xref: /freebsd/sys/kern/vfs_subr.c (revision b6a05070fa77edc7ce6e60b61623fd806e807be6)
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 		nbp = gbincore(bo, lblkno);
1656 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1657 		    != xflags)
1658 			break;			/* nbp invalid */
1659 	}
1660 	return (retval);
1661 }
1662 
1663 int
1664 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1665 {
1666 	struct buf *bp;
1667 	int error;
1668 	daddr_t lblkno;
1669 
1670 	ASSERT_BO_LOCKED(bo);
1671 
1672 	for (lblkno = startn;; lblkno++) {
1673 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1674 		if (bp == NULL || bp->b_lblkno >= endn)
1675 			break;
1676 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1677 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1678 		if (error != 0) {
1679 			BO_RLOCK(bo);
1680 			return (error != ENOLCK ? error : EAGAIN);
1681 		}
1682 		KASSERT(bp->b_bufobj == bo,
1683 		    ("bp %p wrong b_bufobj %p should be %p",
1684 		    bp, bp->b_bufobj, bo));
1685 		if ((bp->b_flags & B_MANAGED) == 0)
1686 			bremfree(bp);
1687 		bp->b_flags |= B_RELBUF;
1688 		/*
1689 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1690 		 * pages backing each buffer in the range are unlikely to be
1691 		 * reused.  Dirty buffers will have the hint applied once
1692 		 * they've been written.
1693 		 */
1694 		if (bp->b_vp->v_object != NULL)
1695 			bp->b_flags |= B_NOREUSE;
1696 		brelse(bp);
1697 		BO_RLOCK(bo);
1698 	}
1699 	return (0);
1700 }
1701 
1702 /*
1703  * Truncate a file's buffer and pages to a specified length.  This
1704  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1705  * sync activity.
1706  */
1707 int
1708 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1709 {
1710 	struct buf *bp, *nbp;
1711 	int anyfreed;
1712 	int trunclbn;
1713 	struct bufobj *bo;
1714 
1715 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1716 	    vp, cred, blksize, (uintmax_t)length);
1717 
1718 	/*
1719 	 * Round up to the *next* lbn.
1720 	 */
1721 	trunclbn = (length + blksize - 1) / blksize;
1722 
1723 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1724 restart:
1725 	bo = &vp->v_bufobj;
1726 	BO_LOCK(bo);
1727 	anyfreed = 1;
1728 	for (;anyfreed;) {
1729 		anyfreed = 0;
1730 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1731 			if (bp->b_lblkno < trunclbn)
1732 				continue;
1733 			if (BUF_LOCK(bp,
1734 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1735 			    BO_LOCKPTR(bo)) == ENOLCK)
1736 				goto restart;
1737 
1738 			bremfree(bp);
1739 			bp->b_flags |= (B_INVAL | B_RELBUF);
1740 			bp->b_flags &= ~B_ASYNC;
1741 			brelse(bp);
1742 			anyfreed = 1;
1743 
1744 			BO_LOCK(bo);
1745 			if (nbp != NULL &&
1746 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1747 			    (nbp->b_vp != vp) ||
1748 			    (nbp->b_flags & B_DELWRI))) {
1749 				BO_UNLOCK(bo);
1750 				goto restart;
1751 			}
1752 		}
1753 
1754 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1755 			if (bp->b_lblkno < trunclbn)
1756 				continue;
1757 			if (BUF_LOCK(bp,
1758 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1759 			    BO_LOCKPTR(bo)) == ENOLCK)
1760 				goto restart;
1761 			bremfree(bp);
1762 			bp->b_flags |= (B_INVAL | B_RELBUF);
1763 			bp->b_flags &= ~B_ASYNC;
1764 			brelse(bp);
1765 			anyfreed = 1;
1766 
1767 			BO_LOCK(bo);
1768 			if (nbp != NULL &&
1769 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1770 			    (nbp->b_vp != vp) ||
1771 			    (nbp->b_flags & B_DELWRI) == 0)) {
1772 				BO_UNLOCK(bo);
1773 				goto restart;
1774 			}
1775 		}
1776 	}
1777 
1778 	if (length > 0) {
1779 restartsync:
1780 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1781 			if (bp->b_lblkno > 0)
1782 				continue;
1783 			/*
1784 			 * Since we hold the vnode lock this should only
1785 			 * fail if we're racing with the buf daemon.
1786 			 */
1787 			if (BUF_LOCK(bp,
1788 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1789 			    BO_LOCKPTR(bo)) == ENOLCK) {
1790 				goto restart;
1791 			}
1792 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1793 			    ("buf(%p) on dirty queue without DELWRI", bp));
1794 
1795 			bremfree(bp);
1796 			bawrite(bp);
1797 			BO_LOCK(bo);
1798 			goto restartsync;
1799 		}
1800 	}
1801 
1802 	bufobj_wwait(bo, 0, 0);
1803 	BO_UNLOCK(bo);
1804 	vnode_pager_setsize(vp, length);
1805 
1806 	return (0);
1807 }
1808 
1809 static void
1810 buf_vlist_remove(struct buf *bp)
1811 {
1812 	struct bufv *bv;
1813 
1814 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1815 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1816 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1817 	    (BX_VNDIRTY|BX_VNCLEAN),
1818 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1819 	if (bp->b_xflags & BX_VNDIRTY)
1820 		bv = &bp->b_bufobj->bo_dirty;
1821 	else
1822 		bv = &bp->b_bufobj->bo_clean;
1823 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1824 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1825 	bv->bv_cnt--;
1826 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1827 }
1828 
1829 /*
1830  * Add the buffer to the sorted clean or dirty block list.
1831  *
1832  * NOTE: xflags is passed as a constant, optimizing this inline function!
1833  */
1834 static void
1835 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1836 {
1837 	struct bufv *bv;
1838 	struct buf *n;
1839 	int error;
1840 
1841 	ASSERT_BO_WLOCKED(bo);
1842 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1843 	    ("dead bo %p", bo));
1844 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1845 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1846 	bp->b_xflags |= xflags;
1847 	if (xflags & BX_VNDIRTY)
1848 		bv = &bo->bo_dirty;
1849 	else
1850 		bv = &bo->bo_clean;
1851 
1852 	/*
1853 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1854 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1855 	 * than _ge.
1856 	 */
1857 	if (bv->bv_cnt == 0 ||
1858 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1859 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1860 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1861 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1862 	else
1863 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1864 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1865 	if (error)
1866 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1867 	bv->bv_cnt++;
1868 }
1869 
1870 /*
1871  * Look up a buffer using the buffer tries.
1872  */
1873 struct buf *
1874 gbincore(struct bufobj *bo, daddr_t lblkno)
1875 {
1876 	struct buf *bp;
1877 
1878 	ASSERT_BO_LOCKED(bo);
1879 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1880 	if (bp != NULL)
1881 		return (bp);
1882 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1883 }
1884 
1885 /*
1886  * Associate a buffer with a vnode.
1887  */
1888 void
1889 bgetvp(struct vnode *vp, struct buf *bp)
1890 {
1891 	struct bufobj *bo;
1892 
1893 	bo = &vp->v_bufobj;
1894 	ASSERT_BO_WLOCKED(bo);
1895 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1896 
1897 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1898 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1899 	    ("bgetvp: bp already attached! %p", bp));
1900 
1901 	vhold(vp);
1902 	bp->b_vp = vp;
1903 	bp->b_bufobj = bo;
1904 	/*
1905 	 * Insert onto list for new vnode.
1906 	 */
1907 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1908 }
1909 
1910 /*
1911  * Disassociate a buffer from a vnode.
1912  */
1913 void
1914 brelvp(struct buf *bp)
1915 {
1916 	struct bufobj *bo;
1917 	struct vnode *vp;
1918 
1919 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1920 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1921 
1922 	/*
1923 	 * Delete from old vnode list, if on one.
1924 	 */
1925 	vp = bp->b_vp;		/* XXX */
1926 	bo = bp->b_bufobj;
1927 	BO_LOCK(bo);
1928 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1929 		buf_vlist_remove(bp);
1930 	else
1931 		panic("brelvp: Buffer %p not on queue.", bp);
1932 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1933 		bo->bo_flag &= ~BO_ONWORKLST;
1934 		mtx_lock(&sync_mtx);
1935 		LIST_REMOVE(bo, bo_synclist);
1936 		syncer_worklist_len--;
1937 		mtx_unlock(&sync_mtx);
1938 	}
1939 	bp->b_vp = NULL;
1940 	bp->b_bufobj = NULL;
1941 	BO_UNLOCK(bo);
1942 	vdrop(vp);
1943 }
1944 
1945 /*
1946  * Add an item to the syncer work queue.
1947  */
1948 static void
1949 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1950 {
1951 	int slot;
1952 
1953 	ASSERT_BO_WLOCKED(bo);
1954 
1955 	mtx_lock(&sync_mtx);
1956 	if (bo->bo_flag & BO_ONWORKLST)
1957 		LIST_REMOVE(bo, bo_synclist);
1958 	else {
1959 		bo->bo_flag |= BO_ONWORKLST;
1960 		syncer_worklist_len++;
1961 	}
1962 
1963 	if (delay > syncer_maxdelay - 2)
1964 		delay = syncer_maxdelay - 2;
1965 	slot = (syncer_delayno + delay) & syncer_mask;
1966 
1967 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1968 	mtx_unlock(&sync_mtx);
1969 }
1970 
1971 static int
1972 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1973 {
1974 	int error, len;
1975 
1976 	mtx_lock(&sync_mtx);
1977 	len = syncer_worklist_len - sync_vnode_count;
1978 	mtx_unlock(&sync_mtx);
1979 	error = SYSCTL_OUT(req, &len, sizeof(len));
1980 	return (error);
1981 }
1982 
1983 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1984     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1985 
1986 static struct proc *updateproc;
1987 static void sched_sync(void);
1988 static struct kproc_desc up_kp = {
1989 	"syncer",
1990 	sched_sync,
1991 	&updateproc
1992 };
1993 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
1994 
1995 static int
1996 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
1997 {
1998 	struct vnode *vp;
1999 	struct mount *mp;
2000 
2001 	*bo = LIST_FIRST(slp);
2002 	if (*bo == NULL)
2003 		return (0);
2004 	vp = (*bo)->__bo_vnode;	/* XXX */
2005 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2006 		return (1);
2007 	/*
2008 	 * We use vhold in case the vnode does not
2009 	 * successfully sync.  vhold prevents the vnode from
2010 	 * going away when we unlock the sync_mtx so that
2011 	 * we can acquire the vnode interlock.
2012 	 */
2013 	vholdl(vp);
2014 	mtx_unlock(&sync_mtx);
2015 	VI_UNLOCK(vp);
2016 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2017 		vdrop(vp);
2018 		mtx_lock(&sync_mtx);
2019 		return (*bo == LIST_FIRST(slp));
2020 	}
2021 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2022 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2023 	VOP_UNLOCK(vp, 0);
2024 	vn_finished_write(mp);
2025 	BO_LOCK(*bo);
2026 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2027 		/*
2028 		 * Put us back on the worklist.  The worklist
2029 		 * routine will remove us from our current
2030 		 * position and then add us back in at a later
2031 		 * position.
2032 		 */
2033 		vn_syncer_add_to_worklist(*bo, syncdelay);
2034 	}
2035 	BO_UNLOCK(*bo);
2036 	vdrop(vp);
2037 	mtx_lock(&sync_mtx);
2038 	return (0);
2039 }
2040 
2041 static int first_printf = 1;
2042 
2043 /*
2044  * System filesystem synchronizer daemon.
2045  */
2046 static void
2047 sched_sync(void)
2048 {
2049 	struct synclist *next, *slp;
2050 	struct bufobj *bo;
2051 	long starttime;
2052 	struct thread *td = curthread;
2053 	int last_work_seen;
2054 	int net_worklist_len;
2055 	int syncer_final_iter;
2056 	int error;
2057 
2058 	last_work_seen = 0;
2059 	syncer_final_iter = 0;
2060 	syncer_state = SYNCER_RUNNING;
2061 	starttime = time_uptime;
2062 	td->td_pflags |= TDP_NORUNNINGBUF;
2063 
2064 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2065 	    SHUTDOWN_PRI_LAST);
2066 
2067 	mtx_lock(&sync_mtx);
2068 	for (;;) {
2069 		if (syncer_state == SYNCER_FINAL_DELAY &&
2070 		    syncer_final_iter == 0) {
2071 			mtx_unlock(&sync_mtx);
2072 			kproc_suspend_check(td->td_proc);
2073 			mtx_lock(&sync_mtx);
2074 		}
2075 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2076 		if (syncer_state != SYNCER_RUNNING &&
2077 		    starttime != time_uptime) {
2078 			if (first_printf) {
2079 				printf("\nSyncing disks, vnodes remaining...");
2080 				first_printf = 0;
2081 			}
2082 			printf("%d ", net_worklist_len);
2083 		}
2084 		starttime = time_uptime;
2085 
2086 		/*
2087 		 * Push files whose dirty time has expired.  Be careful
2088 		 * of interrupt race on slp queue.
2089 		 *
2090 		 * Skip over empty worklist slots when shutting down.
2091 		 */
2092 		do {
2093 			slp = &syncer_workitem_pending[syncer_delayno];
2094 			syncer_delayno += 1;
2095 			if (syncer_delayno == syncer_maxdelay)
2096 				syncer_delayno = 0;
2097 			next = &syncer_workitem_pending[syncer_delayno];
2098 			/*
2099 			 * If the worklist has wrapped since the
2100 			 * it was emptied of all but syncer vnodes,
2101 			 * switch to the FINAL_DELAY state and run
2102 			 * for one more second.
2103 			 */
2104 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2105 			    net_worklist_len == 0 &&
2106 			    last_work_seen == syncer_delayno) {
2107 				syncer_state = SYNCER_FINAL_DELAY;
2108 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2109 			}
2110 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2111 		    syncer_worklist_len > 0);
2112 
2113 		/*
2114 		 * Keep track of the last time there was anything
2115 		 * on the worklist other than syncer vnodes.
2116 		 * Return to the SHUTTING_DOWN state if any
2117 		 * new work appears.
2118 		 */
2119 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2120 			last_work_seen = syncer_delayno;
2121 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2122 			syncer_state = SYNCER_SHUTTING_DOWN;
2123 		while (!LIST_EMPTY(slp)) {
2124 			error = sync_vnode(slp, &bo, td);
2125 			if (error == 1) {
2126 				LIST_REMOVE(bo, bo_synclist);
2127 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2128 				continue;
2129 			}
2130 
2131 			if (first_printf == 0) {
2132 				/*
2133 				 * Drop the sync mutex, because some watchdog
2134 				 * drivers need to sleep while patting
2135 				 */
2136 				mtx_unlock(&sync_mtx);
2137 				wdog_kern_pat(WD_LASTVAL);
2138 				mtx_lock(&sync_mtx);
2139 			}
2140 
2141 		}
2142 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2143 			syncer_final_iter--;
2144 		/*
2145 		 * The variable rushjob allows the kernel to speed up the
2146 		 * processing of the filesystem syncer process. A rushjob
2147 		 * value of N tells the filesystem syncer to process the next
2148 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2149 		 * is used by the soft update code to speed up the filesystem
2150 		 * syncer process when the incore state is getting so far
2151 		 * ahead of the disk that the kernel memory pool is being
2152 		 * threatened with exhaustion.
2153 		 */
2154 		if (rushjob > 0) {
2155 			rushjob -= 1;
2156 			continue;
2157 		}
2158 		/*
2159 		 * Just sleep for a short period of time between
2160 		 * iterations when shutting down to allow some I/O
2161 		 * to happen.
2162 		 *
2163 		 * If it has taken us less than a second to process the
2164 		 * current work, then wait. Otherwise start right over
2165 		 * again. We can still lose time if any single round
2166 		 * takes more than two seconds, but it does not really
2167 		 * matter as we are just trying to generally pace the
2168 		 * filesystem activity.
2169 		 */
2170 		if (syncer_state != SYNCER_RUNNING ||
2171 		    time_uptime == starttime) {
2172 			thread_lock(td);
2173 			sched_prio(td, PPAUSE);
2174 			thread_unlock(td);
2175 		}
2176 		if (syncer_state != SYNCER_RUNNING)
2177 			cv_timedwait(&sync_wakeup, &sync_mtx,
2178 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2179 		else if (time_uptime == starttime)
2180 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2181 	}
2182 }
2183 
2184 /*
2185  * Request the syncer daemon to speed up its work.
2186  * We never push it to speed up more than half of its
2187  * normal turn time, otherwise it could take over the cpu.
2188  */
2189 int
2190 speedup_syncer(void)
2191 {
2192 	int ret = 0;
2193 
2194 	mtx_lock(&sync_mtx);
2195 	if (rushjob < syncdelay / 2) {
2196 		rushjob += 1;
2197 		stat_rush_requests += 1;
2198 		ret = 1;
2199 	}
2200 	mtx_unlock(&sync_mtx);
2201 	cv_broadcast(&sync_wakeup);
2202 	return (ret);
2203 }
2204 
2205 /*
2206  * Tell the syncer to speed up its work and run though its work
2207  * list several times, then tell it to shut down.
2208  */
2209 static void
2210 syncer_shutdown(void *arg, int howto)
2211 {
2212 
2213 	if (howto & RB_NOSYNC)
2214 		return;
2215 	mtx_lock(&sync_mtx);
2216 	syncer_state = SYNCER_SHUTTING_DOWN;
2217 	rushjob = 0;
2218 	mtx_unlock(&sync_mtx);
2219 	cv_broadcast(&sync_wakeup);
2220 	kproc_shutdown(arg, howto);
2221 }
2222 
2223 void
2224 syncer_suspend(void)
2225 {
2226 
2227 	syncer_shutdown(updateproc, 0);
2228 }
2229 
2230 void
2231 syncer_resume(void)
2232 {
2233 
2234 	mtx_lock(&sync_mtx);
2235 	first_printf = 1;
2236 	syncer_state = SYNCER_RUNNING;
2237 	mtx_unlock(&sync_mtx);
2238 	cv_broadcast(&sync_wakeup);
2239 	kproc_resume(updateproc);
2240 }
2241 
2242 /*
2243  * Reassign a buffer from one vnode to another.
2244  * Used to assign file specific control information
2245  * (indirect blocks) to the vnode to which they belong.
2246  */
2247 void
2248 reassignbuf(struct buf *bp)
2249 {
2250 	struct vnode *vp;
2251 	struct bufobj *bo;
2252 	int delay;
2253 #ifdef INVARIANTS
2254 	struct bufv *bv;
2255 #endif
2256 
2257 	vp = bp->b_vp;
2258 	bo = bp->b_bufobj;
2259 	++reassignbufcalls;
2260 
2261 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2262 	    bp, bp->b_vp, bp->b_flags);
2263 	/*
2264 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2265 	 * is not fully linked in.
2266 	 */
2267 	if (bp->b_flags & B_PAGING)
2268 		panic("cannot reassign paging buffer");
2269 
2270 	/*
2271 	 * Delete from old vnode list, if on one.
2272 	 */
2273 	BO_LOCK(bo);
2274 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2275 		buf_vlist_remove(bp);
2276 	else
2277 		panic("reassignbuf: Buffer %p not on queue.", bp);
2278 	/*
2279 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2280 	 * of clean buffers.
2281 	 */
2282 	if (bp->b_flags & B_DELWRI) {
2283 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2284 			switch (vp->v_type) {
2285 			case VDIR:
2286 				delay = dirdelay;
2287 				break;
2288 			case VCHR:
2289 				delay = metadelay;
2290 				break;
2291 			default:
2292 				delay = filedelay;
2293 			}
2294 			vn_syncer_add_to_worklist(bo, delay);
2295 		}
2296 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2297 	} else {
2298 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2299 
2300 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2301 			mtx_lock(&sync_mtx);
2302 			LIST_REMOVE(bo, bo_synclist);
2303 			syncer_worklist_len--;
2304 			mtx_unlock(&sync_mtx);
2305 			bo->bo_flag &= ~BO_ONWORKLST;
2306 		}
2307 	}
2308 #ifdef INVARIANTS
2309 	bv = &bo->bo_clean;
2310 	bp = TAILQ_FIRST(&bv->bv_hd);
2311 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2312 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2313 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2314 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2315 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2316 	bv = &bo->bo_dirty;
2317 	bp = TAILQ_FIRST(&bv->bv_hd);
2318 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2319 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2320 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2321 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2322 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2323 #endif
2324 	BO_UNLOCK(bo);
2325 }
2326 
2327 /*
2328  * A temporary hack until refcount_* APIs are sorted out.
2329  */
2330 static __inline int
2331 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2332 {
2333 	u_int old;
2334 
2335 	for (;;) {
2336 		old = *count;
2337 		if (old == 0)
2338 			return (0);
2339 		if (atomic_cmpset_int(count, old, old + 1))
2340 			return (1);
2341 	}
2342 }
2343 
2344 static __inline int
2345 vfs_refcount_release_if_not_last(volatile u_int *count)
2346 {
2347 	u_int old;
2348 
2349 	for (;;) {
2350 		old = *count;
2351 		if (old == 1)
2352 			return (0);
2353 		if (atomic_cmpset_int(count, old, old - 1))
2354 			return (1);
2355 	}
2356 }
2357 
2358 static void
2359 v_init_counters(struct vnode *vp)
2360 {
2361 
2362 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2363 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2364 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2365 
2366 	refcount_init(&vp->v_holdcnt, 1);
2367 	refcount_init(&vp->v_usecount, 1);
2368 }
2369 
2370 /*
2371  * Increment the use and hold counts on the vnode, taking care to reference
2372  * the driver's usecount if this is a chardev.  The _vhold() will remove
2373  * the vnode from the free list if it is presently free.
2374  */
2375 static void
2376 v_incr_usecount(struct vnode *vp)
2377 {
2378 
2379 	ASSERT_VI_UNLOCKED(vp, __func__);
2380 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2381 
2382 	if (vp->v_type == VCHR) {
2383 		VI_LOCK(vp);
2384 		_vhold(vp, true);
2385 		if (vp->v_iflag & VI_OWEINACT) {
2386 			VNASSERT(vp->v_usecount == 0, vp,
2387 			    ("vnode with usecount and VI_OWEINACT set"));
2388 			vp->v_iflag &= ~VI_OWEINACT;
2389 		}
2390 		refcount_acquire(&vp->v_usecount);
2391 		v_incr_devcount(vp);
2392 		VI_UNLOCK(vp);
2393 		return;
2394 	}
2395 
2396 	_vhold(vp, false);
2397 	if (vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2398 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2399 		    ("vnode with usecount and VI_OWEINACT set"));
2400 	} else {
2401 		VI_LOCK(vp);
2402 		if (vp->v_iflag & VI_OWEINACT)
2403 			vp->v_iflag &= ~VI_OWEINACT;
2404 		refcount_acquire(&vp->v_usecount);
2405 		VI_UNLOCK(vp);
2406 	}
2407 }
2408 
2409 /*
2410  * Increment si_usecount of the associated device, if any.
2411  */
2412 static void
2413 v_incr_devcount(struct vnode *vp)
2414 {
2415 
2416 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2417 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2418 		dev_lock();
2419 		vp->v_rdev->si_usecount++;
2420 		dev_unlock();
2421 	}
2422 }
2423 
2424 /*
2425  * Decrement si_usecount of the associated device, if any.
2426  */
2427 static void
2428 v_decr_devcount(struct vnode *vp)
2429 {
2430 
2431 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2432 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2433 		dev_lock();
2434 		vp->v_rdev->si_usecount--;
2435 		dev_unlock();
2436 	}
2437 }
2438 
2439 /*
2440  * Grab a particular vnode from the free list, increment its
2441  * reference count and lock it.  VI_DOOMED is set if the vnode
2442  * is being destroyed.  Only callers who specify LK_RETRY will
2443  * see doomed vnodes.  If inactive processing was delayed in
2444  * vput try to do it here.
2445  *
2446  * Notes on lockless counter manipulation:
2447  * _vhold, vputx and other routines make various decisions based
2448  * on either holdcnt or usecount being 0. As long as either contuner
2449  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2450  * with atomic operations. Otherwise the interlock is taken.
2451  */
2452 int
2453 vget(struct vnode *vp, int flags, struct thread *td)
2454 {
2455 	int error, oweinact;
2456 
2457 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2458 	    ("vget: invalid lock operation"));
2459 
2460 	if ((flags & LK_INTERLOCK) != 0)
2461 		ASSERT_VI_LOCKED(vp, __func__);
2462 	else
2463 		ASSERT_VI_UNLOCKED(vp, __func__);
2464 	if ((flags & LK_VNHELD) != 0)
2465 		VNASSERT((vp->v_holdcnt > 0), vp,
2466 		    ("vget: LK_VNHELD passed but vnode not held"));
2467 
2468 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2469 
2470 	if ((flags & LK_VNHELD) == 0)
2471 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2472 
2473 	if ((error = vn_lock(vp, flags)) != 0) {
2474 		vdrop(vp);
2475 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2476 		    vp);
2477 		return (error);
2478 	}
2479 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2480 		panic("vget: vn_lock failed to return ENOENT\n");
2481 	/*
2482 	 * We don't guarantee that any particular close will
2483 	 * trigger inactive processing so just make a best effort
2484 	 * here at preventing a reference to a removed file.  If
2485 	 * we don't succeed no harm is done.
2486 	 *
2487 	 * Upgrade our holdcnt to a usecount.
2488 	 */
2489 	if (vp->v_type != VCHR &&
2490 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2491 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2492 		    ("vnode with usecount and VI_OWEINACT set"));
2493 	} else {
2494 		VI_LOCK(vp);
2495 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2496 			oweinact = 0;
2497 		} else {
2498 			oweinact = 1;
2499 			vp->v_iflag &= ~VI_OWEINACT;
2500 		}
2501 		refcount_acquire(&vp->v_usecount);
2502 		v_incr_devcount(vp);
2503 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2504 		    (flags & LK_NOWAIT) == 0)
2505 			vinactive(vp, td);
2506 		VI_UNLOCK(vp);
2507 	}
2508 	return (0);
2509 }
2510 
2511 /*
2512  * Increase the reference count of a vnode.
2513  */
2514 void
2515 vref(struct vnode *vp)
2516 {
2517 
2518 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2519 	v_incr_usecount(vp);
2520 }
2521 
2522 /*
2523  * Return reference count of a vnode.
2524  *
2525  * The results of this call are only guaranteed when some mechanism is used to
2526  * stop other processes from gaining references to the vnode.  This may be the
2527  * case if the caller holds the only reference.  This is also useful when stale
2528  * data is acceptable as race conditions may be accounted for by some other
2529  * means.
2530  */
2531 int
2532 vrefcnt(struct vnode *vp)
2533 {
2534 
2535 	return (vp->v_usecount);
2536 }
2537 
2538 #define	VPUTX_VRELE	1
2539 #define	VPUTX_VPUT	2
2540 #define	VPUTX_VUNREF	3
2541 
2542 /*
2543  * Decrement the use and hold counts for a vnode.
2544  *
2545  * See an explanation near vget() as to why atomic operation is safe.
2546  */
2547 static void
2548 vputx(struct vnode *vp, int func)
2549 {
2550 	int error;
2551 
2552 	KASSERT(vp != NULL, ("vputx: null vp"));
2553 	if (func == VPUTX_VUNREF)
2554 		ASSERT_VOP_LOCKED(vp, "vunref");
2555 	else if (func == VPUTX_VPUT)
2556 		ASSERT_VOP_LOCKED(vp, "vput");
2557 	else
2558 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2559 	ASSERT_VI_UNLOCKED(vp, __func__);
2560 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2561 
2562 	if (vp->v_type != VCHR &&
2563 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2564 		if (func == VPUTX_VPUT)
2565 			VOP_UNLOCK(vp, 0);
2566 		vdrop(vp);
2567 		return;
2568 	}
2569 
2570 	VI_LOCK(vp);
2571 
2572 	/*
2573 	 * We want to hold the vnode until the inactive finishes to
2574 	 * prevent vgone() races.  We drop the use count here and the
2575 	 * hold count below when we're done.
2576 	 */
2577 	if (!refcount_release(&vp->v_usecount) ||
2578 	    (vp->v_iflag & VI_DOINGINACT)) {
2579 		if (func == VPUTX_VPUT)
2580 			VOP_UNLOCK(vp, 0);
2581 		v_decr_devcount(vp);
2582 		vdropl(vp);
2583 		return;
2584 	}
2585 
2586 	v_decr_devcount(vp);
2587 
2588 	error = 0;
2589 
2590 	if (vp->v_usecount != 0) {
2591 		vprint("vputx: usecount not zero", vp);
2592 		panic("vputx: usecount not zero");
2593 	}
2594 
2595 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2596 
2597 	/*
2598 	 * We must call VOP_INACTIVE with the node locked. Mark
2599 	 * as VI_DOINGINACT to avoid recursion.
2600 	 */
2601 	vp->v_iflag |= VI_OWEINACT;
2602 	switch (func) {
2603 	case VPUTX_VRELE:
2604 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2605 		VI_LOCK(vp);
2606 		break;
2607 	case VPUTX_VPUT:
2608 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2609 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2610 			    LK_NOWAIT);
2611 			VI_LOCK(vp);
2612 		}
2613 		break;
2614 	case VPUTX_VUNREF:
2615 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2616 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2617 			VI_LOCK(vp);
2618 		}
2619 		break;
2620 	}
2621 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2622 	    ("vnode with usecount and VI_OWEINACT set"));
2623 	if (error == 0) {
2624 		if (vp->v_iflag & VI_OWEINACT)
2625 			vinactive(vp, curthread);
2626 		if (func != VPUTX_VUNREF)
2627 			VOP_UNLOCK(vp, 0);
2628 	}
2629 	vdropl(vp);
2630 }
2631 
2632 /*
2633  * Vnode put/release.
2634  * If count drops to zero, call inactive routine and return to freelist.
2635  */
2636 void
2637 vrele(struct vnode *vp)
2638 {
2639 
2640 	vputx(vp, VPUTX_VRELE);
2641 }
2642 
2643 /*
2644  * Release an already locked vnode.  This give the same effects as
2645  * unlock+vrele(), but takes less time and avoids releasing and
2646  * re-aquiring the lock (as vrele() acquires the lock internally.)
2647  */
2648 void
2649 vput(struct vnode *vp)
2650 {
2651 
2652 	vputx(vp, VPUTX_VPUT);
2653 }
2654 
2655 /*
2656  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2657  */
2658 void
2659 vunref(struct vnode *vp)
2660 {
2661 
2662 	vputx(vp, VPUTX_VUNREF);
2663 }
2664 
2665 /*
2666  * Increase the hold count and activate if this is the first reference.
2667  */
2668 void
2669 _vhold(struct vnode *vp, bool locked)
2670 {
2671 	struct mount *mp;
2672 
2673 	if (locked)
2674 		ASSERT_VI_LOCKED(vp, __func__);
2675 	else
2676 		ASSERT_VI_UNLOCKED(vp, __func__);
2677 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2678 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2679 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2680 		    ("_vhold: vnode with holdcnt is free"));
2681 		return;
2682 	}
2683 
2684 	if (!locked)
2685 		VI_LOCK(vp);
2686 	if ((vp->v_iflag & VI_FREE) == 0) {
2687 		refcount_acquire(&vp->v_holdcnt);
2688 		if (!locked)
2689 			VI_UNLOCK(vp);
2690 		return;
2691 	}
2692 	VNASSERT(vp->v_holdcnt == 0, vp,
2693 	    ("%s: wrong hold count", __func__));
2694 	VNASSERT(vp->v_op != NULL, vp,
2695 	    ("%s: vnode already reclaimed.", __func__));
2696 	/*
2697 	 * Remove a vnode from the free list, mark it as in use,
2698 	 * and put it on the active list.
2699 	 */
2700 	mtx_lock(&vnode_free_list_mtx);
2701 	TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2702 	freevnodes--;
2703 	vp->v_iflag &= ~VI_FREE;
2704 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2705 	    ("Activating already active vnode"));
2706 	vp->v_iflag |= VI_ACTIVE;
2707 	mp = vp->v_mount;
2708 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2709 	mp->mnt_activevnodelistsize++;
2710 	mtx_unlock(&vnode_free_list_mtx);
2711 	refcount_acquire(&vp->v_holdcnt);
2712 	if (!locked)
2713 		VI_UNLOCK(vp);
2714 }
2715 
2716 /*
2717  * Drop the hold count of the vnode.  If this is the last reference to
2718  * the vnode we place it on the free list unless it has been vgone'd
2719  * (marked VI_DOOMED) in which case we will free it.
2720  *
2721  * Because the vnode vm object keeps a hold reference on the vnode if
2722  * there is at least one resident non-cached page, the vnode cannot
2723  * leave the active list without the page cleanup done.
2724  */
2725 void
2726 _vdrop(struct vnode *vp, bool locked)
2727 {
2728 	struct bufobj *bo;
2729 	struct mount *mp;
2730 	int active;
2731 
2732 	if (locked)
2733 		ASSERT_VI_LOCKED(vp, __func__);
2734 	else
2735 		ASSERT_VI_UNLOCKED(vp, __func__);
2736 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2737 	if ((int)vp->v_holdcnt <= 0)
2738 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2739 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2740 		if (locked)
2741 			VI_UNLOCK(vp);
2742 		return;
2743 	}
2744 
2745 	if (!locked)
2746 		VI_LOCK(vp);
2747 	if (refcount_release(&vp->v_holdcnt) == 0) {
2748 		VI_UNLOCK(vp);
2749 		return;
2750 	}
2751 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2752 		/*
2753 		 * Mark a vnode as free: remove it from its active list
2754 		 * and put it up for recycling on the freelist.
2755 		 */
2756 		VNASSERT(vp->v_op != NULL, vp,
2757 		    ("vdropl: vnode already reclaimed."));
2758 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2759 		    ("vnode already free"));
2760 		VNASSERT(vp->v_holdcnt == 0, vp,
2761 		    ("vdropl: freeing when we shouldn't"));
2762 		active = vp->v_iflag & VI_ACTIVE;
2763 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2764 			vp->v_iflag &= ~VI_ACTIVE;
2765 			mp = vp->v_mount;
2766 			mtx_lock(&vnode_free_list_mtx);
2767 			if (active) {
2768 				TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2769 				    v_actfreelist);
2770 				mp->mnt_activevnodelistsize--;
2771 			}
2772 			TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2773 			    v_actfreelist);
2774 			freevnodes++;
2775 			vp->v_iflag |= VI_FREE;
2776 			mtx_unlock(&vnode_free_list_mtx);
2777 		} else {
2778 			atomic_add_long(&free_owe_inact, 1);
2779 		}
2780 		VI_UNLOCK(vp);
2781 		return;
2782 	}
2783 	/*
2784 	 * The vnode has been marked for destruction, so free it.
2785 	 *
2786 	 * The vnode will be returned to the zone where it will
2787 	 * normally remain until it is needed for another vnode. We
2788 	 * need to cleanup (or verify that the cleanup has already
2789 	 * been done) any residual data left from its current use
2790 	 * so as not to contaminate the freshly allocated vnode.
2791 	 */
2792 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2793 	atomic_subtract_long(&numvnodes, 1);
2794 	bo = &vp->v_bufobj;
2795 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2796 	    ("cleaned vnode still on the free list."));
2797 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2798 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2799 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2800 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2801 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2802 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2803 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2804 	    ("clean blk trie not empty"));
2805 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2806 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2807 	    ("dirty blk trie not empty"));
2808 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2809 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2810 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2811 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2812 	    ("Dangling rangelock waiters"));
2813 	VI_UNLOCK(vp);
2814 #ifdef MAC
2815 	mac_vnode_destroy(vp);
2816 #endif
2817 	if (vp->v_pollinfo != NULL) {
2818 		destroy_vpollinfo(vp->v_pollinfo);
2819 		vp->v_pollinfo = NULL;
2820 	}
2821 #ifdef INVARIANTS
2822 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2823 	vp->v_op = NULL;
2824 #endif
2825 	bzero(&vp->v_un, sizeof(vp->v_un));
2826 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2827 	vp->v_iflag = 0;
2828 	vp->v_vflag = 0;
2829 	bo->bo_flag = 0;
2830 	uma_zfree(vnode_zone, vp);
2831 }
2832 
2833 /*
2834  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2835  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2836  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2837  * failed lock upgrade.
2838  */
2839 void
2840 vinactive(struct vnode *vp, struct thread *td)
2841 {
2842 	struct vm_object *obj;
2843 
2844 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2845 	ASSERT_VI_LOCKED(vp, "vinactive");
2846 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2847 	    ("vinactive: recursed on VI_DOINGINACT"));
2848 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2849 	vp->v_iflag |= VI_DOINGINACT;
2850 	vp->v_iflag &= ~VI_OWEINACT;
2851 	VI_UNLOCK(vp);
2852 	/*
2853 	 * Before moving off the active list, we must be sure that any
2854 	 * modified pages are converted into the vnode's dirty
2855 	 * buffers, since these will no longer be checked once the
2856 	 * vnode is on the inactive list.
2857 	 *
2858 	 * The write-out of the dirty pages is asynchronous.  At the
2859 	 * point that VOP_INACTIVE() is called, there could still be
2860 	 * pending I/O and dirty pages in the object.
2861 	 */
2862 	obj = vp->v_object;
2863 	if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2864 		VM_OBJECT_WLOCK(obj);
2865 		vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC);
2866 		VM_OBJECT_WUNLOCK(obj);
2867 	}
2868 	VOP_INACTIVE(vp, td);
2869 	VI_LOCK(vp);
2870 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2871 	    ("vinactive: lost VI_DOINGINACT"));
2872 	vp->v_iflag &= ~VI_DOINGINACT;
2873 }
2874 
2875 /*
2876  * Remove any vnodes in the vnode table belonging to mount point mp.
2877  *
2878  * If FORCECLOSE is not specified, there should not be any active ones,
2879  * return error if any are found (nb: this is a user error, not a
2880  * system error). If FORCECLOSE is specified, detach any active vnodes
2881  * that are found.
2882  *
2883  * If WRITECLOSE is set, only flush out regular file vnodes open for
2884  * writing.
2885  *
2886  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2887  *
2888  * `rootrefs' specifies the base reference count for the root vnode
2889  * of this filesystem. The root vnode is considered busy if its
2890  * v_usecount exceeds this value. On a successful return, vflush(, td)
2891  * will call vrele() on the root vnode exactly rootrefs times.
2892  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2893  * be zero.
2894  */
2895 #ifdef DIAGNOSTIC
2896 static int busyprt = 0;		/* print out busy vnodes */
2897 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2898 #endif
2899 
2900 int
2901 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2902 {
2903 	struct vnode *vp, *mvp, *rootvp = NULL;
2904 	struct vattr vattr;
2905 	int busy = 0, error;
2906 
2907 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2908 	    rootrefs, flags);
2909 	if (rootrefs > 0) {
2910 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2911 		    ("vflush: bad args"));
2912 		/*
2913 		 * Get the filesystem root vnode. We can vput() it
2914 		 * immediately, since with rootrefs > 0, it won't go away.
2915 		 */
2916 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2917 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2918 			    __func__, error);
2919 			return (error);
2920 		}
2921 		vput(rootvp);
2922 	}
2923 loop:
2924 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2925 		vholdl(vp);
2926 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2927 		if (error) {
2928 			vdrop(vp);
2929 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2930 			goto loop;
2931 		}
2932 		/*
2933 		 * Skip over a vnodes marked VV_SYSTEM.
2934 		 */
2935 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2936 			VOP_UNLOCK(vp, 0);
2937 			vdrop(vp);
2938 			continue;
2939 		}
2940 		/*
2941 		 * If WRITECLOSE is set, flush out unlinked but still open
2942 		 * files (even if open only for reading) and regular file
2943 		 * vnodes open for writing.
2944 		 */
2945 		if (flags & WRITECLOSE) {
2946 			if (vp->v_object != NULL) {
2947 				VM_OBJECT_WLOCK(vp->v_object);
2948 				vm_object_page_clean(vp->v_object, 0, 0, 0);
2949 				VM_OBJECT_WUNLOCK(vp->v_object);
2950 			}
2951 			error = VOP_FSYNC(vp, MNT_WAIT, td);
2952 			if (error != 0) {
2953 				VOP_UNLOCK(vp, 0);
2954 				vdrop(vp);
2955 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2956 				return (error);
2957 			}
2958 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
2959 			VI_LOCK(vp);
2960 
2961 			if ((vp->v_type == VNON ||
2962 			    (error == 0 && vattr.va_nlink > 0)) &&
2963 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2964 				VOP_UNLOCK(vp, 0);
2965 				vdropl(vp);
2966 				continue;
2967 			}
2968 		} else
2969 			VI_LOCK(vp);
2970 		/*
2971 		 * With v_usecount == 0, all we need to do is clear out the
2972 		 * vnode data structures and we are done.
2973 		 *
2974 		 * If FORCECLOSE is set, forcibly close the vnode.
2975 		 */
2976 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2977 			vgonel(vp);
2978 		} else {
2979 			busy++;
2980 #ifdef DIAGNOSTIC
2981 			if (busyprt)
2982 				vprint("vflush: busy vnode", vp);
2983 #endif
2984 		}
2985 		VOP_UNLOCK(vp, 0);
2986 		vdropl(vp);
2987 	}
2988 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2989 		/*
2990 		 * If just the root vnode is busy, and if its refcount
2991 		 * is equal to `rootrefs', then go ahead and kill it.
2992 		 */
2993 		VI_LOCK(rootvp);
2994 		KASSERT(busy > 0, ("vflush: not busy"));
2995 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2996 		    ("vflush: usecount %d < rootrefs %d",
2997 		     rootvp->v_usecount, rootrefs));
2998 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2999 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3000 			vgone(rootvp);
3001 			VOP_UNLOCK(rootvp, 0);
3002 			busy = 0;
3003 		} else
3004 			VI_UNLOCK(rootvp);
3005 	}
3006 	if (busy) {
3007 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3008 		    busy);
3009 		return (EBUSY);
3010 	}
3011 	for (; rootrefs > 0; rootrefs--)
3012 		vrele(rootvp);
3013 	return (0);
3014 }
3015 
3016 /*
3017  * Recycle an unused vnode to the front of the free list.
3018  */
3019 int
3020 vrecycle(struct vnode *vp)
3021 {
3022 	int recycled;
3023 
3024 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
3025 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3026 	recycled = 0;
3027 	VI_LOCK(vp);
3028 	if (vp->v_usecount == 0) {
3029 		recycled = 1;
3030 		vgonel(vp);
3031 	}
3032 	VI_UNLOCK(vp);
3033 	return (recycled);
3034 }
3035 
3036 /*
3037  * Eliminate all activity associated with a vnode
3038  * in preparation for reuse.
3039  */
3040 void
3041 vgone(struct vnode *vp)
3042 {
3043 	VI_LOCK(vp);
3044 	vgonel(vp);
3045 	VI_UNLOCK(vp);
3046 }
3047 
3048 static void
3049 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3050     struct vnode *lowervp __unused)
3051 {
3052 }
3053 
3054 /*
3055  * Notify upper mounts about reclaimed or unlinked vnode.
3056  */
3057 void
3058 vfs_notify_upper(struct vnode *vp, int event)
3059 {
3060 	static struct vfsops vgonel_vfsops = {
3061 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3062 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3063 	};
3064 	struct mount *mp, *ump, *mmp;
3065 
3066 	mp = vp->v_mount;
3067 	if (mp == NULL)
3068 		return;
3069 
3070 	MNT_ILOCK(mp);
3071 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3072 		goto unlock;
3073 	MNT_IUNLOCK(mp);
3074 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3075 	mmp->mnt_op = &vgonel_vfsops;
3076 	mmp->mnt_kern_flag |= MNTK_MARKER;
3077 	MNT_ILOCK(mp);
3078 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3079 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3080 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3081 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3082 			continue;
3083 		}
3084 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3085 		MNT_IUNLOCK(mp);
3086 		switch (event) {
3087 		case VFS_NOTIFY_UPPER_RECLAIM:
3088 			VFS_RECLAIM_LOWERVP(ump, vp);
3089 			break;
3090 		case VFS_NOTIFY_UPPER_UNLINK:
3091 			VFS_UNLINK_LOWERVP(ump, vp);
3092 			break;
3093 		default:
3094 			KASSERT(0, ("invalid event %d", event));
3095 			break;
3096 		}
3097 		MNT_ILOCK(mp);
3098 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3099 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3100 	}
3101 	free(mmp, M_TEMP);
3102 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3103 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3104 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3105 		wakeup(&mp->mnt_uppers);
3106 	}
3107 unlock:
3108 	MNT_IUNLOCK(mp);
3109 }
3110 
3111 /*
3112  * vgone, with the vp interlock held.
3113  */
3114 static void
3115 vgonel(struct vnode *vp)
3116 {
3117 	struct thread *td;
3118 	int oweinact;
3119 	int active;
3120 	struct mount *mp;
3121 
3122 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3123 	ASSERT_VI_LOCKED(vp, "vgonel");
3124 	VNASSERT(vp->v_holdcnt, vp,
3125 	    ("vgonel: vp %p has no reference.", vp));
3126 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3127 	td = curthread;
3128 
3129 	/*
3130 	 * Don't vgonel if we're already doomed.
3131 	 */
3132 	if (vp->v_iflag & VI_DOOMED)
3133 		return;
3134 	vp->v_iflag |= VI_DOOMED;
3135 
3136 	/*
3137 	 * Check to see if the vnode is in use.  If so, we have to call
3138 	 * VOP_CLOSE() and VOP_INACTIVE().
3139 	 */
3140 	active = vp->v_usecount;
3141 	oweinact = (vp->v_iflag & VI_OWEINACT);
3142 	VI_UNLOCK(vp);
3143 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3144 
3145 	/*
3146 	 * If purging an active vnode, it must be closed and
3147 	 * deactivated before being reclaimed.
3148 	 */
3149 	if (active)
3150 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3151 	if (oweinact || active) {
3152 		VI_LOCK(vp);
3153 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3154 			vinactive(vp, td);
3155 		VI_UNLOCK(vp);
3156 	}
3157 	if (vp->v_type == VSOCK)
3158 		vfs_unp_reclaim(vp);
3159 
3160 	/*
3161 	 * Clean out any buffers associated with the vnode.
3162 	 * If the flush fails, just toss the buffers.
3163 	 */
3164 	mp = NULL;
3165 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3166 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3167 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3168 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3169 			;
3170 	}
3171 
3172 	BO_LOCK(&vp->v_bufobj);
3173 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3174 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3175 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3176 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3177 	    ("vp %p bufobj not invalidated", vp));
3178 	vp->v_bufobj.bo_flag |= BO_DEAD;
3179 	BO_UNLOCK(&vp->v_bufobj);
3180 
3181 	/*
3182 	 * Reclaim the vnode.
3183 	 */
3184 	if (VOP_RECLAIM(vp, td))
3185 		panic("vgone: cannot reclaim");
3186 	if (mp != NULL)
3187 		vn_finished_secondary_write(mp);
3188 	VNASSERT(vp->v_object == NULL, vp,
3189 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3190 	/*
3191 	 * Clear the advisory locks and wake up waiting threads.
3192 	 */
3193 	(void)VOP_ADVLOCKPURGE(vp);
3194 	vp->v_lockf = NULL;
3195 	/*
3196 	 * Delete from old mount point vnode list.
3197 	 */
3198 	delmntque(vp);
3199 	cache_purge(vp);
3200 	/*
3201 	 * Done with purge, reset to the standard lock and invalidate
3202 	 * the vnode.
3203 	 */
3204 	VI_LOCK(vp);
3205 	vp->v_vnlock = &vp->v_lock;
3206 	vp->v_op = &dead_vnodeops;
3207 	vp->v_tag = "none";
3208 	vp->v_type = VBAD;
3209 }
3210 
3211 /*
3212  * Calculate the total number of references to a special device.
3213  */
3214 int
3215 vcount(struct vnode *vp)
3216 {
3217 	int count;
3218 
3219 	dev_lock();
3220 	count = vp->v_rdev->si_usecount;
3221 	dev_unlock();
3222 	return (count);
3223 }
3224 
3225 /*
3226  * Same as above, but using the struct cdev *as argument
3227  */
3228 int
3229 count_dev(struct cdev *dev)
3230 {
3231 	int count;
3232 
3233 	dev_lock();
3234 	count = dev->si_usecount;
3235 	dev_unlock();
3236 	return(count);
3237 }
3238 
3239 /*
3240  * Print out a description of a vnode.
3241  */
3242 static char *typename[] =
3243 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3244  "VMARKER"};
3245 
3246 void
3247 vn_printf(struct vnode *vp, const char *fmt, ...)
3248 {
3249 	va_list ap;
3250 	char buf[256], buf2[16];
3251 	u_long flags;
3252 
3253 	va_start(ap, fmt);
3254 	vprintf(fmt, ap);
3255 	va_end(ap);
3256 	printf("%p: ", (void *)vp);
3257 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3258 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3259 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3260 	buf[0] = '\0';
3261 	buf[1] = '\0';
3262 	if (vp->v_vflag & VV_ROOT)
3263 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3264 	if (vp->v_vflag & VV_ISTTY)
3265 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3266 	if (vp->v_vflag & VV_NOSYNC)
3267 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3268 	if (vp->v_vflag & VV_ETERNALDEV)
3269 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3270 	if (vp->v_vflag & VV_CACHEDLABEL)
3271 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3272 	if (vp->v_vflag & VV_TEXT)
3273 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3274 	if (vp->v_vflag & VV_COPYONWRITE)
3275 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3276 	if (vp->v_vflag & VV_SYSTEM)
3277 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3278 	if (vp->v_vflag & VV_PROCDEP)
3279 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3280 	if (vp->v_vflag & VV_NOKNOTE)
3281 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3282 	if (vp->v_vflag & VV_DELETED)
3283 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3284 	if (vp->v_vflag & VV_MD)
3285 		strlcat(buf, "|VV_MD", sizeof(buf));
3286 	if (vp->v_vflag & VV_FORCEINSMQ)
3287 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3288 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3289 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3290 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3291 	if (flags != 0) {
3292 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3293 		strlcat(buf, buf2, sizeof(buf));
3294 	}
3295 	if (vp->v_iflag & VI_MOUNT)
3296 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3297 	if (vp->v_iflag & VI_DOOMED)
3298 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3299 	if (vp->v_iflag & VI_FREE)
3300 		strlcat(buf, "|VI_FREE", sizeof(buf));
3301 	if (vp->v_iflag & VI_ACTIVE)
3302 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3303 	if (vp->v_iflag & VI_DOINGINACT)
3304 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3305 	if (vp->v_iflag & VI_OWEINACT)
3306 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3307 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3308 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3309 	if (flags != 0) {
3310 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3311 		strlcat(buf, buf2, sizeof(buf));
3312 	}
3313 	printf("    flags (%s)\n", buf + 1);
3314 	if (mtx_owned(VI_MTX(vp)))
3315 		printf(" VI_LOCKed");
3316 	if (vp->v_object != NULL)
3317 		printf("    v_object %p ref %d pages %d "
3318 		    "cleanbuf %d dirtybuf %d\n",
3319 		    vp->v_object, vp->v_object->ref_count,
3320 		    vp->v_object->resident_page_count,
3321 		    vp->v_bufobj.bo_clean.bv_cnt,
3322 		    vp->v_bufobj.bo_dirty.bv_cnt);
3323 	printf("    ");
3324 	lockmgr_printinfo(vp->v_vnlock);
3325 	if (vp->v_data != NULL)
3326 		VOP_PRINT(vp);
3327 }
3328 
3329 #ifdef DDB
3330 /*
3331  * List all of the locked vnodes in the system.
3332  * Called when debugging the kernel.
3333  */
3334 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3335 {
3336 	struct mount *mp;
3337 	struct vnode *vp;
3338 
3339 	/*
3340 	 * Note: because this is DDB, we can't obey the locking semantics
3341 	 * for these structures, which means we could catch an inconsistent
3342 	 * state and dereference a nasty pointer.  Not much to be done
3343 	 * about that.
3344 	 */
3345 	db_printf("Locked vnodes\n");
3346 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3347 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3348 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3349 				vprint("", vp);
3350 		}
3351 	}
3352 }
3353 
3354 /*
3355  * Show details about the given vnode.
3356  */
3357 DB_SHOW_COMMAND(vnode, db_show_vnode)
3358 {
3359 	struct vnode *vp;
3360 
3361 	if (!have_addr)
3362 		return;
3363 	vp = (struct vnode *)addr;
3364 	vn_printf(vp, "vnode ");
3365 }
3366 
3367 /*
3368  * Show details about the given mount point.
3369  */
3370 DB_SHOW_COMMAND(mount, db_show_mount)
3371 {
3372 	struct mount *mp;
3373 	struct vfsopt *opt;
3374 	struct statfs *sp;
3375 	struct vnode *vp;
3376 	char buf[512];
3377 	uint64_t mflags;
3378 	u_int flags;
3379 
3380 	if (!have_addr) {
3381 		/* No address given, print short info about all mount points. */
3382 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3383 			db_printf("%p %s on %s (%s)\n", mp,
3384 			    mp->mnt_stat.f_mntfromname,
3385 			    mp->mnt_stat.f_mntonname,
3386 			    mp->mnt_stat.f_fstypename);
3387 			if (db_pager_quit)
3388 				break;
3389 		}
3390 		db_printf("\nMore info: show mount <addr>\n");
3391 		return;
3392 	}
3393 
3394 	mp = (struct mount *)addr;
3395 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3396 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3397 
3398 	buf[0] = '\0';
3399 	mflags = mp->mnt_flag;
3400 #define	MNT_FLAG(flag)	do {						\
3401 	if (mflags & (flag)) {						\
3402 		if (buf[0] != '\0')					\
3403 			strlcat(buf, ", ", sizeof(buf));		\
3404 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3405 		mflags &= ~(flag);					\
3406 	}								\
3407 } while (0)
3408 	MNT_FLAG(MNT_RDONLY);
3409 	MNT_FLAG(MNT_SYNCHRONOUS);
3410 	MNT_FLAG(MNT_NOEXEC);
3411 	MNT_FLAG(MNT_NOSUID);
3412 	MNT_FLAG(MNT_NFS4ACLS);
3413 	MNT_FLAG(MNT_UNION);
3414 	MNT_FLAG(MNT_ASYNC);
3415 	MNT_FLAG(MNT_SUIDDIR);
3416 	MNT_FLAG(MNT_SOFTDEP);
3417 	MNT_FLAG(MNT_NOSYMFOLLOW);
3418 	MNT_FLAG(MNT_GJOURNAL);
3419 	MNT_FLAG(MNT_MULTILABEL);
3420 	MNT_FLAG(MNT_ACLS);
3421 	MNT_FLAG(MNT_NOATIME);
3422 	MNT_FLAG(MNT_NOCLUSTERR);
3423 	MNT_FLAG(MNT_NOCLUSTERW);
3424 	MNT_FLAG(MNT_SUJ);
3425 	MNT_FLAG(MNT_EXRDONLY);
3426 	MNT_FLAG(MNT_EXPORTED);
3427 	MNT_FLAG(MNT_DEFEXPORTED);
3428 	MNT_FLAG(MNT_EXPORTANON);
3429 	MNT_FLAG(MNT_EXKERB);
3430 	MNT_FLAG(MNT_EXPUBLIC);
3431 	MNT_FLAG(MNT_LOCAL);
3432 	MNT_FLAG(MNT_QUOTA);
3433 	MNT_FLAG(MNT_ROOTFS);
3434 	MNT_FLAG(MNT_USER);
3435 	MNT_FLAG(MNT_IGNORE);
3436 	MNT_FLAG(MNT_UPDATE);
3437 	MNT_FLAG(MNT_DELEXPORT);
3438 	MNT_FLAG(MNT_RELOAD);
3439 	MNT_FLAG(MNT_FORCE);
3440 	MNT_FLAG(MNT_SNAPSHOT);
3441 	MNT_FLAG(MNT_BYFSID);
3442 #undef MNT_FLAG
3443 	if (mflags != 0) {
3444 		if (buf[0] != '\0')
3445 			strlcat(buf, ", ", sizeof(buf));
3446 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3447 		    "0x%016jx", mflags);
3448 	}
3449 	db_printf("    mnt_flag = %s\n", buf);
3450 
3451 	buf[0] = '\0';
3452 	flags = mp->mnt_kern_flag;
3453 #define	MNT_KERN_FLAG(flag)	do {					\
3454 	if (flags & (flag)) {						\
3455 		if (buf[0] != '\0')					\
3456 			strlcat(buf, ", ", sizeof(buf));		\
3457 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3458 		flags &= ~(flag);					\
3459 	}								\
3460 } while (0)
3461 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3462 	MNT_KERN_FLAG(MNTK_ASYNC);
3463 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3464 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3465 	MNT_KERN_FLAG(MNTK_DRAINING);
3466 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3467 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3468 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3469 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3470 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3471 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3472 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3473 	MNT_KERN_FLAG(MNTK_MARKER);
3474 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3475 	MNT_KERN_FLAG(MNTK_NOASYNC);
3476 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3477 	MNT_KERN_FLAG(MNTK_MWAIT);
3478 	MNT_KERN_FLAG(MNTK_SUSPEND);
3479 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3480 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3481 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3482 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3483 #undef MNT_KERN_FLAG
3484 	if (flags != 0) {
3485 		if (buf[0] != '\0')
3486 			strlcat(buf, ", ", sizeof(buf));
3487 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3488 		    "0x%08x", flags);
3489 	}
3490 	db_printf("    mnt_kern_flag = %s\n", buf);
3491 
3492 	db_printf("    mnt_opt = ");
3493 	opt = TAILQ_FIRST(mp->mnt_opt);
3494 	if (opt != NULL) {
3495 		db_printf("%s", opt->name);
3496 		opt = TAILQ_NEXT(opt, link);
3497 		while (opt != NULL) {
3498 			db_printf(", %s", opt->name);
3499 			opt = TAILQ_NEXT(opt, link);
3500 		}
3501 	}
3502 	db_printf("\n");
3503 
3504 	sp = &mp->mnt_stat;
3505 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3506 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3507 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3508 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3509 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3510 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3511 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3512 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3513 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3514 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3515 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3516 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3517 
3518 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3519 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3520 	if (jailed(mp->mnt_cred))
3521 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3522 	db_printf(" }\n");
3523 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3524 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3525 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3526 	db_printf("    mnt_activevnodelistsize = %d\n",
3527 	    mp->mnt_activevnodelistsize);
3528 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3529 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3530 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3531 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3532 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3533 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3534 	db_printf("    mnt_secondary_accwrites = %d\n",
3535 	    mp->mnt_secondary_accwrites);
3536 	db_printf("    mnt_gjprovider = %s\n",
3537 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3538 
3539 	db_printf("\n\nList of active vnodes\n");
3540 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3541 		if (vp->v_type != VMARKER) {
3542 			vn_printf(vp, "vnode ");
3543 			if (db_pager_quit)
3544 				break;
3545 		}
3546 	}
3547 	db_printf("\n\nList of inactive vnodes\n");
3548 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3549 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3550 			vn_printf(vp, "vnode ");
3551 			if (db_pager_quit)
3552 				break;
3553 		}
3554 	}
3555 }
3556 #endif	/* DDB */
3557 
3558 /*
3559  * Fill in a struct xvfsconf based on a struct vfsconf.
3560  */
3561 static int
3562 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3563 {
3564 	struct xvfsconf xvfsp;
3565 
3566 	bzero(&xvfsp, sizeof(xvfsp));
3567 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3568 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3569 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3570 	xvfsp.vfc_flags = vfsp->vfc_flags;
3571 	/*
3572 	 * These are unused in userland, we keep them
3573 	 * to not break binary compatibility.
3574 	 */
3575 	xvfsp.vfc_vfsops = NULL;
3576 	xvfsp.vfc_next = NULL;
3577 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3578 }
3579 
3580 #ifdef COMPAT_FREEBSD32
3581 struct xvfsconf32 {
3582 	uint32_t	vfc_vfsops;
3583 	char		vfc_name[MFSNAMELEN];
3584 	int32_t		vfc_typenum;
3585 	int32_t		vfc_refcount;
3586 	int32_t		vfc_flags;
3587 	uint32_t	vfc_next;
3588 };
3589 
3590 static int
3591 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3592 {
3593 	struct xvfsconf32 xvfsp;
3594 
3595 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3596 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3597 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3598 	xvfsp.vfc_flags = vfsp->vfc_flags;
3599 	xvfsp.vfc_vfsops = 0;
3600 	xvfsp.vfc_next = 0;
3601 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3602 }
3603 #endif
3604 
3605 /*
3606  * Top level filesystem related information gathering.
3607  */
3608 static int
3609 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3610 {
3611 	struct vfsconf *vfsp;
3612 	int error;
3613 
3614 	error = 0;
3615 	vfsconf_slock();
3616 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3617 #ifdef COMPAT_FREEBSD32
3618 		if (req->flags & SCTL_MASK32)
3619 			error = vfsconf2x32(req, vfsp);
3620 		else
3621 #endif
3622 			error = vfsconf2x(req, vfsp);
3623 		if (error)
3624 			break;
3625 	}
3626 	vfsconf_sunlock();
3627 	return (error);
3628 }
3629 
3630 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3631     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3632     "S,xvfsconf", "List of all configured filesystems");
3633 
3634 #ifndef BURN_BRIDGES
3635 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3636 
3637 static int
3638 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3639 {
3640 	int *name = (int *)arg1 - 1;	/* XXX */
3641 	u_int namelen = arg2 + 1;	/* XXX */
3642 	struct vfsconf *vfsp;
3643 
3644 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3645 	    "please rebuild world\n");
3646 
3647 #if 1 || defined(COMPAT_PRELITE2)
3648 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3649 	if (namelen == 1)
3650 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3651 #endif
3652 
3653 	switch (name[1]) {
3654 	case VFS_MAXTYPENUM:
3655 		if (namelen != 2)
3656 			return (ENOTDIR);
3657 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3658 	case VFS_CONF:
3659 		if (namelen != 3)
3660 			return (ENOTDIR);	/* overloaded */
3661 		vfsconf_slock();
3662 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3663 			if (vfsp->vfc_typenum == name[2])
3664 				break;
3665 		}
3666 		vfsconf_sunlock();
3667 		if (vfsp == NULL)
3668 			return (EOPNOTSUPP);
3669 #ifdef COMPAT_FREEBSD32
3670 		if (req->flags & SCTL_MASK32)
3671 			return (vfsconf2x32(req, vfsp));
3672 		else
3673 #endif
3674 			return (vfsconf2x(req, vfsp));
3675 	}
3676 	return (EOPNOTSUPP);
3677 }
3678 
3679 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3680     CTLFLAG_MPSAFE, vfs_sysctl,
3681     "Generic filesystem");
3682 
3683 #if 1 || defined(COMPAT_PRELITE2)
3684 
3685 static int
3686 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3687 {
3688 	int error;
3689 	struct vfsconf *vfsp;
3690 	struct ovfsconf ovfs;
3691 
3692 	vfsconf_slock();
3693 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3694 		bzero(&ovfs, sizeof(ovfs));
3695 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3696 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3697 		ovfs.vfc_index = vfsp->vfc_typenum;
3698 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3699 		ovfs.vfc_flags = vfsp->vfc_flags;
3700 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3701 		if (error != 0) {
3702 			vfsconf_sunlock();
3703 			return (error);
3704 		}
3705 	}
3706 	vfsconf_sunlock();
3707 	return (0);
3708 }
3709 
3710 #endif /* 1 || COMPAT_PRELITE2 */
3711 #endif /* !BURN_BRIDGES */
3712 
3713 #define KINFO_VNODESLOP		10
3714 #ifdef notyet
3715 /*
3716  * Dump vnode list (via sysctl).
3717  */
3718 /* ARGSUSED */
3719 static int
3720 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3721 {
3722 	struct xvnode *xvn;
3723 	struct mount *mp;
3724 	struct vnode *vp;
3725 	int error, len, n;
3726 
3727 	/*
3728 	 * Stale numvnodes access is not fatal here.
3729 	 */
3730 	req->lock = 0;
3731 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3732 	if (!req->oldptr)
3733 		/* Make an estimate */
3734 		return (SYSCTL_OUT(req, 0, len));
3735 
3736 	error = sysctl_wire_old_buffer(req, 0);
3737 	if (error != 0)
3738 		return (error);
3739 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3740 	n = 0;
3741 	mtx_lock(&mountlist_mtx);
3742 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3743 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3744 			continue;
3745 		MNT_ILOCK(mp);
3746 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3747 			if (n == len)
3748 				break;
3749 			vref(vp);
3750 			xvn[n].xv_size = sizeof *xvn;
3751 			xvn[n].xv_vnode = vp;
3752 			xvn[n].xv_id = 0;	/* XXX compat */
3753 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3754 			XV_COPY(usecount);
3755 			XV_COPY(writecount);
3756 			XV_COPY(holdcnt);
3757 			XV_COPY(mount);
3758 			XV_COPY(numoutput);
3759 			XV_COPY(type);
3760 #undef XV_COPY
3761 			xvn[n].xv_flag = vp->v_vflag;
3762 
3763 			switch (vp->v_type) {
3764 			case VREG:
3765 			case VDIR:
3766 			case VLNK:
3767 				break;
3768 			case VBLK:
3769 			case VCHR:
3770 				if (vp->v_rdev == NULL) {
3771 					vrele(vp);
3772 					continue;
3773 				}
3774 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3775 				break;
3776 			case VSOCK:
3777 				xvn[n].xv_socket = vp->v_socket;
3778 				break;
3779 			case VFIFO:
3780 				xvn[n].xv_fifo = vp->v_fifoinfo;
3781 				break;
3782 			case VNON:
3783 			case VBAD:
3784 			default:
3785 				/* shouldn't happen? */
3786 				vrele(vp);
3787 				continue;
3788 			}
3789 			vrele(vp);
3790 			++n;
3791 		}
3792 		MNT_IUNLOCK(mp);
3793 		mtx_lock(&mountlist_mtx);
3794 		vfs_unbusy(mp);
3795 		if (n == len)
3796 			break;
3797 	}
3798 	mtx_unlock(&mountlist_mtx);
3799 
3800 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3801 	free(xvn, M_TEMP);
3802 	return (error);
3803 }
3804 
3805 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3806     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3807     "");
3808 #endif
3809 
3810 static void
3811 unmount_or_warn(struct mount *mp)
3812 {
3813 	int error;
3814 
3815 	error = dounmount(mp, MNT_FORCE, curthread);
3816 	if (error != 0) {
3817 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3818 		if (error == EBUSY)
3819 			printf("BUSY)\n");
3820 		else
3821 			printf("%d)\n", error);
3822 	}
3823 }
3824 
3825 /*
3826  * Unmount all filesystems. The list is traversed in reverse order
3827  * of mounting to avoid dependencies.
3828  */
3829 void
3830 vfs_unmountall(void)
3831 {
3832 	struct mount *mp, *tmp;
3833 
3834 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3835 
3836 	/*
3837 	 * Since this only runs when rebooting, it is not interlocked.
3838 	 */
3839 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3840 		vfs_ref(mp);
3841 
3842 		/*
3843 		 * Forcibly unmounting "/dev" before "/" would prevent clean
3844 		 * unmount of the latter.
3845 		 */
3846 		if (mp == rootdevmp)
3847 			continue;
3848 
3849 		unmount_or_warn(mp);
3850 	}
3851 
3852 	if (rootdevmp != NULL)
3853 		unmount_or_warn(rootdevmp);
3854 }
3855 
3856 /*
3857  * perform msync on all vnodes under a mount point
3858  * the mount point must be locked.
3859  */
3860 void
3861 vfs_msync(struct mount *mp, int flags)
3862 {
3863 	struct vnode *vp, *mvp;
3864 	struct vm_object *obj;
3865 
3866 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3867 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3868 		obj = vp->v_object;
3869 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3870 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3871 			if (!vget(vp,
3872 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3873 			    curthread)) {
3874 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3875 					vput(vp);
3876 					continue;
3877 				}
3878 
3879 				obj = vp->v_object;
3880 				if (obj != NULL) {
3881 					VM_OBJECT_WLOCK(obj);
3882 					vm_object_page_clean(obj, 0, 0,
3883 					    flags == MNT_WAIT ?
3884 					    OBJPC_SYNC : OBJPC_NOSYNC);
3885 					VM_OBJECT_WUNLOCK(obj);
3886 				}
3887 				vput(vp);
3888 			}
3889 		} else
3890 			VI_UNLOCK(vp);
3891 	}
3892 }
3893 
3894 static void
3895 destroy_vpollinfo_free(struct vpollinfo *vi)
3896 {
3897 
3898 	knlist_destroy(&vi->vpi_selinfo.si_note);
3899 	mtx_destroy(&vi->vpi_lock);
3900 	uma_zfree(vnodepoll_zone, vi);
3901 }
3902 
3903 static void
3904 destroy_vpollinfo(struct vpollinfo *vi)
3905 {
3906 
3907 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
3908 	seldrain(&vi->vpi_selinfo);
3909 	destroy_vpollinfo_free(vi);
3910 }
3911 
3912 /*
3913  * Initalize per-vnode helper structure to hold poll-related state.
3914  */
3915 void
3916 v_addpollinfo(struct vnode *vp)
3917 {
3918 	struct vpollinfo *vi;
3919 
3920 	if (vp->v_pollinfo != NULL)
3921 		return;
3922 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
3923 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3924 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3925 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3926 	VI_LOCK(vp);
3927 	if (vp->v_pollinfo != NULL) {
3928 		VI_UNLOCK(vp);
3929 		destroy_vpollinfo_free(vi);
3930 		return;
3931 	}
3932 	vp->v_pollinfo = vi;
3933 	VI_UNLOCK(vp);
3934 }
3935 
3936 /*
3937  * Record a process's interest in events which might happen to
3938  * a vnode.  Because poll uses the historic select-style interface
3939  * internally, this routine serves as both the ``check for any
3940  * pending events'' and the ``record my interest in future events''
3941  * functions.  (These are done together, while the lock is held,
3942  * to avoid race conditions.)
3943  */
3944 int
3945 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3946 {
3947 
3948 	v_addpollinfo(vp);
3949 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3950 	if (vp->v_pollinfo->vpi_revents & events) {
3951 		/*
3952 		 * This leaves events we are not interested
3953 		 * in available for the other process which
3954 		 * which presumably had requested them
3955 		 * (otherwise they would never have been
3956 		 * recorded).
3957 		 */
3958 		events &= vp->v_pollinfo->vpi_revents;
3959 		vp->v_pollinfo->vpi_revents &= ~events;
3960 
3961 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3962 		return (events);
3963 	}
3964 	vp->v_pollinfo->vpi_events |= events;
3965 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3966 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3967 	return (0);
3968 }
3969 
3970 /*
3971  * Routine to create and manage a filesystem syncer vnode.
3972  */
3973 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3974 static int	sync_fsync(struct  vop_fsync_args *);
3975 static int	sync_inactive(struct  vop_inactive_args *);
3976 static int	sync_reclaim(struct  vop_reclaim_args *);
3977 
3978 static struct vop_vector sync_vnodeops = {
3979 	.vop_bypass =	VOP_EOPNOTSUPP,
3980 	.vop_close =	sync_close,		/* close */
3981 	.vop_fsync =	sync_fsync,		/* fsync */
3982 	.vop_inactive =	sync_inactive,	/* inactive */
3983 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3984 	.vop_lock1 =	vop_stdlock,	/* lock */
3985 	.vop_unlock =	vop_stdunlock,	/* unlock */
3986 	.vop_islocked =	vop_stdislocked,	/* islocked */
3987 };
3988 
3989 /*
3990  * Create a new filesystem syncer vnode for the specified mount point.
3991  */
3992 void
3993 vfs_allocate_syncvnode(struct mount *mp)
3994 {
3995 	struct vnode *vp;
3996 	struct bufobj *bo;
3997 	static long start, incr, next;
3998 	int error;
3999 
4000 	/* Allocate a new vnode */
4001 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4002 	if (error != 0)
4003 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4004 	vp->v_type = VNON;
4005 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4006 	vp->v_vflag |= VV_FORCEINSMQ;
4007 	error = insmntque(vp, mp);
4008 	if (error != 0)
4009 		panic("vfs_allocate_syncvnode: insmntque() failed");
4010 	vp->v_vflag &= ~VV_FORCEINSMQ;
4011 	VOP_UNLOCK(vp, 0);
4012 	/*
4013 	 * Place the vnode onto the syncer worklist. We attempt to
4014 	 * scatter them about on the list so that they will go off
4015 	 * at evenly distributed times even if all the filesystems
4016 	 * are mounted at once.
4017 	 */
4018 	next += incr;
4019 	if (next == 0 || next > syncer_maxdelay) {
4020 		start /= 2;
4021 		incr /= 2;
4022 		if (start == 0) {
4023 			start = syncer_maxdelay / 2;
4024 			incr = syncer_maxdelay;
4025 		}
4026 		next = start;
4027 	}
4028 	bo = &vp->v_bufobj;
4029 	BO_LOCK(bo);
4030 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4031 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4032 	mtx_lock(&sync_mtx);
4033 	sync_vnode_count++;
4034 	if (mp->mnt_syncer == NULL) {
4035 		mp->mnt_syncer = vp;
4036 		vp = NULL;
4037 	}
4038 	mtx_unlock(&sync_mtx);
4039 	BO_UNLOCK(bo);
4040 	if (vp != NULL) {
4041 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4042 		vgone(vp);
4043 		vput(vp);
4044 	}
4045 }
4046 
4047 void
4048 vfs_deallocate_syncvnode(struct mount *mp)
4049 {
4050 	struct vnode *vp;
4051 
4052 	mtx_lock(&sync_mtx);
4053 	vp = mp->mnt_syncer;
4054 	if (vp != NULL)
4055 		mp->mnt_syncer = NULL;
4056 	mtx_unlock(&sync_mtx);
4057 	if (vp != NULL)
4058 		vrele(vp);
4059 }
4060 
4061 /*
4062  * Do a lazy sync of the filesystem.
4063  */
4064 static int
4065 sync_fsync(struct vop_fsync_args *ap)
4066 {
4067 	struct vnode *syncvp = ap->a_vp;
4068 	struct mount *mp = syncvp->v_mount;
4069 	int error, save;
4070 	struct bufobj *bo;
4071 
4072 	/*
4073 	 * We only need to do something if this is a lazy evaluation.
4074 	 */
4075 	if (ap->a_waitfor != MNT_LAZY)
4076 		return (0);
4077 
4078 	/*
4079 	 * Move ourselves to the back of the sync list.
4080 	 */
4081 	bo = &syncvp->v_bufobj;
4082 	BO_LOCK(bo);
4083 	vn_syncer_add_to_worklist(bo, syncdelay);
4084 	BO_UNLOCK(bo);
4085 
4086 	/*
4087 	 * Walk the list of vnodes pushing all that are dirty and
4088 	 * not already on the sync list.
4089 	 */
4090 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4091 		return (0);
4092 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4093 		vfs_unbusy(mp);
4094 		return (0);
4095 	}
4096 	save = curthread_pflags_set(TDP_SYNCIO);
4097 	vfs_msync(mp, MNT_NOWAIT);
4098 	error = VFS_SYNC(mp, MNT_LAZY);
4099 	curthread_pflags_restore(save);
4100 	vn_finished_write(mp);
4101 	vfs_unbusy(mp);
4102 	return (error);
4103 }
4104 
4105 /*
4106  * The syncer vnode is no referenced.
4107  */
4108 static int
4109 sync_inactive(struct vop_inactive_args *ap)
4110 {
4111 
4112 	vgone(ap->a_vp);
4113 	return (0);
4114 }
4115 
4116 /*
4117  * The syncer vnode is no longer needed and is being decommissioned.
4118  *
4119  * Modifications to the worklist must be protected by sync_mtx.
4120  */
4121 static int
4122 sync_reclaim(struct vop_reclaim_args *ap)
4123 {
4124 	struct vnode *vp = ap->a_vp;
4125 	struct bufobj *bo;
4126 
4127 	bo = &vp->v_bufobj;
4128 	BO_LOCK(bo);
4129 	mtx_lock(&sync_mtx);
4130 	if (vp->v_mount->mnt_syncer == vp)
4131 		vp->v_mount->mnt_syncer = NULL;
4132 	if (bo->bo_flag & BO_ONWORKLST) {
4133 		LIST_REMOVE(bo, bo_synclist);
4134 		syncer_worklist_len--;
4135 		sync_vnode_count--;
4136 		bo->bo_flag &= ~BO_ONWORKLST;
4137 	}
4138 	mtx_unlock(&sync_mtx);
4139 	BO_UNLOCK(bo);
4140 
4141 	return (0);
4142 }
4143 
4144 /*
4145  * Check if vnode represents a disk device
4146  */
4147 int
4148 vn_isdisk(struct vnode *vp, int *errp)
4149 {
4150 	int error;
4151 
4152 	if (vp->v_type != VCHR) {
4153 		error = ENOTBLK;
4154 		goto out;
4155 	}
4156 	error = 0;
4157 	dev_lock();
4158 	if (vp->v_rdev == NULL)
4159 		error = ENXIO;
4160 	else if (vp->v_rdev->si_devsw == NULL)
4161 		error = ENXIO;
4162 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4163 		error = ENOTBLK;
4164 	dev_unlock();
4165 out:
4166 	if (errp != NULL)
4167 		*errp = error;
4168 	return (error == 0);
4169 }
4170 
4171 /*
4172  * Common filesystem object access control check routine.  Accepts a
4173  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4174  * and optional call-by-reference privused argument allowing vaccess()
4175  * to indicate to the caller whether privilege was used to satisfy the
4176  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4177  */
4178 int
4179 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4180     accmode_t accmode, struct ucred *cred, int *privused)
4181 {
4182 	accmode_t dac_granted;
4183 	accmode_t priv_granted;
4184 
4185 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4186 	    ("invalid bit in accmode"));
4187 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4188 	    ("VAPPEND without VWRITE"));
4189 
4190 	/*
4191 	 * Look for a normal, non-privileged way to access the file/directory
4192 	 * as requested.  If it exists, go with that.
4193 	 */
4194 
4195 	if (privused != NULL)
4196 		*privused = 0;
4197 
4198 	dac_granted = 0;
4199 
4200 	/* Check the owner. */
4201 	if (cred->cr_uid == file_uid) {
4202 		dac_granted |= VADMIN;
4203 		if (file_mode & S_IXUSR)
4204 			dac_granted |= VEXEC;
4205 		if (file_mode & S_IRUSR)
4206 			dac_granted |= VREAD;
4207 		if (file_mode & S_IWUSR)
4208 			dac_granted |= (VWRITE | VAPPEND);
4209 
4210 		if ((accmode & dac_granted) == accmode)
4211 			return (0);
4212 
4213 		goto privcheck;
4214 	}
4215 
4216 	/* Otherwise, check the groups (first match) */
4217 	if (groupmember(file_gid, cred)) {
4218 		if (file_mode & S_IXGRP)
4219 			dac_granted |= VEXEC;
4220 		if (file_mode & S_IRGRP)
4221 			dac_granted |= VREAD;
4222 		if (file_mode & S_IWGRP)
4223 			dac_granted |= (VWRITE | VAPPEND);
4224 
4225 		if ((accmode & dac_granted) == accmode)
4226 			return (0);
4227 
4228 		goto privcheck;
4229 	}
4230 
4231 	/* Otherwise, check everyone else. */
4232 	if (file_mode & S_IXOTH)
4233 		dac_granted |= VEXEC;
4234 	if (file_mode & S_IROTH)
4235 		dac_granted |= VREAD;
4236 	if (file_mode & S_IWOTH)
4237 		dac_granted |= (VWRITE | VAPPEND);
4238 	if ((accmode & dac_granted) == accmode)
4239 		return (0);
4240 
4241 privcheck:
4242 	/*
4243 	 * Build a privilege mask to determine if the set of privileges
4244 	 * satisfies the requirements when combined with the granted mask
4245 	 * from above.  For each privilege, if the privilege is required,
4246 	 * bitwise or the request type onto the priv_granted mask.
4247 	 */
4248 	priv_granted = 0;
4249 
4250 	if (type == VDIR) {
4251 		/*
4252 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4253 		 * requests, instead of PRIV_VFS_EXEC.
4254 		 */
4255 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4256 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4257 			priv_granted |= VEXEC;
4258 	} else {
4259 		/*
4260 		 * Ensure that at least one execute bit is on. Otherwise,
4261 		 * a privileged user will always succeed, and we don't want
4262 		 * this to happen unless the file really is executable.
4263 		 */
4264 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4265 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4266 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4267 			priv_granted |= VEXEC;
4268 	}
4269 
4270 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4271 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4272 		priv_granted |= VREAD;
4273 
4274 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4275 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4276 		priv_granted |= (VWRITE | VAPPEND);
4277 
4278 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4279 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4280 		priv_granted |= VADMIN;
4281 
4282 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4283 		/* XXX audit: privilege used */
4284 		if (privused != NULL)
4285 			*privused = 1;
4286 		return (0);
4287 	}
4288 
4289 	return ((accmode & VADMIN) ? EPERM : EACCES);
4290 }
4291 
4292 /*
4293  * Credential check based on process requesting service, and per-attribute
4294  * permissions.
4295  */
4296 int
4297 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4298     struct thread *td, accmode_t accmode)
4299 {
4300 
4301 	/*
4302 	 * Kernel-invoked always succeeds.
4303 	 */
4304 	if (cred == NOCRED)
4305 		return (0);
4306 
4307 	/*
4308 	 * Do not allow privileged processes in jail to directly manipulate
4309 	 * system attributes.
4310 	 */
4311 	switch (attrnamespace) {
4312 	case EXTATTR_NAMESPACE_SYSTEM:
4313 		/* Potentially should be: return (EPERM); */
4314 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4315 	case EXTATTR_NAMESPACE_USER:
4316 		return (VOP_ACCESS(vp, accmode, cred, td));
4317 	default:
4318 		return (EPERM);
4319 	}
4320 }
4321 
4322 #ifdef DEBUG_VFS_LOCKS
4323 /*
4324  * This only exists to supress warnings from unlocked specfs accesses.  It is
4325  * no longer ok to have an unlocked VFS.
4326  */
4327 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4328 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4329 
4330 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4331 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4332     "Drop into debugger on lock violation");
4333 
4334 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4335 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4336     0, "Check for interlock across VOPs");
4337 
4338 int vfs_badlock_print = 1;	/* Print lock violations. */
4339 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4340     0, "Print lock violations");
4341 
4342 #ifdef KDB
4343 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4344 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4345     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4346 #endif
4347 
4348 static void
4349 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4350 {
4351 
4352 #ifdef KDB
4353 	if (vfs_badlock_backtrace)
4354 		kdb_backtrace();
4355 #endif
4356 	if (vfs_badlock_print)
4357 		printf("%s: %p %s\n", str, (void *)vp, msg);
4358 	if (vfs_badlock_ddb)
4359 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4360 }
4361 
4362 void
4363 assert_vi_locked(struct vnode *vp, const char *str)
4364 {
4365 
4366 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4367 		vfs_badlock("interlock is not locked but should be", str, vp);
4368 }
4369 
4370 void
4371 assert_vi_unlocked(struct vnode *vp, const char *str)
4372 {
4373 
4374 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4375 		vfs_badlock("interlock is locked but should not be", str, vp);
4376 }
4377 
4378 void
4379 assert_vop_locked(struct vnode *vp, const char *str)
4380 {
4381 	int locked;
4382 
4383 	if (!IGNORE_LOCK(vp)) {
4384 		locked = VOP_ISLOCKED(vp);
4385 		if (locked == 0 || locked == LK_EXCLOTHER)
4386 			vfs_badlock("is not locked but should be", str, vp);
4387 	}
4388 }
4389 
4390 void
4391 assert_vop_unlocked(struct vnode *vp, const char *str)
4392 {
4393 
4394 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4395 		vfs_badlock("is locked but should not be", str, vp);
4396 }
4397 
4398 void
4399 assert_vop_elocked(struct vnode *vp, const char *str)
4400 {
4401 
4402 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4403 		vfs_badlock("is not exclusive locked but should be", str, vp);
4404 }
4405 
4406 #if 0
4407 void
4408 assert_vop_elocked_other(struct vnode *vp, const char *str)
4409 {
4410 
4411 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
4412 		vfs_badlock("is not exclusive locked by another thread",
4413 		    str, vp);
4414 }
4415 
4416 void
4417 assert_vop_slocked(struct vnode *vp, const char *str)
4418 {
4419 
4420 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
4421 		vfs_badlock("is not locked shared but should be", str, vp);
4422 }
4423 #endif /* 0 */
4424 #endif /* DEBUG_VFS_LOCKS */
4425 
4426 void
4427 vop_rename_fail(struct vop_rename_args *ap)
4428 {
4429 
4430 	if (ap->a_tvp != NULL)
4431 		vput(ap->a_tvp);
4432 	if (ap->a_tdvp == ap->a_tvp)
4433 		vrele(ap->a_tdvp);
4434 	else
4435 		vput(ap->a_tdvp);
4436 	vrele(ap->a_fdvp);
4437 	vrele(ap->a_fvp);
4438 }
4439 
4440 void
4441 vop_rename_pre(void *ap)
4442 {
4443 	struct vop_rename_args *a = ap;
4444 
4445 #ifdef DEBUG_VFS_LOCKS
4446 	if (a->a_tvp)
4447 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4448 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4449 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4450 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4451 
4452 	/* Check the source (from). */
4453 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4454 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4455 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4456 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4457 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4458 
4459 	/* Check the target. */
4460 	if (a->a_tvp)
4461 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4462 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4463 #endif
4464 	if (a->a_tdvp != a->a_fdvp)
4465 		vhold(a->a_fdvp);
4466 	if (a->a_tvp != a->a_fvp)
4467 		vhold(a->a_fvp);
4468 	vhold(a->a_tdvp);
4469 	if (a->a_tvp)
4470 		vhold(a->a_tvp);
4471 }
4472 
4473 void
4474 vop_strategy_pre(void *ap)
4475 {
4476 #ifdef DEBUG_VFS_LOCKS
4477 	struct vop_strategy_args *a;
4478 	struct buf *bp;
4479 
4480 	a = ap;
4481 	bp = a->a_bp;
4482 
4483 	/*
4484 	 * Cluster ops lock their component buffers but not the IO container.
4485 	 */
4486 	if ((bp->b_flags & B_CLUSTER) != 0)
4487 		return;
4488 
4489 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4490 		if (vfs_badlock_print)
4491 			printf(
4492 			    "VOP_STRATEGY: bp is not locked but should be\n");
4493 		if (vfs_badlock_ddb)
4494 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4495 	}
4496 #endif
4497 }
4498 
4499 void
4500 vop_lock_pre(void *ap)
4501 {
4502 #ifdef DEBUG_VFS_LOCKS
4503 	struct vop_lock1_args *a = ap;
4504 
4505 	if ((a->a_flags & LK_INTERLOCK) == 0)
4506 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4507 	else
4508 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4509 #endif
4510 }
4511 
4512 void
4513 vop_lock_post(void *ap, int rc)
4514 {
4515 #ifdef DEBUG_VFS_LOCKS
4516 	struct vop_lock1_args *a = ap;
4517 
4518 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4519 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4520 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4521 #endif
4522 }
4523 
4524 void
4525 vop_unlock_pre(void *ap)
4526 {
4527 #ifdef DEBUG_VFS_LOCKS
4528 	struct vop_unlock_args *a = ap;
4529 
4530 	if (a->a_flags & LK_INTERLOCK)
4531 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4532 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4533 #endif
4534 }
4535 
4536 void
4537 vop_unlock_post(void *ap, int rc)
4538 {
4539 #ifdef DEBUG_VFS_LOCKS
4540 	struct vop_unlock_args *a = ap;
4541 
4542 	if (a->a_flags & LK_INTERLOCK)
4543 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4544 #endif
4545 }
4546 
4547 void
4548 vop_create_post(void *ap, int rc)
4549 {
4550 	struct vop_create_args *a = ap;
4551 
4552 	if (!rc)
4553 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4554 }
4555 
4556 void
4557 vop_deleteextattr_post(void *ap, int rc)
4558 {
4559 	struct vop_deleteextattr_args *a = ap;
4560 
4561 	if (!rc)
4562 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4563 }
4564 
4565 void
4566 vop_link_post(void *ap, int rc)
4567 {
4568 	struct vop_link_args *a = ap;
4569 
4570 	if (!rc) {
4571 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4572 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4573 	}
4574 }
4575 
4576 void
4577 vop_mkdir_post(void *ap, int rc)
4578 {
4579 	struct vop_mkdir_args *a = ap;
4580 
4581 	if (!rc)
4582 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4583 }
4584 
4585 void
4586 vop_mknod_post(void *ap, int rc)
4587 {
4588 	struct vop_mknod_args *a = ap;
4589 
4590 	if (!rc)
4591 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4592 }
4593 
4594 void
4595 vop_reclaim_post(void *ap, int rc)
4596 {
4597 	struct vop_reclaim_args *a = ap;
4598 
4599 	if (!rc)
4600 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4601 }
4602 
4603 void
4604 vop_remove_post(void *ap, int rc)
4605 {
4606 	struct vop_remove_args *a = ap;
4607 
4608 	if (!rc) {
4609 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4610 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4611 	}
4612 }
4613 
4614 void
4615 vop_rename_post(void *ap, int rc)
4616 {
4617 	struct vop_rename_args *a = ap;
4618 
4619 	if (!rc) {
4620 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
4621 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
4622 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4623 		if (a->a_tvp)
4624 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4625 	}
4626 	if (a->a_tdvp != a->a_fdvp)
4627 		vdrop(a->a_fdvp);
4628 	if (a->a_tvp != a->a_fvp)
4629 		vdrop(a->a_fvp);
4630 	vdrop(a->a_tdvp);
4631 	if (a->a_tvp)
4632 		vdrop(a->a_tvp);
4633 }
4634 
4635 void
4636 vop_rmdir_post(void *ap, int rc)
4637 {
4638 	struct vop_rmdir_args *a = ap;
4639 
4640 	if (!rc) {
4641 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4642 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4643 	}
4644 }
4645 
4646 void
4647 vop_setattr_post(void *ap, int rc)
4648 {
4649 	struct vop_setattr_args *a = ap;
4650 
4651 	if (!rc)
4652 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4653 }
4654 
4655 void
4656 vop_setextattr_post(void *ap, int rc)
4657 {
4658 	struct vop_setextattr_args *a = ap;
4659 
4660 	if (!rc)
4661 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4662 }
4663 
4664 void
4665 vop_symlink_post(void *ap, int rc)
4666 {
4667 	struct vop_symlink_args *a = ap;
4668 
4669 	if (!rc)
4670 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4671 }
4672 
4673 static struct knlist fs_knlist;
4674 
4675 static void
4676 vfs_event_init(void *arg)
4677 {
4678 	knlist_init_mtx(&fs_knlist, NULL);
4679 }
4680 /* XXX - correct order? */
4681 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4682 
4683 void
4684 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4685 {
4686 
4687 	KNOTE_UNLOCKED(&fs_knlist, event);
4688 }
4689 
4690 static int	filt_fsattach(struct knote *kn);
4691 static void	filt_fsdetach(struct knote *kn);
4692 static int	filt_fsevent(struct knote *kn, long hint);
4693 
4694 struct filterops fs_filtops = {
4695 	.f_isfd = 0,
4696 	.f_attach = filt_fsattach,
4697 	.f_detach = filt_fsdetach,
4698 	.f_event = filt_fsevent
4699 };
4700 
4701 static int
4702 filt_fsattach(struct knote *kn)
4703 {
4704 
4705 	kn->kn_flags |= EV_CLEAR;
4706 	knlist_add(&fs_knlist, kn, 0);
4707 	return (0);
4708 }
4709 
4710 static void
4711 filt_fsdetach(struct knote *kn)
4712 {
4713 
4714 	knlist_remove(&fs_knlist, kn, 0);
4715 }
4716 
4717 static int
4718 filt_fsevent(struct knote *kn, long hint)
4719 {
4720 
4721 	kn->kn_fflags |= hint;
4722 	return (kn->kn_fflags != 0);
4723 }
4724 
4725 static int
4726 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4727 {
4728 	struct vfsidctl vc;
4729 	int error;
4730 	struct mount *mp;
4731 
4732 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4733 	if (error)
4734 		return (error);
4735 	if (vc.vc_vers != VFS_CTL_VERS1)
4736 		return (EINVAL);
4737 	mp = vfs_getvfs(&vc.vc_fsid);
4738 	if (mp == NULL)
4739 		return (ENOENT);
4740 	/* ensure that a specific sysctl goes to the right filesystem. */
4741 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4742 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4743 		vfs_rel(mp);
4744 		return (EINVAL);
4745 	}
4746 	VCTLTOREQ(&vc, req);
4747 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4748 	vfs_rel(mp);
4749 	return (error);
4750 }
4751 
4752 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4753     NULL, 0, sysctl_vfs_ctl, "",
4754     "Sysctl by fsid");
4755 
4756 /*
4757  * Function to initialize a va_filerev field sensibly.
4758  * XXX: Wouldn't a random number make a lot more sense ??
4759  */
4760 u_quad_t
4761 init_va_filerev(void)
4762 {
4763 	struct bintime bt;
4764 
4765 	getbinuptime(&bt);
4766 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4767 }
4768 
4769 static int	filt_vfsread(struct knote *kn, long hint);
4770 static int	filt_vfswrite(struct knote *kn, long hint);
4771 static int	filt_vfsvnode(struct knote *kn, long hint);
4772 static void	filt_vfsdetach(struct knote *kn);
4773 static struct filterops vfsread_filtops = {
4774 	.f_isfd = 1,
4775 	.f_detach = filt_vfsdetach,
4776 	.f_event = filt_vfsread
4777 };
4778 static struct filterops vfswrite_filtops = {
4779 	.f_isfd = 1,
4780 	.f_detach = filt_vfsdetach,
4781 	.f_event = filt_vfswrite
4782 };
4783 static struct filterops vfsvnode_filtops = {
4784 	.f_isfd = 1,
4785 	.f_detach = filt_vfsdetach,
4786 	.f_event = filt_vfsvnode
4787 };
4788 
4789 static void
4790 vfs_knllock(void *arg)
4791 {
4792 	struct vnode *vp = arg;
4793 
4794 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4795 }
4796 
4797 static void
4798 vfs_knlunlock(void *arg)
4799 {
4800 	struct vnode *vp = arg;
4801 
4802 	VOP_UNLOCK(vp, 0);
4803 }
4804 
4805 static void
4806 vfs_knl_assert_locked(void *arg)
4807 {
4808 #ifdef DEBUG_VFS_LOCKS
4809 	struct vnode *vp = arg;
4810 
4811 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4812 #endif
4813 }
4814 
4815 static void
4816 vfs_knl_assert_unlocked(void *arg)
4817 {
4818 #ifdef DEBUG_VFS_LOCKS
4819 	struct vnode *vp = arg;
4820 
4821 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4822 #endif
4823 }
4824 
4825 int
4826 vfs_kqfilter(struct vop_kqfilter_args *ap)
4827 {
4828 	struct vnode *vp = ap->a_vp;
4829 	struct knote *kn = ap->a_kn;
4830 	struct knlist *knl;
4831 
4832 	switch (kn->kn_filter) {
4833 	case EVFILT_READ:
4834 		kn->kn_fop = &vfsread_filtops;
4835 		break;
4836 	case EVFILT_WRITE:
4837 		kn->kn_fop = &vfswrite_filtops;
4838 		break;
4839 	case EVFILT_VNODE:
4840 		kn->kn_fop = &vfsvnode_filtops;
4841 		break;
4842 	default:
4843 		return (EINVAL);
4844 	}
4845 
4846 	kn->kn_hook = (caddr_t)vp;
4847 
4848 	v_addpollinfo(vp);
4849 	if (vp->v_pollinfo == NULL)
4850 		return (ENOMEM);
4851 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4852 	vhold(vp);
4853 	knlist_add(knl, kn, 0);
4854 
4855 	return (0);
4856 }
4857 
4858 /*
4859  * Detach knote from vnode
4860  */
4861 static void
4862 filt_vfsdetach(struct knote *kn)
4863 {
4864 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4865 
4866 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4867 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4868 	vdrop(vp);
4869 }
4870 
4871 /*ARGSUSED*/
4872 static int
4873 filt_vfsread(struct knote *kn, long hint)
4874 {
4875 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4876 	struct vattr va;
4877 	int res;
4878 
4879 	/*
4880 	 * filesystem is gone, so set the EOF flag and schedule
4881 	 * the knote for deletion.
4882 	 */
4883 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4884 		VI_LOCK(vp);
4885 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4886 		VI_UNLOCK(vp);
4887 		return (1);
4888 	}
4889 
4890 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4891 		return (0);
4892 
4893 	VI_LOCK(vp);
4894 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4895 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
4896 	VI_UNLOCK(vp);
4897 	return (res);
4898 }
4899 
4900 /*ARGSUSED*/
4901 static int
4902 filt_vfswrite(struct knote *kn, long hint)
4903 {
4904 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4905 
4906 	VI_LOCK(vp);
4907 
4908 	/*
4909 	 * filesystem is gone, so set the EOF flag and schedule
4910 	 * the knote for deletion.
4911 	 */
4912 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
4913 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4914 
4915 	kn->kn_data = 0;
4916 	VI_UNLOCK(vp);
4917 	return (1);
4918 }
4919 
4920 static int
4921 filt_vfsvnode(struct knote *kn, long hint)
4922 {
4923 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4924 	int res;
4925 
4926 	VI_LOCK(vp);
4927 	if (kn->kn_sfflags & hint)
4928 		kn->kn_fflags |= hint;
4929 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4930 		kn->kn_flags |= EV_EOF;
4931 		VI_UNLOCK(vp);
4932 		return (1);
4933 	}
4934 	res = (kn->kn_fflags != 0);
4935 	VI_UNLOCK(vp);
4936 	return (res);
4937 }
4938 
4939 int
4940 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
4941 {
4942 	int error;
4943 
4944 	if (dp->d_reclen > ap->a_uio->uio_resid)
4945 		return (ENAMETOOLONG);
4946 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
4947 	if (error) {
4948 		if (ap->a_ncookies != NULL) {
4949 			if (ap->a_cookies != NULL)
4950 				free(ap->a_cookies, M_TEMP);
4951 			ap->a_cookies = NULL;
4952 			*ap->a_ncookies = 0;
4953 		}
4954 		return (error);
4955 	}
4956 	if (ap->a_ncookies == NULL)
4957 		return (0);
4958 
4959 	KASSERT(ap->a_cookies,
4960 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
4961 
4962 	*ap->a_cookies = realloc(*ap->a_cookies,
4963 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
4964 	(*ap->a_cookies)[*ap->a_ncookies] = off;
4965 	return (0);
4966 }
4967 
4968 /*
4969  * Mark for update the access time of the file if the filesystem
4970  * supports VOP_MARKATIME.  This functionality is used by execve and
4971  * mmap, so we want to avoid the I/O implied by directly setting
4972  * va_atime for the sake of efficiency.
4973  */
4974 void
4975 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
4976 {
4977 	struct mount *mp;
4978 
4979 	mp = vp->v_mount;
4980 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
4981 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
4982 		(void)VOP_MARKATIME(vp);
4983 }
4984 
4985 /*
4986  * The purpose of this routine is to remove granularity from accmode_t,
4987  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
4988  * VADMIN and VAPPEND.
4989  *
4990  * If it returns 0, the caller is supposed to continue with the usual
4991  * access checks using 'accmode' as modified by this routine.  If it
4992  * returns nonzero value, the caller is supposed to return that value
4993  * as errno.
4994  *
4995  * Note that after this routine runs, accmode may be zero.
4996  */
4997 int
4998 vfs_unixify_accmode(accmode_t *accmode)
4999 {
5000 	/*
5001 	 * There is no way to specify explicit "deny" rule using
5002 	 * file mode or POSIX.1e ACLs.
5003 	 */
5004 	if (*accmode & VEXPLICIT_DENY) {
5005 		*accmode = 0;
5006 		return (0);
5007 	}
5008 
5009 	/*
5010 	 * None of these can be translated into usual access bits.
5011 	 * Also, the common case for NFSv4 ACLs is to not contain
5012 	 * either of these bits. Caller should check for VWRITE
5013 	 * on the containing directory instead.
5014 	 */
5015 	if (*accmode & (VDELETE_CHILD | VDELETE))
5016 		return (EPERM);
5017 
5018 	if (*accmode & VADMIN_PERMS) {
5019 		*accmode &= ~VADMIN_PERMS;
5020 		*accmode |= VADMIN;
5021 	}
5022 
5023 	/*
5024 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5025 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5026 	 */
5027 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5028 
5029 	return (0);
5030 }
5031 
5032 /*
5033  * These are helper functions for filesystems to traverse all
5034  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5035  *
5036  * This interface replaces MNT_VNODE_FOREACH.
5037  */
5038 
5039 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5040 
5041 struct vnode *
5042 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5043 {
5044 	struct vnode *vp;
5045 
5046 	if (should_yield())
5047 		kern_yield(PRI_USER);
5048 	MNT_ILOCK(mp);
5049 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5050 	vp = TAILQ_NEXT(*mvp, v_nmntvnodes);
5051 	while (vp != NULL && (vp->v_type == VMARKER ||
5052 	    (vp->v_iflag & VI_DOOMED) != 0))
5053 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5054 
5055 	/* Check if we are done */
5056 	if (vp == NULL) {
5057 		__mnt_vnode_markerfree_all(mvp, mp);
5058 		/* MNT_IUNLOCK(mp); -- done in above function */
5059 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5060 		return (NULL);
5061 	}
5062 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5063 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5064 	VI_LOCK(vp);
5065 	MNT_IUNLOCK(mp);
5066 	return (vp);
5067 }
5068 
5069 struct vnode *
5070 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5071 {
5072 	struct vnode *vp;
5073 
5074 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5075 	MNT_ILOCK(mp);
5076 	MNT_REF(mp);
5077 	(*mvp)->v_type = VMARKER;
5078 
5079 	vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
5080 	while (vp != NULL && (vp->v_type == VMARKER ||
5081 	    (vp->v_iflag & VI_DOOMED) != 0))
5082 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5083 
5084 	/* Check if we are done */
5085 	if (vp == NULL) {
5086 		MNT_REL(mp);
5087 		MNT_IUNLOCK(mp);
5088 		free(*mvp, M_VNODE_MARKER);
5089 		*mvp = NULL;
5090 		return (NULL);
5091 	}
5092 	(*mvp)->v_mount = mp;
5093 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5094 	VI_LOCK(vp);
5095 	MNT_IUNLOCK(mp);
5096 	return (vp);
5097 }
5098 
5099 
5100 void
5101 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5102 {
5103 
5104 	if (*mvp == NULL) {
5105 		MNT_IUNLOCK(mp);
5106 		return;
5107 	}
5108 
5109 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5110 
5111 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5112 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5113 	MNT_REL(mp);
5114 	MNT_IUNLOCK(mp);
5115 	free(*mvp, M_VNODE_MARKER);
5116 	*mvp = NULL;
5117 }
5118 
5119 /*
5120  * These are helper functions for filesystems to traverse their
5121  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5122  */
5123 static void
5124 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5125 {
5126 
5127 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5128 
5129 	MNT_ILOCK(mp);
5130 	MNT_REL(mp);
5131 	MNT_IUNLOCK(mp);
5132 	free(*mvp, M_VNODE_MARKER);
5133 	*mvp = NULL;
5134 }
5135 
5136 static struct vnode *
5137 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5138 {
5139 	struct vnode *vp, *nvp;
5140 
5141 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5142 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5143 restart:
5144 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5145 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5146 	while (vp != NULL) {
5147 		if (vp->v_type == VMARKER) {
5148 			vp = TAILQ_NEXT(vp, v_actfreelist);
5149 			continue;
5150 		}
5151 		if (!VI_TRYLOCK(vp)) {
5152 			if (mp_ncpus == 1 || should_yield()) {
5153 				TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5154 				mtx_unlock(&vnode_free_list_mtx);
5155 				pause("vnacti", 1);
5156 				mtx_lock(&vnode_free_list_mtx);
5157 				goto restart;
5158 			}
5159 			continue;
5160 		}
5161 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5162 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5163 		    ("alien vnode on the active list %p %p", vp, mp));
5164 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5165 			break;
5166 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5167 		VI_UNLOCK(vp);
5168 		vp = nvp;
5169 	}
5170 
5171 	/* Check if we are done */
5172 	if (vp == NULL) {
5173 		mtx_unlock(&vnode_free_list_mtx);
5174 		mnt_vnode_markerfree_active(mvp, mp);
5175 		return (NULL);
5176 	}
5177 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5178 	mtx_unlock(&vnode_free_list_mtx);
5179 	ASSERT_VI_LOCKED(vp, "active iter");
5180 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5181 	return (vp);
5182 }
5183 
5184 struct vnode *
5185 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5186 {
5187 
5188 	if (should_yield())
5189 		kern_yield(PRI_USER);
5190 	mtx_lock(&vnode_free_list_mtx);
5191 	return (mnt_vnode_next_active(mvp, mp));
5192 }
5193 
5194 struct vnode *
5195 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5196 {
5197 	struct vnode *vp;
5198 
5199 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5200 	MNT_ILOCK(mp);
5201 	MNT_REF(mp);
5202 	MNT_IUNLOCK(mp);
5203 	(*mvp)->v_type = VMARKER;
5204 	(*mvp)->v_mount = mp;
5205 
5206 	mtx_lock(&vnode_free_list_mtx);
5207 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5208 	if (vp == NULL) {
5209 		mtx_unlock(&vnode_free_list_mtx);
5210 		mnt_vnode_markerfree_active(mvp, mp);
5211 		return (NULL);
5212 	}
5213 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5214 	return (mnt_vnode_next_active(mvp, mp));
5215 }
5216 
5217 void
5218 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5219 {
5220 
5221 	if (*mvp == NULL)
5222 		return;
5223 
5224 	mtx_lock(&vnode_free_list_mtx);
5225 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5226 	mtx_unlock(&vnode_free_list_mtx);
5227 	mnt_vnode_markerfree_active(mvp, mp);
5228 }
5229