xref: /freebsd/sys/kern/vfs_subr.c (revision a0eb270b125e75d3e36c2533749c4c196d4dbc39)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_compat.h"
45 #include "opt_ddb.h"
46 #include "opt_watchdog.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/condvar.h>
53 #include <sys/conf.h>
54 #include <sys/dirent.h>
55 #include <sys/event.h>
56 #include <sys/eventhandler.h>
57 #include <sys/extattr.h>
58 #include <sys/file.h>
59 #include <sys/fcntl.h>
60 #include <sys/jail.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/kthread.h>
64 #include <sys/lockf.h>
65 #include <sys/malloc.h>
66 #include <sys/mount.h>
67 #include <sys/namei.h>
68 #include <sys/pctrie.h>
69 #include <sys/priv.h>
70 #include <sys/reboot.h>
71 #include <sys/refcount.h>
72 #include <sys/rwlock.h>
73 #include <sys/sched.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/smp.h>
76 #include <sys/stat.h>
77 #include <sys/sysctl.h>
78 #include <sys/syslog.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 
83 #include <machine/stdarg.h>
84 
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_extern.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_kern.h>
94 #include <vm/uma.h>
95 
96 #ifdef DDB
97 #include <ddb/ddb.h>
98 #endif
99 
100 static void	delmntque(struct vnode *vp);
101 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
102 		    int slpflag, int slptimeo);
103 static void	syncer_shutdown(void *arg, int howto);
104 static int	vtryrecycle(struct vnode *vp);
105 static void	v_init_counters(struct vnode *);
106 static void	v_incr_usecount(struct vnode *);
107 static void	v_incr_devcount(struct vnode *);
108 static void	v_decr_devcount(struct vnode *);
109 static void	vnlru_free(int);
110 static void	vgonel(struct vnode *);
111 static void	vfs_knllock(void *arg);
112 static void	vfs_knlunlock(void *arg);
113 static void	vfs_knl_assert_locked(void *arg);
114 static void	vfs_knl_assert_unlocked(void *arg);
115 static void	destroy_vpollinfo(struct vpollinfo *vi);
116 
117 /*
118  * Number of vnodes in existence.  Increased whenever getnewvnode()
119  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
120  */
121 static unsigned long	numvnodes;
122 
123 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
124     "Number of vnodes in existence");
125 
126 static u_long vnodes_created;
127 SYSCTL_ULONG(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
128     0, "Number of vnodes created by getnewvnode");
129 
130 /*
131  * Conversion tables for conversion from vnode types to inode formats
132  * and back.
133  */
134 enum vtype iftovt_tab[16] = {
135 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
136 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
137 };
138 int vttoif_tab[10] = {
139 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
140 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
141 };
142 
143 /*
144  * List of vnodes that are ready for recycling.
145  */
146 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
147 
148 /*
149  * "Free" vnode target.  Free vnodes are rarely completely free, but are
150  * just ones that are cheap to recycle.  Usually they are for files which
151  * have been stat'd but not read; these usually have inode and namecache
152  * data attached to them.  This target is the preferred minimum size of a
153  * sub-cache consisting mostly of such files. The system balances the size
154  * of this sub-cache with its complement to try to prevent either from
155  * thrashing while the other is relatively inactive.  The targets express
156  * a preference for the best balance.
157  *
158  * "Above" this target there are 2 further targets (watermarks) related
159  * to recyling of free vnodes.  In the best-operating case, the cache is
160  * exactly full, the free list has size between vlowat and vhiwat above the
161  * free target, and recycling from it and normal use maintains this state.
162  * Sometimes the free list is below vlowat or even empty, but this state
163  * is even better for immediate use provided the cache is not full.
164  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
165  * ones) to reach one of these states.  The watermarks are currently hard-
166  * coded as 4% and 9% of the available space higher.  These and the default
167  * of 25% for wantfreevnodes are too large if the memory size is large.
168  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
169  * whenever vnlru_proc() becomes active.
170  */
171 static u_long wantfreevnodes;
172 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
173     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
174 static u_long freevnodes;
175 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
176     &freevnodes, 0, "Number of \"free\" vnodes");
177 
178 static u_long recycles_count;
179 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 0,
180     "Number of vnodes recycled to meet vnode cache targets");
181 
182 /*
183  * Various variables used for debugging the new implementation of
184  * reassignbuf().
185  * XXX these are probably of (very) limited utility now.
186  */
187 static int reassignbufcalls;
188 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
189     "Number of calls to reassignbuf");
190 
191 static u_long free_owe_inact;
192 SYSCTL_ULONG(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact, 0,
193     "Number of times free vnodes kept on active list due to VFS "
194     "owing inactivation");
195 
196 /* To keep more than one thread at a time from running vfs_getnewfsid */
197 static struct mtx mntid_mtx;
198 
199 /*
200  * Lock for any access to the following:
201  *	vnode_free_list
202  *	numvnodes
203  *	freevnodes
204  */
205 static struct mtx vnode_free_list_mtx;
206 
207 /* Publicly exported FS */
208 struct nfs_public nfs_pub;
209 
210 static uma_zone_t buf_trie_zone;
211 
212 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
213 static uma_zone_t vnode_zone;
214 static uma_zone_t vnodepoll_zone;
215 
216 /*
217  * The workitem queue.
218  *
219  * It is useful to delay writes of file data and filesystem metadata
220  * for tens of seconds so that quickly created and deleted files need
221  * not waste disk bandwidth being created and removed. To realize this,
222  * we append vnodes to a "workitem" queue. When running with a soft
223  * updates implementation, most pending metadata dependencies should
224  * not wait for more than a few seconds. Thus, mounted on block devices
225  * are delayed only about a half the time that file data is delayed.
226  * Similarly, directory updates are more critical, so are only delayed
227  * about a third the time that file data is delayed. Thus, there are
228  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
229  * one each second (driven off the filesystem syncer process). The
230  * syncer_delayno variable indicates the next queue that is to be processed.
231  * Items that need to be processed soon are placed in this queue:
232  *
233  *	syncer_workitem_pending[syncer_delayno]
234  *
235  * A delay of fifteen seconds is done by placing the request fifteen
236  * entries later in the queue:
237  *
238  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
239  *
240  */
241 static int syncer_delayno;
242 static long syncer_mask;
243 LIST_HEAD(synclist, bufobj);
244 static struct synclist *syncer_workitem_pending;
245 /*
246  * The sync_mtx protects:
247  *	bo->bo_synclist
248  *	sync_vnode_count
249  *	syncer_delayno
250  *	syncer_state
251  *	syncer_workitem_pending
252  *	syncer_worklist_len
253  *	rushjob
254  */
255 static struct mtx sync_mtx;
256 static struct cv sync_wakeup;
257 
258 #define SYNCER_MAXDELAY		32
259 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
260 static int syncdelay = 30;		/* max time to delay syncing data */
261 static int filedelay = 30;		/* time to delay syncing files */
262 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
263     "Time to delay syncing files (in seconds)");
264 static int dirdelay = 29;		/* time to delay syncing directories */
265 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
266     "Time to delay syncing directories (in seconds)");
267 static int metadelay = 28;		/* time to delay syncing metadata */
268 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
269     "Time to delay syncing metadata (in seconds)");
270 static int rushjob;		/* number of slots to run ASAP */
271 static int stat_rush_requests;	/* number of times I/O speeded up */
272 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
273     "Number of times I/O speeded up (rush requests)");
274 
275 /*
276  * When shutting down the syncer, run it at four times normal speed.
277  */
278 #define SYNCER_SHUTDOWN_SPEEDUP		4
279 static int sync_vnode_count;
280 static int syncer_worklist_len;
281 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
282     syncer_state;
283 
284 /* Target for maximum number of vnodes. */
285 int desiredvnodes;
286 static int gapvnodes;		/* gap between wanted and desired */
287 static int vhiwat;		/* enough extras after expansion */
288 static int vlowat;		/* minimal extras before expansion */
289 static int vstir;		/* nonzero to stir non-free vnodes */
290 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
291 
292 static int
293 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
294 {
295 	int error, old_desiredvnodes;
296 
297 	old_desiredvnodes = desiredvnodes;
298 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
299 		return (error);
300 	if (old_desiredvnodes != desiredvnodes) {
301 		wantfreevnodes = desiredvnodes / 4;
302 		/* XXX locking seems to be incomplete. */
303 		vfs_hash_changesize(desiredvnodes);
304 		cache_changesize(desiredvnodes);
305 	}
306 	return (0);
307 }
308 
309 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
310     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
311     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
312 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
313     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
314 static int vnlru_nowhere;
315 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
316     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
317 
318 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
319 static int vnsz2log;
320 
321 /*
322  * Support for the bufobj clean & dirty pctrie.
323  */
324 static void *
325 buf_trie_alloc(struct pctrie *ptree)
326 {
327 
328 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
329 }
330 
331 static void
332 buf_trie_free(struct pctrie *ptree, void *node)
333 {
334 
335 	uma_zfree(buf_trie_zone, node);
336 }
337 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
338 
339 /*
340  * Initialize the vnode management data structures.
341  *
342  * Reevaluate the following cap on the number of vnodes after the physical
343  * memory size exceeds 512GB.  In the limit, as the physical memory size
344  * grows, the ratio of the memory size in KB to to vnodes approaches 64:1.
345  */
346 #ifndef	MAXVNODES_MAX
347 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
348 #endif
349 
350 /*
351  * Initialize a vnode as it first enters the zone.
352  */
353 static int
354 vnode_init(void *mem, int size, int flags)
355 {
356 	struct vnode *vp;
357 	struct bufobj *bo;
358 
359 	vp = mem;
360 	bzero(vp, size);
361 	/*
362 	 * Setup locks.
363 	 */
364 	vp->v_vnlock = &vp->v_lock;
365 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
366 	/*
367 	 * By default, don't allow shared locks unless filesystems opt-in.
368 	 */
369 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
370 	    LK_NOSHARE | LK_IS_VNODE);
371 	/*
372 	 * Initialize bufobj.
373 	 */
374 	bo = &vp->v_bufobj;
375 	bo->__bo_vnode = vp;
376 	rw_init(BO_LOCKPTR(bo), "bufobj interlock");
377 	bo->bo_private = vp;
378 	TAILQ_INIT(&bo->bo_clean.bv_hd);
379 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
380 	/*
381 	 * Initialize namecache.
382 	 */
383 	LIST_INIT(&vp->v_cache_src);
384 	TAILQ_INIT(&vp->v_cache_dst);
385 	/*
386 	 * Initialize rangelocks.
387 	 */
388 	rangelock_init(&vp->v_rl);
389 	return (0);
390 }
391 
392 /*
393  * Free a vnode when it is cleared from the zone.
394  */
395 static void
396 vnode_fini(void *mem, int size)
397 {
398 	struct vnode *vp;
399 	struct bufobj *bo;
400 
401 	vp = mem;
402 	rangelock_destroy(&vp->v_rl);
403 	lockdestroy(vp->v_vnlock);
404 	mtx_destroy(&vp->v_interlock);
405 	bo = &vp->v_bufobj;
406 	rw_destroy(BO_LOCKPTR(bo));
407 }
408 
409 static void
410 vntblinit(void *dummy __unused)
411 {
412 	u_int i;
413 	int physvnodes, virtvnodes;
414 
415 	/*
416 	 * Desiredvnodes is a function of the physical memory size and the
417 	 * kernel's heap size.  Generally speaking, it scales with the
418 	 * physical memory size.  The ratio of desiredvnodes to the physical
419 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
420 	 * Thereafter, the
421 	 * marginal ratio of desiredvnodes to the physical memory size is
422 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
423 	 * size.  The memory required by desiredvnodes vnodes and vm objects
424 	 * must not exceed 1/7th of the kernel's heap size.
425 	 */
426 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
427 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
428 	virtvnodes = vm_kmem_size / (7 * (sizeof(struct vm_object) +
429 	    sizeof(struct vnode)));
430 	desiredvnodes = min(physvnodes, virtvnodes);
431 	if (desiredvnodes > MAXVNODES_MAX) {
432 		if (bootverbose)
433 			printf("Reducing kern.maxvnodes %d -> %d\n",
434 			    desiredvnodes, MAXVNODES_MAX);
435 		desiredvnodes = MAXVNODES_MAX;
436 	}
437 	wantfreevnodes = desiredvnodes / 4;
438 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
439 	TAILQ_INIT(&vnode_free_list);
440 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
441 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
442 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
443 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
444 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
445 	/*
446 	 * Preallocate enough nodes to support one-per buf so that
447 	 * we can not fail an insert.  reassignbuf() callers can not
448 	 * tolerate the insertion failure.
449 	 */
450 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
451 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
452 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
453 	uma_prealloc(buf_trie_zone, nbuf);
454 	/*
455 	 * Initialize the filesystem syncer.
456 	 */
457 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
458 	    &syncer_mask);
459 	syncer_maxdelay = syncer_mask + 1;
460 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
461 	cv_init(&sync_wakeup, "syncer");
462 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
463 		vnsz2log++;
464 	vnsz2log--;
465 }
466 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
467 
468 
469 /*
470  * Mark a mount point as busy. Used to synchronize access and to delay
471  * unmounting. Eventually, mountlist_mtx is not released on failure.
472  *
473  * vfs_busy() is a custom lock, it can block the caller.
474  * vfs_busy() only sleeps if the unmount is active on the mount point.
475  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
476  * vnode belonging to mp.
477  *
478  * Lookup uses vfs_busy() to traverse mount points.
479  * root fs			var fs
480  * / vnode lock		A	/ vnode lock (/var)		D
481  * /var vnode lock	B	/log vnode lock(/var/log)	E
482  * vfs_busy lock	C	vfs_busy lock			F
483  *
484  * Within each file system, the lock order is C->A->B and F->D->E.
485  *
486  * When traversing across mounts, the system follows that lock order:
487  *
488  *        C->A->B
489  *              |
490  *              +->F->D->E
491  *
492  * The lookup() process for namei("/var") illustrates the process:
493  *  VOP_LOOKUP() obtains B while A is held
494  *  vfs_busy() obtains a shared lock on F while A and B are held
495  *  vput() releases lock on B
496  *  vput() releases lock on A
497  *  VFS_ROOT() obtains lock on D while shared lock on F is held
498  *  vfs_unbusy() releases shared lock on F
499  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
500  *    Attempt to lock A (instead of vp_crossmp) while D is held would
501  *    violate the global order, causing deadlocks.
502  *
503  * dounmount() locks B while F is drained.
504  */
505 int
506 vfs_busy(struct mount *mp, int flags)
507 {
508 
509 	MPASS((flags & ~MBF_MASK) == 0);
510 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
511 
512 	MNT_ILOCK(mp);
513 	MNT_REF(mp);
514 	/*
515 	 * If mount point is currenly being unmounted, sleep until the
516 	 * mount point fate is decided.  If thread doing the unmounting fails,
517 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
518 	 * that this mount point has survived the unmount attempt and vfs_busy
519 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
520 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
521 	 * about to be really destroyed.  vfs_busy needs to release its
522 	 * reference on the mount point in this case and return with ENOENT,
523 	 * telling the caller that mount mount it tried to busy is no longer
524 	 * valid.
525 	 */
526 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
527 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
528 			MNT_REL(mp);
529 			MNT_IUNLOCK(mp);
530 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
531 			    __func__);
532 			return (ENOENT);
533 		}
534 		if (flags & MBF_MNTLSTLOCK)
535 			mtx_unlock(&mountlist_mtx);
536 		mp->mnt_kern_flag |= MNTK_MWAIT;
537 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
538 		if (flags & MBF_MNTLSTLOCK)
539 			mtx_lock(&mountlist_mtx);
540 		MNT_ILOCK(mp);
541 	}
542 	if (flags & MBF_MNTLSTLOCK)
543 		mtx_unlock(&mountlist_mtx);
544 	mp->mnt_lockref++;
545 	MNT_IUNLOCK(mp);
546 	return (0);
547 }
548 
549 /*
550  * Free a busy filesystem.
551  */
552 void
553 vfs_unbusy(struct mount *mp)
554 {
555 
556 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
557 	MNT_ILOCK(mp);
558 	MNT_REL(mp);
559 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
560 	mp->mnt_lockref--;
561 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
562 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
563 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
564 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
565 		wakeup(&mp->mnt_lockref);
566 	}
567 	MNT_IUNLOCK(mp);
568 }
569 
570 /*
571  * Lookup a mount point by filesystem identifier.
572  */
573 struct mount *
574 vfs_getvfs(fsid_t *fsid)
575 {
576 	struct mount *mp;
577 
578 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
579 	mtx_lock(&mountlist_mtx);
580 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
581 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
582 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
583 			vfs_ref(mp);
584 			mtx_unlock(&mountlist_mtx);
585 			return (mp);
586 		}
587 	}
588 	mtx_unlock(&mountlist_mtx);
589 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
590 	return ((struct mount *) 0);
591 }
592 
593 /*
594  * Lookup a mount point by filesystem identifier, busying it before
595  * returning.
596  *
597  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
598  * cache for popular filesystem identifiers.  The cache is lockess, using
599  * the fact that struct mount's are never freed.  In worst case we may
600  * get pointer to unmounted or even different filesystem, so we have to
601  * check what we got, and go slow way if so.
602  */
603 struct mount *
604 vfs_busyfs(fsid_t *fsid)
605 {
606 #define	FSID_CACHE_SIZE	256
607 	typedef struct mount * volatile vmp_t;
608 	static vmp_t cache[FSID_CACHE_SIZE];
609 	struct mount *mp;
610 	int error;
611 	uint32_t hash;
612 
613 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
614 	hash = fsid->val[0] ^ fsid->val[1];
615 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
616 	mp = cache[hash];
617 	if (mp == NULL ||
618 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
619 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
620 		goto slow;
621 	if (vfs_busy(mp, 0) != 0) {
622 		cache[hash] = NULL;
623 		goto slow;
624 	}
625 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
626 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
627 		return (mp);
628 	else
629 	    vfs_unbusy(mp);
630 
631 slow:
632 	mtx_lock(&mountlist_mtx);
633 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
634 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
635 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
636 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
637 			if (error) {
638 				cache[hash] = NULL;
639 				mtx_unlock(&mountlist_mtx);
640 				return (NULL);
641 			}
642 			cache[hash] = mp;
643 			return (mp);
644 		}
645 	}
646 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
647 	mtx_unlock(&mountlist_mtx);
648 	return ((struct mount *) 0);
649 }
650 
651 /*
652  * Check if a user can access privileged mount options.
653  */
654 int
655 vfs_suser(struct mount *mp, struct thread *td)
656 {
657 	int error;
658 
659 	/*
660 	 * If the thread is jailed, but this is not a jail-friendly file
661 	 * system, deny immediately.
662 	 */
663 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
664 		return (EPERM);
665 
666 	/*
667 	 * If the file system was mounted outside the jail of the calling
668 	 * thread, deny immediately.
669 	 */
670 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
671 		return (EPERM);
672 
673 	/*
674 	 * If file system supports delegated administration, we don't check
675 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
676 	 * by the file system itself.
677 	 * If this is not the user that did original mount, we check for
678 	 * the PRIV_VFS_MOUNT_OWNER privilege.
679 	 */
680 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
681 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
682 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
683 			return (error);
684 	}
685 	return (0);
686 }
687 
688 /*
689  * Get a new unique fsid.  Try to make its val[0] unique, since this value
690  * will be used to create fake device numbers for stat().  Also try (but
691  * not so hard) make its val[0] unique mod 2^16, since some emulators only
692  * support 16-bit device numbers.  We end up with unique val[0]'s for the
693  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
694  *
695  * Keep in mind that several mounts may be running in parallel.  Starting
696  * the search one past where the previous search terminated is both a
697  * micro-optimization and a defense against returning the same fsid to
698  * different mounts.
699  */
700 void
701 vfs_getnewfsid(struct mount *mp)
702 {
703 	static uint16_t mntid_base;
704 	struct mount *nmp;
705 	fsid_t tfsid;
706 	int mtype;
707 
708 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
709 	mtx_lock(&mntid_mtx);
710 	mtype = mp->mnt_vfc->vfc_typenum;
711 	tfsid.val[1] = mtype;
712 	mtype = (mtype & 0xFF) << 24;
713 	for (;;) {
714 		tfsid.val[0] = makedev(255,
715 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
716 		mntid_base++;
717 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
718 			break;
719 		vfs_rel(nmp);
720 	}
721 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
722 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
723 	mtx_unlock(&mntid_mtx);
724 }
725 
726 /*
727  * Knob to control the precision of file timestamps:
728  *
729  *   0 = seconds only; nanoseconds zeroed.
730  *   1 = seconds and nanoseconds, accurate within 1/HZ.
731  *   2 = seconds and nanoseconds, truncated to microseconds.
732  * >=3 = seconds and nanoseconds, maximum precision.
733  */
734 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
735 
736 static int timestamp_precision = TSP_USEC;
737 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
738     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
739     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, "
740     "3+: sec + ns (max. precision))");
741 
742 /*
743  * Get a current timestamp.
744  */
745 void
746 vfs_timestamp(struct timespec *tsp)
747 {
748 	struct timeval tv;
749 
750 	switch (timestamp_precision) {
751 	case TSP_SEC:
752 		tsp->tv_sec = time_second;
753 		tsp->tv_nsec = 0;
754 		break;
755 	case TSP_HZ:
756 		getnanotime(tsp);
757 		break;
758 	case TSP_USEC:
759 		microtime(&tv);
760 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
761 		break;
762 	case TSP_NSEC:
763 	default:
764 		nanotime(tsp);
765 		break;
766 	}
767 }
768 
769 /*
770  * Set vnode attributes to VNOVAL
771  */
772 void
773 vattr_null(struct vattr *vap)
774 {
775 
776 	vap->va_type = VNON;
777 	vap->va_size = VNOVAL;
778 	vap->va_bytes = VNOVAL;
779 	vap->va_mode = VNOVAL;
780 	vap->va_nlink = VNOVAL;
781 	vap->va_uid = VNOVAL;
782 	vap->va_gid = VNOVAL;
783 	vap->va_fsid = VNOVAL;
784 	vap->va_fileid = VNOVAL;
785 	vap->va_blocksize = VNOVAL;
786 	vap->va_rdev = VNOVAL;
787 	vap->va_atime.tv_sec = VNOVAL;
788 	vap->va_atime.tv_nsec = VNOVAL;
789 	vap->va_mtime.tv_sec = VNOVAL;
790 	vap->va_mtime.tv_nsec = VNOVAL;
791 	vap->va_ctime.tv_sec = VNOVAL;
792 	vap->va_ctime.tv_nsec = VNOVAL;
793 	vap->va_birthtime.tv_sec = VNOVAL;
794 	vap->va_birthtime.tv_nsec = VNOVAL;
795 	vap->va_flags = VNOVAL;
796 	vap->va_gen = VNOVAL;
797 	vap->va_vaflags = 0;
798 }
799 
800 /*
801  * This routine is called when we have too many vnodes.  It attempts
802  * to free <count> vnodes and will potentially free vnodes that still
803  * have VM backing store (VM backing store is typically the cause
804  * of a vnode blowout so we want to do this).  Therefore, this operation
805  * is not considered cheap.
806  *
807  * A number of conditions may prevent a vnode from being reclaimed.
808  * the buffer cache may have references on the vnode, a directory
809  * vnode may still have references due to the namei cache representing
810  * underlying files, or the vnode may be in active use.   It is not
811  * desireable to reuse such vnodes.  These conditions may cause the
812  * number of vnodes to reach some minimum value regardless of what
813  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
814  */
815 static int
816 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
817 {
818 	struct vnode *vp;
819 	int count, done, target;
820 
821 	done = 0;
822 	vn_start_write(NULL, &mp, V_WAIT);
823 	MNT_ILOCK(mp);
824 	count = mp->mnt_nvnodelistsize;
825 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
826 	target = target / 10 + 1;
827 	while (count != 0 && done < target) {
828 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
829 		while (vp != NULL && vp->v_type == VMARKER)
830 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
831 		if (vp == NULL)
832 			break;
833 		/*
834 		 * XXX LRU is completely broken for non-free vnodes.  First
835 		 * by calling here in mountpoint order, then by moving
836 		 * unselected vnodes to the end here, and most grossly by
837 		 * removing the vlruvp() function that was supposed to
838 		 * maintain the order.  (This function was born broken
839 		 * since syncer problems prevented it doing anything.)  The
840 		 * order is closer to LRC (C = Created).
841 		 *
842 		 * LRU reclaiming of vnodes seems to have last worked in
843 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
844 		 * Then there was no hold count, and inactive vnodes were
845 		 * simply put on the free list in LRU order.  The separate
846 		 * lists also break LRU.  We prefer to reclaim from the
847 		 * free list for technical reasons.  This tends to thrash
848 		 * the free list to keep very unrecently used held vnodes.
849 		 * The problem is mitigated by keeping the free list large.
850 		 */
851 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
852 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
853 		--count;
854 		if (!VI_TRYLOCK(vp))
855 			goto next_iter;
856 		/*
857 		 * If it's been deconstructed already, it's still
858 		 * referenced, or it exceeds the trigger, skip it.
859 		 * Also skip free vnodes.  We are trying to make space
860 		 * to expand the free list, not reduce it.
861 		 */
862 		if (vp->v_usecount ||
863 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
864 		    ((vp->v_iflag & VI_FREE) != 0) ||
865 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
866 		    vp->v_object->resident_page_count > trigger)) {
867 			VI_UNLOCK(vp);
868 			goto next_iter;
869 		}
870 		MNT_IUNLOCK(mp);
871 		vholdl(vp);
872 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
873 			vdrop(vp);
874 			goto next_iter_mntunlocked;
875 		}
876 		VI_LOCK(vp);
877 		/*
878 		 * v_usecount may have been bumped after VOP_LOCK() dropped
879 		 * the vnode interlock and before it was locked again.
880 		 *
881 		 * It is not necessary to recheck VI_DOOMED because it can
882 		 * only be set by another thread that holds both the vnode
883 		 * lock and vnode interlock.  If another thread has the
884 		 * vnode lock before we get to VOP_LOCK() and obtains the
885 		 * vnode interlock after VOP_LOCK() drops the vnode
886 		 * interlock, the other thread will be unable to drop the
887 		 * vnode lock before our VOP_LOCK() call fails.
888 		 */
889 		if (vp->v_usecount ||
890 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
891 		    (vp->v_iflag & VI_FREE) != 0 ||
892 		    (vp->v_object != NULL &&
893 		    vp->v_object->resident_page_count > trigger)) {
894 			VOP_UNLOCK(vp, LK_INTERLOCK);
895 			vdrop(vp);
896 			goto next_iter_mntunlocked;
897 		}
898 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
899 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
900 		atomic_add_long(&recycles_count, 1);
901 		vgonel(vp);
902 		VOP_UNLOCK(vp, 0);
903 		vdropl(vp);
904 		done++;
905 next_iter_mntunlocked:
906 		if (!should_yield())
907 			goto relock_mnt;
908 		goto yield;
909 next_iter:
910 		if (!should_yield())
911 			continue;
912 		MNT_IUNLOCK(mp);
913 yield:
914 		kern_yield(PRI_USER);
915 relock_mnt:
916 		MNT_ILOCK(mp);
917 	}
918 	MNT_IUNLOCK(mp);
919 	vn_finished_write(mp);
920 	return done;
921 }
922 
923 /*
924  * Attempt to reduce the free list by the requested amount.
925  */
926 static void
927 vnlru_free(int count)
928 {
929 	struct vnode *vp;
930 
931 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
932 	for (; count > 0; count--) {
933 		vp = TAILQ_FIRST(&vnode_free_list);
934 		/*
935 		 * The list can be modified while the free_list_mtx
936 		 * has been dropped and vp could be NULL here.
937 		 */
938 		if (!vp)
939 			break;
940 		VNASSERT(vp->v_op != NULL, vp,
941 		    ("vnlru_free: vnode already reclaimed."));
942 		KASSERT((vp->v_iflag & VI_FREE) != 0,
943 		    ("Removing vnode not on freelist"));
944 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
945 		    ("Mangling active vnode"));
946 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
947 		/*
948 		 * Don't recycle if we can't get the interlock.
949 		 */
950 		if (!VI_TRYLOCK(vp)) {
951 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
952 			continue;
953 		}
954 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
955 		    vp, ("vp inconsistent on freelist"));
956 
957 		/*
958 		 * The clear of VI_FREE prevents activation of the
959 		 * vnode.  There is no sense in putting the vnode on
960 		 * the mount point active list, only to remove it
961 		 * later during recycling.  Inline the relevant part
962 		 * of vholdl(), to avoid triggering assertions or
963 		 * activating.
964 		 */
965 		freevnodes--;
966 		vp->v_iflag &= ~VI_FREE;
967 		refcount_acquire(&vp->v_holdcnt);
968 
969 		mtx_unlock(&vnode_free_list_mtx);
970 		VI_UNLOCK(vp);
971 		vtryrecycle(vp);
972 		/*
973 		 * If the recycled succeeded this vdrop will actually free
974 		 * the vnode.  If not it will simply place it back on
975 		 * the free list.
976 		 */
977 		vdrop(vp);
978 		mtx_lock(&vnode_free_list_mtx);
979 	}
980 }
981 
982 /* XXX some names and initialization are bad for limits and watermarks. */
983 static int
984 vspace(void)
985 {
986 	int space;
987 
988 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
989 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
990 	vlowat = vhiwat / 2;
991 	if (numvnodes > desiredvnodes)
992 		return (0);
993 	space = desiredvnodes - numvnodes;
994 	if (freevnodes > wantfreevnodes)
995 		space += freevnodes - wantfreevnodes;
996 	return (space);
997 }
998 
999 /*
1000  * Attempt to recycle vnodes in a context that is always safe to block.
1001  * Calling vlrurecycle() from the bowels of filesystem code has some
1002  * interesting deadlock problems.
1003  */
1004 static struct proc *vnlruproc;
1005 static int vnlruproc_sig;
1006 
1007 static void
1008 vnlru_proc(void)
1009 {
1010 	struct mount *mp, *nmp;
1011 	unsigned long ofreevnodes, onumvnodes;
1012 	int done, force, reclaim_nc_src, trigger, usevnodes;
1013 
1014 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1015 	    SHUTDOWN_PRI_FIRST);
1016 
1017 	force = 0;
1018 	for (;;) {
1019 		kproc_suspend_check(vnlruproc);
1020 		mtx_lock(&vnode_free_list_mtx);
1021 		/*
1022 		 * If numvnodes is too large (due to desiredvnodes being
1023 		 * adjusted using its sysctl, or emergency growth), first
1024 		 * try to reduce it by discarding from the free list.
1025 		 */
1026 		if (numvnodes > desiredvnodes && freevnodes > 0)
1027 			vnlru_free(ulmin(numvnodes - desiredvnodes,
1028 			    freevnodes));
1029 		/*
1030 		 * Sleep if the vnode cache is in a good state.  This is
1031 		 * when it is not over-full and has space for about a 4%
1032 		 * or 9% expansion (by growing its size or inexcessively
1033 		 * reducing its free list).  Otherwise, try to reclaim
1034 		 * space for a 10% expansion.
1035 		 */
1036 		if (vstir && force == 0) {
1037 			force = 1;
1038 			vstir = 0;
1039 		}
1040 		if (vspace() >= vlowat && force == 0) {
1041 			vnlruproc_sig = 0;
1042 			wakeup(&vnlruproc_sig);
1043 			msleep(vnlruproc, &vnode_free_list_mtx,
1044 			    PVFS|PDROP, "vlruwt", hz);
1045 			continue;
1046 		}
1047 		mtx_unlock(&vnode_free_list_mtx);
1048 		done = 0;
1049 		ofreevnodes = freevnodes;
1050 		onumvnodes = numvnodes;
1051 		/*
1052 		 * Calculate parameters for recycling.  These are the same
1053 		 * throughout the loop to give some semblance of fairness.
1054 		 * The trigger point is to avoid recycling vnodes with lots
1055 		 * of resident pages.  We aren't trying to free memory; we
1056 		 * are trying to recycle or at least free vnodes.
1057 		 */
1058 		if (numvnodes <= desiredvnodes)
1059 			usevnodes = numvnodes - freevnodes;
1060 		else
1061 			usevnodes = numvnodes;
1062 		if (usevnodes <= 0)
1063 			usevnodes = 1;
1064 		/*
1065 		 * The trigger value is is chosen to give a conservatively
1066 		 * large value to ensure that it alone doesn't prevent
1067 		 * making progress.  The value can easily be so large that
1068 		 * it is effectively infinite in some congested and
1069 		 * misconfigured cases, and this is necessary.  Normally
1070 		 * it is about 8 to 100 (pages), which is quite large.
1071 		 */
1072 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1073 		if (force < 2)
1074 			trigger = vsmalltrigger;
1075 		reclaim_nc_src = force >= 3;
1076 		mtx_lock(&mountlist_mtx);
1077 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1078 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1079 				nmp = TAILQ_NEXT(mp, mnt_list);
1080 				continue;
1081 			}
1082 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1083 			mtx_lock(&mountlist_mtx);
1084 			nmp = TAILQ_NEXT(mp, mnt_list);
1085 			vfs_unbusy(mp);
1086 		}
1087 		mtx_unlock(&mountlist_mtx);
1088 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1089 			uma_reclaim();
1090 		if (done == 0) {
1091 			if (force == 0 || force == 1) {
1092 				force = 2;
1093 				continue;
1094 			}
1095 			if (force == 2) {
1096 				force = 3;
1097 				continue;
1098 			}
1099 			force = 0;
1100 			vnlru_nowhere++;
1101 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1102 		} else
1103 			kern_yield(PRI_USER);
1104 		/*
1105 		 * After becoming active to expand above low water, keep
1106 		 * active until above high water.
1107 		 */
1108 		force = vspace() < vhiwat;
1109 	}
1110 }
1111 
1112 static struct kproc_desc vnlru_kp = {
1113 	"vnlru",
1114 	vnlru_proc,
1115 	&vnlruproc
1116 };
1117 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1118     &vnlru_kp);
1119 
1120 /*
1121  * Routines having to do with the management of the vnode table.
1122  */
1123 
1124 /*
1125  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1126  * before we actually vgone().  This function must be called with the vnode
1127  * held to prevent the vnode from being returned to the free list midway
1128  * through vgone().
1129  */
1130 static int
1131 vtryrecycle(struct vnode *vp)
1132 {
1133 	struct mount *vnmp;
1134 
1135 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1136 	VNASSERT(vp->v_holdcnt, vp,
1137 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1138 	/*
1139 	 * This vnode may found and locked via some other list, if so we
1140 	 * can't recycle it yet.
1141 	 */
1142 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1143 		CTR2(KTR_VFS,
1144 		    "%s: impossible to recycle, vp %p lock is already held",
1145 		    __func__, vp);
1146 		return (EWOULDBLOCK);
1147 	}
1148 	/*
1149 	 * Don't recycle if its filesystem is being suspended.
1150 	 */
1151 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1152 		VOP_UNLOCK(vp, 0);
1153 		CTR2(KTR_VFS,
1154 		    "%s: impossible to recycle, cannot start the write for %p",
1155 		    __func__, vp);
1156 		return (EBUSY);
1157 	}
1158 	/*
1159 	 * If we got this far, we need to acquire the interlock and see if
1160 	 * anyone picked up this vnode from another list.  If not, we will
1161 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1162 	 * will skip over it.
1163 	 */
1164 	VI_LOCK(vp);
1165 	if (vp->v_usecount) {
1166 		VOP_UNLOCK(vp, LK_INTERLOCK);
1167 		vn_finished_write(vnmp);
1168 		CTR2(KTR_VFS,
1169 		    "%s: impossible to recycle, %p is already referenced",
1170 		    __func__, vp);
1171 		return (EBUSY);
1172 	}
1173 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1174 		atomic_add_long(&recycles_count, 1);
1175 		vgonel(vp);
1176 	}
1177 	VOP_UNLOCK(vp, LK_INTERLOCK);
1178 	vn_finished_write(vnmp);
1179 	return (0);
1180 }
1181 
1182 static void
1183 vcheckspace(void)
1184 {
1185 
1186 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1187 		vnlruproc_sig = 1;
1188 		wakeup(vnlruproc);
1189 	}
1190 }
1191 
1192 /*
1193  * Wait if necessary for space for a new vnode.
1194  */
1195 static int
1196 getnewvnode_wait(int suspended)
1197 {
1198 
1199 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1200 	if (numvnodes >= desiredvnodes) {
1201 		if (suspended) {
1202 			/*
1203 			 * The file system is being suspended.  We cannot
1204 			 * risk a deadlock here, so allow allocation of
1205 			 * another vnode even if this would give too many.
1206 			 */
1207 			return (0);
1208 		}
1209 		if (vnlruproc_sig == 0) {
1210 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1211 			wakeup(vnlruproc);
1212 		}
1213 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1214 		    "vlruwk", hz);
1215 	}
1216 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1217 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1218 		vnlru_free(1);
1219 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1220 }
1221 
1222 /*
1223  * This hack is fragile, and probably not needed any more now that the
1224  * watermark handling works.
1225  */
1226 void
1227 getnewvnode_reserve(u_int count)
1228 {
1229 	struct thread *td;
1230 
1231 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1232 	/* XXX no longer so quick, but this part is not racy. */
1233 	mtx_lock(&vnode_free_list_mtx);
1234 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1235 		vnlru_free(ulmin(numvnodes + count - desiredvnodes,
1236 		    freevnodes - wantfreevnodes));
1237 	mtx_unlock(&vnode_free_list_mtx);
1238 
1239 	td = curthread;
1240 	/* First try to be quick and racy. */
1241 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1242 		td->td_vp_reserv += count;
1243 		vcheckspace();	/* XXX no longer so quick, but more racy */
1244 		return;
1245 	} else
1246 		atomic_subtract_long(&numvnodes, count);
1247 
1248 	mtx_lock(&vnode_free_list_mtx);
1249 	while (count > 0) {
1250 		if (getnewvnode_wait(0) == 0) {
1251 			count--;
1252 			td->td_vp_reserv++;
1253 			atomic_add_long(&numvnodes, 1);
1254 		}
1255 	}
1256 	vcheckspace();
1257 	mtx_unlock(&vnode_free_list_mtx);
1258 }
1259 
1260 /*
1261  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1262  * misconfgured or changed significantly.  Reducing desiredvnodes below
1263  * the reserved amount should cause bizarre behaviour like reducing it
1264  * below the number of active vnodes -- the system will try to reduce
1265  * numvnodes to match, but should fail, so the subtraction below should
1266  * not overflow.
1267  */
1268 void
1269 getnewvnode_drop_reserve(void)
1270 {
1271 	struct thread *td;
1272 
1273 	td = curthread;
1274 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1275 	td->td_vp_reserv = 0;
1276 }
1277 
1278 /*
1279  * Return the next vnode from the free list.
1280  */
1281 int
1282 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1283     struct vnode **vpp)
1284 {
1285 	struct vnode *vp;
1286 	struct thread *td;
1287 	struct lock_object *lo;
1288 	static int cyclecount;
1289 	int error;
1290 
1291 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1292 	vp = NULL;
1293 	td = curthread;
1294 	if (td->td_vp_reserv > 0) {
1295 		td->td_vp_reserv -= 1;
1296 		goto alloc;
1297 	}
1298 	mtx_lock(&vnode_free_list_mtx);
1299 	if (numvnodes < desiredvnodes)
1300 		cyclecount = 0;
1301 	else if (cyclecount++ >= freevnodes) {
1302 		cyclecount = 0;
1303 		vstir = 1;
1304 	}
1305 	/*
1306 	 * Grow the vnode cache if it will not be above its target max
1307 	 * after growing.  Otherwise, if the free list is nonempty, try
1308 	 * to reclaim 1 item from it before growing the cache (possibly
1309 	 * above its target max if the reclamation failed or is delayed).
1310 	 * Otherwise, wait for some space.  In all cases, schedule
1311 	 * vnlru_proc() if we are getting short of space.  The watermarks
1312 	 * should be chosen so that we never wait or even reclaim from
1313 	 * the free list to below its target minimum.
1314 	 */
1315 	if (numvnodes + 1 <= desiredvnodes)
1316 		;
1317 	else if (freevnodes > 0)
1318 		vnlru_free(1);
1319 	else {
1320 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1321 		    MNTK_SUSPEND));
1322 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1323 		if (error != 0) {
1324 			mtx_unlock(&vnode_free_list_mtx);
1325 			return (error);
1326 		}
1327 #endif
1328 	}
1329 	vcheckspace();
1330 	atomic_add_long(&numvnodes, 1);
1331 	mtx_unlock(&vnode_free_list_mtx);
1332 alloc:
1333 	atomic_add_long(&vnodes_created, 1);
1334 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1335 	/*
1336 	 * Locks are given the generic name "vnode" when created.
1337 	 * Follow the historic practice of using the filesystem
1338 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1339 	 *
1340 	 * Locks live in a witness group keyed on their name. Thus,
1341 	 * when a lock is renamed, it must also move from the witness
1342 	 * group of its old name to the witness group of its new name.
1343 	 *
1344 	 * The change only needs to be made when the vnode moves
1345 	 * from one filesystem type to another. We ensure that each
1346 	 * filesystem use a single static name pointer for its tag so
1347 	 * that we can compare pointers rather than doing a strcmp().
1348 	 */
1349 	lo = &vp->v_vnlock->lock_object;
1350 	if (lo->lo_name != tag) {
1351 		lo->lo_name = tag;
1352 		WITNESS_DESTROY(lo);
1353 		WITNESS_INIT(lo, tag);
1354 	}
1355 	/*
1356 	 * By default, don't allow shared locks unless filesystems opt-in.
1357 	 */
1358 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1359 	/*
1360 	 * Finalize various vnode identity bits.
1361 	 */
1362 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1363 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1364 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1365 	vp->v_type = VNON;
1366 	vp->v_tag = tag;
1367 	vp->v_op = vops;
1368 	v_init_counters(vp);
1369 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1370 #ifdef MAC
1371 	mac_vnode_init(vp);
1372 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1373 		mac_vnode_associate_singlelabel(mp, vp);
1374 	else if (mp == NULL && vops != &dead_vnodeops)
1375 		printf("NULL mp in getnewvnode()\n");
1376 #endif
1377 	if (mp != NULL) {
1378 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1379 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1380 			vp->v_vflag |= VV_NOKNOTE;
1381 	}
1382 
1383 	/*
1384 	 * For the filesystems which do not use vfs_hash_insert(),
1385 	 * still initialize v_hash to have vfs_hash_index() useful.
1386 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1387 	 * its own hashing.
1388 	 */
1389 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1390 
1391 	*vpp = vp;
1392 	return (0);
1393 }
1394 
1395 /*
1396  * Delete from old mount point vnode list, if on one.
1397  */
1398 static void
1399 delmntque(struct vnode *vp)
1400 {
1401 	struct mount *mp;
1402 	int active;
1403 
1404 	mp = vp->v_mount;
1405 	if (mp == NULL)
1406 		return;
1407 	MNT_ILOCK(mp);
1408 	VI_LOCK(vp);
1409 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1410 	    ("Active vnode list size %d > Vnode list size %d",
1411 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1412 	active = vp->v_iflag & VI_ACTIVE;
1413 	vp->v_iflag &= ~VI_ACTIVE;
1414 	if (active) {
1415 		mtx_lock(&vnode_free_list_mtx);
1416 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1417 		mp->mnt_activevnodelistsize--;
1418 		mtx_unlock(&vnode_free_list_mtx);
1419 	}
1420 	vp->v_mount = NULL;
1421 	VI_UNLOCK(vp);
1422 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1423 		("bad mount point vnode list size"));
1424 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1425 	mp->mnt_nvnodelistsize--;
1426 	MNT_REL(mp);
1427 	MNT_IUNLOCK(mp);
1428 }
1429 
1430 static void
1431 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1432 {
1433 
1434 	vp->v_data = NULL;
1435 	vp->v_op = &dead_vnodeops;
1436 	vgone(vp);
1437 	vput(vp);
1438 }
1439 
1440 /*
1441  * Insert into list of vnodes for the new mount point, if available.
1442  */
1443 int
1444 insmntque1(struct vnode *vp, struct mount *mp,
1445 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1446 {
1447 
1448 	KASSERT(vp->v_mount == NULL,
1449 		("insmntque: vnode already on per mount vnode list"));
1450 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1451 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1452 
1453 	/*
1454 	 * We acquire the vnode interlock early to ensure that the
1455 	 * vnode cannot be recycled by another process releasing a
1456 	 * holdcnt on it before we get it on both the vnode list
1457 	 * and the active vnode list. The mount mutex protects only
1458 	 * manipulation of the vnode list and the vnode freelist
1459 	 * mutex protects only manipulation of the active vnode list.
1460 	 * Hence the need to hold the vnode interlock throughout.
1461 	 */
1462 	MNT_ILOCK(mp);
1463 	VI_LOCK(vp);
1464 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1465 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1466 	    mp->mnt_nvnodelistsize == 0)) &&
1467 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1468 		VI_UNLOCK(vp);
1469 		MNT_IUNLOCK(mp);
1470 		if (dtr != NULL)
1471 			dtr(vp, dtr_arg);
1472 		return (EBUSY);
1473 	}
1474 	vp->v_mount = mp;
1475 	MNT_REF(mp);
1476 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1477 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1478 		("neg mount point vnode list size"));
1479 	mp->mnt_nvnodelistsize++;
1480 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1481 	    ("Activating already active vnode"));
1482 	vp->v_iflag |= VI_ACTIVE;
1483 	mtx_lock(&vnode_free_list_mtx);
1484 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1485 	mp->mnt_activevnodelistsize++;
1486 	mtx_unlock(&vnode_free_list_mtx);
1487 	VI_UNLOCK(vp);
1488 	MNT_IUNLOCK(mp);
1489 	return (0);
1490 }
1491 
1492 int
1493 insmntque(struct vnode *vp, struct mount *mp)
1494 {
1495 
1496 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1497 }
1498 
1499 /*
1500  * Flush out and invalidate all buffers associated with a bufobj
1501  * Called with the underlying object locked.
1502  */
1503 int
1504 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1505 {
1506 	int error;
1507 
1508 	BO_LOCK(bo);
1509 	if (flags & V_SAVE) {
1510 		error = bufobj_wwait(bo, slpflag, slptimeo);
1511 		if (error) {
1512 			BO_UNLOCK(bo);
1513 			return (error);
1514 		}
1515 		if (bo->bo_dirty.bv_cnt > 0) {
1516 			BO_UNLOCK(bo);
1517 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1518 				return (error);
1519 			/*
1520 			 * XXX We could save a lock/unlock if this was only
1521 			 * enabled under INVARIANTS
1522 			 */
1523 			BO_LOCK(bo);
1524 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1525 				panic("vinvalbuf: dirty bufs");
1526 		}
1527 	}
1528 	/*
1529 	 * If you alter this loop please notice that interlock is dropped and
1530 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1531 	 * no race conditions occur from this.
1532 	 */
1533 	do {
1534 		error = flushbuflist(&bo->bo_clean,
1535 		    flags, bo, slpflag, slptimeo);
1536 		if (error == 0 && !(flags & V_CLEANONLY))
1537 			error = flushbuflist(&bo->bo_dirty,
1538 			    flags, bo, slpflag, slptimeo);
1539 		if (error != 0 && error != EAGAIN) {
1540 			BO_UNLOCK(bo);
1541 			return (error);
1542 		}
1543 	} while (error != 0);
1544 
1545 	/*
1546 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1547 	 * have write I/O in-progress but if there is a VM object then the
1548 	 * VM object can also have read-I/O in-progress.
1549 	 */
1550 	do {
1551 		bufobj_wwait(bo, 0, 0);
1552 		BO_UNLOCK(bo);
1553 		if (bo->bo_object != NULL) {
1554 			VM_OBJECT_WLOCK(bo->bo_object);
1555 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1556 			VM_OBJECT_WUNLOCK(bo->bo_object);
1557 		}
1558 		BO_LOCK(bo);
1559 	} while (bo->bo_numoutput > 0);
1560 	BO_UNLOCK(bo);
1561 
1562 	/*
1563 	 * Destroy the copy in the VM cache, too.
1564 	 */
1565 	if (bo->bo_object != NULL &&
1566 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) {
1567 		VM_OBJECT_WLOCK(bo->bo_object);
1568 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1569 		    OBJPR_CLEANONLY : 0);
1570 		VM_OBJECT_WUNLOCK(bo->bo_object);
1571 	}
1572 
1573 #ifdef INVARIANTS
1574 	BO_LOCK(bo);
1575 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 &&
1576 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1577 		panic("vinvalbuf: flush failed");
1578 	BO_UNLOCK(bo);
1579 #endif
1580 	return (0);
1581 }
1582 
1583 /*
1584  * Flush out and invalidate all buffers associated with a vnode.
1585  * Called with the underlying object locked.
1586  */
1587 int
1588 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1589 {
1590 
1591 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1592 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1593 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1594 		return (0);
1595 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1596 }
1597 
1598 /*
1599  * Flush out buffers on the specified list.
1600  *
1601  */
1602 static int
1603 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1604     int slptimeo)
1605 {
1606 	struct buf *bp, *nbp;
1607 	int retval, error;
1608 	daddr_t lblkno;
1609 	b_xflags_t xflags;
1610 
1611 	ASSERT_BO_WLOCKED(bo);
1612 
1613 	retval = 0;
1614 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1615 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1616 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1617 			continue;
1618 		}
1619 		lblkno = 0;
1620 		xflags = 0;
1621 		if (nbp != NULL) {
1622 			lblkno = nbp->b_lblkno;
1623 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1624 		}
1625 		retval = EAGAIN;
1626 		error = BUF_TIMELOCK(bp,
1627 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1628 		    "flushbuf", slpflag, slptimeo);
1629 		if (error) {
1630 			BO_LOCK(bo);
1631 			return (error != ENOLCK ? error : EAGAIN);
1632 		}
1633 		KASSERT(bp->b_bufobj == bo,
1634 		    ("bp %p wrong b_bufobj %p should be %p",
1635 		    bp, bp->b_bufobj, bo));
1636 		/*
1637 		 * XXX Since there are no node locks for NFS, I
1638 		 * believe there is a slight chance that a delayed
1639 		 * write will occur while sleeping just above, so
1640 		 * check for it.
1641 		 */
1642 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1643 		    (flags & V_SAVE)) {
1644 			bremfree(bp);
1645 			bp->b_flags |= B_ASYNC;
1646 			bwrite(bp);
1647 			BO_LOCK(bo);
1648 			return (EAGAIN);	/* XXX: why not loop ? */
1649 		}
1650 		bremfree(bp);
1651 		bp->b_flags |= (B_INVAL | B_RELBUF);
1652 		bp->b_flags &= ~B_ASYNC;
1653 		brelse(bp);
1654 		BO_LOCK(bo);
1655 		if (nbp != NULL &&
1656 		    (nbp->b_bufobj != bo ||
1657 		     nbp->b_lblkno != lblkno ||
1658 		     (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1659 			break;			/* nbp invalid */
1660 	}
1661 	return (retval);
1662 }
1663 
1664 /*
1665  * Truncate a file's buffer and pages to a specified length.  This
1666  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1667  * sync activity.
1668  */
1669 int
1670 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1671 {
1672 	struct buf *bp, *nbp;
1673 	int anyfreed;
1674 	int trunclbn;
1675 	struct bufobj *bo;
1676 
1677 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1678 	    vp, cred, blksize, (uintmax_t)length);
1679 
1680 	/*
1681 	 * Round up to the *next* lbn.
1682 	 */
1683 	trunclbn = (length + blksize - 1) / blksize;
1684 
1685 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1686 restart:
1687 	bo = &vp->v_bufobj;
1688 	BO_LOCK(bo);
1689 	anyfreed = 1;
1690 	for (;anyfreed;) {
1691 		anyfreed = 0;
1692 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1693 			if (bp->b_lblkno < trunclbn)
1694 				continue;
1695 			if (BUF_LOCK(bp,
1696 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1697 			    BO_LOCKPTR(bo)) == ENOLCK)
1698 				goto restart;
1699 
1700 			bremfree(bp);
1701 			bp->b_flags |= (B_INVAL | B_RELBUF);
1702 			bp->b_flags &= ~B_ASYNC;
1703 			brelse(bp);
1704 			anyfreed = 1;
1705 
1706 			BO_LOCK(bo);
1707 			if (nbp != NULL &&
1708 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1709 			    (nbp->b_vp != vp) ||
1710 			    (nbp->b_flags & B_DELWRI))) {
1711 				BO_UNLOCK(bo);
1712 				goto restart;
1713 			}
1714 		}
1715 
1716 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1717 			if (bp->b_lblkno < trunclbn)
1718 				continue;
1719 			if (BUF_LOCK(bp,
1720 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1721 			    BO_LOCKPTR(bo)) == ENOLCK)
1722 				goto restart;
1723 			bremfree(bp);
1724 			bp->b_flags |= (B_INVAL | B_RELBUF);
1725 			bp->b_flags &= ~B_ASYNC;
1726 			brelse(bp);
1727 			anyfreed = 1;
1728 
1729 			BO_LOCK(bo);
1730 			if (nbp != NULL &&
1731 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1732 			    (nbp->b_vp != vp) ||
1733 			    (nbp->b_flags & B_DELWRI) == 0)) {
1734 				BO_UNLOCK(bo);
1735 				goto restart;
1736 			}
1737 		}
1738 	}
1739 
1740 	if (length > 0) {
1741 restartsync:
1742 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1743 			if (bp->b_lblkno > 0)
1744 				continue;
1745 			/*
1746 			 * Since we hold the vnode lock this should only
1747 			 * fail if we're racing with the buf daemon.
1748 			 */
1749 			if (BUF_LOCK(bp,
1750 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1751 			    BO_LOCKPTR(bo)) == ENOLCK) {
1752 				goto restart;
1753 			}
1754 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1755 			    ("buf(%p) on dirty queue without DELWRI", bp));
1756 
1757 			bremfree(bp);
1758 			bawrite(bp);
1759 			BO_LOCK(bo);
1760 			goto restartsync;
1761 		}
1762 	}
1763 
1764 	bufobj_wwait(bo, 0, 0);
1765 	BO_UNLOCK(bo);
1766 	vnode_pager_setsize(vp, length);
1767 
1768 	return (0);
1769 }
1770 
1771 static void
1772 buf_vlist_remove(struct buf *bp)
1773 {
1774 	struct bufv *bv;
1775 
1776 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1777 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1778 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1779 	    (BX_VNDIRTY|BX_VNCLEAN),
1780 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1781 	if (bp->b_xflags & BX_VNDIRTY)
1782 		bv = &bp->b_bufobj->bo_dirty;
1783 	else
1784 		bv = &bp->b_bufobj->bo_clean;
1785 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1786 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1787 	bv->bv_cnt--;
1788 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1789 }
1790 
1791 /*
1792  * Add the buffer to the sorted clean or dirty block list.
1793  *
1794  * NOTE: xflags is passed as a constant, optimizing this inline function!
1795  */
1796 static void
1797 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1798 {
1799 	struct bufv *bv;
1800 	struct buf *n;
1801 	int error;
1802 
1803 	ASSERT_BO_WLOCKED(bo);
1804 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1805 	    ("dead bo %p", bo));
1806 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1807 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1808 	bp->b_xflags |= xflags;
1809 	if (xflags & BX_VNDIRTY)
1810 		bv = &bo->bo_dirty;
1811 	else
1812 		bv = &bo->bo_clean;
1813 
1814 	/*
1815 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1816 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1817 	 * than _ge.
1818 	 */
1819 	if (bv->bv_cnt == 0 ||
1820 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1821 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1822 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1823 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1824 	else
1825 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1826 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1827 	if (error)
1828 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1829 	bv->bv_cnt++;
1830 }
1831 
1832 /*
1833  * Look up a buffer using the buffer tries.
1834  */
1835 struct buf *
1836 gbincore(struct bufobj *bo, daddr_t lblkno)
1837 {
1838 	struct buf *bp;
1839 
1840 	ASSERT_BO_LOCKED(bo);
1841 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1842 	if (bp != NULL)
1843 		return (bp);
1844 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1845 }
1846 
1847 /*
1848  * Associate a buffer with a vnode.
1849  */
1850 void
1851 bgetvp(struct vnode *vp, struct buf *bp)
1852 {
1853 	struct bufobj *bo;
1854 
1855 	bo = &vp->v_bufobj;
1856 	ASSERT_BO_WLOCKED(bo);
1857 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1858 
1859 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1860 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1861 	    ("bgetvp: bp already attached! %p", bp));
1862 
1863 	vhold(vp);
1864 	bp->b_vp = vp;
1865 	bp->b_bufobj = bo;
1866 	/*
1867 	 * Insert onto list for new vnode.
1868 	 */
1869 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1870 }
1871 
1872 /*
1873  * Disassociate a buffer from a vnode.
1874  */
1875 void
1876 brelvp(struct buf *bp)
1877 {
1878 	struct bufobj *bo;
1879 	struct vnode *vp;
1880 
1881 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1882 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1883 
1884 	/*
1885 	 * Delete from old vnode list, if on one.
1886 	 */
1887 	vp = bp->b_vp;		/* XXX */
1888 	bo = bp->b_bufobj;
1889 	BO_LOCK(bo);
1890 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1891 		buf_vlist_remove(bp);
1892 	else
1893 		panic("brelvp: Buffer %p not on queue.", bp);
1894 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1895 		bo->bo_flag &= ~BO_ONWORKLST;
1896 		mtx_lock(&sync_mtx);
1897 		LIST_REMOVE(bo, bo_synclist);
1898 		syncer_worklist_len--;
1899 		mtx_unlock(&sync_mtx);
1900 	}
1901 	bp->b_vp = NULL;
1902 	bp->b_bufobj = NULL;
1903 	BO_UNLOCK(bo);
1904 	vdrop(vp);
1905 }
1906 
1907 /*
1908  * Add an item to the syncer work queue.
1909  */
1910 static void
1911 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1912 {
1913 	int slot;
1914 
1915 	ASSERT_BO_WLOCKED(bo);
1916 
1917 	mtx_lock(&sync_mtx);
1918 	if (bo->bo_flag & BO_ONWORKLST)
1919 		LIST_REMOVE(bo, bo_synclist);
1920 	else {
1921 		bo->bo_flag |= BO_ONWORKLST;
1922 		syncer_worklist_len++;
1923 	}
1924 
1925 	if (delay > syncer_maxdelay - 2)
1926 		delay = syncer_maxdelay - 2;
1927 	slot = (syncer_delayno + delay) & syncer_mask;
1928 
1929 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1930 	mtx_unlock(&sync_mtx);
1931 }
1932 
1933 static int
1934 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1935 {
1936 	int error, len;
1937 
1938 	mtx_lock(&sync_mtx);
1939 	len = syncer_worklist_len - sync_vnode_count;
1940 	mtx_unlock(&sync_mtx);
1941 	error = SYSCTL_OUT(req, &len, sizeof(len));
1942 	return (error);
1943 }
1944 
1945 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1946     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1947 
1948 static struct proc *updateproc;
1949 static void sched_sync(void);
1950 static struct kproc_desc up_kp = {
1951 	"syncer",
1952 	sched_sync,
1953 	&updateproc
1954 };
1955 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
1956 
1957 static int
1958 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
1959 {
1960 	struct vnode *vp;
1961 	struct mount *mp;
1962 
1963 	*bo = LIST_FIRST(slp);
1964 	if (*bo == NULL)
1965 		return (0);
1966 	vp = (*bo)->__bo_vnode;	/* XXX */
1967 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
1968 		return (1);
1969 	/*
1970 	 * We use vhold in case the vnode does not
1971 	 * successfully sync.  vhold prevents the vnode from
1972 	 * going away when we unlock the sync_mtx so that
1973 	 * we can acquire the vnode interlock.
1974 	 */
1975 	vholdl(vp);
1976 	mtx_unlock(&sync_mtx);
1977 	VI_UNLOCK(vp);
1978 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1979 		vdrop(vp);
1980 		mtx_lock(&sync_mtx);
1981 		return (*bo == LIST_FIRST(slp));
1982 	}
1983 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1984 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1985 	VOP_UNLOCK(vp, 0);
1986 	vn_finished_write(mp);
1987 	BO_LOCK(*bo);
1988 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
1989 		/*
1990 		 * Put us back on the worklist.  The worklist
1991 		 * routine will remove us from our current
1992 		 * position and then add us back in at a later
1993 		 * position.
1994 		 */
1995 		vn_syncer_add_to_worklist(*bo, syncdelay);
1996 	}
1997 	BO_UNLOCK(*bo);
1998 	vdrop(vp);
1999 	mtx_lock(&sync_mtx);
2000 	return (0);
2001 }
2002 
2003 static int first_printf = 1;
2004 
2005 /*
2006  * System filesystem synchronizer daemon.
2007  */
2008 static void
2009 sched_sync(void)
2010 {
2011 	struct synclist *next, *slp;
2012 	struct bufobj *bo;
2013 	long starttime;
2014 	struct thread *td = curthread;
2015 	int last_work_seen;
2016 	int net_worklist_len;
2017 	int syncer_final_iter;
2018 	int error;
2019 
2020 	last_work_seen = 0;
2021 	syncer_final_iter = 0;
2022 	syncer_state = SYNCER_RUNNING;
2023 	starttime = time_uptime;
2024 	td->td_pflags |= TDP_NORUNNINGBUF;
2025 
2026 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2027 	    SHUTDOWN_PRI_LAST);
2028 
2029 	mtx_lock(&sync_mtx);
2030 	for (;;) {
2031 		if (syncer_state == SYNCER_FINAL_DELAY &&
2032 		    syncer_final_iter == 0) {
2033 			mtx_unlock(&sync_mtx);
2034 			kproc_suspend_check(td->td_proc);
2035 			mtx_lock(&sync_mtx);
2036 		}
2037 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2038 		if (syncer_state != SYNCER_RUNNING &&
2039 		    starttime != time_uptime) {
2040 			if (first_printf) {
2041 				printf("\nSyncing disks, vnodes remaining...");
2042 				first_printf = 0;
2043 			}
2044 			printf("%d ", net_worklist_len);
2045 		}
2046 		starttime = time_uptime;
2047 
2048 		/*
2049 		 * Push files whose dirty time has expired.  Be careful
2050 		 * of interrupt race on slp queue.
2051 		 *
2052 		 * Skip over empty worklist slots when shutting down.
2053 		 */
2054 		do {
2055 			slp = &syncer_workitem_pending[syncer_delayno];
2056 			syncer_delayno += 1;
2057 			if (syncer_delayno == syncer_maxdelay)
2058 				syncer_delayno = 0;
2059 			next = &syncer_workitem_pending[syncer_delayno];
2060 			/*
2061 			 * If the worklist has wrapped since the
2062 			 * it was emptied of all but syncer vnodes,
2063 			 * switch to the FINAL_DELAY state and run
2064 			 * for one more second.
2065 			 */
2066 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2067 			    net_worklist_len == 0 &&
2068 			    last_work_seen == syncer_delayno) {
2069 				syncer_state = SYNCER_FINAL_DELAY;
2070 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2071 			}
2072 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2073 		    syncer_worklist_len > 0);
2074 
2075 		/*
2076 		 * Keep track of the last time there was anything
2077 		 * on the worklist other than syncer vnodes.
2078 		 * Return to the SHUTTING_DOWN state if any
2079 		 * new work appears.
2080 		 */
2081 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2082 			last_work_seen = syncer_delayno;
2083 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2084 			syncer_state = SYNCER_SHUTTING_DOWN;
2085 		while (!LIST_EMPTY(slp)) {
2086 			error = sync_vnode(slp, &bo, td);
2087 			if (error == 1) {
2088 				LIST_REMOVE(bo, bo_synclist);
2089 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2090 				continue;
2091 			}
2092 
2093 			if (first_printf == 0) {
2094 				/*
2095 				 * Drop the sync mutex, because some watchdog
2096 				 * drivers need to sleep while patting
2097 				 */
2098 				mtx_unlock(&sync_mtx);
2099 				wdog_kern_pat(WD_LASTVAL);
2100 				mtx_lock(&sync_mtx);
2101 			}
2102 
2103 		}
2104 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2105 			syncer_final_iter--;
2106 		/*
2107 		 * The variable rushjob allows the kernel to speed up the
2108 		 * processing of the filesystem syncer process. A rushjob
2109 		 * value of N tells the filesystem syncer to process the next
2110 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2111 		 * is used by the soft update code to speed up the filesystem
2112 		 * syncer process when the incore state is getting so far
2113 		 * ahead of the disk that the kernel memory pool is being
2114 		 * threatened with exhaustion.
2115 		 */
2116 		if (rushjob > 0) {
2117 			rushjob -= 1;
2118 			continue;
2119 		}
2120 		/*
2121 		 * Just sleep for a short period of time between
2122 		 * iterations when shutting down to allow some I/O
2123 		 * to happen.
2124 		 *
2125 		 * If it has taken us less than a second to process the
2126 		 * current work, then wait. Otherwise start right over
2127 		 * again. We can still lose time if any single round
2128 		 * takes more than two seconds, but it does not really
2129 		 * matter as we are just trying to generally pace the
2130 		 * filesystem activity.
2131 		 */
2132 		if (syncer_state != SYNCER_RUNNING ||
2133 		    time_uptime == starttime) {
2134 			thread_lock(td);
2135 			sched_prio(td, PPAUSE);
2136 			thread_unlock(td);
2137 		}
2138 		if (syncer_state != SYNCER_RUNNING)
2139 			cv_timedwait(&sync_wakeup, &sync_mtx,
2140 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2141 		else if (time_uptime == starttime)
2142 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2143 	}
2144 }
2145 
2146 /*
2147  * Request the syncer daemon to speed up its work.
2148  * We never push it to speed up more than half of its
2149  * normal turn time, otherwise it could take over the cpu.
2150  */
2151 int
2152 speedup_syncer(void)
2153 {
2154 	int ret = 0;
2155 
2156 	mtx_lock(&sync_mtx);
2157 	if (rushjob < syncdelay / 2) {
2158 		rushjob += 1;
2159 		stat_rush_requests += 1;
2160 		ret = 1;
2161 	}
2162 	mtx_unlock(&sync_mtx);
2163 	cv_broadcast(&sync_wakeup);
2164 	return (ret);
2165 }
2166 
2167 /*
2168  * Tell the syncer to speed up its work and run though its work
2169  * list several times, then tell it to shut down.
2170  */
2171 static void
2172 syncer_shutdown(void *arg, int howto)
2173 {
2174 
2175 	if (howto & RB_NOSYNC)
2176 		return;
2177 	mtx_lock(&sync_mtx);
2178 	syncer_state = SYNCER_SHUTTING_DOWN;
2179 	rushjob = 0;
2180 	mtx_unlock(&sync_mtx);
2181 	cv_broadcast(&sync_wakeup);
2182 	kproc_shutdown(arg, howto);
2183 }
2184 
2185 void
2186 syncer_suspend(void)
2187 {
2188 
2189 	syncer_shutdown(updateproc, 0);
2190 }
2191 
2192 void
2193 syncer_resume(void)
2194 {
2195 
2196 	mtx_lock(&sync_mtx);
2197 	first_printf = 1;
2198 	syncer_state = SYNCER_RUNNING;
2199 	mtx_unlock(&sync_mtx);
2200 	cv_broadcast(&sync_wakeup);
2201 	kproc_resume(updateproc);
2202 }
2203 
2204 /*
2205  * Reassign a buffer from one vnode to another.
2206  * Used to assign file specific control information
2207  * (indirect blocks) to the vnode to which they belong.
2208  */
2209 void
2210 reassignbuf(struct buf *bp)
2211 {
2212 	struct vnode *vp;
2213 	struct bufobj *bo;
2214 	int delay;
2215 #ifdef INVARIANTS
2216 	struct bufv *bv;
2217 #endif
2218 
2219 	vp = bp->b_vp;
2220 	bo = bp->b_bufobj;
2221 	++reassignbufcalls;
2222 
2223 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2224 	    bp, bp->b_vp, bp->b_flags);
2225 	/*
2226 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2227 	 * is not fully linked in.
2228 	 */
2229 	if (bp->b_flags & B_PAGING)
2230 		panic("cannot reassign paging buffer");
2231 
2232 	/*
2233 	 * Delete from old vnode list, if on one.
2234 	 */
2235 	BO_LOCK(bo);
2236 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2237 		buf_vlist_remove(bp);
2238 	else
2239 		panic("reassignbuf: Buffer %p not on queue.", bp);
2240 	/*
2241 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2242 	 * of clean buffers.
2243 	 */
2244 	if (bp->b_flags & B_DELWRI) {
2245 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2246 			switch (vp->v_type) {
2247 			case VDIR:
2248 				delay = dirdelay;
2249 				break;
2250 			case VCHR:
2251 				delay = metadelay;
2252 				break;
2253 			default:
2254 				delay = filedelay;
2255 			}
2256 			vn_syncer_add_to_worklist(bo, delay);
2257 		}
2258 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2259 	} else {
2260 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2261 
2262 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2263 			mtx_lock(&sync_mtx);
2264 			LIST_REMOVE(bo, bo_synclist);
2265 			syncer_worklist_len--;
2266 			mtx_unlock(&sync_mtx);
2267 			bo->bo_flag &= ~BO_ONWORKLST;
2268 		}
2269 	}
2270 #ifdef INVARIANTS
2271 	bv = &bo->bo_clean;
2272 	bp = TAILQ_FIRST(&bv->bv_hd);
2273 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2274 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2275 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2276 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2277 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2278 	bv = &bo->bo_dirty;
2279 	bp = TAILQ_FIRST(&bv->bv_hd);
2280 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2281 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2282 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2283 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2284 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2285 #endif
2286 	BO_UNLOCK(bo);
2287 }
2288 
2289 /*
2290  * A temporary hack until refcount_* APIs are sorted out.
2291  */
2292 static __inline int
2293 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2294 {
2295 	u_int old;
2296 
2297 	for (;;) {
2298 		old = *count;
2299 		if (old == 0)
2300 			return (0);
2301 		if (atomic_cmpset_int(count, old, old + 1))
2302 			return (1);
2303 	}
2304 }
2305 
2306 static __inline int
2307 vfs_refcount_release_if_not_last(volatile u_int *count)
2308 {
2309 	u_int old;
2310 
2311 	for (;;) {
2312 		old = *count;
2313 		if (old == 1)
2314 			return (0);
2315 		if (atomic_cmpset_int(count, old, old - 1))
2316 			return (1);
2317 	}
2318 }
2319 
2320 static void
2321 v_init_counters(struct vnode *vp)
2322 {
2323 
2324 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2325 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2326 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2327 
2328 	refcount_init(&vp->v_holdcnt, 1);
2329 	refcount_init(&vp->v_usecount, 1);
2330 }
2331 
2332 /*
2333  * Increment the use and hold counts on the vnode, taking care to reference
2334  * the driver's usecount if this is a chardev.  The _vhold() will remove
2335  * the vnode from the free list if it is presently free.
2336  */
2337 static void
2338 v_incr_usecount(struct vnode *vp)
2339 {
2340 
2341 	ASSERT_VI_UNLOCKED(vp, __func__);
2342 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2343 
2344 	if (vp->v_type == VCHR) {
2345 		VI_LOCK(vp);
2346 		_vhold(vp, true);
2347 		if (vp->v_iflag & VI_OWEINACT) {
2348 			VNASSERT(vp->v_usecount == 0, vp,
2349 			    ("vnode with usecount and VI_OWEINACT set"));
2350 			vp->v_iflag &= ~VI_OWEINACT;
2351 		}
2352 		refcount_acquire(&vp->v_usecount);
2353 		v_incr_devcount(vp);
2354 		VI_UNLOCK(vp);
2355 		return;
2356 	}
2357 
2358 	_vhold(vp, false);
2359 	if (vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2360 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2361 		    ("vnode with usecount and VI_OWEINACT set"));
2362 	} else {
2363 		VI_LOCK(vp);
2364 		if (vp->v_iflag & VI_OWEINACT)
2365 			vp->v_iflag &= ~VI_OWEINACT;
2366 		refcount_acquire(&vp->v_usecount);
2367 		VI_UNLOCK(vp);
2368 	}
2369 }
2370 
2371 /*
2372  * Increment si_usecount of the associated device, if any.
2373  */
2374 static void
2375 v_incr_devcount(struct vnode *vp)
2376 {
2377 
2378 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2379 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2380 		dev_lock();
2381 		vp->v_rdev->si_usecount++;
2382 		dev_unlock();
2383 	}
2384 }
2385 
2386 /*
2387  * Decrement si_usecount of the associated device, if any.
2388  */
2389 static void
2390 v_decr_devcount(struct vnode *vp)
2391 {
2392 
2393 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2394 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2395 		dev_lock();
2396 		vp->v_rdev->si_usecount--;
2397 		dev_unlock();
2398 	}
2399 }
2400 
2401 /*
2402  * Grab a particular vnode from the free list, increment its
2403  * reference count and lock it.  VI_DOOMED is set if the vnode
2404  * is being destroyed.  Only callers who specify LK_RETRY will
2405  * see doomed vnodes.  If inactive processing was delayed in
2406  * vput try to do it here.
2407  *
2408  * Notes on lockless counter manipulation:
2409  * _vhold, vputx and other routines make various decisions based
2410  * on either holdcnt or usecount being 0. As long as either contuner
2411  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2412  * with atomic operations. Otherwise the interlock is taken.
2413  */
2414 int
2415 vget(struct vnode *vp, int flags, struct thread *td)
2416 {
2417 	int error, oweinact;
2418 
2419 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2420 	    ("vget: invalid lock operation"));
2421 
2422 	if ((flags & LK_INTERLOCK) != 0)
2423 		ASSERT_VI_LOCKED(vp, __func__);
2424 	else
2425 		ASSERT_VI_UNLOCKED(vp, __func__);
2426 	if ((flags & LK_VNHELD) != 0)
2427 		VNASSERT((vp->v_holdcnt > 0), vp,
2428 		    ("vget: LK_VNHELD passed but vnode not held"));
2429 
2430 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2431 
2432 	if ((flags & LK_VNHELD) == 0)
2433 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2434 
2435 	if ((error = vn_lock(vp, flags)) != 0) {
2436 		vdrop(vp);
2437 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2438 		    vp);
2439 		return (error);
2440 	}
2441 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2442 		panic("vget: vn_lock failed to return ENOENT\n");
2443 	/*
2444 	 * We don't guarantee that any particular close will
2445 	 * trigger inactive processing so just make a best effort
2446 	 * here at preventing a reference to a removed file.  If
2447 	 * we don't succeed no harm is done.
2448 	 *
2449 	 * Upgrade our holdcnt to a usecount.
2450 	 */
2451 	if (vp->v_type != VCHR &&
2452 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2453 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2454 		    ("vnode with usecount and VI_OWEINACT set"));
2455 	} else {
2456 		VI_LOCK(vp);
2457 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2458 			oweinact = 0;
2459 		} else {
2460 			oweinact = 1;
2461 			vp->v_iflag &= ~VI_OWEINACT;
2462 		}
2463 		refcount_acquire(&vp->v_usecount);
2464 		v_incr_devcount(vp);
2465 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2466 		    (flags & LK_NOWAIT) == 0)
2467 			vinactive(vp, td);
2468 		VI_UNLOCK(vp);
2469 	}
2470 	return (0);
2471 }
2472 
2473 /*
2474  * Increase the reference count of a vnode.
2475  */
2476 void
2477 vref(struct vnode *vp)
2478 {
2479 
2480 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2481 	v_incr_usecount(vp);
2482 }
2483 
2484 /*
2485  * Return reference count of a vnode.
2486  *
2487  * The results of this call are only guaranteed when some mechanism is used to
2488  * stop other processes from gaining references to the vnode.  This may be the
2489  * case if the caller holds the only reference.  This is also useful when stale
2490  * data is acceptable as race conditions may be accounted for by some other
2491  * means.
2492  */
2493 int
2494 vrefcnt(struct vnode *vp)
2495 {
2496 
2497 	return (vp->v_usecount);
2498 }
2499 
2500 #define	VPUTX_VRELE	1
2501 #define	VPUTX_VPUT	2
2502 #define	VPUTX_VUNREF	3
2503 
2504 /*
2505  * Decrement the use and hold counts for a vnode.
2506  *
2507  * See an explanation near vget() as to why atomic operation is safe.
2508  */
2509 static void
2510 vputx(struct vnode *vp, int func)
2511 {
2512 	int error;
2513 
2514 	KASSERT(vp != NULL, ("vputx: null vp"));
2515 	if (func == VPUTX_VUNREF)
2516 		ASSERT_VOP_LOCKED(vp, "vunref");
2517 	else if (func == VPUTX_VPUT)
2518 		ASSERT_VOP_LOCKED(vp, "vput");
2519 	else
2520 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2521 	ASSERT_VI_UNLOCKED(vp, __func__);
2522 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2523 
2524 	if (vp->v_type != VCHR &&
2525 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2526 		if (func == VPUTX_VPUT)
2527 			VOP_UNLOCK(vp, 0);
2528 		vdrop(vp);
2529 		return;
2530 	}
2531 
2532 	VI_LOCK(vp);
2533 
2534 	/*
2535 	 * We want to hold the vnode until the inactive finishes to
2536 	 * prevent vgone() races.  We drop the use count here and the
2537 	 * hold count below when we're done.
2538 	 */
2539 	if (!refcount_release(&vp->v_usecount) ||
2540 	    (vp->v_iflag & VI_DOINGINACT)) {
2541 		if (func == VPUTX_VPUT)
2542 			VOP_UNLOCK(vp, 0);
2543 		v_decr_devcount(vp);
2544 		vdropl(vp);
2545 		return;
2546 	}
2547 
2548 	v_decr_devcount(vp);
2549 
2550 	error = 0;
2551 
2552 	if (vp->v_usecount != 0) {
2553 		vprint("vputx: usecount not zero", vp);
2554 		panic("vputx: usecount not zero");
2555 	}
2556 
2557 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2558 
2559 	/*
2560 	 * We must call VOP_INACTIVE with the node locked. Mark
2561 	 * as VI_DOINGINACT to avoid recursion.
2562 	 */
2563 	vp->v_iflag |= VI_OWEINACT;
2564 	switch (func) {
2565 	case VPUTX_VRELE:
2566 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2567 		VI_LOCK(vp);
2568 		break;
2569 	case VPUTX_VPUT:
2570 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2571 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2572 			    LK_NOWAIT);
2573 			VI_LOCK(vp);
2574 		}
2575 		break;
2576 	case VPUTX_VUNREF:
2577 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2578 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2579 			VI_LOCK(vp);
2580 		}
2581 		break;
2582 	}
2583 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2584 	    ("vnode with usecount and VI_OWEINACT set"));
2585 	if (error == 0) {
2586 		if (vp->v_iflag & VI_OWEINACT)
2587 			vinactive(vp, curthread);
2588 		if (func != VPUTX_VUNREF)
2589 			VOP_UNLOCK(vp, 0);
2590 	}
2591 	vdropl(vp);
2592 }
2593 
2594 /*
2595  * Vnode put/release.
2596  * If count drops to zero, call inactive routine and return to freelist.
2597  */
2598 void
2599 vrele(struct vnode *vp)
2600 {
2601 
2602 	vputx(vp, VPUTX_VRELE);
2603 }
2604 
2605 /*
2606  * Release an already locked vnode.  This give the same effects as
2607  * unlock+vrele(), but takes less time and avoids releasing and
2608  * re-aquiring the lock (as vrele() acquires the lock internally.)
2609  */
2610 void
2611 vput(struct vnode *vp)
2612 {
2613 
2614 	vputx(vp, VPUTX_VPUT);
2615 }
2616 
2617 /*
2618  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2619  */
2620 void
2621 vunref(struct vnode *vp)
2622 {
2623 
2624 	vputx(vp, VPUTX_VUNREF);
2625 }
2626 
2627 /*
2628  * Increase the hold count and activate if this is the first reference.
2629  */
2630 void
2631 _vhold(struct vnode *vp, bool locked)
2632 {
2633 	struct mount *mp;
2634 
2635 	if (locked)
2636 		ASSERT_VI_LOCKED(vp, __func__);
2637 	else
2638 		ASSERT_VI_UNLOCKED(vp, __func__);
2639 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2640 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2641 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2642 		    ("_vhold: vnode with holdcnt is free"));
2643 		return;
2644 	}
2645 
2646 	if (!locked)
2647 		VI_LOCK(vp);
2648 	if ((vp->v_iflag & VI_FREE) == 0) {
2649 		refcount_acquire(&vp->v_holdcnt);
2650 		if (!locked)
2651 			VI_UNLOCK(vp);
2652 		return;
2653 	}
2654 	VNASSERT(vp->v_holdcnt == 0, vp,
2655 	    ("%s: wrong hold count", __func__));
2656 	VNASSERT(vp->v_op != NULL, vp,
2657 	    ("%s: vnode already reclaimed.", __func__));
2658 	/*
2659 	 * Remove a vnode from the free list, mark it as in use,
2660 	 * and put it on the active list.
2661 	 */
2662 	mtx_lock(&vnode_free_list_mtx);
2663 	TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2664 	freevnodes--;
2665 	vp->v_iflag &= ~VI_FREE;
2666 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2667 	    ("Activating already active vnode"));
2668 	vp->v_iflag |= VI_ACTIVE;
2669 	mp = vp->v_mount;
2670 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2671 	mp->mnt_activevnodelistsize++;
2672 	mtx_unlock(&vnode_free_list_mtx);
2673 	refcount_acquire(&vp->v_holdcnt);
2674 	if (!locked)
2675 		VI_UNLOCK(vp);
2676 }
2677 
2678 /*
2679  * Drop the hold count of the vnode.  If this is the last reference to
2680  * the vnode we place it on the free list unless it has been vgone'd
2681  * (marked VI_DOOMED) in which case we will free it.
2682  *
2683  * Because the vnode vm object keeps a hold reference on the vnode if
2684  * there is at least one resident non-cached page, the vnode cannot
2685  * leave the active list without the page cleanup done.
2686  */
2687 void
2688 _vdrop(struct vnode *vp, bool locked)
2689 {
2690 	struct bufobj *bo;
2691 	struct mount *mp;
2692 	int active;
2693 
2694 	if (locked)
2695 		ASSERT_VI_LOCKED(vp, __func__);
2696 	else
2697 		ASSERT_VI_UNLOCKED(vp, __func__);
2698 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2699 	if ((int)vp->v_holdcnt <= 0)
2700 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2701 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2702 		if (locked)
2703 			VI_UNLOCK(vp);
2704 		return;
2705 	}
2706 
2707 	if (!locked)
2708 		VI_LOCK(vp);
2709 	if (refcount_release(&vp->v_holdcnt) == 0) {
2710 		VI_UNLOCK(vp);
2711 		return;
2712 	}
2713 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2714 		/*
2715 		 * Mark a vnode as free: remove it from its active list
2716 		 * and put it up for recycling on the freelist.
2717 		 */
2718 		VNASSERT(vp->v_op != NULL, vp,
2719 		    ("vdropl: vnode already reclaimed."));
2720 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2721 		    ("vnode already free"));
2722 		VNASSERT(vp->v_holdcnt == 0, vp,
2723 		    ("vdropl: freeing when we shouldn't"));
2724 		active = vp->v_iflag & VI_ACTIVE;
2725 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2726 			vp->v_iflag &= ~VI_ACTIVE;
2727 			mp = vp->v_mount;
2728 			mtx_lock(&vnode_free_list_mtx);
2729 			if (active) {
2730 				TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2731 				    v_actfreelist);
2732 				mp->mnt_activevnodelistsize--;
2733 			}
2734 			TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2735 			    v_actfreelist);
2736 			freevnodes++;
2737 			vp->v_iflag |= VI_FREE;
2738 			mtx_unlock(&vnode_free_list_mtx);
2739 		} else {
2740 			atomic_add_long(&free_owe_inact, 1);
2741 		}
2742 		VI_UNLOCK(vp);
2743 		return;
2744 	}
2745 	/*
2746 	 * The vnode has been marked for destruction, so free it.
2747 	 *
2748 	 * The vnode will be returned to the zone where it will
2749 	 * normally remain until it is needed for another vnode. We
2750 	 * need to cleanup (or verify that the cleanup has already
2751 	 * been done) any residual data left from its current use
2752 	 * so as not to contaminate the freshly allocated vnode.
2753 	 */
2754 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2755 	atomic_subtract_long(&numvnodes, 1);
2756 	bo = &vp->v_bufobj;
2757 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2758 	    ("cleaned vnode still on the free list."));
2759 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2760 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2761 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2762 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2763 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2764 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2765 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2766 	    ("clean blk trie not empty"));
2767 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2768 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2769 	    ("dirty blk trie not empty"));
2770 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2771 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2772 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2773 	VI_UNLOCK(vp);
2774 #ifdef MAC
2775 	mac_vnode_destroy(vp);
2776 #endif
2777 	if (vp->v_pollinfo != NULL) {
2778 		destroy_vpollinfo(vp->v_pollinfo);
2779 		vp->v_pollinfo = NULL;
2780 	}
2781 #ifdef INVARIANTS
2782 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2783 	vp->v_op = NULL;
2784 #endif
2785 	vp->v_iflag = 0;
2786 	vp->v_vflag = 0;
2787 	bo->bo_flag = 0;
2788 	uma_zfree(vnode_zone, vp);
2789 }
2790 
2791 /*
2792  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2793  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2794  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2795  * failed lock upgrade.
2796  */
2797 void
2798 vinactive(struct vnode *vp, struct thread *td)
2799 {
2800 	struct vm_object *obj;
2801 
2802 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2803 	ASSERT_VI_LOCKED(vp, "vinactive");
2804 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2805 	    ("vinactive: recursed on VI_DOINGINACT"));
2806 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2807 	vp->v_iflag |= VI_DOINGINACT;
2808 	vp->v_iflag &= ~VI_OWEINACT;
2809 	VI_UNLOCK(vp);
2810 	/*
2811 	 * Before moving off the active list, we must be sure that any
2812 	 * modified pages are converted into the vnode's dirty
2813 	 * buffers, since these will no longer be checked once the
2814 	 * vnode is on the inactive list.
2815 	 *
2816 	 * The write-out of the dirty pages is asynchronous.  At the
2817 	 * point that VOP_INACTIVE() is called, there could still be
2818 	 * pending I/O and dirty pages in the object.
2819 	 */
2820 	obj = vp->v_object;
2821 	if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2822 		VM_OBJECT_WLOCK(obj);
2823 		vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC);
2824 		VM_OBJECT_WUNLOCK(obj);
2825 	}
2826 	VOP_INACTIVE(vp, td);
2827 	VI_LOCK(vp);
2828 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2829 	    ("vinactive: lost VI_DOINGINACT"));
2830 	vp->v_iflag &= ~VI_DOINGINACT;
2831 }
2832 
2833 /*
2834  * Remove any vnodes in the vnode table belonging to mount point mp.
2835  *
2836  * If FORCECLOSE is not specified, there should not be any active ones,
2837  * return error if any are found (nb: this is a user error, not a
2838  * system error). If FORCECLOSE is specified, detach any active vnodes
2839  * that are found.
2840  *
2841  * If WRITECLOSE is set, only flush out regular file vnodes open for
2842  * writing.
2843  *
2844  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2845  *
2846  * `rootrefs' specifies the base reference count for the root vnode
2847  * of this filesystem. The root vnode is considered busy if its
2848  * v_usecount exceeds this value. On a successful return, vflush(, td)
2849  * will call vrele() on the root vnode exactly rootrefs times.
2850  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2851  * be zero.
2852  */
2853 #ifdef DIAGNOSTIC
2854 static int busyprt = 0;		/* print out busy vnodes */
2855 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2856 #endif
2857 
2858 int
2859 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2860 {
2861 	struct vnode *vp, *mvp, *rootvp = NULL;
2862 	struct vattr vattr;
2863 	int busy = 0, error;
2864 
2865 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2866 	    rootrefs, flags);
2867 	if (rootrefs > 0) {
2868 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2869 		    ("vflush: bad args"));
2870 		/*
2871 		 * Get the filesystem root vnode. We can vput() it
2872 		 * immediately, since with rootrefs > 0, it won't go away.
2873 		 */
2874 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2875 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2876 			    __func__, error);
2877 			return (error);
2878 		}
2879 		vput(rootvp);
2880 	}
2881 loop:
2882 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2883 		vholdl(vp);
2884 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2885 		if (error) {
2886 			vdrop(vp);
2887 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2888 			goto loop;
2889 		}
2890 		/*
2891 		 * Skip over a vnodes marked VV_SYSTEM.
2892 		 */
2893 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2894 			VOP_UNLOCK(vp, 0);
2895 			vdrop(vp);
2896 			continue;
2897 		}
2898 		/*
2899 		 * If WRITECLOSE is set, flush out unlinked but still open
2900 		 * files (even if open only for reading) and regular file
2901 		 * vnodes open for writing.
2902 		 */
2903 		if (flags & WRITECLOSE) {
2904 			if (vp->v_object != NULL) {
2905 				VM_OBJECT_WLOCK(vp->v_object);
2906 				vm_object_page_clean(vp->v_object, 0, 0, 0);
2907 				VM_OBJECT_WUNLOCK(vp->v_object);
2908 			}
2909 			error = VOP_FSYNC(vp, MNT_WAIT, td);
2910 			if (error != 0) {
2911 				VOP_UNLOCK(vp, 0);
2912 				vdrop(vp);
2913 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2914 				return (error);
2915 			}
2916 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
2917 			VI_LOCK(vp);
2918 
2919 			if ((vp->v_type == VNON ||
2920 			    (error == 0 && vattr.va_nlink > 0)) &&
2921 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2922 				VOP_UNLOCK(vp, 0);
2923 				vdropl(vp);
2924 				continue;
2925 			}
2926 		} else
2927 			VI_LOCK(vp);
2928 		/*
2929 		 * With v_usecount == 0, all we need to do is clear out the
2930 		 * vnode data structures and we are done.
2931 		 *
2932 		 * If FORCECLOSE is set, forcibly close the vnode.
2933 		 */
2934 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2935 			vgonel(vp);
2936 		} else {
2937 			busy++;
2938 #ifdef DIAGNOSTIC
2939 			if (busyprt)
2940 				vprint("vflush: busy vnode", vp);
2941 #endif
2942 		}
2943 		VOP_UNLOCK(vp, 0);
2944 		vdropl(vp);
2945 	}
2946 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2947 		/*
2948 		 * If just the root vnode is busy, and if its refcount
2949 		 * is equal to `rootrefs', then go ahead and kill it.
2950 		 */
2951 		VI_LOCK(rootvp);
2952 		KASSERT(busy > 0, ("vflush: not busy"));
2953 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2954 		    ("vflush: usecount %d < rootrefs %d",
2955 		     rootvp->v_usecount, rootrefs));
2956 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2957 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
2958 			vgone(rootvp);
2959 			VOP_UNLOCK(rootvp, 0);
2960 			busy = 0;
2961 		} else
2962 			VI_UNLOCK(rootvp);
2963 	}
2964 	if (busy) {
2965 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
2966 		    busy);
2967 		return (EBUSY);
2968 	}
2969 	for (; rootrefs > 0; rootrefs--)
2970 		vrele(rootvp);
2971 	return (0);
2972 }
2973 
2974 /*
2975  * Recycle an unused vnode to the front of the free list.
2976  */
2977 int
2978 vrecycle(struct vnode *vp)
2979 {
2980 	int recycled;
2981 
2982 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
2983 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2984 	recycled = 0;
2985 	VI_LOCK(vp);
2986 	if (vp->v_usecount == 0) {
2987 		recycled = 1;
2988 		vgonel(vp);
2989 	}
2990 	VI_UNLOCK(vp);
2991 	return (recycled);
2992 }
2993 
2994 /*
2995  * Eliminate all activity associated with a vnode
2996  * in preparation for reuse.
2997  */
2998 void
2999 vgone(struct vnode *vp)
3000 {
3001 	VI_LOCK(vp);
3002 	vgonel(vp);
3003 	VI_UNLOCK(vp);
3004 }
3005 
3006 static void
3007 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3008     struct vnode *lowervp __unused)
3009 {
3010 }
3011 
3012 /*
3013  * Notify upper mounts about reclaimed or unlinked vnode.
3014  */
3015 void
3016 vfs_notify_upper(struct vnode *vp, int event)
3017 {
3018 	static struct vfsops vgonel_vfsops = {
3019 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3020 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3021 	};
3022 	struct mount *mp, *ump, *mmp;
3023 
3024 	mp = vp->v_mount;
3025 	if (mp == NULL)
3026 		return;
3027 
3028 	MNT_ILOCK(mp);
3029 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3030 		goto unlock;
3031 	MNT_IUNLOCK(mp);
3032 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3033 	mmp->mnt_op = &vgonel_vfsops;
3034 	mmp->mnt_kern_flag |= MNTK_MARKER;
3035 	MNT_ILOCK(mp);
3036 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3037 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3038 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3039 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3040 			continue;
3041 		}
3042 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3043 		MNT_IUNLOCK(mp);
3044 		switch (event) {
3045 		case VFS_NOTIFY_UPPER_RECLAIM:
3046 			VFS_RECLAIM_LOWERVP(ump, vp);
3047 			break;
3048 		case VFS_NOTIFY_UPPER_UNLINK:
3049 			VFS_UNLINK_LOWERVP(ump, vp);
3050 			break;
3051 		default:
3052 			KASSERT(0, ("invalid event %d", event));
3053 			break;
3054 		}
3055 		MNT_ILOCK(mp);
3056 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3057 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3058 	}
3059 	free(mmp, M_TEMP);
3060 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3061 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3062 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3063 		wakeup(&mp->mnt_uppers);
3064 	}
3065 unlock:
3066 	MNT_IUNLOCK(mp);
3067 }
3068 
3069 /*
3070  * vgone, with the vp interlock held.
3071  */
3072 static void
3073 vgonel(struct vnode *vp)
3074 {
3075 	struct thread *td;
3076 	int oweinact;
3077 	int active;
3078 	struct mount *mp;
3079 
3080 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3081 	ASSERT_VI_LOCKED(vp, "vgonel");
3082 	VNASSERT(vp->v_holdcnt, vp,
3083 	    ("vgonel: vp %p has no reference.", vp));
3084 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3085 	td = curthread;
3086 
3087 	/*
3088 	 * Don't vgonel if we're already doomed.
3089 	 */
3090 	if (vp->v_iflag & VI_DOOMED)
3091 		return;
3092 	vp->v_iflag |= VI_DOOMED;
3093 
3094 	/*
3095 	 * Check to see if the vnode is in use.  If so, we have to call
3096 	 * VOP_CLOSE() and VOP_INACTIVE().
3097 	 */
3098 	active = vp->v_usecount;
3099 	oweinact = (vp->v_iflag & VI_OWEINACT);
3100 	VI_UNLOCK(vp);
3101 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3102 
3103 	/*
3104 	 * If purging an active vnode, it must be closed and
3105 	 * deactivated before being reclaimed.
3106 	 */
3107 	if (active)
3108 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3109 	if (oweinact || active) {
3110 		VI_LOCK(vp);
3111 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3112 			vinactive(vp, td);
3113 		VI_UNLOCK(vp);
3114 	}
3115 	if (vp->v_type == VSOCK)
3116 		vfs_unp_reclaim(vp);
3117 
3118 	/*
3119 	 * Clean out any buffers associated with the vnode.
3120 	 * If the flush fails, just toss the buffers.
3121 	 */
3122 	mp = NULL;
3123 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3124 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3125 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3126 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3127 			;
3128 	}
3129 
3130 	BO_LOCK(&vp->v_bufobj);
3131 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3132 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3133 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3134 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3135 	    ("vp %p bufobj not invalidated", vp));
3136 	vp->v_bufobj.bo_flag |= BO_DEAD;
3137 	BO_UNLOCK(&vp->v_bufobj);
3138 
3139 	/*
3140 	 * Reclaim the vnode.
3141 	 */
3142 	if (VOP_RECLAIM(vp, td))
3143 		panic("vgone: cannot reclaim");
3144 	if (mp != NULL)
3145 		vn_finished_secondary_write(mp);
3146 	VNASSERT(vp->v_object == NULL, vp,
3147 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3148 	/*
3149 	 * Clear the advisory locks and wake up waiting threads.
3150 	 */
3151 	(void)VOP_ADVLOCKPURGE(vp);
3152 	vp->v_lockf = NULL;
3153 	/*
3154 	 * Delete from old mount point vnode list.
3155 	 */
3156 	delmntque(vp);
3157 	cache_purge(vp);
3158 	/*
3159 	 * Done with purge, reset to the standard lock and invalidate
3160 	 * the vnode.
3161 	 */
3162 	VI_LOCK(vp);
3163 	vp->v_vnlock = &vp->v_lock;
3164 	vp->v_op = &dead_vnodeops;
3165 	vp->v_tag = "none";
3166 	vp->v_type = VBAD;
3167 }
3168 
3169 /*
3170  * Calculate the total number of references to a special device.
3171  */
3172 int
3173 vcount(struct vnode *vp)
3174 {
3175 	int count;
3176 
3177 	dev_lock();
3178 	count = vp->v_rdev->si_usecount;
3179 	dev_unlock();
3180 	return (count);
3181 }
3182 
3183 /*
3184  * Same as above, but using the struct cdev *as argument
3185  */
3186 int
3187 count_dev(struct cdev *dev)
3188 {
3189 	int count;
3190 
3191 	dev_lock();
3192 	count = dev->si_usecount;
3193 	dev_unlock();
3194 	return(count);
3195 }
3196 
3197 /*
3198  * Print out a description of a vnode.
3199  */
3200 static char *typename[] =
3201 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3202  "VMARKER"};
3203 
3204 void
3205 vn_printf(struct vnode *vp, const char *fmt, ...)
3206 {
3207 	va_list ap;
3208 	char buf[256], buf2[16];
3209 	u_long flags;
3210 
3211 	va_start(ap, fmt);
3212 	vprintf(fmt, ap);
3213 	va_end(ap);
3214 	printf("%p: ", (void *)vp);
3215 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3216 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3217 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3218 	buf[0] = '\0';
3219 	buf[1] = '\0';
3220 	if (vp->v_vflag & VV_ROOT)
3221 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3222 	if (vp->v_vflag & VV_ISTTY)
3223 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3224 	if (vp->v_vflag & VV_NOSYNC)
3225 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3226 	if (vp->v_vflag & VV_ETERNALDEV)
3227 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3228 	if (vp->v_vflag & VV_CACHEDLABEL)
3229 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3230 	if (vp->v_vflag & VV_TEXT)
3231 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3232 	if (vp->v_vflag & VV_COPYONWRITE)
3233 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3234 	if (vp->v_vflag & VV_SYSTEM)
3235 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3236 	if (vp->v_vflag & VV_PROCDEP)
3237 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3238 	if (vp->v_vflag & VV_NOKNOTE)
3239 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3240 	if (vp->v_vflag & VV_DELETED)
3241 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3242 	if (vp->v_vflag & VV_MD)
3243 		strlcat(buf, "|VV_MD", sizeof(buf));
3244 	if (vp->v_vflag & VV_FORCEINSMQ)
3245 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3246 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3247 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3248 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3249 	if (flags != 0) {
3250 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3251 		strlcat(buf, buf2, sizeof(buf));
3252 	}
3253 	if (vp->v_iflag & VI_MOUNT)
3254 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3255 	if (vp->v_iflag & VI_DOOMED)
3256 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3257 	if (vp->v_iflag & VI_FREE)
3258 		strlcat(buf, "|VI_FREE", sizeof(buf));
3259 	if (vp->v_iflag & VI_ACTIVE)
3260 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3261 	if (vp->v_iflag & VI_DOINGINACT)
3262 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3263 	if (vp->v_iflag & VI_OWEINACT)
3264 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3265 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3266 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3267 	if (flags != 0) {
3268 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3269 		strlcat(buf, buf2, sizeof(buf));
3270 	}
3271 	printf("    flags (%s)\n", buf + 1);
3272 	if (mtx_owned(VI_MTX(vp)))
3273 		printf(" VI_LOCKed");
3274 	if (vp->v_object != NULL)
3275 		printf("    v_object %p ref %d pages %d "
3276 		    "cleanbuf %d dirtybuf %d\n",
3277 		    vp->v_object, vp->v_object->ref_count,
3278 		    vp->v_object->resident_page_count,
3279 		    vp->v_bufobj.bo_clean.bv_cnt,
3280 		    vp->v_bufobj.bo_dirty.bv_cnt);
3281 	printf("    ");
3282 	lockmgr_printinfo(vp->v_vnlock);
3283 	if (vp->v_data != NULL)
3284 		VOP_PRINT(vp);
3285 }
3286 
3287 #ifdef DDB
3288 /*
3289  * List all of the locked vnodes in the system.
3290  * Called when debugging the kernel.
3291  */
3292 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3293 {
3294 	struct mount *mp;
3295 	struct vnode *vp;
3296 
3297 	/*
3298 	 * Note: because this is DDB, we can't obey the locking semantics
3299 	 * for these structures, which means we could catch an inconsistent
3300 	 * state and dereference a nasty pointer.  Not much to be done
3301 	 * about that.
3302 	 */
3303 	db_printf("Locked vnodes\n");
3304 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3305 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3306 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3307 				vprint("", vp);
3308 		}
3309 	}
3310 }
3311 
3312 /*
3313  * Show details about the given vnode.
3314  */
3315 DB_SHOW_COMMAND(vnode, db_show_vnode)
3316 {
3317 	struct vnode *vp;
3318 
3319 	if (!have_addr)
3320 		return;
3321 	vp = (struct vnode *)addr;
3322 	vn_printf(vp, "vnode ");
3323 }
3324 
3325 /*
3326  * Show details about the given mount point.
3327  */
3328 DB_SHOW_COMMAND(mount, db_show_mount)
3329 {
3330 	struct mount *mp;
3331 	struct vfsopt *opt;
3332 	struct statfs *sp;
3333 	struct vnode *vp;
3334 	char buf[512];
3335 	uint64_t mflags;
3336 	u_int flags;
3337 
3338 	if (!have_addr) {
3339 		/* No address given, print short info about all mount points. */
3340 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3341 			db_printf("%p %s on %s (%s)\n", mp,
3342 			    mp->mnt_stat.f_mntfromname,
3343 			    mp->mnt_stat.f_mntonname,
3344 			    mp->mnt_stat.f_fstypename);
3345 			if (db_pager_quit)
3346 				break;
3347 		}
3348 		db_printf("\nMore info: show mount <addr>\n");
3349 		return;
3350 	}
3351 
3352 	mp = (struct mount *)addr;
3353 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3354 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3355 
3356 	buf[0] = '\0';
3357 	mflags = mp->mnt_flag;
3358 #define	MNT_FLAG(flag)	do {						\
3359 	if (mflags & (flag)) {						\
3360 		if (buf[0] != '\0')					\
3361 			strlcat(buf, ", ", sizeof(buf));		\
3362 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3363 		mflags &= ~(flag);					\
3364 	}								\
3365 } while (0)
3366 	MNT_FLAG(MNT_RDONLY);
3367 	MNT_FLAG(MNT_SYNCHRONOUS);
3368 	MNT_FLAG(MNT_NOEXEC);
3369 	MNT_FLAG(MNT_NOSUID);
3370 	MNT_FLAG(MNT_NFS4ACLS);
3371 	MNT_FLAG(MNT_UNION);
3372 	MNT_FLAG(MNT_ASYNC);
3373 	MNT_FLAG(MNT_SUIDDIR);
3374 	MNT_FLAG(MNT_SOFTDEP);
3375 	MNT_FLAG(MNT_NOSYMFOLLOW);
3376 	MNT_FLAG(MNT_GJOURNAL);
3377 	MNT_FLAG(MNT_MULTILABEL);
3378 	MNT_FLAG(MNT_ACLS);
3379 	MNT_FLAG(MNT_NOATIME);
3380 	MNT_FLAG(MNT_NOCLUSTERR);
3381 	MNT_FLAG(MNT_NOCLUSTERW);
3382 	MNT_FLAG(MNT_SUJ);
3383 	MNT_FLAG(MNT_EXRDONLY);
3384 	MNT_FLAG(MNT_EXPORTED);
3385 	MNT_FLAG(MNT_DEFEXPORTED);
3386 	MNT_FLAG(MNT_EXPORTANON);
3387 	MNT_FLAG(MNT_EXKERB);
3388 	MNT_FLAG(MNT_EXPUBLIC);
3389 	MNT_FLAG(MNT_LOCAL);
3390 	MNT_FLAG(MNT_QUOTA);
3391 	MNT_FLAG(MNT_ROOTFS);
3392 	MNT_FLAG(MNT_USER);
3393 	MNT_FLAG(MNT_IGNORE);
3394 	MNT_FLAG(MNT_UPDATE);
3395 	MNT_FLAG(MNT_DELEXPORT);
3396 	MNT_FLAG(MNT_RELOAD);
3397 	MNT_FLAG(MNT_FORCE);
3398 	MNT_FLAG(MNT_SNAPSHOT);
3399 	MNT_FLAG(MNT_BYFSID);
3400 #undef MNT_FLAG
3401 	if (mflags != 0) {
3402 		if (buf[0] != '\0')
3403 			strlcat(buf, ", ", sizeof(buf));
3404 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3405 		    "0x%016jx", mflags);
3406 	}
3407 	db_printf("    mnt_flag = %s\n", buf);
3408 
3409 	buf[0] = '\0';
3410 	flags = mp->mnt_kern_flag;
3411 #define	MNT_KERN_FLAG(flag)	do {					\
3412 	if (flags & (flag)) {						\
3413 		if (buf[0] != '\0')					\
3414 			strlcat(buf, ", ", sizeof(buf));		\
3415 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3416 		flags &= ~(flag);					\
3417 	}								\
3418 } while (0)
3419 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3420 	MNT_KERN_FLAG(MNTK_ASYNC);
3421 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3422 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3423 	MNT_KERN_FLAG(MNTK_DRAINING);
3424 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3425 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3426 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3427 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3428 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3429 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3430 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3431 	MNT_KERN_FLAG(MNTK_MARKER);
3432 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3433 	MNT_KERN_FLAG(MNTK_NOASYNC);
3434 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3435 	MNT_KERN_FLAG(MNTK_MWAIT);
3436 	MNT_KERN_FLAG(MNTK_SUSPEND);
3437 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3438 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3439 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3440 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3441 #undef MNT_KERN_FLAG
3442 	if (flags != 0) {
3443 		if (buf[0] != '\0')
3444 			strlcat(buf, ", ", sizeof(buf));
3445 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3446 		    "0x%08x", flags);
3447 	}
3448 	db_printf("    mnt_kern_flag = %s\n", buf);
3449 
3450 	db_printf("    mnt_opt = ");
3451 	opt = TAILQ_FIRST(mp->mnt_opt);
3452 	if (opt != NULL) {
3453 		db_printf("%s", opt->name);
3454 		opt = TAILQ_NEXT(opt, link);
3455 		while (opt != NULL) {
3456 			db_printf(", %s", opt->name);
3457 			opt = TAILQ_NEXT(opt, link);
3458 		}
3459 	}
3460 	db_printf("\n");
3461 
3462 	sp = &mp->mnt_stat;
3463 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3464 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3465 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3466 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3467 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3468 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3469 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3470 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3471 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3472 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3473 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3474 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3475 
3476 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3477 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3478 	if (jailed(mp->mnt_cred))
3479 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3480 	db_printf(" }\n");
3481 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3482 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3483 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3484 	db_printf("    mnt_activevnodelistsize = %d\n",
3485 	    mp->mnt_activevnodelistsize);
3486 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3487 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3488 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3489 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3490 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3491 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3492 	db_printf("    mnt_secondary_accwrites = %d\n",
3493 	    mp->mnt_secondary_accwrites);
3494 	db_printf("    mnt_gjprovider = %s\n",
3495 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3496 
3497 	db_printf("\n\nList of active vnodes\n");
3498 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3499 		if (vp->v_type != VMARKER) {
3500 			vn_printf(vp, "vnode ");
3501 			if (db_pager_quit)
3502 				break;
3503 		}
3504 	}
3505 	db_printf("\n\nList of inactive vnodes\n");
3506 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3507 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3508 			vn_printf(vp, "vnode ");
3509 			if (db_pager_quit)
3510 				break;
3511 		}
3512 	}
3513 }
3514 #endif	/* DDB */
3515 
3516 /*
3517  * Fill in a struct xvfsconf based on a struct vfsconf.
3518  */
3519 static int
3520 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3521 {
3522 	struct xvfsconf xvfsp;
3523 
3524 	bzero(&xvfsp, sizeof(xvfsp));
3525 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3526 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3527 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3528 	xvfsp.vfc_flags = vfsp->vfc_flags;
3529 	/*
3530 	 * These are unused in userland, we keep them
3531 	 * to not break binary compatibility.
3532 	 */
3533 	xvfsp.vfc_vfsops = NULL;
3534 	xvfsp.vfc_next = NULL;
3535 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3536 }
3537 
3538 #ifdef COMPAT_FREEBSD32
3539 struct xvfsconf32 {
3540 	uint32_t	vfc_vfsops;
3541 	char		vfc_name[MFSNAMELEN];
3542 	int32_t		vfc_typenum;
3543 	int32_t		vfc_refcount;
3544 	int32_t		vfc_flags;
3545 	uint32_t	vfc_next;
3546 };
3547 
3548 static int
3549 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3550 {
3551 	struct xvfsconf32 xvfsp;
3552 
3553 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3554 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3555 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3556 	xvfsp.vfc_flags = vfsp->vfc_flags;
3557 	xvfsp.vfc_vfsops = 0;
3558 	xvfsp.vfc_next = 0;
3559 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3560 }
3561 #endif
3562 
3563 /*
3564  * Top level filesystem related information gathering.
3565  */
3566 static int
3567 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3568 {
3569 	struct vfsconf *vfsp;
3570 	int error;
3571 
3572 	error = 0;
3573 	vfsconf_slock();
3574 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3575 #ifdef COMPAT_FREEBSD32
3576 		if (req->flags & SCTL_MASK32)
3577 			error = vfsconf2x32(req, vfsp);
3578 		else
3579 #endif
3580 			error = vfsconf2x(req, vfsp);
3581 		if (error)
3582 			break;
3583 	}
3584 	vfsconf_sunlock();
3585 	return (error);
3586 }
3587 
3588 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3589     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3590     "S,xvfsconf", "List of all configured filesystems");
3591 
3592 #ifndef BURN_BRIDGES
3593 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3594 
3595 static int
3596 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3597 {
3598 	int *name = (int *)arg1 - 1;	/* XXX */
3599 	u_int namelen = arg2 + 1;	/* XXX */
3600 	struct vfsconf *vfsp;
3601 
3602 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3603 	    "please rebuild world\n");
3604 
3605 #if 1 || defined(COMPAT_PRELITE2)
3606 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3607 	if (namelen == 1)
3608 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3609 #endif
3610 
3611 	switch (name[1]) {
3612 	case VFS_MAXTYPENUM:
3613 		if (namelen != 2)
3614 			return (ENOTDIR);
3615 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3616 	case VFS_CONF:
3617 		if (namelen != 3)
3618 			return (ENOTDIR);	/* overloaded */
3619 		vfsconf_slock();
3620 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3621 			if (vfsp->vfc_typenum == name[2])
3622 				break;
3623 		}
3624 		vfsconf_sunlock();
3625 		if (vfsp == NULL)
3626 			return (EOPNOTSUPP);
3627 #ifdef COMPAT_FREEBSD32
3628 		if (req->flags & SCTL_MASK32)
3629 			return (vfsconf2x32(req, vfsp));
3630 		else
3631 #endif
3632 			return (vfsconf2x(req, vfsp));
3633 	}
3634 	return (EOPNOTSUPP);
3635 }
3636 
3637 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3638     CTLFLAG_MPSAFE, vfs_sysctl,
3639     "Generic filesystem");
3640 
3641 #if 1 || defined(COMPAT_PRELITE2)
3642 
3643 static int
3644 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3645 {
3646 	int error;
3647 	struct vfsconf *vfsp;
3648 	struct ovfsconf ovfs;
3649 
3650 	vfsconf_slock();
3651 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3652 		bzero(&ovfs, sizeof(ovfs));
3653 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3654 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3655 		ovfs.vfc_index = vfsp->vfc_typenum;
3656 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3657 		ovfs.vfc_flags = vfsp->vfc_flags;
3658 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3659 		if (error != 0) {
3660 			vfsconf_sunlock();
3661 			return (error);
3662 		}
3663 	}
3664 	vfsconf_sunlock();
3665 	return (0);
3666 }
3667 
3668 #endif /* 1 || COMPAT_PRELITE2 */
3669 #endif /* !BURN_BRIDGES */
3670 
3671 #define KINFO_VNODESLOP		10
3672 #ifdef notyet
3673 /*
3674  * Dump vnode list (via sysctl).
3675  */
3676 /* ARGSUSED */
3677 static int
3678 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3679 {
3680 	struct xvnode *xvn;
3681 	struct mount *mp;
3682 	struct vnode *vp;
3683 	int error, len, n;
3684 
3685 	/*
3686 	 * Stale numvnodes access is not fatal here.
3687 	 */
3688 	req->lock = 0;
3689 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3690 	if (!req->oldptr)
3691 		/* Make an estimate */
3692 		return (SYSCTL_OUT(req, 0, len));
3693 
3694 	error = sysctl_wire_old_buffer(req, 0);
3695 	if (error != 0)
3696 		return (error);
3697 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3698 	n = 0;
3699 	mtx_lock(&mountlist_mtx);
3700 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3701 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3702 			continue;
3703 		MNT_ILOCK(mp);
3704 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3705 			if (n == len)
3706 				break;
3707 			vref(vp);
3708 			xvn[n].xv_size = sizeof *xvn;
3709 			xvn[n].xv_vnode = vp;
3710 			xvn[n].xv_id = 0;	/* XXX compat */
3711 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3712 			XV_COPY(usecount);
3713 			XV_COPY(writecount);
3714 			XV_COPY(holdcnt);
3715 			XV_COPY(mount);
3716 			XV_COPY(numoutput);
3717 			XV_COPY(type);
3718 #undef XV_COPY
3719 			xvn[n].xv_flag = vp->v_vflag;
3720 
3721 			switch (vp->v_type) {
3722 			case VREG:
3723 			case VDIR:
3724 			case VLNK:
3725 				break;
3726 			case VBLK:
3727 			case VCHR:
3728 				if (vp->v_rdev == NULL) {
3729 					vrele(vp);
3730 					continue;
3731 				}
3732 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3733 				break;
3734 			case VSOCK:
3735 				xvn[n].xv_socket = vp->v_socket;
3736 				break;
3737 			case VFIFO:
3738 				xvn[n].xv_fifo = vp->v_fifoinfo;
3739 				break;
3740 			case VNON:
3741 			case VBAD:
3742 			default:
3743 				/* shouldn't happen? */
3744 				vrele(vp);
3745 				continue;
3746 			}
3747 			vrele(vp);
3748 			++n;
3749 		}
3750 		MNT_IUNLOCK(mp);
3751 		mtx_lock(&mountlist_mtx);
3752 		vfs_unbusy(mp);
3753 		if (n == len)
3754 			break;
3755 	}
3756 	mtx_unlock(&mountlist_mtx);
3757 
3758 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3759 	free(xvn, M_TEMP);
3760 	return (error);
3761 }
3762 
3763 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3764     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3765     "");
3766 #endif
3767 
3768 static void
3769 unmount_or_warn(struct mount *mp)
3770 {
3771 	int error;
3772 
3773 	error = dounmount(mp, MNT_FORCE, curthread);
3774 	if (error != 0) {
3775 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3776 		if (error == EBUSY)
3777 			printf("BUSY)\n");
3778 		else
3779 			printf("%d)\n", error);
3780 	}
3781 }
3782 
3783 /*
3784  * Unmount all filesystems. The list is traversed in reverse order
3785  * of mounting to avoid dependencies.
3786  */
3787 void
3788 vfs_unmountall(void)
3789 {
3790 	struct mount *mp, *tmp;
3791 
3792 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3793 
3794 	/*
3795 	 * Since this only runs when rebooting, it is not interlocked.
3796 	 */
3797 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3798 		vfs_ref(mp);
3799 
3800 		/*
3801 		 * Forcibly unmounting "/dev" before "/" would prevent clean
3802 		 * unmount of the latter.
3803 		 */
3804 		if (mp == rootdevmp)
3805 			continue;
3806 
3807 		unmount_or_warn(mp);
3808 	}
3809 
3810 	if (rootdevmp != NULL)
3811 		unmount_or_warn(rootdevmp);
3812 }
3813 
3814 /*
3815  * perform msync on all vnodes under a mount point
3816  * the mount point must be locked.
3817  */
3818 void
3819 vfs_msync(struct mount *mp, int flags)
3820 {
3821 	struct vnode *vp, *mvp;
3822 	struct vm_object *obj;
3823 
3824 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3825 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3826 		obj = vp->v_object;
3827 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3828 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3829 			if (!vget(vp,
3830 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3831 			    curthread)) {
3832 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3833 					vput(vp);
3834 					continue;
3835 				}
3836 
3837 				obj = vp->v_object;
3838 				if (obj != NULL) {
3839 					VM_OBJECT_WLOCK(obj);
3840 					vm_object_page_clean(obj, 0, 0,
3841 					    flags == MNT_WAIT ?
3842 					    OBJPC_SYNC : OBJPC_NOSYNC);
3843 					VM_OBJECT_WUNLOCK(obj);
3844 				}
3845 				vput(vp);
3846 			}
3847 		} else
3848 			VI_UNLOCK(vp);
3849 	}
3850 }
3851 
3852 static void
3853 destroy_vpollinfo_free(struct vpollinfo *vi)
3854 {
3855 
3856 	knlist_destroy(&vi->vpi_selinfo.si_note);
3857 	mtx_destroy(&vi->vpi_lock);
3858 	uma_zfree(vnodepoll_zone, vi);
3859 }
3860 
3861 static void
3862 destroy_vpollinfo(struct vpollinfo *vi)
3863 {
3864 
3865 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
3866 	seldrain(&vi->vpi_selinfo);
3867 	destroy_vpollinfo_free(vi);
3868 }
3869 
3870 /*
3871  * Initalize per-vnode helper structure to hold poll-related state.
3872  */
3873 void
3874 v_addpollinfo(struct vnode *vp)
3875 {
3876 	struct vpollinfo *vi;
3877 
3878 	if (vp->v_pollinfo != NULL)
3879 		return;
3880 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
3881 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3882 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3883 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3884 	VI_LOCK(vp);
3885 	if (vp->v_pollinfo != NULL) {
3886 		VI_UNLOCK(vp);
3887 		destroy_vpollinfo_free(vi);
3888 		return;
3889 	}
3890 	vp->v_pollinfo = vi;
3891 	VI_UNLOCK(vp);
3892 }
3893 
3894 /*
3895  * Record a process's interest in events which might happen to
3896  * a vnode.  Because poll uses the historic select-style interface
3897  * internally, this routine serves as both the ``check for any
3898  * pending events'' and the ``record my interest in future events''
3899  * functions.  (These are done together, while the lock is held,
3900  * to avoid race conditions.)
3901  */
3902 int
3903 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3904 {
3905 
3906 	v_addpollinfo(vp);
3907 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3908 	if (vp->v_pollinfo->vpi_revents & events) {
3909 		/*
3910 		 * This leaves events we are not interested
3911 		 * in available for the other process which
3912 		 * which presumably had requested them
3913 		 * (otherwise they would never have been
3914 		 * recorded).
3915 		 */
3916 		events &= vp->v_pollinfo->vpi_revents;
3917 		vp->v_pollinfo->vpi_revents &= ~events;
3918 
3919 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3920 		return (events);
3921 	}
3922 	vp->v_pollinfo->vpi_events |= events;
3923 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3924 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3925 	return (0);
3926 }
3927 
3928 /*
3929  * Routine to create and manage a filesystem syncer vnode.
3930  */
3931 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3932 static int	sync_fsync(struct  vop_fsync_args *);
3933 static int	sync_inactive(struct  vop_inactive_args *);
3934 static int	sync_reclaim(struct  vop_reclaim_args *);
3935 
3936 static struct vop_vector sync_vnodeops = {
3937 	.vop_bypass =	VOP_EOPNOTSUPP,
3938 	.vop_close =	sync_close,		/* close */
3939 	.vop_fsync =	sync_fsync,		/* fsync */
3940 	.vop_inactive =	sync_inactive,	/* inactive */
3941 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3942 	.vop_lock1 =	vop_stdlock,	/* lock */
3943 	.vop_unlock =	vop_stdunlock,	/* unlock */
3944 	.vop_islocked =	vop_stdislocked,	/* islocked */
3945 };
3946 
3947 /*
3948  * Create a new filesystem syncer vnode for the specified mount point.
3949  */
3950 void
3951 vfs_allocate_syncvnode(struct mount *mp)
3952 {
3953 	struct vnode *vp;
3954 	struct bufobj *bo;
3955 	static long start, incr, next;
3956 	int error;
3957 
3958 	/* Allocate a new vnode */
3959 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
3960 	if (error != 0)
3961 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
3962 	vp->v_type = VNON;
3963 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3964 	vp->v_vflag |= VV_FORCEINSMQ;
3965 	error = insmntque(vp, mp);
3966 	if (error != 0)
3967 		panic("vfs_allocate_syncvnode: insmntque() failed");
3968 	vp->v_vflag &= ~VV_FORCEINSMQ;
3969 	VOP_UNLOCK(vp, 0);
3970 	/*
3971 	 * Place the vnode onto the syncer worklist. We attempt to
3972 	 * scatter them about on the list so that they will go off
3973 	 * at evenly distributed times even if all the filesystems
3974 	 * are mounted at once.
3975 	 */
3976 	next += incr;
3977 	if (next == 0 || next > syncer_maxdelay) {
3978 		start /= 2;
3979 		incr /= 2;
3980 		if (start == 0) {
3981 			start = syncer_maxdelay / 2;
3982 			incr = syncer_maxdelay;
3983 		}
3984 		next = start;
3985 	}
3986 	bo = &vp->v_bufobj;
3987 	BO_LOCK(bo);
3988 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
3989 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3990 	mtx_lock(&sync_mtx);
3991 	sync_vnode_count++;
3992 	if (mp->mnt_syncer == NULL) {
3993 		mp->mnt_syncer = vp;
3994 		vp = NULL;
3995 	}
3996 	mtx_unlock(&sync_mtx);
3997 	BO_UNLOCK(bo);
3998 	if (vp != NULL) {
3999 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4000 		vgone(vp);
4001 		vput(vp);
4002 	}
4003 }
4004 
4005 void
4006 vfs_deallocate_syncvnode(struct mount *mp)
4007 {
4008 	struct vnode *vp;
4009 
4010 	mtx_lock(&sync_mtx);
4011 	vp = mp->mnt_syncer;
4012 	if (vp != NULL)
4013 		mp->mnt_syncer = NULL;
4014 	mtx_unlock(&sync_mtx);
4015 	if (vp != NULL)
4016 		vrele(vp);
4017 }
4018 
4019 /*
4020  * Do a lazy sync of the filesystem.
4021  */
4022 static int
4023 sync_fsync(struct vop_fsync_args *ap)
4024 {
4025 	struct vnode *syncvp = ap->a_vp;
4026 	struct mount *mp = syncvp->v_mount;
4027 	int error, save;
4028 	struct bufobj *bo;
4029 
4030 	/*
4031 	 * We only need to do something if this is a lazy evaluation.
4032 	 */
4033 	if (ap->a_waitfor != MNT_LAZY)
4034 		return (0);
4035 
4036 	/*
4037 	 * Move ourselves to the back of the sync list.
4038 	 */
4039 	bo = &syncvp->v_bufobj;
4040 	BO_LOCK(bo);
4041 	vn_syncer_add_to_worklist(bo, syncdelay);
4042 	BO_UNLOCK(bo);
4043 
4044 	/*
4045 	 * Walk the list of vnodes pushing all that are dirty and
4046 	 * not already on the sync list.
4047 	 */
4048 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4049 		return (0);
4050 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4051 		vfs_unbusy(mp);
4052 		return (0);
4053 	}
4054 	save = curthread_pflags_set(TDP_SYNCIO);
4055 	vfs_msync(mp, MNT_NOWAIT);
4056 	error = VFS_SYNC(mp, MNT_LAZY);
4057 	curthread_pflags_restore(save);
4058 	vn_finished_write(mp);
4059 	vfs_unbusy(mp);
4060 	return (error);
4061 }
4062 
4063 /*
4064  * The syncer vnode is no referenced.
4065  */
4066 static int
4067 sync_inactive(struct vop_inactive_args *ap)
4068 {
4069 
4070 	vgone(ap->a_vp);
4071 	return (0);
4072 }
4073 
4074 /*
4075  * The syncer vnode is no longer needed and is being decommissioned.
4076  *
4077  * Modifications to the worklist must be protected by sync_mtx.
4078  */
4079 static int
4080 sync_reclaim(struct vop_reclaim_args *ap)
4081 {
4082 	struct vnode *vp = ap->a_vp;
4083 	struct bufobj *bo;
4084 
4085 	bo = &vp->v_bufobj;
4086 	BO_LOCK(bo);
4087 	mtx_lock(&sync_mtx);
4088 	if (vp->v_mount->mnt_syncer == vp)
4089 		vp->v_mount->mnt_syncer = NULL;
4090 	if (bo->bo_flag & BO_ONWORKLST) {
4091 		LIST_REMOVE(bo, bo_synclist);
4092 		syncer_worklist_len--;
4093 		sync_vnode_count--;
4094 		bo->bo_flag &= ~BO_ONWORKLST;
4095 	}
4096 	mtx_unlock(&sync_mtx);
4097 	BO_UNLOCK(bo);
4098 
4099 	return (0);
4100 }
4101 
4102 /*
4103  * Check if vnode represents a disk device
4104  */
4105 int
4106 vn_isdisk(struct vnode *vp, int *errp)
4107 {
4108 	int error;
4109 
4110 	if (vp->v_type != VCHR) {
4111 		error = ENOTBLK;
4112 		goto out;
4113 	}
4114 	error = 0;
4115 	dev_lock();
4116 	if (vp->v_rdev == NULL)
4117 		error = ENXIO;
4118 	else if (vp->v_rdev->si_devsw == NULL)
4119 		error = ENXIO;
4120 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4121 		error = ENOTBLK;
4122 	dev_unlock();
4123 out:
4124 	if (errp != NULL)
4125 		*errp = error;
4126 	return (error == 0);
4127 }
4128 
4129 /*
4130  * Common filesystem object access control check routine.  Accepts a
4131  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4132  * and optional call-by-reference privused argument allowing vaccess()
4133  * to indicate to the caller whether privilege was used to satisfy the
4134  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4135  */
4136 int
4137 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4138     accmode_t accmode, struct ucred *cred, int *privused)
4139 {
4140 	accmode_t dac_granted;
4141 	accmode_t priv_granted;
4142 
4143 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4144 	    ("invalid bit in accmode"));
4145 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4146 	    ("VAPPEND without VWRITE"));
4147 
4148 	/*
4149 	 * Look for a normal, non-privileged way to access the file/directory
4150 	 * as requested.  If it exists, go with that.
4151 	 */
4152 
4153 	if (privused != NULL)
4154 		*privused = 0;
4155 
4156 	dac_granted = 0;
4157 
4158 	/* Check the owner. */
4159 	if (cred->cr_uid == file_uid) {
4160 		dac_granted |= VADMIN;
4161 		if (file_mode & S_IXUSR)
4162 			dac_granted |= VEXEC;
4163 		if (file_mode & S_IRUSR)
4164 			dac_granted |= VREAD;
4165 		if (file_mode & S_IWUSR)
4166 			dac_granted |= (VWRITE | VAPPEND);
4167 
4168 		if ((accmode & dac_granted) == accmode)
4169 			return (0);
4170 
4171 		goto privcheck;
4172 	}
4173 
4174 	/* Otherwise, check the groups (first match) */
4175 	if (groupmember(file_gid, cred)) {
4176 		if (file_mode & S_IXGRP)
4177 			dac_granted |= VEXEC;
4178 		if (file_mode & S_IRGRP)
4179 			dac_granted |= VREAD;
4180 		if (file_mode & S_IWGRP)
4181 			dac_granted |= (VWRITE | VAPPEND);
4182 
4183 		if ((accmode & dac_granted) == accmode)
4184 			return (0);
4185 
4186 		goto privcheck;
4187 	}
4188 
4189 	/* Otherwise, check everyone else. */
4190 	if (file_mode & S_IXOTH)
4191 		dac_granted |= VEXEC;
4192 	if (file_mode & S_IROTH)
4193 		dac_granted |= VREAD;
4194 	if (file_mode & S_IWOTH)
4195 		dac_granted |= (VWRITE | VAPPEND);
4196 	if ((accmode & dac_granted) == accmode)
4197 		return (0);
4198 
4199 privcheck:
4200 	/*
4201 	 * Build a privilege mask to determine if the set of privileges
4202 	 * satisfies the requirements when combined with the granted mask
4203 	 * from above.  For each privilege, if the privilege is required,
4204 	 * bitwise or the request type onto the priv_granted mask.
4205 	 */
4206 	priv_granted = 0;
4207 
4208 	if (type == VDIR) {
4209 		/*
4210 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4211 		 * requests, instead of PRIV_VFS_EXEC.
4212 		 */
4213 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4214 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4215 			priv_granted |= VEXEC;
4216 	} else {
4217 		/*
4218 		 * Ensure that at least one execute bit is on. Otherwise,
4219 		 * a privileged user will always succeed, and we don't want
4220 		 * this to happen unless the file really is executable.
4221 		 */
4222 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4223 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4224 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4225 			priv_granted |= VEXEC;
4226 	}
4227 
4228 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4229 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4230 		priv_granted |= VREAD;
4231 
4232 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4233 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4234 		priv_granted |= (VWRITE | VAPPEND);
4235 
4236 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4237 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4238 		priv_granted |= VADMIN;
4239 
4240 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4241 		/* XXX audit: privilege used */
4242 		if (privused != NULL)
4243 			*privused = 1;
4244 		return (0);
4245 	}
4246 
4247 	return ((accmode & VADMIN) ? EPERM : EACCES);
4248 }
4249 
4250 /*
4251  * Credential check based on process requesting service, and per-attribute
4252  * permissions.
4253  */
4254 int
4255 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4256     struct thread *td, accmode_t accmode)
4257 {
4258 
4259 	/*
4260 	 * Kernel-invoked always succeeds.
4261 	 */
4262 	if (cred == NOCRED)
4263 		return (0);
4264 
4265 	/*
4266 	 * Do not allow privileged processes in jail to directly manipulate
4267 	 * system attributes.
4268 	 */
4269 	switch (attrnamespace) {
4270 	case EXTATTR_NAMESPACE_SYSTEM:
4271 		/* Potentially should be: return (EPERM); */
4272 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4273 	case EXTATTR_NAMESPACE_USER:
4274 		return (VOP_ACCESS(vp, accmode, cred, td));
4275 	default:
4276 		return (EPERM);
4277 	}
4278 }
4279 
4280 #ifdef DEBUG_VFS_LOCKS
4281 /*
4282  * This only exists to supress warnings from unlocked specfs accesses.  It is
4283  * no longer ok to have an unlocked VFS.
4284  */
4285 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4286 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4287 
4288 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4289 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4290     "Drop into debugger on lock violation");
4291 
4292 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4293 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4294     0, "Check for interlock across VOPs");
4295 
4296 int vfs_badlock_print = 1;	/* Print lock violations. */
4297 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4298     0, "Print lock violations");
4299 
4300 #ifdef KDB
4301 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4302 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4303     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4304 #endif
4305 
4306 static void
4307 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4308 {
4309 
4310 #ifdef KDB
4311 	if (vfs_badlock_backtrace)
4312 		kdb_backtrace();
4313 #endif
4314 	if (vfs_badlock_print)
4315 		printf("%s: %p %s\n", str, (void *)vp, msg);
4316 	if (vfs_badlock_ddb)
4317 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4318 }
4319 
4320 void
4321 assert_vi_locked(struct vnode *vp, const char *str)
4322 {
4323 
4324 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4325 		vfs_badlock("interlock is not locked but should be", str, vp);
4326 }
4327 
4328 void
4329 assert_vi_unlocked(struct vnode *vp, const char *str)
4330 {
4331 
4332 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4333 		vfs_badlock("interlock is locked but should not be", str, vp);
4334 }
4335 
4336 void
4337 assert_vop_locked(struct vnode *vp, const char *str)
4338 {
4339 	int locked;
4340 
4341 	if (!IGNORE_LOCK(vp)) {
4342 		locked = VOP_ISLOCKED(vp);
4343 		if (locked == 0 || locked == LK_EXCLOTHER)
4344 			vfs_badlock("is not locked but should be", str, vp);
4345 	}
4346 }
4347 
4348 void
4349 assert_vop_unlocked(struct vnode *vp, const char *str)
4350 {
4351 
4352 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4353 		vfs_badlock("is locked but should not be", str, vp);
4354 }
4355 
4356 void
4357 assert_vop_elocked(struct vnode *vp, const char *str)
4358 {
4359 
4360 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4361 		vfs_badlock("is not exclusive locked but should be", str, vp);
4362 }
4363 
4364 #if 0
4365 void
4366 assert_vop_elocked_other(struct vnode *vp, const char *str)
4367 {
4368 
4369 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
4370 		vfs_badlock("is not exclusive locked by another thread",
4371 		    str, vp);
4372 }
4373 
4374 void
4375 assert_vop_slocked(struct vnode *vp, const char *str)
4376 {
4377 
4378 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
4379 		vfs_badlock("is not locked shared but should be", str, vp);
4380 }
4381 #endif /* 0 */
4382 #endif /* DEBUG_VFS_LOCKS */
4383 
4384 void
4385 vop_rename_fail(struct vop_rename_args *ap)
4386 {
4387 
4388 	if (ap->a_tvp != NULL)
4389 		vput(ap->a_tvp);
4390 	if (ap->a_tdvp == ap->a_tvp)
4391 		vrele(ap->a_tdvp);
4392 	else
4393 		vput(ap->a_tdvp);
4394 	vrele(ap->a_fdvp);
4395 	vrele(ap->a_fvp);
4396 }
4397 
4398 void
4399 vop_rename_pre(void *ap)
4400 {
4401 	struct vop_rename_args *a = ap;
4402 
4403 #ifdef DEBUG_VFS_LOCKS
4404 	if (a->a_tvp)
4405 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4406 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4407 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4408 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4409 
4410 	/* Check the source (from). */
4411 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4412 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4413 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4414 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4415 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4416 
4417 	/* Check the target. */
4418 	if (a->a_tvp)
4419 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4420 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4421 #endif
4422 	if (a->a_tdvp != a->a_fdvp)
4423 		vhold(a->a_fdvp);
4424 	if (a->a_tvp != a->a_fvp)
4425 		vhold(a->a_fvp);
4426 	vhold(a->a_tdvp);
4427 	if (a->a_tvp)
4428 		vhold(a->a_tvp);
4429 }
4430 
4431 void
4432 vop_strategy_pre(void *ap)
4433 {
4434 #ifdef DEBUG_VFS_LOCKS
4435 	struct vop_strategy_args *a;
4436 	struct buf *bp;
4437 
4438 	a = ap;
4439 	bp = a->a_bp;
4440 
4441 	/*
4442 	 * Cluster ops lock their component buffers but not the IO container.
4443 	 */
4444 	if ((bp->b_flags & B_CLUSTER) != 0)
4445 		return;
4446 
4447 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4448 		if (vfs_badlock_print)
4449 			printf(
4450 			    "VOP_STRATEGY: bp is not locked but should be\n");
4451 		if (vfs_badlock_ddb)
4452 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4453 	}
4454 #endif
4455 }
4456 
4457 void
4458 vop_lock_pre(void *ap)
4459 {
4460 #ifdef DEBUG_VFS_LOCKS
4461 	struct vop_lock1_args *a = ap;
4462 
4463 	if ((a->a_flags & LK_INTERLOCK) == 0)
4464 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4465 	else
4466 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4467 #endif
4468 }
4469 
4470 void
4471 vop_lock_post(void *ap, int rc)
4472 {
4473 #ifdef DEBUG_VFS_LOCKS
4474 	struct vop_lock1_args *a = ap;
4475 
4476 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4477 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4478 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4479 #endif
4480 }
4481 
4482 void
4483 vop_unlock_pre(void *ap)
4484 {
4485 #ifdef DEBUG_VFS_LOCKS
4486 	struct vop_unlock_args *a = ap;
4487 
4488 	if (a->a_flags & LK_INTERLOCK)
4489 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4490 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4491 #endif
4492 }
4493 
4494 void
4495 vop_unlock_post(void *ap, int rc)
4496 {
4497 #ifdef DEBUG_VFS_LOCKS
4498 	struct vop_unlock_args *a = ap;
4499 
4500 	if (a->a_flags & LK_INTERLOCK)
4501 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4502 #endif
4503 }
4504 
4505 void
4506 vop_create_post(void *ap, int rc)
4507 {
4508 	struct vop_create_args *a = ap;
4509 
4510 	if (!rc)
4511 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4512 }
4513 
4514 void
4515 vop_deleteextattr_post(void *ap, int rc)
4516 {
4517 	struct vop_deleteextattr_args *a = ap;
4518 
4519 	if (!rc)
4520 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4521 }
4522 
4523 void
4524 vop_link_post(void *ap, int rc)
4525 {
4526 	struct vop_link_args *a = ap;
4527 
4528 	if (!rc) {
4529 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4530 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4531 	}
4532 }
4533 
4534 void
4535 vop_mkdir_post(void *ap, int rc)
4536 {
4537 	struct vop_mkdir_args *a = ap;
4538 
4539 	if (!rc)
4540 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4541 }
4542 
4543 void
4544 vop_mknod_post(void *ap, int rc)
4545 {
4546 	struct vop_mknod_args *a = ap;
4547 
4548 	if (!rc)
4549 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4550 }
4551 
4552 void
4553 vop_reclaim_post(void *ap, int rc)
4554 {
4555 	struct vop_reclaim_args *a = ap;
4556 
4557 	if (!rc)
4558 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4559 }
4560 
4561 void
4562 vop_remove_post(void *ap, int rc)
4563 {
4564 	struct vop_remove_args *a = ap;
4565 
4566 	if (!rc) {
4567 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4568 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4569 	}
4570 }
4571 
4572 void
4573 vop_rename_post(void *ap, int rc)
4574 {
4575 	struct vop_rename_args *a = ap;
4576 
4577 	if (!rc) {
4578 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
4579 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
4580 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4581 		if (a->a_tvp)
4582 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4583 	}
4584 	if (a->a_tdvp != a->a_fdvp)
4585 		vdrop(a->a_fdvp);
4586 	if (a->a_tvp != a->a_fvp)
4587 		vdrop(a->a_fvp);
4588 	vdrop(a->a_tdvp);
4589 	if (a->a_tvp)
4590 		vdrop(a->a_tvp);
4591 }
4592 
4593 void
4594 vop_rmdir_post(void *ap, int rc)
4595 {
4596 	struct vop_rmdir_args *a = ap;
4597 
4598 	if (!rc) {
4599 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4600 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4601 	}
4602 }
4603 
4604 void
4605 vop_setattr_post(void *ap, int rc)
4606 {
4607 	struct vop_setattr_args *a = ap;
4608 
4609 	if (!rc)
4610 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4611 }
4612 
4613 void
4614 vop_setextattr_post(void *ap, int rc)
4615 {
4616 	struct vop_setextattr_args *a = ap;
4617 
4618 	if (!rc)
4619 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4620 }
4621 
4622 void
4623 vop_symlink_post(void *ap, int rc)
4624 {
4625 	struct vop_symlink_args *a = ap;
4626 
4627 	if (!rc)
4628 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4629 }
4630 
4631 static struct knlist fs_knlist;
4632 
4633 static void
4634 vfs_event_init(void *arg)
4635 {
4636 	knlist_init_mtx(&fs_knlist, NULL);
4637 }
4638 /* XXX - correct order? */
4639 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4640 
4641 void
4642 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4643 {
4644 
4645 	KNOTE_UNLOCKED(&fs_knlist, event);
4646 }
4647 
4648 static int	filt_fsattach(struct knote *kn);
4649 static void	filt_fsdetach(struct knote *kn);
4650 static int	filt_fsevent(struct knote *kn, long hint);
4651 
4652 struct filterops fs_filtops = {
4653 	.f_isfd = 0,
4654 	.f_attach = filt_fsattach,
4655 	.f_detach = filt_fsdetach,
4656 	.f_event = filt_fsevent
4657 };
4658 
4659 static int
4660 filt_fsattach(struct knote *kn)
4661 {
4662 
4663 	kn->kn_flags |= EV_CLEAR;
4664 	knlist_add(&fs_knlist, kn, 0);
4665 	return (0);
4666 }
4667 
4668 static void
4669 filt_fsdetach(struct knote *kn)
4670 {
4671 
4672 	knlist_remove(&fs_knlist, kn, 0);
4673 }
4674 
4675 static int
4676 filt_fsevent(struct knote *kn, long hint)
4677 {
4678 
4679 	kn->kn_fflags |= hint;
4680 	return (kn->kn_fflags != 0);
4681 }
4682 
4683 static int
4684 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4685 {
4686 	struct vfsidctl vc;
4687 	int error;
4688 	struct mount *mp;
4689 
4690 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4691 	if (error)
4692 		return (error);
4693 	if (vc.vc_vers != VFS_CTL_VERS1)
4694 		return (EINVAL);
4695 	mp = vfs_getvfs(&vc.vc_fsid);
4696 	if (mp == NULL)
4697 		return (ENOENT);
4698 	/* ensure that a specific sysctl goes to the right filesystem. */
4699 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4700 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4701 		vfs_rel(mp);
4702 		return (EINVAL);
4703 	}
4704 	VCTLTOREQ(&vc, req);
4705 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4706 	vfs_rel(mp);
4707 	return (error);
4708 }
4709 
4710 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4711     NULL, 0, sysctl_vfs_ctl, "",
4712     "Sysctl by fsid");
4713 
4714 /*
4715  * Function to initialize a va_filerev field sensibly.
4716  * XXX: Wouldn't a random number make a lot more sense ??
4717  */
4718 u_quad_t
4719 init_va_filerev(void)
4720 {
4721 	struct bintime bt;
4722 
4723 	getbinuptime(&bt);
4724 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4725 }
4726 
4727 static int	filt_vfsread(struct knote *kn, long hint);
4728 static int	filt_vfswrite(struct knote *kn, long hint);
4729 static int	filt_vfsvnode(struct knote *kn, long hint);
4730 static void	filt_vfsdetach(struct knote *kn);
4731 static struct filterops vfsread_filtops = {
4732 	.f_isfd = 1,
4733 	.f_detach = filt_vfsdetach,
4734 	.f_event = filt_vfsread
4735 };
4736 static struct filterops vfswrite_filtops = {
4737 	.f_isfd = 1,
4738 	.f_detach = filt_vfsdetach,
4739 	.f_event = filt_vfswrite
4740 };
4741 static struct filterops vfsvnode_filtops = {
4742 	.f_isfd = 1,
4743 	.f_detach = filt_vfsdetach,
4744 	.f_event = filt_vfsvnode
4745 };
4746 
4747 static void
4748 vfs_knllock(void *arg)
4749 {
4750 	struct vnode *vp = arg;
4751 
4752 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4753 }
4754 
4755 static void
4756 vfs_knlunlock(void *arg)
4757 {
4758 	struct vnode *vp = arg;
4759 
4760 	VOP_UNLOCK(vp, 0);
4761 }
4762 
4763 static void
4764 vfs_knl_assert_locked(void *arg)
4765 {
4766 #ifdef DEBUG_VFS_LOCKS
4767 	struct vnode *vp = arg;
4768 
4769 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4770 #endif
4771 }
4772 
4773 static void
4774 vfs_knl_assert_unlocked(void *arg)
4775 {
4776 #ifdef DEBUG_VFS_LOCKS
4777 	struct vnode *vp = arg;
4778 
4779 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4780 #endif
4781 }
4782 
4783 int
4784 vfs_kqfilter(struct vop_kqfilter_args *ap)
4785 {
4786 	struct vnode *vp = ap->a_vp;
4787 	struct knote *kn = ap->a_kn;
4788 	struct knlist *knl;
4789 
4790 	switch (kn->kn_filter) {
4791 	case EVFILT_READ:
4792 		kn->kn_fop = &vfsread_filtops;
4793 		break;
4794 	case EVFILT_WRITE:
4795 		kn->kn_fop = &vfswrite_filtops;
4796 		break;
4797 	case EVFILT_VNODE:
4798 		kn->kn_fop = &vfsvnode_filtops;
4799 		break;
4800 	default:
4801 		return (EINVAL);
4802 	}
4803 
4804 	kn->kn_hook = (caddr_t)vp;
4805 
4806 	v_addpollinfo(vp);
4807 	if (vp->v_pollinfo == NULL)
4808 		return (ENOMEM);
4809 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4810 	vhold(vp);
4811 	knlist_add(knl, kn, 0);
4812 
4813 	return (0);
4814 }
4815 
4816 /*
4817  * Detach knote from vnode
4818  */
4819 static void
4820 filt_vfsdetach(struct knote *kn)
4821 {
4822 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4823 
4824 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4825 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4826 	vdrop(vp);
4827 }
4828 
4829 /*ARGSUSED*/
4830 static int
4831 filt_vfsread(struct knote *kn, long hint)
4832 {
4833 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4834 	struct vattr va;
4835 	int res;
4836 
4837 	/*
4838 	 * filesystem is gone, so set the EOF flag and schedule
4839 	 * the knote for deletion.
4840 	 */
4841 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4842 		VI_LOCK(vp);
4843 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4844 		VI_UNLOCK(vp);
4845 		return (1);
4846 	}
4847 
4848 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4849 		return (0);
4850 
4851 	VI_LOCK(vp);
4852 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4853 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
4854 	VI_UNLOCK(vp);
4855 	return (res);
4856 }
4857 
4858 /*ARGSUSED*/
4859 static int
4860 filt_vfswrite(struct knote *kn, long hint)
4861 {
4862 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4863 
4864 	VI_LOCK(vp);
4865 
4866 	/*
4867 	 * filesystem is gone, so set the EOF flag and schedule
4868 	 * the knote for deletion.
4869 	 */
4870 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
4871 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4872 
4873 	kn->kn_data = 0;
4874 	VI_UNLOCK(vp);
4875 	return (1);
4876 }
4877 
4878 static int
4879 filt_vfsvnode(struct knote *kn, long hint)
4880 {
4881 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4882 	int res;
4883 
4884 	VI_LOCK(vp);
4885 	if (kn->kn_sfflags & hint)
4886 		kn->kn_fflags |= hint;
4887 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4888 		kn->kn_flags |= EV_EOF;
4889 		VI_UNLOCK(vp);
4890 		return (1);
4891 	}
4892 	res = (kn->kn_fflags != 0);
4893 	VI_UNLOCK(vp);
4894 	return (res);
4895 }
4896 
4897 int
4898 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
4899 {
4900 	int error;
4901 
4902 	if (dp->d_reclen > ap->a_uio->uio_resid)
4903 		return (ENAMETOOLONG);
4904 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
4905 	if (error) {
4906 		if (ap->a_ncookies != NULL) {
4907 			if (ap->a_cookies != NULL)
4908 				free(ap->a_cookies, M_TEMP);
4909 			ap->a_cookies = NULL;
4910 			*ap->a_ncookies = 0;
4911 		}
4912 		return (error);
4913 	}
4914 	if (ap->a_ncookies == NULL)
4915 		return (0);
4916 
4917 	KASSERT(ap->a_cookies,
4918 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
4919 
4920 	*ap->a_cookies = realloc(*ap->a_cookies,
4921 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
4922 	(*ap->a_cookies)[*ap->a_ncookies] = off;
4923 	return (0);
4924 }
4925 
4926 /*
4927  * Mark for update the access time of the file if the filesystem
4928  * supports VOP_MARKATIME.  This functionality is used by execve and
4929  * mmap, so we want to avoid the I/O implied by directly setting
4930  * va_atime for the sake of efficiency.
4931  */
4932 void
4933 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
4934 {
4935 	struct mount *mp;
4936 
4937 	mp = vp->v_mount;
4938 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
4939 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
4940 		(void)VOP_MARKATIME(vp);
4941 }
4942 
4943 /*
4944  * The purpose of this routine is to remove granularity from accmode_t,
4945  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
4946  * VADMIN and VAPPEND.
4947  *
4948  * If it returns 0, the caller is supposed to continue with the usual
4949  * access checks using 'accmode' as modified by this routine.  If it
4950  * returns nonzero value, the caller is supposed to return that value
4951  * as errno.
4952  *
4953  * Note that after this routine runs, accmode may be zero.
4954  */
4955 int
4956 vfs_unixify_accmode(accmode_t *accmode)
4957 {
4958 	/*
4959 	 * There is no way to specify explicit "deny" rule using
4960 	 * file mode or POSIX.1e ACLs.
4961 	 */
4962 	if (*accmode & VEXPLICIT_DENY) {
4963 		*accmode = 0;
4964 		return (0);
4965 	}
4966 
4967 	/*
4968 	 * None of these can be translated into usual access bits.
4969 	 * Also, the common case for NFSv4 ACLs is to not contain
4970 	 * either of these bits. Caller should check for VWRITE
4971 	 * on the containing directory instead.
4972 	 */
4973 	if (*accmode & (VDELETE_CHILD | VDELETE))
4974 		return (EPERM);
4975 
4976 	if (*accmode & VADMIN_PERMS) {
4977 		*accmode &= ~VADMIN_PERMS;
4978 		*accmode |= VADMIN;
4979 	}
4980 
4981 	/*
4982 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
4983 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
4984 	 */
4985 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
4986 
4987 	return (0);
4988 }
4989 
4990 /*
4991  * These are helper functions for filesystems to traverse all
4992  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
4993  *
4994  * This interface replaces MNT_VNODE_FOREACH.
4995  */
4996 
4997 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
4998 
4999 struct vnode *
5000 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5001 {
5002 	struct vnode *vp;
5003 
5004 	if (should_yield())
5005 		kern_yield(PRI_USER);
5006 	MNT_ILOCK(mp);
5007 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5008 	vp = TAILQ_NEXT(*mvp, v_nmntvnodes);
5009 	while (vp != NULL && (vp->v_type == VMARKER ||
5010 	    (vp->v_iflag & VI_DOOMED) != 0))
5011 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5012 
5013 	/* Check if we are done */
5014 	if (vp == NULL) {
5015 		__mnt_vnode_markerfree_all(mvp, mp);
5016 		/* MNT_IUNLOCK(mp); -- done in above function */
5017 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5018 		return (NULL);
5019 	}
5020 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5021 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5022 	VI_LOCK(vp);
5023 	MNT_IUNLOCK(mp);
5024 	return (vp);
5025 }
5026 
5027 struct vnode *
5028 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5029 {
5030 	struct vnode *vp;
5031 
5032 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5033 	MNT_ILOCK(mp);
5034 	MNT_REF(mp);
5035 	(*mvp)->v_type = VMARKER;
5036 
5037 	vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
5038 	while (vp != NULL && (vp->v_type == VMARKER ||
5039 	    (vp->v_iflag & VI_DOOMED) != 0))
5040 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5041 
5042 	/* Check if we are done */
5043 	if (vp == NULL) {
5044 		MNT_REL(mp);
5045 		MNT_IUNLOCK(mp);
5046 		free(*mvp, M_VNODE_MARKER);
5047 		*mvp = NULL;
5048 		return (NULL);
5049 	}
5050 	(*mvp)->v_mount = mp;
5051 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5052 	VI_LOCK(vp);
5053 	MNT_IUNLOCK(mp);
5054 	return (vp);
5055 }
5056 
5057 
5058 void
5059 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5060 {
5061 
5062 	if (*mvp == NULL) {
5063 		MNT_IUNLOCK(mp);
5064 		return;
5065 	}
5066 
5067 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5068 
5069 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5070 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5071 	MNT_REL(mp);
5072 	MNT_IUNLOCK(mp);
5073 	free(*mvp, M_VNODE_MARKER);
5074 	*mvp = NULL;
5075 }
5076 
5077 /*
5078  * These are helper functions for filesystems to traverse their
5079  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5080  */
5081 static void
5082 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5083 {
5084 
5085 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5086 
5087 	MNT_ILOCK(mp);
5088 	MNT_REL(mp);
5089 	MNT_IUNLOCK(mp);
5090 	free(*mvp, M_VNODE_MARKER);
5091 	*mvp = NULL;
5092 }
5093 
5094 static struct vnode *
5095 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5096 {
5097 	struct vnode *vp, *nvp;
5098 
5099 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5100 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5101 restart:
5102 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5103 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5104 	while (vp != NULL) {
5105 		if (vp->v_type == VMARKER) {
5106 			vp = TAILQ_NEXT(vp, v_actfreelist);
5107 			continue;
5108 		}
5109 		if (!VI_TRYLOCK(vp)) {
5110 			if (mp_ncpus == 1 || should_yield()) {
5111 				TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5112 				mtx_unlock(&vnode_free_list_mtx);
5113 				pause("vnacti", 1);
5114 				mtx_lock(&vnode_free_list_mtx);
5115 				goto restart;
5116 			}
5117 			continue;
5118 		}
5119 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5120 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5121 		    ("alien vnode on the active list %p %p", vp, mp));
5122 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5123 			break;
5124 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5125 		VI_UNLOCK(vp);
5126 		vp = nvp;
5127 	}
5128 
5129 	/* Check if we are done */
5130 	if (vp == NULL) {
5131 		mtx_unlock(&vnode_free_list_mtx);
5132 		mnt_vnode_markerfree_active(mvp, mp);
5133 		return (NULL);
5134 	}
5135 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5136 	mtx_unlock(&vnode_free_list_mtx);
5137 	ASSERT_VI_LOCKED(vp, "active iter");
5138 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5139 	return (vp);
5140 }
5141 
5142 struct vnode *
5143 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5144 {
5145 
5146 	if (should_yield())
5147 		kern_yield(PRI_USER);
5148 	mtx_lock(&vnode_free_list_mtx);
5149 	return (mnt_vnode_next_active(mvp, mp));
5150 }
5151 
5152 struct vnode *
5153 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5154 {
5155 	struct vnode *vp;
5156 
5157 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5158 	MNT_ILOCK(mp);
5159 	MNT_REF(mp);
5160 	MNT_IUNLOCK(mp);
5161 	(*mvp)->v_type = VMARKER;
5162 	(*mvp)->v_mount = mp;
5163 
5164 	mtx_lock(&vnode_free_list_mtx);
5165 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5166 	if (vp == NULL) {
5167 		mtx_unlock(&vnode_free_list_mtx);
5168 		mnt_vnode_markerfree_active(mvp, mp);
5169 		return (NULL);
5170 	}
5171 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5172 	return (mnt_vnode_next_active(mvp, mp));
5173 }
5174 
5175 void
5176 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5177 {
5178 
5179 	if (*mvp == NULL)
5180 		return;
5181 
5182 	mtx_lock(&vnode_free_list_mtx);
5183 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5184 	mtx_unlock(&vnode_free_list_mtx);
5185 	mnt_vnode_markerfree_active(mvp, mp);
5186 }
5187