
2975 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2976 	if (!locked) {
2977 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2978 			VNODE_REFCOUNT_FENCE_ACQ();
2979 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2980 			    ("_vhold: vnode with holdcnt is free"));
2981 			return;
2982 		}
2983 		VI_LOCK(vp);
2984 	}
2985 	if ((vp->v_iflag & VI_FREE) == 0) {
2986 		refcount_acquire(&vp->v_holdcnt);
2987 		if (!locked)
2988 			VI_UNLOCK(vp);
2989 		return;
2990 	}
2991 	VNASSERT(vp->v_holdcnt == 0, vp,
2992 	    ("%s: wrong hold count", __func__));
2993 	VNASSERT(vp->v_op != NULL, vp,
2994 	    ("%s: vnode already reclaimed.", __func__));
2995 	/*
2996 	 * Remove a vnode from the free list, mark it as in use,
2997 	 * and put it on the active list.
2998 	 */
2999 	VNASSERT(vp->v_mount != NULL, vp,
3000 	    ("_vhold: vnode not on per mount vnode list"));
3001 	mp = vp->v_mount;
3002 	mtx_lock(&mp->mnt_listmtx);
3003 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
3004 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
3005 		mp->mnt_tmpfreevnodelistsize--;
3006 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
3007 	} else {
3008 		mtx_lock(&vnode_free_list_mtx);
3009 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3010 		freevnodes--;
3011 		mtx_unlock(&vnode_free_list_mtx);
3012 	}
3013 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3014 	    ("Activating already active vnode"));
3015 	vp->v_iflag &= ~VI_FREE;
3016 	vp->v_iflag |= VI_ACTIVE;
3017 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3018 	mp->mnt_activevnodelistsize++;
3019 	mtx_unlock(&mp->mnt_listmtx);
3020 	refcount_acquire(&vp->v_holdcnt);
3021 	if (!locked)
3022 		VI_UNLOCK(vp);
3023 }
3024 
3025 /*
3026  * Drop the hold count of the vnode.  If this is the last reference to
3027  * the vnode we place it on the free list unless it has been vgone'd
3028  * (marked VI_DOOMED) in which case we will free it.
3029  *
3030  * Because the vnode vm object keeps a hold reference on the vnode if
3031  * there is at least one resident non-cached page, the vnode cannot
3032  * leave the active list without the page cleanup done.
3033  */
3034 void
3035 _vdrop(struct vnode *vp, bool locked)
3036 {
3037 	struct bufobj *bo;
3038 	struct mount *mp;
3039 	int active;
3040 
3041 	if (locked)
3042 		ASSERT_VI_LOCKED(vp, __func__);
3043 	else
3044 		ASSERT_VI_UNLOCKED(vp, __func__);
3045 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3046 	if ((int)vp->v_holdcnt <= 0)
3047 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
3048 	if (!locked) {
3049 		if (refcount_release_if_not_last(&vp->v_holdcnt))
3050 			return;
3051 		VI_LOCK(vp);
3052 	}
3053 	if (refcount_release(&vp->v_holdcnt) == 0) {
3054 		VI_UNLOCK(vp);
3055 		return;
3056 	}
3057 	if ((vp->v_iflag & VI_DOOMED) == 0) {
3058 		/*
3059 		 * Mark a vnode as free: remove it from its active list
3060 		 * and put it up for recycling on the freelist.
3061 		 */
3062 		VNASSERT(vp->v_op != NULL, vp,
3063 		    ("vdropl: vnode already reclaimed."));
3064 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3065 		    ("vnode already free"));
3066 		VNASSERT(vp->v_holdcnt == 0, vp,
3067 		    ("vdropl: freeing when we shouldn't"));
3068 		active = vp->v_iflag & VI_ACTIVE;
3069 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
3070 			vp->v_iflag &= ~VI_ACTIVE;
3071 			mp = vp->v_mount;
3072 			if (mp != NULL) {
3073 				mtx_lock(&mp->mnt_listmtx);
3074 				if (active) {
3075 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
3076 					    vp, v_actfreelist);
3077 					mp->mnt_activevnodelistsize--;
3078 				}
3079 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3080 				    vp, v_actfreelist);
3081 				mp->mnt_tmpfreevnodelistsize++;
3082 				vp->v_iflag |= VI_FREE;
3083 				vp->v_mflag |= VMP_TMPMNTFREELIST;
3084 				VI_UNLOCK(vp);
3085 				if (mp->mnt_tmpfreevnodelistsize >=
3086 				    mnt_free_list_batch)
3087 					vnlru_return_batch_locked(mp);
3088 				mtx_unlock(&mp->mnt_listmtx);
3089 			} else {
3090 				VNASSERT(active == 0, vp,
3091 				    ("vdropl: active vnode not on per mount "
3092 				    "vnode list"));
3093 				mtx_lock(&vnode_free_list_mtx);
3094 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3095 				    v_actfreelist);
3096 				freevnodes++;
3097 				vp->v_iflag |= VI_FREE;
3098 				VI_UNLOCK(vp);
3099 				mtx_unlock(&vnode_free_list_mtx);
3100 			}
3101 		} else {
3102 			VI_UNLOCK(vp);
3103 			counter_u64_add(free_owe_inact, 1);
3104 		}
3105 		return;
3106 	}
3107 	/*
3108 	 * The vnode has been marked for destruction, so free it.
3109 	 *
3110 	 * The vnode will be returned to the zone where it will
3111 	 * normally remain until it is needed for another vnode. We
3112 	 * need to cleanup (or verify that the cleanup has already
3113 	 * been done) any residual data left from its current use
3114 	 * so as not to contaminate the freshly allocated vnode.
3115 	 */
3116 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
3117 	atomic_subtract_long(&numvnodes, 1);
3118 	bo = &vp->v_bufobj;
3119 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3120 	    ("cleaned vnode still on the free list."));
3121 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
3122 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
3123 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
3124 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
3125 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
3126 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
3127 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
3128 	    ("clean blk trie not empty"));
3129 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3130 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3131 	    ("dirty blk trie not empty"));
3132 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3133 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3134 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3135 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3136 	    ("Dangling rangelock waiters"));
3137 	VI_UNLOCK(vp);
3138 #ifdef MAC
3139 	mac_vnode_destroy(vp);
3140 #endif
3141 	if (vp->v_pollinfo != NULL) {
3142 		destroy_vpollinfo(vp->v_pollinfo);
3143 		vp->v_pollinfo = NULL;
3144 	}
3145 #ifdef INVARIANTS
3146 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3147 	vp->v_op = NULL;
3148 #endif
3149 	vp->v_mountedhere = NULL;
3150 	vp->v_unpcb = NULL;
3151 	vp->v_rdev = NULL;
3152 	vp->v_fifoinfo = NULL;
3153 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3154 	vp->v_iflag = 0;
3155 	vp->v_vflag = 0;
3156 	bo->bo_flag = 0;
3157 	uma_zfree(vnode_zone, vp);
3158 }
3159 
3160 /*
3161  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3162  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3163  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3164  * failed lock upgrade.
3165  */
3166 void
3167 vinactive(struct vnode *vp, struct thread *td)
3168 {
3169 	struct vm_object *obj;
3170 
3171 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3172 	ASSERT_VI_LOCKED(vp, "vinactive");
3173 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3174 	    ("vinactive: recursed on VI_DOINGINACT"));
3175 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3176 	vp->v_iflag |= VI_DOINGINACT;
3177 	vp->v_iflag &= ~VI_OWEINACT;
3178 	VI_UNLOCK(vp);
3179 	/*
3180 	 * Before moving off the active list, we must be sure that any
3181 	 * modified pages are converted into the vnode's dirty
3182 	 * buffers, since these will no longer be checked once the
3183 	 * vnode is on the inactive list.
3184 	 *
3185 	 * The write-out of the dirty pages is asynchronous.  At the
3186 	 * point that VOP_INACTIVE() is called, there could still be
3187 	 * pending I/O and dirty pages in the object.
3188 	 */
3189 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3190 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3191 		VM_OBJECT_WLOCK(obj);
3192 		vm_object_page_clean(obj, 0, 0, 0);
3193 		VM_OBJECT_WUNLOCK(obj);
3194 	}
3195 	VOP_INACTIVE(vp, td);
3196 	VI_LOCK(vp);
3197 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3198 	    ("vinactive: lost VI_DOINGINACT"));
3199 	vp->v_iflag &= ~VI_DOINGINACT;
3200 }
3201 
3202 /*
3203  * Remove any vnodes in the vnode table belonging to mount point mp.
3204  *
3205  * If FORCECLOSE is not specified, there should not be any active ones,
3206  * return error if any are found (nb: this is a user error, not a
3207  * system error). If FORCECLOSE is specified, detach any active vnodes
3208  * that are found.
3209  *
3210  * If WRITECLOSE is set, only flush out regular file vnodes open for
3211  * writing.
3212  *
3213  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3214  *
3215  * `rootrefs' specifies the base reference count for the root vnode
3216  * of this filesystem. The root vnode is considered busy if its
3217  * v_usecount exceeds this value. On a successful return, vflush(, td)
3218  * will call vrele() on the root vnode exactly rootrefs times.
3219  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3220  * be zero.
3221  */
3222 #ifdef DIAGNOSTIC
3223 static int busyprt = 0;		/* print out busy vnodes */
3224 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3225 #endif
3226 
3227 int
3228 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3229 {
3230 	struct vnode *vp, *mvp, *rootvp = NULL;
3231 	struct vattr vattr;
3232 	int busy = 0, error;
3233 
3234 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3235 	    rootrefs, flags);
3236 	if (rootrefs > 0) {
3237 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3238 		    ("vflush: bad args"));
3239 		/*
3240 		 * Get the filesystem root vnode. We can vput() it
3241 		 * immediately, since with rootrefs > 0, it won't go away.
3242 		 */
3243 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3244 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3245 			    __func__, error);
3246 			return (error);
3247 		}
3248 		vput(rootvp);
3249 	}
3250 loop:
3251 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3252 		vholdl(vp);
3253 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3254 		if (error) {
3255 			vdrop(vp);
3256 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3257 			goto loop;
3258 		}
3259 		/*
3260 		 * Skip over a vnodes marked VV_SYSTEM.
3261 		 */
3262 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3263 			VOP_UNLOCK(vp, 0);
3264 			vdrop(vp);
3265 			continue;
3266 		}
3267 		/*
3268 		 * If WRITECLOSE is set, flush out unlinked but still open
3269 		 * files (even if open only for reading) and regular file
3270 		 * vnodes open for writing.
3271 		 */
3272 		if (flags & WRITECLOSE) {
3273 			if (vp->v_object != NULL) {
3274 				VM_OBJECT_WLOCK(vp->v_object);
3275 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3276 				VM_OBJECT_WUNLOCK(vp->v_object);
3277 			}
3278 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3279 			if (error != 0) {
3280 				VOP_UNLOCK(vp, 0);
3281 				vdrop(vp);
3282 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3283 				return (error);
3284 			}
3285 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3286 			VI_LOCK(vp);
3287 
3288 			if ((vp->v_type == VNON ||
3289 			    (error == 0 && vattr.va_nlink > 0)) &&
3290 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3291 				VOP_UNLOCK(vp, 0);
3292 				vdropl(vp);
3293 				continue;
3294 			}
3295 		} else
3296 			VI_LOCK(vp);
3297 		/*
3298 		 * With v_usecount == 0, all we need to do is clear out the
3299 		 * vnode data structures and we are done.
3300 		 *
3301 		 * If FORCECLOSE is set, forcibly close the vnode.
3302 		 */
3303 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3304 			vgonel(vp);
3305 		} else {
3306 			busy++;
3307 #ifdef DIAGNOSTIC
3308 			if (busyprt)
3309 				vn_printf(vp, "vflush: busy vnode ");
3310 #endif
3311 		}
3312 		VOP_UNLOCK(vp, 0);
3313 		vdropl(vp);
3314 	}
3315 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3316 		/*
3317 		 * If just the root vnode is busy, and if its refcount
3318 		 * is equal to `rootrefs', then go ahead and kill it.
3319 		 */
3320 		VI_LOCK(rootvp);
3321 		KASSERT(busy > 0, ("vflush: not busy"));
3322 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3323 		    ("vflush: usecount %d < rootrefs %d",
3324 		     rootvp->v_usecount, rootrefs));
3325 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3326 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3327 			vgone(rootvp);
3328 			VOP_UNLOCK(rootvp, 0);
3329 			busy = 0;
3330 		} else
3331 			VI_UNLOCK(rootvp);
3332 	}
3333 	if (busy) {
3334 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3335 		    busy);
3336 		return (EBUSY);
3337 	}
3338 	for (; rootrefs > 0; rootrefs--)
3339 		vrele(rootvp);
3340 	return (0);
3341 }
3342 
3343 /*
3344  * Recycle an unused vnode to the front of the free list.
3345  */
3346 int
3347 vrecycle(struct vnode *vp)
3348 {
3349 	int recycled;
3350 
3351 	VI_LOCK(vp);
3352 	recycled = vrecyclel(vp);
3353 	VI_UNLOCK(vp);
3354 	return (recycled);
3355 }
3356 
3357 /*
3358  * vrecycle, with the vp interlock held.
3359  */
3360 int
3361 vrecyclel(struct vnode *vp)
3362 {
3363 	int recycled;
3364 
3365 	ASSERT_VOP_ELOCKED(vp, __func__);
3366 	ASSERT_VI_LOCKED(vp, __func__);
3367 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3368 	recycled = 0;
3369 	if (vp->v_usecount == 0) {
3370 		recycled = 1;
3371 		vgonel(vp);
3372 	}
3373 	return (recycled);
3374 }
3375 
3376 /*
3377  * Eliminate all activity associated with a vnode
3378  * in preparation for reuse.
3379  */
3380 void
3381 vgone(struct vnode *vp)
3382 {
3383 	VI_LOCK(vp);
3384 	vgonel(vp);
3385 	VI_UNLOCK(vp);
3386 }
3387 
3388 static void
3389 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3390     struct vnode *lowervp __unused)
3391 {
3392 }
3393 
3394 /*
3395  * Notify upper mounts about reclaimed or unlinked vnode.
3396  */
3397 void
3398 vfs_notify_upper(struct vnode *vp, int event)
3399 {
3400 	static struct vfsops vgonel_vfsops = {
3401 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3402 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3403 	};
3404 	struct mount *mp, *ump, *mmp;
3405 
3406 	mp = vp->v_mount;
3407 	if (mp == NULL)
3408 		return;
3409 
3410 	MNT_ILOCK(mp);
3411 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3412 		goto unlock;
3413 	MNT_IUNLOCK(mp);
3414 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3415 	mmp->mnt_op = &vgonel_vfsops;
3416 	mmp->mnt_kern_flag |= MNTK_MARKER;
3417 	MNT_ILOCK(mp);
3418 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3419 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3420 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3421 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3422 			continue;
3423 		}
3424 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3425 		MNT_IUNLOCK(mp);
3426 		switch (event) {
3427 		case VFS_NOTIFY_UPPER_RECLAIM:
3428 			VFS_RECLAIM_LOWERVP(ump, vp);
3429 			break;
3430 		case VFS_NOTIFY_UPPER_UNLINK:
3431 			VFS_UNLINK_LOWERVP(ump, vp);
3432 			break;
3433 		default:
3434 			KASSERT(0, ("invalid event %d", event));
3435 			break;
3436 		}
3437 		MNT_ILOCK(mp);
3438 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3439 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3440 	}
3441 	free(mmp, M_TEMP);
3442 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3443 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3444 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3445 		wakeup(&mp->mnt_uppers);
3446 	}
3447 unlock:
3448 	MNT_IUNLOCK(mp);
3449 }
3450 
3451 /*
3452  * vgone, with the vp interlock held.
3453  */
3454 static void
3455 vgonel(struct vnode *vp)
3456 {
3457 	struct thread *td;
3458 	int oweinact;
3459 	int active;
3460 	struct mount *mp;
3461 
3462 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3463 	ASSERT_VI_LOCKED(vp, "vgonel");
3464 	VNASSERT(vp->v_holdcnt, vp,
3465 	    ("vgonel: vp %p has no reference.", vp));
3466 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3467 	td = curthread;
3468 
3469 	/*
3470 	 * Don't vgonel if we're already doomed.
3471 	 */
3472 	if (vp->v_iflag & VI_DOOMED)
3473 		return;
3474 	vp->v_iflag |= VI_DOOMED;
3475 
3476 	/*
3477 	 * Check to see if the vnode is in use.  If so, we have to call
3478 	 * VOP_CLOSE() and VOP_INACTIVE().
3479 	 */
3480 	active = vp->v_usecount;
3481 	oweinact = (vp->v_iflag & VI_OWEINACT);
3482 	VI_UNLOCK(vp);
3483 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3484 
3485 	/*
3486 	 * If purging an active vnode, it must be closed and
3487 	 * deactivated before being reclaimed.
3488 	 */
3489 	if (active)
3490 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3491 	if (oweinact || active) {
3492 		VI_LOCK(vp);
3493 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3494 			vinactive(vp, td);
3495 		VI_UNLOCK(vp);
3496 	}
3497 	if (vp->v_type == VSOCK)
3498 		vfs_unp_reclaim(vp);
3499 
3500 	/*
3501 	 * Clean out any buffers associated with the vnode.
3502 	 * If the flush fails, just toss the buffers.
3503 	 */
3504 	mp = NULL;
3505 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3506 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3507 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3508 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3509 			;
3510 	}
3511 
3512 	BO_LOCK(&vp->v_bufobj);
3513 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3514 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3515 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3516 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3517 	    ("vp %p bufobj not invalidated", vp));
3518 
3519 	/*
3520 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3521 	 * after the object's page queue is flushed.
3522 	 */
3523 	if (vp->v_bufobj.bo_object == NULL)
3524 		vp->v_bufobj.bo_flag |= BO_DEAD;
3525 	BO_UNLOCK(&vp->v_bufobj);
3526 
3527 	/*
3528 	 * Reclaim the vnode.
3529 	 */
3530 	if (VOP_RECLAIM(vp, td))
3531 		panic("vgone: cannot reclaim");
3532 	if (mp != NULL)
3533 		vn_finished_secondary_write(mp);
3534 	VNASSERT(vp->v_object == NULL, vp,
3535 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3536 	/*
3537 	 * Clear the advisory locks and wake up waiting threads.
3538 	 */
3539 	(void)VOP_ADVLOCKPURGE(vp);
3540 	vp->v_lockf = NULL;
3541 	/*
3542 	 * Delete from old mount point vnode list.
3543 	 */
3544 	delmntque(vp);
3545 	cache_purge(vp);
3546 	/*
3547 	 * Done with purge, reset to the standard lock and invalidate
3548 	 * the vnode.
3549 	 */
3550 	VI_LOCK(vp);
3551 	vp->v_vnlock = &vp->v_lock;
3552 	vp->v_op = &dead_vnodeops;
3553 	vp->v_tag = "none";
3554 	vp->v_type = VBAD;
3555 }
3556 
3557 /*
3558  * Calculate the total number of references to a special device.
3559  */
3560 int
3561 vcount(struct vnode *vp)
3562 {
3563 	int count;
3564 
3565 	dev_lock();
3566 	count = vp->v_rdev->si_usecount;
3567 	dev_unlock();
3568 	return (count);
3569 }
3570 
3571 /*
3572  * Same as above, but using the struct cdev *as argument
3573  */
3574 int
3575 count_dev(struct cdev *dev)
3576 {
3577 	int count;
3578 
3579 	dev_lock();
3580 	count = dev->si_usecount;
3581 	dev_unlock();
3582 	return(count);
3583 }
3584 
3585 /*
3586  * Print out a description of a vnode.
3587  */
3588 static char *typename[] =
3589 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3590  "VMARKER"};
3591 
3592 void
3593 vn_printf(struct vnode *vp, const char *fmt, ...)
3594 {
3595 	va_list ap;
3596 	char buf[256], buf2[16];
3597 	u_long flags;
3598 
3599 	va_start(ap, fmt);
3600 	vprintf(fmt, ap);
3601 	va_end(ap);
3602 	printf("%p: ", (void *)vp);
3603 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3604 	printf("    usecount %d, writecount %d, refcount %d",
3605 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3606 	switch (vp->v_type) {
3607 	case VDIR:
3608 		printf(" mountedhere %p\n", vp->v_mountedhere);
3609 		break;
3610 	case VCHR:
3611 		printf(" rdev %p\n", vp->v_rdev);
3612 		break;
3613 	case VSOCK:
3614 		printf(" socket %p\n", vp->v_unpcb);
3615 		break;
3616 	case VFIFO:
3617 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3618 		break;
3619 	default:
3620 		printf("\n");
3621 		break;
3622 	}
3623 	buf[0] = '\0';
3624 	buf[1] = '\0';
3625 	if (vp->v_vflag & VV_ROOT)
3626 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3627 	if (vp->v_vflag & VV_ISTTY)
3628 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3629 	if (vp->v_vflag & VV_NOSYNC)
3630 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3631 	if (vp->v_vflag & VV_ETERNALDEV)
3632 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3633 	if (vp->v_vflag & VV_CACHEDLABEL)
3634 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3635 	if (vp->v_vflag & VV_COPYONWRITE)
3636 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3637 	if (vp->v_vflag & VV_SYSTEM)
3638 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3639 	if (vp->v_vflag & VV_PROCDEP)
3640 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3641 	if (vp->v_vflag & VV_NOKNOTE)
3642 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3643 	if (vp->v_vflag & VV_DELETED)
3644 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3645 	if (vp->v_vflag & VV_MD)
3646 		strlcat(buf, "|VV_MD", sizeof(buf));
3647 	if (vp->v_vflag & VV_FORCEINSMQ)
3648 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3649 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3650 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3651 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3652 	if (flags != 0) {
3653 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3654 		strlcat(buf, buf2, sizeof(buf));
3655 	}
3656 	if (vp->v_iflag & VI_MOUNT)
3657 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3658 	if (vp->v_iflag & VI_DOOMED)
3659 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3660 	if (vp->v_iflag & VI_FREE)
3661 		strlcat(buf, "|VI_FREE", sizeof(buf));
3662 	if (vp->v_iflag & VI_ACTIVE)
3663 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3664 	if (vp->v_iflag & VI_DOINGINACT)
3665 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3666 	if (vp->v_iflag & VI_OWEINACT)
3667 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3668 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3669 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3670 	if (flags != 0) {
3671 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3672 		strlcat(buf, buf2, sizeof(buf));
3673 	}
3674 	printf("    flags (%s)\n", buf + 1);
3675 	if (mtx_owned(VI_MTX(vp)))
3676 		printf(" VI_LOCKed");
3677 	if (vp->v_object != NULL)
3678 		printf("    v_object %p ref %d pages %d "
3679 		    "cleanbuf %d dirtybuf %d\n",
3680 		    vp->v_object, vp->v_object->ref_count,
3681 		    vp->v_object->resident_page_count,
3682 		    vp->v_bufobj.bo_clean.bv_cnt,
3683 		    vp->v_bufobj.bo_dirty.bv_cnt);
3684 	printf("    ");
3685 	lockmgr_printinfo(vp->v_vnlock);
3686 	if (vp->v_data != NULL)
3687 		VOP_PRINT(vp);
3688 }
3689 
3690 #ifdef DDB
3691 /*
3692  * List all of the locked vnodes in the system.
3693  * Called when debugging the kernel.
3694  */
3695 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3696 {
3697 	struct mount *mp;
3698 	struct vnode *vp;
3699 
3700 	/*
3701 	 * Note: because this is DDB, we can't obey the locking semantics
3702 	 * for these structures, which means we could catch an inconsistent
3703 	 * state and dereference a nasty pointer.  Not much to be done
3704 	 * about that.
3705 	 */
3706 	db_printf("Locked vnodes\n");
3707 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3708 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3709 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3710 				vn_printf(vp, "vnode ");
3711 		}
3712 	}
3713 }
3714 
3715 /*
3716  * Show details about the given vnode.
3717  */
3718 DB_SHOW_COMMAND(vnode, db_show_vnode)
3719 {
3720 	struct vnode *vp;
3721 
3722 	if (!have_addr)
3723 		return;
3724 	vp = (struct vnode *)addr;
3725 	vn_printf(vp, "vnode ");
3726 }
3727 
3728 /*
3729  * Show details about the given mount point.
3730  */
3731 DB_SHOW_COMMAND(mount, db_show_mount)
3732 {
3733 	struct mount *mp;
3734 	struct vfsopt *opt;
3735 	struct statfs *sp;
3736 	struct vnode *vp;
3737 	char buf[512];
3738 	uint64_t mflags;
3739 	u_int flags;
3740 
3741 	if (!have_addr) {
3742 		/* No address given, print short info about all mount points. */
3743 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3744 			db_printf("%p %s on %s (%s)\n", mp,
3745 			    mp->mnt_stat.f_mntfromname,
3746 			    mp->mnt_stat.f_mntonname,
3747 			    mp->mnt_stat.f_fstypename);
3748 			if (db_pager_quit)
3749 				break;
3750 		}
3751 		db_printf("\nMore info: show mount <addr>\n");
3752 		return;
3753 	}
3754 
3755 	mp = (struct mount *)addr;
3756 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3757 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3758 
3759 	buf[0] = '\0';
3760 	mflags = mp->mnt_flag;
3761 #define	MNT_FLAG(flag)	do {						\
3762 	if (mflags & (flag)) {						\
3763 		if (buf[0] != '\0')					\
3764 			strlcat(buf, ", ", sizeof(buf));		\
3765 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3766 		mflags &= ~(flag);					\
3767 	}								\
3768 } while (0)
3769 	MNT_FLAG(MNT_RDONLY);
3770 	MNT_FLAG(MNT_SYNCHRONOUS);
3771 	MNT_FLAG(MNT_NOEXEC);
3772 	MNT_FLAG(MNT_NOSUID);
3773 	MNT_FLAG(MNT_NFS4ACLS);
3774 	MNT_FLAG(MNT_UNION);
3775 	MNT_FLAG(MNT_ASYNC);
3776 	MNT_FLAG(MNT_SUIDDIR);
3777 	MNT_FLAG(MNT_SOFTDEP);
3778 	MNT_FLAG(MNT_NOSYMFOLLOW);
3779 	MNT_FLAG(MNT_GJOURNAL);
3780 	MNT_FLAG(MNT_MULTILABEL);
3781 	MNT_FLAG(MNT_ACLS);
3782 	MNT_FLAG(MNT_NOATIME);
3783 	MNT_FLAG(MNT_NOCLUSTERR);
3784 	MNT_FLAG(MNT_NOCLUSTERW);
3785 	MNT_FLAG(MNT_SUJ);
3786 	MNT_FLAG(MNT_EXRDONLY);
3787 	MNT_FLAG(MNT_EXPORTED);
3788 	MNT_FLAG(MNT_DEFEXPORTED);
3789 	MNT_FLAG(MNT_EXPORTANON);
3790 	MNT_FLAG(MNT_EXKERB);
3791 	MNT_FLAG(MNT_EXPUBLIC);
3792 	MNT_FLAG(MNT_LOCAL);
3793 	MNT_FLAG(MNT_QUOTA);
3794 	MNT_FLAG(MNT_ROOTFS);
3795 	MNT_FLAG(MNT_USER);
3796 	MNT_FLAG(MNT_IGNORE);
3797 	MNT_FLAG(MNT_UPDATE);
3798 	MNT_FLAG(MNT_DELEXPORT);
3799 	MNT_FLAG(MNT_RELOAD);
3800 	MNT_FLAG(MNT_FORCE);
3801 	MNT_FLAG(MNT_SNAPSHOT);
3802 	MNT_FLAG(MNT_BYFSID);
3803 #undef MNT_FLAG
3804 	if (mflags != 0) {
3805 		if (buf[0] != '\0')
3806 			strlcat(buf, ", ", sizeof(buf));
3807 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3808 		    "0x%016jx", mflags);
3809 	}
3810 	db_printf("    mnt_flag = %s\n", buf);
3811 
3812 	buf[0] = '\0';
3813 	flags = mp->mnt_kern_flag;
3814 #define	MNT_KERN_FLAG(flag)	do {					\
3815 	if (flags & (flag)) {						\
3816 		if (buf[0] != '\0')					\
3817 			strlcat(buf, ", ", sizeof(buf));		\
3818 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3819 		flags &= ~(flag);					\
3820 	}								\
3821 } while (0)
3822 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3823 	MNT_KERN_FLAG(MNTK_ASYNC);
3824 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3825 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3826 	MNT_KERN_FLAG(MNTK_DRAINING);
3827 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3828 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3829 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3830 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3831 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3832 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3833 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3834 	MNT_KERN_FLAG(MNTK_MARKER);
3835 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3836 	MNT_KERN_FLAG(MNTK_NOASYNC);
3837 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3838 	MNT_KERN_FLAG(MNTK_MWAIT);
3839 	MNT_KERN_FLAG(MNTK_SUSPEND);
3840 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3841 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3842 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3843 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3844 #undef MNT_KERN_FLAG
3845 	if (flags != 0) {
3846 		if (buf[0] != '\0')
3847 			strlcat(buf, ", ", sizeof(buf));
3848 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3849 		    "0x%08x", flags);
3850 	}
3851 	db_printf("    mnt_kern_flag = %s\n", buf);
3852 
3853 	db_printf("    mnt_opt = ");
3854 	opt = TAILQ_FIRST(mp->mnt_opt);
3855 	if (opt != NULL) {
3856 		db_printf("%s", opt->name);
3857 		opt = TAILQ_NEXT(opt, link);
3858 		while (opt != NULL) {
3859 			db_printf(", %s", opt->name);
3860 			opt = TAILQ_NEXT(opt, link);
3861 		}
3862 	}
3863 	db_printf("\n");
3864 
3865 	sp = &mp->mnt_stat;
3866 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3867 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3868 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3869 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3870 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3871 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3872 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3873 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3874 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3875 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3876 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3877 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3878 
3879 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3880 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3881 	if (jailed(mp->mnt_cred))
3882 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3883 	db_printf(" }\n");
3884 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3885 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3886 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3887 	db_printf("    mnt_activevnodelistsize = %d\n",
3888 	    mp->mnt_activevnodelistsize);
3889 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3890 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3891 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3892 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3893 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3894 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3895 	db_printf("    mnt_secondary_accwrites = %d\n",
3896 	    mp->mnt_secondary_accwrites);
3897 	db_printf("    mnt_gjprovider = %s\n",
3898 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3899 
3900 	db_printf("\n\nList of active vnodes\n");
3901 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3902 		if (vp->v_type != VMARKER) {
3903 			vn_printf(vp, "vnode ");
3904 			if (db_pager_quit)
3905 				break;
3906 		}
3907 	}
3908 	db_printf("\n\nList of inactive vnodes\n");
3909 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3910 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3911 			vn_printf(vp, "vnode ");
3912 			if (db_pager_quit)
3913 				break;
3914 		}
3915 	}
3916 }
3917 #endif	/* DDB */
3918 
3919 /*
3920  * Fill in a struct xvfsconf based on a struct vfsconf.
3921  */
3922 static int
3923 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3924 {
3925 	struct xvfsconf xvfsp;
3926 
3927 	bzero(&xvfsp, sizeof(xvfsp));
3928 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3929 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3930 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3931 	xvfsp.vfc_flags = vfsp->vfc_flags;
3932 	/*
3933 	 * These are unused in userland, we keep them
3934 	 * to not break binary compatibility.
3935 	 */
3936 	xvfsp.vfc_vfsops = NULL;
3937 	xvfsp.vfc_next = NULL;
3938 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3939 }
3940 
3941 #ifdef COMPAT_FREEBSD32
3942 struct xvfsconf32 {
3943 	uint32_t	vfc_vfsops;
3944 	char		vfc_name[MFSNAMELEN];
3945 	int32_t		vfc_typenum;
3946 	int32_t		vfc_refcount;
3947 	int32_t		vfc_flags;
3948 	uint32_t	vfc_next;
3949 };
3950 
3951 static int
3952 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3953 {
3954 	struct xvfsconf32 xvfsp;
3955 
3956 	bzero(&xvfsp, sizeof(xvfsp));
3957 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3958 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3959 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3960 	xvfsp.vfc_flags = vfsp->vfc_flags;
3961 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3962 }
3963 #endif
3964 
3965 /*
3966  * Top level filesystem related information gathering.
3967  */
3968 static int
3969 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3970 {
3971 	struct vfsconf *vfsp;
3972 	int error;
3973 
3974 	error = 0;
3975 	vfsconf_slock();
3976 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3977 #ifdef COMPAT_FREEBSD32
3978 		if (req->flags & SCTL_MASK32)
3979 			error = vfsconf2x32(req, vfsp);
3980 		else
3981 #endif
3982 			error = vfsconf2x(req, vfsp);
3983 		if (error)
3984 			break;
3985 	}
3986 	vfsconf_sunlock();
3987 	return (error);
3988 }
3989 
3990 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3991     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3992     "S,xvfsconf", "List of all configured filesystems");
3993 
3994 #ifndef BURN_BRIDGES
3995 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3996 
3997 static int
3998 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3999 {
4000 	int *name = (int *)arg1 - 1;	/* XXX */
4001 	u_int namelen = arg2 + 1;	/* XXX */
4002 	struct vfsconf *vfsp;
4003 
4004 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4005 	    "please rebuild world\n");
4006 
4007 #if 1 || defined(COMPAT_PRELITE2)
4008 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4009 	if (namelen == 1)
4010 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4011 #endif
4012 
4013 	switch (name[1]) {
4014 	case VFS_MAXTYPENUM:
4015 		if (namelen != 2)
4016 			return (ENOTDIR);
4017 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4018 	case VFS_CONF:
4019 		if (namelen != 3)
4020 			return (ENOTDIR);	/* overloaded */
4021 		vfsconf_slock();
4022 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4023 			if (vfsp->vfc_typenum == name[2])
4024 				break;
4025 		}
4026 		vfsconf_sunlock();
4027 		if (vfsp == NULL)
4028 			return (EOPNOTSUPP);
4029 #ifdef COMPAT_FREEBSD32
4030 		if (req->flags & SCTL_MASK32)
4031 			return (vfsconf2x32(req, vfsp));
4032 		else
4033 #endif
4034 			return (vfsconf2x(req, vfsp));
4035 	}
4036 	return (EOPNOTSUPP);
4037 }
4038 
4039 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4040     CTLFLAG_MPSAFE, vfs_sysctl,
4041     "Generic filesystem");
4042 
4043 #if 1 || defined(COMPAT_PRELITE2)
4044 
4045 static int
4046 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4047 {
4048 	int error;
4049 	struct vfsconf *vfsp;
4050 	struct ovfsconf ovfs;
4051 
4052 	vfsconf_slock();
4053 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4054 		bzero(&ovfs, sizeof(ovfs));
4055 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4056 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4057 		ovfs.vfc_index = vfsp->vfc_typenum;
4058 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4059 		ovfs.vfc_flags = vfsp->vfc_flags;
4060 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4061 		if (error != 0) {
4062 			vfsconf_sunlock();
4063 			return (error);
4064 		}
4065 	}
4066 	vfsconf_sunlock();
4067 	return (0);
4068 }
4069 
4070 #endif /* 1 || COMPAT_PRELITE2 */
4071 #endif /* !BURN_BRIDGES */
4072 
4073 #define KINFO_VNODESLOP		10
4074 #ifdef notyet
4075 /*
4076  * Dump vnode list (via sysctl).
4077  */
4078 /* ARGSUSED */
4079 static int
4080 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4081 {
4082 	struct xvnode *xvn;
4083 	struct mount *mp;
4084 	struct vnode *vp;
4085 	int error, len, n;
4086 
4087 	/*
4088 	 * Stale numvnodes access is not fatal here.
4089 	 */
4090 	req->lock = 0;
4091 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4092 	if (!req->oldptr)
4093 		/* Make an estimate */
4094 		return (SYSCTL_OUT(req, 0, len));
4095 
4096 	error = sysctl_wire_old_buffer(req, 0);
4097 	if (error != 0)
4098 		return (error);
4099 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4100 	n = 0;
4101 	mtx_lock(&mountlist_mtx);
4102 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4103 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4104 			continue;
4105 		MNT_ILOCK(mp);
4106 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4107 			if (n == len)
4108 				break;
4109 			vref(vp);
4110 			xvn[n].xv_size = sizeof *xvn;
4111 			xvn[n].xv_vnode = vp;
4112 			xvn[n].xv_id = 0;	/* XXX compat */
4113 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4114 			XV_COPY(usecount);
4115 			XV_COPY(writecount);
4116 			XV_COPY(holdcnt);
4117 			XV_COPY(mount);
4118 			XV_COPY(numoutput);
4119 			XV_COPY(type);
4120 #undef XV_COPY
4121 			xvn[n].xv_flag = vp->v_vflag;
4122 
4123 			switch (vp->v_type) {
4124 			case VREG:
4125 			case VDIR:
4126 			case VLNK:
4127 				break;
4128 			case VBLK:
4129 			case VCHR:
4130 				if (vp->v_rdev == NULL) {
4131 					vrele(vp);
4132 					continue;
4133 				}
4134 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4135 				break;
4136 			case VSOCK:
4137 				xvn[n].xv_socket = vp->v_socket;
4138 				break;
4139 			case VFIFO:
4140 				xvn[n].xv_fifo = vp->v_fifoinfo;
4141 				break;
4142 			case VNON:
4143 			case VBAD:
4144 			default:
4145 				/* shouldn't happen? */
4146 				vrele(vp);
4147 				continue;
4148 			}
4149 			vrele(vp);
4150 			++n;
4151 		}
4152 		MNT_IUNLOCK(mp);
4153 		mtx_lock(&mountlist_mtx);
4154 		vfs_unbusy(mp);
4155 		if (n == len)
4156 			break;
4157 	}
4158 	mtx_unlock(&mountlist_mtx);
4159 
4160 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4161 	free(xvn, M_TEMP);
4162 	return (error);
4163 }
4164 
4165 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4166     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4167     "");
4168 #endif
4169 
4170 static void
4171 unmount_or_warn(struct mount *mp)
4172 {
4173 	int error;
4174 
4175 	error = dounmount(mp, MNT_FORCE, curthread);
4176 	if (error != 0) {
4177 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4178 		if (error == EBUSY)
4179 			printf("BUSY)\n");
4180 		else
4181 			printf("%d)\n", error);
4182 	}
4183 }
4184 
4185 /*
4186  * Unmount all filesystems. The list is traversed in reverse order
4187  * of mounting to avoid dependencies.
4188  */
4189 void
4190 vfs_unmountall(void)
4191 {
4192 	struct mount *mp, *tmp;
4193 
4194 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4195 
4196 	/*
4197 	 * Since this only runs when rebooting, it is not interlocked.
4198 	 */
4199 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4200 		vfs_ref(mp);
4201 
4202 		/*
4203 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4204 		 * unmount of the latter.
4205 		 */
4206 		if (mp == rootdevmp)
4207 			continue;
4208 
4209 		unmount_or_warn(mp);
4210 	}
4211 
4212 	if (rootdevmp != NULL)
4213 		unmount_or_warn(rootdevmp);
4214 }
4215 
4216 /*
4217  * perform msync on all vnodes under a mount point
4218  * the mount point must be locked.
4219  */
4220 void
4221 vfs_msync(struct mount *mp, int flags)
4222 {
4223 	struct vnode *vp, *mvp;
4224 	struct vm_object *obj;
4225 
4226 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4227 
4228 	vnlru_return_batch(mp);
4229 
4230 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4231 		obj = vp->v_object;
4232 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4233 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4234 			if (!vget(vp,
4235 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4236 			    curthread)) {
4237 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4238 					vput(vp);
4239 					continue;
4240 				}
4241 
4242 				obj = vp->v_object;
4243 				if (obj != NULL) {
4244 					VM_OBJECT_WLOCK(obj);
4245 					vm_object_page_clean(obj, 0, 0,
4246 					    flags == MNT_WAIT ?
4247 					    OBJPC_SYNC : OBJPC_NOSYNC);
4248 					VM_OBJECT_WUNLOCK(obj);
4249 				}
4250 				vput(vp);
4251 			}
4252 		} else
4253 			VI_UNLOCK(vp);
4254 	}
4255 }
4256 
4257 static void
4258 destroy_vpollinfo_free(struct vpollinfo *vi)
4259 {
4260 
4261 	knlist_destroy(&vi->vpi_selinfo.si_note);
4262 	mtx_destroy(&vi->vpi_lock);
4263 	uma_zfree(vnodepoll_zone, vi);
4264 }
4265 
4266 static void
4267 destroy_vpollinfo(struct vpollinfo *vi)
4268 {
4269 
4270 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4271 	seldrain(&vi->vpi_selinfo);
4272 	destroy_vpollinfo_free(vi);
4273 }
4274 
4275 /*
4276  * Initialize per-vnode helper structure to hold poll-related state.
4277  */
4278 void
4279 v_addpollinfo(struct vnode *vp)
4280 {
4281 	struct vpollinfo *vi;
4282 
4283 	if (vp->v_pollinfo != NULL)
4284 		return;
4285 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4286 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4287 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4288 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4289 	VI_LOCK(vp);
4290 	if (vp->v_pollinfo != NULL) {
4291 		VI_UNLOCK(vp);
4292 		destroy_vpollinfo_free(vi);
4293 		return;
4294 	}
4295 	vp->v_pollinfo = vi;
4296 	VI_UNLOCK(vp);
4297 }
4298 
4299 /*
4300  * Record a process's interest in events which might happen to
4301  * a vnode.  Because poll uses the historic select-style interface
4302  * internally, this routine serves as both the ``check for any
4303  * pending events'' and the ``record my interest in future events''
4304  * functions.  (These are done together, while the lock is held,
4305  * to avoid race conditions.)
4306  */
4307 int
4308 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4309 {
4310 
4311 	v_addpollinfo(vp);
4312 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4313 	if (vp->v_pollinfo->vpi_revents & events) {
4314 		/*
4315 		 * This leaves events we are not interested
4316 		 * in available for the other process which
4317 		 * which presumably had requested them
4318 		 * (otherwise they would never have been
4319 		 * recorded).
4320 		 */
4321 		events &= vp->v_pollinfo->vpi_revents;
4322 		vp->v_pollinfo->vpi_revents &= ~events;
4323 
4324 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4325 		return (events);
4326 	}
4327 	vp->v_pollinfo->vpi_events |= events;
4328 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4329 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4330 	return (0);
4331 }
4332 
4333 /*
4334  * Routine to create and manage a filesystem syncer vnode.
4335  */
4336 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4337 static int	sync_fsync(struct  vop_fsync_args *);
4338 static int	sync_inactive(struct  vop_inactive_args *);
4339 static int	sync_reclaim(struct  vop_reclaim_args *);
4340 
4341 static struct vop_vector sync_vnodeops = {
4342 	.vop_bypass =	VOP_EOPNOTSUPP,
4343 	.vop_close =	sync_close,		/* close */
4344 	.vop_fsync =	sync_fsync,		/* fsync */
4345 	.vop_inactive =	sync_inactive,	/* inactive */
4346 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4347 	.vop_lock1 =	vop_stdlock,	/* lock */
4348 	.vop_unlock =	vop_stdunlock,	/* unlock */
4349 	.vop_islocked =	vop_stdislocked,	/* islocked */
4350 };
4351 
4352 /*
4353  * Create a new filesystem syncer vnode for the specified mount point.
4354  */
4355 void
4356 vfs_allocate_syncvnode(struct mount *mp)
4357 {
4358 	struct vnode *vp;
4359 	struct bufobj *bo;
4360 	static long start, incr, next;
4361 	int error;
4362 
4363 	/* Allocate a new vnode */
4364 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4365 	if (error != 0)
4366 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4367 	vp->v_type = VNON;
4368 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4369 	vp->v_vflag |= VV_FORCEINSMQ;
4370 	error = insmntque(vp, mp);
4371 	if (error != 0)
4372 		panic("vfs_allocate_syncvnode: insmntque() failed");
4373 	vp->v_vflag &= ~VV_FORCEINSMQ;
4374 	VOP_UNLOCK(vp, 0);
4375 	/*
4376 	 * Place the vnode onto the syncer worklist. We attempt to
4377 	 * scatter them about on the list so that they will go off
4378 	 * at evenly distributed times even if all the filesystems
4379 	 * are mounted at once.
4380 	 */
4381 	next += incr;
4382 	if (next == 0 || next > syncer_maxdelay) {
4383 		start /= 2;
4384 		incr /= 2;
4385 		if (start == 0) {
4386 			start = syncer_maxdelay / 2;
4387 			incr = syncer_maxdelay;
4388 		}
4389 		next = start;
4390 	}
4391 	bo = &vp->v_bufobj;
4392 	BO_LOCK(bo);
4393 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4394 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4395 	mtx_lock(&sync_mtx);
4396 	sync_vnode_count++;
4397 	if (mp->mnt_syncer == NULL) {
4398 		mp->mnt_syncer = vp;
4399 		vp = NULL;
4400 	}
4401 	mtx_unlock(&sync_mtx);
4402 	BO_UNLOCK(bo);
4403 	if (vp != NULL) {
4404 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4405 		vgone(vp);
4406 		vput(vp);
4407 	}
4408 }
4409 
4410 void
4411 vfs_deallocate_syncvnode(struct mount *mp)
4412 {
4413 	struct vnode *vp;
4414 
4415 	mtx_lock(&sync_mtx);
4416 	vp = mp->mnt_syncer;
4417 	if (vp != NULL)
4418 		mp->mnt_syncer = NULL;
4419 	mtx_unlock(&sync_mtx);
4420 	if (vp != NULL)
4421 		vrele(vp);
4422 }
4423 
4424 /*
4425  * Do a lazy sync of the filesystem.
4426  */
4427 static int
4428 sync_fsync(struct vop_fsync_args *ap)
4429 {
4430 	struct vnode *syncvp = ap->a_vp;
4431 	struct mount *mp = syncvp->v_mount;
4432 	int error, save;
4433 	struct bufobj *bo;
4434 
4435 	/*
4436 	 * We only need to do something if this is a lazy evaluation.
4437 	 */
4438 	if (ap->a_waitfor != MNT_LAZY)
4439 		return (0);
4440 
4441 	/*
4442 	 * Move ourselves to the back of the sync list.
4443 	 */
4444 	bo = &syncvp->v_bufobj;
4445 	BO_LOCK(bo);
4446 	vn_syncer_add_to_worklist(bo, syncdelay);
4447 	BO_UNLOCK(bo);
4448 
4449 	/*
4450 	 * Walk the list of vnodes pushing all that are dirty and
4451 	 * not already on the sync list.
4452 	 */
4453 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4454 		return (0);
4455 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4456 		vfs_unbusy(mp);
4457 		return (0);
4458 	}
4459 	save = curthread_pflags_set(TDP_SYNCIO);
4460 	vfs_msync(mp, MNT_NOWAIT);
4461 	error = VFS_SYNC(mp, MNT_LAZY);
4462 	curthread_pflags_restore(save);
4463 	vn_finished_write(mp);
4464 	vfs_unbusy(mp);
4465 	return (error);
4466 }
4467 
4468 /*
4469  * The syncer vnode is no referenced.
4470  */
4471 static int
4472 sync_inactive(struct vop_inactive_args *ap)
4473 {
4474 
4475 	vgone(ap->a_vp);
4476 	return (0);
4477 }
4478 
4479 /*
4480  * The syncer vnode is no longer needed and is being decommissioned.
4481  *
4482  * Modifications to the worklist must be protected by sync_mtx.
4483  */
4484 static int
4485 sync_reclaim(struct vop_reclaim_args *ap)
4486 {
4487 	struct vnode *vp = ap->a_vp;
4488 	struct bufobj *bo;
4489 
4490 	bo = &vp->v_bufobj;
4491 	BO_LOCK(bo);
4492 	mtx_lock(&sync_mtx);
4493 	if (vp->v_mount->mnt_syncer == vp)
4494 		vp->v_mount->mnt_syncer = NULL;
4495 	if (bo->bo_flag & BO_ONWORKLST) {
4496 		LIST_REMOVE(bo, bo_synclist);
4497 		syncer_worklist_len--;
4498 		sync_vnode_count--;
4499 		bo->bo_flag &= ~BO_ONWORKLST;
4500 	}
4501 	mtx_unlock(&sync_mtx);
4502 	BO_UNLOCK(bo);
4503 
4504 	return (0);
4505 }
4506 
4507 /*
4508  * Check if vnode represents a disk device
4509  */
4510 int
4511 vn_isdisk(struct vnode *vp, int *errp)
4512 {
4513 	int error;
4514 
4515 	if (vp->v_type != VCHR) {
4516 		error = ENOTBLK;
4517 		goto out;
4518 	}
4519 	error = 0;
4520 	dev_lock();
4521 	if (vp->v_rdev == NULL)
4522 		error = ENXIO;
4523 	else if (vp->v_rdev->si_devsw == NULL)
4524 		error = ENXIO;
4525 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4526 		error = ENOTBLK;
4527 	dev_unlock();
4528 out:
4529 	if (errp != NULL)
4530 		*errp = error;
4531 	return (error == 0);
4532 }
4533 
4534 /*
4535  * Common filesystem object access control check routine.  Accepts a
4536  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4537  * and optional call-by-reference privused argument allowing vaccess()
4538  * to indicate to the caller whether privilege was used to satisfy the
4539  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4540  */
4541 int
4542 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4543     accmode_t accmode, struct ucred *cred, int *privused)
4544 {
4545 	accmode_t dac_granted;
4546 	accmode_t priv_granted;
4547 
4548 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4549 	    ("invalid bit in accmode"));
4550 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4551 	    ("VAPPEND without VWRITE"));
4552 
4553 	/*
4554 	 * Look for a normal, non-privileged way to access the file/directory
4555 	 * as requested.  If it exists, go with that.
4556 	 */
4557 
4558 	if (privused != NULL)
4559 		*privused = 0;
4560 
4561 	dac_granted = 0;
4562 
4563 	/* Check the owner. */
4564 	if (cred->cr_uid == file_uid) {
4565 		dac_granted |= VADMIN;
4566 		if (file_mode & S_IXUSR)
4567 			dac_granted |= VEXEC;
4568 		if (file_mode & S_IRUSR)
4569 			dac_granted |= VREAD;
4570 		if (file_mode & S_IWUSR)
4571 			dac_granted |= (VWRITE | VAPPEND);
4572 
4573 		if ((accmode & dac_granted) == accmode)
4574 			return (0);
4575 
4576 		goto privcheck;
4577 	}
4578 
4579 	/* Otherwise, check the groups (first match) */
4580 	if (groupmember(file_gid, cred)) {
4581 		if (file_mode & S_IXGRP)
4582 			dac_granted |= VEXEC;
4583 		if (file_mode & S_IRGRP)
4584 			dac_granted |= VREAD;
4585 		if (file_mode & S_IWGRP)
4586 			dac_granted |= (VWRITE | VAPPEND);
4587 
4588 		if ((accmode & dac_granted) == accmode)
4589 			return (0);
4590 
4591 		goto privcheck;
4592 	}
4593 
4594 	/* Otherwise, check everyone else. */
4595 	if (file_mode & S_IXOTH)
4596 		dac_granted |= VEXEC;
4597 	if (file_mode & S_IROTH)
4598 		dac_granted |= VREAD;
4599 	if (file_mode & S_IWOTH)
4600 		dac_granted |= (VWRITE | VAPPEND);
4601 	if ((accmode & dac_granted) == accmode)
4602 		return (0);
4603 
4604 privcheck:
4605 	/*
4606 	 * Build a privilege mask to determine if the set of privileges
4607 	 * satisfies the requirements when combined with the granted mask
4608 	 * from above.  For each privilege, if the privilege is required,
4609 	 * bitwise or the request type onto the priv_granted mask.
4610 	 */
4611 	priv_granted = 0;
4612 
4613 	if (type == VDIR) {
4614 		/*
4615 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4616 		 * requests, instead of PRIV_VFS_EXEC.
4617 		 */
4618 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4619 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4620 			priv_granted |= VEXEC;
4621 	} else {
4622 		/*
4623 		 * Ensure that at least one execute bit is on. Otherwise,
4624 		 * a privileged user will always succeed, and we don't want
4625 		 * this to happen unless the file really is executable.
4626 		 */
4627 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4628 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4629 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4630 			priv_granted |= VEXEC;
4631 	}
4632 
4633 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4634 	    !priv_check_cred(cred, PRIV_VFS_READ))
4635 		priv_granted |= VREAD;
4636 
4637 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4638 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4639 		priv_granted |= (VWRITE | VAPPEND);
4640 
4641 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4642 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4643 		priv_granted |= VADMIN;
4644 
4645 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4646 		/* XXX audit: privilege used */
4647 		if (privused != NULL)
4648 			*privused = 1;
4649 		return (0);
4650 	}
4651 
4652 	return ((accmode & VADMIN) ? EPERM : EACCES);
4653 }
4654 
4655 /*
4656  * Credential check based on process requesting service, and per-attribute
4657  * permissions.
4658  */
4659 int
4660 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4661     struct thread *td, accmode_t accmode)
4662 {
4663 
4664 	/*
4665 	 * Kernel-invoked always succeeds.
4666 	 */
4667 	if (cred == NOCRED)
4668 		return (0);
4669 
4670 	/*
4671 	 * Do not allow privileged processes in jail to directly manipulate
4672 	 * system attributes.
4673 	 */
4674 	switch (attrnamespace) {
4675 	case EXTATTR_NAMESPACE_SYSTEM:
4676 		/* Potentially should be: return (EPERM); */
4677 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4678 	case EXTATTR_NAMESPACE_USER:
4679 		return (VOP_ACCESS(vp, accmode, cred, td));
4680 	default:
4681 		return (EPERM);
4682 	}
4683 }
4684 
4685 #ifdef DEBUG_VFS_LOCKS
4686 /*
4687  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4688  * no longer ok to have an unlocked VFS.
4689  */
4690 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4691 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4692 
4693 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4694 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4695     "Drop into debugger on lock violation");
4696 
4697 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4698 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4699     0, "Check for interlock across VOPs");
4700 
4701 int vfs_badlock_print = 1;	/* Print lock violations. */
4702 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4703     0, "Print lock violations");
4704 
4705 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4706 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4707     0, "Print vnode details on lock violations");
4708 
4709 #ifdef KDB
4710 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4711 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4712     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4713 #endif
4714 
4715 static void
4716 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4717 {
4718 
4719 #ifdef KDB
4720 	if (vfs_badlock_backtrace)
4721 		kdb_backtrace();
4722 #endif
4723 	if (vfs_badlock_vnode)
4724 		vn_printf(vp, "vnode ");
4725 	if (vfs_badlock_print)
4726 		printf("%s: %p %s\n", str, (void *)vp, msg);
4727 	if (vfs_badlock_ddb)
4728 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4729 }
4730 
4731 void
4732 assert_vi_locked(struct vnode *vp, const char *str)
4733 {
4734 
4735 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4736 		vfs_badlock("interlock is not locked but should be", str, vp);
4737 }
4738 
4739 void
4740 assert_vi_unlocked(struct vnode *vp, const char *str)
4741 {
4742 
4743 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4744 		vfs_badlock("interlock is locked but should not be", str, vp);
4745 }
4746 
4747 void
4748 assert_vop_locked(struct vnode *vp, const char *str)
4749 {
4750 	int locked;
4751 
4752 	if (!IGNORE_LOCK(vp)) {
4753 		locked = VOP_ISLOCKED(vp);
4754 		if (locked == 0 || locked == LK_EXCLOTHER)
4755 			vfs_badlock("is not locked but should be", str, vp);
4756 	}
4757 }
4758 
4759 void
4760 assert_vop_unlocked(struct vnode *vp, const char *str)
4761 {
4762 
4763 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4764 		vfs_badlock("is locked but should not be", str, vp);
4765 }
4766 
4767 void
4768 assert_vop_elocked(struct vnode *vp, const char *str)
4769 {
4770 
4771 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4772 		vfs_badlock("is not exclusive locked but should be", str, vp);
4773 }
4774 #endif /* DEBUG_VFS_LOCKS */
4775 
4776 void
4777 vop_rename_fail(struct vop_rename_args *ap)
4778 {
4779 
4780 	if (ap->a_tvp != NULL)
4781 		vput(ap->a_tvp);
4782 	if (ap->a_tdvp == ap->a_tvp)
4783 		vrele(ap->a_tdvp);
4784 	else
4785 		vput(ap->a_tdvp);
4786 	vrele(ap->a_fdvp);
4787 	vrele(ap->a_fvp);
4788 }
4789 
4790 void
4791 vop_rename_pre(void *ap)
4792 {
4793 	struct vop_rename_args *a = ap;
4794 
4795 #ifdef DEBUG_VFS_LOCKS
4796 	if (a->a_tvp)
4797 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4798 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4799 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4800 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4801 
4802 	/* Check the source (from). */
4803 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4804 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4805 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4806 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4807 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4808 
4809 	/* Check the target. */
4810 	if (a->a_tvp)
4811 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4812 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4813 #endif
4814 	if (a->a_tdvp != a->a_fdvp)
4815 		vhold(a->a_fdvp);
4816 	if (a->a_tvp != a->a_fvp)
4817 		vhold(a->a_fvp);
4818 	vhold(a->a_tdvp);
4819 	if (a->a_tvp)
4820 		vhold(a->a_tvp);
4821 }
4822 
4823 #ifdef DEBUG_VFS_LOCKS
4824 void
4825 vop_strategy_pre(void *ap)
4826 {
4827 	struct vop_strategy_args *a;
4828 	struct buf *bp;
4829 
4830 	a = ap;
4831 	bp = a->a_bp;
4832 
4833 	/*
4834 	 * Cluster ops lock their component buffers but not the IO container.
4835 	 */
4836 	if ((bp->b_flags & B_CLUSTER) != 0)
4837 		return;
4838 
4839 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4840 		if (vfs_badlock_print)
4841 			printf(
4842 			    "VOP_STRATEGY: bp is not locked but should be\n");
4843 		if (vfs_badlock_ddb)
4844 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4845 	}
4846 }
4847 
4848 void
4849 vop_lock_pre(void *ap)
4850 {
4851 	struct vop_lock1_args *a = ap;
4852 
4853 	if ((a->a_flags & LK_INTERLOCK) == 0)
4854 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4855 	else
4856 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4857 }
4858 
4859 void
4860 vop_lock_post(void *ap, int rc)
4861 {
4862 	struct vop_lock1_args *a = ap;
4863 
4864 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4865 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4866 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4867 }
4868 
4869 void
4870 vop_unlock_pre(void *ap)
4871 {
4872 	struct vop_unlock_args *a = ap;
4873 
4874 	if (a->a_flags & LK_INTERLOCK)
4875 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4876 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4877 }
4878 
4879 void
4880 vop_unlock_post(void *ap, int rc)
4881 {
4882 	struct vop_unlock_args *a = ap;
4883 
4884 	if (a->a_flags & LK_INTERLOCK)
4885 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4886 }
4887 #endif
4888 
4889 void
4890 vop_create_post(void *ap, int rc)
4891 {
4892 	struct vop_create_args *a = ap;
4893 
4894 	if (!rc)
4895 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4896 }
4897 
4898 void
4899 vop_deleteextattr_post(void *ap, int rc)
4900 {
4901 	struct vop_deleteextattr_args *a = ap;
4902 
4903 	if (!rc)
4904 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4905 }
4906 
4907 void
4908 vop_link_post(void *ap, int rc)
4909 {
4910 	struct vop_link_args *a = ap;
4911 
4912 	if (!rc) {
4913 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4914 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4915 	}
4916 }
4917 
4918 void
4919 vop_mkdir_post(void *ap, int rc)
4920 {
4921 	struct vop_mkdir_args *a = ap;
4922 
4923 	if (!rc)
4924 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4925 }
4926 
4927 void
4928 vop_mknod_post(void *ap, int rc)
4929 {
4930 	struct vop_mknod_args *a = ap;
4931 
4932 	if (!rc)
4933 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4934 }
4935 
4936 void
4937 vop_reclaim_post(void *ap, int rc)
4938 {
4939 	struct vop_reclaim_args *a = ap;
4940 
4941 	if (!rc)
4942 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4943 }
4944 
4945 void
4946 vop_remove_post(void *ap, int rc)
4947 {
4948 	struct vop_remove_args *a = ap;
4949 
4950 	if (!rc) {
4951 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4952 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4953 	}
4954 }
4955 
4956 void
4957 vop_rename_post(void *ap, int rc)
4958 {
4959 	struct vop_rename_args *a = ap;
4960 	long hint;
4961 
4962 	if (!rc) {
4963 		hint = NOTE_WRITE;
4964 		if (a->a_fdvp == a->a_tdvp) {
4965 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4966 				hint |= NOTE_LINK;
4967 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4968 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4969 		} else {
4970 			hint |= NOTE_EXTEND;
4971 			if (a->a_fvp->v_type == VDIR)
4972 				hint |= NOTE_LINK;
4973 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4974 
4975 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4976 			    a->a_tvp->v_type == VDIR)
4977 				hint &= ~NOTE_LINK;
4978 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4979 		}
4980 
4981 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4982 		if (a->a_tvp)
4983 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4984 	}
4985 	if (a->a_tdvp != a->a_fdvp)
4986 		vdrop(a->a_fdvp);
4987 	if (a->a_tvp != a->a_fvp)
4988 		vdrop(a->a_fvp);
4989 	vdrop(a->a_tdvp);
4990 	if (a->a_tvp)
4991 		vdrop(a->a_tvp);
4992 }
4993 
4994 void
4995 vop_rmdir_post(void *ap, int rc)
4996 {
4997 	struct vop_rmdir_args *a = ap;
4998 
4999 	if (!rc) {
5000 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5001 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5002 	}
5003 }
5004 
5005 void
5006 vop_setattr_post(void *ap, int rc)
5007 {
5008 	struct vop_setattr_args *a = ap;
5009 
5010 	if (!rc)
5011 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5012 }
5013 
5014 void
5015 vop_setextattr_post(void *ap, int rc)
5016 {
5017 	struct vop_setextattr_args *a = ap;
5018 
5019 	if (!rc)
5020 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5021 }
5022 
5023 void
5024 vop_symlink_post(void *ap, int rc)
5025 {
5026 	struct vop_symlink_args *a = ap;
5027 
5028 	if (!rc)
5029 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5030 }
5031 
5032 void
5033 vop_open_post(void *ap, int rc)
5034 {
5035 	struct vop_open_args *a = ap;
5036 
5037 	if (!rc)
5038 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5039 }
5040 
5041 void
5042 vop_close_post(void *ap, int rc)
5043 {
5044 	struct vop_close_args *a = ap;
5045 
5046 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5047 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
5048 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5049 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5050 	}
5051 }
5052 
5053 void
5054 vop_read_post(void *ap, int rc)
5055 {
5056 	struct vop_read_args *a = ap;
5057 
5058 	if (!rc)
5059 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5060 }
5061 
5062 void
5063 vop_readdir_post(void *ap, int rc)
5064 {
5065 	struct vop_readdir_args *a = ap;
5066 
5067 	if (!rc)
5068 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5069 }
5070 
5071 static struct knlist fs_knlist;
5072 
5073 static void
5074 vfs_event_init(void *arg)
5075 {
5076 	knlist_init_mtx(&fs_knlist, NULL);
5077 }
5078 /* XXX - correct order? */
5079 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5080 
5081 void
5082 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5083 {
5084 
5085 	KNOTE_UNLOCKED(&fs_knlist, event);
5086 }
5087 
5088 static int	filt_fsattach(struct knote *kn);
5089 static void	filt_fsdetach(struct knote *kn);
5090 static int	filt_fsevent(struct knote *kn, long hint);
5091 
5092 struct filterops fs_filtops = {
5093 	.f_isfd = 0,
5094 	.f_attach = filt_fsattach,
5095 	.f_detach = filt_fsdetach,
5096 	.f_event = filt_fsevent
5097 };
5098 
5099 static int
5100 filt_fsattach(struct knote *kn)
5101 {
5102 
5103 	kn->kn_flags |= EV_CLEAR;
5104 	knlist_add(&fs_knlist, kn, 0);
5105 	return (0);
5106 }
5107 
5108 static void
5109 filt_fsdetach(struct knote *kn)
5110 {
5111 
5112 	knlist_remove(&fs_knlist, kn, 0);
5113 }
5114 
5115 static int
5116 filt_fsevent(struct knote *kn, long hint)
5117 {
5118 
5119 	kn->kn_fflags |= hint;
5120 	return (kn->kn_fflags != 0);
5121 }
5122 
5123 static int
5124 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5125 {
5126 	struct vfsidctl vc;
5127 	int error;
5128 	struct mount *mp;
5129 
5130 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5131 	if (error)
5132 		return (error);
5133 	if (vc.vc_vers != VFS_CTL_VERS1)
5134 		return (EINVAL);
5135 	mp = vfs_getvfs(&vc.vc_fsid);
5136 	if (mp == NULL)
5137 		return (ENOENT);
5138 	/* ensure that a specific sysctl goes to the right filesystem. */
5139 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5140 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5141 		vfs_rel(mp);
5142 		return (EINVAL);
5143 	}
5144 	VCTLTOREQ(&vc, req);
5145 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5146 	vfs_rel(mp);
5147 	return (error);
5148 }
5149 
5150 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5151     NULL, 0, sysctl_vfs_ctl, "",
5152     "Sysctl by fsid");
5153 
5154 /*
5155  * Function to initialize a va_filerev field sensibly.
5156  * XXX: Wouldn't a random number make a lot more sense ??
5157  */
5158 u_quad_t
5159 init_va_filerev(void)
5160 {
5161 	struct bintime bt;
5162 
5163 	getbinuptime(&bt);
5164 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5165 }
5166 
5167 static int	filt_vfsread(struct knote *kn, long hint);
5168 static int	filt_vfswrite(struct knote *kn, long hint);
5169 static int	filt_vfsvnode(struct knote *kn, long hint);
5170 static void	filt_vfsdetach(struct knote *kn);
5171 static struct filterops vfsread_filtops = {
5172 	.f_isfd = 1,
5173 	.f_detach = filt_vfsdetach,
5174 	.f_event = filt_vfsread
5175 };
5176 static struct filterops vfswrite_filtops = {
5177 	.f_isfd = 1,
5178 	.f_detach = filt_vfsdetach,
5179 	.f_event = filt_vfswrite
5180 };
5181 static struct filterops vfsvnode_filtops = {
5182 	.f_isfd = 1,
5183 	.f_detach = filt_vfsdetach,
5184 	.f_event = filt_vfsvnode
5185 };
5186 
5187 static void
5188 vfs_knllock(void *arg)
5189 {
5190 	struct vnode *vp = arg;
5191 
5192 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5193 }
5194 
5195 static void
5196 vfs_knlunlock(void *arg)
5197 {
5198 	struct vnode *vp = arg;
5199 
5200 	VOP_UNLOCK(vp, 0);
5201 }
5202 
5203 static void
5204 vfs_knl_assert_locked(void *arg)
5205 {
5206 #ifdef DEBUG_VFS_LOCKS
5207 	struct vnode *vp = arg;
5208 
5209 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5210 #endif
5211 }
5212 
5213 static void
5214 vfs_knl_assert_unlocked(void *arg)
5215 {
5216 #ifdef DEBUG_VFS_LOCKS
5217 	struct vnode *vp = arg;
5218 
5219 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5220 #endif
5221 }
5222 
5223 int
5224 vfs_kqfilter(struct vop_kqfilter_args *ap)
5225 {
5226 	struct vnode *vp = ap->a_vp;
5227 	struct knote *kn = ap->a_kn;
5228 	struct knlist *knl;
5229 
5230 	switch (kn->kn_filter) {
5231 	case EVFILT_READ:
5232 		kn->kn_fop = &vfsread_filtops;
5233 		break;
5234 	case EVFILT_WRITE:
5235 		kn->kn_fop = &vfswrite_filtops;
5236 		break;
5237 	case EVFILT_VNODE:
5238 		kn->kn_fop = &vfsvnode_filtops;
5239 		break;
5240 	default:
5241 		return (EINVAL);
5242 	}
5243 
5244 	kn->kn_hook = (caddr_t)vp;
5245 
5246 	v_addpollinfo(vp);
5247 	if (vp->v_pollinfo == NULL)
5248 		return (ENOMEM);
5249 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5250 	vhold(vp);
5251 	knlist_add(knl, kn, 0);
5252 
5253 	return (0);
5254 }
5255 
5256 /*
5257  * Detach knote from vnode
5258  */
5259 static void
5260 filt_vfsdetach(struct knote *kn)
5261 {
5262 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5263 
5264 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5265 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5266 	vdrop(vp);
5267 }
5268 
5269 /*ARGSUSED*/
5270 static int
5271 filt_vfsread(struct knote *kn, long hint)
5272 {
5273 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5274 	struct vattr va;
5275 	int res;
5276 
5277 	/*
5278 	 * filesystem is gone, so set the EOF flag and schedule
5279 	 * the knote for deletion.
5280 	 */
5281 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5282 		VI_LOCK(vp);
5283 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5284 		VI_UNLOCK(vp);
5285 		return (1);
5286 	}
5287 
5288 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5289 		return (0);
5290 
5291 	VI_LOCK(vp);
5292 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5293 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5294 	VI_UNLOCK(vp);
5295 	return (res);
5296 }
5297 
5298 /*ARGSUSED*/
5299 static int
5300 filt_vfswrite(struct knote *kn, long hint)
5301 {
5302 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5303 
5304 	VI_LOCK(vp);
5305 
5306 	/*
5307 	 * filesystem is gone, so set the EOF flag and schedule
5308 	 * the knote for deletion.
5309 	 */
5310 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5311 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5312 
5313 	kn->kn_data = 0;
5314 	VI_UNLOCK(vp);
5315 	return (1);
5316 }
5317 
5318 static int
5319 filt_vfsvnode(struct knote *kn, long hint)
5320 {
5321 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5322 	int res;
5323 
5324 	VI_LOCK(vp);
5325 	if (kn->kn_sfflags & hint)
5326 		kn->kn_fflags |= hint;
5327 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5328 		kn->kn_flags |= EV_EOF;
5329 		VI_UNLOCK(vp);
5330 		return (1);
5331 	}
5332 	res = (kn->kn_fflags != 0);
5333 	VI_UNLOCK(vp);
5334 	return (res);
5335 }
5336 
5337 int
5338 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5339 {
5340 	int error;
5341 
5342 	if (dp->d_reclen > ap->a_uio->uio_resid)
5343 		return (ENAMETOOLONG);
5344 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5345 	if (error) {
5346 		if (ap->a_ncookies != NULL) {
5347 			if (ap->a_cookies != NULL)
5348 				free(ap->a_cookies, M_TEMP);
5349 			ap->a_cookies = NULL;
5350 			*ap->a_ncookies = 0;
5351 		}
5352 		return (error);
5353 	}
5354 	if (ap->a_ncookies == NULL)
5355 		return (0);
5356 
5357 	KASSERT(ap->a_cookies,
5358 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5359 
5360 	*ap->a_cookies = realloc(*ap->a_cookies,
5361 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5362 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5363 	*ap->a_ncookies += 1;
5364 	return (0);
5365 }
5366 
5367 /*
5368  * Mark for update the access time of the file if the filesystem
5369  * supports VOP_MARKATIME.  This functionality is used by execve and
5370  * mmap, so we want to avoid the I/O implied by directly setting
5371  * va_atime for the sake of efficiency.
5372  */
5373 void
5374 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5375 {
5376 	struct mount *mp;
5377 
5378 	mp = vp->v_mount;
5379 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5380 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5381 		(void)VOP_MARKATIME(vp);
5382 }
5383 
5384 /*
5385  * The purpose of this routine is to remove granularity from accmode_t,
5386  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5387  * VADMIN and VAPPEND.
5388  *
5389  * If it returns 0, the caller is supposed to continue with the usual
5390  * access checks using 'accmode' as modified by this routine.  If it
5391  * returns nonzero value, the caller is supposed to return that value
5392  * as errno.
5393  *
5394  * Note that after this routine runs, accmode may be zero.
5395  */
5396 int
5397 vfs_unixify_accmode(accmode_t *accmode)
5398 {
5399 	/*
5400 	 * There is no way to specify explicit "deny" rule using
5401 	 * file mode or POSIX.1e ACLs.
5402 	 */
5403 	if (*accmode & VEXPLICIT_DENY) {
5404 		*accmode = 0;
5405 		return (0);
5406 	}
5407 
5408 	/*
5409 	 * None of these can be translated into usual access bits.
5410 	 * Also, the common case for NFSv4 ACLs is to not contain
5411 	 * either of these bits. Caller should check for VWRITE
5412 	 * on the containing directory instead.
5413 	 */
5414 	if (*accmode & (VDELETE_CHILD | VDELETE))
5415 		return (EPERM);
5416 
5417 	if (*accmode & VADMIN_PERMS) {
5418 		*accmode &= ~VADMIN_PERMS;
5419 		*accmode |= VADMIN;
5420 	}
5421 
5422 	/*
5423 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5424 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5425 	 */
5426 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5427 
5428 	return (0);
5429 }
5430 
5431 /*
5432  * These are helper functions for filesystems to traverse all
5433  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5434  *
5435  * This interface replaces MNT_VNODE_FOREACH.
5436  */
5437 
5438 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5439 
5440 struct vnode *
5441 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5442 {
5443 	struct vnode *vp;
5444 
5445 	if (should_yield())
5446 		kern_yield(PRI_USER);
5447 	MNT_ILOCK(mp);
5448 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5449 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5450 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5451 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5452 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5453 			continue;
5454 		VI_LOCK(vp);
5455 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5456 			VI_UNLOCK(vp);
5457 			continue;
5458 		}
5459 		break;
5460 	}
5461 	if (vp == NULL) {
5462 		__mnt_vnode_markerfree_all(mvp, mp);
5463 		/* MNT_IUNLOCK(mp); -- done in above function */
5464 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5465 		return (NULL);
5466 	}
5467 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5468 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5469 	MNT_IUNLOCK(mp);
5470 	return (vp);
5471 }
5472 
5473 struct vnode *
5474 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5475 {
5476 	struct vnode *vp;
5477 
5478 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5479 	MNT_ILOCK(mp);
5480 	MNT_REF(mp);
5481 	(*mvp)->v_mount = mp;
5482 	(*mvp)->v_type = VMARKER;
5483 
5484 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5485 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5486 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5487 			continue;
5488 		VI_LOCK(vp);
5489 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5490 			VI_UNLOCK(vp);
5491 			continue;
5492 		}
5493 		break;
5494 	}
5495 	if (vp == NULL) {
5496 		MNT_REL(mp);
5497 		MNT_IUNLOCK(mp);
5498 		free(*mvp, M_VNODE_MARKER);
5499 		*mvp = NULL;
5500 		return (NULL);
5501 	}
5502 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5503 	MNT_IUNLOCK(mp);
5504 	return (vp);
5505 }
5506 
5507 void
5508 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5509 {
5510 
5511 	if (*mvp == NULL) {
5512 		MNT_IUNLOCK(mp);
5513 		return;
5514 	}
5515 
5516 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5517 
5518 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5519 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5520 	MNT_REL(mp);
5521 	MNT_IUNLOCK(mp);
5522 	free(*mvp, M_VNODE_MARKER);
5523 	*mvp = NULL;
5524 }
5525 
5526 /*
5527  * These are helper functions for filesystems to traverse their
5528  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5529  */
5530 static void
5531 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5532 {
5533 
5534 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5535 
5536 	MNT_ILOCK(mp);
5537 	MNT_REL(mp);
5538 	MNT_IUNLOCK(mp);
5539 	free(*mvp, M_VNODE_MARKER);
5540 	*mvp = NULL;
5541 }
5542 
5543 /*
5544  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5545  * conventional lock order during mnt_vnode_next_active iteration.
5546  *
5547  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5548  * The list lock is dropped and reacquired.  On success, both locks are held.
5549  * On failure, the mount vnode list lock is held but the vnode interlock is
5550  * not, and the procedure may have yielded.
5551  */
5552 static bool
5553 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5554     struct vnode *vp)
5555 {
5556 	const struct vnode *tmp;
5557 	bool held, ret;
5558 
5559 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5560 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5561 	    ("%s: bad marker", __func__));
5562 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5563 	    ("%s: inappropriate vnode", __func__));
5564 	ASSERT_VI_UNLOCKED(vp, __func__);
5565 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5566 
5567 	ret = false;
5568 
5569 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5570 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5571 
5572 	/*
5573 	 * Use a hold to prevent vp from disappearing while the mount vnode
5574 	 * list lock is dropped and reacquired.  Normally a hold would be
5575 	 * acquired with vhold(), but that might try to acquire the vnode
5576 	 * interlock, which would be a LOR with the mount vnode list lock.
5577 	 */
5578 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5579 	mtx_unlock(&mp->mnt_listmtx);
5580 	if (!held)
5581 		goto abort;
5582 	VI_LOCK(vp);
5583 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5584 		vdropl(vp);
5585 		goto abort;
5586 	}
5587 	mtx_lock(&mp->mnt_listmtx);
5588 
5589 	/*
5590 	 * Determine whether the vnode is still the next one after the marker,
5591 	 * excepting any other markers.  If the vnode has not been doomed by
5592 	 * vgone() then the hold should have ensured that it remained on the
5593 	 * active list.  If it has been doomed but is still on the active list,
5594 	 * don't abort, but rather skip over it (avoid spinning on doomed
5595 	 * vnodes).
5596 	 */
5597 	tmp = mvp;
5598 	do {
5599 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5600 	} while (tmp != NULL && tmp->v_type == VMARKER);
5601 	if (tmp != vp) {
5602 		mtx_unlock(&mp->mnt_listmtx);
5603 		VI_UNLOCK(vp);
5604 		goto abort;
5605 	}
5606 
5607 	ret = true;
5608 	goto out;
5609 abort:
5610 	maybe_yield();
5611 	mtx_lock(&mp->mnt_listmtx);
5612 out:
5613 	if (ret)
5614 		ASSERT_VI_LOCKED(vp, __func__);
5615 	else
5616 		ASSERT_VI_UNLOCKED(vp, __func__);
5617 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5618 	return (ret);
5619 }
5620 
5621 static struct vnode *
5622 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5623 {
5624 	struct vnode *vp, *nvp;
5625 
5626 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5627 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5628 restart:
5629 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5630 	while (vp != NULL) {
5631 		if (vp->v_type == VMARKER) {
5632 			vp = TAILQ_NEXT(vp, v_actfreelist);
5633 			continue;
5634 		}
5635 		/*
5636 		 * Try-lock because this is the wrong lock order.  If that does
5637 		 * not succeed, drop the mount vnode list lock and try to
5638 		 * reacquire it and the vnode interlock in the right order.
5639 		 */
5640 		if (!VI_TRYLOCK(vp) &&
5641 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5642 			goto restart;
5643 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5644 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5645 		    ("alien vnode on the active list %p %p", vp, mp));
5646 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5647 			break;
5648 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5649 		VI_UNLOCK(vp);
5650 		vp = nvp;
5651 	}
5652 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5653 
5654 	/* Check if we are done */
5655 	if (vp == NULL) {
5656 		mtx_unlock(&mp->mnt_listmtx);
5657 		mnt_vnode_markerfree_active(mvp, mp);
5658 		return (NULL);
5659 	}
5660 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5661 	mtx_unlock(&mp->mnt_listmtx);
5662 	ASSERT_VI_LOCKED(vp, "active iter");
5663 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5664 	return (vp);
5665 }
5666 
5667 struct vnode *
5668 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5669 {
5670 
5671 	if (should_yield())
5672 		kern_yield(PRI_USER);
5673 	mtx_lock(&mp->mnt_listmtx);
5674 	return (mnt_vnode_next_active(mvp, mp));
5675 }
5676 
5677 struct vnode *
5678 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5679 {
5680 	struct vnode *vp;
5681 
5682 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5683 	MNT_ILOCK(mp);
5684 	MNT_REF(mp);
5685 	MNT_IUNLOCK(mp);
5686 	(*mvp)->v_type = VMARKER;
5687 	(*mvp)->v_mount = mp;
5688 
5689 	mtx_lock(&mp->mnt_listmtx);
5690 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5691 	if (vp == NULL) {
5692 		mtx_unlock(&mp->mnt_listmtx);
5693 		mnt_vnode_markerfree_active(mvp, mp);
5694 		return (NULL);
5695 	}
5696 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5697 	return (mnt_vnode_next_active(mvp, mp));
5698 }
5699 
5700 void
5701 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5702 {
5703 
5704 	if (*mvp == NULL)
5705 		return;
5706 
5707 	mtx_lock(&mp->mnt_listmtx);
5708 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5709 	mtx_unlock(&mp->mnt_listmtx);
5710 	mnt_vnode_markerfree_active(mvp, mp);
5711 }
5712