xref: /freebsd/sys/kern/vfs_subr.c (revision cec50dea12481dc578c0805c887ab2097e1c06c5)
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/reboot.h>
64 #include <sys/sleepqueue.h>
65 #include <sys/stat.h>
66 #include <sys/sysctl.h>
67 #include <sys/syslog.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 
71 #include <vm/vm.h>
72 #include <vm/vm_object.h>
73 #include <vm/vm_extern.h>
74 #include <vm/pmap.h>
75 #include <vm/vm_map.h>
76 #include <vm/vm_page.h>
77 #include <vm/vm_kern.h>
78 #include <vm/uma.h>
79 
80 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
81 
82 static void	addalias(struct vnode *vp, struct cdev *nvp_rdev);
83 static void	delmntque(struct vnode *vp);
84 static void	insmntque(struct vnode *vp, struct mount *mp);
85 static void	vclean(struct vnode *vp, int flags, struct thread *td);
86 static void	vlruvp(struct vnode *vp);
87 static int	flushbuflist(struct buf *blist, int flags, struct vnode *vp,
88 		    int slpflag, int slptimeo, int *errorp);
89 static void	syncer_shutdown(void *arg, int howto);
90 static int	vtryrecycle(struct vnode *vp);
91 static void	vx_lock(struct vnode *vp);
92 static void	vx_unlock(struct vnode *vp);
93 static void	vgonechrl(struct vnode *vp, struct thread *td);
94 
95 
96 /*
97  * Number of vnodes in existence.  Increased whenever getnewvnode()
98  * allocates a new vnode, never decreased.
99  */
100 static unsigned long	numvnodes;
101 
102 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
103 
104 /*
105  * Conversion tables for conversion from vnode types to inode formats
106  * and back.
107  */
108 enum vtype iftovt_tab[16] = {
109 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
110 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
111 };
112 int vttoif_tab[9] = {
113 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
114 	S_IFSOCK, S_IFIFO, S_IFMT,
115 };
116 
117 /*
118  * List of vnodes that are ready for recycling.
119  */
120 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
121 
122 /*
123  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
124  * getnewvnode() will return a newly allocated vnode.
125  */
126 static u_long wantfreevnodes = 25;
127 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
128 /* Number of vnodes in the free list. */
129 static u_long freevnodes;
130 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
131 
132 /*
133  * Various variables used for debugging the new implementation of
134  * reassignbuf().
135  * XXX these are probably of (very) limited utility now.
136  */
137 static int reassignbufcalls;
138 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
139 static int nameileafonly;
140 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
141 
142 /*
143  * Cache for the mount type id assigned to NFS.  This is used for
144  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
145  */
146 int	nfs_mount_type = -1;
147 
148 /* To keep more than one thread at a time from running vfs_getnewfsid */
149 static struct mtx mntid_mtx;
150 
151 /*
152  * Lock for any access to the following:
153  *	vnode_free_list
154  *	numvnodes
155  *	freevnodes
156  */
157 static struct mtx vnode_free_list_mtx;
158 
159 /*
160  * For any iteration/modification of dev->si_hlist (linked through
161  * v_specnext)
162  */
163 static struct mtx spechash_mtx;
164 
165 /* Publicly exported FS */
166 struct nfs_public nfs_pub;
167 
168 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
169 static uma_zone_t vnode_zone;
170 static uma_zone_t vnodepoll_zone;
171 
172 /* Set to 1 to print out reclaim of active vnodes */
173 int	prtactive;
174 
175 /*
176  * The workitem queue.
177  *
178  * It is useful to delay writes of file data and filesystem metadata
179  * for tens of seconds so that quickly created and deleted files need
180  * not waste disk bandwidth being created and removed. To realize this,
181  * we append vnodes to a "workitem" queue. When running with a soft
182  * updates implementation, most pending metadata dependencies should
183  * not wait for more than a few seconds. Thus, mounted on block devices
184  * are delayed only about a half the time that file data is delayed.
185  * Similarly, directory updates are more critical, so are only delayed
186  * about a third the time that file data is delayed. Thus, there are
187  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
188  * one each second (driven off the filesystem syncer process). The
189  * syncer_delayno variable indicates the next queue that is to be processed.
190  * Items that need to be processed soon are placed in this queue:
191  *
192  *	syncer_workitem_pending[syncer_delayno]
193  *
194  * A delay of fifteen seconds is done by placing the request fifteen
195  * entries later in the queue:
196  *
197  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
198  *
199  */
200 static int syncer_delayno;
201 static long syncer_mask;
202 LIST_HEAD(synclist, vnode);
203 static struct synclist *syncer_workitem_pending;
204 /*
205  * The sync_mtx protects:
206  *	vp->v_synclist
207  *	sync_vnode_count
208  *	syncer_delayno
209  *	syncer_state
210  *	syncer_workitem_pending
211  *	syncer_worklist_len
212  *	rushjob
213  */
214 static struct mtx sync_mtx;
215 
216 #define SYNCER_MAXDELAY		32
217 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
218 static int syncdelay = 30;		/* max time to delay syncing data */
219 static int filedelay = 30;		/* time to delay syncing files */
220 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
221 static int dirdelay = 29;		/* time to delay syncing directories */
222 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
223 static int metadelay = 28;		/* time to delay syncing metadata */
224 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
225 static int rushjob;		/* number of slots to run ASAP */
226 static int stat_rush_requests;	/* number of times I/O speeded up */
227 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
228 
229 /*
230  * When shutting down the syncer, run it at four times normal speed.
231  */
232 #define SYNCER_SHUTDOWN_SPEEDUP		4
233 static int sync_vnode_count;
234 static int syncer_worklist_len;
235 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
236     syncer_state;
237 
238 /*
239  * Number of vnodes we want to exist at any one time.  This is mostly used
240  * to size hash tables in vnode-related code.  It is normally not used in
241  * getnewvnode(), as wantfreevnodes is normally nonzero.)
242  *
243  * XXX desiredvnodes is historical cruft and should not exist.
244  */
245 int desiredvnodes;
246 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
247     &desiredvnodes, 0, "Maximum number of vnodes");
248 static int minvnodes;
249 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
250     &minvnodes, 0, "Minimum number of vnodes");
251 static int vnlru_nowhere;
252 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
253     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
254 
255 /* Hook for calling soft updates. */
256 int (*softdep_process_worklist_hook)(struct mount *);
257 
258 /*
259  * Initialize the vnode management data structures.
260  */
261 #ifndef	MAXVNODES_MAX
262 #define	MAXVNODES_MAX	100000
263 #endif
264 static void
265 vntblinit(void *dummy __unused)
266 {
267 
268 	/*
269 	 * Desiredvnodes is a function of the physical memory size and
270 	 * the kernel's heap size.  Specifically, desiredvnodes scales
271 	 * in proportion to the physical memory size until two fifths
272 	 * of the kernel's heap size is consumed by vnodes and vm
273 	 * objects.
274 	 */
275 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
276 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
277 	if (desiredvnodes > MAXVNODES_MAX) {
278 		if (bootverbose)
279 			printf("Reducing kern.maxvnodes %d -> %d\n",
280 			    desiredvnodes, MAXVNODES_MAX);
281 		desiredvnodes = MAXVNODES_MAX;
282 	}
283 	minvnodes = desiredvnodes / 4;
284 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
285 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
286 	mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF);
287 	TAILQ_INIT(&vnode_free_list);
288 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
289 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
290 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
291 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
292 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
293 	/*
294 	 * Initialize the filesystem syncer.
295 	 */
296 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
297 		&syncer_mask);
298 	syncer_maxdelay = syncer_mask + 1;
299 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
300 }
301 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
302 
303 
304 /*
305  * Mark a mount point as busy. Used to synchronize access and to delay
306  * unmounting. Interlock is not released on failure.
307  */
308 int
309 vfs_busy(mp, flags, interlkp, td)
310 	struct mount *mp;
311 	int flags;
312 	struct mtx *interlkp;
313 	struct thread *td;
314 {
315 	int lkflags;
316 
317 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
318 		if (flags & LK_NOWAIT)
319 			return (ENOENT);
320 		mp->mnt_kern_flag |= MNTK_MWAIT;
321 		/*
322 		 * Since all busy locks are shared except the exclusive
323 		 * lock granted when unmounting, the only place that a
324 		 * wakeup needs to be done is at the release of the
325 		 * exclusive lock at the end of dounmount.
326 		 */
327 		msleep(mp, interlkp, PVFS, "vfs_busy", 0);
328 		return (ENOENT);
329 	}
330 	lkflags = LK_SHARED | LK_NOPAUSE;
331 	if (interlkp)
332 		lkflags |= LK_INTERLOCK;
333 	if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td))
334 		panic("vfs_busy: unexpected lock failure");
335 	return (0);
336 }
337 
338 /*
339  * Free a busy filesystem.
340  */
341 void
342 vfs_unbusy(mp, td)
343 	struct mount *mp;
344 	struct thread *td;
345 {
346 
347 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
348 }
349 
350 /*
351  * Lookup a mount point by filesystem identifier.
352  */
353 struct mount *
354 vfs_getvfs(fsid)
355 	fsid_t *fsid;
356 {
357 	register struct mount *mp;
358 
359 	mtx_lock(&mountlist_mtx);
360 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
361 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
362 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
363 			mtx_unlock(&mountlist_mtx);
364 			return (mp);
365 		}
366 	}
367 	mtx_unlock(&mountlist_mtx);
368 	return ((struct mount *) 0);
369 }
370 
371 /*
372  * Check if a user can access priveledged mount options.
373  */
374 int
375 vfs_suser(struct mount *mp, struct thread *td)
376 {
377 	int error;
378 
379 	if ((mp->mnt_flag & MNT_USER) == 0 ||
380 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
381 		if ((error = suser(td)) != 0)
382 			return (error);
383 	}
384 	return (0);
385 }
386 
387 /*
388  * Get a new unique fsid.  Try to make its val[0] unique, since this value
389  * will be used to create fake device numbers for stat().  Also try (but
390  * not so hard) make its val[0] unique mod 2^16, since some emulators only
391  * support 16-bit device numbers.  We end up with unique val[0]'s for the
392  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
393  *
394  * Keep in mind that several mounts may be running in parallel.  Starting
395  * the search one past where the previous search terminated is both a
396  * micro-optimization and a defense against returning the same fsid to
397  * different mounts.
398  */
399 void
400 vfs_getnewfsid(mp)
401 	struct mount *mp;
402 {
403 	static u_int16_t mntid_base;
404 	fsid_t tfsid;
405 	int mtype;
406 
407 	mtx_lock(&mntid_mtx);
408 	mtype = mp->mnt_vfc->vfc_typenum;
409 	tfsid.val[1] = mtype;
410 	mtype = (mtype & 0xFF) << 24;
411 	for (;;) {
412 		tfsid.val[0] = makedev(255,
413 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
414 		mntid_base++;
415 		if (vfs_getvfs(&tfsid) == NULL)
416 			break;
417 	}
418 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
419 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
420 	mtx_unlock(&mntid_mtx);
421 }
422 
423 /*
424  * Knob to control the precision of file timestamps:
425  *
426  *   0 = seconds only; nanoseconds zeroed.
427  *   1 = seconds and nanoseconds, accurate within 1/HZ.
428  *   2 = seconds and nanoseconds, truncated to microseconds.
429  * >=3 = seconds and nanoseconds, maximum precision.
430  */
431 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
432 
433 static int timestamp_precision = TSP_SEC;
434 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
435     &timestamp_precision, 0, "");
436 
437 /*
438  * Get a current timestamp.
439  */
440 void
441 vfs_timestamp(tsp)
442 	struct timespec *tsp;
443 {
444 	struct timeval tv;
445 
446 	switch (timestamp_precision) {
447 	case TSP_SEC:
448 		tsp->tv_sec = time_second;
449 		tsp->tv_nsec = 0;
450 		break;
451 	case TSP_HZ:
452 		getnanotime(tsp);
453 		break;
454 	case TSP_USEC:
455 		microtime(&tv);
456 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
457 		break;
458 	case TSP_NSEC:
459 	default:
460 		nanotime(tsp);
461 		break;
462 	}
463 }
464 
465 /*
466  * Set vnode attributes to VNOVAL
467  */
468 void
469 vattr_null(vap)
470 	register struct vattr *vap;
471 {
472 
473 	vap->va_type = VNON;
474 	vap->va_size = VNOVAL;
475 	vap->va_bytes = VNOVAL;
476 	vap->va_mode = VNOVAL;
477 	vap->va_nlink = VNOVAL;
478 	vap->va_uid = VNOVAL;
479 	vap->va_gid = VNOVAL;
480 	vap->va_fsid = VNOVAL;
481 	vap->va_fileid = VNOVAL;
482 	vap->va_blocksize = VNOVAL;
483 	vap->va_rdev = VNOVAL;
484 	vap->va_atime.tv_sec = VNOVAL;
485 	vap->va_atime.tv_nsec = VNOVAL;
486 	vap->va_mtime.tv_sec = VNOVAL;
487 	vap->va_mtime.tv_nsec = VNOVAL;
488 	vap->va_ctime.tv_sec = VNOVAL;
489 	vap->va_ctime.tv_nsec = VNOVAL;
490 	vap->va_birthtime.tv_sec = VNOVAL;
491 	vap->va_birthtime.tv_nsec = VNOVAL;
492 	vap->va_flags = VNOVAL;
493 	vap->va_gen = VNOVAL;
494 	vap->va_vaflags = 0;
495 }
496 
497 /*
498  * This routine is called when we have too many vnodes.  It attempts
499  * to free <count> vnodes and will potentially free vnodes that still
500  * have VM backing store (VM backing store is typically the cause
501  * of a vnode blowout so we want to do this).  Therefore, this operation
502  * is not considered cheap.
503  *
504  * A number of conditions may prevent a vnode from being reclaimed.
505  * the buffer cache may have references on the vnode, a directory
506  * vnode may still have references due to the namei cache representing
507  * underlying files, or the vnode may be in active use.   It is not
508  * desireable to reuse such vnodes.  These conditions may cause the
509  * number of vnodes to reach some minimum value regardless of what
510  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
511  */
512 static int
513 vlrureclaim(struct mount *mp)
514 {
515 	struct vnode *vp;
516 	int done;
517 	int trigger;
518 	int usevnodes;
519 	int count;
520 
521 	/*
522 	 * Calculate the trigger point, don't allow user
523 	 * screwups to blow us up.   This prevents us from
524 	 * recycling vnodes with lots of resident pages.  We
525 	 * aren't trying to free memory, we are trying to
526 	 * free vnodes.
527 	 */
528 	usevnodes = desiredvnodes;
529 	if (usevnodes <= 0)
530 		usevnodes = 1;
531 	trigger = cnt.v_page_count * 2 / usevnodes;
532 
533 	done = 0;
534 	MNT_ILOCK(mp);
535 	count = mp->mnt_nvnodelistsize / 10 + 1;
536 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
537 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
538 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
539 
540 		if (vp->v_type != VNON &&
541 		    vp->v_type != VBAD &&
542 		    VI_TRYLOCK(vp)) {
543 			if (VMIGHTFREE(vp) &&           /* critical path opt */
544 			    (vp->v_object == NULL ||
545 			    vp->v_object->resident_page_count < trigger)) {
546 				MNT_IUNLOCK(mp);
547 				vgonel(vp, curthread);
548 				done++;
549 				MNT_ILOCK(mp);
550 			} else
551 				VI_UNLOCK(vp);
552 		}
553 		--count;
554 	}
555 	MNT_IUNLOCK(mp);
556 	return done;
557 }
558 
559 /*
560  * Attempt to recycle vnodes in a context that is always safe to block.
561  * Calling vlrurecycle() from the bowels of filesystem code has some
562  * interesting deadlock problems.
563  */
564 static struct proc *vnlruproc;
565 static int vnlruproc_sig;
566 
567 static void
568 vnlru_proc(void)
569 {
570 	struct mount *mp, *nmp;
571 	int done;
572 	struct proc *p = vnlruproc;
573 	struct thread *td = FIRST_THREAD_IN_PROC(p);
574 
575 	mtx_lock(&Giant);
576 
577 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
578 	    SHUTDOWN_PRI_FIRST);
579 
580 	for (;;) {
581 		kthread_suspend_check(p);
582 		mtx_lock(&vnode_free_list_mtx);
583 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
584 			mtx_unlock(&vnode_free_list_mtx);
585 			vnlruproc_sig = 0;
586 			wakeup(&vnlruproc_sig);
587 			tsleep(vnlruproc, PVFS, "vlruwt", hz);
588 			continue;
589 		}
590 		mtx_unlock(&vnode_free_list_mtx);
591 		done = 0;
592 		mtx_lock(&mountlist_mtx);
593 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
594 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
595 				nmp = TAILQ_NEXT(mp, mnt_list);
596 				continue;
597 			}
598 			done += vlrureclaim(mp);
599 			mtx_lock(&mountlist_mtx);
600 			nmp = TAILQ_NEXT(mp, mnt_list);
601 			vfs_unbusy(mp, td);
602 		}
603 		mtx_unlock(&mountlist_mtx);
604 		if (done == 0) {
605 #if 0
606 			/* These messages are temporary debugging aids */
607 			if (vnlru_nowhere < 5)
608 				printf("vnlru process getting nowhere..\n");
609 			else if (vnlru_nowhere == 5)
610 				printf("vnlru process messages stopped.\n");
611 #endif
612 			vnlru_nowhere++;
613 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
614 		}
615 	}
616 }
617 
618 static struct kproc_desc vnlru_kp = {
619 	"vnlru",
620 	vnlru_proc,
621 	&vnlruproc
622 };
623 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
624 
625 
626 /*
627  * Routines having to do with the management of the vnode table.
628  */
629 
630 /*
631  * Check to see if a free vnode can be recycled. If it can,
632  * recycle it and return it with the vnode interlock held.
633  */
634 static int
635 vtryrecycle(struct vnode *vp)
636 {
637 	struct thread *td = curthread;
638 	vm_object_t object;
639 	struct mount *vnmp;
640 	int error;
641 
642 	/* Don't recycle if we can't get the interlock */
643 	if (!VI_TRYLOCK(vp))
644 		return (EWOULDBLOCK);
645 	/*
646 	 * This vnode may found and locked via some other list, if so we
647 	 * can't recycle it yet.
648 	 */
649 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
650 		return (EWOULDBLOCK);
651 	/*
652 	 * Don't recycle if its filesystem is being suspended.
653 	 */
654 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
655 		VOP_UNLOCK(vp, 0, td);
656 		return (EBUSY);
657 	}
658 
659 	/*
660 	 * Don't recycle if we still have cached pages.
661 	 */
662 	if (VOP_GETVOBJECT(vp, &object) == 0) {
663 		VM_OBJECT_LOCK(object);
664 		if (object->resident_page_count ||
665 		    object->ref_count) {
666 			VM_OBJECT_UNLOCK(object);
667 			error = EBUSY;
668 			goto done;
669 		}
670 		VM_OBJECT_UNLOCK(object);
671 	}
672 	if (LIST_FIRST(&vp->v_cache_src)) {
673 		/*
674 		 * note: nameileafonly sysctl is temporary,
675 		 * for debugging only, and will eventually be
676 		 * removed.
677 		 */
678 		if (nameileafonly > 0) {
679 			/*
680 			 * Do not reuse namei-cached directory
681 			 * vnodes that have cached
682 			 * subdirectories.
683 			 */
684 			if (cache_leaf_test(vp) < 0) {
685 				error = EISDIR;
686 				goto done;
687 			}
688 		} else if (nameileafonly < 0 ||
689 			    vmiodirenable == 0) {
690 			/*
691 			 * Do not reuse namei-cached directory
692 			 * vnodes if nameileafonly is -1 or
693 			 * if VMIO backing for directories is
694 			 * turned off (otherwise we reuse them
695 			 * too quickly).
696 			 */
697 			error = EBUSY;
698 			goto done;
699 		}
700 	}
701 	/*
702 	 * If we got this far, we need to acquire the interlock and see if
703 	 * anyone picked up this vnode from another list.  If not, we will
704 	 * mark it with XLOCK via vgonel() so that anyone who does find it
705 	 * will skip over it.
706 	 */
707 	VI_LOCK(vp);
708 	if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) {
709 		VI_UNLOCK(vp);
710 		error = EBUSY;
711 		goto done;
712 	}
713 	mtx_lock(&vnode_free_list_mtx);
714 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
715 	vp->v_iflag &= ~VI_FREE;
716 	mtx_unlock(&vnode_free_list_mtx);
717 	vp->v_iflag |= VI_DOOMED;
718 	if (vp->v_type != VBAD) {
719 		VOP_UNLOCK(vp, 0, td);
720 		vgonel(vp, td);
721 		VI_LOCK(vp);
722 	} else
723 		VOP_UNLOCK(vp, 0, td);
724 	vn_finished_write(vnmp);
725 	return (0);
726 done:
727 	VOP_UNLOCK(vp, 0, td);
728 	vn_finished_write(vnmp);
729 	return (error);
730 }
731 
732 /*
733  * Return the next vnode from the free list.
734  */
735 int
736 getnewvnode(tag, mp, vops, vpp)
737 	const char *tag;
738 	struct mount *mp;
739 	vop_t **vops;
740 	struct vnode **vpp;
741 {
742 	struct vnode *vp = NULL;
743 	struct vpollinfo *pollinfo = NULL;
744 
745 	mtx_lock(&vnode_free_list_mtx);
746 
747 	/*
748 	 * Try to reuse vnodes if we hit the max.  This situation only
749 	 * occurs in certain large-memory (2G+) situations.  We cannot
750 	 * attempt to directly reclaim vnodes due to nasty recursion
751 	 * problems.
752 	 */
753 	while (numvnodes - freevnodes > desiredvnodes) {
754 		if (vnlruproc_sig == 0) {
755 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
756 			wakeup(vnlruproc);
757 		}
758 		mtx_unlock(&vnode_free_list_mtx);
759 		tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz);
760 		mtx_lock(&vnode_free_list_mtx);
761 	}
762 
763 	/*
764 	 * Attempt to reuse a vnode already on the free list, allocating
765 	 * a new vnode if we can't find one or if we have not reached a
766 	 * good minimum for good LRU performance.
767 	 */
768 
769 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
770 		int error;
771 		int count;
772 
773 		for (count = 0; count < freevnodes; count++) {
774 			vp = TAILQ_FIRST(&vnode_free_list);
775 
776 			KASSERT(vp->v_usecount == 0 &&
777 			    (vp->v_iflag & VI_DOINGINACT) == 0,
778 			    ("getnewvnode: free vnode isn't"));
779 
780 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
781 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
782 			mtx_unlock(&vnode_free_list_mtx);
783 			error = vtryrecycle(vp);
784 			mtx_lock(&vnode_free_list_mtx);
785 			if (error == 0)
786 				break;
787 			vp = NULL;
788 		}
789 	}
790 	if (vp) {
791 		freevnodes--;
792 		mtx_unlock(&vnode_free_list_mtx);
793 
794 #ifdef INVARIANTS
795 		{
796 			if (vp->v_data)
797 				panic("cleaned vnode isn't");
798 			if (vp->v_numoutput)
799 				panic("Clean vnode has pending I/O's");
800 			if (vp->v_writecount != 0)
801 				panic("Non-zero write count");
802 		}
803 #endif
804 		if ((pollinfo = vp->v_pollinfo) != NULL) {
805 			/*
806 			 * To avoid lock order reversals, the call to
807 			 * uma_zfree() must be delayed until the vnode
808 			 * interlock is released.
809 			 */
810 			vp->v_pollinfo = NULL;
811 		}
812 #ifdef MAC
813 		mac_destroy_vnode(vp);
814 #endif
815 		vp->v_iflag = 0;
816 		vp->v_vflag = 0;
817 		vp->v_lastw = 0;
818 		vp->v_lasta = 0;
819 		vp->v_cstart = 0;
820 		vp->v_clen = 0;
821 		vp->v_socket = 0;
822 		lockdestroy(vp->v_vnlock);
823 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
824 		KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0"));
825 		KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL"));
826 		KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0"));
827 		KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL"));
828 	} else {
829 		numvnodes++;
830 		mtx_unlock(&vnode_free_list_mtx);
831 
832 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
833 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
834 		VI_LOCK(vp);
835 		vp->v_dd = vp;
836 		vp->v_vnlock = &vp->v_lock;
837 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
838 		cache_purge(vp);		/* Sets up v_id. */
839 		LIST_INIT(&vp->v_cache_src);
840 		TAILQ_INIT(&vp->v_cache_dst);
841 	}
842 
843 	TAILQ_INIT(&vp->v_cleanblkhd);
844 	TAILQ_INIT(&vp->v_dirtyblkhd);
845 	vp->v_type = VNON;
846 	vp->v_tag = tag;
847 	vp->v_op = vops;
848 	*vpp = vp;
849 	vp->v_usecount = 1;
850 	vp->v_data = 0;
851 	vp->v_cachedid = -1;
852 	VI_UNLOCK(vp);
853 	if (pollinfo != NULL) {
854 		knlist_destroy(&pollinfo->vpi_selinfo.si_note);
855 		mtx_destroy(&pollinfo->vpi_lock);
856 		uma_zfree(vnodepoll_zone, pollinfo);
857 	}
858 #ifdef MAC
859 	mac_init_vnode(vp);
860 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
861 		mac_associate_vnode_singlelabel(mp, vp);
862 #endif
863 	delmntque(vp);
864 	if (mp != NULL) {
865 		insmntque(vp, mp);
866 		vp->v_bsize = mp->mnt_stat.f_iosize;
867 	}
868 
869 	return (0);
870 }
871 
872 /*
873  * Delete from old mount point vnode list, if on one.
874  */
875 static void
876 delmntque(struct vnode *vp)
877 {
878 	struct mount *mp;
879 
880 	if (vp->v_mount == NULL)
881 		return;
882 	mp = vp->v_mount;
883 	MNT_ILOCK(mp);
884 	vp->v_mount = NULL;
885 	KASSERT(mp->mnt_nvnodelistsize > 0,
886 		("bad mount point vnode list size"));
887 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
888 	mp->mnt_nvnodelistsize--;
889 	MNT_IUNLOCK(mp);
890 }
891 
892 /*
893  * Insert into list of vnodes for the new mount point, if available.
894  */
895 static void
896 insmntque(struct vnode *vp, struct mount *mp)
897 {
898 
899 	vp->v_mount = mp;
900 	KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)"));
901 	MNT_ILOCK(vp->v_mount);
902 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
903 	mp->mnt_nvnodelistsize++;
904 	MNT_IUNLOCK(vp->v_mount);
905 }
906 
907 /*
908  * Update outstanding I/O count and do wakeup if requested.
909  */
910 void
911 vwakeup(bp)
912 	register struct buf *bp;
913 {
914 	register struct vnode *vp;
915 
916 	if ((vp = bp->b_vp)) {
917 		VI_LOCK(vp);
918 		vp->v_numoutput--;
919 		if (vp->v_numoutput < 0)
920 			panic("vwakeup: neg numoutput");
921 		if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) {
922 			vp->v_iflag &= ~VI_BWAIT;
923 			wakeup(&vp->v_numoutput);
924 		}
925 		VI_UNLOCK(vp);
926 	}
927 }
928 
929 /*
930  * Flush out and invalidate all buffers associated with a vnode.
931  * Called with the underlying object locked.
932  */
933 int
934 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo)
935 	struct vnode *vp;
936 	int flags;
937 	struct ucred *cred;
938 	struct thread *td;
939 	int slpflag, slptimeo;
940 {
941 	struct buf *blist;
942 	int error;
943 	vm_object_t object;
944 
945 	GIANT_REQUIRED;
946 
947 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
948 
949 	VI_LOCK(vp);
950 	if (flags & V_SAVE) {
951 		while (vp->v_numoutput) {
952 			vp->v_iflag |= VI_BWAIT;
953 			error = msleep(&vp->v_numoutput, VI_MTX(vp),
954 			    slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo);
955 			if (error) {
956 				VI_UNLOCK(vp);
957 				return (error);
958 			}
959 		}
960 		if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
961 			VI_UNLOCK(vp);
962 			if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0)
963 				return (error);
964 			/*
965 			 * XXX We could save a lock/unlock if this was only
966 			 * enabled under INVARIANTS
967 			 */
968 			VI_LOCK(vp);
969 			if (vp->v_numoutput > 0 ||
970 			    !TAILQ_EMPTY(&vp->v_dirtyblkhd))
971 				panic("vinvalbuf: dirty bufs");
972 		}
973 	}
974 	/*
975 	 * If you alter this loop please notice that interlock is dropped and
976 	 * reacquired in flushbuflist.  Special care is needed to ensure that
977 	 * no race conditions occur from this.
978 	 */
979 	for (error = 0;;) {
980 		if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 &&
981 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
982 			if (error)
983 				break;
984 			continue;
985 		}
986 		if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 &&
987 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
988 			if (error)
989 				break;
990 			continue;
991 		}
992 		break;
993 	}
994 	if (error) {
995 		VI_UNLOCK(vp);
996 		return (error);
997 	}
998 
999 	/*
1000 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1001 	 * have write I/O in-progress but if there is a VM object then the
1002 	 * VM object can also have read-I/O in-progress.
1003 	 */
1004 	do {
1005 		while (vp->v_numoutput > 0) {
1006 			vp->v_iflag |= VI_BWAIT;
1007 			msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0);
1008 		}
1009 		VI_UNLOCK(vp);
1010 		if (VOP_GETVOBJECT(vp, &object) == 0) {
1011 			VM_OBJECT_LOCK(object);
1012 			vm_object_pip_wait(object, "vnvlbx");
1013 			VM_OBJECT_UNLOCK(object);
1014 		}
1015 		VI_LOCK(vp);
1016 	} while (vp->v_numoutput > 0);
1017 	VI_UNLOCK(vp);
1018 
1019 	/*
1020 	 * Destroy the copy in the VM cache, too.
1021 	 */
1022 	if (VOP_GETVOBJECT(vp, &object) == 0) {
1023 		VM_OBJECT_LOCK(object);
1024 		vm_object_page_remove(object, 0, 0,
1025 			(flags & V_SAVE) ? TRUE : FALSE);
1026 		VM_OBJECT_UNLOCK(object);
1027 	}
1028 
1029 #ifdef INVARIANTS
1030 	VI_LOCK(vp);
1031 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1032 	    (!TAILQ_EMPTY(&vp->v_dirtyblkhd) ||
1033 	     !TAILQ_EMPTY(&vp->v_cleanblkhd)))
1034 		panic("vinvalbuf: flush failed");
1035 	VI_UNLOCK(vp);
1036 #endif
1037 	return (0);
1038 }
1039 
1040 /*
1041  * Flush out buffers on the specified list.
1042  *
1043  */
1044 static int
1045 flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp)
1046 	struct buf *blist;
1047 	int flags;
1048 	struct vnode *vp;
1049 	int slpflag, slptimeo;
1050 	int *errorp;
1051 {
1052 	struct buf *bp, *nbp;
1053 	int found, error;
1054 
1055 	ASSERT_VI_LOCKED(vp, "flushbuflist");
1056 
1057 	for (found = 0, bp = blist; bp; bp = nbp) {
1058 		nbp = TAILQ_NEXT(bp, b_vnbufs);
1059 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1060 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1061 			continue;
1062 		}
1063 		found += 1;
1064 		error = BUF_TIMELOCK(bp,
1065 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp),
1066 		    "flushbuf", slpflag, slptimeo);
1067 		if (error) {
1068 			if (error != ENOLCK)
1069 				*errorp = error;
1070 			goto done;
1071 		}
1072 		/*
1073 		 * XXX Since there are no node locks for NFS, I
1074 		 * believe there is a slight chance that a delayed
1075 		 * write will occur while sleeping just above, so
1076 		 * check for it.  Note that vfs_bio_awrite expects
1077 		 * buffers to reside on a queue, while bwrite and
1078 		 * brelse do not.
1079 		 */
1080 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1081 			(flags & V_SAVE)) {
1082 
1083 			if (bp->b_vp == vp) {
1084 				if (bp->b_flags & B_CLUSTEROK) {
1085 					vfs_bio_awrite(bp);
1086 				} else {
1087 					bremfree(bp);
1088 					bp->b_flags |= B_ASYNC;
1089 					bwrite(bp);
1090 				}
1091 			} else {
1092 				bremfree(bp);
1093 				(void) bwrite(bp);
1094 			}
1095 			goto done;
1096 		}
1097 		bremfree(bp);
1098 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1099 		bp->b_flags &= ~B_ASYNC;
1100 		brelse(bp);
1101 		VI_LOCK(vp);
1102 	}
1103 	return (found);
1104 done:
1105 	VI_LOCK(vp);
1106 	return (found);
1107 }
1108 
1109 /*
1110  * Truncate a file's buffer and pages to a specified length.  This
1111  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1112  * sync activity.
1113  */
1114 int
1115 vtruncbuf(vp, cred, td, length, blksize)
1116 	register struct vnode *vp;
1117 	struct ucred *cred;
1118 	struct thread *td;
1119 	off_t length;
1120 	int blksize;
1121 {
1122 	register struct buf *bp;
1123 	struct buf *nbp;
1124 	int anyfreed;
1125 	int trunclbn;
1126 
1127 	/*
1128 	 * Round up to the *next* lbn.
1129 	 */
1130 	trunclbn = (length + blksize - 1) / blksize;
1131 
1132 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1133 restart:
1134 	VI_LOCK(vp);
1135 	anyfreed = 1;
1136 	for (;anyfreed;) {
1137 		anyfreed = 0;
1138 		for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) {
1139 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1140 			if (bp->b_lblkno >= trunclbn) {
1141 				if (BUF_LOCK(bp,
1142 				    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1143 				    VI_MTX(vp)) == ENOLCK)
1144 					goto restart;
1145 
1146 				bremfree(bp);
1147 				bp->b_flags |= (B_INVAL | B_RELBUF);
1148 				bp->b_flags &= ~B_ASYNC;
1149 				brelse(bp);
1150 				anyfreed = 1;
1151 
1152 				if (nbp &&
1153 				    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1154 				    (nbp->b_vp != vp) ||
1155 				    (nbp->b_flags & B_DELWRI))) {
1156 					goto restart;
1157 				}
1158 				VI_LOCK(vp);
1159 			}
1160 		}
1161 
1162 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1163 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1164 			if (bp->b_lblkno >= trunclbn) {
1165 				if (BUF_LOCK(bp,
1166 				    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1167 				    VI_MTX(vp)) == ENOLCK)
1168 					goto restart;
1169 				bremfree(bp);
1170 				bp->b_flags |= (B_INVAL | B_RELBUF);
1171 				bp->b_flags &= ~B_ASYNC;
1172 				brelse(bp);
1173 				anyfreed = 1;
1174 				if (nbp &&
1175 				    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1176 				    (nbp->b_vp != vp) ||
1177 				    (nbp->b_flags & B_DELWRI) == 0)) {
1178 					goto restart;
1179 				}
1180 				VI_LOCK(vp);
1181 			}
1182 		}
1183 	}
1184 
1185 	if (length > 0) {
1186 restartsync:
1187 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1188 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1189 			if (bp->b_lblkno > 0)
1190 				continue;
1191 			/*
1192 			 * Since we hold the vnode lock this should only
1193 			 * fail if we're racing with the buf daemon.
1194 			 */
1195 			if (BUF_LOCK(bp,
1196 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1197 			    VI_MTX(vp)) == ENOLCK) {
1198 				goto restart;
1199 			}
1200 			KASSERT((bp->b_flags & B_DELWRI),
1201 			    ("buf(%p) on dirty queue without DELWRI", bp));
1202 
1203 			bremfree(bp);
1204 			bawrite(bp);
1205 			VI_LOCK(vp);
1206 			goto restartsync;
1207 		}
1208 	}
1209 
1210 	while (vp->v_numoutput > 0) {
1211 		vp->v_iflag |= VI_BWAIT;
1212 		msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0);
1213 	}
1214 	VI_UNLOCK(vp);
1215 	vnode_pager_setsize(vp, length);
1216 
1217 	return (0);
1218 }
1219 
1220 /*
1221  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1222  * 		 a vnode.
1223  *
1224  *	NOTE: We have to deal with the special case of a background bitmap
1225  *	buffer, a situation where two buffers will have the same logical
1226  *	block offset.  We want (1) only the foreground buffer to be accessed
1227  *	in a lookup and (2) must differentiate between the foreground and
1228  *	background buffer in the splay tree algorithm because the splay
1229  *	tree cannot normally handle multiple entities with the same 'index'.
1230  *	We accomplish this by adding differentiating flags to the splay tree's
1231  *	numerical domain.
1232  */
1233 static
1234 struct buf *
1235 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1236 {
1237 	struct buf dummy;
1238 	struct buf *lefttreemax, *righttreemin, *y;
1239 
1240 	if (root == NULL)
1241 		return (NULL);
1242 	lefttreemax = righttreemin = &dummy;
1243 	for (;;) {
1244 		if (lblkno < root->b_lblkno ||
1245 		    (lblkno == root->b_lblkno &&
1246 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1247 			if ((y = root->b_left) == NULL)
1248 				break;
1249 			if (lblkno < y->b_lblkno) {
1250 				/* Rotate right. */
1251 				root->b_left = y->b_right;
1252 				y->b_right = root;
1253 				root = y;
1254 				if ((y = root->b_left) == NULL)
1255 					break;
1256 			}
1257 			/* Link into the new root's right tree. */
1258 			righttreemin->b_left = root;
1259 			righttreemin = root;
1260 		} else if (lblkno > root->b_lblkno ||
1261 		    (lblkno == root->b_lblkno &&
1262 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1263 			if ((y = root->b_right) == NULL)
1264 				break;
1265 			if (lblkno > y->b_lblkno) {
1266 				/* Rotate left. */
1267 				root->b_right = y->b_left;
1268 				y->b_left = root;
1269 				root = y;
1270 				if ((y = root->b_right) == NULL)
1271 					break;
1272 			}
1273 			/* Link into the new root's left tree. */
1274 			lefttreemax->b_right = root;
1275 			lefttreemax = root;
1276 		} else {
1277 			break;
1278 		}
1279 		root = y;
1280 	}
1281 	/* Assemble the new root. */
1282 	lefttreemax->b_right = root->b_left;
1283 	righttreemin->b_left = root->b_right;
1284 	root->b_left = dummy.b_right;
1285 	root->b_right = dummy.b_left;
1286 	return (root);
1287 }
1288 
1289 static
1290 void
1291 buf_vlist_remove(struct buf *bp)
1292 {
1293 	struct vnode *vp = bp->b_vp;
1294 	struct buf *root;
1295 
1296 	ASSERT_VI_LOCKED(vp, "buf_vlist_remove");
1297 	if (bp->b_xflags & BX_VNDIRTY) {
1298 		if (bp != vp->v_dirtyblkroot) {
1299 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1300 			    vp->v_dirtyblkroot);
1301 			KASSERT(root == bp,
1302 			    ("splay lookup failed during dirty remove"));
1303 		}
1304 		if (bp->b_left == NULL) {
1305 			root = bp->b_right;
1306 		} else {
1307 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1308 			    bp->b_left);
1309 			root->b_right = bp->b_right;
1310 		}
1311 		vp->v_dirtyblkroot = root;
1312 		TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs);
1313 		vp->v_dirtybufcnt--;
1314 	} else {
1315 		/* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */
1316 		if (bp != vp->v_cleanblkroot) {
1317 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1318 			    vp->v_cleanblkroot);
1319 			KASSERT(root == bp,
1320 			    ("splay lookup failed during clean remove"));
1321 		}
1322 		if (bp->b_left == NULL) {
1323 			root = bp->b_right;
1324 		} else {
1325 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1326 			    bp->b_left);
1327 			root->b_right = bp->b_right;
1328 		}
1329 		vp->v_cleanblkroot = root;
1330 		TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs);
1331 		vp->v_cleanbufcnt--;
1332 	}
1333 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1334 }
1335 
1336 /*
1337  * Add the buffer to the sorted clean or dirty block list using a
1338  * splay tree algorithm.
1339  *
1340  * NOTE: xflags is passed as a constant, optimizing this inline function!
1341  */
1342 static
1343 void
1344 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags)
1345 {
1346 	struct buf *root;
1347 
1348 	ASSERT_VI_LOCKED(vp, "buf_vlist_add");
1349 	bp->b_xflags |= xflags;
1350 	if (xflags & BX_VNDIRTY) {
1351 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1352 		if (root == NULL) {
1353 			bp->b_left = NULL;
1354 			bp->b_right = NULL;
1355 			TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs);
1356 		} else if (bp->b_lblkno < root->b_lblkno ||
1357 		    (bp->b_lblkno == root->b_lblkno &&
1358 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1359 			bp->b_left = root->b_left;
1360 			bp->b_right = root;
1361 			root->b_left = NULL;
1362 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1363 		} else {
1364 			bp->b_right = root->b_right;
1365 			bp->b_left = root;
1366 			root->b_right = NULL;
1367 			TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd,
1368 			    root, bp, b_vnbufs);
1369 		}
1370 		vp->v_dirtybufcnt++;
1371 		vp->v_dirtyblkroot = bp;
1372 	} else {
1373 		/* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */
1374 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1375 		if (root == NULL) {
1376 			bp->b_left = NULL;
1377 			bp->b_right = NULL;
1378 			TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs);
1379 		} else if (bp->b_lblkno < root->b_lblkno ||
1380 		    (bp->b_lblkno == root->b_lblkno &&
1381 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1382 			bp->b_left = root->b_left;
1383 			bp->b_right = root;
1384 			root->b_left = NULL;
1385 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1386 		} else {
1387 			bp->b_right = root->b_right;
1388 			bp->b_left = root;
1389 			root->b_right = NULL;
1390 			TAILQ_INSERT_AFTER(&vp->v_cleanblkhd,
1391 			    root, bp, b_vnbufs);
1392 		}
1393 		vp->v_cleanbufcnt++;
1394 		vp->v_cleanblkroot = bp;
1395 	}
1396 }
1397 
1398 /*
1399  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1400  * shadow buffers used in background bitmap writes.
1401  *
1402  * This code isn't quite efficient as it could be because we are maintaining
1403  * two sorted lists and do not know which list the block resides in.
1404  *
1405  * During a "make buildworld" the desired buffer is found at one of
1406  * the roots more than 60% of the time.  Thus, checking both roots
1407  * before performing either splay eliminates unnecessary splays on the
1408  * first tree splayed.
1409  */
1410 struct buf *
1411 gbincore(struct vnode *vp, daddr_t lblkno)
1412 {
1413 	struct buf *bp;
1414 
1415 	GIANT_REQUIRED;
1416 
1417 	ASSERT_VI_LOCKED(vp, "gbincore");
1418 	if ((bp = vp->v_cleanblkroot) != NULL &&
1419 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1420 		return (bp);
1421 	if ((bp = vp->v_dirtyblkroot) != NULL &&
1422 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1423 		return (bp);
1424 	if ((bp = vp->v_cleanblkroot) != NULL) {
1425 		vp->v_cleanblkroot = bp = buf_splay(lblkno, 0, bp);
1426 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1427 			return (bp);
1428 	}
1429 	if ((bp = vp->v_dirtyblkroot) != NULL) {
1430 		vp->v_dirtyblkroot = bp = buf_splay(lblkno, 0, bp);
1431 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1432 			return (bp);
1433 	}
1434 	return (NULL);
1435 }
1436 
1437 /*
1438  * Associate a buffer with a vnode.
1439  */
1440 void
1441 bgetvp(vp, bp)
1442 	register struct vnode *vp;
1443 	register struct buf *bp;
1444 {
1445 
1446 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1447 
1448 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1449 	    ("bgetvp: bp already attached! %p", bp));
1450 
1451 	ASSERT_VI_LOCKED(vp, "bgetvp");
1452 	vholdl(vp);
1453 	bp->b_vp = vp;
1454 	bp->b_dev = vn_todev(vp);
1455 	/*
1456 	 * Insert onto list for new vnode.
1457 	 */
1458 	buf_vlist_add(bp, vp, BX_VNCLEAN);
1459 }
1460 
1461 /*
1462  * Disassociate a buffer from a vnode.
1463  */
1464 void
1465 brelvp(bp)
1466 	register struct buf *bp;
1467 {
1468 	struct vnode *vp;
1469 
1470 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1471 
1472 	/*
1473 	 * Delete from old vnode list, if on one.
1474 	 */
1475 	vp = bp->b_vp;
1476 	VI_LOCK(vp);
1477 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1478 		buf_vlist_remove(bp);
1479 	if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1480 		vp->v_iflag &= ~VI_ONWORKLST;
1481 		mtx_lock(&sync_mtx);
1482 		LIST_REMOVE(vp, v_synclist);
1483  		syncer_worklist_len--;
1484 		mtx_unlock(&sync_mtx);
1485 	}
1486 	vdropl(vp);
1487 	bp->b_vp = (struct vnode *) 0;
1488 	if (bp->b_object)
1489 		bp->b_object = NULL;
1490 	VI_UNLOCK(vp);
1491 }
1492 
1493 /*
1494  * Add an item to the syncer work queue.
1495  */
1496 static void
1497 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
1498 {
1499 	int slot;
1500 
1501 	ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist");
1502 
1503 	mtx_lock(&sync_mtx);
1504 	if (vp->v_iflag & VI_ONWORKLST)
1505 		LIST_REMOVE(vp, v_synclist);
1506 	else {
1507 		vp->v_iflag |= VI_ONWORKLST;
1508  		syncer_worklist_len++;
1509 	}
1510 
1511 	if (delay > syncer_maxdelay - 2)
1512 		delay = syncer_maxdelay - 2;
1513 	slot = (syncer_delayno + delay) & syncer_mask;
1514 
1515 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
1516 	mtx_unlock(&sync_mtx);
1517 }
1518 
1519 static int
1520 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1521 {
1522 	int error, len;
1523 
1524 	mtx_lock(&sync_mtx);
1525 	len = syncer_worklist_len - sync_vnode_count;
1526 	mtx_unlock(&sync_mtx);
1527 	error = SYSCTL_OUT(req, &len, sizeof(len));
1528 	return (error);
1529 }
1530 
1531 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1532     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1533 
1534 struct  proc *updateproc;
1535 static void sched_sync(void);
1536 static struct kproc_desc up_kp = {
1537 	"syncer",
1538 	sched_sync,
1539 	&updateproc
1540 };
1541 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1542 
1543 /*
1544  * System filesystem synchronizer daemon.
1545  */
1546 static void
1547 sched_sync(void)
1548 {
1549 	struct synclist *next;
1550 	struct synclist *slp;
1551 	struct vnode *vp;
1552 	struct mount *mp;
1553 	long starttime;
1554 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1555 	static int dummychan;
1556 	int last_work_seen;
1557 	int net_worklist_len;
1558 	int syncer_final_iter;
1559 	int first_printf;
1560 
1561 	mtx_lock(&Giant);
1562 	last_work_seen = 0;
1563 	syncer_final_iter = 0;
1564 	first_printf = 1;
1565 	syncer_state = SYNCER_RUNNING;
1566 	starttime = time_second;
1567 
1568 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1569 	    SHUTDOWN_PRI_LAST);
1570 
1571 	for (;;) {
1572 		mtx_lock(&sync_mtx);
1573 		if (syncer_state == SYNCER_FINAL_DELAY &&
1574 		    syncer_final_iter == 0) {
1575 			mtx_unlock(&sync_mtx);
1576 			kthread_suspend_check(td->td_proc);
1577 			mtx_lock(&sync_mtx);
1578 		}
1579 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1580 		if (syncer_state != SYNCER_RUNNING &&
1581 		    starttime != time_second) {
1582 			if (first_printf) {
1583 				printf("\nSyncing disks, vnodes remaining...");
1584 				first_printf = 0;
1585 			}
1586 			printf("%d ", net_worklist_len);
1587 		}
1588 		starttime = time_second;
1589 
1590 		/*
1591 		 * Push files whose dirty time has expired.  Be careful
1592 		 * of interrupt race on slp queue.
1593 		 *
1594 		 * Skip over empty worklist slots when shutting down.
1595 		 */
1596 		do {
1597 			slp = &syncer_workitem_pending[syncer_delayno];
1598 			syncer_delayno += 1;
1599 			if (syncer_delayno == syncer_maxdelay)
1600 				syncer_delayno = 0;
1601 			next = &syncer_workitem_pending[syncer_delayno];
1602 			/*
1603 			 * If the worklist has wrapped since the
1604 			 * it was emptied of all but syncer vnodes,
1605 			 * switch to the FINAL_DELAY state and run
1606 			 * for one more second.
1607 			 */
1608 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1609 			    net_worklist_len == 0 &&
1610 			    last_work_seen == syncer_delayno) {
1611 				syncer_state = SYNCER_FINAL_DELAY;
1612 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1613 			}
1614 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1615 		    syncer_worklist_len > 0);
1616 
1617 		/*
1618 		 * Keep track of the last time there was anything
1619 		 * on the worklist other than syncer vnodes.
1620 		 * Return to the SHUTTING_DOWN state if any
1621 		 * new work appears.
1622 		 */
1623 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1624 			last_work_seen = syncer_delayno;
1625 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1626 			syncer_state = SYNCER_SHUTTING_DOWN;
1627 		while ((vp = LIST_FIRST(slp)) != NULL) {
1628 			if (VOP_ISLOCKED(vp, NULL) != 0 ||
1629 			    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1630 				LIST_REMOVE(vp, v_synclist);
1631 				LIST_INSERT_HEAD(next, vp, v_synclist);
1632 				continue;
1633 			}
1634 			if (VI_TRYLOCK(vp) == 0) {
1635 				LIST_REMOVE(vp, v_synclist);
1636 				LIST_INSERT_HEAD(next, vp, v_synclist);
1637 				vn_finished_write(mp);
1638 				continue;
1639 			}
1640 			/*
1641 			 * We use vhold in case the vnode does not
1642 			 * successfully sync.  vhold prevents the vnode from
1643 			 * going away when we unlock the sync_mtx so that
1644 			 * we can acquire the vnode interlock.
1645 			 */
1646 			vholdl(vp);
1647 			mtx_unlock(&sync_mtx);
1648 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td);
1649 			(void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td);
1650 			VOP_UNLOCK(vp, 0, td);
1651 			vn_finished_write(mp);
1652 			VI_LOCK(vp);
1653 			if ((vp->v_iflag & VI_ONWORKLST) != 0) {
1654 				/*
1655 				 * Put us back on the worklist.  The worklist
1656 				 * routine will remove us from our current
1657 				 * position and then add us back in at a later
1658 				 * position.
1659 				 */
1660 				vn_syncer_add_to_worklist(vp, syncdelay);
1661 			}
1662 			vdropl(vp);
1663 			VI_UNLOCK(vp);
1664 			mtx_lock(&sync_mtx);
1665 		}
1666 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1667 			syncer_final_iter--;
1668 		mtx_unlock(&sync_mtx);
1669 
1670 		/*
1671 		 * Do soft update processing.
1672 		 */
1673 		if (softdep_process_worklist_hook != NULL)
1674 			(*softdep_process_worklist_hook)(NULL);
1675 
1676 		/*
1677 		 * The variable rushjob allows the kernel to speed up the
1678 		 * processing of the filesystem syncer process. A rushjob
1679 		 * value of N tells the filesystem syncer to process the next
1680 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1681 		 * is used by the soft update code to speed up the filesystem
1682 		 * syncer process when the incore state is getting so far
1683 		 * ahead of the disk that the kernel memory pool is being
1684 		 * threatened with exhaustion.
1685 		 */
1686 		mtx_lock(&sync_mtx);
1687 		if (rushjob > 0) {
1688 			rushjob -= 1;
1689 			mtx_unlock(&sync_mtx);
1690 			continue;
1691 		}
1692 		mtx_unlock(&sync_mtx);
1693 		/*
1694 		 * Just sleep for a short period if time between
1695 		 * iterations when shutting down to allow some I/O
1696 		 * to happen.
1697 		 *
1698 		 * If it has taken us less than a second to process the
1699 		 * current work, then wait. Otherwise start right over
1700 		 * again. We can still lose time if any single round
1701 		 * takes more than two seconds, but it does not really
1702 		 * matter as we are just trying to generally pace the
1703 		 * filesystem activity.
1704 		 */
1705 		if (syncer_state != SYNCER_RUNNING)
1706 			tsleep(&dummychan, PPAUSE, "syncfnl",
1707 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1708 		else if (time_second == starttime)
1709 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1710 	}
1711 }
1712 
1713 /*
1714  * Request the syncer daemon to speed up its work.
1715  * We never push it to speed up more than half of its
1716  * normal turn time, otherwise it could take over the cpu.
1717  */
1718 int
1719 speedup_syncer()
1720 {
1721 	struct thread *td;
1722 	int ret = 0;
1723 
1724 	td = FIRST_THREAD_IN_PROC(updateproc);
1725 	sleepq_remove(td, &lbolt);
1726 	mtx_lock(&sync_mtx);
1727 	if (rushjob < syncdelay / 2) {
1728 		rushjob += 1;
1729 		stat_rush_requests += 1;
1730 		ret = 1;
1731 	}
1732 	mtx_unlock(&sync_mtx);
1733 	return (ret);
1734 }
1735 
1736 /*
1737  * Tell the syncer to speed up its work and run though its work
1738  * list several times, then tell it to shut down.
1739  */
1740 static void
1741 syncer_shutdown(void *arg, int howto)
1742 {
1743 	struct thread *td;
1744 
1745 	if (howto & RB_NOSYNC)
1746 		return;
1747 	td = FIRST_THREAD_IN_PROC(updateproc);
1748 	sleepq_remove(td, &lbolt);
1749 	mtx_lock(&sync_mtx);
1750 	syncer_state = SYNCER_SHUTTING_DOWN;
1751 	rushjob = 0;
1752 	mtx_unlock(&sync_mtx);
1753 	kproc_shutdown(arg, howto);
1754 }
1755 
1756 /*
1757  * Associate a p-buffer with a vnode.
1758  *
1759  * Also sets B_PAGING flag to indicate that vnode is not fully associated
1760  * with the buffer.  i.e. the bp has not been linked into the vnode or
1761  * ref-counted.
1762  */
1763 void
1764 pbgetvp(vp, bp)
1765 	register struct vnode *vp;
1766 	register struct buf *bp;
1767 {
1768 
1769 	KASSERT(bp->b_vp == NULL, ("pbgetvp: not free"));
1770 
1771 	bp->b_vp = vp;
1772 	bp->b_object = vp->v_object;
1773 	bp->b_flags |= B_PAGING;
1774 	bp->b_dev = vn_todev(vp);
1775 }
1776 
1777 /*
1778  * Disassociate a p-buffer from a vnode.
1779  */
1780 void
1781 pbrelvp(bp)
1782 	register struct buf *bp;
1783 {
1784 
1785 	KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL"));
1786 
1787 	/* XXX REMOVE ME */
1788 	VI_LOCK(bp->b_vp);
1789 	if (TAILQ_NEXT(bp, b_vnbufs) != NULL) {
1790 		panic(
1791 		    "relpbuf(): b_vp was probably reassignbuf()d %p %x",
1792 		    bp,
1793 		    (int)bp->b_flags
1794 		);
1795 	}
1796 	VI_UNLOCK(bp->b_vp);
1797 	bp->b_vp = (struct vnode *) 0;
1798 	bp->b_object = NULL;
1799 	bp->b_flags &= ~B_PAGING;
1800 }
1801 
1802 /*
1803  * Reassign a buffer from one vnode to another.
1804  * Used to assign file specific control information
1805  * (indirect blocks) to the vnode to which they belong.
1806  */
1807 void
1808 reassignbuf(struct buf *bp)
1809 {
1810 	struct vnode *vp;
1811 	int delay;
1812 
1813 	vp = bp->b_vp;
1814 	++reassignbufcalls;
1815 
1816 	/*
1817 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1818 	 * is not fully linked in.
1819 	 */
1820 	if (bp->b_flags & B_PAGING)
1821 		panic("cannot reassign paging buffer");
1822 
1823 	/*
1824 	 * Delete from old vnode list, if on one.
1825 	 */
1826 	VI_LOCK(vp);
1827 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1828 		buf_vlist_remove(bp);
1829 	/*
1830 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1831 	 * of clean buffers.
1832 	 */
1833 	if (bp->b_flags & B_DELWRI) {
1834 		if ((vp->v_iflag & VI_ONWORKLST) == 0) {
1835 			switch (vp->v_type) {
1836 			case VDIR:
1837 				delay = dirdelay;
1838 				break;
1839 			case VCHR:
1840 				delay = metadelay;
1841 				break;
1842 			default:
1843 				delay = filedelay;
1844 			}
1845 			vn_syncer_add_to_worklist(vp, delay);
1846 		}
1847 		buf_vlist_add(bp, vp, BX_VNDIRTY);
1848 	} else {
1849 		buf_vlist_add(bp, vp, BX_VNCLEAN);
1850 
1851 		if ((vp->v_iflag & VI_ONWORKLST) &&
1852 		    TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1853 			mtx_lock(&sync_mtx);
1854 			LIST_REMOVE(vp, v_synclist);
1855  			syncer_worklist_len--;
1856 			mtx_unlock(&sync_mtx);
1857 			vp->v_iflag &= ~VI_ONWORKLST;
1858 		}
1859 	}
1860 	VI_UNLOCK(vp);
1861 }
1862 
1863 /*
1864  * Create a vnode for a device.
1865  * Used for mounting the root filesystem.
1866  */
1867 int
1868 bdevvp(dev, vpp)
1869 	struct cdev *dev;
1870 	struct vnode **vpp;
1871 {
1872 	register struct vnode *vp;
1873 	struct vnode *nvp;
1874 	int error;
1875 
1876 	if (dev == NULL) {
1877 		*vpp = NULLVP;
1878 		return (ENXIO);
1879 	}
1880 	if (vfinddev(dev, vpp))
1881 		return (0);
1882 
1883 	error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp);
1884 	if (error) {
1885 		*vpp = NULLVP;
1886 		return (error);
1887 	}
1888 	vp = nvp;
1889 	vp->v_type = VCHR;
1890 	vp->v_bsize = DEV_BSIZE;
1891 	addalias(vp, dev);
1892 	*vpp = vp;
1893 	return (0);
1894 }
1895 
1896 static void
1897 v_incr_usecount(struct vnode *vp, int delta)
1898 {
1899 
1900 	vp->v_usecount += delta;
1901 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1902 		mtx_lock(&spechash_mtx);
1903 		vp->v_rdev->si_usecount += delta;
1904 		mtx_unlock(&spechash_mtx);
1905 	}
1906 }
1907 
1908 /*
1909  * Add vnode to the alias list hung off the struct cdev *.
1910  *
1911  * The reason for this gunk is that multiple vnodes can reference
1912  * the same physical device, so checking vp->v_usecount to see
1913  * how many users there are is inadequate; the v_usecount for
1914  * the vnodes need to be accumulated.  vcount() does that.
1915  */
1916 struct vnode *
1917 addaliasu(nvp, nvp_rdev)
1918 	struct vnode *nvp;
1919 	dev_t nvp_rdev;
1920 {
1921 	struct vnode *ovp;
1922 	vop_t **ops;
1923 	struct cdev *dev;
1924 
1925 	if (nvp->v_type == VBLK)
1926 		return (nvp);
1927 	if (nvp->v_type != VCHR)
1928 		panic("addaliasu on non-special vnode");
1929 	dev = findcdev(nvp_rdev);
1930 	if (dev == NULL)
1931 		return (nvp);
1932 	/*
1933 	 * Check to see if we have a bdevvp vnode with no associated
1934 	 * filesystem. If so, we want to associate the filesystem of
1935 	 * the new newly instigated vnode with the bdevvp vnode and
1936 	 * discard the newly created vnode rather than leaving the
1937 	 * bdevvp vnode lying around with no associated filesystem.
1938 	 */
1939 	if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) {
1940 		addalias(nvp, dev);
1941 		return (nvp);
1942 	}
1943 	/*
1944 	 * Discard unneeded vnode, but save its node specific data.
1945 	 * Note that if there is a lock, it is carried over in the
1946 	 * node specific data to the replacement vnode.
1947 	 */
1948 	vref(ovp);
1949 	ovp->v_data = nvp->v_data;
1950 	ovp->v_tag = nvp->v_tag;
1951 	nvp->v_data = NULL;
1952 	lockdestroy(ovp->v_vnlock);
1953 	lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg,
1954 	    nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK);
1955 	ops = ovp->v_op;
1956 	ovp->v_op = nvp->v_op;
1957 	if (VOP_ISLOCKED(nvp, curthread)) {
1958 		VOP_UNLOCK(nvp, 0, curthread);
1959 		vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread);
1960 	}
1961 	nvp->v_op = ops;
1962 	delmntque(ovp);
1963 	insmntque(ovp, nvp->v_mount);
1964 	vrele(nvp);
1965 	vgone(nvp);
1966 	return (ovp);
1967 }
1968 
1969 /* This is a local helper function that do the same as addaliasu, but for a
1970  * struct cdev *instead of an dev_t. */
1971 static void
1972 addalias(nvp, dev)
1973 	struct vnode *nvp;
1974 	struct cdev *dev;
1975 {
1976 
1977 	KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode"));
1978 	dev_ref(dev);
1979 	nvp->v_rdev = dev;
1980 	VI_LOCK(nvp);
1981 	mtx_lock(&spechash_mtx);
1982 	SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext);
1983 	dev->si_usecount += nvp->v_usecount;
1984 	mtx_unlock(&spechash_mtx);
1985 	VI_UNLOCK(nvp);
1986 }
1987 
1988 /*
1989  * Grab a particular vnode from the free list, increment its
1990  * reference count and lock it. The vnode lock bit is set if the
1991  * vnode is being eliminated in vgone. The process is awakened
1992  * when the transition is completed, and an error returned to
1993  * indicate that the vnode is no longer usable (possibly having
1994  * been changed to a new filesystem type).
1995  */
1996 int
1997 vget(vp, flags, td)
1998 	register struct vnode *vp;
1999 	int flags;
2000 	struct thread *td;
2001 {
2002 	int error;
2003 
2004 	/*
2005 	 * If the vnode is in the process of being cleaned out for
2006 	 * another use, we wait for the cleaning to finish and then
2007 	 * return failure. Cleaning is determined by checking that
2008 	 * the VI_XLOCK flag is set.
2009 	 */
2010 	if ((flags & LK_INTERLOCK) == 0)
2011 		VI_LOCK(vp);
2012 	if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) {
2013 		if ((flags & LK_NOWAIT) == 0) {
2014 			vp->v_iflag |= VI_XWANT;
2015 			msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0);
2016 			return (ENOENT);
2017 		}
2018 		VI_UNLOCK(vp);
2019 		return (EBUSY);
2020 	}
2021 
2022 	v_incr_usecount(vp, 1);
2023 
2024 	if (VSHOULDBUSY(vp))
2025 		vbusy(vp);
2026 	if (flags & LK_TYPE_MASK) {
2027 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2028 			/*
2029 			 * must expand vrele here because we do not want
2030 			 * to call VOP_INACTIVE if the reference count
2031 			 * drops back to zero since it was never really
2032 			 * active. We must remove it from the free list
2033 			 * before sleeping so that multiple processes do
2034 			 * not try to recycle it.
2035 			 */
2036 			VI_LOCK(vp);
2037 			v_incr_usecount(vp, -1);
2038 			if (VSHOULDFREE(vp))
2039 				vfree(vp);
2040 			else
2041 				vlruvp(vp);
2042 			VI_UNLOCK(vp);
2043 		}
2044 		return (error);
2045 	}
2046 	VI_UNLOCK(vp);
2047 	return (0);
2048 }
2049 
2050 /*
2051  * Increase the reference count of a vnode.
2052  */
2053 void
2054 vref(struct vnode *vp)
2055 {
2056 
2057 	VI_LOCK(vp);
2058 	v_incr_usecount(vp, 1);
2059 	VI_UNLOCK(vp);
2060 }
2061 
2062 /*
2063  * Return reference count of a vnode.
2064  *
2065  * The results of this call are only guaranteed when some mechanism other
2066  * than the VI lock is used to stop other processes from gaining references
2067  * to the vnode.  This may be the case if the caller holds the only reference.
2068  * This is also useful when stale data is acceptable as race conditions may
2069  * be accounted for by some other means.
2070  */
2071 int
2072 vrefcnt(struct vnode *vp)
2073 {
2074 	int usecnt;
2075 
2076 	VI_LOCK(vp);
2077 	usecnt = vp->v_usecount;
2078 	VI_UNLOCK(vp);
2079 
2080 	return (usecnt);
2081 }
2082 
2083 
2084 /*
2085  * Vnode put/release.
2086  * If count drops to zero, call inactive routine and return to freelist.
2087  */
2088 void
2089 vrele(vp)
2090 	struct vnode *vp;
2091 {
2092 	struct thread *td = curthread;	/* XXX */
2093 
2094 	GIANT_REQUIRED;
2095 
2096 	KASSERT(vp != NULL, ("vrele: null vp"));
2097 
2098 	VI_LOCK(vp);
2099 
2100 	/* Skip this v_writecount check if we're going to panic below. */
2101 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2102 	    ("vrele: missed vn_close"));
2103 
2104 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2105 	    vp->v_usecount == 1)) {
2106 		v_incr_usecount(vp, -1);
2107 		VI_UNLOCK(vp);
2108 
2109 		return;
2110 	}
2111 
2112 	if (vp->v_usecount == 1) {
2113 		v_incr_usecount(vp, -1);
2114 		/*
2115 		 * We must call VOP_INACTIVE with the node locked. Mark
2116 		 * as VI_DOINGINACT to avoid recursion.
2117 		 */
2118 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2119 			VI_LOCK(vp);
2120 			vp->v_iflag |= VI_DOINGINACT;
2121 			VI_UNLOCK(vp);
2122 			VOP_INACTIVE(vp, td);
2123 			VI_LOCK(vp);
2124 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2125 			    ("vrele: lost VI_DOINGINACT"));
2126 			vp->v_iflag &= ~VI_DOINGINACT;
2127 		} else
2128 			VI_LOCK(vp);
2129 		if (VSHOULDFREE(vp))
2130 			vfree(vp);
2131 		else
2132 			vlruvp(vp);
2133 		VI_UNLOCK(vp);
2134 
2135 	} else {
2136 #ifdef DIAGNOSTIC
2137 		vprint("vrele: negative ref count", vp);
2138 #endif
2139 		VI_UNLOCK(vp);
2140 		panic("vrele: negative ref cnt");
2141 	}
2142 }
2143 
2144 /*
2145  * Release an already locked vnode.  This give the same effects as
2146  * unlock+vrele(), but takes less time and avoids releasing and
2147  * re-aquiring the lock (as vrele() aquires the lock internally.)
2148  */
2149 void
2150 vput(vp)
2151 	struct vnode *vp;
2152 {
2153 	struct thread *td = curthread;	/* XXX */
2154 
2155 	GIANT_REQUIRED;
2156 
2157 	KASSERT(vp != NULL, ("vput: null vp"));
2158 	VI_LOCK(vp);
2159 	/* Skip this v_writecount check if we're going to panic below. */
2160 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2161 	    ("vput: missed vn_close"));
2162 
2163 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2164 	    vp->v_usecount == 1)) {
2165 		v_incr_usecount(vp, -1);
2166 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
2167 		return;
2168 	}
2169 
2170 	if (vp->v_usecount == 1) {
2171 		v_incr_usecount(vp, -1);
2172 		/*
2173 		 * We must call VOP_INACTIVE with the node locked, so
2174 		 * we just need to release the vnode mutex. Mark as
2175 		 * as VI_DOINGINACT to avoid recursion.
2176 		 */
2177 		vp->v_iflag |= VI_DOINGINACT;
2178 		VI_UNLOCK(vp);
2179 		VOP_INACTIVE(vp, td);
2180 		VI_LOCK(vp);
2181 		KASSERT(vp->v_iflag & VI_DOINGINACT,
2182 		    ("vput: lost VI_DOINGINACT"));
2183 		vp->v_iflag &= ~VI_DOINGINACT;
2184 		if (VSHOULDFREE(vp))
2185 			vfree(vp);
2186 		else
2187 			vlruvp(vp);
2188 		VI_UNLOCK(vp);
2189 
2190 	} else {
2191 #ifdef DIAGNOSTIC
2192 		vprint("vput: negative ref count", vp);
2193 #endif
2194 		panic("vput: negative ref cnt");
2195 	}
2196 }
2197 
2198 /*
2199  * Somebody doesn't want the vnode recycled.
2200  */
2201 void
2202 vhold(struct vnode *vp)
2203 {
2204 
2205 	VI_LOCK(vp);
2206 	vholdl(vp);
2207 	VI_UNLOCK(vp);
2208 }
2209 
2210 void
2211 vholdl(vp)
2212 	register struct vnode *vp;
2213 {
2214 
2215 	vp->v_holdcnt++;
2216 	if (VSHOULDBUSY(vp))
2217 		vbusy(vp);
2218 }
2219 
2220 /*
2221  * Note that there is one less who cares about this vnode.  vdrop() is the
2222  * opposite of vhold().
2223  */
2224 void
2225 vdrop(struct vnode *vp)
2226 {
2227 
2228 	VI_LOCK(vp);
2229 	vdropl(vp);
2230 	VI_UNLOCK(vp);
2231 }
2232 
2233 void
2234 vdropl(vp)
2235 	register struct vnode *vp;
2236 {
2237 
2238 	if (vp->v_holdcnt <= 0)
2239 		panic("vdrop: holdcnt");
2240 	vp->v_holdcnt--;
2241 	if (VSHOULDFREE(vp))
2242 		vfree(vp);
2243 	else
2244 		vlruvp(vp);
2245 }
2246 
2247 /*
2248  * Remove any vnodes in the vnode table belonging to mount point mp.
2249  *
2250  * If FORCECLOSE is not specified, there should not be any active ones,
2251  * return error if any are found (nb: this is a user error, not a
2252  * system error). If FORCECLOSE is specified, detach any active vnodes
2253  * that are found.
2254  *
2255  * If WRITECLOSE is set, only flush out regular file vnodes open for
2256  * writing.
2257  *
2258  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2259  *
2260  * `rootrefs' specifies the base reference count for the root vnode
2261  * of this filesystem. The root vnode is considered busy if its
2262  * v_usecount exceeds this value. On a successful return, vflush(, td)
2263  * will call vrele() on the root vnode exactly rootrefs times.
2264  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2265  * be zero.
2266  */
2267 #ifdef DIAGNOSTIC
2268 static int busyprt = 0;		/* print out busy vnodes */
2269 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2270 #endif
2271 
2272 int
2273 vflush(mp, rootrefs, flags, td)
2274 	struct mount *mp;
2275 	int rootrefs;
2276 	int flags;
2277 	struct thread *td;
2278 {
2279 	struct vnode *vp, *nvp, *rootvp = NULL;
2280 	struct vattr vattr;
2281 	int busy = 0, error;
2282 
2283 	if (rootrefs > 0) {
2284 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2285 		    ("vflush: bad args"));
2286 		/*
2287 		 * Get the filesystem root vnode. We can vput() it
2288 		 * immediately, since with rootrefs > 0, it won't go away.
2289 		 */
2290 		if ((error = VFS_ROOT(mp, &rootvp, td)) != 0)
2291 			return (error);
2292 		vput(rootvp);
2293 
2294 	}
2295 	MNT_ILOCK(mp);
2296 loop:
2297 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2298 
2299 		VI_LOCK(vp);
2300 		MNT_IUNLOCK(mp);
2301 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2302 		if (error) {
2303 			MNT_ILOCK(mp);
2304 			goto loop;
2305 		}
2306 		/*
2307 		 * Skip over a vnodes marked VV_SYSTEM.
2308 		 */
2309 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2310 			VOP_UNLOCK(vp, 0, td);
2311 			MNT_ILOCK(mp);
2312 			continue;
2313 		}
2314 		/*
2315 		 * If WRITECLOSE is set, flush out unlinked but still open
2316 		 * files (even if open only for reading) and regular file
2317 		 * vnodes open for writing.
2318 		 */
2319 		if (flags & WRITECLOSE) {
2320 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2321 			VI_LOCK(vp);
2322 
2323 			if ((vp->v_type == VNON ||
2324 			    (error == 0 && vattr.va_nlink > 0)) &&
2325 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2326 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2327 				MNT_ILOCK(mp);
2328 				continue;
2329 			}
2330 		} else
2331 			VI_LOCK(vp);
2332 
2333 		VOP_UNLOCK(vp, 0, td);
2334 
2335 		/*
2336 		 * With v_usecount == 0, all we need to do is clear out the
2337 		 * vnode data structures and we are done.
2338 		 */
2339 		if (vp->v_usecount == 0) {
2340 			vgonel(vp, td);
2341 			MNT_ILOCK(mp);
2342 			continue;
2343 		}
2344 
2345 		/*
2346 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2347 		 * or character devices, revert to an anonymous device. For
2348 		 * all other files, just kill them.
2349 		 */
2350 		if (flags & FORCECLOSE) {
2351 			if (vp->v_type != VCHR)
2352 				vgonel(vp, td);
2353 			else
2354 				vgonechrl(vp, td);
2355 			MNT_ILOCK(mp);
2356 			continue;
2357 		}
2358 #ifdef DIAGNOSTIC
2359 		if (busyprt)
2360 			vprint("vflush: busy vnode", vp);
2361 #endif
2362 		VI_UNLOCK(vp);
2363 		MNT_ILOCK(mp);
2364 		busy++;
2365 	}
2366 	MNT_IUNLOCK(mp);
2367 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2368 		/*
2369 		 * If just the root vnode is busy, and if its refcount
2370 		 * is equal to `rootrefs', then go ahead and kill it.
2371 		 */
2372 		VI_LOCK(rootvp);
2373 		KASSERT(busy > 0, ("vflush: not busy"));
2374 		KASSERT(rootvp->v_usecount >= rootrefs,
2375 		    ("vflush: usecount %d < rootrefs %d",
2376 		     rootvp->v_usecount, rootrefs));
2377 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2378 			vgonel(rootvp, td);
2379 			busy = 0;
2380 		} else
2381 			VI_UNLOCK(rootvp);
2382 	}
2383 	if (busy)
2384 		return (EBUSY);
2385 	for (; rootrefs > 0; rootrefs--)
2386 		vrele(rootvp);
2387 	return (0);
2388 }
2389 
2390 /*
2391  * This moves a now (likely recyclable) vnode to the end of the
2392  * mountlist.  XXX However, it is temporarily disabled until we
2393  * can clean up ffs_sync() and friends, which have loop restart
2394  * conditions which this code causes to operate O(N^2).
2395  */
2396 static void
2397 vlruvp(struct vnode *vp)
2398 {
2399 #if 0
2400 	struct mount *mp;
2401 
2402 	if ((mp = vp->v_mount) != NULL) {
2403 		MNT_ILOCK(mp);
2404 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2405 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2406 		MNT_IUNLOCK(mp);
2407 	}
2408 #endif
2409 }
2410 
2411 static void
2412 vx_lock(struct vnode *vp)
2413 {
2414 
2415 	ASSERT_VI_LOCKED(vp, "vx_lock");
2416 
2417 	/*
2418 	 * Prevent the vnode from being recycled or brought into use while we
2419 	 * clean it out.
2420 	 */
2421 	if (vp->v_iflag & VI_XLOCK)
2422 		panic("vclean: deadlock");
2423 	vp->v_iflag |= VI_XLOCK;
2424 	vp->v_vxthread = curthread;
2425 }
2426 
2427 static void
2428 vx_unlock(struct vnode *vp)
2429 {
2430 	ASSERT_VI_LOCKED(vp, "vx_unlock");
2431 	vp->v_iflag &= ~VI_XLOCK;
2432 	vp->v_vxthread = NULL;
2433 	if (vp->v_iflag & VI_XWANT) {
2434 		vp->v_iflag &= ~VI_XWANT;
2435 		wakeup(vp);
2436 	}
2437 }
2438 
2439 /*
2440  * Disassociate the underlying filesystem from a vnode.
2441  */
2442 static void
2443 vclean(vp, flags, td)
2444 	struct vnode *vp;
2445 	int flags;
2446 	struct thread *td;
2447 {
2448 	int active;
2449 
2450 	ASSERT_VI_LOCKED(vp, "vclean");
2451 	/*
2452 	 * Check to see if the vnode is in use. If so we have to reference it
2453 	 * before we clean it out so that its count cannot fall to zero and
2454 	 * generate a race against ourselves to recycle it.
2455 	 */
2456 	if ((active = vp->v_usecount))
2457 		v_incr_usecount(vp, 1);
2458 
2459 	/*
2460 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2461 	 * have the object locked while it cleans it out. The VOP_LOCK
2462 	 * ensures that the VOP_INACTIVE routine is done with its work.
2463 	 * For active vnodes, it ensures that no other activity can
2464 	 * occur while the underlying object is being cleaned out.
2465 	 */
2466 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2467 
2468 	/*
2469 	 * Clean out any buffers associated with the vnode.
2470 	 * If the flush fails, just toss the buffers.
2471 	 */
2472 	if (flags & DOCLOSE) {
2473 		struct buf *bp;
2474 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
2475 		if (bp != NULL)
2476 			(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2477 		if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0)
2478 			vinvalbuf(vp, 0, NOCRED, td, 0, 0);
2479 	}
2480 
2481 	VOP_DESTROYVOBJECT(vp);
2482 
2483 	/*
2484 	 * Any other processes trying to obtain this lock must first
2485 	 * wait for VXLOCK to clear, then call the new lock operation.
2486 	 */
2487 	VOP_UNLOCK(vp, 0, td);
2488 
2489 	/*
2490 	 * If purging an active vnode, it must be closed and
2491 	 * deactivated before being reclaimed. Note that the
2492 	 * VOP_INACTIVE will unlock the vnode.
2493 	 */
2494 	if (active) {
2495 		if (flags & DOCLOSE)
2496 			VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2497 		VI_LOCK(vp);
2498 		if ((vp->v_iflag & VI_DOINGINACT) == 0) {
2499 			vp->v_iflag |= VI_DOINGINACT;
2500 			VI_UNLOCK(vp);
2501 			if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2502 				panic("vclean: cannot relock.");
2503 			VOP_INACTIVE(vp, td);
2504 			VI_LOCK(vp);
2505 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2506 			    ("vclean: lost VI_DOINGINACT"));
2507 			vp->v_iflag &= ~VI_DOINGINACT;
2508 		}
2509 		VI_UNLOCK(vp);
2510 	}
2511 	/*
2512 	 * Reclaim the vnode.
2513 	 */
2514 	if (VOP_RECLAIM(vp, td))
2515 		panic("vclean: cannot reclaim");
2516 
2517 	if (active) {
2518 		/*
2519 		 * Inline copy of vrele() since VOP_INACTIVE
2520 		 * has already been called.
2521 		 */
2522 		VI_LOCK(vp);
2523 		v_incr_usecount(vp, -1);
2524 		if (vp->v_usecount <= 0) {
2525 #ifdef INVARIANTS
2526 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2527 				vprint("vclean: bad ref count", vp);
2528 				panic("vclean: ref cnt");
2529 			}
2530 #endif
2531 			if (VSHOULDFREE(vp))
2532 				vfree(vp);
2533 		}
2534 		VI_UNLOCK(vp);
2535 	}
2536 	/*
2537 	 * Delete from old mount point vnode list.
2538 	 */
2539 	delmntque(vp);
2540 	cache_purge(vp);
2541 	VI_LOCK(vp);
2542 	if (VSHOULDFREE(vp))
2543 		vfree(vp);
2544 
2545 	/*
2546 	 * Done with purge, reset to the standard lock and
2547 	 * notify sleepers of the grim news.
2548 	 */
2549 	vp->v_vnlock = &vp->v_lock;
2550 	vp->v_op = dead_vnodeop_p;
2551 	if (vp->v_pollinfo != NULL)
2552 		vn_pollgone(vp);
2553 	vp->v_tag = "none";
2554 }
2555 
2556 /*
2557  * Eliminate all activity associated with the requested vnode
2558  * and with all vnodes aliased to the requested vnode.
2559  */
2560 int
2561 vop_revoke(ap)
2562 	struct vop_revoke_args /* {
2563 		struct vnode *a_vp;
2564 		int a_flags;
2565 	} */ *ap;
2566 {
2567 	struct vnode *vp, *vq;
2568 	struct cdev *dev;
2569 
2570 	KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke"));
2571 	vp = ap->a_vp;
2572 	KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR"));
2573 
2574 	VI_LOCK(vp);
2575 	/*
2576 	 * If a vgone (or vclean) is already in progress,
2577 	 * wait until it is done and return.
2578 	 */
2579 	if (vp->v_iflag & VI_XLOCK) {
2580 		vp->v_iflag |= VI_XWANT;
2581 		msleep(vp, VI_MTX(vp), PINOD | PDROP,
2582 		    "vop_revokeall", 0);
2583 		return (0);
2584 	}
2585 	VI_UNLOCK(vp);
2586 	dev = vp->v_rdev;
2587 	for (;;) {
2588 		mtx_lock(&spechash_mtx);
2589 		vq = SLIST_FIRST(&dev->si_hlist);
2590 		mtx_unlock(&spechash_mtx);
2591 		if (vq == NULL)
2592 			break;
2593 		vgone(vq);
2594 	}
2595 	return (0);
2596 }
2597 
2598 /*
2599  * Recycle an unused vnode to the front of the free list.
2600  * Release the passed interlock if the vnode will be recycled.
2601  */
2602 int
2603 vrecycle(vp, inter_lkp, td)
2604 	struct vnode *vp;
2605 	struct mtx *inter_lkp;
2606 	struct thread *td;
2607 {
2608 
2609 	VI_LOCK(vp);
2610 	if (vp->v_usecount == 0) {
2611 		if (inter_lkp) {
2612 			mtx_unlock(inter_lkp);
2613 		}
2614 		vgonel(vp, td);
2615 		return (1);
2616 	}
2617 	VI_UNLOCK(vp);
2618 	return (0);
2619 }
2620 
2621 /*
2622  * Eliminate all activity associated with a vnode
2623  * in preparation for reuse.
2624  */
2625 void
2626 vgone(vp)
2627 	register struct vnode *vp;
2628 {
2629 	struct thread *td = curthread;	/* XXX */
2630 
2631 	VI_LOCK(vp);
2632 	vgonel(vp, td);
2633 }
2634 
2635 /*
2636  * Disassociate a character device from the its underlying filesystem and
2637  * attach it to spec.  This is for use when the chr device is still active
2638  * and the filesystem is going away.
2639  */
2640 static void
2641 vgonechrl(struct vnode *vp, struct thread *td)
2642 {
2643 	ASSERT_VI_LOCKED(vp, "vgonechrl");
2644 	vx_lock(vp);
2645 	/*
2646 	 * This is a custom version of vclean() which does not tearm down
2647 	 * the bufs or vm objects held by this vnode.  This allows filesystems
2648 	 * to continue using devices which were discovered via another
2649 	 * filesystem that has been unmounted.
2650 	 */
2651 	if (vp->v_usecount != 0) {
2652 		v_incr_usecount(vp, 1);
2653 		/*
2654 		 * Ensure that no other activity can occur while the
2655 		 * underlying object is being cleaned out.
2656 		 */
2657 		VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2658 		/*
2659 		 * Any other processes trying to obtain this lock must first
2660 		 * wait for VXLOCK to clear, then call the new lock operation.
2661 		 */
2662 		VOP_UNLOCK(vp, 0, td);
2663 		vp->v_vnlock = &vp->v_lock;
2664 		vp->v_tag = "orphanchr";
2665 		vp->v_op = spec_vnodeop_p;
2666 		delmntque(vp);
2667 		cache_purge(vp);
2668 		vrele(vp);
2669 		VI_LOCK(vp);
2670 	} else
2671 		vclean(vp, 0, td);
2672 	vp->v_op = spec_vnodeop_p;
2673 	vx_unlock(vp);
2674 	VI_UNLOCK(vp);
2675 }
2676 
2677 /*
2678  * vgone, with the vp interlock held.
2679  */
2680 void
2681 vgonel(vp, td)
2682 	struct vnode *vp;
2683 	struct thread *td;
2684 {
2685 	/*
2686 	 * If a vgone (or vclean) is already in progress,
2687 	 * wait until it is done and return.
2688 	 */
2689 	ASSERT_VI_LOCKED(vp, "vgonel");
2690 	if (vp->v_iflag & VI_XLOCK) {
2691 		vp->v_iflag |= VI_XWANT;
2692 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0);
2693 		return;
2694 	}
2695 	vx_lock(vp);
2696 
2697 	/*
2698 	 * Clean out the filesystem specific data.
2699 	 */
2700 	vclean(vp, DOCLOSE, td);
2701 	VI_UNLOCK(vp);
2702 
2703 	/*
2704 	 * If special device, remove it from special device alias list
2705 	 * if it is on one.
2706 	 */
2707 	VI_LOCK(vp);
2708 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2709 		mtx_lock(&spechash_mtx);
2710 		SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext);
2711 		vp->v_rdev->si_usecount -= vp->v_usecount;
2712 		mtx_unlock(&spechash_mtx);
2713 		dev_rel(vp->v_rdev);
2714 		vp->v_rdev = NULL;
2715 	}
2716 
2717 	/*
2718 	 * If it is on the freelist and not already at the head,
2719 	 * move it to the head of the list. The test of the
2720 	 * VDOOMED flag and the reference count of zero is because
2721 	 * it will be removed from the free list by getnewvnode,
2722 	 * but will not have its reference count incremented until
2723 	 * after calling vgone. If the reference count were
2724 	 * incremented first, vgone would (incorrectly) try to
2725 	 * close the previous instance of the underlying object.
2726 	 */
2727 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2728 		mtx_lock(&vnode_free_list_mtx);
2729 		if (vp->v_iflag & VI_FREE) {
2730 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2731 		} else {
2732 			vp->v_iflag |= VI_FREE;
2733 			freevnodes++;
2734 		}
2735 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2736 		mtx_unlock(&vnode_free_list_mtx);
2737 	}
2738 
2739 	vp->v_type = VBAD;
2740 	vx_unlock(vp);
2741 	VI_UNLOCK(vp);
2742 }
2743 
2744 /*
2745  * Lookup a vnode by device number.
2746  */
2747 int
2748 vfinddev(dev, vpp)
2749 	struct cdev *dev;
2750 	struct vnode **vpp;
2751 {
2752 	struct vnode *vp;
2753 
2754 	mtx_lock(&spechash_mtx);
2755 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2756 		*vpp = vp;
2757 		mtx_unlock(&spechash_mtx);
2758 		return (1);
2759 	}
2760 	mtx_unlock(&spechash_mtx);
2761 	return (0);
2762 }
2763 
2764 /*
2765  * Calculate the total number of references to a special device.
2766  */
2767 int
2768 vcount(vp)
2769 	struct vnode *vp;
2770 {
2771 	int count;
2772 
2773 	mtx_lock(&spechash_mtx);
2774 	count = vp->v_rdev->si_usecount;
2775 	mtx_unlock(&spechash_mtx);
2776 	return (count);
2777 }
2778 
2779 /*
2780  * Same as above, but using the struct cdev *as argument
2781  */
2782 int
2783 count_dev(dev)
2784 	struct cdev *dev;
2785 {
2786 	int count;
2787 
2788 	mtx_lock(&spechash_mtx);
2789 	count = dev->si_usecount;
2790 	mtx_unlock(&spechash_mtx);
2791 	return(count);
2792 }
2793 
2794 /*
2795  * Print out a description of a vnode.
2796  */
2797 static char *typename[] =
2798 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2799 
2800 void
2801 vprint(label, vp)
2802 	char *label;
2803 	struct vnode *vp;
2804 {
2805 	char buf[96];
2806 
2807 	if (label != NULL)
2808 		printf("%s: %p: ", label, (void *)vp);
2809 	else
2810 		printf("%p: ", (void *)vp);
2811 	printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,",
2812 	    vp->v_tag, typename[vp->v_type], vp->v_usecount,
2813 	    vp->v_writecount, vp->v_holdcnt);
2814 	buf[0] = '\0';
2815 	if (vp->v_vflag & VV_ROOT)
2816 		strcat(buf, "|VV_ROOT");
2817 	if (vp->v_vflag & VV_TEXT)
2818 		strcat(buf, "|VV_TEXT");
2819 	if (vp->v_vflag & VV_SYSTEM)
2820 		strcat(buf, "|VV_SYSTEM");
2821 	if (vp->v_iflag & VI_XLOCK)
2822 		strcat(buf, "|VI_XLOCK");
2823 	if (vp->v_iflag & VI_XWANT)
2824 		strcat(buf, "|VI_XWANT");
2825 	if (vp->v_iflag & VI_BWAIT)
2826 		strcat(buf, "|VI_BWAIT");
2827 	if (vp->v_iflag & VI_DOOMED)
2828 		strcat(buf, "|VI_DOOMED");
2829 	if (vp->v_iflag & VI_FREE)
2830 		strcat(buf, "|VI_FREE");
2831 	if (vp->v_vflag & VV_OBJBUF)
2832 		strcat(buf, "|VV_OBJBUF");
2833 	if (buf[0] != '\0')
2834 		printf(" flags (%s),", &buf[1]);
2835 	lockmgr_printinfo(vp->v_vnlock);
2836 	printf("\n");
2837 	if (vp->v_data != NULL)
2838 		VOP_PRINT(vp);
2839 }
2840 
2841 #ifdef DDB
2842 #include <ddb/ddb.h>
2843 /*
2844  * List all of the locked vnodes in the system.
2845  * Called when debugging the kernel.
2846  */
2847 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2848 {
2849 	struct mount *mp, *nmp;
2850 	struct vnode *vp;
2851 
2852 	/*
2853 	 * Note: because this is DDB, we can't obey the locking semantics
2854 	 * for these structures, which means we could catch an inconsistent
2855 	 * state and dereference a nasty pointer.  Not much to be done
2856 	 * about that.
2857 	 */
2858 	printf("Locked vnodes\n");
2859 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2860 		nmp = TAILQ_NEXT(mp, mnt_list);
2861 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2862 			if (VOP_ISLOCKED(vp, NULL))
2863 				vprint(NULL, vp);
2864 		}
2865 		nmp = TAILQ_NEXT(mp, mnt_list);
2866 	}
2867 }
2868 #endif
2869 
2870 /*
2871  * Fill in a struct xvfsconf based on a struct vfsconf.
2872  */
2873 static void
2874 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2875 {
2876 
2877 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2878 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2879 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2880 	xvfsp->vfc_flags = vfsp->vfc_flags;
2881 	/*
2882 	 * These are unused in userland, we keep them
2883 	 * to not break binary compatibility.
2884 	 */
2885 	xvfsp->vfc_vfsops = NULL;
2886 	xvfsp->vfc_next = NULL;
2887 }
2888 
2889 /*
2890  * Top level filesystem related information gathering.
2891  */
2892 static int
2893 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2894 {
2895 	struct vfsconf *vfsp;
2896 	struct xvfsconf xvfsp;
2897 	int error;
2898 
2899 	error = 0;
2900 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2901 		vfsconf2x(vfsp, &xvfsp);
2902 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2903 		if (error)
2904 			break;
2905 	}
2906 	return (error);
2907 }
2908 
2909 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2910     "S,xvfsconf", "List of all configured filesystems");
2911 
2912 #ifndef BURN_BRIDGES
2913 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2914 
2915 static int
2916 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2917 {
2918 	int *name = (int *)arg1 - 1;	/* XXX */
2919 	u_int namelen = arg2 + 1;	/* XXX */
2920 	struct vfsconf *vfsp;
2921 	struct xvfsconf xvfsp;
2922 
2923 	printf("WARNING: userland calling deprecated sysctl, "
2924 	    "please rebuild world\n");
2925 
2926 #if 1 || defined(COMPAT_PRELITE2)
2927 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2928 	if (namelen == 1)
2929 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2930 #endif
2931 
2932 	switch (name[1]) {
2933 	case VFS_MAXTYPENUM:
2934 		if (namelen != 2)
2935 			return (ENOTDIR);
2936 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2937 	case VFS_CONF:
2938 		if (namelen != 3)
2939 			return (ENOTDIR);	/* overloaded */
2940 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2941 			if (vfsp->vfc_typenum == name[2])
2942 				break;
2943 		if (vfsp == NULL)
2944 			return (EOPNOTSUPP);
2945 		vfsconf2x(vfsp, &xvfsp);
2946 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2947 	}
2948 	return (EOPNOTSUPP);
2949 }
2950 
2951 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl,
2952 	"Generic filesystem");
2953 
2954 #if 1 || defined(COMPAT_PRELITE2)
2955 
2956 static int
2957 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2958 {
2959 	int error;
2960 	struct vfsconf *vfsp;
2961 	struct ovfsconf ovfs;
2962 
2963 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2964 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2965 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2966 		ovfs.vfc_index = vfsp->vfc_typenum;
2967 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2968 		ovfs.vfc_flags = vfsp->vfc_flags;
2969 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2970 		if (error)
2971 			return error;
2972 	}
2973 	return 0;
2974 }
2975 
2976 #endif /* 1 || COMPAT_PRELITE2 */
2977 #endif /* !BURN_BRIDGES */
2978 
2979 #define KINFO_VNODESLOP		10
2980 #ifdef notyet
2981 /*
2982  * Dump vnode list (via sysctl).
2983  */
2984 /* ARGSUSED */
2985 static int
2986 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2987 {
2988 	struct xvnode *xvn;
2989 	struct thread *td = req->td;
2990 	struct mount *mp;
2991 	struct vnode *vp;
2992 	int error, len, n;
2993 
2994 	/*
2995 	 * Stale numvnodes access is not fatal here.
2996 	 */
2997 	req->lock = 0;
2998 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2999 	if (!req->oldptr)
3000 		/* Make an estimate */
3001 		return (SYSCTL_OUT(req, 0, len));
3002 
3003 	error = sysctl_wire_old_buffer(req, 0);
3004 	if (error != 0)
3005 		return (error);
3006 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3007 	n = 0;
3008 	mtx_lock(&mountlist_mtx);
3009 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3010 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
3011 			continue;
3012 		MNT_ILOCK(mp);
3013 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3014 			if (n == len)
3015 				break;
3016 			vref(vp);
3017 			xvn[n].xv_size = sizeof *xvn;
3018 			xvn[n].xv_vnode = vp;
3019 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3020 			XV_COPY(usecount);
3021 			XV_COPY(writecount);
3022 			XV_COPY(holdcnt);
3023 			XV_COPY(id);
3024 			XV_COPY(mount);
3025 			XV_COPY(numoutput);
3026 			XV_COPY(type);
3027 #undef XV_COPY
3028 			xvn[n].xv_flag = vp->v_vflag;
3029 
3030 			switch (vp->v_type) {
3031 			case VREG:
3032 			case VDIR:
3033 			case VLNK:
3034 				xvn[n].xv_dev = vp->v_cachedfs;
3035 				xvn[n].xv_ino = vp->v_cachedid;
3036 				break;
3037 			case VBLK:
3038 			case VCHR:
3039 				if (vp->v_rdev == NULL) {
3040 					vrele(vp);
3041 					continue;
3042 				}
3043 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3044 				break;
3045 			case VSOCK:
3046 				xvn[n].xv_socket = vp->v_socket;
3047 				break;
3048 			case VFIFO:
3049 				xvn[n].xv_fifo = vp->v_fifoinfo;
3050 				break;
3051 			case VNON:
3052 			case VBAD:
3053 			default:
3054 				/* shouldn't happen? */
3055 				vrele(vp);
3056 				continue;
3057 			}
3058 			vrele(vp);
3059 			++n;
3060 		}
3061 		MNT_IUNLOCK(mp);
3062 		mtx_lock(&mountlist_mtx);
3063 		vfs_unbusy(mp, td);
3064 		if (n == len)
3065 			break;
3066 	}
3067 	mtx_unlock(&mountlist_mtx);
3068 
3069 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3070 	free(xvn, M_TEMP);
3071 	return (error);
3072 }
3073 
3074 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
3075 	0, 0, sysctl_vnode, "S,xvnode", "");
3076 #endif
3077 
3078 /*
3079  * Check to see if a filesystem is mounted on a block device.
3080  */
3081 int
3082 vfs_mountedon(vp)
3083 	struct vnode *vp;
3084 {
3085 
3086 	if (vp->v_rdev->si_mountpoint != NULL)
3087 		return (EBUSY);
3088 	return (0);
3089 }
3090 
3091 /*
3092  * Unmount all filesystems. The list is traversed in reverse order
3093  * of mounting to avoid dependencies.
3094  */
3095 void
3096 vfs_unmountall()
3097 {
3098 	struct mount *mp;
3099 	struct thread *td;
3100 	int error;
3101 
3102 	if (curthread != NULL)
3103 		td = curthread;
3104 	else
3105 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
3106 	/*
3107 	 * Since this only runs when rebooting, it is not interlocked.
3108 	 */
3109 	while(!TAILQ_EMPTY(&mountlist)) {
3110 		mp = TAILQ_LAST(&mountlist, mntlist);
3111 		error = dounmount(mp, MNT_FORCE, td);
3112 		if (error) {
3113 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
3114 			printf("unmount of %s failed (",
3115 			    mp->mnt_stat.f_mntonname);
3116 			if (error == EBUSY)
3117 				printf("BUSY)\n");
3118 			else
3119 				printf("%d)\n", error);
3120 		} else {
3121 			/* The unmount has removed mp from the mountlist */
3122 		}
3123 	}
3124 }
3125 
3126 /*
3127  * perform msync on all vnodes under a mount point
3128  * the mount point must be locked.
3129  */
3130 void
3131 vfs_msync(struct mount *mp, int flags)
3132 {
3133 	struct vnode *vp, *nvp;
3134 	struct vm_object *obj;
3135 	int tries;
3136 
3137 	GIANT_REQUIRED;
3138 
3139 	tries = 5;
3140 	MNT_ILOCK(mp);
3141 loop:
3142 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
3143 		if (vp->v_mount != mp) {
3144 			if (--tries > 0)
3145 				goto loop;
3146 			break;
3147 		}
3148 
3149 		VI_LOCK(vp);
3150 		if (vp->v_iflag & VI_XLOCK) {
3151 			VI_UNLOCK(vp);
3152 			continue;
3153 		}
3154 
3155 		if ((vp->v_iflag & VI_OBJDIRTY) &&
3156 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
3157 			MNT_IUNLOCK(mp);
3158 			if (!vget(vp,
3159 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3160 			    curthread)) {
3161 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3162 					vput(vp);
3163 					MNT_ILOCK(mp);
3164 					continue;
3165 				}
3166 
3167 				if (VOP_GETVOBJECT(vp, &obj) == 0) {
3168 					VM_OBJECT_LOCK(obj);
3169 					vm_object_page_clean(obj, 0, 0,
3170 					    flags == MNT_WAIT ?
3171 					    OBJPC_SYNC : OBJPC_NOSYNC);
3172 					VM_OBJECT_UNLOCK(obj);
3173 				}
3174 				vput(vp);
3175 			}
3176 			MNT_ILOCK(mp);
3177 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
3178 				if (--tries > 0)
3179 					goto loop;
3180 				break;
3181 			}
3182 		} else
3183 			VI_UNLOCK(vp);
3184 	}
3185 	MNT_IUNLOCK(mp);
3186 }
3187 
3188 /*
3189  * Create the VM object needed for VMIO and mmap support.  This
3190  * is done for all VREG files in the system.  Some filesystems might
3191  * afford the additional metadata buffering capability of the
3192  * VMIO code by making the device node be VMIO mode also.
3193  *
3194  * vp must be locked when vfs_object_create is called.
3195  */
3196 int
3197 vfs_object_create(vp, td, cred)
3198 	struct vnode *vp;
3199 	struct thread *td;
3200 	struct ucred *cred;
3201 {
3202 
3203 	GIANT_REQUIRED;
3204 	return (VOP_CREATEVOBJECT(vp, cred, td));
3205 }
3206 
3207 /*
3208  * Mark a vnode as free, putting it up for recycling.
3209  */
3210 void
3211 vfree(vp)
3212 	struct vnode *vp;
3213 {
3214 
3215 	ASSERT_VI_LOCKED(vp, "vfree");
3216 	mtx_lock(&vnode_free_list_mtx);
3217 	KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free"));
3218 	if (vp->v_iflag & VI_AGE) {
3219 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
3220 	} else {
3221 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
3222 	}
3223 	freevnodes++;
3224 	mtx_unlock(&vnode_free_list_mtx);
3225 	vp->v_iflag &= ~VI_AGE;
3226 	vp->v_iflag |= VI_FREE;
3227 }
3228 
3229 /*
3230  * Opposite of vfree() - mark a vnode as in use.
3231  */
3232 void
3233 vbusy(vp)
3234 	struct vnode *vp;
3235 {
3236 
3237 	ASSERT_VI_LOCKED(vp, "vbusy");
3238 	KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free"));
3239 
3240 	mtx_lock(&vnode_free_list_mtx);
3241 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
3242 	freevnodes--;
3243 	mtx_unlock(&vnode_free_list_mtx);
3244 
3245 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
3246 }
3247 
3248 /*
3249  * Initalize per-vnode helper structure to hold poll-related state.
3250  */
3251 void
3252 v_addpollinfo(struct vnode *vp)
3253 {
3254 	struct vpollinfo *vi;
3255 
3256 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
3257 	if (vp->v_pollinfo != NULL) {
3258 		uma_zfree(vnodepoll_zone, vi);
3259 		return;
3260 	}
3261 	vp->v_pollinfo = vi;
3262 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3263 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note,
3264 	    &vp->v_pollinfo->vpi_lock);
3265 }
3266 
3267 /*
3268  * Record a process's interest in events which might happen to
3269  * a vnode.  Because poll uses the historic select-style interface
3270  * internally, this routine serves as both the ``check for any
3271  * pending events'' and the ``record my interest in future events''
3272  * functions.  (These are done together, while the lock is held,
3273  * to avoid race conditions.)
3274  */
3275 int
3276 vn_pollrecord(vp, td, events)
3277 	struct vnode *vp;
3278 	struct thread *td;
3279 	short events;
3280 {
3281 
3282 	if (vp->v_pollinfo == NULL)
3283 		v_addpollinfo(vp);
3284 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3285 	if (vp->v_pollinfo->vpi_revents & events) {
3286 		/*
3287 		 * This leaves events we are not interested
3288 		 * in available for the other process which
3289 		 * which presumably had requested them
3290 		 * (otherwise they would never have been
3291 		 * recorded).
3292 		 */
3293 		events &= vp->v_pollinfo->vpi_revents;
3294 		vp->v_pollinfo->vpi_revents &= ~events;
3295 
3296 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3297 		return events;
3298 	}
3299 	vp->v_pollinfo->vpi_events |= events;
3300 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3301 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3302 	return 0;
3303 }
3304 
3305 /*
3306  * Note the occurrence of an event.  If the VN_POLLEVENT macro is used,
3307  * it is possible for us to miss an event due to race conditions, but
3308  * that condition is expected to be rare, so for the moment it is the
3309  * preferred interface.
3310  */
3311 void
3312 vn_pollevent(vp, events)
3313 	struct vnode *vp;
3314 	short events;
3315 {
3316 
3317 	if (vp->v_pollinfo == NULL)
3318 		v_addpollinfo(vp);
3319 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3320 	if (vp->v_pollinfo->vpi_events & events) {
3321 		/*
3322 		 * We clear vpi_events so that we don't
3323 		 * call selwakeup() twice if two events are
3324 		 * posted before the polling process(es) is
3325 		 * awakened.  This also ensures that we take at
3326 		 * most one selwakeup() if the polling process
3327 		 * is no longer interested.  However, it does
3328 		 * mean that only one event can be noticed at
3329 		 * a time.  (Perhaps we should only clear those
3330 		 * event bits which we note?) XXX
3331 		 */
3332 		vp->v_pollinfo->vpi_events = 0;	/* &= ~events ??? */
3333 		vp->v_pollinfo->vpi_revents |= events;
3334 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
3335 	}
3336 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3337 }
3338 
3339 /*
3340  * Wake up anyone polling on vp because it is being revoked.
3341  * This depends on dead_poll() returning POLLHUP for correct
3342  * behavior.
3343  */
3344 void
3345 vn_pollgone(vp)
3346 	struct vnode *vp;
3347 {
3348 
3349 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3350 	VN_KNOTE_LOCKED(vp, NOTE_REVOKE);
3351 	if (vp->v_pollinfo->vpi_events) {
3352 		vp->v_pollinfo->vpi_events = 0;
3353 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
3354 	}
3355 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3356 }
3357 
3358 
3359 
3360 /*
3361  * Routine to create and manage a filesystem syncer vnode.
3362  */
3363 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3364 static int	sync_fsync(struct  vop_fsync_args *);
3365 static int	sync_inactive(struct  vop_inactive_args *);
3366 static int	sync_reclaim(struct  vop_reclaim_args *);
3367 
3368 static vop_t **sync_vnodeop_p;
3369 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = {
3370 	{ &vop_default_desc,	(vop_t *) vop_eopnotsupp },
3371 	{ &vop_close_desc,	(vop_t *) sync_close },		/* close */
3372 	{ &vop_fsync_desc,	(vop_t *) sync_fsync },		/* fsync */
3373 	{ &vop_inactive_desc,	(vop_t *) sync_inactive },	/* inactive */
3374 	{ &vop_reclaim_desc,	(vop_t *) sync_reclaim },	/* reclaim */
3375 	{ &vop_lock_desc,	(vop_t *) vop_stdlock },	/* lock */
3376 	{ &vop_unlock_desc,	(vop_t *) vop_stdunlock },	/* unlock */
3377 	{ &vop_islocked_desc,	(vop_t *) vop_stdislocked },	/* islocked */
3378 	{ NULL, NULL }
3379 };
3380 static struct vnodeopv_desc sync_vnodeop_opv_desc =
3381 	{ &sync_vnodeop_p, sync_vnodeop_entries };
3382 
3383 VNODEOP_SET(sync_vnodeop_opv_desc);
3384 
3385 /*
3386  * Create a new filesystem syncer vnode for the specified mount point.
3387  */
3388 int
3389 vfs_allocate_syncvnode(mp)
3390 	struct mount *mp;
3391 {
3392 	struct vnode *vp;
3393 	static long start, incr, next;
3394 	int error;
3395 
3396 	/* Allocate a new vnode */
3397 	if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) {
3398 		mp->mnt_syncer = NULL;
3399 		return (error);
3400 	}
3401 	vp->v_type = VNON;
3402 	/*
3403 	 * Place the vnode onto the syncer worklist. We attempt to
3404 	 * scatter them about on the list so that they will go off
3405 	 * at evenly distributed times even if all the filesystems
3406 	 * are mounted at once.
3407 	 */
3408 	next += incr;
3409 	if (next == 0 || next > syncer_maxdelay) {
3410 		start /= 2;
3411 		incr /= 2;
3412 		if (start == 0) {
3413 			start = syncer_maxdelay / 2;
3414 			incr = syncer_maxdelay;
3415 		}
3416 		next = start;
3417 	}
3418 	VI_LOCK(vp);
3419 	vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0);
3420 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3421 	mtx_lock(&sync_mtx);
3422 	sync_vnode_count++;
3423 	mtx_unlock(&sync_mtx);
3424 	VI_UNLOCK(vp);
3425 	mp->mnt_syncer = vp;
3426 	return (0);
3427 }
3428 
3429 /*
3430  * Do a lazy sync of the filesystem.
3431  */
3432 static int
3433 sync_fsync(ap)
3434 	struct vop_fsync_args /* {
3435 		struct vnode *a_vp;
3436 		struct ucred *a_cred;
3437 		int a_waitfor;
3438 		struct thread *a_td;
3439 	} */ *ap;
3440 {
3441 	struct vnode *syncvp = ap->a_vp;
3442 	struct mount *mp = syncvp->v_mount;
3443 	struct thread *td = ap->a_td;
3444 	int error, asyncflag;
3445 
3446 	/*
3447 	 * We only need to do something if this is a lazy evaluation.
3448 	 */
3449 	if (ap->a_waitfor != MNT_LAZY)
3450 		return (0);
3451 
3452 	/*
3453 	 * Move ourselves to the back of the sync list.
3454 	 */
3455 	VI_LOCK(syncvp);
3456 	vn_syncer_add_to_worklist(syncvp, syncdelay);
3457 	VI_UNLOCK(syncvp);
3458 
3459 	/*
3460 	 * Walk the list of vnodes pushing all that are dirty and
3461 	 * not already on the sync list.
3462 	 */
3463 	mtx_lock(&mountlist_mtx);
3464 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3465 		mtx_unlock(&mountlist_mtx);
3466 		return (0);
3467 	}
3468 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3469 		vfs_unbusy(mp, td);
3470 		return (0);
3471 	}
3472 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3473 	mp->mnt_flag &= ~MNT_ASYNC;
3474 	vfs_msync(mp, MNT_NOWAIT);
3475 	error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td);
3476 	if (asyncflag)
3477 		mp->mnt_flag |= MNT_ASYNC;
3478 	vn_finished_write(mp);
3479 	vfs_unbusy(mp, td);
3480 	return (error);
3481 }
3482 
3483 /*
3484  * The syncer vnode is no referenced.
3485  */
3486 static int
3487 sync_inactive(ap)
3488 	struct vop_inactive_args /* {
3489 		struct vnode *a_vp;
3490 		struct thread *a_td;
3491 	} */ *ap;
3492 {
3493 
3494 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3495 	vgone(ap->a_vp);
3496 	return (0);
3497 }
3498 
3499 /*
3500  * The syncer vnode is no longer needed and is being decommissioned.
3501  *
3502  * Modifications to the worklist must be protected by sync_mtx.
3503  */
3504 static int
3505 sync_reclaim(ap)
3506 	struct vop_reclaim_args /* {
3507 		struct vnode *a_vp;
3508 	} */ *ap;
3509 {
3510 	struct vnode *vp = ap->a_vp;
3511 
3512 	VI_LOCK(vp);
3513 	vp->v_mount->mnt_syncer = NULL;
3514 	if (vp->v_iflag & VI_ONWORKLST) {
3515 		mtx_lock(&sync_mtx);
3516 		LIST_REMOVE(vp, v_synclist);
3517  		syncer_worklist_len--;
3518 		sync_vnode_count--;
3519 		mtx_unlock(&sync_mtx);
3520 		vp->v_iflag &= ~VI_ONWORKLST;
3521 	}
3522 	VI_UNLOCK(vp);
3523 
3524 	return (0);
3525 }
3526 
3527 /*
3528  * extract the struct cdev *from a VCHR
3529  */
3530 struct cdev *
3531 vn_todev(vp)
3532 	struct vnode *vp;
3533 {
3534 
3535 	if (vp->v_type != VCHR)
3536 		return (NULL);
3537 	return (vp->v_rdev);
3538 }
3539 
3540 /*
3541  * Check if vnode represents a disk device
3542  */
3543 int
3544 vn_isdisk(vp, errp)
3545 	struct vnode *vp;
3546 	int *errp;
3547 {
3548 	int error;
3549 
3550 	error = 0;
3551 	if (vp->v_type != VCHR)
3552 		error = ENOTBLK;
3553 	else if (vp->v_rdev == NULL)
3554 		error = ENXIO;
3555 	else if (!(devsw(vp->v_rdev)->d_flags & D_DISK))
3556 		error = ENOTBLK;
3557 	if (errp != NULL)
3558 		*errp = error;
3559 	return (error == 0);
3560 }
3561 
3562 /*
3563  * Free data allocated by namei(); see namei(9) for details.
3564  */
3565 void
3566 NDFREE(ndp, flags)
3567      struct nameidata *ndp;
3568      const u_int flags;
3569 {
3570 
3571 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3572 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3573 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3574 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3575 	}
3576 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3577 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3578 	    ndp->ni_dvp != ndp->ni_vp)
3579 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3580 	if (!(flags & NDF_NO_DVP_RELE) &&
3581 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3582 		vrele(ndp->ni_dvp);
3583 		ndp->ni_dvp = NULL;
3584 	}
3585 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3586 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3587 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3588 	if (!(flags & NDF_NO_VP_RELE) &&
3589 	    ndp->ni_vp) {
3590 		vrele(ndp->ni_vp);
3591 		ndp->ni_vp = NULL;
3592 	}
3593 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3594 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3595 		vrele(ndp->ni_startdir);
3596 		ndp->ni_startdir = NULL;
3597 	}
3598 }
3599 
3600 /*
3601  * Common filesystem object access control check routine.  Accepts a
3602  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3603  * and optional call-by-reference privused argument allowing vaccess()
3604  * to indicate to the caller whether privilege was used to satisfy the
3605  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3606  */
3607 int
3608 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3609 	enum vtype type;
3610 	mode_t file_mode;
3611 	uid_t file_uid;
3612 	gid_t file_gid;
3613 	mode_t acc_mode;
3614 	struct ucred *cred;
3615 	int *privused;
3616 {
3617 	mode_t dac_granted;
3618 #ifdef CAPABILITIES
3619 	mode_t cap_granted;
3620 #endif
3621 
3622 	/*
3623 	 * Look for a normal, non-privileged way to access the file/directory
3624 	 * as requested.  If it exists, go with that.
3625 	 */
3626 
3627 	if (privused != NULL)
3628 		*privused = 0;
3629 
3630 	dac_granted = 0;
3631 
3632 	/* Check the owner. */
3633 	if (cred->cr_uid == file_uid) {
3634 		dac_granted |= VADMIN;
3635 		if (file_mode & S_IXUSR)
3636 			dac_granted |= VEXEC;
3637 		if (file_mode & S_IRUSR)
3638 			dac_granted |= VREAD;
3639 		if (file_mode & S_IWUSR)
3640 			dac_granted |= (VWRITE | VAPPEND);
3641 
3642 		if ((acc_mode & dac_granted) == acc_mode)
3643 			return (0);
3644 
3645 		goto privcheck;
3646 	}
3647 
3648 	/* Otherwise, check the groups (first match) */
3649 	if (groupmember(file_gid, cred)) {
3650 		if (file_mode & S_IXGRP)
3651 			dac_granted |= VEXEC;
3652 		if (file_mode & S_IRGRP)
3653 			dac_granted |= VREAD;
3654 		if (file_mode & S_IWGRP)
3655 			dac_granted |= (VWRITE | VAPPEND);
3656 
3657 		if ((acc_mode & dac_granted) == acc_mode)
3658 			return (0);
3659 
3660 		goto privcheck;
3661 	}
3662 
3663 	/* Otherwise, check everyone else. */
3664 	if (file_mode & S_IXOTH)
3665 		dac_granted |= VEXEC;
3666 	if (file_mode & S_IROTH)
3667 		dac_granted |= VREAD;
3668 	if (file_mode & S_IWOTH)
3669 		dac_granted |= (VWRITE | VAPPEND);
3670 	if ((acc_mode & dac_granted) == acc_mode)
3671 		return (0);
3672 
3673 privcheck:
3674 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3675 		/* XXX audit: privilege used */
3676 		if (privused != NULL)
3677 			*privused = 1;
3678 		return (0);
3679 	}
3680 
3681 #ifdef CAPABILITIES
3682 	/*
3683 	 * Build a capability mask to determine if the set of capabilities
3684 	 * satisfies the requirements when combined with the granted mask
3685 	 * from above.
3686 	 * For each capability, if the capability is required, bitwise
3687 	 * or the request type onto the cap_granted mask.
3688 	 */
3689 	cap_granted = 0;
3690 
3691 	if (type == VDIR) {
3692 		/*
3693 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3694 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3695 		 */
3696 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3697 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3698 			cap_granted |= VEXEC;
3699 	} else {
3700 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3701 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3702 			cap_granted |= VEXEC;
3703 	}
3704 
3705 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3706 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3707 		cap_granted |= VREAD;
3708 
3709 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3710 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3711 		cap_granted |= (VWRITE | VAPPEND);
3712 
3713 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3714 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3715 		cap_granted |= VADMIN;
3716 
3717 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3718 		/* XXX audit: privilege used */
3719 		if (privused != NULL)
3720 			*privused = 1;
3721 		return (0);
3722 	}
3723 #endif
3724 
3725 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3726 }
3727 
3728 /*
3729  * Credential check based on process requesting service, and per-attribute
3730  * permissions.
3731  */
3732 int
3733 extattr_check_cred(struct vnode *vp, int attrnamespace,
3734     struct ucred *cred, struct thread *td, int access)
3735 {
3736 
3737 	/*
3738 	 * Kernel-invoked always succeeds.
3739 	 */
3740 	if (cred == NOCRED)
3741 		return (0);
3742 
3743 	/*
3744 	 * Do not allow privileged processes in jail to directly
3745 	 * manipulate system attributes.
3746 	 *
3747 	 * XXX What capability should apply here?
3748 	 * Probably CAP_SYS_SETFFLAG.
3749 	 */
3750 	switch (attrnamespace) {
3751 	case EXTATTR_NAMESPACE_SYSTEM:
3752 		/* Potentially should be: return (EPERM); */
3753 		return (suser_cred(cred, 0));
3754 	case EXTATTR_NAMESPACE_USER:
3755 		return (VOP_ACCESS(vp, access, cred, td));
3756 	default:
3757 		return (EPERM);
3758 	}
3759 }
3760 
3761 #ifdef DEBUG_VFS_LOCKS
3762 /*
3763  * This only exists to supress warnings from unlocked specfs accesses.  It is
3764  * no longer ok to have an unlocked VFS.
3765  */
3766 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3767 
3768 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3769 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3770 
3771 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3772 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3773 
3774 int vfs_badlock_print = 1;	/* Print lock violations. */
3775 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3776 
3777 #ifdef KDB
3778 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3779 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3780 #endif
3781 
3782 static void
3783 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3784 {
3785 
3786 #ifdef KDB
3787 	if (vfs_badlock_backtrace)
3788 		kdb_backtrace();
3789 #endif
3790 	if (vfs_badlock_print)
3791 		printf("%s: %p %s\n", str, (void *)vp, msg);
3792 	if (vfs_badlock_ddb)
3793 		kdb_enter("lock violation");
3794 }
3795 
3796 void
3797 assert_vi_locked(struct vnode *vp, const char *str)
3798 {
3799 
3800 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3801 		vfs_badlock("interlock is not locked but should be", str, vp);
3802 }
3803 
3804 void
3805 assert_vi_unlocked(struct vnode *vp, const char *str)
3806 {
3807 
3808 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3809 		vfs_badlock("interlock is locked but should not be", str, vp);
3810 }
3811 
3812 void
3813 assert_vop_locked(struct vnode *vp, const char *str)
3814 {
3815 
3816 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3817 		vfs_badlock("is not locked but should be", str, vp);
3818 }
3819 
3820 void
3821 assert_vop_unlocked(struct vnode *vp, const char *str)
3822 {
3823 
3824 	if (vp && !IGNORE_LOCK(vp) &&
3825 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3826 		vfs_badlock("is locked but should not be", str, vp);
3827 }
3828 
3829 #if 0
3830 void
3831 assert_vop_elocked(struct vnode *vp, const char *str)
3832 {
3833 
3834 	if (vp && !IGNORE_LOCK(vp) &&
3835 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3836 		vfs_badlock("is not exclusive locked but should be", str, vp);
3837 }
3838 
3839 void
3840 assert_vop_elocked_other(struct vnode *vp, const char *str)
3841 {
3842 
3843 	if (vp && !IGNORE_LOCK(vp) &&
3844 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3845 		vfs_badlock("is not exclusive locked by another thread",
3846 		    str, vp);
3847 }
3848 
3849 void
3850 assert_vop_slocked(struct vnode *vp, const char *str)
3851 {
3852 
3853 	if (vp && !IGNORE_LOCK(vp) &&
3854 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3855 		vfs_badlock("is not locked shared but should be", str, vp);
3856 }
3857 #endif /* 0 */
3858 
3859 void
3860 vop_rename_pre(void *ap)
3861 {
3862 	struct vop_rename_args *a = ap;
3863 
3864 	if (a->a_tvp)
3865 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3866 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3867 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3868 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3869 
3870 	/* Check the source (from). */
3871 	if (a->a_tdvp != a->a_fdvp)
3872 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3873 	if (a->a_tvp != a->a_fvp)
3874 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3875 
3876 	/* Check the target. */
3877 	if (a->a_tvp)
3878 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3879 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3880 }
3881 
3882 void
3883 vop_strategy_pre(void *ap)
3884 {
3885 	struct vop_strategy_args *a;
3886 	struct buf *bp;
3887 
3888 	a = ap;
3889 	bp = a->a_bp;
3890 
3891 	/*
3892 	 * Cluster ops lock their component buffers but not the IO container.
3893 	 */
3894 	if ((bp->b_flags & B_CLUSTER) != 0)
3895 		return;
3896 
3897 	if (BUF_REFCNT(bp) < 1) {
3898 		if (vfs_badlock_print)
3899 			printf(
3900 			    "VOP_STRATEGY: bp is not locked but should be\n");
3901 		if (vfs_badlock_ddb)
3902 			kdb_enter("lock violation");
3903 	}
3904 }
3905 
3906 void
3907 vop_lookup_pre(void *ap)
3908 {
3909 	struct vop_lookup_args *a;
3910 	struct vnode *dvp;
3911 
3912 	a = ap;
3913 	dvp = a->a_dvp;
3914 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3915 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3916 }
3917 
3918 void
3919 vop_lookup_post(void *ap, int rc)
3920 {
3921 	struct vop_lookup_args *a;
3922 	struct componentname *cnp;
3923 	struct vnode *dvp;
3924 	struct vnode *vp;
3925 	int flags;
3926 
3927 	a = ap;
3928 	dvp = a->a_dvp;
3929 	cnp = a->a_cnp;
3930 	vp = *(a->a_vpp);
3931 	flags = cnp->cn_flags;
3932 
3933 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3934 
3935 	/*
3936 	 * If this is the last path component for this lookup and LOCKPARENT
3937 	 * is set, OR if there is an error the directory has to be locked.
3938 	 */
3939 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
3940 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
3941 	else if (rc != 0)
3942 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
3943 	else if (dvp != vp)
3944 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
3945 	if (flags & PDIRUNLOCK)
3946 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
3947 }
3948 
3949 void
3950 vop_lock_pre(void *ap)
3951 {
3952 	struct vop_lock_args *a = ap;
3953 
3954 	if ((a->a_flags & LK_INTERLOCK) == 0)
3955 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3956 	else
3957 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3958 }
3959 
3960 void
3961 vop_lock_post(void *ap, int rc)
3962 {
3963 	struct vop_lock_args *a = ap;
3964 
3965 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3966 	if (rc == 0)
3967 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3968 }
3969 
3970 void
3971 vop_unlock_pre(void *ap)
3972 {
3973 	struct vop_unlock_args *a = ap;
3974 
3975 	if (a->a_flags & LK_INTERLOCK)
3976 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3977 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3978 }
3979 
3980 void
3981 vop_unlock_post(void *ap, int rc)
3982 {
3983 	struct vop_unlock_args *a = ap;
3984 
3985 	if (a->a_flags & LK_INTERLOCK)
3986 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3987 }
3988 #endif /* DEBUG_VFS_LOCKS */
3989 
3990 static struct knlist fs_knlist;
3991 
3992 static void
3993 vfs_event_init(void *arg)
3994 {
3995 	knlist_init(&fs_knlist, NULL);
3996 }
3997 /* XXX - correct order? */
3998 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3999 
4000 void
4001 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
4002 {
4003 
4004 	KNOTE_UNLOCKED(&fs_knlist, event);
4005 }
4006 
4007 static int	filt_fsattach(struct knote *kn);
4008 static void	filt_fsdetach(struct knote *kn);
4009 static int	filt_fsevent(struct knote *kn, long hint);
4010 
4011 struct filterops fs_filtops =
4012 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
4013 
4014 static int
4015 filt_fsattach(struct knote *kn)
4016 {
4017 
4018 	kn->kn_flags |= EV_CLEAR;
4019 	knlist_add(&fs_knlist, kn, 0);
4020 	return (0);
4021 }
4022 
4023 static void
4024 filt_fsdetach(struct knote *kn)
4025 {
4026 
4027 	knlist_remove(&fs_knlist, kn, 0);
4028 }
4029 
4030 static int
4031 filt_fsevent(struct knote *kn, long hint)
4032 {
4033 
4034 	kn->kn_fflags |= hint;
4035 	return (kn->kn_fflags != 0);
4036 }
4037 
4038 static int
4039 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4040 {
4041 	struct vfsidctl vc;
4042 	int error;
4043 	struct mount *mp;
4044 
4045 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4046 	if (error)
4047 		return (error);
4048 	if (vc.vc_vers != VFS_CTL_VERS1)
4049 		return (EINVAL);
4050 	mp = vfs_getvfs(&vc.vc_fsid);
4051 	if (mp == NULL)
4052 		return (ENOENT);
4053 	/* ensure that a specific sysctl goes to the right filesystem. */
4054 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4055 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4056 		return (EINVAL);
4057 	}
4058 	VCTLTOREQ(&vc, req);
4059 	return (VFS_SYSCTL(mp, vc.vc_op, req));
4060 }
4061 
4062 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
4063         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
4064 
4065 /*
4066  * Function to initialize a va_filerev field sensibly.
4067  * XXX: Wouldn't a random number make a lot more sense ??
4068  */
4069 u_quad_t
4070 init_va_filerev(void)
4071 {
4072 	struct bintime bt;
4073 
4074 	getbinuptime(&bt);
4075 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4076 }
4077