xref: /freebsd/sys/kern/vfs_subr.c (revision fbf96e52bbd90bbbb9c9e2ae6fbc101fa6ebd080)
1 /*
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ddb.h"
45 #include "opt_mac.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/conf.h>
52 #include <sys/event.h>
53 #include <sys/eventhandler.h>
54 #include <sys/extattr.h>
55 #include <sys/fcntl.h>
56 #include <sys/kdb.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/mac.h>
60 #include <sys/malloc.h>
61 #include <sys/mount.h>
62 #include <sys/namei.h>
63 #include <sys/sleepqueue.h>
64 #include <sys/stat.h>
65 #include <sys/sysctl.h>
66 #include <sys/syslog.h>
67 #include <sys/vmmeter.h>
68 #include <sys/vnode.h>
69 
70 #include <vm/vm.h>
71 #include <vm/vm_object.h>
72 #include <vm/vm_extern.h>
73 #include <vm/pmap.h>
74 #include <vm/vm_map.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_kern.h>
77 #include <vm/uma.h>
78 
79 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
80 
81 static void	addalias(struct vnode *vp, struct cdev *nvp_rdev);
82 static void	delmntque(struct vnode *vp);
83 static void	insmntque(struct vnode *vp, struct mount *mp);
84 static void	vclean(struct vnode *vp, int flags, struct thread *td);
85 static void	vlruvp(struct vnode *vp);
86 static int	flushbuflist(struct buf *blist, int flags, struct vnode *vp,
87 		    int slpflag, int slptimeo, int *errorp);
88 static void	syncer_shutdown(void *arg, int howto);
89 static int	vtryrecycle(struct vnode *vp);
90 static void	vx_lock(struct vnode *vp);
91 static void	vx_unlock(struct vnode *vp);
92 static void	vgonechrl(struct vnode *vp, struct thread *td);
93 
94 
95 /*
96  * Number of vnodes in existence.  Increased whenever getnewvnode()
97  * allocates a new vnode, never decreased.
98  */
99 static unsigned long	numvnodes;
100 
101 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
102 
103 /*
104  * Conversion tables for conversion from vnode types to inode formats
105  * and back.
106  */
107 enum vtype iftovt_tab[16] = {
108 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
109 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
110 };
111 int vttoif_tab[9] = {
112 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
113 	S_IFSOCK, S_IFIFO, S_IFMT,
114 };
115 
116 /*
117  * List of vnodes that are ready for recycling.
118  */
119 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
120 
121 /*
122  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
123  * getnewvnode() will return a newly allocated vnode.
124  */
125 static u_long wantfreevnodes = 25;
126 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
127 /* Number of vnodes in the free list. */
128 static u_long freevnodes;
129 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
130 
131 /*
132  * Various variables used for debugging the new implementation of
133  * reassignbuf().
134  * XXX these are probably of (very) limited utility now.
135  */
136 static int reassignbufcalls;
137 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
138 static int nameileafonly;
139 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
140 
141 /*
142  * Cache for the mount type id assigned to NFS.  This is used for
143  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
144  */
145 int	nfs_mount_type = -1;
146 
147 /* To keep more than one thread at a time from running vfs_getnewfsid */
148 static struct mtx mntid_mtx;
149 
150 /*
151  * Lock for any access to the following:
152  *	vnode_free_list
153  *	numvnodes
154  *	freevnodes
155  */
156 static struct mtx vnode_free_list_mtx;
157 
158 /*
159  * For any iteration/modification of dev->si_hlist (linked through
160  * v_specnext)
161  */
162 static struct mtx spechash_mtx;
163 
164 /* Publicly exported FS */
165 struct nfs_public nfs_pub;
166 
167 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
168 static uma_zone_t vnode_zone;
169 static uma_zone_t vnodepoll_zone;
170 
171 /* Set to 1 to print out reclaim of active vnodes */
172 int	prtactive;
173 
174 /*
175  * The workitem queue.
176  *
177  * It is useful to delay writes of file data and filesystem metadata
178  * for tens of seconds so that quickly created and deleted files need
179  * not waste disk bandwidth being created and removed. To realize this,
180  * we append vnodes to a "workitem" queue. When running with a soft
181  * updates implementation, most pending metadata dependencies should
182  * not wait for more than a few seconds. Thus, mounted on block devices
183  * are delayed only about a half the time that file data is delayed.
184  * Similarly, directory updates are more critical, so are only delayed
185  * about a third the time that file data is delayed. Thus, there are
186  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
187  * one each second (driven off the filesystem syncer process). The
188  * syncer_delayno variable indicates the next queue that is to be processed.
189  * Items that need to be processed soon are placed in this queue:
190  *
191  *	syncer_workitem_pending[syncer_delayno]
192  *
193  * A delay of fifteen seconds is done by placing the request fifteen
194  * entries later in the queue:
195  *
196  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
197  *
198  */
199 static int syncer_delayno;
200 static long syncer_mask;
201 LIST_HEAD(synclist, vnode);
202 static struct synclist *syncer_workitem_pending;
203 /*
204  * The sync_mtx protects:
205  *	vp->v_synclist
206  *	sync_vnode_count
207  *	syncer_delayno
208  *	syncer_state
209  *	syncer_workitem_pending
210  *	syncer_worklist_len
211  *	rushjob
212  */
213 static struct mtx sync_mtx;
214 
215 #define SYNCER_MAXDELAY		32
216 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
217 static int syncdelay = 30;		/* max time to delay syncing data */
218 static int filedelay = 30;		/* time to delay syncing files */
219 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
220 static int dirdelay = 29;		/* time to delay syncing directories */
221 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
222 static int metadelay = 28;		/* time to delay syncing metadata */
223 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
224 static int rushjob;		/* number of slots to run ASAP */
225 static int stat_rush_requests;	/* number of times I/O speeded up */
226 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
227 
228 /*
229  * When shutting down the syncer, run it at four times normal speed.
230  */
231 #define SYNCER_SHUTDOWN_SPEEDUP		4
232 static int sync_vnode_count;
233 static int syncer_worklist_len;
234 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
235     syncer_state;
236 
237 /*
238  * Number of vnodes we want to exist at any one time.  This is mostly used
239  * to size hash tables in vnode-related code.  It is normally not used in
240  * getnewvnode(), as wantfreevnodes is normally nonzero.)
241  *
242  * XXX desiredvnodes is historical cruft and should not exist.
243  */
244 int desiredvnodes;
245 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
246     &desiredvnodes, 0, "Maximum number of vnodes");
247 static int minvnodes;
248 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
249     &minvnodes, 0, "Minimum number of vnodes");
250 static int vnlru_nowhere;
251 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
252     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
253 
254 /* Hook for calling soft updates. */
255 int (*softdep_process_worklist_hook)(struct mount *);
256 
257 /*
258  * Initialize the vnode management data structures.
259  */
260 static void
261 vntblinit(void *dummy __unused)
262 {
263 
264 	/*
265 	 * Desiredvnodes is a function of the physical memory size and
266 	 * the kernel's heap size.  Specifically, desiredvnodes scales
267 	 * in proportion to the physical memory size until two fifths
268 	 * of the kernel's heap size is consumed by vnodes and vm
269 	 * objects.
270 	 */
271 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
272 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
273 	minvnodes = desiredvnodes / 4;
274 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
275 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
276 	mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF);
277 	TAILQ_INIT(&vnode_free_list);
278 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
279 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
280 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
281 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
282 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
283 	/*
284 	 * Initialize the filesystem syncer.
285 	 */
286 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
287 		&syncer_mask);
288 	syncer_maxdelay = syncer_mask + 1;
289 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
290 }
291 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
292 
293 
294 /*
295  * Mark a mount point as busy. Used to synchronize access and to delay
296  * unmounting. Interlock is not released on failure.
297  */
298 int
299 vfs_busy(mp, flags, interlkp, td)
300 	struct mount *mp;
301 	int flags;
302 	struct mtx *interlkp;
303 	struct thread *td;
304 {
305 	int lkflags;
306 
307 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
308 		if (flags & LK_NOWAIT)
309 			return (ENOENT);
310 		mp->mnt_kern_flag |= MNTK_MWAIT;
311 		/*
312 		 * Since all busy locks are shared except the exclusive
313 		 * lock granted when unmounting, the only place that a
314 		 * wakeup needs to be done is at the release of the
315 		 * exclusive lock at the end of dounmount.
316 		 */
317 		msleep(mp, interlkp, PVFS, "vfs_busy", 0);
318 		return (ENOENT);
319 	}
320 	lkflags = LK_SHARED | LK_NOPAUSE;
321 	if (interlkp)
322 		lkflags |= LK_INTERLOCK;
323 	if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td))
324 		panic("vfs_busy: unexpected lock failure");
325 	return (0);
326 }
327 
328 /*
329  * Free a busy filesystem.
330  */
331 void
332 vfs_unbusy(mp, td)
333 	struct mount *mp;
334 	struct thread *td;
335 {
336 
337 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
338 }
339 
340 /*
341  * Lookup a mount point by filesystem identifier.
342  */
343 struct mount *
344 vfs_getvfs(fsid)
345 	fsid_t *fsid;
346 {
347 	register struct mount *mp;
348 
349 	mtx_lock(&mountlist_mtx);
350 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
351 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
352 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
353 			mtx_unlock(&mountlist_mtx);
354 			return (mp);
355 		}
356 	}
357 	mtx_unlock(&mountlist_mtx);
358 	return ((struct mount *) 0);
359 }
360 
361 /*
362  * Check if a user can access priveledged mount options.
363  */
364 int
365 vfs_suser(struct mount *mp, struct thread *td)
366 {
367 	int error;
368 
369 	if ((mp->mnt_flag & MNT_USER) == 0 ||
370 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
371 		if ((error = suser(td)) != 0)
372 			return (error);
373 	}
374 	return (0);
375 }
376 
377 /*
378  * Get a new unique fsid.  Try to make its val[0] unique, since this value
379  * will be used to create fake device numbers for stat().  Also try (but
380  * not so hard) make its val[0] unique mod 2^16, since some emulators only
381  * support 16-bit device numbers.  We end up with unique val[0]'s for the
382  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
383  *
384  * Keep in mind that several mounts may be running in parallel.  Starting
385  * the search one past where the previous search terminated is both a
386  * micro-optimization and a defense against returning the same fsid to
387  * different mounts.
388  */
389 void
390 vfs_getnewfsid(mp)
391 	struct mount *mp;
392 {
393 	static u_int16_t mntid_base;
394 	fsid_t tfsid;
395 	int mtype;
396 
397 	mtx_lock(&mntid_mtx);
398 	mtype = mp->mnt_vfc->vfc_typenum;
399 	tfsid.val[1] = mtype;
400 	mtype = (mtype & 0xFF) << 24;
401 	for (;;) {
402 		tfsid.val[0] = makedev(255,
403 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
404 		mntid_base++;
405 		if (vfs_getvfs(&tfsid) == NULL)
406 			break;
407 	}
408 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
409 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
410 	mtx_unlock(&mntid_mtx);
411 }
412 
413 /*
414  * Knob to control the precision of file timestamps:
415  *
416  *   0 = seconds only; nanoseconds zeroed.
417  *   1 = seconds and nanoseconds, accurate within 1/HZ.
418  *   2 = seconds and nanoseconds, truncated to microseconds.
419  * >=3 = seconds and nanoseconds, maximum precision.
420  */
421 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
422 
423 static int timestamp_precision = TSP_SEC;
424 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
425     &timestamp_precision, 0, "");
426 
427 /*
428  * Get a current timestamp.
429  */
430 void
431 vfs_timestamp(tsp)
432 	struct timespec *tsp;
433 {
434 	struct timeval tv;
435 
436 	switch (timestamp_precision) {
437 	case TSP_SEC:
438 		tsp->tv_sec = time_second;
439 		tsp->tv_nsec = 0;
440 		break;
441 	case TSP_HZ:
442 		getnanotime(tsp);
443 		break;
444 	case TSP_USEC:
445 		microtime(&tv);
446 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
447 		break;
448 	case TSP_NSEC:
449 	default:
450 		nanotime(tsp);
451 		break;
452 	}
453 }
454 
455 /*
456  * Set vnode attributes to VNOVAL
457  */
458 void
459 vattr_null(vap)
460 	register struct vattr *vap;
461 {
462 
463 	vap->va_type = VNON;
464 	vap->va_size = VNOVAL;
465 	vap->va_bytes = VNOVAL;
466 	vap->va_mode = VNOVAL;
467 	vap->va_nlink = VNOVAL;
468 	vap->va_uid = VNOVAL;
469 	vap->va_gid = VNOVAL;
470 	vap->va_fsid = VNOVAL;
471 	vap->va_fileid = VNOVAL;
472 	vap->va_blocksize = VNOVAL;
473 	vap->va_rdev = VNOVAL;
474 	vap->va_atime.tv_sec = VNOVAL;
475 	vap->va_atime.tv_nsec = VNOVAL;
476 	vap->va_mtime.tv_sec = VNOVAL;
477 	vap->va_mtime.tv_nsec = VNOVAL;
478 	vap->va_ctime.tv_sec = VNOVAL;
479 	vap->va_ctime.tv_nsec = VNOVAL;
480 	vap->va_birthtime.tv_sec = VNOVAL;
481 	vap->va_birthtime.tv_nsec = VNOVAL;
482 	vap->va_flags = VNOVAL;
483 	vap->va_gen = VNOVAL;
484 	vap->va_vaflags = 0;
485 }
486 
487 /*
488  * This routine is called when we have too many vnodes.  It attempts
489  * to free <count> vnodes and will potentially free vnodes that still
490  * have VM backing store (VM backing store is typically the cause
491  * of a vnode blowout so we want to do this).  Therefore, this operation
492  * is not considered cheap.
493  *
494  * A number of conditions may prevent a vnode from being reclaimed.
495  * the buffer cache may have references on the vnode, a directory
496  * vnode may still have references due to the namei cache representing
497  * underlying files, or the vnode may be in active use.   It is not
498  * desireable to reuse such vnodes.  These conditions may cause the
499  * number of vnodes to reach some minimum value regardless of what
500  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
501  */
502 static int
503 vlrureclaim(struct mount *mp)
504 {
505 	struct vnode *vp;
506 	int done;
507 	int trigger;
508 	int usevnodes;
509 	int count;
510 
511 	/*
512 	 * Calculate the trigger point, don't allow user
513 	 * screwups to blow us up.   This prevents us from
514 	 * recycling vnodes with lots of resident pages.  We
515 	 * aren't trying to free memory, we are trying to
516 	 * free vnodes.
517 	 */
518 	usevnodes = desiredvnodes;
519 	if (usevnodes <= 0)
520 		usevnodes = 1;
521 	trigger = cnt.v_page_count * 2 / usevnodes;
522 
523 	done = 0;
524 	MNT_ILOCK(mp);
525 	count = mp->mnt_nvnodelistsize / 10 + 1;
526 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
527 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
528 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
529 
530 		if (vp->v_type != VNON &&
531 		    vp->v_type != VBAD &&
532 		    VI_TRYLOCK(vp)) {
533 			if (VMIGHTFREE(vp) &&           /* critical path opt */
534 			    (vp->v_object == NULL ||
535 			    vp->v_object->resident_page_count < trigger)) {
536 				MNT_IUNLOCK(mp);
537 				vgonel(vp, curthread);
538 				done++;
539 				MNT_ILOCK(mp);
540 			} else
541 				VI_UNLOCK(vp);
542 		}
543 		--count;
544 	}
545 	MNT_IUNLOCK(mp);
546 	return done;
547 }
548 
549 /*
550  * Attempt to recycle vnodes in a context that is always safe to block.
551  * Calling vlrurecycle() from the bowels of filesystem code has some
552  * interesting deadlock problems.
553  */
554 static struct proc *vnlruproc;
555 static int vnlruproc_sig;
556 
557 static void
558 vnlru_proc(void)
559 {
560 	struct mount *mp, *nmp;
561 	int done;
562 	struct proc *p = vnlruproc;
563 	struct thread *td = FIRST_THREAD_IN_PROC(p);
564 
565 	mtx_lock(&Giant);
566 
567 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
568 	    SHUTDOWN_PRI_FIRST);
569 
570 	for (;;) {
571 		kthread_suspend_check(p);
572 		mtx_lock(&vnode_free_list_mtx);
573 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
574 			mtx_unlock(&vnode_free_list_mtx);
575 			vnlruproc_sig = 0;
576 			wakeup(&vnlruproc_sig);
577 			tsleep(vnlruproc, PVFS, "vlruwt", hz);
578 			continue;
579 		}
580 		mtx_unlock(&vnode_free_list_mtx);
581 		done = 0;
582 		mtx_lock(&mountlist_mtx);
583 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
584 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
585 				nmp = TAILQ_NEXT(mp, mnt_list);
586 				continue;
587 			}
588 			done += vlrureclaim(mp);
589 			mtx_lock(&mountlist_mtx);
590 			nmp = TAILQ_NEXT(mp, mnt_list);
591 			vfs_unbusy(mp, td);
592 		}
593 		mtx_unlock(&mountlist_mtx);
594 		if (done == 0) {
595 #if 0
596 			/* These messages are temporary debugging aids */
597 			if (vnlru_nowhere < 5)
598 				printf("vnlru process getting nowhere..\n");
599 			else if (vnlru_nowhere == 5)
600 				printf("vnlru process messages stopped.\n");
601 #endif
602 			vnlru_nowhere++;
603 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
604 		}
605 	}
606 }
607 
608 static struct kproc_desc vnlru_kp = {
609 	"vnlru",
610 	vnlru_proc,
611 	&vnlruproc
612 };
613 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
614 
615 
616 /*
617  * Routines having to do with the management of the vnode table.
618  */
619 
620 /*
621  * Check to see if a free vnode can be recycled. If it can,
622  * recycle it and return it with the vnode interlock held.
623  */
624 static int
625 vtryrecycle(struct vnode *vp)
626 {
627 	struct thread *td = curthread;
628 	vm_object_t object;
629 	struct mount *vnmp;
630 	int error;
631 
632 	/* Don't recycle if we can't get the interlock */
633 	if (!VI_TRYLOCK(vp))
634 		return (EWOULDBLOCK);
635 	/*
636 	 * This vnode may found and locked via some other list, if so we
637 	 * can't recycle it yet.
638 	 */
639 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
640 		return (EWOULDBLOCK);
641 	/*
642 	 * Don't recycle if its filesystem is being suspended.
643 	 */
644 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
645 		VOP_UNLOCK(vp, 0, td);
646 		return (EBUSY);
647 	}
648 
649 	/*
650 	 * Don't recycle if we still have cached pages.
651 	 */
652 	if (VOP_GETVOBJECT(vp, &object) == 0) {
653 		VM_OBJECT_LOCK(object);
654 		if (object->resident_page_count ||
655 		    object->ref_count) {
656 			VM_OBJECT_UNLOCK(object);
657 			error = EBUSY;
658 			goto done;
659 		}
660 		VM_OBJECT_UNLOCK(object);
661 	}
662 	if (LIST_FIRST(&vp->v_cache_src)) {
663 		/*
664 		 * note: nameileafonly sysctl is temporary,
665 		 * for debugging only, and will eventually be
666 		 * removed.
667 		 */
668 		if (nameileafonly > 0) {
669 			/*
670 			 * Do not reuse namei-cached directory
671 			 * vnodes that have cached
672 			 * subdirectories.
673 			 */
674 			if (cache_leaf_test(vp) < 0) {
675 				error = EISDIR;
676 				goto done;
677 			}
678 		} else if (nameileafonly < 0 ||
679 			    vmiodirenable == 0) {
680 			/*
681 			 * Do not reuse namei-cached directory
682 			 * vnodes if nameileafonly is -1 or
683 			 * if VMIO backing for directories is
684 			 * turned off (otherwise we reuse them
685 			 * too quickly).
686 			 */
687 			error = EBUSY;
688 			goto done;
689 		}
690 	}
691 	/*
692 	 * If we got this far, we need to acquire the interlock and see if
693 	 * anyone picked up this vnode from another list.  If not, we will
694 	 * mark it with XLOCK via vgonel() so that anyone who does find it
695 	 * will skip over it.
696 	 */
697 	VI_LOCK(vp);
698 	if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) {
699 		VI_UNLOCK(vp);
700 		error = EBUSY;
701 		goto done;
702 	}
703 	mtx_lock(&vnode_free_list_mtx);
704 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
705 	vp->v_iflag &= ~VI_FREE;
706 	mtx_unlock(&vnode_free_list_mtx);
707 	vp->v_iflag |= VI_DOOMED;
708 	if (vp->v_type != VBAD) {
709 		VOP_UNLOCK(vp, 0, td);
710 		vgonel(vp, td);
711 		VI_LOCK(vp);
712 	} else
713 		VOP_UNLOCK(vp, 0, td);
714 	vn_finished_write(vnmp);
715 	return (0);
716 done:
717 	VOP_UNLOCK(vp, 0, td);
718 	vn_finished_write(vnmp);
719 	return (error);
720 }
721 
722 /*
723  * Return the next vnode from the free list.
724  */
725 int
726 getnewvnode(tag, mp, vops, vpp)
727 	const char *tag;
728 	struct mount *mp;
729 	vop_t **vops;
730 	struct vnode **vpp;
731 {
732 	struct vnode *vp = NULL;
733 	struct vpollinfo *pollinfo = NULL;
734 
735 	mtx_lock(&vnode_free_list_mtx);
736 
737 	/*
738 	 * Try to reuse vnodes if we hit the max.  This situation only
739 	 * occurs in certain large-memory (2G+) situations.  We cannot
740 	 * attempt to directly reclaim vnodes due to nasty recursion
741 	 * problems.
742 	 */
743 	while (numvnodes - freevnodes > desiredvnodes) {
744 		if (vnlruproc_sig == 0) {
745 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
746 			wakeup(vnlruproc);
747 		}
748 		mtx_unlock(&vnode_free_list_mtx);
749 		tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz);
750 		mtx_lock(&vnode_free_list_mtx);
751 	}
752 
753 	/*
754 	 * Attempt to reuse a vnode already on the free list, allocating
755 	 * a new vnode if we can't find one or if we have not reached a
756 	 * good minimum for good LRU performance.
757 	 */
758 
759 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
760 		int error;
761 		int count;
762 
763 		for (count = 0; count < freevnodes; count++) {
764 			vp = TAILQ_FIRST(&vnode_free_list);
765 
766 			KASSERT(vp->v_usecount == 0 &&
767 			    (vp->v_iflag & VI_DOINGINACT) == 0,
768 			    ("getnewvnode: free vnode isn't"));
769 
770 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
771 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
772 			mtx_unlock(&vnode_free_list_mtx);
773 			error = vtryrecycle(vp);
774 			mtx_lock(&vnode_free_list_mtx);
775 			if (error == 0)
776 				break;
777 			vp = NULL;
778 		}
779 	}
780 	if (vp) {
781 		freevnodes--;
782 		mtx_unlock(&vnode_free_list_mtx);
783 
784 #ifdef INVARIANTS
785 		{
786 			if (vp->v_data)
787 				panic("cleaned vnode isn't");
788 			if (vp->v_numoutput)
789 				panic("Clean vnode has pending I/O's");
790 			if (vp->v_writecount != 0)
791 				panic("Non-zero write count");
792 		}
793 #endif
794 		if ((pollinfo = vp->v_pollinfo) != NULL) {
795 			/*
796 			 * To avoid lock order reversals, the call to
797 			 * uma_zfree() must be delayed until the vnode
798 			 * interlock is released.
799 			 */
800 			vp->v_pollinfo = NULL;
801 		}
802 #ifdef MAC
803 		mac_destroy_vnode(vp);
804 #endif
805 		vp->v_iflag = 0;
806 		vp->v_vflag = 0;
807 		vp->v_lastw = 0;
808 		vp->v_lasta = 0;
809 		vp->v_cstart = 0;
810 		vp->v_clen = 0;
811 		vp->v_socket = 0;
812 		lockdestroy(vp->v_vnlock);
813 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
814 		KASSERT(vp->v_cleanbufcnt == 0, ("cleanbufcnt not 0"));
815 		KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL"));
816 		KASSERT(vp->v_dirtybufcnt == 0, ("dirtybufcnt not 0"));
817 		KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL"));
818 	} else {
819 		numvnodes++;
820 		mtx_unlock(&vnode_free_list_mtx);
821 
822 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
823 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
824 		VI_LOCK(vp);
825 		vp->v_dd = vp;
826 		vp->v_vnlock = &vp->v_lock;
827 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
828 		cache_purge(vp);		/* Sets up v_id. */
829 		LIST_INIT(&vp->v_cache_src);
830 		TAILQ_INIT(&vp->v_cache_dst);
831 	}
832 
833 	TAILQ_INIT(&vp->v_cleanblkhd);
834 	TAILQ_INIT(&vp->v_dirtyblkhd);
835 	vp->v_type = VNON;
836 	vp->v_tag = tag;
837 	vp->v_op = vops;
838 	*vpp = vp;
839 	vp->v_usecount = 1;
840 	vp->v_data = 0;
841 	vp->v_cachedid = -1;
842 	VI_UNLOCK(vp);
843 	if (pollinfo != NULL) {
844 		mtx_destroy(&pollinfo->vpi_lock);
845 		uma_zfree(vnodepoll_zone, pollinfo);
846 	}
847 #ifdef MAC
848 	mac_init_vnode(vp);
849 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
850 		mac_associate_vnode_singlelabel(mp, vp);
851 #endif
852 	delmntque(vp);
853 	if (mp != NULL) {
854 		insmntque(vp, mp);
855 		vp->v_bsize = mp->mnt_stat.f_iosize;
856 	}
857 
858 	return (0);
859 }
860 
861 /*
862  * Delete from old mount point vnode list, if on one.
863  */
864 static void
865 delmntque(struct vnode *vp)
866 {
867 	struct mount *mp;
868 
869 	if (vp->v_mount == NULL)
870 		return;
871 	mp = vp->v_mount;
872 	MNT_ILOCK(mp);
873 	vp->v_mount = NULL;
874 	KASSERT(mp->mnt_nvnodelistsize > 0,
875 		("bad mount point vnode list size"));
876 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
877 	mp->mnt_nvnodelistsize--;
878 	MNT_IUNLOCK(mp);
879 }
880 
881 /*
882  * Insert into list of vnodes for the new mount point, if available.
883  */
884 static void
885 insmntque(struct vnode *vp, struct mount *mp)
886 {
887 
888 	vp->v_mount = mp;
889 	KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)"));
890 	MNT_ILOCK(vp->v_mount);
891 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
892 	mp->mnt_nvnodelistsize++;
893 	MNT_IUNLOCK(vp->v_mount);
894 }
895 
896 /*
897  * Update outstanding I/O count and do wakeup if requested.
898  */
899 void
900 vwakeup(bp)
901 	register struct buf *bp;
902 {
903 	register struct vnode *vp;
904 
905 	bp->b_flags &= ~B_WRITEINPROG;
906 	if ((vp = bp->b_vp)) {
907 		VI_LOCK(vp);
908 		vp->v_numoutput--;
909 		if (vp->v_numoutput < 0)
910 			panic("vwakeup: neg numoutput");
911 		if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) {
912 			vp->v_iflag &= ~VI_BWAIT;
913 			wakeup(&vp->v_numoutput);
914 		}
915 		VI_UNLOCK(vp);
916 	}
917 }
918 
919 /*
920  * Flush out and invalidate all buffers associated with a vnode.
921  * Called with the underlying object locked.
922  */
923 int
924 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo)
925 	struct vnode *vp;
926 	int flags;
927 	struct ucred *cred;
928 	struct thread *td;
929 	int slpflag, slptimeo;
930 {
931 	struct buf *blist;
932 	int error;
933 	vm_object_t object;
934 
935 	GIANT_REQUIRED;
936 
937 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
938 
939 	VI_LOCK(vp);
940 	if (flags & V_SAVE) {
941 		while (vp->v_numoutput) {
942 			vp->v_iflag |= VI_BWAIT;
943 			error = msleep(&vp->v_numoutput, VI_MTX(vp),
944 			    slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo);
945 			if (error) {
946 				VI_UNLOCK(vp);
947 				return (error);
948 			}
949 		}
950 		if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
951 			VI_UNLOCK(vp);
952 			if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0)
953 				return (error);
954 			/*
955 			 * XXX We could save a lock/unlock if this was only
956 			 * enabled under INVARIANTS
957 			 */
958 			VI_LOCK(vp);
959 			if (vp->v_numoutput > 0 ||
960 			    !TAILQ_EMPTY(&vp->v_dirtyblkhd))
961 				panic("vinvalbuf: dirty bufs");
962 		}
963 	}
964 	/*
965 	 * If you alter this loop please notice that interlock is dropped and
966 	 * reacquired in flushbuflist.  Special care is needed to ensure that
967 	 * no race conditions occur from this.
968 	 */
969 	for (error = 0;;) {
970 		if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 &&
971 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
972 			if (error)
973 				break;
974 			continue;
975 		}
976 		if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 &&
977 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
978 			if (error)
979 				break;
980 			continue;
981 		}
982 		break;
983 	}
984 	if (error) {
985 		VI_UNLOCK(vp);
986 		return (error);
987 	}
988 
989 	/*
990 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
991 	 * have write I/O in-progress but if there is a VM object then the
992 	 * VM object can also have read-I/O in-progress.
993 	 */
994 	do {
995 		while (vp->v_numoutput > 0) {
996 			vp->v_iflag |= VI_BWAIT;
997 			msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0);
998 		}
999 		VI_UNLOCK(vp);
1000 		if (VOP_GETVOBJECT(vp, &object) == 0) {
1001 			VM_OBJECT_LOCK(object);
1002 			vm_object_pip_wait(object, "vnvlbx");
1003 			VM_OBJECT_UNLOCK(object);
1004 		}
1005 		VI_LOCK(vp);
1006 	} while (vp->v_numoutput > 0);
1007 	VI_UNLOCK(vp);
1008 
1009 	/*
1010 	 * Destroy the copy in the VM cache, too.
1011 	 */
1012 	if (VOP_GETVOBJECT(vp, &object) == 0) {
1013 		VM_OBJECT_LOCK(object);
1014 		vm_object_page_remove(object, 0, 0,
1015 			(flags & V_SAVE) ? TRUE : FALSE);
1016 		VM_OBJECT_UNLOCK(object);
1017 	}
1018 
1019 #ifdef INVARIANTS
1020 	VI_LOCK(vp);
1021 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1022 	    (!TAILQ_EMPTY(&vp->v_dirtyblkhd) ||
1023 	     !TAILQ_EMPTY(&vp->v_cleanblkhd)))
1024 		panic("vinvalbuf: flush failed");
1025 	VI_UNLOCK(vp);
1026 #endif
1027 	return (0);
1028 }
1029 
1030 /*
1031  * Flush out buffers on the specified list.
1032  *
1033  */
1034 static int
1035 flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp)
1036 	struct buf *blist;
1037 	int flags;
1038 	struct vnode *vp;
1039 	int slpflag, slptimeo;
1040 	int *errorp;
1041 {
1042 	struct buf *bp, *nbp;
1043 	int found, error;
1044 
1045 	ASSERT_VI_LOCKED(vp, "flushbuflist");
1046 
1047 	for (found = 0, bp = blist; bp; bp = nbp) {
1048 		nbp = TAILQ_NEXT(bp, b_vnbufs);
1049 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1050 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1051 			continue;
1052 		}
1053 		found += 1;
1054 		error = BUF_TIMELOCK(bp,
1055 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, VI_MTX(vp),
1056 		    "flushbuf", slpflag, slptimeo);
1057 		if (error) {
1058 			if (error != ENOLCK)
1059 				*errorp = error;
1060 			goto done;
1061 		}
1062 		/*
1063 		 * XXX Since there are no node locks for NFS, I
1064 		 * believe there is a slight chance that a delayed
1065 		 * write will occur while sleeping just above, so
1066 		 * check for it.  Note that vfs_bio_awrite expects
1067 		 * buffers to reside on a queue, while bwrite and
1068 		 * brelse do not.
1069 		 */
1070 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1071 			(flags & V_SAVE)) {
1072 
1073 			if (bp->b_vp == vp) {
1074 				if (bp->b_flags & B_CLUSTEROK) {
1075 					vfs_bio_awrite(bp);
1076 				} else {
1077 					bremfree(bp);
1078 					bp->b_flags |= B_ASYNC;
1079 					bwrite(bp);
1080 				}
1081 			} else {
1082 				bremfree(bp);
1083 				(void) bwrite(bp);
1084 			}
1085 			goto done;
1086 		}
1087 		bremfree(bp);
1088 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1089 		bp->b_flags &= ~B_ASYNC;
1090 		brelse(bp);
1091 		VI_LOCK(vp);
1092 	}
1093 	return (found);
1094 done:
1095 	VI_LOCK(vp);
1096 	return (found);
1097 }
1098 
1099 /*
1100  * Truncate a file's buffer and pages to a specified length.  This
1101  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1102  * sync activity.
1103  */
1104 int
1105 vtruncbuf(vp, cred, td, length, blksize)
1106 	register struct vnode *vp;
1107 	struct ucred *cred;
1108 	struct thread *td;
1109 	off_t length;
1110 	int blksize;
1111 {
1112 	register struct buf *bp;
1113 	struct buf *nbp;
1114 	int anyfreed;
1115 	int trunclbn;
1116 
1117 	/*
1118 	 * Round up to the *next* lbn.
1119 	 */
1120 	trunclbn = (length + blksize - 1) / blksize;
1121 
1122 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1123 restart:
1124 	VI_LOCK(vp);
1125 	anyfreed = 1;
1126 	for (;anyfreed;) {
1127 		anyfreed = 0;
1128 		for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) {
1129 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1130 			if (bp->b_lblkno >= trunclbn) {
1131 				if (BUF_LOCK(bp,
1132 				    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1133 				    VI_MTX(vp)) == ENOLCK)
1134 					goto restart;
1135 
1136 				bremfree(bp);
1137 				bp->b_flags |= (B_INVAL | B_RELBUF);
1138 				bp->b_flags &= ~B_ASYNC;
1139 				brelse(bp);
1140 				anyfreed = 1;
1141 
1142 				if (nbp &&
1143 				    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1144 				    (nbp->b_vp != vp) ||
1145 				    (nbp->b_flags & B_DELWRI))) {
1146 					goto restart;
1147 				}
1148 				VI_LOCK(vp);
1149 			}
1150 		}
1151 
1152 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1153 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1154 			if (bp->b_lblkno >= trunclbn) {
1155 				if (BUF_LOCK(bp,
1156 				    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1157 				    VI_MTX(vp)) == ENOLCK)
1158 					goto restart;
1159 				bremfree(bp);
1160 				bp->b_flags |= (B_INVAL | B_RELBUF);
1161 				bp->b_flags &= ~B_ASYNC;
1162 				brelse(bp);
1163 				anyfreed = 1;
1164 				if (nbp &&
1165 				    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1166 				    (nbp->b_vp != vp) ||
1167 				    (nbp->b_flags & B_DELWRI) == 0)) {
1168 					goto restart;
1169 				}
1170 				VI_LOCK(vp);
1171 			}
1172 		}
1173 	}
1174 
1175 	if (length > 0) {
1176 restartsync:
1177 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1178 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1179 			if (bp->b_lblkno > 0)
1180 				continue;
1181 			/*
1182 			 * Since we hold the vnode lock this should only
1183 			 * fail if we're racing with the buf daemon.
1184 			 */
1185 			if (BUF_LOCK(bp,
1186 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1187 			    VI_MTX(vp)) == ENOLCK) {
1188 				goto restart;
1189 			}
1190 			KASSERT((bp->b_flags & B_DELWRI),
1191 			    ("buf(%p) on dirty queue without DELWRI", bp));
1192 
1193 			bremfree(bp);
1194 			bawrite(bp);
1195 			VI_LOCK(vp);
1196 			goto restartsync;
1197 		}
1198 	}
1199 
1200 	while (vp->v_numoutput > 0) {
1201 		vp->v_iflag |= VI_BWAIT;
1202 		msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0);
1203 	}
1204 	VI_UNLOCK(vp);
1205 	vnode_pager_setsize(vp, length);
1206 
1207 	return (0);
1208 }
1209 
1210 /*
1211  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1212  * 		 a vnode.
1213  *
1214  *	NOTE: We have to deal with the special case of a background bitmap
1215  *	buffer, a situation where two buffers will have the same logical
1216  *	block offset.  We want (1) only the foreground buffer to be accessed
1217  *	in a lookup and (2) must differentiate between the foreground and
1218  *	background buffer in the splay tree algorithm because the splay
1219  *	tree cannot normally handle multiple entities with the same 'index'.
1220  *	We accomplish this by adding differentiating flags to the splay tree's
1221  *	numerical domain.
1222  */
1223 static
1224 struct buf *
1225 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1226 {
1227 	struct buf dummy;
1228 	struct buf *lefttreemax, *righttreemin, *y;
1229 
1230 	if (root == NULL)
1231 		return (NULL);
1232 	lefttreemax = righttreemin = &dummy;
1233 	for (;;) {
1234 		if (lblkno < root->b_lblkno ||
1235 		    (lblkno == root->b_lblkno &&
1236 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1237 			if ((y = root->b_left) == NULL)
1238 				break;
1239 			if (lblkno < y->b_lblkno) {
1240 				/* Rotate right. */
1241 				root->b_left = y->b_right;
1242 				y->b_right = root;
1243 				root = y;
1244 				if ((y = root->b_left) == NULL)
1245 					break;
1246 			}
1247 			/* Link into the new root's right tree. */
1248 			righttreemin->b_left = root;
1249 			righttreemin = root;
1250 		} else if (lblkno > root->b_lblkno ||
1251 		    (lblkno == root->b_lblkno &&
1252 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1253 			if ((y = root->b_right) == NULL)
1254 				break;
1255 			if (lblkno > y->b_lblkno) {
1256 				/* Rotate left. */
1257 				root->b_right = y->b_left;
1258 				y->b_left = root;
1259 				root = y;
1260 				if ((y = root->b_right) == NULL)
1261 					break;
1262 			}
1263 			/* Link into the new root's left tree. */
1264 			lefttreemax->b_right = root;
1265 			lefttreemax = root;
1266 		} else {
1267 			break;
1268 		}
1269 		root = y;
1270 	}
1271 	/* Assemble the new root. */
1272 	lefttreemax->b_right = root->b_left;
1273 	righttreemin->b_left = root->b_right;
1274 	root->b_left = dummy.b_right;
1275 	root->b_right = dummy.b_left;
1276 	return (root);
1277 }
1278 
1279 static
1280 void
1281 buf_vlist_remove(struct buf *bp)
1282 {
1283 	struct vnode *vp = bp->b_vp;
1284 	struct buf *root;
1285 
1286 	ASSERT_VI_LOCKED(vp, "buf_vlist_remove");
1287 	if (bp->b_xflags & BX_VNDIRTY) {
1288 		if (bp != vp->v_dirtyblkroot) {
1289 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1290 			    vp->v_dirtyblkroot);
1291 			KASSERT(root == bp,
1292 			    ("splay lookup failed during dirty remove"));
1293 		}
1294 		if (bp->b_left == NULL) {
1295 			root = bp->b_right;
1296 		} else {
1297 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1298 			    bp->b_left);
1299 			root->b_right = bp->b_right;
1300 		}
1301 		vp->v_dirtyblkroot = root;
1302 		TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs);
1303 		vp->v_dirtybufcnt--;
1304 	} else {
1305 		/* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */
1306 		if (bp != vp->v_cleanblkroot) {
1307 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1308 			    vp->v_cleanblkroot);
1309 			KASSERT(root == bp,
1310 			    ("splay lookup failed during clean remove"));
1311 		}
1312 		if (bp->b_left == NULL) {
1313 			root = bp->b_right;
1314 		} else {
1315 			root = buf_splay(bp->b_lblkno, bp->b_xflags,
1316 			    bp->b_left);
1317 			root->b_right = bp->b_right;
1318 		}
1319 		vp->v_cleanblkroot = root;
1320 		TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs);
1321 		vp->v_cleanbufcnt--;
1322 	}
1323 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1324 }
1325 
1326 /*
1327  * Add the buffer to the sorted clean or dirty block list using a
1328  * splay tree algorithm.
1329  *
1330  * NOTE: xflags is passed as a constant, optimizing this inline function!
1331  */
1332 static
1333 void
1334 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags)
1335 {
1336 	struct buf *root;
1337 
1338 	ASSERT_VI_LOCKED(vp, "buf_vlist_add");
1339 	bp->b_xflags |= xflags;
1340 	if (xflags & BX_VNDIRTY) {
1341 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1342 		if (root == NULL) {
1343 			bp->b_left = NULL;
1344 			bp->b_right = NULL;
1345 			TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs);
1346 		} else if (bp->b_lblkno < root->b_lblkno ||
1347 		    (bp->b_lblkno == root->b_lblkno &&
1348 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1349 			bp->b_left = root->b_left;
1350 			bp->b_right = root;
1351 			root->b_left = NULL;
1352 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1353 		} else {
1354 			bp->b_right = root->b_right;
1355 			bp->b_left = root;
1356 			root->b_right = NULL;
1357 			TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd,
1358 			    root, bp, b_vnbufs);
1359 		}
1360 		vp->v_dirtybufcnt++;
1361 		vp->v_dirtyblkroot = bp;
1362 	} else {
1363 		/* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */
1364 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1365 		if (root == NULL) {
1366 			bp->b_left = NULL;
1367 			bp->b_right = NULL;
1368 			TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs);
1369 		} else if (bp->b_lblkno < root->b_lblkno ||
1370 		    (bp->b_lblkno == root->b_lblkno &&
1371 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1372 			bp->b_left = root->b_left;
1373 			bp->b_right = root;
1374 			root->b_left = NULL;
1375 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1376 		} else {
1377 			bp->b_right = root->b_right;
1378 			bp->b_left = root;
1379 			root->b_right = NULL;
1380 			TAILQ_INSERT_AFTER(&vp->v_cleanblkhd,
1381 			    root, bp, b_vnbufs);
1382 		}
1383 		vp->v_cleanbufcnt++;
1384 		vp->v_cleanblkroot = bp;
1385 	}
1386 }
1387 
1388 /*
1389  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1390  * shadow buffers used in background bitmap writes.
1391  *
1392  * This code isn't quite efficient as it could be because we are maintaining
1393  * two sorted lists and do not know which list the block resides in.
1394  *
1395  * During a "make buildworld" the desired buffer is found at one of
1396  * the roots more than 60% of the time.  Thus, checking both roots
1397  * before performing either splay eliminates unnecessary splays on the
1398  * first tree splayed.
1399  */
1400 struct buf *
1401 gbincore(struct vnode *vp, daddr_t lblkno)
1402 {
1403 	struct buf *bp;
1404 
1405 	GIANT_REQUIRED;
1406 
1407 	ASSERT_VI_LOCKED(vp, "gbincore");
1408 	if ((bp = vp->v_cleanblkroot) != NULL &&
1409 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1410 		return (bp);
1411 	if ((bp = vp->v_dirtyblkroot) != NULL &&
1412 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1413 		return (bp);
1414 	if ((bp = vp->v_cleanblkroot) != NULL) {
1415 		vp->v_cleanblkroot = bp = buf_splay(lblkno, 0, bp);
1416 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1417 			return (bp);
1418 	}
1419 	if ((bp = vp->v_dirtyblkroot) != NULL) {
1420 		vp->v_dirtyblkroot = bp = buf_splay(lblkno, 0, bp);
1421 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1422 			return (bp);
1423 	}
1424 	return (NULL);
1425 }
1426 
1427 /*
1428  * Associate a buffer with a vnode.
1429  */
1430 void
1431 bgetvp(vp, bp)
1432 	register struct vnode *vp;
1433 	register struct buf *bp;
1434 {
1435 
1436 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1437 
1438 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1439 	    ("bgetvp: bp already attached! %p", bp));
1440 
1441 	ASSERT_VI_LOCKED(vp, "bgetvp");
1442 	vholdl(vp);
1443 	bp->b_vp = vp;
1444 	bp->b_dev = vn_todev(vp);
1445 	/*
1446 	 * Insert onto list for new vnode.
1447 	 */
1448 	buf_vlist_add(bp, vp, BX_VNCLEAN);
1449 }
1450 
1451 /*
1452  * Disassociate a buffer from a vnode.
1453  */
1454 void
1455 brelvp(bp)
1456 	register struct buf *bp;
1457 {
1458 	struct vnode *vp;
1459 
1460 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1461 
1462 	/*
1463 	 * Delete from old vnode list, if on one.
1464 	 */
1465 	vp = bp->b_vp;
1466 	VI_LOCK(vp);
1467 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1468 		buf_vlist_remove(bp);
1469 	if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1470 		vp->v_iflag &= ~VI_ONWORKLST;
1471 		mtx_lock(&sync_mtx);
1472 		LIST_REMOVE(vp, v_synclist);
1473  		syncer_worklist_len--;
1474 		mtx_unlock(&sync_mtx);
1475 	}
1476 	vdropl(vp);
1477 	bp->b_vp = (struct vnode *) 0;
1478 	if (bp->b_object)
1479 		bp->b_object = NULL;
1480 	VI_UNLOCK(vp);
1481 }
1482 
1483 /*
1484  * Add an item to the syncer work queue.
1485  */
1486 static void
1487 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
1488 {
1489 	int slot;
1490 
1491 	ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist");
1492 
1493 	mtx_lock(&sync_mtx);
1494 	if (vp->v_iflag & VI_ONWORKLST)
1495 		LIST_REMOVE(vp, v_synclist);
1496 	else {
1497 		vp->v_iflag |= VI_ONWORKLST;
1498  		syncer_worklist_len++;
1499 	}
1500 
1501 	if (delay > syncer_maxdelay - 2)
1502 		delay = syncer_maxdelay - 2;
1503 	slot = (syncer_delayno + delay) & syncer_mask;
1504 
1505 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
1506 	mtx_unlock(&sync_mtx);
1507 }
1508 
1509 static int
1510 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1511 {
1512 	int error, len;
1513 
1514 	mtx_lock(&sync_mtx);
1515 	len = syncer_worklist_len - sync_vnode_count;
1516 	mtx_unlock(&sync_mtx);
1517 	error = SYSCTL_OUT(req, &len, sizeof(len));
1518 	return (error);
1519 }
1520 
1521 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1522     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1523 
1524 struct  proc *updateproc;
1525 static void sched_sync(void);
1526 static struct kproc_desc up_kp = {
1527 	"syncer",
1528 	sched_sync,
1529 	&updateproc
1530 };
1531 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1532 
1533 /*
1534  * System filesystem synchronizer daemon.
1535  */
1536 static void
1537 sched_sync(void)
1538 {
1539 	struct synclist *next;
1540 	struct synclist *slp;
1541 	struct vnode *vp;
1542 	struct mount *mp;
1543 	long starttime;
1544 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1545 	static int dummychan;
1546 	int last_work_seen;
1547 	int net_worklist_len;
1548 	int syncer_final_iter;
1549 
1550 	mtx_lock(&Giant);
1551 	last_work_seen = 0;
1552 	syncer_final_iter = 0;
1553 	syncer_state = SYNCER_RUNNING;
1554 	starttime = time_second;
1555 
1556 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1557 	    SHUTDOWN_PRI_LAST);
1558 
1559 	for (;;) {
1560 		mtx_lock(&sync_mtx);
1561 		if (syncer_state == SYNCER_FINAL_DELAY &&
1562 		    syncer_final_iter == 0) {
1563 			mtx_unlock(&sync_mtx);
1564 			kthread_suspend_check(td->td_proc);
1565 			mtx_lock(&sync_mtx);
1566 		}
1567 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1568 		if (syncer_state != SYNCER_RUNNING && starttime != time_second)
1569 			printf("%d ", net_worklist_len);
1570 		starttime = time_second;
1571 
1572 		/*
1573 		 * Push files whose dirty time has expired.  Be careful
1574 		 * of interrupt race on slp queue.
1575 		 *
1576 		 * Skip over empty worklist slots when shutting down.
1577 		 */
1578 		do {
1579 			slp = &syncer_workitem_pending[syncer_delayno];
1580 			syncer_delayno += 1;
1581 			if (syncer_delayno == syncer_maxdelay)
1582 				syncer_delayno = 0;
1583 			next = &syncer_workitem_pending[syncer_delayno];
1584 			/*
1585 			 * If the worklist has wrapped since the
1586 			 * it was emptied of all but syncer vnodes,
1587 			 * switch to the FINAL_DELAY state and run
1588 			 * for one more second.
1589 			 */
1590 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1591 			    net_worklist_len == 0 &&
1592 			    last_work_seen == syncer_delayno) {
1593 				syncer_state = SYNCER_FINAL_DELAY;
1594 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1595 			}
1596 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1597 		    syncer_worklist_len > 0);
1598 
1599 		/*
1600 		 * Keep track of the last time there was anything
1601 		 * on the worklist other than syncer vnodes.
1602 		 * Return to the SHUTTING_DOWN state if any
1603 		 * new work appears.
1604 		 */
1605 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1606 			last_work_seen = syncer_delayno;
1607 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1608 			syncer_state = SYNCER_SHUTTING_DOWN;
1609 		while ((vp = LIST_FIRST(slp)) != NULL) {
1610 			if (VOP_ISLOCKED(vp, NULL) != 0 ||
1611 			    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1612 				LIST_REMOVE(vp, v_synclist);
1613 				LIST_INSERT_HEAD(next, vp, v_synclist);
1614 				continue;
1615 			}
1616 			if (VI_TRYLOCK(vp) == 0) {
1617 				LIST_REMOVE(vp, v_synclist);
1618 				LIST_INSERT_HEAD(next, vp, v_synclist);
1619 				vn_finished_write(mp);
1620 				continue;
1621 			}
1622 			/*
1623 			 * We use vhold in case the vnode does not
1624 			 * successfully sync.  vhold prevents the vnode from
1625 			 * going away when we unlock the sync_mtx so that
1626 			 * we can acquire the vnode interlock.
1627 			 */
1628 			vholdl(vp);
1629 			mtx_unlock(&sync_mtx);
1630 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td);
1631 			(void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td);
1632 			VOP_UNLOCK(vp, 0, td);
1633 			vn_finished_write(mp);
1634 			VI_LOCK(vp);
1635 			if ((vp->v_iflag & VI_ONWORKLST) != 0) {
1636 				/*
1637 				 * Put us back on the worklist.  The worklist
1638 				 * routine will remove us from our current
1639 				 * position and then add us back in at a later
1640 				 * position.
1641 				 */
1642 				vn_syncer_add_to_worklist(vp, syncdelay);
1643 			}
1644 			vdropl(vp);
1645 			VI_UNLOCK(vp);
1646 			mtx_lock(&sync_mtx);
1647 		}
1648 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1649 			syncer_final_iter--;
1650 		mtx_unlock(&sync_mtx);
1651 
1652 		/*
1653 		 * Do soft update processing.
1654 		 */
1655 		if (softdep_process_worklist_hook != NULL)
1656 			(*softdep_process_worklist_hook)(NULL);
1657 
1658 		/*
1659 		 * The variable rushjob allows the kernel to speed up the
1660 		 * processing of the filesystem syncer process. A rushjob
1661 		 * value of N tells the filesystem syncer to process the next
1662 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1663 		 * is used by the soft update code to speed up the filesystem
1664 		 * syncer process when the incore state is getting so far
1665 		 * ahead of the disk that the kernel memory pool is being
1666 		 * threatened with exhaustion.
1667 		 */
1668 		mtx_lock(&sync_mtx);
1669 		if (rushjob > 0) {
1670 			rushjob -= 1;
1671 			mtx_unlock(&sync_mtx);
1672 			continue;
1673 		}
1674 		mtx_unlock(&sync_mtx);
1675 		/*
1676 		 * Just sleep for a short period if time between
1677 		 * iterations when shutting down to allow some I/O
1678 		 * to happen.
1679 		 *
1680 		 * If it has taken us less than a second to process the
1681 		 * current work, then wait. Otherwise start right over
1682 		 * again. We can still lose time if any single round
1683 		 * takes more than two seconds, but it does not really
1684 		 * matter as we are just trying to generally pace the
1685 		 * filesystem activity.
1686 		 */
1687 		if (syncer_state != SYNCER_RUNNING)
1688 			tsleep(&dummychan, PPAUSE, "syncfnl",
1689 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1690 		else if (time_second == starttime)
1691 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1692 	}
1693 }
1694 
1695 /*
1696  * Request the syncer daemon to speed up its work.
1697  * We never push it to speed up more than half of its
1698  * normal turn time, otherwise it could take over the cpu.
1699  */
1700 int
1701 speedup_syncer()
1702 {
1703 	struct thread *td;
1704 	int ret = 0;
1705 
1706 	td = FIRST_THREAD_IN_PROC(updateproc);
1707 	sleepq_remove(td, &lbolt);
1708 	mtx_lock(&sync_mtx);
1709 	if (rushjob < syncdelay / 2) {
1710 		rushjob += 1;
1711 		stat_rush_requests += 1;
1712 		ret = 1;
1713 	}
1714 	mtx_unlock(&sync_mtx);
1715 	return (ret);
1716 }
1717 
1718 /*
1719  * Tell the syncer to speed up its work and run though its work
1720  * list several times, then tell it to shut down.
1721  */
1722 static void
1723 syncer_shutdown(void *arg, int howto)
1724 {
1725 	struct thread *td;
1726 
1727 	td = FIRST_THREAD_IN_PROC(updateproc);
1728 	sleepq_remove(td, &lbolt);
1729 	mtx_lock(&sync_mtx);
1730 	syncer_state = SYNCER_SHUTTING_DOWN;
1731 	rushjob = 0;
1732 	mtx_unlock(&sync_mtx);
1733 	kproc_shutdown(arg, howto);
1734 }
1735 
1736 /*
1737  * Associate a p-buffer with a vnode.
1738  *
1739  * Also sets B_PAGING flag to indicate that vnode is not fully associated
1740  * with the buffer.  i.e. the bp has not been linked into the vnode or
1741  * ref-counted.
1742  */
1743 void
1744 pbgetvp(vp, bp)
1745 	register struct vnode *vp;
1746 	register struct buf *bp;
1747 {
1748 
1749 	KASSERT(bp->b_vp == NULL, ("pbgetvp: not free"));
1750 
1751 	bp->b_vp = vp;
1752 	bp->b_object = vp->v_object;
1753 	bp->b_flags |= B_PAGING;
1754 	bp->b_dev = vn_todev(vp);
1755 }
1756 
1757 /*
1758  * Disassociate a p-buffer from a vnode.
1759  */
1760 void
1761 pbrelvp(bp)
1762 	register struct buf *bp;
1763 {
1764 
1765 	KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL"));
1766 
1767 	/* XXX REMOVE ME */
1768 	VI_LOCK(bp->b_vp);
1769 	if (TAILQ_NEXT(bp, b_vnbufs) != NULL) {
1770 		panic(
1771 		    "relpbuf(): b_vp was probably reassignbuf()d %p %x",
1772 		    bp,
1773 		    (int)bp->b_flags
1774 		);
1775 	}
1776 	VI_UNLOCK(bp->b_vp);
1777 	bp->b_vp = (struct vnode *) 0;
1778 	bp->b_object = NULL;
1779 	bp->b_flags &= ~B_PAGING;
1780 }
1781 
1782 /*
1783  * Reassign a buffer from one vnode to another.
1784  * Used to assign file specific control information
1785  * (indirect blocks) to the vnode to which they belong.
1786  */
1787 void
1788 reassignbuf(bp, newvp)
1789 	register struct buf *bp;
1790 	register struct vnode *newvp;
1791 {
1792 	struct vnode *vp;
1793 	int delay;
1794 
1795 	if (newvp == NULL) {
1796 		printf("reassignbuf: NULL");
1797 		return;
1798 	}
1799 	vp = bp->b_vp;
1800 	++reassignbufcalls;
1801 
1802 	/*
1803 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1804 	 * is not fully linked in.
1805 	 */
1806 	if (bp->b_flags & B_PAGING)
1807 		panic("cannot reassign paging buffer");
1808 
1809 	/*
1810 	 * Delete from old vnode list, if on one.
1811 	 */
1812 	VI_LOCK(vp);
1813 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) {
1814 		buf_vlist_remove(bp);
1815 		if (vp != newvp) {
1816 			vdropl(bp->b_vp);
1817 			bp->b_vp = NULL;	/* for clarification */
1818 		}
1819 	}
1820 	if (vp != newvp) {
1821 		VI_UNLOCK(vp);
1822 		VI_LOCK(newvp);
1823 	}
1824 	/*
1825 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1826 	 * of clean buffers.
1827 	 */
1828 	if (bp->b_flags & B_DELWRI) {
1829 		if ((newvp->v_iflag & VI_ONWORKLST) == 0) {
1830 			switch (newvp->v_type) {
1831 			case VDIR:
1832 				delay = dirdelay;
1833 				break;
1834 			case VCHR:
1835 				delay = metadelay;
1836 				break;
1837 			default:
1838 				delay = filedelay;
1839 			}
1840 			vn_syncer_add_to_worklist(newvp, delay);
1841 		}
1842 		buf_vlist_add(bp, newvp, BX_VNDIRTY);
1843 	} else {
1844 		buf_vlist_add(bp, newvp, BX_VNCLEAN);
1845 
1846 		if ((newvp->v_iflag & VI_ONWORKLST) &&
1847 		    TAILQ_EMPTY(&newvp->v_dirtyblkhd)) {
1848 			mtx_lock(&sync_mtx);
1849 			LIST_REMOVE(newvp, v_synclist);
1850  			syncer_worklist_len--;
1851 			mtx_unlock(&sync_mtx);
1852 			newvp->v_iflag &= ~VI_ONWORKLST;
1853 		}
1854 	}
1855 	if (bp->b_vp != newvp) {
1856 		bp->b_vp = newvp;
1857 		vholdl(bp->b_vp);
1858 	}
1859 	VI_UNLOCK(newvp);
1860 }
1861 
1862 /*
1863  * Create a vnode for a device.
1864  * Used for mounting the root filesystem.
1865  */
1866 int
1867 bdevvp(dev, vpp)
1868 	struct cdev *dev;
1869 	struct vnode **vpp;
1870 {
1871 	register struct vnode *vp;
1872 	struct vnode *nvp;
1873 	int error;
1874 
1875 	if (dev == NULL) {
1876 		*vpp = NULLVP;
1877 		return (ENXIO);
1878 	}
1879 	if (vfinddev(dev, vpp))
1880 		return (0);
1881 
1882 	error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp);
1883 	if (error) {
1884 		*vpp = NULLVP;
1885 		return (error);
1886 	}
1887 	vp = nvp;
1888 	vp->v_type = VCHR;
1889 	vp->v_bsize = DEV_BSIZE;
1890 	addalias(vp, dev);
1891 	*vpp = vp;
1892 	return (0);
1893 }
1894 
1895 static void
1896 v_incr_usecount(struct vnode *vp, int delta)
1897 {
1898 
1899 	vp->v_usecount += delta;
1900 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1901 		mtx_lock(&spechash_mtx);
1902 		vp->v_rdev->si_usecount += delta;
1903 		mtx_unlock(&spechash_mtx);
1904 	}
1905 }
1906 
1907 /*
1908  * Add vnode to the alias list hung off the struct cdev *.
1909  *
1910  * The reason for this gunk is that multiple vnodes can reference
1911  * the same physical device, so checking vp->v_usecount to see
1912  * how many users there are is inadequate; the v_usecount for
1913  * the vnodes need to be accumulated.  vcount() does that.
1914  */
1915 struct vnode *
1916 addaliasu(nvp, nvp_rdev)
1917 	struct vnode *nvp;
1918 	dev_t nvp_rdev;
1919 {
1920 	struct vnode *ovp;
1921 	vop_t **ops;
1922 	struct cdev *dev;
1923 
1924 	if (nvp->v_type == VBLK)
1925 		return (nvp);
1926 	if (nvp->v_type != VCHR)
1927 		panic("addaliasu on non-special vnode");
1928 	dev = findcdev(nvp_rdev);
1929 	if (dev == NULL)
1930 		return (nvp);
1931 	/*
1932 	 * Check to see if we have a bdevvp vnode with no associated
1933 	 * filesystem. If so, we want to associate the filesystem of
1934 	 * the new newly instigated vnode with the bdevvp vnode and
1935 	 * discard the newly created vnode rather than leaving the
1936 	 * bdevvp vnode lying around with no associated filesystem.
1937 	 */
1938 	if (vfinddev(dev, &ovp) == 0 || ovp->v_data != NULL) {
1939 		addalias(nvp, dev);
1940 		return (nvp);
1941 	}
1942 	/*
1943 	 * Discard unneeded vnode, but save its node specific data.
1944 	 * Note that if there is a lock, it is carried over in the
1945 	 * node specific data to the replacement vnode.
1946 	 */
1947 	vref(ovp);
1948 	ovp->v_data = nvp->v_data;
1949 	ovp->v_tag = nvp->v_tag;
1950 	nvp->v_data = NULL;
1951 	lockdestroy(ovp->v_vnlock);
1952 	lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg,
1953 	    nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK);
1954 	ops = ovp->v_op;
1955 	ovp->v_op = nvp->v_op;
1956 	if (VOP_ISLOCKED(nvp, curthread)) {
1957 		VOP_UNLOCK(nvp, 0, curthread);
1958 		vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread);
1959 	}
1960 	nvp->v_op = ops;
1961 	delmntque(ovp);
1962 	insmntque(ovp, nvp->v_mount);
1963 	vrele(nvp);
1964 	vgone(nvp);
1965 	return (ovp);
1966 }
1967 
1968 /* This is a local helper function that do the same as addaliasu, but for a
1969  * struct cdev *instead of an dev_t. */
1970 static void
1971 addalias(nvp, dev)
1972 	struct vnode *nvp;
1973 	struct cdev *dev;
1974 {
1975 
1976 	KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode"));
1977 	dev_ref(dev);
1978 	nvp->v_rdev = dev;
1979 	VI_LOCK(nvp);
1980 	mtx_lock(&spechash_mtx);
1981 	SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext);
1982 	dev->si_usecount += nvp->v_usecount;
1983 	mtx_unlock(&spechash_mtx);
1984 	VI_UNLOCK(nvp);
1985 }
1986 
1987 /*
1988  * Grab a particular vnode from the free list, increment its
1989  * reference count and lock it. The vnode lock bit is set if the
1990  * vnode is being eliminated in vgone. The process is awakened
1991  * when the transition is completed, and an error returned to
1992  * indicate that the vnode is no longer usable (possibly having
1993  * been changed to a new filesystem type).
1994  */
1995 int
1996 vget(vp, flags, td)
1997 	register struct vnode *vp;
1998 	int flags;
1999 	struct thread *td;
2000 {
2001 	int error;
2002 
2003 	/*
2004 	 * If the vnode is in the process of being cleaned out for
2005 	 * another use, we wait for the cleaning to finish and then
2006 	 * return failure. Cleaning is determined by checking that
2007 	 * the VI_XLOCK flag is set.
2008 	 */
2009 	if ((flags & LK_INTERLOCK) == 0)
2010 		VI_LOCK(vp);
2011 	if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) {
2012 		if ((flags & LK_NOWAIT) == 0) {
2013 			vp->v_iflag |= VI_XWANT;
2014 			msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0);
2015 			return (ENOENT);
2016 		}
2017 		VI_UNLOCK(vp);
2018 		return (EBUSY);
2019 	}
2020 
2021 	v_incr_usecount(vp, 1);
2022 
2023 	if (VSHOULDBUSY(vp))
2024 		vbusy(vp);
2025 	if (flags & LK_TYPE_MASK) {
2026 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2027 			/*
2028 			 * must expand vrele here because we do not want
2029 			 * to call VOP_INACTIVE if the reference count
2030 			 * drops back to zero since it was never really
2031 			 * active. We must remove it from the free list
2032 			 * before sleeping so that multiple processes do
2033 			 * not try to recycle it.
2034 			 */
2035 			VI_LOCK(vp);
2036 			v_incr_usecount(vp, -1);
2037 			if (VSHOULDFREE(vp))
2038 				vfree(vp);
2039 			else
2040 				vlruvp(vp);
2041 			VI_UNLOCK(vp);
2042 		}
2043 		return (error);
2044 	}
2045 	VI_UNLOCK(vp);
2046 	return (0);
2047 }
2048 
2049 /*
2050  * Increase the reference count of a vnode.
2051  */
2052 void
2053 vref(struct vnode *vp)
2054 {
2055 
2056 	VI_LOCK(vp);
2057 	v_incr_usecount(vp, 1);
2058 	VI_UNLOCK(vp);
2059 }
2060 
2061 /*
2062  * Return reference count of a vnode.
2063  *
2064  * The results of this call are only guaranteed when some mechanism other
2065  * than the VI lock is used to stop other processes from gaining references
2066  * to the vnode.  This may be the case if the caller holds the only reference.
2067  * This is also useful when stale data is acceptable as race conditions may
2068  * be accounted for by some other means.
2069  */
2070 int
2071 vrefcnt(struct vnode *vp)
2072 {
2073 	int usecnt;
2074 
2075 	VI_LOCK(vp);
2076 	usecnt = vp->v_usecount;
2077 	VI_UNLOCK(vp);
2078 
2079 	return (usecnt);
2080 }
2081 
2082 
2083 /*
2084  * Vnode put/release.
2085  * If count drops to zero, call inactive routine and return to freelist.
2086  */
2087 void
2088 vrele(vp)
2089 	struct vnode *vp;
2090 {
2091 	struct thread *td = curthread;	/* XXX */
2092 
2093 	GIANT_REQUIRED;
2094 
2095 	KASSERT(vp != NULL, ("vrele: null vp"));
2096 
2097 	VI_LOCK(vp);
2098 
2099 	/* Skip this v_writecount check if we're going to panic below. */
2100 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2101 	    ("vrele: missed vn_close"));
2102 
2103 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2104 	    vp->v_usecount == 1)) {
2105 		v_incr_usecount(vp, -1);
2106 		VI_UNLOCK(vp);
2107 
2108 		return;
2109 	}
2110 
2111 	if (vp->v_usecount == 1) {
2112 		v_incr_usecount(vp, -1);
2113 		/*
2114 		 * We must call VOP_INACTIVE with the node locked. Mark
2115 		 * as VI_DOINGINACT to avoid recursion.
2116 		 */
2117 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2118 			VI_LOCK(vp);
2119 			vp->v_iflag |= VI_DOINGINACT;
2120 			VI_UNLOCK(vp);
2121 			VOP_INACTIVE(vp, td);
2122 			VI_LOCK(vp);
2123 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2124 			    ("vrele: lost VI_DOINGINACT"));
2125 			vp->v_iflag &= ~VI_DOINGINACT;
2126 		} else
2127 			VI_LOCK(vp);
2128 		if (VSHOULDFREE(vp))
2129 			vfree(vp);
2130 		else
2131 			vlruvp(vp);
2132 		VI_UNLOCK(vp);
2133 
2134 	} else {
2135 #ifdef DIAGNOSTIC
2136 		vprint("vrele: negative ref count", vp);
2137 #endif
2138 		VI_UNLOCK(vp);
2139 		panic("vrele: negative ref cnt");
2140 	}
2141 }
2142 
2143 /*
2144  * Release an already locked vnode.  This give the same effects as
2145  * unlock+vrele(), but takes less time and avoids releasing and
2146  * re-aquiring the lock (as vrele() aquires the lock internally.)
2147  */
2148 void
2149 vput(vp)
2150 	struct vnode *vp;
2151 {
2152 	struct thread *td = curthread;	/* XXX */
2153 
2154 	GIANT_REQUIRED;
2155 
2156 	KASSERT(vp != NULL, ("vput: null vp"));
2157 	VI_LOCK(vp);
2158 	/* Skip this v_writecount check if we're going to panic below. */
2159 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2160 	    ("vput: missed vn_close"));
2161 
2162 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2163 	    vp->v_usecount == 1)) {
2164 		v_incr_usecount(vp, -1);
2165 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
2166 		return;
2167 	}
2168 
2169 	if (vp->v_usecount == 1) {
2170 		v_incr_usecount(vp, -1);
2171 		/*
2172 		 * We must call VOP_INACTIVE with the node locked, so
2173 		 * we just need to release the vnode mutex. Mark as
2174 		 * as VI_DOINGINACT to avoid recursion.
2175 		 */
2176 		vp->v_iflag |= VI_DOINGINACT;
2177 		VI_UNLOCK(vp);
2178 		VOP_INACTIVE(vp, td);
2179 		VI_LOCK(vp);
2180 		KASSERT(vp->v_iflag & VI_DOINGINACT,
2181 		    ("vput: lost VI_DOINGINACT"));
2182 		vp->v_iflag &= ~VI_DOINGINACT;
2183 		if (VSHOULDFREE(vp))
2184 			vfree(vp);
2185 		else
2186 			vlruvp(vp);
2187 		VI_UNLOCK(vp);
2188 
2189 	} else {
2190 #ifdef DIAGNOSTIC
2191 		vprint("vput: negative ref count", vp);
2192 #endif
2193 		panic("vput: negative ref cnt");
2194 	}
2195 }
2196 
2197 /*
2198  * Somebody doesn't want the vnode recycled.
2199  */
2200 void
2201 vhold(struct vnode *vp)
2202 {
2203 
2204 	VI_LOCK(vp);
2205 	vholdl(vp);
2206 	VI_UNLOCK(vp);
2207 }
2208 
2209 void
2210 vholdl(vp)
2211 	register struct vnode *vp;
2212 {
2213 
2214 	vp->v_holdcnt++;
2215 	if (VSHOULDBUSY(vp))
2216 		vbusy(vp);
2217 }
2218 
2219 /*
2220  * Note that there is one less who cares about this vnode.  vdrop() is the
2221  * opposite of vhold().
2222  */
2223 void
2224 vdrop(struct vnode *vp)
2225 {
2226 
2227 	VI_LOCK(vp);
2228 	vdropl(vp);
2229 	VI_UNLOCK(vp);
2230 }
2231 
2232 void
2233 vdropl(vp)
2234 	register struct vnode *vp;
2235 {
2236 
2237 	if (vp->v_holdcnt <= 0)
2238 		panic("vdrop: holdcnt");
2239 	vp->v_holdcnt--;
2240 	if (VSHOULDFREE(vp))
2241 		vfree(vp);
2242 	else
2243 		vlruvp(vp);
2244 }
2245 
2246 /*
2247  * Remove any vnodes in the vnode table belonging to mount point mp.
2248  *
2249  * If FORCECLOSE is not specified, there should not be any active ones,
2250  * return error if any are found (nb: this is a user error, not a
2251  * system error). If FORCECLOSE is specified, detach any active vnodes
2252  * that are found.
2253  *
2254  * If WRITECLOSE is set, only flush out regular file vnodes open for
2255  * writing.
2256  *
2257  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2258  *
2259  * `rootrefs' specifies the base reference count for the root vnode
2260  * of this filesystem. The root vnode is considered busy if its
2261  * v_usecount exceeds this value. On a successful return, vflush(, td)
2262  * will call vrele() on the root vnode exactly rootrefs times.
2263  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2264  * be zero.
2265  */
2266 #ifdef DIAGNOSTIC
2267 static int busyprt = 0;		/* print out busy vnodes */
2268 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2269 #endif
2270 
2271 int
2272 vflush(mp, rootrefs, flags, td)
2273 	struct mount *mp;
2274 	int rootrefs;
2275 	int flags;
2276 	struct thread *td;
2277 {
2278 	struct vnode *vp, *nvp, *rootvp = NULL;
2279 	struct vattr vattr;
2280 	int busy = 0, error;
2281 
2282 	if (rootrefs > 0) {
2283 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2284 		    ("vflush: bad args"));
2285 		/*
2286 		 * Get the filesystem root vnode. We can vput() it
2287 		 * immediately, since with rootrefs > 0, it won't go away.
2288 		 */
2289 		if ((error = VFS_ROOT(mp, &rootvp, td)) != 0)
2290 			return (error);
2291 		vput(rootvp);
2292 
2293 	}
2294 	MNT_ILOCK(mp);
2295 loop:
2296 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2297 
2298 		VI_LOCK(vp);
2299 		MNT_IUNLOCK(mp);
2300 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2301 		if (error) {
2302 			MNT_ILOCK(mp);
2303 			goto loop;
2304 		}
2305 		/*
2306 		 * Skip over a vnodes marked VV_SYSTEM.
2307 		 */
2308 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2309 			VOP_UNLOCK(vp, 0, td);
2310 			MNT_ILOCK(mp);
2311 			continue;
2312 		}
2313 		/*
2314 		 * If WRITECLOSE is set, flush out unlinked but still open
2315 		 * files (even if open only for reading) and regular file
2316 		 * vnodes open for writing.
2317 		 */
2318 		if (flags & WRITECLOSE) {
2319 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2320 			VI_LOCK(vp);
2321 
2322 			if ((vp->v_type == VNON ||
2323 			    (error == 0 && vattr.va_nlink > 0)) &&
2324 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2325 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2326 				MNT_ILOCK(mp);
2327 				continue;
2328 			}
2329 		} else
2330 			VI_LOCK(vp);
2331 
2332 		VOP_UNLOCK(vp, 0, td);
2333 
2334 		/*
2335 		 * With v_usecount == 0, all we need to do is clear out the
2336 		 * vnode data structures and we are done.
2337 		 */
2338 		if (vp->v_usecount == 0) {
2339 			vgonel(vp, td);
2340 			MNT_ILOCK(mp);
2341 			continue;
2342 		}
2343 
2344 		/*
2345 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2346 		 * or character devices, revert to an anonymous device. For
2347 		 * all other files, just kill them.
2348 		 */
2349 		if (flags & FORCECLOSE) {
2350 			if (vp->v_type != VCHR)
2351 				vgonel(vp, td);
2352 			else
2353 				vgonechrl(vp, td);
2354 			MNT_ILOCK(mp);
2355 			continue;
2356 		}
2357 #ifdef DIAGNOSTIC
2358 		if (busyprt)
2359 			vprint("vflush: busy vnode", vp);
2360 #endif
2361 		VI_UNLOCK(vp);
2362 		MNT_ILOCK(mp);
2363 		busy++;
2364 	}
2365 	MNT_IUNLOCK(mp);
2366 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2367 		/*
2368 		 * If just the root vnode is busy, and if its refcount
2369 		 * is equal to `rootrefs', then go ahead and kill it.
2370 		 */
2371 		VI_LOCK(rootvp);
2372 		KASSERT(busy > 0, ("vflush: not busy"));
2373 		KASSERT(rootvp->v_usecount >= rootrefs,
2374 		    ("vflush: usecount %d < rootrefs %d",
2375 		     rootvp->v_usecount, rootrefs));
2376 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2377 			vgonel(rootvp, td);
2378 			busy = 0;
2379 		} else
2380 			VI_UNLOCK(rootvp);
2381 	}
2382 	if (busy)
2383 		return (EBUSY);
2384 	for (; rootrefs > 0; rootrefs--)
2385 		vrele(rootvp);
2386 	return (0);
2387 }
2388 
2389 /*
2390  * This moves a now (likely recyclable) vnode to the end of the
2391  * mountlist.  XXX However, it is temporarily disabled until we
2392  * can clean up ffs_sync() and friends, which have loop restart
2393  * conditions which this code causes to operate O(N^2).
2394  */
2395 static void
2396 vlruvp(struct vnode *vp)
2397 {
2398 #if 0
2399 	struct mount *mp;
2400 
2401 	if ((mp = vp->v_mount) != NULL) {
2402 		MNT_ILOCK(mp);
2403 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2404 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2405 		MNT_IUNLOCK(mp);
2406 	}
2407 #endif
2408 }
2409 
2410 static void
2411 vx_lock(struct vnode *vp)
2412 {
2413 
2414 	ASSERT_VI_LOCKED(vp, "vx_lock");
2415 
2416 	/*
2417 	 * Prevent the vnode from being recycled or brought into use while we
2418 	 * clean it out.
2419 	 */
2420 	if (vp->v_iflag & VI_XLOCK)
2421 		panic("vclean: deadlock");
2422 	vp->v_iflag |= VI_XLOCK;
2423 	vp->v_vxthread = curthread;
2424 }
2425 
2426 static void
2427 vx_unlock(struct vnode *vp)
2428 {
2429 	ASSERT_VI_LOCKED(vp, "vx_unlock");
2430 	vp->v_iflag &= ~VI_XLOCK;
2431 	vp->v_vxthread = NULL;
2432 	if (vp->v_iflag & VI_XWANT) {
2433 		vp->v_iflag &= ~VI_XWANT;
2434 		wakeup(vp);
2435 	}
2436 }
2437 
2438 /*
2439  * Disassociate the underlying filesystem from a vnode.
2440  */
2441 static void
2442 vclean(vp, flags, td)
2443 	struct vnode *vp;
2444 	int flags;
2445 	struct thread *td;
2446 {
2447 	int active;
2448 
2449 	ASSERT_VI_LOCKED(vp, "vclean");
2450 	/*
2451 	 * Check to see if the vnode is in use. If so we have to reference it
2452 	 * before we clean it out so that its count cannot fall to zero and
2453 	 * generate a race against ourselves to recycle it.
2454 	 */
2455 	if ((active = vp->v_usecount))
2456 		v_incr_usecount(vp, 1);
2457 
2458 	/*
2459 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2460 	 * have the object locked while it cleans it out. The VOP_LOCK
2461 	 * ensures that the VOP_INACTIVE routine is done with its work.
2462 	 * For active vnodes, it ensures that no other activity can
2463 	 * occur while the underlying object is being cleaned out.
2464 	 */
2465 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2466 
2467 	/*
2468 	 * Clean out any buffers associated with the vnode.
2469 	 * If the flush fails, just toss the buffers.
2470 	 */
2471 	if (flags & DOCLOSE) {
2472 		struct buf *bp;
2473 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
2474 		if (bp != NULL)
2475 			(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2476 		if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0)
2477 			vinvalbuf(vp, 0, NOCRED, td, 0, 0);
2478 	}
2479 
2480 	VOP_DESTROYVOBJECT(vp);
2481 
2482 	/*
2483 	 * Any other processes trying to obtain this lock must first
2484 	 * wait for VXLOCK to clear, then call the new lock operation.
2485 	 */
2486 	VOP_UNLOCK(vp, 0, td);
2487 
2488 	/*
2489 	 * If purging an active vnode, it must be closed and
2490 	 * deactivated before being reclaimed. Note that the
2491 	 * VOP_INACTIVE will unlock the vnode.
2492 	 */
2493 	if (active) {
2494 		if (flags & DOCLOSE)
2495 			VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2496 		VI_LOCK(vp);
2497 		if ((vp->v_iflag & VI_DOINGINACT) == 0) {
2498 			vp->v_iflag |= VI_DOINGINACT;
2499 			VI_UNLOCK(vp);
2500 			if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2501 				panic("vclean: cannot relock.");
2502 			VOP_INACTIVE(vp, td);
2503 			VI_LOCK(vp);
2504 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2505 			    ("vclean: lost VI_DOINGINACT"));
2506 			vp->v_iflag &= ~VI_DOINGINACT;
2507 		}
2508 		VI_UNLOCK(vp);
2509 	}
2510 	/*
2511 	 * Reclaim the vnode.
2512 	 */
2513 	if (VOP_RECLAIM(vp, td))
2514 		panic("vclean: cannot reclaim");
2515 
2516 	if (active) {
2517 		/*
2518 		 * Inline copy of vrele() since VOP_INACTIVE
2519 		 * has already been called.
2520 		 */
2521 		VI_LOCK(vp);
2522 		v_incr_usecount(vp, -1);
2523 		if (vp->v_usecount <= 0) {
2524 #ifdef INVARIANTS
2525 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2526 				vprint("vclean: bad ref count", vp);
2527 				panic("vclean: ref cnt");
2528 			}
2529 #endif
2530 			if (VSHOULDFREE(vp))
2531 				vfree(vp);
2532 		}
2533 		VI_UNLOCK(vp);
2534 	}
2535 	/*
2536 	 * Delete from old mount point vnode list.
2537 	 */
2538 	delmntque(vp);
2539 	cache_purge(vp);
2540 	VI_LOCK(vp);
2541 	if (VSHOULDFREE(vp))
2542 		vfree(vp);
2543 
2544 	/*
2545 	 * Done with purge, reset to the standard lock and
2546 	 * notify sleepers of the grim news.
2547 	 */
2548 	vp->v_vnlock = &vp->v_lock;
2549 	vp->v_op = dead_vnodeop_p;
2550 	if (vp->v_pollinfo != NULL)
2551 		vn_pollgone(vp);
2552 	vp->v_tag = "none";
2553 }
2554 
2555 /*
2556  * Eliminate all activity associated with the requested vnode
2557  * and with all vnodes aliased to the requested vnode.
2558  */
2559 int
2560 vop_revoke(ap)
2561 	struct vop_revoke_args /* {
2562 		struct vnode *a_vp;
2563 		int a_flags;
2564 	} */ *ap;
2565 {
2566 	struct vnode *vp, *vq;
2567 	struct cdev *dev;
2568 
2569 	KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke"));
2570 	vp = ap->a_vp;
2571 	KASSERT((vp->v_type == VCHR), ("vop_revoke: not VCHR"));
2572 
2573 	VI_LOCK(vp);
2574 	/*
2575 	 * If a vgone (or vclean) is already in progress,
2576 	 * wait until it is done and return.
2577 	 */
2578 	if (vp->v_iflag & VI_XLOCK) {
2579 		vp->v_iflag |= VI_XWANT;
2580 		msleep(vp, VI_MTX(vp), PINOD | PDROP,
2581 		    "vop_revokeall", 0);
2582 		return (0);
2583 	}
2584 	VI_UNLOCK(vp);
2585 	dev = vp->v_rdev;
2586 	for (;;) {
2587 		mtx_lock(&spechash_mtx);
2588 		vq = SLIST_FIRST(&dev->si_hlist);
2589 		mtx_unlock(&spechash_mtx);
2590 		if (vq == NULL)
2591 			break;
2592 		vgone(vq);
2593 	}
2594 	return (0);
2595 }
2596 
2597 /*
2598  * Recycle an unused vnode to the front of the free list.
2599  * Release the passed interlock if the vnode will be recycled.
2600  */
2601 int
2602 vrecycle(vp, inter_lkp, td)
2603 	struct vnode *vp;
2604 	struct mtx *inter_lkp;
2605 	struct thread *td;
2606 {
2607 
2608 	VI_LOCK(vp);
2609 	if (vp->v_usecount == 0) {
2610 		if (inter_lkp) {
2611 			mtx_unlock(inter_lkp);
2612 		}
2613 		vgonel(vp, td);
2614 		return (1);
2615 	}
2616 	VI_UNLOCK(vp);
2617 	return (0);
2618 }
2619 
2620 /*
2621  * Eliminate all activity associated with a vnode
2622  * in preparation for reuse.
2623  */
2624 void
2625 vgone(vp)
2626 	register struct vnode *vp;
2627 {
2628 	struct thread *td = curthread;	/* XXX */
2629 
2630 	VI_LOCK(vp);
2631 	vgonel(vp, td);
2632 }
2633 
2634 /*
2635  * Disassociate a character device from the its underlying filesystem and
2636  * attach it to spec.  This is for use when the chr device is still active
2637  * and the filesystem is going away.
2638  */
2639 static void
2640 vgonechrl(struct vnode *vp, struct thread *td)
2641 {
2642 	ASSERT_VI_LOCKED(vp, "vgonechrl");
2643 	vx_lock(vp);
2644 	/*
2645 	 * This is a custom version of vclean() which does not tearm down
2646 	 * the bufs or vm objects held by this vnode.  This allows filesystems
2647 	 * to continue using devices which were discovered via another
2648 	 * filesystem that has been unmounted.
2649 	 */
2650 	if (vp->v_usecount != 0) {
2651 		v_incr_usecount(vp, 1);
2652 		/*
2653 		 * Ensure that no other activity can occur while the
2654 		 * underlying object is being cleaned out.
2655 		 */
2656 		VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2657 		/*
2658 		 * Any other processes trying to obtain this lock must first
2659 		 * wait for VXLOCK to clear, then call the new lock operation.
2660 		 */
2661 		VOP_UNLOCK(vp, 0, td);
2662 		vp->v_vnlock = &vp->v_lock;
2663 		vp->v_tag = "orphanchr";
2664 		vp->v_op = spec_vnodeop_p;
2665 		delmntque(vp);
2666 		cache_purge(vp);
2667 		vrele(vp);
2668 		VI_LOCK(vp);
2669 	} else
2670 		vclean(vp, 0, td);
2671 	vp->v_op = spec_vnodeop_p;
2672 	vx_unlock(vp);
2673 	VI_UNLOCK(vp);
2674 }
2675 
2676 /*
2677  * vgone, with the vp interlock held.
2678  */
2679 void
2680 vgonel(vp, td)
2681 	struct vnode *vp;
2682 	struct thread *td;
2683 {
2684 	/*
2685 	 * If a vgone (or vclean) is already in progress,
2686 	 * wait until it is done and return.
2687 	 */
2688 	ASSERT_VI_LOCKED(vp, "vgonel");
2689 	if (vp->v_iflag & VI_XLOCK) {
2690 		vp->v_iflag |= VI_XWANT;
2691 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0);
2692 		return;
2693 	}
2694 	vx_lock(vp);
2695 
2696 	/*
2697 	 * Clean out the filesystem specific data.
2698 	 */
2699 	vclean(vp, DOCLOSE, td);
2700 	VI_UNLOCK(vp);
2701 
2702 	/*
2703 	 * If special device, remove it from special device alias list
2704 	 * if it is on one.
2705 	 */
2706 	VI_LOCK(vp);
2707 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2708 		mtx_lock(&spechash_mtx);
2709 		SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext);
2710 		vp->v_rdev->si_usecount -= vp->v_usecount;
2711 		mtx_unlock(&spechash_mtx);
2712 		dev_rel(vp->v_rdev);
2713 		vp->v_rdev = NULL;
2714 	}
2715 
2716 	/*
2717 	 * If it is on the freelist and not already at the head,
2718 	 * move it to the head of the list. The test of the
2719 	 * VDOOMED flag and the reference count of zero is because
2720 	 * it will be removed from the free list by getnewvnode,
2721 	 * but will not have its reference count incremented until
2722 	 * after calling vgone. If the reference count were
2723 	 * incremented first, vgone would (incorrectly) try to
2724 	 * close the previous instance of the underlying object.
2725 	 */
2726 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2727 		mtx_lock(&vnode_free_list_mtx);
2728 		if (vp->v_iflag & VI_FREE) {
2729 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2730 		} else {
2731 			vp->v_iflag |= VI_FREE;
2732 			freevnodes++;
2733 		}
2734 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2735 		mtx_unlock(&vnode_free_list_mtx);
2736 	}
2737 
2738 	vp->v_type = VBAD;
2739 	vx_unlock(vp);
2740 	VI_UNLOCK(vp);
2741 }
2742 
2743 /*
2744  * Lookup a vnode by device number.
2745  */
2746 int
2747 vfinddev(dev, vpp)
2748 	struct cdev *dev;
2749 	struct vnode **vpp;
2750 {
2751 	struct vnode *vp;
2752 
2753 	mtx_lock(&spechash_mtx);
2754 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2755 		*vpp = vp;
2756 		mtx_unlock(&spechash_mtx);
2757 		return (1);
2758 	}
2759 	mtx_unlock(&spechash_mtx);
2760 	return (0);
2761 }
2762 
2763 /*
2764  * Calculate the total number of references to a special device.
2765  */
2766 int
2767 vcount(vp)
2768 	struct vnode *vp;
2769 {
2770 	int count;
2771 
2772 	mtx_lock(&spechash_mtx);
2773 	count = vp->v_rdev->si_usecount;
2774 	mtx_unlock(&spechash_mtx);
2775 	return (count);
2776 }
2777 
2778 /*
2779  * Same as above, but using the struct cdev *as argument
2780  */
2781 int
2782 count_dev(dev)
2783 	struct cdev *dev;
2784 {
2785 	int count;
2786 
2787 	mtx_lock(&spechash_mtx);
2788 	count = dev->si_usecount;
2789 	mtx_unlock(&spechash_mtx);
2790 	return(count);
2791 }
2792 
2793 /*
2794  * Print out a description of a vnode.
2795  */
2796 static char *typename[] =
2797 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2798 
2799 void
2800 vprint(label, vp)
2801 	char *label;
2802 	struct vnode *vp;
2803 {
2804 	char buf[96];
2805 
2806 	if (label != NULL)
2807 		printf("%s: %p: ", label, (void *)vp);
2808 	else
2809 		printf("%p: ", (void *)vp);
2810 	printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,",
2811 	    vp->v_tag, typename[vp->v_type], vp->v_usecount,
2812 	    vp->v_writecount, vp->v_holdcnt);
2813 	buf[0] = '\0';
2814 	if (vp->v_vflag & VV_ROOT)
2815 		strcat(buf, "|VV_ROOT");
2816 	if (vp->v_vflag & VV_TEXT)
2817 		strcat(buf, "|VV_TEXT");
2818 	if (vp->v_vflag & VV_SYSTEM)
2819 		strcat(buf, "|VV_SYSTEM");
2820 	if (vp->v_iflag & VI_XLOCK)
2821 		strcat(buf, "|VI_XLOCK");
2822 	if (vp->v_iflag & VI_XWANT)
2823 		strcat(buf, "|VI_XWANT");
2824 	if (vp->v_iflag & VI_BWAIT)
2825 		strcat(buf, "|VI_BWAIT");
2826 	if (vp->v_iflag & VI_DOOMED)
2827 		strcat(buf, "|VI_DOOMED");
2828 	if (vp->v_iflag & VI_FREE)
2829 		strcat(buf, "|VI_FREE");
2830 	if (vp->v_vflag & VV_OBJBUF)
2831 		strcat(buf, "|VV_OBJBUF");
2832 	if (buf[0] != '\0')
2833 		printf(" flags (%s),", &buf[1]);
2834 	lockmgr_printinfo(vp->v_vnlock);
2835 	printf("\n");
2836 	if (vp->v_data != NULL)
2837 		VOP_PRINT(vp);
2838 }
2839 
2840 #ifdef DDB
2841 #include <ddb/ddb.h>
2842 /*
2843  * List all of the locked vnodes in the system.
2844  * Called when debugging the kernel.
2845  */
2846 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2847 {
2848 	struct mount *mp, *nmp;
2849 	struct vnode *vp;
2850 
2851 	/*
2852 	 * Note: because this is DDB, we can't obey the locking semantics
2853 	 * for these structures, which means we could catch an inconsistent
2854 	 * state and dereference a nasty pointer.  Not much to be done
2855 	 * about that.
2856 	 */
2857 	printf("Locked vnodes\n");
2858 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2859 		nmp = TAILQ_NEXT(mp, mnt_list);
2860 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2861 			if (VOP_ISLOCKED(vp, NULL))
2862 				vprint(NULL, vp);
2863 		}
2864 		nmp = TAILQ_NEXT(mp, mnt_list);
2865 	}
2866 }
2867 #endif
2868 
2869 /*
2870  * Fill in a struct xvfsconf based on a struct vfsconf.
2871  */
2872 static void
2873 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2874 {
2875 
2876 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2877 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2878 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2879 	xvfsp->vfc_flags = vfsp->vfc_flags;
2880 	/*
2881 	 * These are unused in userland, we keep them
2882 	 * to not break binary compatibility.
2883 	 */
2884 	xvfsp->vfc_vfsops = NULL;
2885 	xvfsp->vfc_next = NULL;
2886 }
2887 
2888 /*
2889  * Top level filesystem related information gathering.
2890  */
2891 static int
2892 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2893 {
2894 	struct vfsconf *vfsp;
2895 	struct xvfsconf *xvfsp;
2896 	int cnt, error, i;
2897 
2898 	cnt = 0;
2899 	for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next)
2900 		cnt++;
2901 	xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK);
2902 	/*
2903 	 * Handle the race that we will have here when struct vfsconf
2904 	 * will be locked down by using both cnt and checking vfc_next
2905 	 * against NULL to determine the end of the loop.  The race will
2906 	 * happen because we will have to unlock before calling malloc().
2907 	 * We are protected by Giant for now.
2908 	 */
2909 	i = 0;
2910 	for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) {
2911 		vfsconf2x(vfsp, xvfsp + i);
2912 		i++;
2913 	}
2914 	error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i);
2915 	free(xvfsp, M_TEMP);
2916 	return (error);
2917 }
2918 
2919 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2920     "S,xvfsconf", "List of all configured filesystems");
2921 
2922 #ifndef BURN_BRIDGES
2923 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2924 
2925 static int
2926 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2927 {
2928 	int *name = (int *)arg1 - 1;	/* XXX */
2929 	u_int namelen = arg2 + 1;	/* XXX */
2930 	struct vfsconf *vfsp;
2931 	struct xvfsconf xvfsp;
2932 
2933 	printf("WARNING: userland calling deprecated sysctl, "
2934 	    "please rebuild world\n");
2935 
2936 #if 1 || defined(COMPAT_PRELITE2)
2937 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2938 	if (namelen == 1)
2939 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2940 #endif
2941 
2942 	switch (name[1]) {
2943 	case VFS_MAXTYPENUM:
2944 		if (namelen != 2)
2945 			return (ENOTDIR);
2946 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2947 	case VFS_CONF:
2948 		if (namelen != 3)
2949 			return (ENOTDIR);	/* overloaded */
2950 		for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next)
2951 			if (vfsp->vfc_typenum == name[2])
2952 				break;
2953 		if (vfsp == NULL)
2954 			return (EOPNOTSUPP);
2955 		vfsconf2x(vfsp, &xvfsp);
2956 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2957 	}
2958 	return (EOPNOTSUPP);
2959 }
2960 
2961 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl,
2962 	"Generic filesystem");
2963 
2964 #if 1 || defined(COMPAT_PRELITE2)
2965 
2966 static int
2967 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2968 {
2969 	int error;
2970 	struct vfsconf *vfsp;
2971 	struct ovfsconf ovfs;
2972 
2973 	for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) {
2974 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2975 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2976 		ovfs.vfc_index = vfsp->vfc_typenum;
2977 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2978 		ovfs.vfc_flags = vfsp->vfc_flags;
2979 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2980 		if (error)
2981 			return error;
2982 	}
2983 	return 0;
2984 }
2985 
2986 #endif /* 1 || COMPAT_PRELITE2 */
2987 #endif /* !BURN_BRIDGES */
2988 
2989 #define KINFO_VNODESLOP		10
2990 #ifdef notyet
2991 /*
2992  * Dump vnode list (via sysctl).
2993  */
2994 /* ARGSUSED */
2995 static int
2996 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2997 {
2998 	struct xvnode *xvn;
2999 	struct thread *td = req->td;
3000 	struct mount *mp;
3001 	struct vnode *vp;
3002 	int error, len, n;
3003 
3004 	/*
3005 	 * Stale numvnodes access is not fatal here.
3006 	 */
3007 	req->lock = 0;
3008 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3009 	if (!req->oldptr)
3010 		/* Make an estimate */
3011 		return (SYSCTL_OUT(req, 0, len));
3012 
3013 	error = sysctl_wire_old_buffer(req, 0);
3014 	if (error != 0)
3015 		return (error);
3016 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3017 	n = 0;
3018 	mtx_lock(&mountlist_mtx);
3019 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3020 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
3021 			continue;
3022 		MNT_ILOCK(mp);
3023 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3024 			if (n == len)
3025 				break;
3026 			vref(vp);
3027 			xvn[n].xv_size = sizeof *xvn;
3028 			xvn[n].xv_vnode = vp;
3029 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3030 			XV_COPY(usecount);
3031 			XV_COPY(writecount);
3032 			XV_COPY(holdcnt);
3033 			XV_COPY(id);
3034 			XV_COPY(mount);
3035 			XV_COPY(numoutput);
3036 			XV_COPY(type);
3037 #undef XV_COPY
3038 			xvn[n].xv_flag = vp->v_vflag;
3039 
3040 			switch (vp->v_type) {
3041 			case VREG:
3042 			case VDIR:
3043 			case VLNK:
3044 				xvn[n].xv_dev = vp->v_cachedfs;
3045 				xvn[n].xv_ino = vp->v_cachedid;
3046 				break;
3047 			case VBLK:
3048 			case VCHR:
3049 				if (vp->v_rdev == NULL) {
3050 					vrele(vp);
3051 					continue;
3052 				}
3053 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3054 				break;
3055 			case VSOCK:
3056 				xvn[n].xv_socket = vp->v_socket;
3057 				break;
3058 			case VFIFO:
3059 				xvn[n].xv_fifo = vp->v_fifoinfo;
3060 				break;
3061 			case VNON:
3062 			case VBAD:
3063 			default:
3064 				/* shouldn't happen? */
3065 				vrele(vp);
3066 				continue;
3067 			}
3068 			vrele(vp);
3069 			++n;
3070 		}
3071 		MNT_IUNLOCK(mp);
3072 		mtx_lock(&mountlist_mtx);
3073 		vfs_unbusy(mp, td);
3074 		if (n == len)
3075 			break;
3076 	}
3077 	mtx_unlock(&mountlist_mtx);
3078 
3079 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3080 	free(xvn, M_TEMP);
3081 	return (error);
3082 }
3083 
3084 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
3085 	0, 0, sysctl_vnode, "S,xvnode", "");
3086 #endif
3087 
3088 /*
3089  * Check to see if a filesystem is mounted on a block device.
3090  */
3091 int
3092 vfs_mountedon(vp)
3093 	struct vnode *vp;
3094 {
3095 
3096 	if (vp->v_rdev->si_mountpoint != NULL)
3097 		return (EBUSY);
3098 	return (0);
3099 }
3100 
3101 /*
3102  * Unmount all filesystems. The list is traversed in reverse order
3103  * of mounting to avoid dependencies.
3104  */
3105 void
3106 vfs_unmountall()
3107 {
3108 	struct mount *mp;
3109 	struct thread *td;
3110 	int error;
3111 
3112 	if (curthread != NULL)
3113 		td = curthread;
3114 	else
3115 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
3116 	/*
3117 	 * Since this only runs when rebooting, it is not interlocked.
3118 	 */
3119 	while(!TAILQ_EMPTY(&mountlist)) {
3120 		mp = TAILQ_LAST(&mountlist, mntlist);
3121 		error = dounmount(mp, MNT_FORCE, td);
3122 		if (error) {
3123 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
3124 			printf("unmount of %s failed (",
3125 			    mp->mnt_stat.f_mntonname);
3126 			if (error == EBUSY)
3127 				printf("BUSY)\n");
3128 			else
3129 				printf("%d)\n", error);
3130 		} else {
3131 			/* The unmount has removed mp from the mountlist */
3132 		}
3133 	}
3134 }
3135 
3136 /*
3137  * perform msync on all vnodes under a mount point
3138  * the mount point must be locked.
3139  */
3140 void
3141 vfs_msync(struct mount *mp, int flags)
3142 {
3143 	struct vnode *vp, *nvp;
3144 	struct vm_object *obj;
3145 	int tries;
3146 
3147 	GIANT_REQUIRED;
3148 
3149 	tries = 5;
3150 	MNT_ILOCK(mp);
3151 loop:
3152 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
3153 		if (vp->v_mount != mp) {
3154 			if (--tries > 0)
3155 				goto loop;
3156 			break;
3157 		}
3158 
3159 		VI_LOCK(vp);
3160 		if (vp->v_iflag & VI_XLOCK) {
3161 			VI_UNLOCK(vp);
3162 			continue;
3163 		}
3164 
3165 		if ((vp->v_iflag & VI_OBJDIRTY) &&
3166 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
3167 			MNT_IUNLOCK(mp);
3168 			if (!vget(vp,
3169 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3170 			    curthread)) {
3171 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3172 					vput(vp);
3173 					MNT_ILOCK(mp);
3174 					continue;
3175 				}
3176 
3177 				if (VOP_GETVOBJECT(vp, &obj) == 0) {
3178 					VM_OBJECT_LOCK(obj);
3179 					vm_object_page_clean(obj, 0, 0,
3180 					    flags == MNT_WAIT ?
3181 					    OBJPC_SYNC : OBJPC_NOSYNC);
3182 					VM_OBJECT_UNLOCK(obj);
3183 				}
3184 				vput(vp);
3185 			}
3186 			MNT_ILOCK(mp);
3187 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
3188 				if (--tries > 0)
3189 					goto loop;
3190 				break;
3191 			}
3192 		} else
3193 			VI_UNLOCK(vp);
3194 	}
3195 	MNT_IUNLOCK(mp);
3196 }
3197 
3198 /*
3199  * Create the VM object needed for VMIO and mmap support.  This
3200  * is done for all VREG files in the system.  Some filesystems might
3201  * afford the additional metadata buffering capability of the
3202  * VMIO code by making the device node be VMIO mode also.
3203  *
3204  * vp must be locked when vfs_object_create is called.
3205  */
3206 int
3207 vfs_object_create(vp, td, cred)
3208 	struct vnode *vp;
3209 	struct thread *td;
3210 	struct ucred *cred;
3211 {
3212 
3213 	GIANT_REQUIRED;
3214 	return (VOP_CREATEVOBJECT(vp, cred, td));
3215 }
3216 
3217 /*
3218  * Mark a vnode as free, putting it up for recycling.
3219  */
3220 void
3221 vfree(vp)
3222 	struct vnode *vp;
3223 {
3224 
3225 	ASSERT_VI_LOCKED(vp, "vfree");
3226 	mtx_lock(&vnode_free_list_mtx);
3227 	KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free"));
3228 	if (vp->v_iflag & VI_AGE) {
3229 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
3230 	} else {
3231 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
3232 	}
3233 	freevnodes++;
3234 	mtx_unlock(&vnode_free_list_mtx);
3235 	vp->v_iflag &= ~VI_AGE;
3236 	vp->v_iflag |= VI_FREE;
3237 }
3238 
3239 /*
3240  * Opposite of vfree() - mark a vnode as in use.
3241  */
3242 void
3243 vbusy(vp)
3244 	struct vnode *vp;
3245 {
3246 
3247 	ASSERT_VI_LOCKED(vp, "vbusy");
3248 	KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free"));
3249 
3250 	mtx_lock(&vnode_free_list_mtx);
3251 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
3252 	freevnodes--;
3253 	mtx_unlock(&vnode_free_list_mtx);
3254 
3255 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
3256 }
3257 
3258 /*
3259  * Initalize per-vnode helper structure to hold poll-related state.
3260  */
3261 void
3262 v_addpollinfo(struct vnode *vp)
3263 {
3264 
3265 	vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK);
3266 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3267 }
3268 
3269 /*
3270  * Record a process's interest in events which might happen to
3271  * a vnode.  Because poll uses the historic select-style interface
3272  * internally, this routine serves as both the ``check for any
3273  * pending events'' and the ``record my interest in future events''
3274  * functions.  (These are done together, while the lock is held,
3275  * to avoid race conditions.)
3276  */
3277 int
3278 vn_pollrecord(vp, td, events)
3279 	struct vnode *vp;
3280 	struct thread *td;
3281 	short events;
3282 {
3283 
3284 	if (vp->v_pollinfo == NULL)
3285 		v_addpollinfo(vp);
3286 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3287 	if (vp->v_pollinfo->vpi_revents & events) {
3288 		/*
3289 		 * This leaves events we are not interested
3290 		 * in available for the other process which
3291 		 * which presumably had requested them
3292 		 * (otherwise they would never have been
3293 		 * recorded).
3294 		 */
3295 		events &= vp->v_pollinfo->vpi_revents;
3296 		vp->v_pollinfo->vpi_revents &= ~events;
3297 
3298 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3299 		return events;
3300 	}
3301 	vp->v_pollinfo->vpi_events |= events;
3302 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3303 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3304 	return 0;
3305 }
3306 
3307 /*
3308  * Note the occurrence of an event.  If the VN_POLLEVENT macro is used,
3309  * it is possible for us to miss an event due to race conditions, but
3310  * that condition is expected to be rare, so for the moment it is the
3311  * preferred interface.
3312  */
3313 void
3314 vn_pollevent(vp, events)
3315 	struct vnode *vp;
3316 	short events;
3317 {
3318 
3319 	if (vp->v_pollinfo == NULL)
3320 		v_addpollinfo(vp);
3321 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3322 	if (vp->v_pollinfo->vpi_events & events) {
3323 		/*
3324 		 * We clear vpi_events so that we don't
3325 		 * call selwakeup() twice if two events are
3326 		 * posted before the polling process(es) is
3327 		 * awakened.  This also ensures that we take at
3328 		 * most one selwakeup() if the polling process
3329 		 * is no longer interested.  However, it does
3330 		 * mean that only one event can be noticed at
3331 		 * a time.  (Perhaps we should only clear those
3332 		 * event bits which we note?) XXX
3333 		 */
3334 		vp->v_pollinfo->vpi_events = 0;	/* &= ~events ??? */
3335 		vp->v_pollinfo->vpi_revents |= events;
3336 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
3337 	}
3338 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3339 }
3340 
3341 /*
3342  * Wake up anyone polling on vp because it is being revoked.
3343  * This depends on dead_poll() returning POLLHUP for correct
3344  * behavior.
3345  */
3346 void
3347 vn_pollgone(vp)
3348 	struct vnode *vp;
3349 {
3350 
3351 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3352 	VN_KNOTE(vp, NOTE_REVOKE);
3353 	if (vp->v_pollinfo->vpi_events) {
3354 		vp->v_pollinfo->vpi_events = 0;
3355 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
3356 	}
3357 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3358 }
3359 
3360 
3361 
3362 /*
3363  * Routine to create and manage a filesystem syncer vnode.
3364  */
3365 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3366 static int	sync_fsync(struct  vop_fsync_args *);
3367 static int	sync_inactive(struct  vop_inactive_args *);
3368 static int	sync_reclaim(struct  vop_reclaim_args *);
3369 
3370 static vop_t **sync_vnodeop_p;
3371 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = {
3372 	{ &vop_default_desc,	(vop_t *) vop_eopnotsupp },
3373 	{ &vop_close_desc,	(vop_t *) sync_close },		/* close */
3374 	{ &vop_fsync_desc,	(vop_t *) sync_fsync },		/* fsync */
3375 	{ &vop_inactive_desc,	(vop_t *) sync_inactive },	/* inactive */
3376 	{ &vop_reclaim_desc,	(vop_t *) sync_reclaim },	/* reclaim */
3377 	{ &vop_lock_desc,	(vop_t *) vop_stdlock },	/* lock */
3378 	{ &vop_unlock_desc,	(vop_t *) vop_stdunlock },	/* unlock */
3379 	{ &vop_islocked_desc,	(vop_t *) vop_stdislocked },	/* islocked */
3380 	{ NULL, NULL }
3381 };
3382 static struct vnodeopv_desc sync_vnodeop_opv_desc =
3383 	{ &sync_vnodeop_p, sync_vnodeop_entries };
3384 
3385 VNODEOP_SET(sync_vnodeop_opv_desc);
3386 
3387 /*
3388  * Create a new filesystem syncer vnode for the specified mount point.
3389  */
3390 int
3391 vfs_allocate_syncvnode(mp)
3392 	struct mount *mp;
3393 {
3394 	struct vnode *vp;
3395 	static long start, incr, next;
3396 	int error;
3397 
3398 	/* Allocate a new vnode */
3399 	if ((error = getnewvnode("syncer", mp, sync_vnodeop_p, &vp)) != 0) {
3400 		mp->mnt_syncer = NULL;
3401 		return (error);
3402 	}
3403 	vp->v_type = VNON;
3404 	/*
3405 	 * Place the vnode onto the syncer worklist. We attempt to
3406 	 * scatter them about on the list so that they will go off
3407 	 * at evenly distributed times even if all the filesystems
3408 	 * are mounted at once.
3409 	 */
3410 	next += incr;
3411 	if (next == 0 || next > syncer_maxdelay) {
3412 		start /= 2;
3413 		incr /= 2;
3414 		if (start == 0) {
3415 			start = syncer_maxdelay / 2;
3416 			incr = syncer_maxdelay;
3417 		}
3418 		next = start;
3419 	}
3420 	VI_LOCK(vp);
3421 	vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0);
3422 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3423 	mtx_lock(&sync_mtx);
3424 	sync_vnode_count++;
3425 	mtx_unlock(&sync_mtx);
3426 	VI_UNLOCK(vp);
3427 	mp->mnt_syncer = vp;
3428 	return (0);
3429 }
3430 
3431 /*
3432  * Do a lazy sync of the filesystem.
3433  */
3434 static int
3435 sync_fsync(ap)
3436 	struct vop_fsync_args /* {
3437 		struct vnode *a_vp;
3438 		struct ucred *a_cred;
3439 		int a_waitfor;
3440 		struct thread *a_td;
3441 	} */ *ap;
3442 {
3443 	struct vnode *syncvp = ap->a_vp;
3444 	struct mount *mp = syncvp->v_mount;
3445 	struct thread *td = ap->a_td;
3446 	int error, asyncflag;
3447 
3448 	/*
3449 	 * We only need to do something if this is a lazy evaluation.
3450 	 */
3451 	if (ap->a_waitfor != MNT_LAZY)
3452 		return (0);
3453 
3454 	/*
3455 	 * Move ourselves to the back of the sync list.
3456 	 */
3457 	VI_LOCK(syncvp);
3458 	vn_syncer_add_to_worklist(syncvp, syncdelay);
3459 	VI_UNLOCK(syncvp);
3460 
3461 	/*
3462 	 * Walk the list of vnodes pushing all that are dirty and
3463 	 * not already on the sync list.
3464 	 */
3465 	mtx_lock(&mountlist_mtx);
3466 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3467 		mtx_unlock(&mountlist_mtx);
3468 		return (0);
3469 	}
3470 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3471 		vfs_unbusy(mp, td);
3472 		return (0);
3473 	}
3474 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3475 	mp->mnt_flag &= ~MNT_ASYNC;
3476 	vfs_msync(mp, MNT_NOWAIT);
3477 	error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td);
3478 	if (asyncflag)
3479 		mp->mnt_flag |= MNT_ASYNC;
3480 	vn_finished_write(mp);
3481 	vfs_unbusy(mp, td);
3482 	return (error);
3483 }
3484 
3485 /*
3486  * The syncer vnode is no referenced.
3487  */
3488 static int
3489 sync_inactive(ap)
3490 	struct vop_inactive_args /* {
3491 		struct vnode *a_vp;
3492 		struct thread *a_td;
3493 	} */ *ap;
3494 {
3495 
3496 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3497 	vgone(ap->a_vp);
3498 	return (0);
3499 }
3500 
3501 /*
3502  * The syncer vnode is no longer needed and is being decommissioned.
3503  *
3504  * Modifications to the worklist must be protected by sync_mtx.
3505  */
3506 static int
3507 sync_reclaim(ap)
3508 	struct vop_reclaim_args /* {
3509 		struct vnode *a_vp;
3510 	} */ *ap;
3511 {
3512 	struct vnode *vp = ap->a_vp;
3513 
3514 	VI_LOCK(vp);
3515 	vp->v_mount->mnt_syncer = NULL;
3516 	if (vp->v_iflag & VI_ONWORKLST) {
3517 		mtx_lock(&sync_mtx);
3518 		LIST_REMOVE(vp, v_synclist);
3519  		syncer_worklist_len--;
3520 		sync_vnode_count--;
3521 		mtx_unlock(&sync_mtx);
3522 		vp->v_iflag &= ~VI_ONWORKLST;
3523 	}
3524 	VI_UNLOCK(vp);
3525 
3526 	return (0);
3527 }
3528 
3529 /*
3530  * extract the struct cdev *from a VCHR
3531  */
3532 struct cdev *
3533 vn_todev(vp)
3534 	struct vnode *vp;
3535 {
3536 
3537 	if (vp->v_type != VCHR)
3538 		return (NULL);
3539 	return (vp->v_rdev);
3540 }
3541 
3542 /*
3543  * Check if vnode represents a disk device
3544  */
3545 int
3546 vn_isdisk(vp, errp)
3547 	struct vnode *vp;
3548 	int *errp;
3549 {
3550 	int error;
3551 
3552 	error = 0;
3553 	if (vp->v_type != VCHR)
3554 		error = ENOTBLK;
3555 	else if (vp->v_rdev == NULL)
3556 		error = ENXIO;
3557 	else if (!(devsw(vp->v_rdev)->d_flags & D_DISK))
3558 		error = ENOTBLK;
3559 	if (errp != NULL)
3560 		*errp = error;
3561 	return (error == 0);
3562 }
3563 
3564 /*
3565  * Free data allocated by namei(); see namei(9) for details.
3566  */
3567 void
3568 NDFREE(ndp, flags)
3569      struct nameidata *ndp;
3570      const u_int flags;
3571 {
3572 
3573 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3574 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3575 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3576 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3577 	}
3578 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3579 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3580 	    ndp->ni_dvp != ndp->ni_vp)
3581 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3582 	if (!(flags & NDF_NO_DVP_RELE) &&
3583 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3584 		vrele(ndp->ni_dvp);
3585 		ndp->ni_dvp = NULL;
3586 	}
3587 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3588 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3589 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3590 	if (!(flags & NDF_NO_VP_RELE) &&
3591 	    ndp->ni_vp) {
3592 		vrele(ndp->ni_vp);
3593 		ndp->ni_vp = NULL;
3594 	}
3595 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3596 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3597 		vrele(ndp->ni_startdir);
3598 		ndp->ni_startdir = NULL;
3599 	}
3600 }
3601 
3602 /*
3603  * Common filesystem object access control check routine.  Accepts a
3604  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3605  * and optional call-by-reference privused argument allowing vaccess()
3606  * to indicate to the caller whether privilege was used to satisfy the
3607  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3608  */
3609 int
3610 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3611 	enum vtype type;
3612 	mode_t file_mode;
3613 	uid_t file_uid;
3614 	gid_t file_gid;
3615 	mode_t acc_mode;
3616 	struct ucred *cred;
3617 	int *privused;
3618 {
3619 	mode_t dac_granted;
3620 #ifdef CAPABILITIES
3621 	mode_t cap_granted;
3622 #endif
3623 
3624 	/*
3625 	 * Look for a normal, non-privileged way to access the file/directory
3626 	 * as requested.  If it exists, go with that.
3627 	 */
3628 
3629 	if (privused != NULL)
3630 		*privused = 0;
3631 
3632 	dac_granted = 0;
3633 
3634 	/* Check the owner. */
3635 	if (cred->cr_uid == file_uid) {
3636 		dac_granted |= VADMIN;
3637 		if (file_mode & S_IXUSR)
3638 			dac_granted |= VEXEC;
3639 		if (file_mode & S_IRUSR)
3640 			dac_granted |= VREAD;
3641 		if (file_mode & S_IWUSR)
3642 			dac_granted |= (VWRITE | VAPPEND);
3643 
3644 		if ((acc_mode & dac_granted) == acc_mode)
3645 			return (0);
3646 
3647 		goto privcheck;
3648 	}
3649 
3650 	/* Otherwise, check the groups (first match) */
3651 	if (groupmember(file_gid, cred)) {
3652 		if (file_mode & S_IXGRP)
3653 			dac_granted |= VEXEC;
3654 		if (file_mode & S_IRGRP)
3655 			dac_granted |= VREAD;
3656 		if (file_mode & S_IWGRP)
3657 			dac_granted |= (VWRITE | VAPPEND);
3658 
3659 		if ((acc_mode & dac_granted) == acc_mode)
3660 			return (0);
3661 
3662 		goto privcheck;
3663 	}
3664 
3665 	/* Otherwise, check everyone else. */
3666 	if (file_mode & S_IXOTH)
3667 		dac_granted |= VEXEC;
3668 	if (file_mode & S_IROTH)
3669 		dac_granted |= VREAD;
3670 	if (file_mode & S_IWOTH)
3671 		dac_granted |= (VWRITE | VAPPEND);
3672 	if ((acc_mode & dac_granted) == acc_mode)
3673 		return (0);
3674 
3675 privcheck:
3676 	if (!suser_cred(cred, PRISON_ROOT)) {
3677 		/* XXX audit: privilege used */
3678 		if (privused != NULL)
3679 			*privused = 1;
3680 		return (0);
3681 	}
3682 
3683 #ifdef CAPABILITIES
3684 	/*
3685 	 * Build a capability mask to determine if the set of capabilities
3686 	 * satisfies the requirements when combined with the granted mask
3687 	 * from above.
3688 	 * For each capability, if the capability is required, bitwise
3689 	 * or the request type onto the cap_granted mask.
3690 	 */
3691 	cap_granted = 0;
3692 
3693 	if (type == VDIR) {
3694 		/*
3695 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3696 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3697 		 */
3698 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3699 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT))
3700 			cap_granted |= VEXEC;
3701 	} else {
3702 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3703 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT))
3704 			cap_granted |= VEXEC;
3705 	}
3706 
3707 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3708 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT))
3709 		cap_granted |= VREAD;
3710 
3711 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3712 	    !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT))
3713 		cap_granted |= (VWRITE | VAPPEND);
3714 
3715 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3716 	    !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT))
3717 		cap_granted |= VADMIN;
3718 
3719 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3720 		/* XXX audit: privilege used */
3721 		if (privused != NULL)
3722 			*privused = 1;
3723 		return (0);
3724 	}
3725 #endif
3726 
3727 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3728 }
3729 
3730 /*
3731  * Credential check based on process requesting service, and per-attribute
3732  * permissions.
3733  */
3734 int
3735 extattr_check_cred(struct vnode *vp, int attrnamespace,
3736     struct ucred *cred, struct thread *td, int access)
3737 {
3738 
3739 	/*
3740 	 * Kernel-invoked always succeeds.
3741 	 */
3742 	if (cred == NOCRED)
3743 		return (0);
3744 
3745 	/*
3746 	 * Do not allow privileged processes in jail to directly
3747 	 * manipulate system attributes.
3748 	 *
3749 	 * XXX What capability should apply here?
3750 	 * Probably CAP_SYS_SETFFLAG.
3751 	 */
3752 	switch (attrnamespace) {
3753 	case EXTATTR_NAMESPACE_SYSTEM:
3754 		/* Potentially should be: return (EPERM); */
3755 		return (suser_cred(cred, 0));
3756 	case EXTATTR_NAMESPACE_USER:
3757 		return (VOP_ACCESS(vp, access, cred, td));
3758 	default:
3759 		return (EPERM);
3760 	}
3761 }
3762 
3763 #ifdef DEBUG_VFS_LOCKS
3764 /*
3765  * This only exists to supress warnings from unlocked specfs accesses.  It is
3766  * no longer ok to have an unlocked VFS.
3767  */
3768 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3769 
3770 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3771 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3772 int vfs_badlock_print = 1;	/* Print lock violations. */
3773 
3774 static void
3775 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3776 {
3777 
3778 	if (vfs_badlock_print)
3779 		printf("%s: %p %s\n", str, (void *)vp, msg);
3780 	if (vfs_badlock_ddb)
3781 		kdb_enter("lock violation");
3782 }
3783 
3784 void
3785 assert_vi_locked(struct vnode *vp, const char *str)
3786 {
3787 
3788 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3789 		vfs_badlock("interlock is not locked but should be", str, vp);
3790 }
3791 
3792 void
3793 assert_vi_unlocked(struct vnode *vp, const char *str)
3794 {
3795 
3796 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3797 		vfs_badlock("interlock is locked but should not be", str, vp);
3798 }
3799 
3800 void
3801 assert_vop_locked(struct vnode *vp, const char *str)
3802 {
3803 
3804 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3805 		vfs_badlock("is not locked but should be", str, vp);
3806 }
3807 
3808 void
3809 assert_vop_unlocked(struct vnode *vp, const char *str)
3810 {
3811 
3812 	if (vp && !IGNORE_LOCK(vp) &&
3813 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3814 		vfs_badlock("is locked but should not be", str, vp);
3815 }
3816 
3817 #if 0
3818 void
3819 assert_vop_elocked(struct vnode *vp, const char *str)
3820 {
3821 
3822 	if (vp && !IGNORE_LOCK(vp) &&
3823 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3824 		vfs_badlock("is not exclusive locked but should be", str, vp);
3825 }
3826 
3827 void
3828 assert_vop_elocked_other(struct vnode *vp, const char *str)
3829 {
3830 
3831 	if (vp && !IGNORE_LOCK(vp) &&
3832 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3833 		vfs_badlock("is not exclusive locked by another thread",
3834 		    str, vp);
3835 }
3836 
3837 void
3838 assert_vop_slocked(struct vnode *vp, const char *str)
3839 {
3840 
3841 	if (vp && !IGNORE_LOCK(vp) &&
3842 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3843 		vfs_badlock("is not locked shared but should be", str, vp);
3844 }
3845 #endif /* 0 */
3846 
3847 void
3848 vop_rename_pre(void *ap)
3849 {
3850 	struct vop_rename_args *a = ap;
3851 
3852 	if (a->a_tvp)
3853 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3854 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3855 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3856 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3857 
3858 	/* Check the source (from). */
3859 	if (a->a_tdvp != a->a_fdvp)
3860 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3861 	if (a->a_tvp != a->a_fvp)
3862 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3863 
3864 	/* Check the target. */
3865 	if (a->a_tvp)
3866 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3867 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3868 }
3869 
3870 void
3871 vop_strategy_pre(void *ap)
3872 {
3873 	struct vop_strategy_args *a;
3874 	struct buf *bp;
3875 
3876 	a = ap;
3877 	bp = a->a_bp;
3878 
3879 	/*
3880 	 * Cluster ops lock their component buffers but not the IO container.
3881 	 */
3882 	if ((bp->b_flags & B_CLUSTER) != 0)
3883 		return;
3884 
3885 	if (BUF_REFCNT(bp) < 1) {
3886 		if (vfs_badlock_print)
3887 			printf(
3888 			    "VOP_STRATEGY: bp is not locked but should be\n");
3889 		if (vfs_badlock_ddb)
3890 			kdb_enter("lock violation");
3891 	}
3892 }
3893 
3894 void
3895 vop_lookup_pre(void *ap)
3896 {
3897 	struct vop_lookup_args *a;
3898 	struct vnode *dvp;
3899 
3900 	a = ap;
3901 	dvp = a->a_dvp;
3902 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3903 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3904 }
3905 
3906 void
3907 vop_lookup_post(void *ap, int rc)
3908 {
3909 	struct vop_lookup_args *a;
3910 	struct componentname *cnp;
3911 	struct vnode *dvp;
3912 	struct vnode *vp;
3913 	int flags;
3914 
3915 	a = ap;
3916 	dvp = a->a_dvp;
3917 	cnp = a->a_cnp;
3918 	vp = *(a->a_vpp);
3919 	flags = cnp->cn_flags;
3920 
3921 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3922 
3923 	/*
3924 	 * If this is the last path component for this lookup and LOCKPARENT
3925 	 * is set, OR if there is an error the directory has to be locked.
3926 	 */
3927 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
3928 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
3929 	else if (rc != 0)
3930 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
3931 	else if (dvp != vp)
3932 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
3933 	if (flags & PDIRUNLOCK)
3934 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
3935 }
3936 
3937 void
3938 vop_lock_pre(void *ap)
3939 {
3940 	struct vop_lock_args *a = ap;
3941 
3942 	if ((a->a_flags & LK_INTERLOCK) == 0)
3943 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3944 	else
3945 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3946 }
3947 
3948 void
3949 vop_lock_post(void *ap, int rc)
3950 {
3951 	struct vop_lock_args *a = ap;
3952 
3953 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3954 	if (rc == 0)
3955 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3956 }
3957 
3958 void
3959 vop_unlock_pre(void *ap)
3960 {
3961 	struct vop_unlock_args *a = ap;
3962 
3963 	if (a->a_flags & LK_INTERLOCK)
3964 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3965 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3966 }
3967 
3968 void
3969 vop_unlock_post(void *ap, int rc)
3970 {
3971 	struct vop_unlock_args *a = ap;
3972 
3973 	if (a->a_flags & LK_INTERLOCK)
3974 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3975 }
3976 #endif /* DEBUG_VFS_LOCKS */
3977 
3978 static struct klist fs_klist = SLIST_HEAD_INITIALIZER(&fs_klist);
3979 
3980 void
3981 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3982 {
3983 
3984 	KNOTE(&fs_klist, event);
3985 }
3986 
3987 static int	filt_fsattach(struct knote *kn);
3988 static void	filt_fsdetach(struct knote *kn);
3989 static int	filt_fsevent(struct knote *kn, long hint);
3990 
3991 struct filterops fs_filtops =
3992 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3993 
3994 static int
3995 filt_fsattach(struct knote *kn)
3996 {
3997 
3998 	kn->kn_flags |= EV_CLEAR;
3999 	SLIST_INSERT_HEAD(&fs_klist, kn, kn_selnext);
4000 	return (0);
4001 }
4002 
4003 static void
4004 filt_fsdetach(struct knote *kn)
4005 {
4006 
4007 	SLIST_REMOVE(&fs_klist, kn, knote, kn_selnext);
4008 }
4009 
4010 static int
4011 filt_fsevent(struct knote *kn, long hint)
4012 {
4013 
4014 	kn->kn_fflags |= hint;
4015 	return (kn->kn_fflags != 0);
4016 }
4017 
4018 static int
4019 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4020 {
4021 	struct vfsidctl vc;
4022 	int error;
4023 	struct mount *mp;
4024 
4025 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4026 	if (error)
4027 		return (error);
4028 	if (vc.vc_vers != VFS_CTL_VERS1)
4029 		return (EINVAL);
4030 	mp = vfs_getvfs(&vc.vc_fsid);
4031 	if (mp == NULL)
4032 		return (ENOENT);
4033 	/* ensure that a specific sysctl goes to the right filesystem. */
4034 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4035 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4036 		return (EINVAL);
4037 	}
4038 	VCTLTOREQ(&vc, req);
4039 	return (VFS_SYSCTL(mp, vc.vc_op, req));
4040 }
4041 
4042 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
4043         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
4044