xref: /freebsd/sys/kern/vfs_subr.c (revision 6b3455a7665208c366849f0b2b3bc916fb97516e)
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_ddb.h"
45 #include "opt_mac.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/conf.h>
52 #include <sys/event.h>
53 #include <sys/eventhandler.h>
54 #include <sys/extattr.h>
55 #include <sys/fcntl.h>
56 #include <sys/kdb.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/mac.h>
60 #include <sys/malloc.h>
61 #include <sys/mount.h>
62 #include <sys/namei.h>
63 #include <sys/sleepqueue.h>
64 #include <sys/stat.h>
65 #include <sys/sysctl.h>
66 #include <sys/syslog.h>
67 #include <sys/vmmeter.h>
68 #include <sys/vnode.h>
69 
70 #include <vm/vm.h>
71 #include <vm/vm_object.h>
72 #include <vm/vm_extern.h>
73 #include <vm/pmap.h>
74 #include <vm/vm_map.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_kern.h>
77 #include <vm/uma.h>
78 
79 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
80 
81 static void	addalias(struct vnode *vp, struct cdev *nvp_rdev);
82 static void	delmntque(struct vnode *vp);
83 static void	insmntque(struct vnode *vp, struct mount *mp);
84 static void	vclean(struct vnode *vp, int flags, struct thread *td);
85 static void	vlruvp(struct vnode *vp);
86 static int	flushbuflist(struct buf *blist, int flags, struct vnode *vp,
87 		    int slpflag, int slptimeo, int *errorp);
88 static void	syncer_shutdown(void *arg, int howto);
89 static int	vtryrecycle(struct vnode *vp);
90 static void	vx_lock(struct vnode *vp);
91 static void	vx_unlock(struct vnode *vp);
92 static void	vgonechrl(struct vnode *vp, struct thread *td);
93 
94 
95 /*
96  * Number of vnodes in existence.  Increased whenever getnewvnode()
97  * allocates a new vnode, never decreased.
98  */
99 static unsigned long	numvnodes;
100 
101 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
102 
103 /*
104  * Conversion tables for conversion from vnode types to inode formats
105  * and back.
106  */
107 enum vtype iftovt_tab[16] = {
108 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
109 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
110 };
111 int vttoif_tab[9] = {
112 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
113 	S_IFSOCK, S_IFIFO, S_IFMT,
114 };
115 
116 /*
117  * List of vnodes that are ready for recycling.
118  */
119 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
120 
121 /*
122  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
123  * getnewvnode() will return a newly allocated vnode.
124  */
125 static u_long wantfreevnodes = 25;
126 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
127 /* Number of vnodes in the free list. */
128 static u_long freevnodes;
129 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
130 
131 /*
132  * Various variables used for debugging the new implementation of
133  * reassignbuf().
134  * XXX these are probably of (very) limited utility now.
135  */
136 static int reassignbufcalls;
137 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
138 static int nameileafonly;
139 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
140 
141 /*
142  * Cache for the mount type id assigned to NFS.  This is used for
143  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
144  */
145 int	nfs_mount_type = -1;
146 
147 /* To keep more than one thread at a time from running vfs_getnewfsid */
148 static struct mtx mntid_mtx;
149 
150 /*
151  * Lock for any access to the following:
152  *	vnode_free_list
153  *	numvnodes
154  *	freevnodes
155  */
156 static struct mtx vnode_free_list_mtx;
157 
158 /*
159  * For any iteration/modification of dev->si_hlist (linked through
160  * v_specnext)
161  */
162 static struct mtx spechash_mtx;
163 
164 /* Publicly exported FS */
165 struct nfs_public nfs_pub;
166 
167 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
168 static uma_zone_t vnode_zone;
169 static uma_zone_t vnodepoll_zone;
170 
171 /* Set to 1 to print out reclaim of active vnodes */
172 int	prtactive;
173 
174 /*
175  * The workitem queue.
176  *
177  * It is useful to delay writes of file data and filesystem metadata
178  * for tens of seconds so that quickly created and deleted files need
179  * not waste disk bandwidth being created and removed. To realize this,
180  * we append vnodes to a "workitem" queue. When running with a soft
181  * updates implementation, most pending metadata dependencies should
182  * not wait for more than a few seconds. Thus, mounted on block devices
183  * are delayed only about a half the time that file data is delayed.
184  * Similarly, directory updates are more critical, so are only delayed
185  * about a third the time that file data is delayed. Thus, there are
186  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
187  * one each second (driven off the filesystem syncer process). The
188  * syncer_delayno variable indicates the next queue that is to be processed.
189  * Items that need to be processed soon are placed in this queue:
190  *
191  *	syncer_workitem_pending[syncer_delayno]
192  *
193  * A delay of fifteen seconds is done by placing the request fifteen
194  * entries later in the queue:
195  *
196  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
197  *
198  */
199 static int syncer_delayno;
200 static long syncer_mask;
201 LIST_HEAD(synclist, vnode);
202 static struct synclist *syncer_workitem_pending;
203 /*
204  * The sync_mtx protects:
205  *	vp->v_synclist
206  *	sync_vnode_count
207  *	syncer_delayno
208  *	syncer_state
209  *	syncer_workitem_pending
210  *	syncer_worklist_len
211  *	rushjob
212  */
213 static struct mtx sync_mtx;
214 
215 #define SYNCER_MAXDELAY		32
216 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
217 static int syncdelay = 30;		/* max time to delay syncing data */
218 static int filedelay = 30;		/* time to delay syncing files */
219 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
220 static int dirdelay = 29;		/* time to delay syncing directories */
221 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
222 static int metadelay = 28;		/* time to delay syncing metadata */
223 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
224 static int rushjob;		/* number of slots to run ASAP */
225 static int stat_rush_requests;	/* number of times I/O speeded up */
226 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
227 
228 /*
229  * When shutting down the syncer, run it at four times normal speed.
230  */
231 #define SYNCER_SHUTDOWN_SPEEDUP		4
232 static int sync_vnode_count;
233 static int syncer_worklist_len;
234 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
235     syncer_state;
236 
237 /*
238  * Number of vnodes we want to exist at any one time.  This is mostly used
239  * to size hash tables in vnode-related code.  It is normally not used in
240  * getnewvnode(), as wantfreevnodes is normally nonzero.)
241  *
242  * XXX desiredvnodes is historical cruft and should not exist.
243  */
244 int desiredvnodes;
245 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
246     &desiredvnodes, 0, "Maximum number of vnodes");
247 static int minvnodes;
248 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
249     &minvnodes, 0, "Minimum number of vnodes");
250 static int vnlru_nowhere;
251 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
252     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
253 
254 /* Hook for calling soft updates. */
255 int (*softdep_process_worklist_hook)(struct mount *);
256 
257 /*
258  * Initialize the vnode management data structures.
259  */
260 static void
261 vntblinit(void *dummy __unused)
262 {
263 
264 	/*
265 	 * Desiredvnodes is a function of the physical memory size and
266 	 * the kernel's heap size.  Specifically, desiredvnodes scales
267 	 * in proportion to the physical memory size until two fifths
268 	 * of the kernel's heap size is consumed by vnodes and vm
269 	 * objects.
270 	 */
271 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
272 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
273 	minvnodes = desiredvnodes / 4;
274 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
275 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
276 	mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF);
277 	TAILQ_INIT(&vnode_free_list);
278 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
279 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
280 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
281 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
282 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
283 	/*
284 	 * Initialize the filesystem syncer.
285 	 */
286 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
287 		&syncer_mask);
288 	syncer_maxdelay = syncer_mask + 1;
289 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
290 }
291 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
292 
293 
294 /*
295  * Mark a mount point as busy. Used to synchronize access and to delay
296  * unmounting. Interlock is not released on failure.
297  */
298 int
299 vfs_busy(mp, flags, interlkp, td)
300 	struct mount *mp;
301 	int flags;
302 	struct mtx *interlkp;
303 	struct thread *td;
304 {
305 	int lkflags;
306 
307 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
308 		if (flags & LK_NOWAIT)
309 			return (ENOENT);
310 		mp->mnt_kern_flag |= MNTK_MWAIT;
311 		/*
312 		 * Since all busy locks are shared except the exclusive
313 		 * lock granted when unmounting, the only place that a
314 		 * wakeup needs to be done is at the release of the
315 		 * exclusive lock at the end of dounmount.
316 		 */
317 		msleep(mp, interlkp, PVFS, "vfs_busy", 0);
318 		return (ENOENT);
319 	}
320 	lkflags = LK_SHARED | LK_NOPAUSE;
321 	if (interlkp)
322 		lkflags |= LK_INTERLOCK;
323 	if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td))
324 		panic("vfs_busy: unexpected lock failure");
325 	return (0);
326 }
327 
328 /*
329  * Free a busy filesystem.
330  */
331 void
332 vfs_unbusy(mp, td)
333 	struct mount *mp;
334 	struct thread *td;
335 {
336 
337 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
338 }
339 
340 /*
341  * Lookup a mount point by filesystem identifier.
342  */
343 struct mount *
344 vfs_getvfs(fsid)
345 	fsid_t *fsid;
346 {
347 	register struct mount *mp;
348 
349 	mtx_lock(&mountlist_mtx);
350 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
351 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
352 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
353 			mtx_unlock(&mountlist_mtx);
354 			return (mp);
355 		}
356 	}
357 	mtx_unlock(&mountlist_mtx);
358 	return ((struct mount *) 0);
359 }
360 
361 /*
362  * Check if a user can access priveledged mount options.
363  */
364 int
365 vfs_suser(struct mount *mp, struct thread *td)
366 {
367 	int error;
368 
369 	if ((mp->mnt_flag & MNT_USER) == 0 ||
370 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
371 		if ((error = suser(td)) != 0)
372 			return (error);
373 	}
374 	return (0);
375 }
376 
377 /*
378  * Get a new unique fsid.  Try to make its val[0] unique, since this value
379  * will be used to create fake device numbers for stat().  Also try (but
380  * not so hard) make its val[0] unique mod 2^16, since some emulators only
381  * support 16-bit device numbers.  We end up with unique val[0]'s for the
382  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
383  *
384  * Keep in mind that several mounts may be running in parallel.  Starting
385  * the search one past where the previous search terminated is both a
386  * micro-optimization and a defense against returning the same fsid to
387  * different mounts.
388  */
389 void
390 vfs_getnewfsid(mp)
391 	struct mount *mp;
392 {
393 	static u_int16_t mntid_base;
394 	fsid_t tfsid;
395 	int mtype;
396 
397 	mtx_lock(&mntid_mtx);
398 	mtype = mp->mnt_vfc->vfc_typenum;
399 	tfsid.val[1] = mtype;
400 	mtype = (mtype & 0xFF) << 24;
401 	for (;;) {
402 		tfsid.val[0] = makedev(255,
403 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
404 		mntid_base++;
405 		if (vfs_getvfs(&tfsid) == NULL)
406 			break;
407 	}
408 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
409 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
410 	mtx_unlock(&mntid_mtx);
411 }
412 
413 /*
414  * Knob to control the precision of file timestamps:
415  *
416  *   0 = seconds only; nanoseconds zeroed.
417  *   1 = seconds and nanoseconds, accurate within 1/HZ.
418  *   2 = seconds and nanoseconds, truncated to microseconds.
419  * >=3 = seconds and nanoseconds, maximum precision.
420  */
421 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
422 
423 static int timestamp_precision = TSP_SEC;
424 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
425     &timestamp_precision, 0, "");
426 
427 /*
428  * Get a current timestamp.
429  */
430 void
431 vfs_timestamp(tsp)
432 	struct timespec *tsp;
433 {
434 	struct timeval tv;
435 
436 	switch (timestamp_precision) {
437 	case TSP_SEC:
438 		tsp->tv_sec = time_second;
439 		tsp->tv_nsec = 0;
440 		break;
441 	case TSP_HZ:
442 		getnanotime(tsp);
443 		break;
444 	case TSP_USEC:
445 		microtime(&tv);
446 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
447 		break;
448 	case TSP_NSEC:
449 	default:
450 		nanotime(tsp);
451 		break;
452 	}
453 }
454 
455 /*
456  * Set vnode attributes to VNOVAL
457  */
458 void
459 vattr_null(vap)
460 	register struct vattr *vap;
461 {
462 
463 	vap->va_type = VNON;
464 	vap->va_size = VNOVAL;
465 	vap->va_bytes = VNOVAL;
466 	vap->va_mode = VNOVAL;
467 	vap->va_nlink = VNOVAL;
468 	vap->va_uid = VNOVAL;
469 	vap->va_gid = VNOVAL;
470 	vap->va_fsid = VNOVAL;
471 	vap->va_fileid = VNOVAL;
472 	vap->va_blocksize = VNOVAL;
473 	vap->va_rdev = VNOVAL;
474 	vap->va_atime.tv_sec = VNOVAL;
475 	vap->va_atime.tv_nsec = VNOVAL;
476 	vap->va_mtime.tv_sec = VNOVAL;
477 	vap->va_mtime.tv_nsec = VNOVAL;
478 	vap->va_ctime.tv_sec = VNOVAL;
479 	vap->va_ctime.tv_nsec = VNOVAL;
480 	vap->va_birthtime.tv_sec = VNOVAL;
481 	vap->va_birthtime.tv_nsec = VNOVAL;
482 	vap->va_flags = VNOVAL;
483 	vap->va_gen = VNOVAL;
484 	vap->va_vaflags = 0;
485 }
486 
487 /*
488  * This routine is called when we have too many vnodes.  It attempts
489  * to free <count> vnodes and will potentially free vnodes that still
490  * have VM backing store (VM backing store is typically the cause
491  * of a vnode blowout so we want to do this).  Therefore, this operation
492  * is not considered cheap.
493  *
494  * A number of conditions may prevent a vnode from being reclaimed.
495  * the buffer cache may have references on the vnode, a directory
496  * vnode may still have references due to the namei cache representing
497  * underlying files, or the vnode may be in active use.   It is not
498  * desireable to reuse such vnodes.  These conditions may cause the
499  * number of vnodes to reach some minimum value regardless of what
500  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
501  */
502 static int
503 vlrureclaim(struct mount *mp)
504 {
505 	struct vnode *vp;
506 	int done;
507 	int trigger;
508 	int usevnodes;
509 	int count;
510 
511 	/*
512 	 * Calculate the trigger point, don't allow user
513 	 * screwups to blow us up.   This prevents us from
514 	 * recycling vnodes with lots of resident pages.  We
515 	 * aren't trying to free memory, we are trying to
516 	 * free vnodes.
517 	 */
518 	usevnodes = desiredvnodes;
519 	if (usevnodes <= 0)
520 		usevnodes = 1;
521 	trigger = cnt.v_page_count * 2 / usevnodes;
522 
523 	done = 0;
524 	MNT_ILOCK(mp);
525 	count = mp->mnt_nvnodelistsize / 10 + 1;
526 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
527 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
528 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
529 
530 		if (vp->v_type != VNON &&
531 		    vp->v_type != VBAD &&
532 		    VI_TRYLOCK(vp)) {
533 			if (VMIGHTFREE(vp) &&           /* critical path opt */
534 			    (vp->v_object == NULL ||
535 			    vp->v_object->resident_page_count < trigger)) {
536 				MNT_IUNLOCK(mp);
537 				vgonel(vp, curthread);
538 				done++;
539 				MNT_ILOCK(mp);
540 			} else
541 				VI_UNLOCK(vp);
542 		}
543 		--count;
544 	}
545 	MNT_IUNLOCK(mp);
546 	return done;
547 }
548 
549 /*
550  * Attempt to recycle vnodes in a context that is always safe to block.
551  * Calling vlrurecycle() from the bowels of filesystem code has some
552  * interesting deadlock problems.
553  */
554 static struct proc *vnlruproc;
555 static int vnlruproc_sig;
556 
557 static void
558 vnlru_proc(void)
559 {
560 	struct mount *mp, *nmp;
561 	int done;
562 	struct proc *p = vnlruproc;
563 	struct thread *td = FIRST_THREAD_IN_PROC(p);
564 
565 	mtx_lock(&Giant);
566 
567 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
568 	    SHUTDOWN_PRI_FIRST);
569 
570 	for (;;) {
571 		kthread_suspend_check(p);
572 		mtx_lock(&vnode_free_list_mtx);
573 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
574 			mtx_unlock(&vnode_free_list_mtx);
575 			vnlruproc_sig = 0;
576 			wakeup(&vnlruproc_sig);
577 			tsleep(vnlruproc, PVFS, "vlruwt", hz);
578 			continue;
579 		}
580 		mtx_unlock(&vnode_free_list_mtx);
581 		done = 0;
582 		mtx_lock(&mountlist_mtx);
583 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
584 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
585 				nmp = TAILQ_NEXT(mp, mnt_list);
586 				continue;
587 			}
588 			done += vlrureclaim(mp);
589 			mtx_lock(&mountlist_mtx);
590 			nmp = TAILQ_NEXT(mp, mnt_list);
591 			vfs_unbusy(mp, td);
592 		}
593 		mtx_unlock(&mountlist_mtx);
594 		if (done == 0) {
595 #if 0
596 			/* These messages are temporary debugging aids */
597 			if (vnlru_nowhere < 5)
598 				printf("vnlru process getting nowhere..\n");
599 			else if (vnlru_nowhere == 5)
600 				printf("vnlru process messages stopped.\n");
601 #endif
602 			vnlru_nowhere++;
603 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
604 		}
605 	}
606 }
607 
608 static struct kproc_desc vnlru_kp = {
609 	"vnlru",
610 	vnlru_proc,
611 	&vnlruproc
612 };
613 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
614 
615 
616 /*
617  * Routines having to do with the management of the vnode table.
618  */
619 
620 /*
621  * Check to see if a free vnode can be recycled. If it can,
622  * recycle it and return it with the vnode interlock held.
623  */
624 static int
625 vtryrecycle(struct vnode *vp)
626 {
627 	struct thread *td = curthread;
628 	vm_object_t object;
629 	struct mount *vnmp;
630 	int error;
631 
632 	/* Don't recycle if we can't get the interlock */
633 	if (!VI_TRYLOCK(vp))
634 		return (EWOULDBLOCK);
635 	/*
636 	 * This vnode may found and locked via some other list, if so we
637 	 * can't recycle it yet.
638 	 */
639 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
640 		return (EWOULDBLOCK);
641 	/*
642 	 * Don't recycle if its filesystem is being suspended.
643 	 */
644 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
645 		VOP_UNLOCK(vp, 0, td);
646 		return (EBUSY);
647 	}
648 
649 	/*
650 	 * Don't recycle if we still have cached pages.
651 	 */
652 	if (VOP_GETVOBJECT(vp, &object) == 0) {
653 		VM_OBJECT_LOCK(object);
654 		if (object->resident_page_count ||
655 		    object->ref_count) {
656 			VM_OBJECT_UNLOCK(object);
657 			error = EBUSY;
658 			goto done;
659 		}
660 		VM_OBJECT_UNLOCK(object);
661 	}
662 	if (LIST_FIRST(&vp->v_cache_src)) {
663 		/*
664 		 * note: nameileafonly sysctl is temporary,
665 		 * for debugging only, and will eventually be
666 		 * removed.
667 		 */
668 		if (nameileafonly > 0) {
669 			/*
670 			 * Do not reuse namei-cached directory
671 			 * vnodes that have cached
672 			 * subdirectories.
673 			 */
674 			if (cache_leaf_test(vp) < 0) {
675 				error = EISDIR;
676 				goto done;
677 			}
678 		} else if (nameileafonly < 0 ||
679 			    vmiodirenable == 0) {
680 			/*
681 			 * Do not reuse namei-cached directory
682 			 * vnodes if nameileafonly is -1 or
683 			 * if VMIO backing for directories is
684 			 * turned off (otherwise we reuse them
685 			 * too quickly).
686 			 */
687 			error = EBUSY;
688 			goto done;
689 		}
690 	}
691 	/*
692 	 * If we got this far, we need to acquire the interlock and see if
693 	 * anyone picked up this vnode from another list.  If not, we will
694 	 * mark it with XLOCK via vgonel() so that anyone who does find it
695 	 * will skip over it.
696 	 */
697 	VI_LOCK(vp);
698 	if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) {
699 		VI_UNLOCK(vp);
700 		error = EBUSY;
701 		goto done;
702 	}
703 	mtx_lock(&vnode_free_list_mtx);
704 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
705 	vp->v_iflag &= ~VI_FREE;
706 	mtx_unlock(&vnode_free_list_mtx);
707 	vp->v_iflag |= VI_DOOMED;
708 	if (vp->v_type != VBAD) {
709 		VOP_UNLOCK(vp, 0, td);
710 		vgonel(vp, td);
711 		VI_LOCK(vp);
712 	} else
713 		VOP_UNLOCK(vp, 0, td);
714 	vn_finished_write(vnmp);
715 	return (0);
716 done:
717 	VOP_UNLOCK(vp, 0, td);
718 	vn_finished_write(vnmp);
719 	return (error);
720 }
721 
722 /*
723  * Return the next vnode from the free list.
724  */
725 int
726 getnewvnode(tag, mp, vops, vpp)
727 	const char *tag;
728 	struct mount *mp;
729 	vop_t **vops;
730 	struct vnode **vpp;
731 {
732 	struct vnode *vp = NULL;
733 	struct vpollinfo *pollinfo = NULL;
734 
735 	mtx_lock(&vnode_free_list_mtx);
736 
737 	/*
738 	 * Try to reuse vnodes if we hit the max.  This situation only
739 	 * occurs in certain large-memory (2G+) situations.  We cannot
740 	 * attempt to directly reclaim vnodes due to nasty recursion
741 	 * problems.
742 	 */
743 	while (numvnodes - freevnodes > desiredvnodes) {
744 		if (vnlruproc_sig == 0) {
745 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
746 			wakeup(vnlruproc);
747 		}
748 		mtx_unlock(&vnode_free_list_mtx);
749 		tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz);
750 		mtx_lock(&vnode_free_list_mtx);
751 	}
752 
753 	/*
754 	 * Attempt to reuse a vnode already on the free list, allocating
755 	 * a new vnode if we can't find one or if we have not reached a
756 	 * good minimum for good LRU performance.
757 	 */
758 
759 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
760 		int error;
761 		int count;
762 
763 		for (count = 0; count < freevnodes; count++) {
764 			vp = TAILQ_FIRST(&vnode_free_list);
765 
766 			KASSERT(vp->v_usecount == 0 &&
767 			    (vp->v_iflag & VI_DOINGINACT) == 0,
768 			    ("getnewvnode: free vnode isn't"));
769 
770 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
771 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
772 			mtx_unlock(&vnode_free_list_mtx);
773 			error = vtryrecycle(vp);
774 			mtx_lock(&vnode_free_list_mtx);
775 			if (error == 0)
776 				break;
777 			vp = NULL;
778 		}
779 	}
780 	if (vp) {
781 		freevnodes--;
782 		mtx_unlock(&vnode_free_list_mtx);
783 
784 #ifdef INVARIANTS
785 		{
786 			if (vp->v_data)
787 				panic("cleaned vnode isn't");
788 			if (vp->v_numoutput)
789 				panic("Clean vnode has pending I/O's");
790 			if (vp->v_writecount != 0)
791 				panic("Non-zero write count");
792 		}
793 #endif
794 		if ((pollinfo = vp->v_pollinfo) != NULL) {
795 			/*
796 			 * To avoid lock order reversals, the call to
797 			 * uma_zfree() must be delayed until the vnode
798 			 * interlock is released.
799 			 */
800 			vp->v_pollinfo = NULL;
801 		}
802 #ifdef MAC
803 		mac_destroy_vnode(vp);
804 #endif
805 		vp->v_iflag = 0;
806 		vp->v_vflag = 0;
807 		vp->v_lastw = 0;
808 		vp->v_lasta = 0;
809 		vp->v_cstart = 0;
810 		vp->v_clen = 0;
811 		vp->v_socket = 0;
812 		lockdestroy(vp->v_vnlock);
813 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
814 		KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0"));
815 		KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL"));
816 		KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0"));
817 		KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL"));
818 	} else {
819 		numvnodes++;
820 		mtx_unlock(&vnode_free_list_mtx);
821 
822 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
823 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
824 		VI_LOCK(vp);
825 		vp->v_dd = vp;
826 		vp->v_vnlock = &vp->v_lock;
827 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
828 		cache_purge(vp);		/* Sets up v_id. */
829 		LIST_INIT(&vp->v_cache_src);
830 		TAILQ_INIT(&vp->v_cache_dst);
831 	}
832 
833 	TAILQ_INIT(&vp->v_cleanblkhd);
834 	TAILQ_INIT(&vp->v_dirtyblkhd);
835 	vp->v_type = VNON;
836 	vp->v_tag = tag;
837 	vp->v_op = vops;
838 	*vpp = vp;
839 	vp->v_usecount = 1;
840 	vp->v_data = 0;
841 	vp->v_cachedid = -1;
842 	VI_UNLOCK(vp);
843 	if (pollinfo != NULL) {
844 		mtx_destroy(&pollinfo->vpi_lock);
845 		uma_zfree(vnodepoll_zone, pollinfo);
846 	}
847 #ifdef MAC
848 	mac_init_vnode(vp);
849 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
850 		mac_associate_vnode_singlelabel(mp, vp);
851 #endif
852 	delmntque(vp);
853 	if (mp != NULL) {
854 		insmntque(vp, mp);
855 		vp->v_bsize = mp->mnt_stat.f_iosize;
856 	}
857 
858 	return (0);
859 }
860 
861 /*
862  * Delete from old mount point vnode list, if on one.
863  */
864 static void
865 delmntque(struct vnode *vp)
866 {
867 	struct mount *mp;
868 
869 	if (vp->v_mount == NULL)
870 		return;
871 	mp = vp->v_mount;
872 	MNT_ILOCK(mp);
873 	vp->v_mount = NULL;
874 	KASSERT(mp->mnt_nvnodelistsize > 0,
875 		("bad mount point vnode list size"));
876 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
877 	mp->mnt_nvnodelistsize--;
878 	MNT_IUNLOCK(mp);
879 }
880 
881 /*
882  * Insert into list of vnodes for the new mount point, if available.
883  */
884 static void
885 insmntque(struct vnode *vp, struct mount *mp)
886 {
887 
888 	vp->v_mount = mp;
889 	KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)"));
890 	MNT_ILOCK(vp->v_mount);
891 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
892 	mp->mnt_nvnodelistsize++;
893 	MNT_IUNLOCK(vp->v_mount);
894 }
895 
896 /*
897  * Update outstanding I/O count and do wakeup if requested.
898  */
899 void
900 vwakeup(bp)
901 	register struct buf *bp;
902 {
903 	register struct vnode *vp;
904 
905 	bp->b_flags &= ~B_WRITEINPROG;
906 	if ((vp = bp->b_vp)) {
907 		VI_LOCK(vp);
908 		vp->v_numoutput--;
909 		if (vp->v_numoutput < 0)
910 			panic("vwakeup: neg numoutput");
911 		if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) {
912 			vp->v_iflag &= ~VI_BWAIT;
913 			wakeup(&vp->v_numoutput);
914 		}
915 		VI_UNLOCK(vp);
916 	}
917 }
918 
919 /*
920  * Flush out and invalidate all buffers associated with a vnode.
921  * Called with the underlying object locked.
922  */
923 int
924 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo)
925 	struct vnode *vp;
926 	int flags;
927 	struct ucred *cred;
928 	struct thread *td;
929 	int slpflag, slptimeo;
930 {
931 	struct buf *blist;
932 	int error;
933 	vm_object_t object;
934 
935 	GIANT_REQUIRED;
936 
937 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
938 
939 	VI_LOCK(vp);
940 	if (flags & V_SAVE) {
941 		while (vp->v_numoutput) {
942 			vp->v_iflag |= VI_BWAIT;
943 			error = msleep(&vp->v_numoutput, VI_MTX(vp),
944 			    slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo);
945 			if (error) {
946 				VI_UNLOCK(vp);
947 				return (error);
948 			}
949 		}
950 		if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
951 			VI_UNLOCK(vp);
952 			if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0)
953 				return (error);
954 			/*
955 			 * XXX We could save a lock/unlock if this was only
956 			 * enabled under INVARIANTS
957 			 */
958 			VI_LOCK(vp);
959 			if (vp->v_numoutput > 0 ||
960 			    !TAILQ_EMPTY(&vp->v_dirtyblkhd))
961 				panic("vinvalbuf: dirty bufs");
962 		}
963 	}
964 	/*
965 	 * If you alter this loop please notice that interlock is dropped and
966 	 * reacquired in flushbuflist.  Special care is needed to ensure that
967 	 * no race conditions occur from this.
968 	 */
969 	for (error = 0;;) {
970 		if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 &&
971 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
972 			if (error)
973 				break;
974 			continue;
975 		}
976 		if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 &&
977 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
978 			if (error)
979 				break;
980 			continue;
981 		}
982 		break;
983 	}
984 	if (error) {
985 		VI_UNLOCK(vp);
986 		return (error);
987 	}
988 
989 	/*
990 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
991 	 * have write I/O in-progress but if there is a VM object then the
992 	 * VM object can also have read-I/O in-progress.
993 	 */
994 	do {
995 		while (vp->v_numoutput > 0) {
996 			vp->v_iflag |= VI_BWAIT;
997 			msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0);
998 		}
999 		VI_UNLOCK(vp);
1000 		if (VOP_GETVOBJECT(vp, &object) == 0) {
1001 			VM_OBJECT_LOCK(object);
1002 			vm_object_pip_wait(object, "vnvlbx");
1003 			VM_OBJECT_UNLOCK(object);
1004 		}
1005 		VI_LOCK(vp);
1006 	} while (vp->v_numoutput > 0);
1007 	VI_UNLOCK(vp);
1008 
1009 	/*
1010 	 * Destroy the copy in the VM cache, too.
1011 	 */
1012 	if (VOP_GETVOBJECT(vp, &object) == 0) {
1013 		VM_OBJECT_LOCK(object);
1014 		vm_object_page_remove(object, 0, 0,
1015 			(flags & V_SAVE) ? TRUE : FALSE);
1016 		VM_OBJECT_UNLOCK(object);
1017 	}
1018 
1019 #ifdef INVARIANTS
1020 	VI_LOCK(vp);
1021 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1022 	    (!TAILQ_EMPTY(&vp->v_dirtyblkhd) ||
1023 	     !TAILQ_EMPTY(&vp->v_cleanblkhd)))
1024 		panic("vinvalbuf: flush failed");
1025 	VI_UNLOCK(vp);
1026 #endif
1027 	return (0);
1028 }
1029 
1030 /*
1031  * Flush out buffers on the specified list.
1032  *
1033  */
1034 static int
1035 flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp)
1036 	struct buf *blist;
1037 	int flags;
1038 	struct vnode *vp;
1039 	int slpflag, slptimeo;
1040 	int *errorp;
1041 {
1042 	struct buf *bp, *nbp;
1043 	int found, error;
1044 
1045 	ASSERT_VI_LOCKED(vp, "flushbuflist");
1046 
1047 	for (found = 0, bp = blist; bp; bp = nbp) {
1048 		nbp = TAILQ_NEXT(bp, b_vnbufs);
1049 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1050 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1051 			continue;
1052 		}
1053 		found += 1;
1054 		error = BUF_TIMELOCK(bp,
1055 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp),
1056 		    "flushbuf", slpflag, slptimeo);
1057 		if (error) {
1058 			if (error != ENOLCK)
1059 				*errorp = error;
1060 			goto done;
1061 		}
1062 		/*
1063 		 * XXX Since there are no node locks for NFS, I
1064 		 * believe there is a slight chance that a delayed
1065 		 * write will occur while sleeping just above, so
1066 		 * check for it.  Note that vfs_bio_awrite expects
1067 		 * buffers to reside on a queue, while bwrite and
1068 		 * brelse do not.
1069 		 */
1070 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1071 			(flags & V_SAVE)) {
1072 
1073 			if (bp->b_vp == vp) {
1074 				if (bp->b_flags & B_CLUSTEROK) {
1075 					vfs_bio_awrite(bp);
1076 				} else {
1077 					bremfree(bp);
1078 					bp->b_flags |= B_ASYNC;
1079 					bwrite(bp);
1080 				}
1081 			} else {
1082 				bremfree(bp);
1083 				(void) bwrite(bp);
1084 			}
1085 			goto done;
1086 		}
1087 		bremfree(bp);
1088 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1089 		bp->b_flags &= ~B_ASYNC;
1090 		brelse(bp);
1091 		VI_LOCK(vp);
1092 	}
1093 	return (found);
1094 done:
1095 	VI_LOCK(vp);
1096 	return (found);
1097 }
1098 
1099 /*
1100  * Truncate a file's buffer and pages to a specified length.  This
1101  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1102  * sync activity.
1103  */
1104 int
1105 vtruncbuf(vp, cred, td, length, blksize)
1106 	register struct vnode *vp;
1107 	struct ucred *cred;
1108 	struct thread *td;
1109 	off_t length;
1110 	int blksize;
1111 {
1112 	register struct buf *bp;
1113 	struct buf *nbp;
1114 	int anyfreed;
1115 	int trunclbn;
1116 
1117 	/*
1118 	 * Round up to the *next* lbn.
1119 	 */
1120 	trunclbn = (length + blksize - 1) / blksize;
1121 
1122 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1123 restart:
1124 	VI_LOCK(vp);
1125 	anyfreed = 1;
1126 	for (;anyfreed;) {
1127 		anyfreed = 0;
1128 		for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) {
1129 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1130 			if (bp->b_lblkno >= trunclbn) {
1131 				if (BUF_LOCK(bp,
1132 				    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1133 				    VI_MTX(vp)) == ENOLCK)
1134 					goto restart;
1135 
1136 				bremfree(bp);
1137 				bp->b_flags |= (B_INVAL | B_RELBUF);
1138 				bp->b_flags &= ~B_ASYNC;
1139 				brelse(bp);
1140 				anyfreed = 1;
1141 
1142 				if (nbp &&
1143 				    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1144 				    (nbp->b_vp != vp) ||
1145 				    (nbp->b_flags & B_DELWRI))) {
1146 					goto restart;
1147 				}
1148 				VI_LOCK(vp);
1149 			}
1150 		}
1151 
1152 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1153 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1154 			if (bp->b_lblkno >= trunclbn) {
1155 				if (BUF_LOCK(bp,
1156 				    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1157 				    VI_MTX(vp)) == ENOLCK)
1158 					goto restart;
1159 				bremfree(bp);
1160 				bp->b_flags |= (B_INVAL | B_RELBUF);
1161 				bp->b_flags &= ~B_ASYNC;
1162 				brelse(bp);
1163 				anyfreed = 1;
1164 				if (nbp &&
1165 				    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1166 				    (nbp->b_vp != vp) ||
1167 				    (nbp->b_flags & B_DELWRI) == 0)) {
1168 					goto restart;
1169 				}
1170 				VI_LOCK(vp);
1171 			}
1172 		}
1173 	}
1174 
1175 	if (length > 0) {
1176 restartsync:
1177 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1178 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1179 			if (bp->b_lblkno > 0)
1180 				continue;
1181 			/*
1182 			 * Since we hold the vnode lock this should only
1183 			 * fail if we're racing with the buf daemon.
1184 			 */
1185 			if (BUF_LOCK(bp,
1186 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1187 			    VI_MTX(vp)) == ENOLCK) {
1188 				goto restart;
1189 			}
1190 			KASSERT((bp->b_flags & B_DELWRI),
1191 			    ("buf(%p) on dirty queue without DELWRI", bp));
1192 
1193 			bremfree(bp);
1194 			bawrite(bp);
1195 			VI_LOCK(vp);
1196 			goto restartsync;
1197 		}
1198 	}
1199 
1200 	while (vp->v_numoutput > 0) {
1201 		vp->v_iflag |= VI_BWAIT;
1202 		msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0);
1203 	}
1204 	VI_UNLOCK(vp);
1205 	vnode_pager_setsize(vp, length);
1206 
1207 	return (0);
1208 }
1209 
1210 /*
1211  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1212  * 		 a vnode.
1213  *
1214  *	NOTE: We have to deal with the special case of a background bitmap
1215  *	buffer, a situation where two buffers will have the same logical
1216  *	block offset.  We want (1) only the foreground buffer to be accessed
1217  *	in a lookup and (2) must differentiate between the foreground and
1218  *	background buffer in the splay tree algorithm because the splay
1219  *	tree cannot normally handle multiple entities with the same 'index'.
1220  *	We accomplish this by adding differentiating flags to the splay tree's
1221  *	numerical domain.
1222  */
1223 static
1224 struct buf *
1225 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1226 {
1227 	struct buf dummy;
1228 	struct buf *lefttreemax, *righttreemin, *y;
1229 
1230 	if (root == NULL)
1231 		return (NULL);
1232 	lefttreemax = righttreemin = &dummy;
1233 	for (;;) {
1234 		if (lblkno < root->b_lblkno ||
1235 		    (lblkno == root->b_lblkno &&
1236 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1237 			if ((y = root->b_left) == NULL)
1238 				break;
1239 			if (lblkno < y->b_lblkno) {
1240 				/* Rotate right. */
1241 				root->b_left = y->b_right;
1242 				y->b_right = root;
1243 				root = y;
1244 				if ((y = root->b_left) == NULL)
1245 					break;
1246 			}
1247 			/* Link into the new root's right tree. */
1248 			righttreemin->b_left = root;
1249 			righttreemin = root;
1250 		} else if (lblkno > root->b_lblkno ||
1251 		    (lblkno == root->b_lblkno &&
1252 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1253 			if ((y = root->b_right) == NULL)
1254 				break;
1255 			if (lblkno > y->b_lblkno) {
1256 				/* Rotate left. */
1257 				root->b_right = y->b_left;
1258 				y->b_left = root;
1259 				root = y;
1260 				if ((y = root->b_right) == NULL)
1261 					break;
1262 			}
1263 			/* Link into the new root's left tree. */
1264 			lefttreemax->b_right = root;
1265 			lefttreemax = root;
1266 		} else {
1267 			break;
1268 		}
1269 		root = y;
1270 	}
1271 	/* Assemble the new root. */
1272 	lefttreemax->b_right = root->b_left;
1273 	righttreemin->b_left = root->b_right;
1274 	root->b_left = dummy.b_right;
1275 	root->b_right = dummy.b_left;
1276 	return (root);
1277 }
1278 
1279 static
1280 void
1281 buf_vlist_remove(struct buf *bp)
1282 {
1283 	struct vnode *vp = bp->b_vp;
1284 	struct buf *root;
1285 
1286 	ASSERT_VI_LOCKED(vp, "buf_vlist_remove");
1287 	if (bp->b_xflags & BX_VNDIRTY) {
1288 		if (bp != vp->v_dirtyblkroot) {
1289 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1290 			    vp->v_dirtyblkroot);
1291 			KASSERT(root == bp,
1292 			    ("splay lookup failed during dirty remove"));
1293 		}
1294 		if (bp->b_left == NULL) {
1295 			root = bp->b_right;
1296 		} else {
1297 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1298 			    bp->b_left);
1299 			root->b_right = bp->b_right;
1300 		}
1301 		vp->v_dirtyblkroot = root;
1302 		TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs);
1303 		vp->v_dirtybufcnt--;
1304 	} else {
1305 		/* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */
1306 		if (bp != vp->v_cleanblkroot) {
1307 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1308 			    vp->v_cleanblkroot);
1309 			KASSERT(root == bp,
1310 			    ("splay lookup failed during clean remove"));
1311 		}
1312 		if (bp->b_left == NULL) {
1313 			root = bp->b_right;
1314 		} else {
1315 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1316 			    bp->b_left);
1317 			root->b_right = bp->b_right;
1318 		}
1319 		vp->v_cleanblkroot = root;
1320 		TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs);
1321 		vp->v_cleanbufcnt--;
1322 	}
1323 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1324 }
1325 
1326 /*
1327  * Add the buffer to the sorted clean or dirty block list using a
1328  * splay tree algorithm.
1329  *
1330  * NOTE: xflags is passed as a constant, optimizing this inline function!
1331  */
1332 static
1333 void
1334 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags)
1335 {
1336 	struct buf *root;
1337 
1338 	ASSERT_VI_LOCKED(vp, "buf_vlist_add");
1339 	bp->b_xflags |= xflags;
1340 	if (xflags & BX_VNDIRTY) {
1341 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1342 		if (root == NULL) {
1343 			bp->b_left = NULL;
1344 			bp->b_right = NULL;
1345 			TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs);
1346 		} else if (bp->b_lblkno < root->b_lblkno ||
1347 		    (bp->b_lblkno == root->b_lblkno &&
1348 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1349 			bp->b_left = root->b_left;
1350 			bp->b_right = root;
1351 			root->b_left = NULL;
1352 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1353 		} else {
1354 			bp->b_right = root->b_right;
1355 			bp->b_left = root;
1356 			root->b_right = NULL;
1357 			TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd,
1358 			    root, bp, b_vnbufs);
1359 		}
1360 		vp->v_dirtybufcnt++;
1361 		vp->v_dirtyblkroot = bp;
1362 	} else {
1363 		/* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */
1364 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1365 		if (root == NULL) {
1366 			bp->b_left = NULL;
1367 			bp->b_right = NULL;
1368 			TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs);
1369 		} else if (bp->b_lblkno < root->b_lblkno ||
1370 		    (bp->b_lblkno == root->b_lblkno &&
1371 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1372 			bp->b_left = root->b_left;
1373 			bp->b_right = root;
1374 			root->b_left = NULL;
1375 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1376 		} else {
1377 			bp->b_right = root->b_right;
1378 			bp->b_left = root;
1379 			root->b_right = NULL;
1380 			TAILQ_INSERT_AFTER(&vp->v_cleanblkhd,
1381 			    root, bp, b_vnbufs);
1382 		}
1383 		vp->v_cleanbufcnt++;
1384 		vp->v_cleanblkroot = bp;
1385 	}
1386 }
1387 
1388 /*
1389  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1390  * shadow buffers used in background bitmap writes.
1391  *
1392  * This code isn't quite efficient as it could be because we are maintaining
1393  * two sorted lists and do not know which list the block resides in.
1394  *
1395  * During a "make buildworld" the desired buffer is found at one of
1396  * the roots more than 60% of the time.  Thus, checking both roots
1397  * before performing either splay eliminates unnecessary splays on the
1398  * first tree splayed.
1399  */
1400 struct buf *
1401 gbincore(struct vnode *vp, daddr_t lblkno)
1402 {
1403 	struct buf *bp;
1404 
1405 	GIANT_REQUIRED;
1406 
1407 	ASSERT_VI_LOCKED(vp, "gbincore");
1408 	if ((bp = vp->v_cleanblkroot) != NULL &&
1409 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1410 		return (bp);
1411 	if ((bp = vp->v_dirtyblkroot) != NULL &&
1412 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1413 		return (bp);
1414 	if ((bp = vp->v_cleanblkroot) != NULL) {
1415 		vp->v_cleanblkroot = bp = buf_splay(lblkno, 0, bp);
1416 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1417 			return (bp);
1418 	}
1419 	if ((bp = vp->v_dirtyblkroot) != NULL) {
1420 		vp->v_dirtyblkroot = bp = buf_splay(lblkno, 0, bp);
1421 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1422 			return (bp);
1423 	}
1424 	return (NULL);
1425 }
1426 
1427 /*
1428  * Associate a buffer with a vnode.
1429  */
1430 void
1431 bgetvp(vp, bp)
1432 	register struct vnode *vp;
1433 	register struct buf *bp;
1434 {
1435 
1436 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1437 
1438 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1439 	    ("bgetvp: bp already attached! %p", bp));
1440 
1441 	ASSERT_VI_LOCKED(vp, "bgetvp");
1442 	vholdl(vp);
1443 	bp->b_vp = vp;
1444 	bp->b_dev = vn_todev(vp);
1445 	/*
1446 	 * Insert onto list for new vnode.
1447 	 */
1448 	buf_vlist_add(bp, vp, BX_VNCLEAN);
1449 }
1450 
1451 /*
1452  * Disassociate a buffer from a vnode.
1453  */
1454 void
1455 brelvp(bp)
1456 	register struct buf *bp;
1457 {
1458 	struct vnode *vp;
1459 
1460 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1461 
1462 	/*
1463 	 * Delete from old vnode list, if on one.
1464 	 */
1465 	vp = bp->b_vp;
1466 	VI_LOCK(vp);
1467 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1468 		buf_vlist_remove(bp);
1469 	if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1470 		vp->v_iflag &= ~VI_ONWORKLST;
1471 		mtx_lock(&sync_mtx);
1472 		LIST_REMOVE(vp, v_synclist);
1473  		syncer_worklist_len--;
1474 		mtx_unlock(&sync_mtx);
1475 	}
1476 	vdropl(vp);
1477 	bp->b_vp = (struct vnode *) 0;
1478 	if (bp->b_object)
1479 		bp->b_object = NULL;
1480 	VI_UNLOCK(vp);
1481 }
1482 
1483 /*
1484  * Add an item to the syncer work queue.
1485  */
1486 static void
1487 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
1488 {
1489 	int slot;
1490 
1491 	ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist");
1492 
1493 	mtx_lock(&sync_mtx);
1494 	if (vp->v_iflag & VI_ONWORKLST)
1495 		LIST_REMOVE(vp, v_synclist);
1496 	else {
1497 		vp->v_iflag |= VI_ONWORKLST;
1498  		syncer_worklist_len++;
1499 	}
1500 
1501 	if (delay > syncer_maxdelay - 2)
1502 		delay = syncer_maxdelay - 2;
1503 	slot = (syncer_delayno + delay) & syncer_mask;
1504 
1505 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
1506 	mtx_unlock(&sync_mtx);
1507 }
1508 
1509 static int
1510 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1511 {
1512 	int error, len;
1513 
1514 	mtx_lock(&sync_mtx);
1515 	len = syncer_worklist_len - sync_vnode_count;
1516 	mtx_unlock(&sync_mtx);
1517 	error = SYSCTL_OUT(req, &len, sizeof(len));
1518 	return (error);
1519 }
1520 
1521 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1522     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1523 
1524 struct  proc *updateproc;
1525 static void sched_sync(void);
1526 static struct kproc_desc up_kp = {
1527 	"syncer",
1528 	sched_sync,
1529 	&updateproc
1530 };
1531 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1532 
1533 /*
1534  * System filesystem synchronizer daemon.
1535  */
1536 static void
1537 sched_sync(void)
1538 {
1539 	struct synclist *next;
1540 	struct synclist *slp;
1541 	struct vnode *vp;
1542 	struct mount *mp;
1543 	long starttime;
1544 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1545 	static int dummychan;
1546 	int last_work_seen;
1547 	int net_worklist_len;
1548 	int syncer_final_iter;
1549 	int first_printf;
1550 
1551 	mtx_lock(&Giant);
1552 	last_work_seen = 0;
1553 	syncer_final_iter = 0;
1554 	first_printf = 1;
1555 	syncer_state = SYNCER_RUNNING;
1556 	starttime = time_second;
1557 
1558 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1559 	    SHUTDOWN_PRI_LAST);
1560 
1561 	for (;;) {
1562 		mtx_lock(&sync_mtx);
1563 		if (syncer_state == SYNCER_FINAL_DELAY &&
1564 		    syncer_final_iter == 0) {
1565 			mtx_unlock(&sync_mtx);
1566 			kthread_suspend_check(td->td_proc);
1567 			mtx_lock(&sync_mtx);
1568 		}
1569 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1570 		if (syncer_state != SYNCER_RUNNING &&
1571 		    starttime != time_second) {
1572 			if (first_printf) {
1573 				printf("\nSyncer syncing, vnodes remaining...");
1574 				first_printf = 0;
1575 			}
1576 			printf("%d ", net_worklist_len);
1577 		}
1578 		starttime = time_second;
1579 
1580 		/*
1581 		 * Push files whose dirty time has expired.  Be careful
1582 		 * of interrupt race on slp queue.
1583 		 *
1584 		 * Skip over empty worklist slots when shutting down.
1585 		 */
1586 		do {
1587 			slp = &syncer_workitem_pending[syncer_delayno];
1588 			syncer_delayno += 1;
1589 			if (syncer_delayno == syncer_maxdelay)
1590 				syncer_delayno = 0;
1591 			next = &syncer_workitem_pending[syncer_delayno];
1592 			/*
1593 			 * If the worklist has wrapped since the
1594 			 * it was emptied of all but syncer vnodes,
1595 			 * switch to the FINAL_DELAY state and run
1596 			 * for one more second.
1597 			 */
1598 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1599 			    net_worklist_len == 0 &&
1600 			    last_work_seen == syncer_delayno) {
1601 				syncer_state = SYNCER_FINAL_DELAY;
1602 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1603 			}
1604 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1605 		    syncer_worklist_len > 0);
1606 
1607 		/*
1608 		 * Keep track of the last time there was anything
1609 		 * on the worklist other than syncer vnodes.
1610 		 * Return to the SHUTTING_DOWN state if any
1611 		 * new work appears.
1612 		 */
1613 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1614 			last_work_seen = syncer_delayno;
1615 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1616 			syncer_state = SYNCER_SHUTTING_DOWN;
1617 		while ((vp = LIST_FIRST(slp)) != NULL) {
1618 			if (VOP_ISLOCKED(vp, NULL) != 0 ||
1619 			    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1620 				LIST_REMOVE(vp, v_synclist);
1621 				LIST_INSERT_HEAD(next, vp, v_synclist);
1622 				continue;
1623 			}
1624 			if (VI_TRYLOCK(vp) == 0) {
1625 				LIST_REMOVE(vp, v_synclist);
1626 				LIST_INSERT_HEAD(next, vp, v_synclist);
1627 				vn_finished_write(mp);
1628 				continue;
1629 			}
1630 			/*
1631 			 * We use vhold in case the vnode does not
1632 			 * successfully sync.  vhold prevents the vnode from
1633 			 * going away when we unlock the sync_mtx so that
1634 			 * we can acquire the vnode interlock.
1635 			 */
1636 			vholdl(vp);
1637 			mtx_unlock(&sync_mtx);
1638 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td);
1639 			(void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td);
1640 			VOP_UNLOCK(vp, 0, td);
1641 			vn_finished_write(mp);
1642 			VI_LOCK(vp);
1643 			if ((vp->v_iflag & VI_ONWORKLST) != 0) {
1644 				/*
1645 				 * Put us back on the worklist.  The worklist
1646 				 * routine will remove us from our current
1647 				 * position and then add us back in at a later
1648 				 * position.
1649 				 */
1650 				vn_syncer_add_to_worklist(vp, syncdelay);
1651 			}
1652 			vdropl(vp);
1653 			VI_UNLOCK(vp);
1654 			mtx_lock(&sync_mtx);
1655 		}
1656 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1657 			syncer_final_iter--;
1658 		mtx_unlock(&sync_mtx);
1659 
1660 		/*
1661 		 * Do soft update processing.
1662 		 */
1663 		if (softdep_process_worklist_hook != NULL)
1664 			(*softdep_process_worklist_hook)(NULL);
1665 
1666 		/*
1667 		 * The variable rushjob allows the kernel to speed up the
1668 		 * processing of the filesystem syncer process. A rushjob
1669 		 * value of N tells the filesystem syncer to process the next
1670 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1671 		 * is used by the soft update code to speed up the filesystem
1672 		 * syncer process when the incore state is getting so far
1673 		 * ahead of the disk that the kernel memory pool is being
1674 		 * threatened with exhaustion.
1675 		 */
1676 		mtx_lock(&sync_mtx);
1677 		if (rushjob > 0) {
1678 			rushjob -= 1;
1679 			mtx_unlock(&sync_mtx);
1680 			continue;
1681 		}
1682 		mtx_unlock(&sync_mtx);
1683 		/*
1684 		 * Just sleep for a short period if time between
1685 		 * iterations when shutting down to allow some I/O
1686 		 * to happen.
1687 		 *
1688 		 * If it has taken us less than a second to process the
1689 		 * current work, then wait. Otherwise start right over
1690 		 * again. We can still lose time if any single round
1691 		 * takes more than two seconds, but it does not really
1692 		 * matter as we are just trying to generally pace the
1693 		 * filesystem activity.
1694 		 */
1695 		if (syncer_state != SYNCER_RUNNING)
1696 			tsleep(&dummychan, PPAUSE, "syncfnl",
1697 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1698 		else if (time_second == starttime)
1699 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1700 	}
1701 }
1702 
1703 /*
1704  * Request the syncer daemon to speed up its work.
1705  * We never push it to speed up more than half of its
1706  * normal turn time, otherwise it could take over the cpu.
1707  */
1708 int
1709 speedup_syncer()
1710 {
1711 	struct thread *td;
1712 	int ret = 0;
1713 
1714 	td = FIRST_THREAD_IN_PROC(updateproc);
1715 	sleepq_remove(td, &lbolt);
1716 	mtx_lock(&sync_mtx);
1717 	if (rushjob < syncdelay / 2) {
1718 		rushjob += 1;
1719 		stat_rush_requests += 1;
1720 		ret = 1;
1721 	}
1722 	mtx_unlock(&sync_mtx);
1723 	return (ret);
1724 }
1725 
1726 /*
1727  * Tell the syncer to speed up its work and run though its work
1728  * list several times, then tell it to shut down.
1729  */
1730 static void
1731 syncer_shutdown(void *arg, int howto)
1732 {
1733 	struct thread *td;
1734 
1735 	td = FIRST_THREAD_IN_PROC(updateproc);
1736 	sleepq_remove(td, &lbolt);
1737 	mtx_lock(&sync_mtx);
1738 	syncer_state = SYNCER_SHUTTING_DOWN;
1739 	rushjob = 0;
1740 	mtx_unlock(&sync_mtx);
1741 	kproc_shutdown(arg, howto);
1742 }
1743 
1744 /*
1745  * Associate a p-buffer with a vnode.
1746  *
1747  * Also sets B_PAGING flag to indicate that vnode is not fully associated
1748  * with the buffer.  i.e. the bp has not been linked into the vnode or
1749  * ref-counted.
1750  */
1751 void
1752 pbgetvp(vp, bp)
1753 	register struct vnode *vp;
1754 	register struct buf *bp;
1755 {
1756 
1757 	KASSERT(bp->b_vp == NULL, ("pbgetvp: not free"));
1758 
1759 	bp->b_vp = vp;
1760 	bp->b_object = vp->v_object;
1761 	bp->b_flags |= B_PAGING;
1762 	bp->b_dev = vn_todev(vp);
1763 }
1764 
1765 /*
1766  * Disassociate a p-buffer from a vnode.
1767  */
1768 void
1769 pbrelvp(bp)
1770 	register struct buf *bp;
1771 {
1772 
1773 	KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL"));
1774 
1775 	/* XXX REMOVE ME */
1776 	VI_LOCK(bp->b_vp);
1777 	if (TAILQ_NEXT(bp, b_vnbufs) != NULL) {
1778 		panic(
1779 		    "relpbuf(): b_vp was probably reassignbuf()d %p %x",
1780 		    bp,
1781 		    (int)bp->b_flags
1782 		);
1783 	}
1784 	VI_UNLOCK(bp->b_vp);
1785 	bp->b_vp = (struct vnode *) 0;
1786 	bp->b_object = NULL;
1787 	bp->b_flags &= ~B_PAGING;
1788 }
1789 
1790 /*
1791  * Reassign a buffer from one vnode to another.
1792  * Used to assign file specific control information
1793  * (indirect blocks) to the vnode to which they belong.
1794  */
1795 void
1796 reassignbuf(struct buf *bp)
1797 {
1798 	struct vnode *vp;
1799 	int delay;
1800 
1801 	vp = bp->b_vp;
1802 	++reassignbufcalls;
1803 
1804 	/*
1805 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1806 	 * is not fully linked in.
1807 	 */
1808 	if (bp->b_flags & B_PAGING)
1809 		panic("cannot reassign paging buffer");
1810 
1811 	/*
1812 	 * Delete from old vnode list, if on one.
1813 	 */
1814 	VI_LOCK(vp);
1815 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1816 		buf_vlist_remove(bp);
1817 	/*
1818 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1819 	 * of clean buffers.
1820 	 */
1821 	if (bp->b_flags & B_DELWRI) {
1822 		if ((vp->v_iflag & VI_ONWORKLST) == 0) {
1823 			switch (vp->v_type) {
1824 			case VDIR:
1825 				delay = dirdelay;
1826 				break;
1827 			case VCHR:
1828 				delay = metadelay;
1829 				break;
1830 			default:
1831 				delay = filedelay;
1832 			}
1833 			vn_syncer_add_to_worklist(vp, delay);
1834 		}
1835 		buf_vlist_add(bp, vp, BX_VNDIRTY);
1836 	} else {
1837 		buf_vlist_add(bp, vp, BX_VNCLEAN);
1838 
1839 		if ((vp->v_iflag & VI_ONWORKLST) &&
1840 		    TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1841 			mtx_lock(&sync_mtx);
1842 			LIST_REMOVE(vp, v_synclist);
1843  			syncer_worklist_len--;
1844 			mtx_unlock(&sync_mtx);
1845 			vp->v_iflag &= ~VI_ONWORKLST;
1846 		}
1847 	}
1848 	VI_UNLOCK(vp);
1849 }
1850 
1851 /*
1852  * Create a vnode for a device.
1853  * Used for mounting the root filesystem.
1854  */
1855 int
1856 bdevvp(dev, vpp)
1857 	struct cdev *dev;
1858 	struct vnode **vpp;
1859 {
1860 	register struct vnode *vp;
1861 	struct vnode *nvp;
1862 	int error;
1863 
1864 	if (dev == NULL) {
1865 		*vpp = NULLVP;
1866 		return (ENXIO);
1867 	}
1868 	if (vfinddev(dev, vpp))
1869 		return (0);
1870 
1871 	error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp);
1872 	if (error) {
1873 		*vpp = NULLVP;
1874 		return (error);
1875 	}
1876 	vp = nvp;
1877 	vp->v_type = VCHR;
1878 	vp->v_bsize = DEV_BSIZE;
1879 	addalias(vp, dev);
1880 	*vpp = vp;
1881 	return (0);
1882 }
1883 
1884 static void
1885 v_incr_usecount(struct vnode *vp, int delta)
1886 {
1887 
1888 	vp->v_usecount += delta;
1889 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1890 		mtx_lock(&spechash_mtx);
1891 		vp->v_rdev->si_usecount += delta;
1892 		mtx_unlock(&spechash_mtx);
1893 	}
1894 }
1895 
1896 /*
1897  * Add vnode to the alias list hung off the struct cdev *.
1898  *
1899  * The reason for this gunk is that multiple vnodes can reference
1900  * the same physical device, so checking vp->v_usecount to see
1901  * how many users there are is inadequate; the v_usecount for
1902  * the vnodes need to be accumulated.  vcount() does that.
1903  */
1904 struct vnode *
1905 addaliasu(nvp, nvp_rdev)
1906 	struct vnode *nvp;
1907 	dev_t nvp_rdev;
1908 {
1909 	struct vnode *ovp;
1910 	vop_t **ops;
1911 	struct cdev *dev;
1912 
1913 	if (nvp->v_type == VBLK)
1914 		return (nvp);
1915 	if (nvp->v_type != VCHR)
1916 		panic("addaliasu on non-special vnode");
1917 	dev = findcdev(nvp_rdev);
1918 	if (dev == NULL)
1919 		return (nvp);
1920 	/*
1921 	 * Check to see if we have a bdevvp vnode with no associated
1922 	 * filesystem. If so, we want to associate the filesystem of
1923 	 * the new newly instigated vnode with the bdevvp vnode and
1924 	 * discard the newly created vnode rather than leaving the
1925 	 * bdevvp vnode lying around with no associated filesystem.
1926 	 */
1927 	if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) {
1928 		addalias(nvp, dev);
1929 		return (nvp);
1930 	}
1931 	/*
1932 	 * Discard unneeded vnode, but save its node specific data.
1933 	 * Note that if there is a lock, it is carried over in the
1934 	 * node specific data to the replacement vnode.
1935 	 */
1936 	vref(ovp);
1937 	ovp->v_data = nvp->v_data;
1938 	ovp->v_tag = nvp->v_tag;
1939 	nvp->v_data = NULL;
1940 	lockdestroy(ovp->v_vnlock);
1941 	lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg,
1942 	    nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK);
1943 	ops = ovp->v_op;
1944 	ovp->v_op = nvp->v_op;
1945 	if (VOP_ISLOCKED(nvp, curthread)) {
1946 		VOP_UNLOCK(nvp, 0, curthread);
1947 		vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread);
1948 	}
1949 	nvp->v_op = ops;
1950 	delmntque(ovp);
1951 	insmntque(ovp, nvp->v_mount);
1952 	vrele(nvp);
1953 	vgone(nvp);
1954 	return (ovp);
1955 }
1956 
1957 /* This is a local helper function that do the same as addaliasu, but for a
1958  * struct cdev *instead of an dev_t. */
1959 static void
1960 addalias(nvp, dev)
1961 	struct vnode *nvp;
1962 	struct cdev *dev;
1963 {
1964 
1965 	KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode"));
1966 	dev_ref(dev);
1967 	nvp->v_rdev = dev;
1968 	VI_LOCK(nvp);
1969 	mtx_lock(&spechash_mtx);
1970 	SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext);
1971 	dev->si_usecount += nvp->v_usecount;
1972 	mtx_unlock(&spechash_mtx);
1973 	VI_UNLOCK(nvp);
1974 }
1975 
1976 /*
1977  * Grab a particular vnode from the free list, increment its
1978  * reference count and lock it. The vnode lock bit is set if the
1979  * vnode is being eliminated in vgone. The process is awakened
1980  * when the transition is completed, and an error returned to
1981  * indicate that the vnode is no longer usable (possibly having
1982  * been changed to a new filesystem type).
1983  */
1984 int
1985 vget(vp, flags, td)
1986 	register struct vnode *vp;
1987 	int flags;
1988 	struct thread *td;
1989 {
1990 	int error;
1991 
1992 	/*
1993 	 * If the vnode is in the process of being cleaned out for
1994 	 * another use, we wait for the cleaning to finish and then
1995 	 * return failure. Cleaning is determined by checking that
1996 	 * the VI_XLOCK flag is set.
1997 	 */
1998 	if ((flags & LK_INTERLOCK) == 0)
1999 		VI_LOCK(vp);
2000 	if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) {
2001 		if ((flags & LK_NOWAIT) == 0) {
2002 			vp->v_iflag |= VI_XWANT;
2003 			msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0);
2004 			return (ENOENT);
2005 		}
2006 		VI_UNLOCK(vp);
2007 		return (EBUSY);
2008 	}
2009 
2010 	v_incr_usecount(vp, 1);
2011 
2012 	if (VSHOULDBUSY(vp))
2013 		vbusy(vp);
2014 	if (flags & LK_TYPE_MASK) {
2015 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2016 			/*
2017 			 * must expand vrele here because we do not want
2018 			 * to call VOP_INACTIVE if the reference count
2019 			 * drops back to zero since it was never really
2020 			 * active. We must remove it from the free list
2021 			 * before sleeping so that multiple processes do
2022 			 * not try to recycle it.
2023 			 */
2024 			VI_LOCK(vp);
2025 			v_incr_usecount(vp, -1);
2026 			if (VSHOULDFREE(vp))
2027 				vfree(vp);
2028 			else
2029 				vlruvp(vp);
2030 			VI_UNLOCK(vp);
2031 		}
2032 		return (error);
2033 	}
2034 	VI_UNLOCK(vp);
2035 	return (0);
2036 }
2037 
2038 /*
2039  * Increase the reference count of a vnode.
2040  */
2041 void
2042 vref(struct vnode *vp)
2043 {
2044 
2045 	VI_LOCK(vp);
2046 	v_incr_usecount(vp, 1);
2047 	VI_UNLOCK(vp);
2048 }
2049 
2050 /*
2051  * Return reference count of a vnode.
2052  *
2053  * The results of this call are only guaranteed when some mechanism other
2054  * than the VI lock is used to stop other processes from gaining references
2055  * to the vnode.  This may be the case if the caller holds the only reference.
2056  * This is also useful when stale data is acceptable as race conditions may
2057  * be accounted for by some other means.
2058  */
2059 int
2060 vrefcnt(struct vnode *vp)
2061 {
2062 	int usecnt;
2063 
2064 	VI_LOCK(vp);
2065 	usecnt = vp->v_usecount;
2066 	VI_UNLOCK(vp);
2067 
2068 	return (usecnt);
2069 }
2070 
2071 
2072 /*
2073  * Vnode put/release.
2074  * If count drops to zero, call inactive routine and return to freelist.
2075  */
2076 void
2077 vrele(vp)
2078 	struct vnode *vp;
2079 {
2080 	struct thread *td = curthread;	/* XXX */
2081 
2082 	GIANT_REQUIRED;
2083 
2084 	KASSERT(vp != NULL, ("vrele: null vp"));
2085 
2086 	VI_LOCK(vp);
2087 
2088 	/* Skip this v_writecount check if we're going to panic below. */
2089 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2090 	    ("vrele: missed vn_close"));
2091 
2092 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2093 	    vp->v_usecount == 1)) {
2094 		v_incr_usecount(vp, -1);
2095 		VI_UNLOCK(vp);
2096 
2097 		return;
2098 	}
2099 
2100 	if (vp->v_usecount == 1) {
2101 		v_incr_usecount(vp, -1);
2102 		/*
2103 		 * We must call VOP_INACTIVE with the node locked. Mark
2104 		 * as VI_DOINGINACT to avoid recursion.
2105 		 */
2106 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2107 			VI_LOCK(vp);
2108 			vp->v_iflag |= VI_DOINGINACT;
2109 			VI_UNLOCK(vp);
2110 			VOP_INACTIVE(vp, td);
2111 			VI_LOCK(vp);
2112 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2113 			    ("vrele: lost VI_DOINGINACT"));
2114 			vp->v_iflag &= ~VI_DOINGINACT;
2115 		} else
2116 			VI_LOCK(vp);
2117 		if (VSHOULDFREE(vp))
2118 			vfree(vp);
2119 		else
2120 			vlruvp(vp);
2121 		VI_UNLOCK(vp);
2122 
2123 	} else {
2124 #ifdef DIAGNOSTIC
2125 		vprint("vrele: negative ref count", vp);
2126 #endif
2127 		VI_UNLOCK(vp);
2128 		panic("vrele: negative ref cnt");
2129 	}
2130 }
2131 
2132 /*
2133  * Release an already locked vnode.  This give the same effects as
2134  * unlock+vrele(), but takes less time and avoids releasing and
2135  * re-aquiring the lock (as vrele() aquires the lock internally.)
2136  */
2137 void
2138 vput(vp)
2139 	struct vnode *vp;
2140 {
2141 	struct thread *td = curthread;	/* XXX */
2142 
2143 	GIANT_REQUIRED;
2144 
2145 	KASSERT(vp != NULL, ("vput: null vp"));
2146 	VI_LOCK(vp);
2147 	/* Skip this v_writecount check if we're going to panic below. */
2148 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2149 	    ("vput: missed vn_close"));
2150 
2151 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2152 	    vp->v_usecount == 1)) {
2153 		v_incr_usecount(vp, -1);
2154 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
2155 		return;
2156 	}
2157 
2158 	if (vp->v_usecount == 1) {
2159 		v_incr_usecount(vp, -1);
2160 		/*
2161 		 * We must call VOP_INACTIVE with the node locked, so
2162 		 * we just need to release the vnode mutex. Mark as
2163 		 * as VI_DOINGINACT to avoid recursion.
2164 		 */
2165 		vp->v_iflag |= VI_DOINGINACT;
2166 		VI_UNLOCK(vp);
2167 		VOP_INACTIVE(vp, td);
2168 		VI_LOCK(vp);
2169 		KASSERT(vp->v_iflag & VI_DOINGINACT,
2170 		    ("vput: lost VI_DOINGINACT"));
2171 		vp->v_iflag &= ~VI_DOINGINACT;
2172 		if (VSHOULDFREE(vp))
2173 			vfree(vp);
2174 		else
2175 			vlruvp(vp);
2176 		VI_UNLOCK(vp);
2177 
2178 	} else {
2179 #ifdef DIAGNOSTIC
2180 		vprint("vput: negative ref count", vp);
2181 #endif
2182 		panic("vput: negative ref cnt");
2183 	}
2184 }
2185 
2186 /*
2187  * Somebody doesn't want the vnode recycled.
2188  */
2189 void
2190 vhold(struct vnode *vp)
2191 {
2192 
2193 	VI_LOCK(vp);
2194 	vholdl(vp);
2195 	VI_UNLOCK(vp);
2196 }
2197 
2198 void
2199 vholdl(vp)
2200 	register struct vnode *vp;
2201 {
2202 
2203 	vp->v_holdcnt++;
2204 	if (VSHOULDBUSY(vp))
2205 		vbusy(vp);
2206 }
2207 
2208 /*
2209  * Note that there is one less who cares about this vnode.  vdrop() is the
2210  * opposite of vhold().
2211  */
2212 void
2213 vdrop(struct vnode *vp)
2214 {
2215 
2216 	VI_LOCK(vp);
2217 	vdropl(vp);
2218 	VI_UNLOCK(vp);
2219 }
2220 
2221 void
2222 vdropl(vp)
2223 	register struct vnode *vp;
2224 {
2225 
2226 	if (vp->v_holdcnt <= 0)
2227 		panic("vdrop: holdcnt");
2228 	vp->v_holdcnt--;
2229 	if (VSHOULDFREE(vp))
2230 		vfree(vp);
2231 	else
2232 		vlruvp(vp);
2233 }
2234 
2235 /*
2236  * Remove any vnodes in the vnode table belonging to mount point mp.
2237  *
2238  * If FORCECLOSE is not specified, there should not be any active ones,
2239  * return error if any are found (nb: this is a user error, not a
2240  * system error). If FORCECLOSE is specified, detach any active vnodes
2241  * that are found.
2242  *
2243  * If WRITECLOSE is set, only flush out regular file vnodes open for
2244  * writing.
2245  *
2246  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2247  *
2248  * `rootrefs' specifies the base reference count for the root vnode
2249  * of this filesystem. The root vnode is considered busy if its
2250  * v_usecount exceeds this value. On a successful return, vflush(, td)
2251  * will call vrele() on the root vnode exactly rootrefs times.
2252  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2253  * be zero.
2254  */
2255 #ifdef DIAGNOSTIC
2256 static int busyprt = 0;		/* print out busy vnodes */
2257 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2258 #endif
2259 
2260 int
2261 vflush(mp, rootrefs, flags, td)
2262 	struct mount *mp;
2263 	int rootrefs;
2264 	int flags;
2265 	struct thread *td;
2266 {
2267 	struct vnode *vp, *nvp, *rootvp = NULL;
2268 	struct vattr vattr;
2269 	int busy = 0, error;
2270 
2271 	if (rootrefs > 0) {
2272 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2273 		    ("vflush: bad args"));
2274 		/*
2275 		 * Get the filesystem root vnode. We can vput() it
2276 		 * immediately, since with rootrefs > 0, it won't go away.
2277 		 */
2278 		if ((error = VFS_ROOT(mp, &rootvp, td)) != 0)
2279 			return (error);
2280 		vput(rootvp);
2281 
2282 	}
2283 	MNT_ILOCK(mp);
2284 loop:
2285 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2286 
2287 		VI_LOCK(vp);
2288 		MNT_IUNLOCK(mp);
2289 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2290 		if (error) {
2291 			MNT_ILOCK(mp);
2292 			goto loop;
2293 		}
2294 		/*
2295 		 * Skip over a vnodes marked VV_SYSTEM.
2296 		 */
2297 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2298 			VOP_UNLOCK(vp, 0, td);
2299 			MNT_ILOCK(mp);
2300 			continue;
2301 		}
2302 		/*
2303 		 * If WRITECLOSE is set, flush out unlinked but still open
2304 		 * files (even if open only for reading) and regular file
2305 		 * vnodes open for writing.
2306 		 */
2307 		if (flags & WRITECLOSE) {
2308 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2309 			VI_LOCK(vp);
2310 
2311 			if ((vp->v_type == VNON ||
2312 			    (error == 0 && vattr.va_nlink > 0)) &&
2313 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2314 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2315 				MNT_ILOCK(mp);
2316 				continue;
2317 			}
2318 		} else
2319 			VI_LOCK(vp);
2320 
2321 		VOP_UNLOCK(vp, 0, td);
2322 
2323 		/*
2324 		 * With v_usecount == 0, all we need to do is clear out the
2325 		 * vnode data structures and we are done.
2326 		 */
2327 		if (vp->v_usecount == 0) {
2328 			vgonel(vp, td);
2329 			MNT_ILOCK(mp);
2330 			continue;
2331 		}
2332 
2333 		/*
2334 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2335 		 * or character devices, revert to an anonymous device. For
2336 		 * all other files, just kill them.
2337 		 */
2338 		if (flags & FORCECLOSE) {
2339 			if (vp->v_type != VCHR)
2340 				vgonel(vp, td);
2341 			else
2342 				vgonechrl(vp, td);
2343 			MNT_ILOCK(mp);
2344 			continue;
2345 		}
2346 #ifdef DIAGNOSTIC
2347 		if (busyprt)
2348 			vprint("vflush: busy vnode", vp);
2349 #endif
2350 		VI_UNLOCK(vp);
2351 		MNT_ILOCK(mp);
2352 		busy++;
2353 	}
2354 	MNT_IUNLOCK(mp);
2355 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2356 		/*
2357 		 * If just the root vnode is busy, and if its refcount
2358 		 * is equal to `rootrefs', then go ahead and kill it.
2359 		 */
2360 		VI_LOCK(rootvp);
2361 		KASSERT(busy > 0, ("vflush: not busy"));
2362 		KASSERT(rootvp->v_usecount >= rootrefs,
2363 		    ("vflush: usecount %d < rootrefs %d",
2364 		     rootvp->v_usecount, rootrefs));
2365 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2366 			vgonel(rootvp, td);
2367 			busy = 0;
2368 		} else
2369 			VI_UNLOCK(rootvp);
2370 	}
2371 	if (busy)
2372 		return (EBUSY);
2373 	for (; rootrefs > 0; rootrefs--)
2374 		vrele(rootvp);
2375 	return (0);
2376 }
2377 
2378 /*
2379  * This moves a now (likely recyclable) vnode to the end of the
2380  * mountlist.  XXX However, it is temporarily disabled until we
2381  * can clean up ffs_sync() and friends, which have loop restart
2382  * conditions which this code causes to operate O(N^2).
2383  */
2384 static void
2385 vlruvp(struct vnode *vp)
2386 {
2387 #if 0
2388 	struct mount *mp;
2389 
2390 	if ((mp = vp->v_mount) != NULL) {
2391 		MNT_ILOCK(mp);
2392 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2393 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2394 		MNT_IUNLOCK(mp);
2395 	}
2396 #endif
2397 }
2398 
2399 static void
2400 vx_lock(struct vnode *vp)
2401 {
2402 
2403 	ASSERT_VI_LOCKED(vp, "vx_lock");
2404 
2405 	/*
2406 	 * Prevent the vnode from being recycled or brought into use while we
2407 	 * clean it out.
2408 	 */
2409 	if (vp->v_iflag & VI_XLOCK)
2410 		panic("vclean: deadlock");
2411 	vp->v_iflag |= VI_XLOCK;
2412 	vp->v_vxthread = curthread;
2413 }
2414 
2415 static void
2416 vx_unlock(struct vnode *vp)
2417 {
2418 	ASSERT_VI_LOCKED(vp, "vx_unlock");
2419 	vp->v_iflag &= ~VI_XLOCK;
2420 	vp->v_vxthread = NULL;
2421 	if (vp->v_iflag & VI_XWANT) {
2422 		vp->v_iflag &= ~VI_XWANT;
2423 		wakeup(vp);
2424 	}
2425 }
2426 
2427 /*
2428  * Disassociate the underlying filesystem from a vnode.
2429  */
2430 static void
2431 vclean(vp, flags, td)
2432 	struct vnode *vp;
2433 	int flags;
2434 	struct thread *td;
2435 {
2436 	int active;
2437 
2438 	ASSERT_VI_LOCKED(vp, "vclean");
2439 	/*
2440 	 * Check to see if the vnode is in use. If so we have to reference it
2441 	 * before we clean it out so that its count cannot fall to zero and
2442 	 * generate a race against ourselves to recycle it.
2443 	 */
2444 	if ((active = vp->v_usecount))
2445 		v_incr_usecount(vp, 1);
2446 
2447 	/*
2448 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2449 	 * have the object locked while it cleans it out. The VOP_LOCK
2450 	 * ensures that the VOP_INACTIVE routine is done with its work.
2451 	 * For active vnodes, it ensures that no other activity can
2452 	 * occur while the underlying object is being cleaned out.
2453 	 */
2454 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2455 
2456 	/*
2457 	 * Clean out any buffers associated with the vnode.
2458 	 * If the flush fails, just toss the buffers.
2459 	 */
2460 	if (flags & DOCLOSE) {
2461 		struct buf *bp;
2462 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
2463 		if (bp != NULL)
2464 			(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2465 		if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0)
2466 			vinvalbuf(vp, 0, NOCRED, td, 0, 0);
2467 	}
2468 
2469 	VOP_DESTROYVOBJECT(vp);
2470 
2471 	/*
2472 	 * Any other processes trying to obtain this lock must first
2473 	 * wait for VXLOCK to clear, then call the new lock operation.
2474 	 */
2475 	VOP_UNLOCK(vp, 0, td);
2476 
2477 	/*
2478 	 * If purging an active vnode, it must be closed and
2479 	 * deactivated before being reclaimed. Note that the
2480 	 * VOP_INACTIVE will unlock the vnode.
2481 	 */
2482 	if (active) {
2483 		if (flags & DOCLOSE)
2484 			VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2485 		VI_LOCK(vp);
2486 		if ((vp->v_iflag & VI_DOINGINACT) == 0) {
2487 			vp->v_iflag |= VI_DOINGINACT;
2488 			VI_UNLOCK(vp);
2489 			if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2490 				panic("vclean: cannot relock.");
2491 			VOP_INACTIVE(vp, td);
2492 			VI_LOCK(vp);
2493 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2494 			    ("vclean: lost VI_DOINGINACT"));
2495 			vp->v_iflag &= ~VI_DOINGINACT;
2496 		}
2497 		VI_UNLOCK(vp);
2498 	}
2499 	/*
2500 	 * Reclaim the vnode.
2501 	 */
2502 	if (VOP_RECLAIM(vp, td))
2503 		panic("vclean: cannot reclaim");
2504 
2505 	if (active) {
2506 		/*
2507 		 * Inline copy of vrele() since VOP_INACTIVE
2508 		 * has already been called.
2509 		 */
2510 		VI_LOCK(vp);
2511 		v_incr_usecount(vp, -1);
2512 		if (vp->v_usecount <= 0) {
2513 #ifdef INVARIANTS
2514 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2515 				vprint("vclean: bad ref count", vp);
2516 				panic("vclean: ref cnt");
2517 			}
2518 #endif
2519 			if (VSHOULDFREE(vp))
2520 				vfree(vp);
2521 		}
2522 		VI_UNLOCK(vp);
2523 	}
2524 	/*
2525 	 * Delete from old mount point vnode list.
2526 	 */
2527 	delmntque(vp);
2528 	cache_purge(vp);
2529 	VI_LOCK(vp);
2530 	if (VSHOULDFREE(vp))
2531 		vfree(vp);
2532 
2533 	/*
2534 	 * Done with purge, reset to the standard lock and
2535 	 * notify sleepers of the grim news.
2536 	 */
2537 	vp->v_vnlock = &vp->v_lock;
2538 	vp->v_op = dead_vnodeop_p;
2539 	if (vp->v_pollinfo != NULL)
2540 		vn_pollgone(vp);
2541 	vp->v_tag = "none";
2542 }
2543 
2544 /*
2545  * Eliminate all activity associated with the requested vnode
2546  * and with all vnodes aliased to the requested vnode.
2547  */
2548 int
2549 vop_revoke(ap)
2550 	struct vop_revoke_args /* {
2551 		struct vnode *a_vp;
2552 		int a_flags;
2553 	} */ *ap;
2554 {
2555 	struct vnode *vp, *vq;
2556 	struct cdev *dev;
2557 
2558 	KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke"));
2559 	vp = ap->a_vp;
2560 	KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR"));
2561 
2562 	VI_LOCK(vp);
2563 	/*
2564 	 * If a vgone (or vclean) is already in progress,
2565 	 * wait until it is done and return.
2566 	 */
2567 	if (vp->v_iflag & VI_XLOCK) {
2568 		vp->v_iflag |= VI_XWANT;
2569 		msleep(vp, VI_MTX(vp), PINOD | PDROP,
2570 		    "vop_revokeall", 0);
2571 		return (0);
2572 	}
2573 	VI_UNLOCK(vp);
2574 	dev = vp->v_rdev;
2575 	for (;;) {
2576 		mtx_lock(&spechash_mtx);
2577 		vq = SLIST_FIRST(&dev->si_hlist);
2578 		mtx_unlock(&spechash_mtx);
2579 		if (vq == NULL)
2580 			break;
2581 		vgone(vq);
2582 	}
2583 	return (0);
2584 }
2585 
2586 /*
2587  * Recycle an unused vnode to the front of the free list.
2588  * Release the passed interlock if the vnode will be recycled.
2589  */
2590 int
2591 vrecycle(vp, inter_lkp, td)
2592 	struct vnode *vp;
2593 	struct mtx *inter_lkp;
2594 	struct thread *td;
2595 {
2596 
2597 	VI_LOCK(vp);
2598 	if (vp->v_usecount == 0) {
2599 		if (inter_lkp) {
2600 			mtx_unlock(inter_lkp);
2601 		}
2602 		vgonel(vp, td);
2603 		return (1);
2604 	}
2605 	VI_UNLOCK(vp);
2606 	return (0);
2607 }
2608 
2609 /*
2610  * Eliminate all activity associated with a vnode
2611  * in preparation for reuse.
2612  */
2613 void
2614 vgone(vp)
2615 	register struct vnode *vp;
2616 {
2617 	struct thread *td = curthread;	/* XXX */
2618 
2619 	VI_LOCK(vp);
2620 	vgonel(vp, td);
2621 }
2622 
2623 /*
2624  * Disassociate a character device from the its underlying filesystem and
2625  * attach it to spec.  This is for use when the chr device is still active
2626  * and the filesystem is going away.
2627  */
2628 static void
2629 vgonechrl(struct vnode *vp, struct thread *td)
2630 {
2631 	ASSERT_VI_LOCKED(vp, "vgonechrl");
2632 	vx_lock(vp);
2633 	/*
2634 	 * This is a custom version of vclean() which does not tearm down
2635 	 * the bufs or vm objects held by this vnode.  This allows filesystems
2636 	 * to continue using devices which were discovered via another
2637 	 * filesystem that has been unmounted.
2638 	 */
2639 	if (vp->v_usecount != 0) {
2640 		v_incr_usecount(vp, 1);
2641 		/*
2642 		 * Ensure that no other activity can occur while the
2643 		 * underlying object is being cleaned out.
2644 		 */
2645 		VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2646 		/*
2647 		 * Any other processes trying to obtain this lock must first
2648 		 * wait for VXLOCK to clear, then call the new lock operation.
2649 		 */
2650 		VOP_UNLOCK(vp, 0, td);
2651 		vp->v_vnlock = &vp->v_lock;
2652 		vp->v_tag = "orphanchr";
2653 		vp->v_op = spec_vnodeop_p;
2654 		delmntque(vp);
2655 		cache_purge(vp);
2656 		vrele(vp);
2657 		VI_LOCK(vp);
2658 	} else
2659 		vclean(vp, 0, td);
2660 	vp->v_op = spec_vnodeop_p;
2661 	vx_unlock(vp);
2662 	VI_UNLOCK(vp);
2663 }
2664 
2665 /*
2666  * vgone, with the vp interlock held.
2667  */
2668 void
2669 vgonel(vp, td)
2670 	struct vnode *vp;
2671 	struct thread *td;
2672 {
2673 	/*
2674 	 * If a vgone (or vclean) is already in progress,
2675 	 * wait until it is done and return.
2676 	 */
2677 	ASSERT_VI_LOCKED(vp, "vgonel");
2678 	if (vp->v_iflag & VI_XLOCK) {
2679 		vp->v_iflag |= VI_XWANT;
2680 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0);
2681 		return;
2682 	}
2683 	vx_lock(vp);
2684 
2685 	/*
2686 	 * Clean out the filesystem specific data.
2687 	 */
2688 	vclean(vp, DOCLOSE, td);
2689 	VI_UNLOCK(vp);
2690 
2691 	/*
2692 	 * If special device, remove it from special device alias list
2693 	 * if it is on one.
2694 	 */
2695 	VI_LOCK(vp);
2696 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2697 		mtx_lock(&spechash_mtx);
2698 		SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext);
2699 		vp->v_rdev->si_usecount -= vp->v_usecount;
2700 		mtx_unlock(&spechash_mtx);
2701 		dev_rel(vp->v_rdev);
2702 		vp->v_rdev = NULL;
2703 	}
2704 
2705 	/*
2706 	 * If it is on the freelist and not already at the head,
2707 	 * move it to the head of the list. The test of the
2708 	 * VDOOMED flag and the reference count of zero is because
2709 	 * it will be removed from the free list by getnewvnode,
2710 	 * but will not have its reference count incremented until
2711 	 * after calling vgone. If the reference count were
2712 	 * incremented first, vgone would (incorrectly) try to
2713 	 * close the previous instance of the underlying object.
2714 	 */
2715 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2716 		mtx_lock(&vnode_free_list_mtx);
2717 		if (vp->v_iflag & VI_FREE) {
2718 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2719 		} else {
2720 			vp->v_iflag |= VI_FREE;
2721 			freevnodes++;
2722 		}
2723 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2724 		mtx_unlock(&vnode_free_list_mtx);
2725 	}
2726 
2727 	vp->v_type = VBAD;
2728 	vx_unlock(vp);
2729 	VI_UNLOCK(vp);
2730 }
2731 
2732 /*
2733  * Lookup a vnode by device number.
2734  */
2735 int
2736 vfinddev(dev, vpp)
2737 	struct cdev *dev;
2738 	struct vnode **vpp;
2739 {
2740 	struct vnode *vp;
2741 
2742 	mtx_lock(&spechash_mtx);
2743 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2744 		*vpp = vp;
2745 		mtx_unlock(&spechash_mtx);
2746 		return (1);
2747 	}
2748 	mtx_unlock(&spechash_mtx);
2749 	return (0);
2750 }
2751 
2752 /*
2753  * Calculate the total number of references to a special device.
2754  */
2755 int
2756 vcount(vp)
2757 	struct vnode *vp;
2758 {
2759 	int count;
2760 
2761 	mtx_lock(&spechash_mtx);
2762 	count = vp->v_rdev->si_usecount;
2763 	mtx_unlock(&spechash_mtx);
2764 	return (count);
2765 }
2766 
2767 /*
2768  * Same as above, but using the struct cdev *as argument
2769  */
2770 int
2771 count_dev(dev)
2772 	struct cdev *dev;
2773 {
2774 	int count;
2775 
2776 	mtx_lock(&spechash_mtx);
2777 	count = dev->si_usecount;
2778 	mtx_unlock(&spechash_mtx);
2779 	return(count);
2780 }
2781 
2782 /*
2783  * Print out a description of a vnode.
2784  */
2785 static char *typename[] =
2786 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2787 
2788 void
2789 vprint(label, vp)
2790 	char *label;
2791 	struct vnode *vp;
2792 {
2793 	char buf[96];
2794 
2795 	if (label != NULL)
2796 		printf("%s: %p: ", label, (void *)vp);
2797 	else
2798 		printf("%p: ", (void *)vp);
2799 	printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,",
2800 	    vp->v_tag, typename[vp->v_type], vp->v_usecount,
2801 	    vp->v_writecount, vp->v_holdcnt);
2802 	buf[0] = '\0';
2803 	if (vp->v_vflag & VV_ROOT)
2804 		strcat(buf, "|VV_ROOT");
2805 	if (vp->v_vflag & VV_TEXT)
2806 		strcat(buf, "|VV_TEXT");
2807 	if (vp->v_vflag & VV_SYSTEM)
2808 		strcat(buf, "|VV_SYSTEM");
2809 	if (vp->v_iflag & VI_XLOCK)
2810 		strcat(buf, "|VI_XLOCK");
2811 	if (vp->v_iflag & VI_XWANT)
2812 		strcat(buf, "|VI_XWANT");
2813 	if (vp->v_iflag & VI_BWAIT)
2814 		strcat(buf, "|VI_BWAIT");
2815 	if (vp->v_iflag & VI_DOOMED)
2816 		strcat(buf, "|VI_DOOMED");
2817 	if (vp->v_iflag & VI_FREE)
2818 		strcat(buf, "|VI_FREE");
2819 	if (vp->v_vflag & VV_OBJBUF)
2820 		strcat(buf, "|VV_OBJBUF");
2821 	if (buf[0] != '\0')
2822 		printf(" flags (%s),", &buf[1]);
2823 	lockmgr_printinfo(vp->v_vnlock);
2824 	printf("\n");
2825 	if (vp->v_data != NULL)
2826 		VOP_PRINT(vp);
2827 }
2828 
2829 #ifdef DDB
2830 #include <ddb/ddb.h>
2831 /*
2832  * List all of the locked vnodes in the system.
2833  * Called when debugging the kernel.
2834  */
2835 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2836 {
2837 	struct mount *mp, *nmp;
2838 	struct vnode *vp;
2839 
2840 	/*
2841 	 * Note: because this is DDB, we can't obey the locking semantics
2842 	 * for these structures, which means we could catch an inconsistent
2843 	 * state and dereference a nasty pointer.  Not much to be done
2844 	 * about that.
2845 	 */
2846 	printf("Locked vnodes\n");
2847 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2848 		nmp = TAILQ_NEXT(mp, mnt_list);
2849 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2850 			if (VOP_ISLOCKED(vp, NULL))
2851 				vprint(NULL, vp);
2852 		}
2853 		nmp = TAILQ_NEXT(mp, mnt_list);
2854 	}
2855 }
2856 #endif
2857 
2858 /*
2859  * Fill in a struct xvfsconf based on a struct vfsconf.
2860  */
2861 static void
2862 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2863 {
2864 
2865 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2866 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2867 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2868 	xvfsp->vfc_flags = vfsp->vfc_flags;
2869 	/*
2870 	 * These are unused in userland, we keep them
2871 	 * to not break binary compatibility.
2872 	 */
2873 	xvfsp->vfc_vfsops = NULL;
2874 	xvfsp->vfc_next = NULL;
2875 }
2876 
2877 /*
2878  * Top level filesystem related information gathering.
2879  */
2880 static int
2881 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2882 {
2883 	struct vfsconf *vfsp;
2884 	struct xvfsconf xvfsp;
2885 	int error;
2886 
2887 	error = 0;
2888 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2889 		vfsconf2x(vfsp, &xvfsp);
2890 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2891 		if (error)
2892 			break;
2893 	}
2894 	return (error);
2895 }
2896 
2897 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2898     "S,xvfsconf", "List of all configured filesystems");
2899 
2900 #ifndef BURN_BRIDGES
2901 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2902 
2903 static int
2904 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2905 {
2906 	int *name = (int *)arg1 - 1;	/* XXX */
2907 	u_int namelen = arg2 + 1;	/* XXX */
2908 	struct vfsconf *vfsp;
2909 	struct xvfsconf xvfsp;
2910 
2911 	printf("WARNING: userland calling deprecated sysctl, "
2912 	    "please rebuild world\n");
2913 
2914 #if 1 || defined(COMPAT_PRELITE2)
2915 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2916 	if (namelen == 1)
2917 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2918 #endif
2919 
2920 	switch (name[1]) {
2921 	case VFS_MAXTYPENUM:
2922 		if (namelen != 2)
2923 			return (ENOTDIR);
2924 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2925 	case VFS_CONF:
2926 		if (namelen != 3)
2927 			return (ENOTDIR);	/* overloaded */
2928 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2929 			if (vfsp->vfc_typenum == name[2])
2930 				break;
2931 		if (vfsp == NULL)
2932 			return (EOPNOTSUPP);
2933 		vfsconf2x(vfsp, &xvfsp);
2934 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2935 	}
2936 	return (EOPNOTSUPP);
2937 }
2938 
2939 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl,
2940 	"Generic filesystem");
2941 
2942 #if 1 || defined(COMPAT_PRELITE2)
2943 
2944 static int
2945 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2946 {
2947 	int error;
2948 	struct vfsconf *vfsp;
2949 	struct ovfsconf ovfs;
2950 
2951 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2952 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2953 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2954 		ovfs.vfc_index = vfsp->vfc_typenum;
2955 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2956 		ovfs.vfc_flags = vfsp->vfc_flags;
2957 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2958 		if (error)
2959 			return error;
2960 	}
2961 	return 0;
2962 }
2963 
2964 #endif /* 1 || COMPAT_PRELITE2 */
2965 #endif /* !BURN_BRIDGES */
2966 
2967 #define KINFO_VNODESLOP		10
2968 #ifdef notyet
2969 /*
2970  * Dump vnode list (via sysctl).
2971  */
2972 /* ARGSUSED */
2973 static int
2974 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2975 {
2976 	struct xvnode *xvn;
2977 	struct thread *td = req->td;
2978 	struct mount *mp;
2979 	struct vnode *vp;
2980 	int error, len, n;
2981 
2982 	/*
2983 	 * Stale numvnodes access is not fatal here.
2984 	 */
2985 	req->lock = 0;
2986 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2987 	if (!req->oldptr)
2988 		/* Make an estimate */
2989 		return (SYSCTL_OUT(req, 0, len));
2990 
2991 	error = sysctl_wire_old_buffer(req, 0);
2992 	if (error != 0)
2993 		return (error);
2994 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2995 	n = 0;
2996 	mtx_lock(&mountlist_mtx);
2997 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2998 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2999 			continue;
3000 		MNT_ILOCK(mp);
3001 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3002 			if (n == len)
3003 				break;
3004 			vref(vp);
3005 			xvn[n].xv_size = sizeof *xvn;
3006 			xvn[n].xv_vnode = vp;
3007 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3008 			XV_COPY(usecount);
3009 			XV_COPY(writecount);
3010 			XV_COPY(holdcnt);
3011 			XV_COPY(id);
3012 			XV_COPY(mount);
3013 			XV_COPY(numoutput);
3014 			XV_COPY(type);
3015 #undef XV_COPY
3016 			xvn[n].xv_flag = vp->v_vflag;
3017 
3018 			switch (vp->v_type) {
3019 			case VREG:
3020 			case VDIR:
3021 			case VLNK:
3022 				xvn[n].xv_dev = vp->v_cachedfs;
3023 				xvn[n].xv_ino = vp->v_cachedid;
3024 				break;
3025 			case VBLK:
3026 			case VCHR:
3027 				if (vp->v_rdev == NULL) {
3028 					vrele(vp);
3029 					continue;
3030 				}
3031 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3032 				break;
3033 			case VSOCK:
3034 				xvn[n].xv_socket = vp->v_socket;
3035 				break;
3036 			case VFIFO:
3037 				xvn[n].xv_fifo = vp->v_fifoinfo;
3038 				break;
3039 			case VNON:
3040 			case VBAD:
3041 			default:
3042 				/* shouldn't happen? */
3043 				vrele(vp);
3044 				continue;
3045 			}
3046 			vrele(vp);
3047 			++n;
3048 		}
3049 		MNT_IUNLOCK(mp);
3050 		mtx_lock(&mountlist_mtx);
3051 		vfs_unbusy(mp, td);
3052 		if (n == len)
3053 			break;
3054 	}
3055 	mtx_unlock(&mountlist_mtx);
3056 
3057 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3058 	free(xvn, M_TEMP);
3059 	return (error);
3060 }
3061 
3062 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
3063 	0, 0, sysctl_vnode, "S,xvnode", "");
3064 #endif
3065 
3066 /*
3067  * Check to see if a filesystem is mounted on a block device.
3068  */
3069 int
3070 vfs_mountedon(vp)
3071 	struct vnode *vp;
3072 {
3073 
3074 	if (vp->v_rdev->si_mountpoint != NULL)
3075 		return (EBUSY);
3076 	return (0);
3077 }
3078 
3079 /*
3080  * Unmount all filesystems. The list is traversed in reverse order
3081  * of mounting to avoid dependencies.
3082  */
3083 void
3084 vfs_unmountall()
3085 {
3086 	struct mount *mp;
3087 	struct thread *td;
3088 	int error;
3089 
3090 	if (curthread != NULL)
3091 		td = curthread;
3092 	else
3093 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
3094 	/*
3095 	 * Since this only runs when rebooting, it is not interlocked.
3096 	 */
3097 	while(!TAILQ_EMPTY(&mountlist)) {
3098 		mp = TAILQ_LAST(&mountlist, mntlist);
3099 		error = dounmount(mp, MNT_FORCE, td);
3100 		if (error) {
3101 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
3102 			printf("unmount of %s failed (",
3103 			    mp->mnt_stat.f_mntonname);
3104 			if (error == EBUSY)
3105 				printf("BUSY)\n");
3106 			else
3107 				printf("%d)\n", error);
3108 		} else {
3109 			/* The unmount has removed mp from the mountlist */
3110 		}
3111 	}
3112 }
3113 
3114 /*
3115  * perform msync on all vnodes under a mount point
3116  * the mount point must be locked.
3117  */
3118 void
3119 vfs_msync(struct mount *mp, int flags)
3120 {
3121 	struct vnode *vp, *nvp;
3122 	struct vm_object *obj;
3123 	int tries;
3124 
3125 	GIANT_REQUIRED;
3126 
3127 	tries = 5;
3128 	MNT_ILOCK(mp);
3129 loop:
3130 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
3131 		if (vp->v_mount != mp) {
3132 			if (--tries > 0)
3133 				goto loop;
3134 			break;
3135 		}
3136 
3137 		VI_LOCK(vp);
3138 		if (vp->v_iflag & VI_XLOCK) {
3139 			VI_UNLOCK(vp);
3140 			continue;
3141 		}
3142 
3143 		if ((vp->v_iflag & VI_OBJDIRTY) &&
3144 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
3145 			MNT_IUNLOCK(mp);
3146 			if (!vget(vp,
3147 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3148 			    curthread)) {
3149 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3150 					vput(vp);
3151 					MNT_ILOCK(mp);
3152 					continue;
3153 				}
3154 
3155 				if (VOP_GETVOBJECT(vp, &obj) == 0) {
3156 					VM_OBJECT_LOCK(obj);
3157 					vm_object_page_clean(obj, 0, 0,
3158 					    flags == MNT_WAIT ?
3159 					    OBJPC_SYNC : OBJPC_NOSYNC);
3160 					VM_OBJECT_UNLOCK(obj);
3161 				}
3162 				vput(vp);
3163 			}
3164 			MNT_ILOCK(mp);
3165 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
3166 				if (--tries > 0)
3167 					goto loop;
3168 				break;
3169 			}
3170 		} else
3171 			VI_UNLOCK(vp);
3172 	}
3173 	MNT_IUNLOCK(mp);
3174 }
3175 
3176 /*
3177  * Create the VM object needed for VMIO and mmap support.  This
3178  * is done for all VREG files in the system.  Some filesystems might
3179  * afford the additional metadata buffering capability of the
3180  * VMIO code by making the device node be VMIO mode also.
3181  *
3182  * vp must be locked when vfs_object_create is called.
3183  */
3184 int
3185 vfs_object_create(vp, td, cred)
3186 	struct vnode *vp;
3187 	struct thread *td;
3188 	struct ucred *cred;
3189 {
3190 
3191 	GIANT_REQUIRED;
3192 	return (VOP_CREATEVOBJECT(vp, cred, td));
3193 }
3194 
3195 /*
3196  * Mark a vnode as free, putting it up for recycling.
3197  */
3198 void
3199 vfree(vp)
3200 	struct vnode *vp;
3201 {
3202 
3203 	ASSERT_VI_LOCKED(vp, "vfree");
3204 	mtx_lock(&vnode_free_list_mtx);
3205 	KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free"));
3206 	if (vp->v_iflag & VI_AGE) {
3207 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
3208 	} else {
3209 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
3210 	}
3211 	freevnodes++;
3212 	mtx_unlock(&vnode_free_list_mtx);
3213 	vp->v_iflag &= ~VI_AGE;
3214 	vp->v_iflag |= VI_FREE;
3215 }
3216 
3217 /*
3218  * Opposite of vfree() - mark a vnode as in use.
3219  */
3220 void
3221 vbusy(vp)
3222 	struct vnode *vp;
3223 {
3224 
3225 	ASSERT_VI_LOCKED(vp, "vbusy");
3226 	KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free"));
3227 
3228 	mtx_lock(&vnode_free_list_mtx);
3229 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
3230 	freevnodes--;
3231 	mtx_unlock(&vnode_free_list_mtx);
3232 
3233 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
3234 }
3235 
3236 /*
3237  * Initalize per-vnode helper structure to hold poll-related state.
3238  */
3239 void
3240 v_addpollinfo(struct vnode *vp)
3241 {
3242 
3243 	vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK);
3244 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3245 }
3246 
3247 /*
3248  * Record a process's interest in events which might happen to
3249  * a vnode.  Because poll uses the historic select-style interface
3250  * internally, this routine serves as both the ``check for any
3251  * pending events'' and the ``record my interest in future events''
3252  * functions.  (These are done together, while the lock is held,
3253  * to avoid race conditions.)
3254  */
3255 int
3256 vn_pollrecord(vp, td, events)
3257 	struct vnode *vp;
3258 	struct thread *td;
3259 	short events;
3260 {
3261 
3262 	if (vp->v_pollinfo == NULL)
3263 		v_addpollinfo(vp);
3264 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3265 	if (vp->v_pollinfo->vpi_revents & events) {
3266 		/*
3267 		 * This leaves events we are not interested
3268 		 * in available for the other process which
3269 		 * which presumably had requested them
3270 		 * (otherwise they would never have been
3271 		 * recorded).
3272 		 */
3273 		events &= vp->v_pollinfo->vpi_revents;
3274 		vp->v_pollinfo->vpi_revents &= ~events;
3275 
3276 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3277 		return events;
3278 	}
3279 	vp->v_pollinfo->vpi_events |= events;
3280 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3281 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3282 	return 0;
3283 }
3284 
3285 /*
3286  * Note the occurrence of an event.  If the VN_POLLEVENT macro is used,
3287  * it is possible for us to miss an event due to race conditions, but
3288  * that condition is expected to be rare, so for the moment it is the
3289  * preferred interface.
3290  */
3291 void
3292 vn_pollevent(vp, events)
3293 	struct vnode *vp;
3294 	short events;
3295 {
3296 
3297 	if (vp->v_pollinfo == NULL)
3298 		v_addpollinfo(vp);
3299 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3300 	if (vp->v_pollinfo->vpi_events & events) {
3301 		/*
3302 		 * We clear vpi_events so that we don't
3303 		 * call selwakeup() twice if two events are
3304 		 * posted before the polling process(es) is
3305 		 * awakened.  This also ensures that we take at
3306 		 * most one selwakeup() if the polling process
3307 		 * is no longer interested.  However, it does
3308 		 * mean that only one event can be noticed at
3309 		 * a time.  (Perhaps we should only clear those
3310 		 * event bits which we note?) XXX
3311 		 */
3312 		vp->v_pollinfo->vpi_events = 0;	/* &= ~events ??? */
3313 		vp->v_pollinfo->vpi_revents |= events;
3314 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
3315 	}
3316 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3317 }
3318 
3319 /*
3320  * Wake up anyone polling on vp because it is being revoked.
3321  * This depends on dead_poll() returning POLLHUP for correct
3322  * behavior.
3323  */
3324 void
3325 vn_pollgone(vp)
3326 	struct vnode *vp;
3327 {
3328 
3329 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3330 	VN_KNOTE(vp, NOTE_REVOKE);
3331 	if (vp->v_pollinfo->vpi_events) {
3332 		vp->v_pollinfo->vpi_events = 0;
3333 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
3334 	}
3335 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3336 }
3337 
3338 
3339 
3340 /*
3341  * Routine to create and manage a filesystem syncer vnode.
3342  */
3343 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3344 static int	sync_fsync(struct  vop_fsync_args *);
3345 static int	sync_inactive(struct  vop_inactive_args *);
3346 static int	sync_reclaim(struct  vop_reclaim_args *);
3347 
3348 static vop_t **sync_vnodeop_p;
3349 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = {
3350 	{ &vop_default_desc,	(vop_t *) vop_eopnotsupp },
3351 	{ &vop_close_desc,	(vop_t *) sync_close },		/* close */
3352 	{ &vop_fsync_desc,	(vop_t *) sync_fsync },		/* fsync */
3353 	{ &vop_inactive_desc,	(vop_t *) sync_inactive },	/* inactive */
3354 	{ &vop_reclaim_desc,	(vop_t *) sync_reclaim },	/* reclaim */
3355 	{ &vop_lock_desc,	(vop_t *) vop_stdlock },	/* lock */
3356 	{ &vop_unlock_desc,	(vop_t *) vop_stdunlock },	/* unlock */
3357 	{ &vop_islocked_desc,	(vop_t *) vop_stdislocked },	/* islocked */
3358 	{ NULL, NULL }
3359 };
3360 static struct vnodeopv_desc sync_vnodeop_opv_desc =
3361 	{ &sync_vnodeop_p, sync_vnodeop_entries };
3362 
3363 VNODEOP_SET(sync_vnodeop_opv_desc);
3364 
3365 /*
3366  * Create a new filesystem syncer vnode for the specified mount point.
3367  */
3368 int
3369 vfs_allocate_syncvnode(mp)
3370 	struct mount *mp;
3371 {
3372 	struct vnode *vp;
3373 	static long start, incr, next;
3374 	int error;
3375 
3376 	/* Allocate a new vnode */
3377 	if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) {
3378 		mp->mnt_syncer = NULL;
3379 		return (error);
3380 	}
3381 	vp->v_type = VNON;
3382 	/*
3383 	 * Place the vnode onto the syncer worklist. We attempt to
3384 	 * scatter them about on the list so that they will go off
3385 	 * at evenly distributed times even if all the filesystems
3386 	 * are mounted at once.
3387 	 */
3388 	next += incr;
3389 	if (next == 0 || next > syncer_maxdelay) {
3390 		start /= 2;
3391 		incr /= 2;
3392 		if (start == 0) {
3393 			start = syncer_maxdelay / 2;
3394 			incr = syncer_maxdelay;
3395 		}
3396 		next = start;
3397 	}
3398 	VI_LOCK(vp);
3399 	vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0);
3400 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3401 	mtx_lock(&sync_mtx);
3402 	sync_vnode_count++;
3403 	mtx_unlock(&sync_mtx);
3404 	VI_UNLOCK(vp);
3405 	mp->mnt_syncer = vp;
3406 	return (0);
3407 }
3408 
3409 /*
3410  * Do a lazy sync of the filesystem.
3411  */
3412 static int
3413 sync_fsync(ap)
3414 	struct vop_fsync_args /* {
3415 		struct vnode *a_vp;
3416 		struct ucred *a_cred;
3417 		int a_waitfor;
3418 		struct thread *a_td;
3419 	} */ *ap;
3420 {
3421 	struct vnode *syncvp = ap->a_vp;
3422 	struct mount *mp = syncvp->v_mount;
3423 	struct thread *td = ap->a_td;
3424 	int error, asyncflag;
3425 
3426 	/*
3427 	 * We only need to do something if this is a lazy evaluation.
3428 	 */
3429 	if (ap->a_waitfor != MNT_LAZY)
3430 		return (0);
3431 
3432 	/*
3433 	 * Move ourselves to the back of the sync list.
3434 	 */
3435 	VI_LOCK(syncvp);
3436 	vn_syncer_add_to_worklist(syncvp, syncdelay);
3437 	VI_UNLOCK(syncvp);
3438 
3439 	/*
3440 	 * Walk the list of vnodes pushing all that are dirty and
3441 	 * not already on the sync list.
3442 	 */
3443 	mtx_lock(&mountlist_mtx);
3444 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3445 		mtx_unlock(&mountlist_mtx);
3446 		return (0);
3447 	}
3448 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3449 		vfs_unbusy(mp, td);
3450 		return (0);
3451 	}
3452 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3453 	mp->mnt_flag &= ~MNT_ASYNC;
3454 	vfs_msync(mp, MNT_NOWAIT);
3455 	error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td);
3456 	if (asyncflag)
3457 		mp->mnt_flag |= MNT_ASYNC;
3458 	vn_finished_write(mp);
3459 	vfs_unbusy(mp, td);
3460 	return (error);
3461 }
3462 
3463 /*
3464  * The syncer vnode is no referenced.
3465  */
3466 static int
3467 sync_inactive(ap)
3468 	struct vop_inactive_args /* {
3469 		struct vnode *a_vp;
3470 		struct thread *a_td;
3471 	} */ *ap;
3472 {
3473 
3474 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3475 	vgone(ap->a_vp);
3476 	return (0);
3477 }
3478 
3479 /*
3480  * The syncer vnode is no longer needed and is being decommissioned.
3481  *
3482  * Modifications to the worklist must be protected by sync_mtx.
3483  */
3484 static int
3485 sync_reclaim(ap)
3486 	struct vop_reclaim_args /* {
3487 		struct vnode *a_vp;
3488 	} */ *ap;
3489 {
3490 	struct vnode *vp = ap->a_vp;
3491 
3492 	VI_LOCK(vp);
3493 	vp->v_mount->mnt_syncer = NULL;
3494 	if (vp->v_iflag & VI_ONWORKLST) {
3495 		mtx_lock(&sync_mtx);
3496 		LIST_REMOVE(vp, v_synclist);
3497  		syncer_worklist_len--;
3498 		sync_vnode_count--;
3499 		mtx_unlock(&sync_mtx);
3500 		vp->v_iflag &= ~VI_ONWORKLST;
3501 	}
3502 	VI_UNLOCK(vp);
3503 
3504 	return (0);
3505 }
3506 
3507 /*
3508  * extract the struct cdev *from a VCHR
3509  */
3510 struct cdev *
3511 vn_todev(vp)
3512 	struct vnode *vp;
3513 {
3514 
3515 	if (vp->v_type != VCHR)
3516 		return (NULL);
3517 	return (vp->v_rdev);
3518 }
3519 
3520 /*
3521  * Check if vnode represents a disk device
3522  */
3523 int
3524 vn_isdisk(vp, errp)
3525 	struct vnode *vp;
3526 	int *errp;
3527 {
3528 	int error;
3529 
3530 	error = 0;
3531 	if (vp->v_type != VCHR)
3532 		error = ENOTBLK;
3533 	else if (vp->v_rdev == NULL)
3534 		error = ENXIO;
3535 	else if (!(devsw(vp->v_rdev)->d_flags & D_DISK))
3536 		error = ENOTBLK;
3537 	if (errp != NULL)
3538 		*errp = error;
3539 	return (error == 0);
3540 }
3541 
3542 /*
3543  * Free data allocated by namei(); see namei(9) for details.
3544  */
3545 void
3546 NDFREE(ndp, flags)
3547      struct nameidata *ndp;
3548      const u_int flags;
3549 {
3550 
3551 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3552 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3553 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3554 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3555 	}
3556 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3557 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3558 	    ndp->ni_dvp != ndp->ni_vp)
3559 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3560 	if (!(flags & NDF_NO_DVP_RELE) &&
3561 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3562 		vrele(ndp->ni_dvp);
3563 		ndp->ni_dvp = NULL;
3564 	}
3565 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3566 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3567 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3568 	if (!(flags & NDF_NO_VP_RELE) &&
3569 	    ndp->ni_vp) {
3570 		vrele(ndp->ni_vp);
3571 		ndp->ni_vp = NULL;
3572 	}
3573 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3574 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3575 		vrele(ndp->ni_startdir);
3576 		ndp->ni_startdir = NULL;
3577 	}
3578 }
3579 
3580 /*
3581  * Common filesystem object access control check routine.  Accepts a
3582  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3583  * and optional call-by-reference privused argument allowing vaccess()
3584  * to indicate to the caller whether privilege was used to satisfy the
3585  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3586  */
3587 int
3588 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3589 	enum vtype type;
3590 	mode_t file_mode;
3591 	uid_t file_uid;
3592 	gid_t file_gid;
3593 	mode_t acc_mode;
3594 	struct ucred *cred;
3595 	int *privused;
3596 {
3597 	mode_t dac_granted;
3598 #ifdef CAPABILITIES
3599 	mode_t cap_granted;
3600 #endif
3601 
3602 	/*
3603 	 * Look for a normal, non-privileged way to access the file/directory
3604 	 * as requested.  If it exists, go with that.
3605 	 */
3606 
3607 	if (privused != NULL)
3608 		*privused = 0;
3609 
3610 	dac_granted = 0;
3611 
3612 	/* Check the owner. */
3613 	if (cred->cr_uid == file_uid) {
3614 		dac_granted |= VADMIN;
3615 		if (file_mode & S_IXUSR)
3616 			dac_granted |= VEXEC;
3617 		if (file_mode & S_IRUSR)
3618 			dac_granted |= VREAD;
3619 		if (file_mode & S_IWUSR)
3620 			dac_granted |= (VWRITE | VAPPEND);
3621 
3622 		if ((acc_mode & dac_granted) == acc_mode)
3623 			return (0);
3624 
3625 		goto privcheck;
3626 	}
3627 
3628 	/* Otherwise, check the groups (first match) */
3629 	if (groupmember(file_gid, cred)) {
3630 		if (file_mode & S_IXGRP)
3631 			dac_granted |= VEXEC;
3632 		if (file_mode & S_IRGRP)
3633 			dac_granted |= VREAD;
3634 		if (file_mode & S_IWGRP)
3635 			dac_granted |= (VWRITE | VAPPEND);
3636 
3637 		if ((acc_mode & dac_granted) == acc_mode)
3638 			return (0);
3639 
3640 		goto privcheck;
3641 	}
3642 
3643 	/* Otherwise, check everyone else. */
3644 	if (file_mode & S_IXOTH)
3645 		dac_granted |= VEXEC;
3646 	if (file_mode & S_IROTH)
3647 		dac_granted |= VREAD;
3648 	if (file_mode & S_IWOTH)
3649 		dac_granted |= (VWRITE | VAPPEND);
3650 	if ((acc_mode & dac_granted) == acc_mode)
3651 		return (0);
3652 
3653 privcheck:
3654 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3655 		/* XXX audit: privilege used */
3656 		if (privused != NULL)
3657 			*privused = 1;
3658 		return (0);
3659 	}
3660 
3661 #ifdef CAPABILITIES
3662 	/*
3663 	 * Build a capability mask to determine if the set of capabilities
3664 	 * satisfies the requirements when combined with the granted mask
3665 	 * from above.
3666 	 * For each capability, if the capability is required, bitwise
3667 	 * or the request type onto the cap_granted mask.
3668 	 */
3669 	cap_granted = 0;
3670 
3671 	if (type == VDIR) {
3672 		/*
3673 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3674 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3675 		 */
3676 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3677 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3678 			cap_granted |= VEXEC;
3679 	} else {
3680 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3681 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3682 			cap_granted |= VEXEC;
3683 	}
3684 
3685 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3686 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3687 		cap_granted |= VREAD;
3688 
3689 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3690 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3691 		cap_granted |= (VWRITE | VAPPEND);
3692 
3693 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3694 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3695 		cap_granted |= VADMIN;
3696 
3697 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3698 		/* XXX audit: privilege used */
3699 		if (privused != NULL)
3700 			*privused = 1;
3701 		return (0);
3702 	}
3703 #endif
3704 
3705 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3706 }
3707 
3708 /*
3709  * Credential check based on process requesting service, and per-attribute
3710  * permissions.
3711  */
3712 int
3713 extattr_check_cred(struct vnode *vp, int attrnamespace,
3714     struct ucred *cred, struct thread *td, int access)
3715 {
3716 
3717 	/*
3718 	 * Kernel-invoked always succeeds.
3719 	 */
3720 	if (cred == NOCRED)
3721 		return (0);
3722 
3723 	/*
3724 	 * Do not allow privileged processes in jail to directly
3725 	 * manipulate system attributes.
3726 	 *
3727 	 * XXX What capability should apply here?
3728 	 * Probably CAP_SYS_SETFFLAG.
3729 	 */
3730 	switch (attrnamespace) {
3731 	case EXTATTR_NAMESPACE_SYSTEM:
3732 		/* Potentially should be: return (EPERM); */
3733 		return (suser_cred(cred, 0));
3734 	case EXTATTR_NAMESPACE_USER:
3735 		return (VOP_ACCESS(vp, access, cred, td));
3736 	default:
3737 		return (EPERM);
3738 	}
3739 }
3740 
3741 #ifdef DEBUG_VFS_LOCKS
3742 /*
3743  * This only exists to supress warnings from unlocked specfs accesses.  It is
3744  * no longer ok to have an unlocked VFS.
3745  */
3746 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3747 
3748 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3749 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3750 
3751 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3752 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3753 
3754 int vfs_badlock_print = 1;	/* Print lock violations. */
3755 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3756 
3757 #ifdef KDB
3758 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3759 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3760 #endif
3761 
3762 static void
3763 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3764 {
3765 
3766 #ifdef KDB
3767 	if (vfs_badlock_backtrace)
3768 		kdb_backtrace();
3769 #endif
3770 	if (vfs_badlock_print)
3771 		printf("%s: %p %s\n", str, (void *)vp, msg);
3772 	if (vfs_badlock_ddb)
3773 		kdb_enter("lock violation");
3774 }
3775 
3776 void
3777 assert_vi_locked(struct vnode *vp, const char *str)
3778 {
3779 
3780 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3781 		vfs_badlock("interlock is not locked but should be", str, vp);
3782 }
3783 
3784 void
3785 assert_vi_unlocked(struct vnode *vp, const char *str)
3786 {
3787 
3788 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3789 		vfs_badlock("interlock is locked but should not be", str, vp);
3790 }
3791 
3792 void
3793 assert_vop_locked(struct vnode *vp, const char *str)
3794 {
3795 
3796 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3797 		vfs_badlock("is not locked but should be", str, vp);
3798 }
3799 
3800 void
3801 assert_vop_unlocked(struct vnode *vp, const char *str)
3802 {
3803 
3804 	if (vp && !IGNORE_LOCK(vp) &&
3805 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3806 		vfs_badlock("is locked but should not be", str, vp);
3807 }
3808 
3809 #if 0
3810 void
3811 assert_vop_elocked(struct vnode *vp, const char *str)
3812 {
3813 
3814 	if (vp && !IGNORE_LOCK(vp) &&
3815 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3816 		vfs_badlock("is not exclusive locked but should be", str, vp);
3817 }
3818 
3819 void
3820 assert_vop_elocked_other(struct vnode *vp, const char *str)
3821 {
3822 
3823 	if (vp && !IGNORE_LOCK(vp) &&
3824 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3825 		vfs_badlock("is not exclusive locked by another thread",
3826 		    str, vp);
3827 }
3828 
3829 void
3830 assert_vop_slocked(struct vnode *vp, const char *str)
3831 {
3832 
3833 	if (vp && !IGNORE_LOCK(vp) &&
3834 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3835 		vfs_badlock("is not locked shared but should be", str, vp);
3836 }
3837 #endif /* 0 */
3838 
3839 void
3840 vop_rename_pre(void *ap)
3841 {
3842 	struct vop_rename_args *a = ap;
3843 
3844 	if (a->a_tvp)
3845 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3846 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3847 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3848 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3849 
3850 	/* Check the source (from). */
3851 	if (a->a_tdvp != a->a_fdvp)
3852 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3853 	if (a->a_tvp != a->a_fvp)
3854 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3855 
3856 	/* Check the target. */
3857 	if (a->a_tvp)
3858 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3859 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3860 }
3861 
3862 void
3863 vop_strategy_pre(void *ap)
3864 {
3865 	struct vop_strategy_args *a;
3866 	struct buf *bp;
3867 
3868 	a = ap;
3869 	bp = a->a_bp;
3870 
3871 	/*
3872 	 * Cluster ops lock their component buffers but not the IO container.
3873 	 */
3874 	if ((bp->b_flags & B_CLUSTER) != 0)
3875 		return;
3876 
3877 	if (BUF_REFCNT(bp) < 1) {
3878 		if (vfs_badlock_print)
3879 			printf(
3880 			    "VOP_STRATEGY: bp is not locked but should be\n");
3881 		if (vfs_badlock_ddb)
3882 			kdb_enter("lock violation");
3883 	}
3884 }
3885 
3886 void
3887 vop_lookup_pre(void *ap)
3888 {
3889 	struct vop_lookup_args *a;
3890 	struct vnode *dvp;
3891 
3892 	a = ap;
3893 	dvp = a->a_dvp;
3894 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3895 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3896 }
3897 
3898 void
3899 vop_lookup_post(void *ap, int rc)
3900 {
3901 	struct vop_lookup_args *a;
3902 	struct componentname *cnp;
3903 	struct vnode *dvp;
3904 	struct vnode *vp;
3905 	int flags;
3906 
3907 	a = ap;
3908 	dvp = a->a_dvp;
3909 	cnp = a->a_cnp;
3910 	vp = *(a->a_vpp);
3911 	flags = cnp->cn_flags;
3912 
3913 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3914 
3915 	/*
3916 	 * If this is the last path component for this lookup and LOCKPARENT
3917 	 * is set, OR if there is an error the directory has to be locked.
3918 	 */
3919 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
3920 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
3921 	else if (rc != 0)
3922 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
3923 	else if (dvp != vp)
3924 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
3925 	if (flags & PDIRUNLOCK)
3926 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
3927 }
3928 
3929 void
3930 vop_lock_pre(void *ap)
3931 {
3932 	struct vop_lock_args *a = ap;
3933 
3934 	if ((a->a_flags & LK_INTERLOCK) == 0)
3935 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3936 	else
3937 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3938 }
3939 
3940 void
3941 vop_lock_post(void *ap, int rc)
3942 {
3943 	struct vop_lock_args *a = ap;
3944 
3945 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3946 	if (rc == 0)
3947 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3948 }
3949 
3950 void
3951 vop_unlock_pre(void *ap)
3952 {
3953 	struct vop_unlock_args *a = ap;
3954 
3955 	if (a->a_flags & LK_INTERLOCK)
3956 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3957 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3958 }
3959 
3960 void
3961 vop_unlock_post(void *ap, int rc)
3962 {
3963 	struct vop_unlock_args *a = ap;
3964 
3965 	if (a->a_flags & LK_INTERLOCK)
3966 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3967 }
3968 #endif /* DEBUG_VFS_LOCKS */
3969 
3970 static struct klist fs_klist = SLIST_HEAD_INITIALIZER(&fs_klist);
3971 
3972 void
3973 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3974 {
3975 
3976 	KNOTE(&fs_klist, event);
3977 }
3978 
3979 static int	filt_fsattach(struct knote *kn);
3980 static void	filt_fsdetach(struct knote *kn);
3981 static int	filt_fsevent(struct knote *kn, long hint);
3982 
3983 struct filterops fs_filtops =
3984 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3985 
3986 static int
3987 filt_fsattach(struct knote *kn)
3988 {
3989 
3990 	kn->kn_flags |= EV_CLEAR;
3991 	SLIST_INSERT_HEAD(&fs_klist, kn, kn_selnext);
3992 	return (0);
3993 }
3994 
3995 static void
3996 filt_fsdetach(struct knote *kn)
3997 {
3998 
3999 	SLIST_REMOVE(&fs_klist, kn, knote, kn_selnext);
4000 }
4001 
4002 static int
4003 filt_fsevent(struct knote *kn, long hint)
4004 {
4005 
4006 	kn->kn_fflags |= hint;
4007 	return (kn->kn_fflags != 0);
4008 }
4009 
4010 static int
4011 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4012 {
4013 	struct vfsidctl vc;
4014 	int error;
4015 	struct mount *mp;
4016 
4017 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4018 	if (error)
4019 		return (error);
4020 	if (vc.vc_vers != VFS_CTL_VERS1)
4021 		return (EINVAL);
4022 	mp = vfs_getvfs(&vc.vc_fsid);
4023 	if (mp == NULL)
4024 		return (ENOENT);
4025 	/* ensure that a specific sysctl goes to the right filesystem. */
4026 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4027 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4028 		return (EINVAL);
4029 	}
4030 	VCTLTOREQ(&vc, req);
4031 	return (VFS_SYSCTL(mp, vc.vc_op, req));
4032 }
4033 
4034 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
4035         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
4036