xref: /freebsd/sys/kern/vfs_subr.c (revision 0b87f79976047c8f4332bbf7dc03146f6b0de79f)
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  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
39  * $FreeBSD$
40  */
41 
42 /*
43  * External virtual filesystem routines
44  */
45 #include "opt_ddb.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/eventhandler.h>
53 #include <sys/fcntl.h>
54 #include <sys/kernel.h>
55 #include <sys/kthread.h>
56 #include <sys/malloc.h>
57 #include <sys/mount.h>
58 #include <sys/namei.h>
59 #include <sys/stat.h>
60 #include <sys/sysctl.h>
61 #include <sys/syslog.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_extern.h>
68 #include <vm/pmap.h>
69 #include <vm/vm_map.h>
70 #include <vm/vm_page.h>
71 #include <vm/uma.h>
72 
73 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
74 
75 static void	addalias(struct vnode *vp, dev_t nvp_rdev);
76 static void	insmntque(struct vnode *vp, struct mount *mp);
77 static void	vclean(struct vnode *vp, int flags, struct thread *td);
78 static void	vlruvp(struct vnode *vp);
79 
80 /*
81  * Number of vnodes in existence.  Increased whenever getnewvnode()
82  * allocates a new vnode, never decreased.
83  */
84 static unsigned long	numvnodes;
85 
86 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
87 
88 /*
89  * Conversion tables for conversion from vnode types to inode formats
90  * and back.
91  */
92 enum vtype iftovt_tab[16] = {
93 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
94 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
95 };
96 int vttoif_tab[9] = {
97 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
98 	S_IFSOCK, S_IFIFO, S_IFMT,
99 };
100 
101 /*
102  * List of vnodes that are ready for recycling.
103  */
104 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
105 
106 /*
107  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
108  * getnewvnode() will return a newly allocated vnode.
109  */
110 static u_long wantfreevnodes = 25;
111 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
112 /* Number of vnodes in the free list. */
113 static u_long freevnodes;
114 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
115 
116 /*
117  * Various variables used for debugging the new implementation of
118  * reassignbuf().
119  * XXX these are probably of (very) limited utility now.
120  */
121 static int reassignbufcalls;
122 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
123 static int nameileafonly;
124 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
125 
126 #ifdef ENABLE_VFS_IOOPT
127 /* See NOTES for a description of this setting. */
128 int vfs_ioopt;
129 SYSCTL_INT(_vfs, OID_AUTO, ioopt, CTLFLAG_RW, &vfs_ioopt, 0, "");
130 #endif
131 
132 /*
133  * Cache for the mount type id assigned to NFS.  This is used for
134  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
135  */
136 int	nfs_mount_type = -1;
137 
138 /* To keep more than one thread at a time from running vfs_getnewfsid */
139 static struct mtx mntid_mtx;
140 
141 /* For any iteration/modification of vnode_free_list */
142 static struct mtx vnode_free_list_mtx;
143 
144 /*
145  * For any iteration/modification of dev->si_hlist (linked through
146  * v_specnext)
147  */
148 static struct mtx spechash_mtx;
149 
150 /* Publicly exported FS */
151 struct nfs_public nfs_pub;
152 
153 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
154 static uma_zone_t vnode_zone;
155 static uma_zone_t vnodepoll_zone;
156 
157 /* Set to 1 to print out reclaim of active vnodes */
158 int	prtactive;
159 
160 /*
161  * The workitem queue.
162  *
163  * It is useful to delay writes of file data and filesystem metadata
164  * for tens of seconds so that quickly created and deleted files need
165  * not waste disk bandwidth being created and removed. To realize this,
166  * we append vnodes to a "workitem" queue. When running with a soft
167  * updates implementation, most pending metadata dependencies should
168  * not wait for more than a few seconds. Thus, mounted on block devices
169  * are delayed only about a half the time that file data is delayed.
170  * Similarly, directory updates are more critical, so are only delayed
171  * about a third the time that file data is delayed. Thus, there are
172  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
173  * one each second (driven off the filesystem syncer process). The
174  * syncer_delayno variable indicates the next queue that is to be processed.
175  * Items that need to be processed soon are placed in this queue:
176  *
177  *	syncer_workitem_pending[syncer_delayno]
178  *
179  * A delay of fifteen seconds is done by placing the request fifteen
180  * entries later in the queue:
181  *
182  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
183  *
184  */
185 static int syncer_delayno;
186 static long syncer_mask;
187 LIST_HEAD(synclist, vnode);
188 static struct synclist *syncer_workitem_pending;
189 
190 #define SYNCER_MAXDELAY		32
191 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
192 static int syncdelay = 30;		/* max time to delay syncing data */
193 static int filedelay = 30;		/* time to delay syncing files */
194 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
195 static int dirdelay = 29;		/* time to delay syncing directories */
196 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
197 static int metadelay = 28;		/* time to delay syncing metadata */
198 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
199 static int rushjob;		/* number of slots to run ASAP */
200 static int stat_rush_requests;	/* number of times I/O speeded up */
201 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
202 
203 /*
204  * Number of vnodes we want to exist at any one time.  This is mostly used
205  * to size hash tables in vnode-related code.  It is normally not used in
206  * getnewvnode(), as wantfreevnodes is normally nonzero.)
207  *
208  * XXX desiredvnodes is historical cruft and should not exist.
209  */
210 int desiredvnodes;
211 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
212     &desiredvnodes, 0, "Maximum number of vnodes");
213 static int minvnodes;
214 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
215     &minvnodes, 0, "Minimum number of vnodes");
216 static int vnlru_nowhere;
217 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0,
218     "Number of times the vnlru process ran without success");
219 
220 /* Hook for calling soft updates */
221 int (*softdep_process_worklist_hook)(struct mount *);
222 
223 #ifdef DEBUG_VFS_LOCKS
224 /* Print lock violations */
225 int vfs_badlock_print = 1;
226 /* Panic on violation */
227 int vfs_badlock_panic = 1;
228 
229 void
230 vop_rename_pre(void *ap)
231 {
232 	struct vop_rename_args *a = ap;
233 
234 	/* Check the source (from) */
235 	if (a->a_tdvp != a->a_fdvp)
236 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked.\n");
237 	if (a->a_tvp != a->a_fvp)
238 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked.\n");
239 
240 	/* Check the target */
241 	if (a->a_tvp)
242 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked.\n");
243 
244 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked.\n");
245 }
246 
247 void
248 vop_strategy_pre(void *ap)
249 {
250 	struct vop_strategy_args *a = ap;
251 	struct buf *bp;
252 
253 	bp = a->a_bp;
254 
255 	/*
256 	 * Cluster ops lock their component buffers but not the IO container.
257 	 */
258 	if ((bp->b_flags & B_CLUSTER) != 0)
259 		return;
260 
261 	if (BUF_REFCNT(bp) < 1) {
262 		if (vfs_badlock_print)
263 			printf("VOP_STRATEGY: bp is not locked but should be.\n");
264 		if (vfs_badlock_panic)
265 			Debugger("Lock violation.\n");
266 	}
267 }
268 
269 void
270 vop_lookup_pre(void *ap)
271 {
272 	struct vop_lookup_args *a = ap;
273 	struct vnode *dvp;
274 
275 	dvp = a->a_dvp;
276 
277 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
278 }
279 
280 void
281 vop_lookup_post(void *ap, int rc)
282 {
283 	struct vop_lookup_args *a = ap;
284 	struct componentname *cnp;
285 	struct vnode *dvp;
286 	struct vnode *vp;
287 	int flags;
288 
289 	dvp = a->a_dvp;
290 	cnp = a->a_cnp;
291 	vp = *(a->a_vpp);
292 	flags = cnp->cn_flags;
293 
294 
295 	/*
296 	 * If this is the last path component for this lookup and LOCPARENT
297 	 * is set, OR if there is an error the directory has to be locked.
298 	 */
299 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
300 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
301 	else if (rc != 0)
302 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
303 	else if (dvp != vp)
304 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
305 
306 	if (flags & PDIRUNLOCK)
307 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
308 
309 	if (rc == 0)
310 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (vpp)");
311 }
312 
313 #endif	/* DEBUG_VFS_LOCKS */
314 
315 void
316 v_addpollinfo(struct vnode *vp)
317 {
318 	vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK);
319 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
320 }
321 
322 /*
323  * Initialize the vnode management data structures.
324  */
325 static void
326 vntblinit(void *dummy __unused)
327 {
328 
329 	desiredvnodes = maxproc + cnt.v_page_count / 4;
330 	minvnodes = desiredvnodes / 4;
331 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
332 	mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF);
333 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
334 	mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF);
335 	TAILQ_INIT(&vnode_free_list);
336 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
337 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
338 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
339 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
340 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
341 	/*
342 	 * Initialize the filesystem syncer.
343 	 */
344 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
345 		&syncer_mask);
346 	syncer_maxdelay = syncer_mask + 1;
347 }
348 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
349 
350 
351 /*
352  * Mark a mount point as busy. Used to synchronize access and to delay
353  * unmounting. Interlock is not released on failure.
354  */
355 int
356 vfs_busy(mp, flags, interlkp, td)
357 	struct mount *mp;
358 	int flags;
359 	struct mtx *interlkp;
360 	struct thread *td;
361 {
362 	int lkflags;
363 
364 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
365 		if (flags & LK_NOWAIT)
366 			return (ENOENT);
367 		mp->mnt_kern_flag |= MNTK_MWAIT;
368 		/*
369 		 * Since all busy locks are shared except the exclusive
370 		 * lock granted when unmounting, the only place that a
371 		 * wakeup needs to be done is at the release of the
372 		 * exclusive lock at the end of dounmount.
373 		 */
374 		msleep(mp, interlkp, PVFS, "vfs_busy", 0);
375 		return (ENOENT);
376 	}
377 	lkflags = LK_SHARED | LK_NOPAUSE;
378 	if (interlkp)
379 		lkflags |= LK_INTERLOCK;
380 	if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td))
381 		panic("vfs_busy: unexpected lock failure");
382 	return (0);
383 }
384 
385 /*
386  * Free a busy filesystem.
387  */
388 void
389 vfs_unbusy(mp, td)
390 	struct mount *mp;
391 	struct thread *td;
392 {
393 
394 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
395 }
396 
397 /*
398  * Lookup a mount point by filesystem identifier.
399  */
400 struct mount *
401 vfs_getvfs(fsid)
402 	fsid_t *fsid;
403 {
404 	register struct mount *mp;
405 
406 	mtx_lock(&mountlist_mtx);
407 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
408 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
409 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
410 			mtx_unlock(&mountlist_mtx);
411 			return (mp);
412 	    }
413 	}
414 	mtx_unlock(&mountlist_mtx);
415 	return ((struct mount *) 0);
416 }
417 
418 /*
419  * Get a new unique fsid.  Try to make its val[0] unique, since this value
420  * will be used to create fake device numbers for stat().  Also try (but
421  * not so hard) make its val[0] unique mod 2^16, since some emulators only
422  * support 16-bit device numbers.  We end up with unique val[0]'s for the
423  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
424  *
425  * Keep in mind that several mounts may be running in parallel.  Starting
426  * the search one past where the previous search terminated is both a
427  * micro-optimization and a defense against returning the same fsid to
428  * different mounts.
429  */
430 void
431 vfs_getnewfsid(mp)
432 	struct mount *mp;
433 {
434 	static u_int16_t mntid_base;
435 	fsid_t tfsid;
436 	int mtype;
437 
438 	mtx_lock(&mntid_mtx);
439 	mtype = mp->mnt_vfc->vfc_typenum;
440 	tfsid.val[1] = mtype;
441 	mtype = (mtype & 0xFF) << 24;
442 	for (;;) {
443 		tfsid.val[0] = makeudev(255,
444 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
445 		mntid_base++;
446 		if (vfs_getvfs(&tfsid) == NULL)
447 			break;
448 	}
449 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
450 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
451 	mtx_unlock(&mntid_mtx);
452 }
453 
454 /*
455  * Knob to control the precision of file timestamps:
456  *
457  *   0 = seconds only; nanoseconds zeroed.
458  *   1 = seconds and nanoseconds, accurate within 1/HZ.
459  *   2 = seconds and nanoseconds, truncated to microseconds.
460  * >=3 = seconds and nanoseconds, maximum precision.
461  */
462 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
463 
464 static int timestamp_precision = TSP_SEC;
465 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
466     &timestamp_precision, 0, "");
467 
468 /*
469  * Get a current timestamp.
470  */
471 void
472 vfs_timestamp(tsp)
473 	struct timespec *tsp;
474 {
475 	struct timeval tv;
476 
477 	switch (timestamp_precision) {
478 	case TSP_SEC:
479 		tsp->tv_sec = time_second;
480 		tsp->tv_nsec = 0;
481 		break;
482 	case TSP_HZ:
483 		getnanotime(tsp);
484 		break;
485 	case TSP_USEC:
486 		microtime(&tv);
487 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
488 		break;
489 	case TSP_NSEC:
490 	default:
491 		nanotime(tsp);
492 		break;
493 	}
494 }
495 
496 /*
497  * Set vnode attributes to VNOVAL
498  */
499 void
500 vattr_null(vap)
501 	register struct vattr *vap;
502 {
503 
504 	vap->va_type = VNON;
505 	vap->va_size = VNOVAL;
506 	vap->va_bytes = VNOVAL;
507 	vap->va_mode = VNOVAL;
508 	vap->va_nlink = VNOVAL;
509 	vap->va_uid = VNOVAL;
510 	vap->va_gid = VNOVAL;
511 	vap->va_fsid = VNOVAL;
512 	vap->va_fileid = VNOVAL;
513 	vap->va_blocksize = VNOVAL;
514 	vap->va_rdev = VNOVAL;
515 	vap->va_atime.tv_sec = VNOVAL;
516 	vap->va_atime.tv_nsec = VNOVAL;
517 	vap->va_mtime.tv_sec = VNOVAL;
518 	vap->va_mtime.tv_nsec = VNOVAL;
519 	vap->va_ctime.tv_sec = VNOVAL;
520 	vap->va_ctime.tv_nsec = VNOVAL;
521 	vap->va_birthtime.tv_sec = VNOVAL;
522 	vap->va_birthtime.tv_nsec = VNOVAL;
523 	vap->va_flags = VNOVAL;
524 	vap->va_gen = VNOVAL;
525 	vap->va_vaflags = 0;
526 }
527 
528 /*
529  * This routine is called when we have too many vnodes.  It attempts
530  * to free <count> vnodes and will potentially free vnodes that still
531  * have VM backing store (VM backing store is typically the cause
532  * of a vnode blowout so we want to do this).  Therefore, this operation
533  * is not considered cheap.
534  *
535  * A number of conditions may prevent a vnode from being reclaimed.
536  * the buffer cache may have references on the vnode, a directory
537  * vnode may still have references due to the namei cache representing
538  * underlying files, or the vnode may be in active use.   It is not
539  * desireable to reuse such vnodes.  These conditions may cause the
540  * number of vnodes to reach some minimum value regardless of what
541  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
542  */
543 static int
544 vlrureclaim(struct mount *mp, int count)
545 {
546 	struct vnode *vp;
547 	int done;
548 	int trigger;
549 	int usevnodes;
550 
551 	/*
552 	 * Calculate the trigger point, don't allow user
553 	 * screwups to blow us up.   This prevents us from
554 	 * recycling vnodes with lots of resident pages.  We
555 	 * aren't trying to free memory, we are trying to
556 	 * free vnodes.
557 	 */
558 	usevnodes = desiredvnodes;
559 	if (usevnodes <= 0)
560 		usevnodes = 1;
561 	trigger = cnt.v_page_count * 2 / usevnodes;
562 
563 	done = 0;
564 	mtx_lock(&mntvnode_mtx);
565 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
566 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
567 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
568 
569 		if (vp->v_type != VNON &&
570 		    vp->v_type != VBAD &&
571 		    VMIGHTFREE(vp) &&           /* critical path opt */
572 		    (vp->v_object == NULL || vp->v_object->resident_page_count < trigger) &&
573 		    mtx_trylock(&vp->v_interlock)
574 		) {
575 			mtx_unlock(&mntvnode_mtx);
576 			if (VMIGHTFREE(vp)) {
577 				vgonel(vp, curthread);
578 				done++;
579 			} else {
580 				mtx_unlock(&vp->v_interlock);
581 			}
582 			mtx_lock(&mntvnode_mtx);
583 		}
584 		--count;
585 	}
586 	mtx_unlock(&mntvnode_mtx);
587 	return done;
588 }
589 
590 /*
591  * Attempt to recycle vnodes in a context that is always safe to block.
592  * Calling vlrurecycle() from the bowels of filesystem code has some
593  * interesting deadlock problems.
594  */
595 static struct proc *vnlruproc;
596 static int vnlruproc_sig;
597 
598 static void
599 vnlru_proc(void)
600 {
601 	struct mount *mp, *nmp;
602 	int s;
603 	int done;
604 	struct proc *p = vnlruproc;
605 	struct thread *td = FIRST_THREAD_IN_PROC(p);	/* XXXKSE */
606 
607 	mtx_lock(&Giant);
608 
609 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
610 	    SHUTDOWN_PRI_FIRST);
611 
612 	s = splbio();
613 	for (;;) {
614 		kthread_suspend_check(p);
615 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
616 			vnlruproc_sig = 0;
617 			tsleep(vnlruproc, PVFS, "vlruwt", 0);
618 			continue;
619 		}
620 		done = 0;
621 		mtx_lock(&mountlist_mtx);
622 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
623 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
624 				nmp = TAILQ_NEXT(mp, mnt_list);
625 				continue;
626 			}
627 			done += vlrureclaim(mp, 10);
628 			mtx_lock(&mountlist_mtx);
629 			nmp = TAILQ_NEXT(mp, mnt_list);
630 			vfs_unbusy(mp, td);
631 		}
632 		mtx_unlock(&mountlist_mtx);
633 		if (done == 0) {
634 #if 0
635 			/* These messages are temporary debugging aids */
636 			if (vnlru_nowhere < 5)
637 				printf("vnlru process getting nowhere..\n");
638 			else if (vnlru_nowhere == 5)
639 				printf("vnlru process messages stopped.\n");
640 #endif
641 			vnlru_nowhere++;
642 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
643 		}
644 	}
645 	splx(s);
646 }
647 
648 static struct kproc_desc vnlru_kp = {
649 	"vnlru",
650 	vnlru_proc,
651 	&vnlruproc
652 };
653 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
654 
655 
656 /*
657  * Routines having to do with the management of the vnode table.
658  */
659 
660 /*
661  * Return the next vnode from the free list.
662  */
663 int
664 getnewvnode(tag, mp, vops, vpp)
665 	enum vtagtype tag;
666 	struct mount *mp;
667 	vop_t **vops;
668 	struct vnode **vpp;
669 {
670 	int s;
671 	struct thread *td = curthread;	/* XXX */
672 	struct vnode *vp = NULL;
673 	struct mount *vnmp;
674 	vm_object_t object;
675 
676 	s = splbio();
677 	/*
678 	 * Try to reuse vnodes if we hit the max.  This situation only
679 	 * occurs in certain large-memory (2G+) situations.  We cannot
680 	 * attempt to directly reclaim vnodes due to nasty recursion
681 	 * problems.
682 	 */
683 	if (vnlruproc_sig == 0 && numvnodes - freevnodes > desiredvnodes) {
684 		vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
685 		wakeup(vnlruproc);
686 	}
687 
688 	/*
689 	 * Attempt to reuse a vnode already on the free list, allocating
690 	 * a new vnode if we can't find one or if we have not reached a
691 	 * good minimum for good LRU performance.
692 	 */
693 
694 	mtx_lock(&vnode_free_list_mtx);
695 
696 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
697 		int count;
698 
699 		for (count = 0; count < freevnodes; count++) {
700 			vp = TAILQ_FIRST(&vnode_free_list);
701 			if (vp == NULL || vp->v_usecount)
702 				panic("getnewvnode: free vnode isn't");
703 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
704 
705 			/* Don't recycle if we can't get the interlock */
706 			if (!mtx_trylock(&vp->v_interlock)) {
707 				vp = NULL;
708 				continue;
709 			}
710 
711 			/* We should be able to immediately acquire this */
712 			if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td) != 0)
713 				continue;
714 			/*
715 			 * Don't recycle if we still have cached pages.
716 			 */
717 			if (VOP_GETVOBJECT(vp, &object) == 0 &&
718 			     (object->resident_page_count ||
719 			      object->ref_count)) {
720 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
721 						    v_freelist);
722 				VOP_UNLOCK(vp, 0, td);
723 				vp = NULL;
724 				continue;
725 			}
726 			if (LIST_FIRST(&vp->v_cache_src)) {
727 				/*
728 				 * note: nameileafonly sysctl is temporary,
729 				 * for debugging only, and will eventually be
730 				 * removed.
731 				 */
732 				if (nameileafonly > 0) {
733 					/*
734 					 * Do not reuse namei-cached directory
735 					 * vnodes that have cached
736 					 * subdirectories.
737 					 */
738 					if (cache_leaf_test(vp) < 0) {
739 						VOP_UNLOCK(vp, 0, td);
740 						TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
741 						vp = NULL;
742 						continue;
743 					}
744 				} else if (nameileafonly < 0 ||
745 					    vmiodirenable == 0) {
746 					/*
747 					 * Do not reuse namei-cached directory
748 					 * vnodes if nameileafonly is -1 or
749 					 * if VMIO backing for directories is
750 					 * turned off (otherwise we reuse them
751 					 * too quickly).
752 					 */
753 					VOP_UNLOCK(vp, 0, td);
754 					TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
755 					vp = NULL;
756 					continue;
757 				}
758 			}
759 			/*
760 			 * Skip over it if its filesystem is being suspended.
761 			 */
762 			if (vn_start_write(vp, &vnmp, V_NOWAIT) == 0)
763 				break;
764 			VOP_UNLOCK(vp, 0, td);
765 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
766 			vp = NULL;
767 		}
768 	}
769 	if (vp) {
770 		vp->v_flag |= VDOOMED;
771 		vp->v_flag &= ~VFREE;
772 		freevnodes--;
773 		mtx_unlock(&vnode_free_list_mtx);
774 		cache_purge(vp);
775 		if (vp->v_type != VBAD) {
776 			VOP_UNLOCK(vp, 0, td);
777 			vgone(vp);
778 		} else {
779 			VOP_UNLOCK(vp, 0, td);
780 		}
781 		vn_finished_write(vnmp);
782 
783 #ifdef INVARIANTS
784 		{
785 			int s;
786 
787 			if (vp->v_data)
788 				panic("cleaned vnode isn't");
789 			s = splbio();
790 			if (vp->v_numoutput)
791 				panic("Clean vnode has pending I/O's");
792 			splx(s);
793 			if (vp->v_writecount != 0)
794 				panic("Non-zero write count");
795 		}
796 #endif
797 		if (vp->v_pollinfo) {
798 			mtx_destroy(&vp->v_pollinfo->vpi_lock);
799 			uma_zfree(vnodepoll_zone, vp->v_pollinfo);
800 		}
801 		vp->v_pollinfo = NULL;
802 		vp->v_flag = 0;
803 		vp->v_lastw = 0;
804 		vp->v_lasta = 0;
805 		vp->v_cstart = 0;
806 		vp->v_clen = 0;
807 		vp->v_socket = 0;
808 		KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL"));
809 		KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL"));
810 	} else {
811 		mtx_unlock(&vnode_free_list_mtx);
812 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
813 		bzero((char *) vp, sizeof *vp);
814 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
815 		vp->v_dd = vp;
816 		cache_purge(vp);
817 		LIST_INIT(&vp->v_cache_src);
818 		TAILQ_INIT(&vp->v_cache_dst);
819 		numvnodes++;
820 	}
821 
822 	TAILQ_INIT(&vp->v_cleanblkhd);
823 	TAILQ_INIT(&vp->v_dirtyblkhd);
824 	vp->v_type = VNON;
825 	vp->v_tag = tag;
826 	vp->v_op = vops;
827 	lockinit(&vp->v_lock, PVFS, "vnlock", VLKTIMEOUT, LK_NOPAUSE);
828 	insmntque(vp, mp);
829 	*vpp = vp;
830 	vp->v_usecount = 1;
831 	vp->v_data = 0;
832 
833 	splx(s);
834 
835 #if 0
836 	vnodeallocs++;
837 	if (vnodeallocs % vnoderecycleperiod == 0 &&
838 	    freevnodes < vnoderecycleminfreevn &&
839 	    vnoderecyclemintotalvn < numvnodes) {
840 		/* Recycle vnodes. */
841 		cache_purgeleafdirs(vnoderecyclenumber);
842 	}
843 #endif
844 
845 	return (0);
846 }
847 
848 /*
849  * Move a vnode from one mount queue to another.
850  */
851 static void
852 insmntque(vp, mp)
853 	register struct vnode *vp;
854 	register struct mount *mp;
855 {
856 
857 	mtx_lock(&mntvnode_mtx);
858 	/*
859 	 * Delete from old mount point vnode list, if on one.
860 	 */
861 	if (vp->v_mount != NULL)
862 		TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
863 	/*
864 	 * Insert into list of vnodes for the new mount point, if available.
865 	 */
866 	if ((vp->v_mount = mp) == NULL) {
867 		mtx_unlock(&mntvnode_mtx);
868 		return;
869 	}
870 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
871 	mtx_unlock(&mntvnode_mtx);
872 }
873 
874 /*
875  * Update outstanding I/O count and do wakeup if requested.
876  */
877 void
878 vwakeup(bp)
879 	register struct buf *bp;
880 {
881 	register struct vnode *vp;
882 
883 	bp->b_flags &= ~B_WRITEINPROG;
884 	if ((vp = bp->b_vp)) {
885 		vp->v_numoutput--;
886 		if (vp->v_numoutput < 0)
887 			panic("vwakeup: neg numoutput");
888 		if ((vp->v_numoutput == 0) && (vp->v_flag & VBWAIT)) {
889 			vp->v_flag &= ~VBWAIT;
890 			wakeup(&vp->v_numoutput);
891 		}
892 	}
893 }
894 
895 /*
896  * Flush out and invalidate all buffers associated with a vnode.
897  * Called with the underlying object locked.
898  */
899 int
900 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo)
901 	register struct vnode *vp;
902 	int flags;
903 	struct ucred *cred;
904 	struct thread *td;
905 	int slpflag, slptimeo;
906 {
907 	register struct buf *bp;
908 	struct buf *nbp, *blist;
909 	int s, error;
910 	vm_object_t object;
911 
912 	GIANT_REQUIRED;
913 
914 	if (flags & V_SAVE) {
915 		s = splbio();
916 		while (vp->v_numoutput) {
917 			vp->v_flag |= VBWAIT;
918 			error = tsleep(&vp->v_numoutput,
919 			    slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo);
920 			if (error) {
921 				splx(s);
922 				return (error);
923 			}
924 		}
925 		if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
926 			splx(s);
927 			if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0)
928 				return (error);
929 			s = splbio();
930 			if (vp->v_numoutput > 0 ||
931 			    !TAILQ_EMPTY(&vp->v_dirtyblkhd))
932 				panic("vinvalbuf: dirty bufs");
933 		}
934 		splx(s);
935 	}
936 	s = splbio();
937 	for (;;) {
938 		blist = TAILQ_FIRST(&vp->v_cleanblkhd);
939 		if (!blist)
940 			blist = TAILQ_FIRST(&vp->v_dirtyblkhd);
941 		if (!blist)
942 			break;
943 
944 		for (bp = blist; bp; bp = nbp) {
945 			nbp = TAILQ_NEXT(bp, b_vnbufs);
946 			if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
947 				error = BUF_TIMELOCK(bp,
948 				    LK_EXCLUSIVE | LK_SLEEPFAIL,
949 				    "vinvalbuf", slpflag, slptimeo);
950 				if (error == ENOLCK)
951 					break;
952 				splx(s);
953 				return (error);
954 			}
955 			/*
956 			 * XXX Since there are no node locks for NFS, I
957 			 * believe there is a slight chance that a delayed
958 			 * write will occur while sleeping just above, so
959 			 * check for it.  Note that vfs_bio_awrite expects
960 			 * buffers to reside on a queue, while BUF_WRITE and
961 			 * brelse do not.
962 			 */
963 			if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
964 				(flags & V_SAVE)) {
965 
966 				if (bp->b_vp == vp) {
967 					if (bp->b_flags & B_CLUSTEROK) {
968 						BUF_UNLOCK(bp);
969 						vfs_bio_awrite(bp);
970 					} else {
971 						bremfree(bp);
972 						bp->b_flags |= B_ASYNC;
973 						BUF_WRITE(bp);
974 					}
975 				} else {
976 					bremfree(bp);
977 					(void) BUF_WRITE(bp);
978 				}
979 				break;
980 			}
981 			bremfree(bp);
982 			bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
983 			bp->b_flags &= ~B_ASYNC;
984 			brelse(bp);
985 		}
986 	}
987 
988 	/*
989 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
990 	 * have write I/O in-progress but if there is a VM object then the
991 	 * VM object can also have read-I/O in-progress.
992 	 */
993 	do {
994 		while (vp->v_numoutput > 0) {
995 			vp->v_flag |= VBWAIT;
996 			tsleep(&vp->v_numoutput, PVM, "vnvlbv", 0);
997 		}
998 		if (VOP_GETVOBJECT(vp, &object) == 0) {
999 			while (object->paging_in_progress)
1000 			vm_object_pip_sleep(object, "vnvlbx");
1001 		}
1002 	} while (vp->v_numoutput > 0);
1003 
1004 	splx(s);
1005 
1006 	/*
1007 	 * Destroy the copy in the VM cache, too.
1008 	 */
1009 	mtx_lock(&vp->v_interlock);
1010 	if (VOP_GETVOBJECT(vp, &object) == 0) {
1011 		vm_object_page_remove(object, 0, 0,
1012 			(flags & V_SAVE) ? TRUE : FALSE);
1013 	}
1014 	mtx_unlock(&vp->v_interlock);
1015 
1016 	if (!TAILQ_EMPTY(&vp->v_dirtyblkhd) || !TAILQ_EMPTY(&vp->v_cleanblkhd))
1017 		panic("vinvalbuf: flush failed");
1018 	return (0);
1019 }
1020 
1021 /*
1022  * Truncate a file's buffer and pages to a specified length.  This
1023  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1024  * sync activity.
1025  */
1026 int
1027 vtruncbuf(vp, cred, td, length, blksize)
1028 	register struct vnode *vp;
1029 	struct ucred *cred;
1030 	struct thread *td;
1031 	off_t length;
1032 	int blksize;
1033 {
1034 	register struct buf *bp;
1035 	struct buf *nbp;
1036 	int s, anyfreed;
1037 	int trunclbn;
1038 
1039 	/*
1040 	 * Round up to the *next* lbn.
1041 	 */
1042 	trunclbn = (length + blksize - 1) / blksize;
1043 
1044 	s = splbio();
1045 restart:
1046 	anyfreed = 1;
1047 	for (;anyfreed;) {
1048 		anyfreed = 0;
1049 		for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) {
1050 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1051 			if (bp->b_lblkno >= trunclbn) {
1052 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1053 					BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL);
1054 					goto restart;
1055 				} else {
1056 					bremfree(bp);
1057 					bp->b_flags |= (B_INVAL | B_RELBUF);
1058 					bp->b_flags &= ~B_ASYNC;
1059 					brelse(bp);
1060 					anyfreed = 1;
1061 				}
1062 				if (nbp &&
1063 				    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1064 				    (nbp->b_vp != vp) ||
1065 				    (nbp->b_flags & B_DELWRI))) {
1066 					goto restart;
1067 				}
1068 			}
1069 		}
1070 
1071 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1072 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1073 			if (bp->b_lblkno >= trunclbn) {
1074 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1075 					BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL);
1076 					goto restart;
1077 				} else {
1078 					bremfree(bp);
1079 					bp->b_flags |= (B_INVAL | B_RELBUF);
1080 					bp->b_flags &= ~B_ASYNC;
1081 					brelse(bp);
1082 					anyfreed = 1;
1083 				}
1084 				if (nbp &&
1085 				    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1086 				    (nbp->b_vp != vp) ||
1087 				    (nbp->b_flags & B_DELWRI) == 0)) {
1088 					goto restart;
1089 				}
1090 			}
1091 		}
1092 	}
1093 
1094 	if (length > 0) {
1095 restartsync:
1096 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1097 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1098 			if ((bp->b_flags & B_DELWRI) && (bp->b_lblkno < 0)) {
1099 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1100 					BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL);
1101 					goto restart;
1102 				} else {
1103 					bremfree(bp);
1104 					if (bp->b_vp == vp) {
1105 						bp->b_flags |= B_ASYNC;
1106 					} else {
1107 						bp->b_flags &= ~B_ASYNC;
1108 					}
1109 					BUF_WRITE(bp);
1110 				}
1111 				goto restartsync;
1112 			}
1113 
1114 		}
1115 	}
1116 
1117 	while (vp->v_numoutput > 0) {
1118 		vp->v_flag |= VBWAIT;
1119 		tsleep(&vp->v_numoutput, PVM, "vbtrunc", 0);
1120 	}
1121 
1122 	splx(s);
1123 
1124 	vnode_pager_setsize(vp, length);
1125 
1126 	return (0);
1127 }
1128 
1129 /*
1130  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1131  * 		 a vnode.
1132  *
1133  *	NOTE: We have to deal with the special case of a background bitmap
1134  *	buffer, a situation where two buffers will have the same logical
1135  *	block offset.  We want (1) only the foreground buffer to be accessed
1136  *	in a lookup and (2) must differentiate between the foreground and
1137  *	background buffer in the splay tree algorithm because the splay
1138  *	tree cannot normally handle multiple entities with the same 'index'.
1139  *	We accomplish this by adding differentiating flags to the splay tree's
1140  *	numerical domain.
1141  */
1142 static
1143 struct buf *
1144 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1145 {
1146 	struct buf dummy;
1147 	struct buf *lefttreemax, *righttreemin, *y;
1148 
1149 	if (root == NULL)
1150 		return (NULL);
1151 	lefttreemax = righttreemin = &dummy;
1152 	for (;;) {
1153 		if (lblkno < root->b_lblkno ||
1154 		    (lblkno == root->b_lblkno &&
1155 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1156 			if ((y = root->b_left) == NULL)
1157 				break;
1158 			if (lblkno < y->b_lblkno) {
1159 				/* Rotate right. */
1160 				root->b_left = y->b_right;
1161 				y->b_right = root;
1162 				root = y;
1163 				if ((y = root->b_left) == NULL)
1164 					break;
1165 			}
1166 			/* Link into the new root's right tree. */
1167 			righttreemin->b_left = root;
1168 			righttreemin = root;
1169 		} else if (lblkno > root->b_lblkno ||
1170 		    (lblkno == root->b_lblkno &&
1171 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1172 			if ((y = root->b_right) == NULL)
1173 				break;
1174 			if (lblkno > y->b_lblkno) {
1175 				/* Rotate left. */
1176 				root->b_right = y->b_left;
1177 				y->b_left = root;
1178 				root = y;
1179 				if ((y = root->b_right) == NULL)
1180 					break;
1181 			}
1182 			/* Link into the new root's left tree. */
1183 			lefttreemax->b_right = root;
1184 			lefttreemax = root;
1185 		} else {
1186 			break;
1187 		}
1188 		root = y;
1189 	}
1190 	/* Assemble the new root. */
1191 	lefttreemax->b_right = root->b_left;
1192 	righttreemin->b_left = root->b_right;
1193 	root->b_left = dummy.b_right;
1194 	root->b_right = dummy.b_left;
1195 	return (root);
1196 }
1197 
1198 static
1199 void
1200 buf_vlist_remove(struct buf *bp)
1201 {
1202 	struct vnode *vp = bp->b_vp;
1203 	struct buf *root;
1204 
1205 	if (bp->b_xflags & BX_VNDIRTY) {
1206 		if (bp != vp->v_dirtyblkroot) {
1207 			root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1208 			KASSERT(root == bp, ("splay lookup failed during dirty remove"));
1209 		}
1210 		if (bp->b_left == NULL) {
1211 			root = bp->b_right;
1212 		} else {
1213 			root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1214 			root->b_right = bp->b_right;
1215 		}
1216 		vp->v_dirtyblkroot = root;
1217 		TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs);
1218 	} else {
1219 		/* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */
1220 		if (bp != vp->v_cleanblkroot) {
1221 			root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1222 			KASSERT(root == bp, ("splay lookup failed during clean remove"));
1223 		}
1224 		if (bp->b_left == NULL) {
1225 			root = bp->b_right;
1226 		} else {
1227 			root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1228 			root->b_right = bp->b_right;
1229 		}
1230 		vp->v_cleanblkroot = root;
1231 		TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs);
1232 	}
1233 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1234 }
1235 
1236 /*
1237  * Add the buffer to the sorted clean or dirty block list using a
1238  * splay tree algorithm.
1239  *
1240  * NOTE: xflags is passed as a constant, optimizing this inline function!
1241  */
1242 static
1243 void
1244 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags)
1245 {
1246 	struct buf *root;
1247 
1248 	bp->b_xflags |= xflags;
1249 	if (xflags & BX_VNDIRTY) {
1250 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1251 		if (root == NULL) {
1252 			bp->b_left = NULL;
1253 			bp->b_right = NULL;
1254 			TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs);
1255 		} else if (bp->b_lblkno < root->b_lblkno ||
1256 		    (bp->b_lblkno == root->b_lblkno &&
1257 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1258 			bp->b_left = root->b_left;
1259 			bp->b_right = root;
1260 			root->b_left = NULL;
1261 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1262 		} else {
1263 			bp->b_right = root->b_right;
1264 			bp->b_left = root;
1265 			root->b_right = NULL;
1266 			TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd,
1267 			    root, bp, b_vnbufs);
1268 		}
1269 		vp->v_dirtyblkroot = bp;
1270 	} else {
1271 		/* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */
1272 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1273 		if (root == NULL) {
1274 			bp->b_left = NULL;
1275 			bp->b_right = NULL;
1276 			TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs);
1277 		} else if (bp->b_lblkno < root->b_lblkno ||
1278 		    (bp->b_lblkno == root->b_lblkno &&
1279 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1280 			bp->b_left = root->b_left;
1281 			bp->b_right = root;
1282 			root->b_left = NULL;
1283 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1284 		} else {
1285 			bp->b_right = root->b_right;
1286 			bp->b_left = root;
1287 			root->b_right = NULL;
1288 			TAILQ_INSERT_AFTER(&vp->v_cleanblkhd,
1289 			    root, bp, b_vnbufs);
1290 		}
1291 		vp->v_cleanblkroot = bp;
1292 	}
1293 }
1294 
1295 #ifndef USE_BUFHASH
1296 
1297 /*
1298  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1299  * shadow buffers used in background bitmap writes.
1300  *
1301  * This code isn't quite efficient as it could be because we are maintaining
1302  * two sorted lists and do not know which list the block resides in.
1303  */
1304 struct buf *
1305 gbincore(struct vnode *vp, daddr_t lblkno)
1306 {
1307 	struct buf *bp;
1308 
1309 	GIANT_REQUIRED;
1310 
1311 	bp = vp->v_cleanblkroot = buf_splay(lblkno, 0, vp->v_cleanblkroot);
1312 	if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1313 		return(bp);
1314 	bp = vp->v_dirtyblkroot = buf_splay(lblkno, 0, vp->v_dirtyblkroot);
1315 	if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1316 		return(bp);
1317 	return(NULL);
1318 }
1319 
1320 #endif
1321 
1322 /*
1323  * Associate a buffer with a vnode.
1324  */
1325 void
1326 bgetvp(vp, bp)
1327 	register struct vnode *vp;
1328 	register struct buf *bp;
1329 {
1330 	int s;
1331 
1332 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1333 
1334 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1335 	    ("bgetvp: bp already attached! %p", bp));
1336 
1337 	vhold(vp);
1338 	bp->b_vp = vp;
1339 	bp->b_dev = vn_todev(vp);
1340 	/*
1341 	 * Insert onto list for new vnode.
1342 	 */
1343 	s = splbio();
1344 	buf_vlist_add(bp, vp, BX_VNCLEAN);
1345 	splx(s);
1346 }
1347 
1348 /*
1349  * Disassociate a buffer from a vnode.
1350  */
1351 void
1352 brelvp(bp)
1353 	register struct buf *bp;
1354 {
1355 	struct vnode *vp;
1356 	int s;
1357 
1358 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1359 
1360 	/*
1361 	 * Delete from old vnode list, if on one.
1362 	 */
1363 	vp = bp->b_vp;
1364 	s = splbio();
1365 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1366 		buf_vlist_remove(bp);
1367 	if ((vp->v_flag & VONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1368 		vp->v_flag &= ~VONWORKLST;
1369 		LIST_REMOVE(vp, v_synclist);
1370 	}
1371 	splx(s);
1372 	bp->b_vp = (struct vnode *) 0;
1373 	vdrop(vp);
1374 	if (bp->b_object)
1375 		bp->b_object = NULL;
1376 }
1377 
1378 /*
1379  * Add an item to the syncer work queue.
1380  */
1381 static void
1382 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
1383 {
1384 	int s, slot;
1385 
1386 	s = splbio();
1387 
1388 	if (vp->v_flag & VONWORKLST) {
1389 		LIST_REMOVE(vp, v_synclist);
1390 	}
1391 
1392 	if (delay > syncer_maxdelay - 2)
1393 		delay = syncer_maxdelay - 2;
1394 	slot = (syncer_delayno + delay) & syncer_mask;
1395 
1396 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
1397 	vp->v_flag |= VONWORKLST;
1398 	splx(s);
1399 }
1400 
1401 struct  proc *updateproc;
1402 static void sched_sync(void);
1403 static struct kproc_desc up_kp = {
1404 	"syncer",
1405 	sched_sync,
1406 	&updateproc
1407 };
1408 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1409 
1410 /*
1411  * System filesystem synchronizer daemon.
1412  */
1413 void
1414 sched_sync(void)
1415 {
1416 	struct synclist *slp;
1417 	struct vnode *vp;
1418 	struct mount *mp;
1419 	long starttime;
1420 	int s;
1421 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);  /* XXXKSE */
1422 
1423 	mtx_lock(&Giant);
1424 
1425 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc,
1426 	    SHUTDOWN_PRI_LAST);
1427 
1428 	for (;;) {
1429 		kthread_suspend_check(td->td_proc);
1430 
1431 		starttime = time_second;
1432 
1433 		/*
1434 		 * Push files whose dirty time has expired.  Be careful
1435 		 * of interrupt race on slp queue.
1436 		 */
1437 		s = splbio();
1438 		slp = &syncer_workitem_pending[syncer_delayno];
1439 		syncer_delayno += 1;
1440 		if (syncer_delayno == syncer_maxdelay)
1441 			syncer_delayno = 0;
1442 		splx(s);
1443 
1444 		while ((vp = LIST_FIRST(slp)) != NULL) {
1445 			if (VOP_ISLOCKED(vp, NULL) == 0 &&
1446 			    vn_start_write(vp, &mp, V_NOWAIT) == 0) {
1447 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1448 				(void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td);
1449 				VOP_UNLOCK(vp, 0, td);
1450 				vn_finished_write(mp);
1451 			}
1452 			s = splbio();
1453 			if (LIST_FIRST(slp) == vp) {
1454 				/*
1455 				 * Note: v_tag VT_VFS vps can remain on the
1456 				 * worklist too with no dirty blocks, but
1457 				 * since sync_fsync() moves it to a different
1458 				 * slot we are safe.
1459 				 */
1460 				if (TAILQ_EMPTY(&vp->v_dirtyblkhd) &&
1461 				    !vn_isdisk(vp, NULL))
1462 					panic("sched_sync: fsync failed vp %p tag %d", vp, vp->v_tag);
1463 				/*
1464 				 * Put us back on the worklist.  The worklist
1465 				 * routine will remove us from our current
1466 				 * position and then add us back in at a later
1467 				 * position.
1468 				 */
1469 				vn_syncer_add_to_worklist(vp, syncdelay);
1470 			}
1471 			splx(s);
1472 		}
1473 
1474 		/*
1475 		 * Do soft update processing.
1476 		 */
1477 		if (softdep_process_worklist_hook != NULL)
1478 			(*softdep_process_worklist_hook)(NULL);
1479 
1480 		/*
1481 		 * The variable rushjob allows the kernel to speed up the
1482 		 * processing of the filesystem syncer process. A rushjob
1483 		 * value of N tells the filesystem syncer to process the next
1484 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1485 		 * is used by the soft update code to speed up the filesystem
1486 		 * syncer process when the incore state is getting so far
1487 		 * ahead of the disk that the kernel memory pool is being
1488 		 * threatened with exhaustion.
1489 		 */
1490 		if (rushjob > 0) {
1491 			rushjob -= 1;
1492 			continue;
1493 		}
1494 		/*
1495 		 * If it has taken us less than a second to process the
1496 		 * current work, then wait. Otherwise start right over
1497 		 * again. We can still lose time if any single round
1498 		 * takes more than two seconds, but it does not really
1499 		 * matter as we are just trying to generally pace the
1500 		 * filesystem activity.
1501 		 */
1502 		if (time_second == starttime)
1503 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1504 	}
1505 }
1506 
1507 /*
1508  * Request the syncer daemon to speed up its work.
1509  * We never push it to speed up more than half of its
1510  * normal turn time, otherwise it could take over the cpu.
1511  * XXXKSE  only one update?
1512  */
1513 int
1514 speedup_syncer()
1515 {
1516 
1517 	mtx_lock_spin(&sched_lock);
1518 	if (FIRST_THREAD_IN_PROC(updateproc)->td_wchan == &lbolt) /* XXXKSE */
1519 		setrunnable(FIRST_THREAD_IN_PROC(updateproc));
1520 	mtx_unlock_spin(&sched_lock);
1521 	if (rushjob < syncdelay / 2) {
1522 		rushjob += 1;
1523 		stat_rush_requests += 1;
1524 		return (1);
1525 	}
1526 	return(0);
1527 }
1528 
1529 /*
1530  * Associate a p-buffer with a vnode.
1531  *
1532  * Also sets B_PAGING flag to indicate that vnode is not fully associated
1533  * with the buffer.  i.e. the bp has not been linked into the vnode or
1534  * ref-counted.
1535  */
1536 void
1537 pbgetvp(vp, bp)
1538 	register struct vnode *vp;
1539 	register struct buf *bp;
1540 {
1541 
1542 	KASSERT(bp->b_vp == NULL, ("pbgetvp: not free"));
1543 
1544 	bp->b_vp = vp;
1545 	bp->b_flags |= B_PAGING;
1546 	bp->b_dev = vn_todev(vp);
1547 }
1548 
1549 /*
1550  * Disassociate a p-buffer from a vnode.
1551  */
1552 void
1553 pbrelvp(bp)
1554 	register struct buf *bp;
1555 {
1556 
1557 	KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL"));
1558 
1559 	/* XXX REMOVE ME */
1560 	if (TAILQ_NEXT(bp, b_vnbufs) != NULL) {
1561 		panic(
1562 		    "relpbuf(): b_vp was probably reassignbuf()d %p %x",
1563 		    bp,
1564 		    (int)bp->b_flags
1565 		);
1566 	}
1567 	bp->b_vp = (struct vnode *) 0;
1568 	bp->b_flags &= ~B_PAGING;
1569 }
1570 
1571 /*
1572  * Reassign a buffer from one vnode to another.
1573  * Used to assign file specific control information
1574  * (indirect blocks) to the vnode to which they belong.
1575  */
1576 void
1577 reassignbuf(bp, newvp)
1578 	register struct buf *bp;
1579 	register struct vnode *newvp;
1580 {
1581 	int delay;
1582 	int s;
1583 
1584 	if (newvp == NULL) {
1585 		printf("reassignbuf: NULL");
1586 		return;
1587 	}
1588 	++reassignbufcalls;
1589 
1590 	/*
1591 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1592 	 * is not fully linked in.
1593 	 */
1594 	if (bp->b_flags & B_PAGING)
1595 		panic("cannot reassign paging buffer");
1596 
1597 	s = splbio();
1598 	/*
1599 	 * Delete from old vnode list, if on one.
1600 	 */
1601 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) {
1602 		buf_vlist_remove(bp);
1603 		if (bp->b_vp != newvp) {
1604 			vdrop(bp->b_vp);
1605 			bp->b_vp = NULL;	/* for clarification */
1606 		}
1607 	}
1608 	/*
1609 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1610 	 * of clean buffers.
1611 	 */
1612 	if (bp->b_flags & B_DELWRI) {
1613 		if ((newvp->v_flag & VONWORKLST) == 0) {
1614 			switch (newvp->v_type) {
1615 			case VDIR:
1616 				delay = dirdelay;
1617 				break;
1618 			case VCHR:
1619 				if (newvp->v_rdev->si_mountpoint != NULL) {
1620 					delay = metadelay;
1621 					break;
1622 				}
1623 				/* fall through */
1624 			default:
1625 				delay = filedelay;
1626 			}
1627 			vn_syncer_add_to_worklist(newvp, delay);
1628 		}
1629 		buf_vlist_add(bp, newvp, BX_VNDIRTY);
1630 	} else {
1631 		buf_vlist_add(bp, newvp, BX_VNCLEAN);
1632 
1633 		if ((newvp->v_flag & VONWORKLST) &&
1634 		    TAILQ_EMPTY(&newvp->v_dirtyblkhd)) {
1635 			newvp->v_flag &= ~VONWORKLST;
1636 			LIST_REMOVE(newvp, v_synclist);
1637 		}
1638 	}
1639 	if (bp->b_vp != newvp) {
1640 		bp->b_vp = newvp;
1641 		vhold(bp->b_vp);
1642 	}
1643 	splx(s);
1644 }
1645 
1646 /*
1647  * Create a vnode for a device.
1648  * Used for mounting the root filesystem.
1649  */
1650 int
1651 bdevvp(dev, vpp)
1652 	dev_t dev;
1653 	struct vnode **vpp;
1654 {
1655 	register struct vnode *vp;
1656 	struct vnode *nvp;
1657 	int error;
1658 
1659 	if (dev == NODEV) {
1660 		*vpp = NULLVP;
1661 		return (ENXIO);
1662 	}
1663 	if (vfinddev(dev, VCHR, vpp))
1664 		return (0);
1665 	error = getnewvnode(VT_NON, (struct mount *)0, spec_vnodeop_p, &nvp);
1666 	if (error) {
1667 		*vpp = NULLVP;
1668 		return (error);
1669 	}
1670 	vp = nvp;
1671 	vp->v_type = VCHR;
1672 	addalias(vp, dev);
1673 	*vpp = vp;
1674 	return (0);
1675 }
1676 
1677 /*
1678  * Add vnode to the alias list hung off the dev_t.
1679  *
1680  * The reason for this gunk is that multiple vnodes can reference
1681  * the same physical device, so checking vp->v_usecount to see
1682  * how many users there are is inadequate; the v_usecount for
1683  * the vnodes need to be accumulated.  vcount() does that.
1684  */
1685 struct vnode *
1686 addaliasu(nvp, nvp_rdev)
1687 	struct vnode *nvp;
1688 	udev_t nvp_rdev;
1689 {
1690 	struct vnode *ovp;
1691 	vop_t **ops;
1692 	dev_t dev;
1693 
1694 	if (nvp->v_type == VBLK)
1695 		return (nvp);
1696 	if (nvp->v_type != VCHR)
1697 		panic("addaliasu on non-special vnode");
1698 	dev = udev2dev(nvp_rdev, 0);
1699 	/*
1700 	 * Check to see if we have a bdevvp vnode with no associated
1701 	 * filesystem. If so, we want to associate the filesystem of
1702 	 * the new newly instigated vnode with the bdevvp vnode and
1703 	 * discard the newly created vnode rather than leaving the
1704 	 * bdevvp vnode lying around with no associated filesystem.
1705 	 */
1706 	if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) {
1707 		addalias(nvp, dev);
1708 		return (nvp);
1709 	}
1710 	/*
1711 	 * Discard unneeded vnode, but save its node specific data.
1712 	 * Note that if there is a lock, it is carried over in the
1713 	 * node specific data to the replacement vnode.
1714 	 */
1715 	vref(ovp);
1716 	ovp->v_data = nvp->v_data;
1717 	ovp->v_tag = nvp->v_tag;
1718 	nvp->v_data = NULL;
1719 	lockinit(&ovp->v_lock, PVFS, nvp->v_lock.lk_wmesg,
1720 	    nvp->v_lock.lk_timo, nvp->v_lock.lk_flags & LK_EXTFLG_MASK);
1721 	if (nvp->v_vnlock)
1722 		ovp->v_vnlock = &ovp->v_lock;
1723 	ops = ovp->v_op;
1724 	ovp->v_op = nvp->v_op;
1725 	if (VOP_ISLOCKED(nvp, curthread)) {
1726 		VOP_UNLOCK(nvp, 0, curthread);
1727 		vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread);
1728 	}
1729 	nvp->v_op = ops;
1730 	insmntque(ovp, nvp->v_mount);
1731 	vrele(nvp);
1732 	vgone(nvp);
1733 	return (ovp);
1734 }
1735 
1736 /* This is a local helper function that do the same as addaliasu, but for a
1737  * dev_t instead of an udev_t. */
1738 static void
1739 addalias(nvp, dev)
1740 	struct vnode *nvp;
1741 	dev_t dev;
1742 {
1743 
1744 	KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode"));
1745 	nvp->v_rdev = dev;
1746 	mtx_lock(&spechash_mtx);
1747 	SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext);
1748 	mtx_unlock(&spechash_mtx);
1749 }
1750 
1751 /*
1752  * Grab a particular vnode from the free list, increment its
1753  * reference count and lock it. The vnode lock bit is set if the
1754  * vnode is being eliminated in vgone. The process is awakened
1755  * when the transition is completed, and an error returned to
1756  * indicate that the vnode is no longer usable (possibly having
1757  * been changed to a new filesystem type).
1758  */
1759 int
1760 vget(vp, flags, td)
1761 	register struct vnode *vp;
1762 	int flags;
1763 	struct thread *td;
1764 {
1765 	int error;
1766 
1767 	/*
1768 	 * If the vnode is in the process of being cleaned out for
1769 	 * another use, we wait for the cleaning to finish and then
1770 	 * return failure. Cleaning is determined by checking that
1771 	 * the VXLOCK flag is set.
1772 	 */
1773 	if ((flags & LK_INTERLOCK) == 0)
1774 		mtx_lock(&vp->v_interlock);
1775 	if (vp->v_flag & VXLOCK) {
1776 		if (vp->v_vxproc == curthread) {
1777 #if 0
1778 			/* this can now occur in normal operation */
1779 			log(LOG_INFO, "VXLOCK interlock avoided\n");
1780 #endif
1781 		} else {
1782 			vp->v_flag |= VXWANT;
1783 			msleep(vp, &vp->v_interlock, PINOD | PDROP, "vget", 0);
1784 			return (ENOENT);
1785 		}
1786 	}
1787 
1788 	vp->v_usecount++;
1789 
1790 	if (VSHOULDBUSY(vp))
1791 		vbusy(vp);
1792 	if (flags & LK_TYPE_MASK) {
1793 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1794 			/*
1795 			 * must expand vrele here because we do not want
1796 			 * to call VOP_INACTIVE if the reference count
1797 			 * drops back to zero since it was never really
1798 			 * active. We must remove it from the free list
1799 			 * before sleeping so that multiple processes do
1800 			 * not try to recycle it.
1801 			 */
1802 			mtx_lock(&vp->v_interlock);
1803 			vp->v_usecount--;
1804 			if (VSHOULDFREE(vp))
1805 				vfree(vp);
1806 			else
1807 				vlruvp(vp);
1808 			mtx_unlock(&vp->v_interlock);
1809 		}
1810 		return (error);
1811 	}
1812 	mtx_unlock(&vp->v_interlock);
1813 	return (0);
1814 }
1815 
1816 /*
1817  * Increase the reference count of a vnode.
1818  */
1819 void
1820 vref(struct vnode *vp)
1821 {
1822 	mtx_lock(&vp->v_interlock);
1823 	vp->v_usecount++;
1824 	mtx_unlock(&vp->v_interlock);
1825 }
1826 
1827 /*
1828  * Vnode put/release.
1829  * If count drops to zero, call inactive routine and return to freelist.
1830  */
1831 void
1832 vrele(vp)
1833 	struct vnode *vp;
1834 {
1835 	struct thread *td = curthread;	/* XXX */
1836 
1837 	KASSERT(vp != NULL, ("vrele: null vp"));
1838 
1839 	mtx_lock(&vp->v_interlock);
1840 
1841 	/* Skip this v_writecount check if we're going to panic below. */
1842 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
1843 	    ("vrele: missed vn_close"));
1844 
1845 	if (vp->v_usecount > 1) {
1846 
1847 		vp->v_usecount--;
1848 		mtx_unlock(&vp->v_interlock);
1849 
1850 		return;
1851 	}
1852 
1853 	if (vp->v_usecount == 1) {
1854 		vp->v_usecount--;
1855 		/*
1856 		 * We must call VOP_INACTIVE with the node locked.
1857 		 * If we are doing a vput, the node is already locked,
1858 		 * but, in the case of vrele, we must explicitly lock
1859 		 * the vnode before calling VOP_INACTIVE.
1860 		 */
1861 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0)
1862 			VOP_INACTIVE(vp, td);
1863 		if (VSHOULDFREE(vp))
1864 			vfree(vp);
1865 		else
1866 			vlruvp(vp);
1867 
1868 	} else {
1869 #ifdef DIAGNOSTIC
1870 		vprint("vrele: negative ref count", vp);
1871 		mtx_unlock(&vp->v_interlock);
1872 #endif
1873 		panic("vrele: negative ref cnt");
1874 	}
1875 }
1876 
1877 /*
1878  * Release an already locked vnode.  This give the same effects as
1879  * unlock+vrele(), but takes less time and avoids releasing and
1880  * re-aquiring the lock (as vrele() aquires the lock internally.)
1881  */
1882 void
1883 vput(vp)
1884 	struct vnode *vp;
1885 {
1886 	struct thread *td = curthread;	/* XXX */
1887 
1888 	GIANT_REQUIRED;
1889 
1890 	KASSERT(vp != NULL, ("vput: null vp"));
1891 	mtx_lock(&vp->v_interlock);
1892 	/* Skip this v_writecount check if we're going to panic below. */
1893 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
1894 	    ("vput: missed vn_close"));
1895 
1896 	if (vp->v_usecount > 1) {
1897 		vp->v_usecount--;
1898 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
1899 		return;
1900 	}
1901 
1902 	if (vp->v_usecount == 1) {
1903 		vp->v_usecount--;
1904 		/*
1905 		 * We must call VOP_INACTIVE with the node locked.
1906 		 * If we are doing a vput, the node is already locked,
1907 		 * so we just need to release the vnode mutex.
1908 		 */
1909 		mtx_unlock(&vp->v_interlock);
1910 		VOP_INACTIVE(vp, td);
1911 		if (VSHOULDFREE(vp))
1912 			vfree(vp);
1913 		else
1914 			vlruvp(vp);
1915 
1916 	} else {
1917 #ifdef DIAGNOSTIC
1918 		vprint("vput: negative ref count", vp);
1919 #endif
1920 		panic("vput: negative ref cnt");
1921 	}
1922 }
1923 
1924 /*
1925  * Somebody doesn't want the vnode recycled.
1926  */
1927 void
1928 vhold(vp)
1929 	register struct vnode *vp;
1930 {
1931 	int s;
1932 
1933 	s = splbio();
1934 	vp->v_holdcnt++;
1935 	if (VSHOULDBUSY(vp))
1936 		vbusy(vp);
1937 	splx(s);
1938 }
1939 
1940 /*
1941  * Note that there is one less who cares about this vnode.  vdrop() is the
1942  * opposite of vhold().
1943  */
1944 void
1945 vdrop(vp)
1946 	register struct vnode *vp;
1947 {
1948 	int s;
1949 
1950 	s = splbio();
1951 	if (vp->v_holdcnt <= 0)
1952 		panic("vdrop: holdcnt");
1953 	vp->v_holdcnt--;
1954 	if (VSHOULDFREE(vp))
1955 		vfree(vp);
1956 	else
1957 		vlruvp(vp);
1958 	splx(s);
1959 }
1960 
1961 /*
1962  * Remove any vnodes in the vnode table belonging to mount point mp.
1963  *
1964  * If FORCECLOSE is not specified, there should not be any active ones,
1965  * return error if any are found (nb: this is a user error, not a
1966  * system error). If FORCECLOSE is specified, detach any active vnodes
1967  * that are found.
1968  *
1969  * If WRITECLOSE is set, only flush out regular file vnodes open for
1970  * writing.
1971  *
1972  * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
1973  *
1974  * `rootrefs' specifies the base reference count for the root vnode
1975  * of this filesystem. The root vnode is considered busy if its
1976  * v_usecount exceeds this value. On a successful return, vflush()
1977  * will call vrele() on the root vnode exactly rootrefs times.
1978  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
1979  * be zero.
1980  */
1981 #ifdef DIAGNOSTIC
1982 static int busyprt = 0;		/* print out busy vnodes */
1983 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
1984 #endif
1985 
1986 int
1987 vflush(mp, rootrefs, flags)
1988 	struct mount *mp;
1989 	int rootrefs;
1990 	int flags;
1991 {
1992 	struct thread *td = curthread;	/* XXX */
1993 	struct vnode *vp, *nvp, *rootvp = NULL;
1994 	struct vattr vattr;
1995 	int busy = 0, error;
1996 
1997 	if (rootrefs > 0) {
1998 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1999 		    ("vflush: bad args"));
2000 		/*
2001 		 * Get the filesystem root vnode. We can vput() it
2002 		 * immediately, since with rootrefs > 0, it won't go away.
2003 		 */
2004 		if ((error = VFS_ROOT(mp, &rootvp)) != 0)
2005 			return (error);
2006 		vput(rootvp);
2007 
2008 	}
2009 	mtx_lock(&mntvnode_mtx);
2010 loop:
2011 	for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) {
2012 		/*
2013 		 * Make sure this vnode wasn't reclaimed in getnewvnode().
2014 		 * Start over if it has (it won't be on the list anymore).
2015 		 */
2016 		if (vp->v_mount != mp)
2017 			goto loop;
2018 		nvp = TAILQ_NEXT(vp, v_nmntvnodes);
2019 
2020 		mtx_unlock(&mntvnode_mtx);
2021 		mtx_lock(&vp->v_interlock);
2022 		/*
2023 		 * Skip over a vnodes marked VSYSTEM.
2024 		 */
2025 		if ((flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
2026 			mtx_unlock(&vp->v_interlock);
2027 			mtx_lock(&mntvnode_mtx);
2028 			continue;
2029 		}
2030 		/*
2031 		 * If WRITECLOSE is set, flush out unlinked but still open
2032 		 * files (even if open only for reading) and regular file
2033 		 * vnodes open for writing.
2034 		 */
2035 		if ((flags & WRITECLOSE) &&
2036 		    (vp->v_type == VNON ||
2037 		    (VOP_GETATTR(vp, &vattr, td->td_ucred, td) == 0 &&
2038 		    vattr.va_nlink > 0)) &&
2039 		    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2040 			mtx_unlock(&vp->v_interlock);
2041 			mtx_lock(&mntvnode_mtx);
2042 			continue;
2043 		}
2044 
2045 		/*
2046 		 * With v_usecount == 0, all we need to do is clear out the
2047 		 * vnode data structures and we are done.
2048 		 */
2049 		if (vp->v_usecount == 0) {
2050 			vgonel(vp, td);
2051 			mtx_lock(&mntvnode_mtx);
2052 			continue;
2053 		}
2054 
2055 		/*
2056 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2057 		 * or character devices, revert to an anonymous device. For
2058 		 * all other files, just kill them.
2059 		 */
2060 		if (flags & FORCECLOSE) {
2061 			if (vp->v_type != VCHR) {
2062 				vgonel(vp, td);
2063 			} else {
2064 				vclean(vp, 0, td);
2065 				vp->v_op = spec_vnodeop_p;
2066 				insmntque(vp, (struct mount *) 0);
2067 			}
2068 			mtx_lock(&mntvnode_mtx);
2069 			continue;
2070 		}
2071 #ifdef DIAGNOSTIC
2072 		if (busyprt)
2073 			vprint("vflush: busy vnode", vp);
2074 #endif
2075 		mtx_unlock(&vp->v_interlock);
2076 		mtx_lock(&mntvnode_mtx);
2077 		busy++;
2078 	}
2079 	mtx_unlock(&mntvnode_mtx);
2080 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2081 		/*
2082 		 * If just the root vnode is busy, and if its refcount
2083 		 * is equal to `rootrefs', then go ahead and kill it.
2084 		 */
2085 		mtx_lock(&rootvp->v_interlock);
2086 		KASSERT(busy > 0, ("vflush: not busy"));
2087 		KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs"));
2088 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2089 			vgonel(rootvp, td);
2090 			busy = 0;
2091 		} else
2092 			mtx_unlock(&rootvp->v_interlock);
2093 	}
2094 	if (busy)
2095 		return (EBUSY);
2096 	for (; rootrefs > 0; rootrefs--)
2097 		vrele(rootvp);
2098 	return (0);
2099 }
2100 
2101 /*
2102  * This moves a now (likely recyclable) vnode to the end of the
2103  * mountlist.  XXX However, it is temporarily disabled until we
2104  * can clean up ffs_sync() and friends, which have loop restart
2105  * conditions which this code causes to operate O(N^2).
2106  */
2107 static void
2108 vlruvp(struct vnode *vp)
2109 {
2110 #if 0
2111 	struct mount *mp;
2112 
2113 	if ((mp = vp->v_mount) != NULL) {
2114 		mtx_lock(&mntvnode_mtx);
2115 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2116 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2117 		mtx_unlock(&mntvnode_mtx);
2118 	}
2119 #endif
2120 }
2121 
2122 /*
2123  * Disassociate the underlying filesystem from a vnode.
2124  */
2125 static void
2126 vclean(vp, flags, td)
2127 	struct vnode *vp;
2128 	int flags;
2129 	struct thread *td;
2130 {
2131 	int active;
2132 
2133 	/*
2134 	 * Check to see if the vnode is in use. If so we have to reference it
2135 	 * before we clean it out so that its count cannot fall to zero and
2136 	 * generate a race against ourselves to recycle it.
2137 	 */
2138 	if ((active = vp->v_usecount))
2139 		vp->v_usecount++;
2140 
2141 	/*
2142 	 * Prevent the vnode from being recycled or brought into use while we
2143 	 * clean it out.
2144 	 */
2145 	if (vp->v_flag & VXLOCK)
2146 		panic("vclean: deadlock");
2147 	vp->v_flag |= VXLOCK;
2148 	vp->v_vxproc = curthread;
2149 	/*
2150 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2151 	 * have the object locked while it cleans it out. The VOP_LOCK
2152 	 * ensures that the VOP_INACTIVE routine is done with its work.
2153 	 * For active vnodes, it ensures that no other activity can
2154 	 * occur while the underlying object is being cleaned out.
2155 	 */
2156 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2157 
2158 	/*
2159 	 * Clean out any buffers associated with the vnode.
2160 	 * If the flush fails, just toss the buffers.
2161 	 */
2162 	if (flags & DOCLOSE) {
2163 		if (TAILQ_FIRST(&vp->v_dirtyblkhd) != NULL)
2164 			(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2165 		if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0)
2166 			vinvalbuf(vp, 0, NOCRED, td, 0, 0);
2167 	}
2168 
2169 	VOP_DESTROYVOBJECT(vp);
2170 
2171 	/*
2172 	 * Any other processes trying to obtain this lock must first
2173 	 * wait for VXLOCK to clear, then call the new lock operation.
2174 	 */
2175 	VOP_UNLOCK(vp, 0, td);
2176 
2177 	/*
2178 	 * If purging an active vnode, it must be closed and
2179 	 * deactivated before being reclaimed. Note that the
2180 	 * VOP_INACTIVE will unlock the vnode.
2181 	 */
2182 	if (active) {
2183 		if (flags & DOCLOSE)
2184 			VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2185 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2186 			panic("vclean: cannot relock.");
2187 		VOP_INACTIVE(vp, td);
2188 	}
2189 
2190 	/*
2191 	 * Reclaim the vnode.
2192 	 */
2193 	if (VOP_RECLAIM(vp, td))
2194 		panic("vclean: cannot reclaim");
2195 
2196 	if (active) {
2197 		/*
2198 		 * Inline copy of vrele() since VOP_INACTIVE
2199 		 * has already been called.
2200 		 */
2201 		mtx_lock(&vp->v_interlock);
2202 		if (--vp->v_usecount <= 0) {
2203 #ifdef DIAGNOSTIC
2204 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2205 				vprint("vclean: bad ref count", vp);
2206 				panic("vclean: ref cnt");
2207 			}
2208 #endif
2209 			vfree(vp);
2210 		}
2211 		mtx_unlock(&vp->v_interlock);
2212 	}
2213 
2214 	cache_purge(vp);
2215 	vp->v_vnlock = NULL;
2216 	lockdestroy(&vp->v_lock);
2217 
2218 	if (VSHOULDFREE(vp))
2219 		vfree(vp);
2220 
2221 	/*
2222 	 * Done with purge, notify sleepers of the grim news.
2223 	 */
2224 	vp->v_op = dead_vnodeop_p;
2225 	if (vp->v_pollinfo != NULL)
2226 		vn_pollgone(vp);
2227 	vp->v_tag = VT_NON;
2228 	vp->v_flag &= ~VXLOCK;
2229 	vp->v_vxproc = NULL;
2230 	if (vp->v_flag & VXWANT) {
2231 		vp->v_flag &= ~VXWANT;
2232 		wakeup(vp);
2233 	}
2234 }
2235 
2236 /*
2237  * Eliminate all activity associated with the requested vnode
2238  * and with all vnodes aliased to the requested vnode.
2239  */
2240 int
2241 vop_revoke(ap)
2242 	struct vop_revoke_args /* {
2243 		struct vnode *a_vp;
2244 		int a_flags;
2245 	} */ *ap;
2246 {
2247 	struct vnode *vp, *vq;
2248 	dev_t dev;
2249 
2250 	KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke"));
2251 
2252 	vp = ap->a_vp;
2253 	/*
2254 	 * If a vgone (or vclean) is already in progress,
2255 	 * wait until it is done and return.
2256 	 */
2257 	if (vp->v_flag & VXLOCK) {
2258 		vp->v_flag |= VXWANT;
2259 		msleep(vp, &vp->v_interlock, PINOD | PDROP,
2260 		    "vop_revokeall", 0);
2261 		return (0);
2262 	}
2263 	dev = vp->v_rdev;
2264 	for (;;) {
2265 		mtx_lock(&spechash_mtx);
2266 		vq = SLIST_FIRST(&dev->si_hlist);
2267 		mtx_unlock(&spechash_mtx);
2268 		if (!vq)
2269 			break;
2270 		vgone(vq);
2271 	}
2272 	return (0);
2273 }
2274 
2275 /*
2276  * Recycle an unused vnode to the front of the free list.
2277  * Release the passed interlock if the vnode will be recycled.
2278  */
2279 int
2280 vrecycle(vp, inter_lkp, td)
2281 	struct vnode *vp;
2282 	struct mtx *inter_lkp;
2283 	struct thread *td;
2284 {
2285 
2286 	mtx_lock(&vp->v_interlock);
2287 	if (vp->v_usecount == 0) {
2288 		if (inter_lkp) {
2289 			mtx_unlock(inter_lkp);
2290 		}
2291 		vgonel(vp, td);
2292 		return (1);
2293 	}
2294 	mtx_unlock(&vp->v_interlock);
2295 	return (0);
2296 }
2297 
2298 /*
2299  * Eliminate all activity associated with a vnode
2300  * in preparation for reuse.
2301  */
2302 void
2303 vgone(vp)
2304 	register struct vnode *vp;
2305 {
2306 	struct thread *td = curthread;	/* XXX */
2307 
2308 	mtx_lock(&vp->v_interlock);
2309 	vgonel(vp, td);
2310 }
2311 
2312 /*
2313  * vgone, with the vp interlock held.
2314  */
2315 void
2316 vgonel(vp, td)
2317 	struct vnode *vp;
2318 	struct thread *td;
2319 {
2320 	int s;
2321 
2322 	/*
2323 	 * If a vgone (or vclean) is already in progress,
2324 	 * wait until it is done and return.
2325 	 */
2326 	if (vp->v_flag & VXLOCK) {
2327 		vp->v_flag |= VXWANT;
2328 		msleep(vp, &vp->v_interlock, PINOD | PDROP, "vgone", 0);
2329 		return;
2330 	}
2331 
2332 	/*
2333 	 * Clean out the filesystem specific data.
2334 	 */
2335 	vclean(vp, DOCLOSE, td);
2336 	mtx_lock(&vp->v_interlock);
2337 
2338 	/*
2339 	 * Delete from old mount point vnode list, if on one.
2340 	 */
2341 	if (vp->v_mount != NULL)
2342 		insmntque(vp, (struct mount *)0);
2343 	/*
2344 	 * If special device, remove it from special device alias list
2345 	 * if it is on one.
2346 	 */
2347 	if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) {
2348 		mtx_lock(&spechash_mtx);
2349 		SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext);
2350 		freedev(vp->v_rdev);
2351 		mtx_unlock(&spechash_mtx);
2352 		vp->v_rdev = NULL;
2353 	}
2354 
2355 	/*
2356 	 * If it is on the freelist and not already at the head,
2357 	 * move it to the head of the list. The test of the
2358 	 * VDOOMED flag and the reference count of zero is because
2359 	 * it will be removed from the free list by getnewvnode,
2360 	 * but will not have its reference count incremented until
2361 	 * after calling vgone. If the reference count were
2362 	 * incremented first, vgone would (incorrectly) try to
2363 	 * close the previous instance of the underlying object.
2364 	 */
2365 	if (vp->v_usecount == 0 && !(vp->v_flag & VDOOMED)) {
2366 		s = splbio();
2367 		mtx_lock(&vnode_free_list_mtx);
2368 		if (vp->v_flag & VFREE)
2369 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2370 		else
2371 			freevnodes++;
2372 		vp->v_flag |= VFREE;
2373 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2374 		mtx_unlock(&vnode_free_list_mtx);
2375 		splx(s);
2376 	}
2377 
2378 	vp->v_type = VBAD;
2379 	mtx_unlock(&vp->v_interlock);
2380 }
2381 
2382 /*
2383  * Lookup a vnode by device number.
2384  */
2385 int
2386 vfinddev(dev, type, vpp)
2387 	dev_t dev;
2388 	enum vtype type;
2389 	struct vnode **vpp;
2390 {
2391 	struct vnode *vp;
2392 
2393 	mtx_lock(&spechash_mtx);
2394 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2395 		if (type == vp->v_type) {
2396 			*vpp = vp;
2397 			mtx_unlock(&spechash_mtx);
2398 			return (1);
2399 		}
2400 	}
2401 	mtx_unlock(&spechash_mtx);
2402 	return (0);
2403 }
2404 
2405 /*
2406  * Calculate the total number of references to a special device.
2407  */
2408 int
2409 vcount(vp)
2410 	struct vnode *vp;
2411 {
2412 	struct vnode *vq;
2413 	int count;
2414 
2415 	count = 0;
2416 	mtx_lock(&spechash_mtx);
2417 	SLIST_FOREACH(vq, &vp->v_rdev->si_hlist, v_specnext)
2418 		count += vq->v_usecount;
2419 	mtx_unlock(&spechash_mtx);
2420 	return (count);
2421 }
2422 
2423 /*
2424  * Same as above, but using the dev_t as argument
2425  */
2426 int
2427 count_dev(dev)
2428 	dev_t dev;
2429 {
2430 	struct vnode *vp;
2431 
2432 	vp = SLIST_FIRST(&dev->si_hlist);
2433 	if (vp == NULL)
2434 		return (0);
2435 	return(vcount(vp));
2436 }
2437 
2438 /*
2439  * Print out a description of a vnode.
2440  */
2441 static char *typename[] =
2442 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2443 
2444 void
2445 vprint(label, vp)
2446 	char *label;
2447 	struct vnode *vp;
2448 {
2449 	char buf[96];
2450 
2451 	if (label != NULL)
2452 		printf("%s: %p: ", label, (void *)vp);
2453 	else
2454 		printf("%p: ", (void *)vp);
2455 	printf("type %s, usecount %d, writecount %d, refcount %d,",
2456 	    typename[vp->v_type], vp->v_usecount, vp->v_writecount,
2457 	    vp->v_holdcnt);
2458 	buf[0] = '\0';
2459 	if (vp->v_flag & VROOT)
2460 		strcat(buf, "|VROOT");
2461 	if (vp->v_flag & VTEXT)
2462 		strcat(buf, "|VTEXT");
2463 	if (vp->v_flag & VSYSTEM)
2464 		strcat(buf, "|VSYSTEM");
2465 	if (vp->v_flag & VXLOCK)
2466 		strcat(buf, "|VXLOCK");
2467 	if (vp->v_flag & VXWANT)
2468 		strcat(buf, "|VXWANT");
2469 	if (vp->v_flag & VBWAIT)
2470 		strcat(buf, "|VBWAIT");
2471 	if (vp->v_flag & VDOOMED)
2472 		strcat(buf, "|VDOOMED");
2473 	if (vp->v_flag & VFREE)
2474 		strcat(buf, "|VFREE");
2475 	if (vp->v_flag & VOBJBUF)
2476 		strcat(buf, "|VOBJBUF");
2477 	if (buf[0] != '\0')
2478 		printf(" flags (%s)", &buf[1]);
2479 	if (vp->v_data == NULL) {
2480 		printf("\n");
2481 	} else {
2482 		printf("\n\t");
2483 		VOP_PRINT(vp);
2484 	}
2485 }
2486 
2487 #ifdef DDB
2488 #include <ddb/ddb.h>
2489 /*
2490  * List all of the locked vnodes in the system.
2491  * Called when debugging the kernel.
2492  */
2493 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2494 {
2495 	struct thread *td = curthread;	/* XXX */
2496 	struct mount *mp, *nmp;
2497 	struct vnode *vp;
2498 
2499 	printf("Locked vnodes\n");
2500 	mtx_lock(&mountlist_mtx);
2501 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2502 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
2503 			nmp = TAILQ_NEXT(mp, mnt_list);
2504 			continue;
2505 		}
2506 		mtx_lock(&mntvnode_mtx);
2507 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2508 			if (VOP_ISLOCKED(vp, NULL))
2509 				vprint((char *)0, vp);
2510 		}
2511 		mtx_unlock(&mntvnode_mtx);
2512 		mtx_lock(&mountlist_mtx);
2513 		nmp = TAILQ_NEXT(mp, mnt_list);
2514 		vfs_unbusy(mp, td);
2515 	}
2516 	mtx_unlock(&mountlist_mtx);
2517 }
2518 #endif
2519 
2520 /*
2521  * Top level filesystem related information gathering.
2522  */
2523 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2524 
2525 static int
2526 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2527 {
2528 	int *name = (int *)arg1 - 1;	/* XXX */
2529 	u_int namelen = arg2 + 1;	/* XXX */
2530 	struct vfsconf *vfsp;
2531 
2532 #if 1 || defined(COMPAT_PRELITE2)
2533 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2534 	if (namelen == 1)
2535 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2536 #endif
2537 
2538 	/* XXX the below code does not compile; vfs_sysctl does not exist. */
2539 #ifdef notyet
2540 	/* all sysctl names at this level are at least name and field */
2541 	if (namelen < 2)
2542 		return (ENOTDIR);		/* overloaded */
2543 	if (name[0] != VFS_GENERIC) {
2544 		for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next)
2545 			if (vfsp->vfc_typenum == name[0])
2546 				break;
2547 		if (vfsp == NULL)
2548 			return (EOPNOTSUPP);
2549 		return ((*vfsp->vfc_vfsops->vfs_sysctl)(&name[1], namelen - 1,
2550 		    oldp, oldlenp, newp, newlen, td));
2551 	}
2552 #endif
2553 	switch (name[1]) {
2554 	case VFS_MAXTYPENUM:
2555 		if (namelen != 2)
2556 			return (ENOTDIR);
2557 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2558 	case VFS_CONF:
2559 		if (namelen != 3)
2560 			return (ENOTDIR);	/* overloaded */
2561 		for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next)
2562 			if (vfsp->vfc_typenum == name[2])
2563 				break;
2564 		if (vfsp == NULL)
2565 			return (EOPNOTSUPP);
2566 		return (SYSCTL_OUT(req, vfsp, sizeof *vfsp));
2567 	}
2568 	return (EOPNOTSUPP);
2569 }
2570 
2571 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD, vfs_sysctl,
2572 	"Generic filesystem");
2573 
2574 #if 1 || defined(COMPAT_PRELITE2)
2575 
2576 static int
2577 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2578 {
2579 	int error;
2580 	struct vfsconf *vfsp;
2581 	struct ovfsconf ovfs;
2582 
2583 	for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) {
2584 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2585 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2586 		ovfs.vfc_index = vfsp->vfc_typenum;
2587 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2588 		ovfs.vfc_flags = vfsp->vfc_flags;
2589 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2590 		if (error)
2591 			return error;
2592 	}
2593 	return 0;
2594 }
2595 
2596 #endif /* 1 || COMPAT_PRELITE2 */
2597 
2598 #if COMPILING_LINT
2599 #define KINFO_VNODESLOP	10
2600 /*
2601  * Dump vnode list (via sysctl).
2602  * Copyout address of vnode followed by vnode.
2603  */
2604 /* ARGSUSED */
2605 static int
2606 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2607 {
2608 	struct thread *td = curthread;	/* XXX */
2609 	struct mount *mp, *nmp;
2610 	struct vnode *nvp, *vp;
2611 	int error;
2612 
2613 #define VPTRSZ	sizeof (struct vnode *)
2614 #define VNODESZ	sizeof (struct vnode)
2615 
2616 	req->lock = 0;
2617 	if (!req->oldptr) /* Make an estimate */
2618 		return (SYSCTL_OUT(req, 0,
2619 			(numvnodes + KINFO_VNODESLOP) * (VPTRSZ + VNODESZ)));
2620 
2621 	mtx_lock(&mountlist_mtx);
2622 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2623 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
2624 			nmp = TAILQ_NEXT(mp, mnt_list);
2625 			continue;
2626 		}
2627 		mtx_lock(&mntvnode_mtx);
2628 again:
2629 		for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
2630 		     vp != NULL;
2631 		     vp = nvp) {
2632 			/*
2633 			 * Check that the vp is still associated with
2634 			 * this filesystem.  RACE: could have been
2635 			 * recycled onto the same filesystem.
2636 			 */
2637 			if (vp->v_mount != mp)
2638 				goto again;
2639 			nvp = TAILQ_NEXT(vp, v_nmntvnodes);
2640 			mtx_unlock(&mntvnode_mtx);
2641 			if ((error = SYSCTL_OUT(req, &vp, VPTRSZ)) ||
2642 			    (error = SYSCTL_OUT(req, vp, VNODESZ)))
2643 				return (error);
2644 			mtx_lock(&mntvnode_mtx);
2645 		}
2646 		mtx_unlock(&mntvnode_mtx);
2647 		mtx_lock(&mountlist_mtx);
2648 		nmp = TAILQ_NEXT(mp, mnt_list);
2649 		vfs_unbusy(mp, td);
2650 	}
2651 	mtx_unlock(&mountlist_mtx);
2652 
2653 	return (0);
2654 }
2655 
2656 /*
2657  * XXX
2658  * Exporting the vnode list on large systems causes them to crash.
2659  * Exporting the vnode list on medium systems causes sysctl to coredump.
2660  */
2661 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2662 	0, 0, sysctl_vnode, "S,vnode", "");
2663 #endif
2664 
2665 /*
2666  * Check to see if a filesystem is mounted on a block device.
2667  */
2668 int
2669 vfs_mountedon(vp)
2670 	struct vnode *vp;
2671 {
2672 
2673 	if (vp->v_rdev->si_mountpoint != NULL)
2674 		return (EBUSY);
2675 	return (0);
2676 }
2677 
2678 /*
2679  * Unmount all filesystems. The list is traversed in reverse order
2680  * of mounting to avoid dependencies.
2681  */
2682 void
2683 vfs_unmountall()
2684 {
2685 	struct mount *mp;
2686 	struct thread *td;
2687 	int error;
2688 
2689 	if (curthread != NULL)
2690 		td = curthread;
2691 	else
2692 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
2693 	/*
2694 	 * Since this only runs when rebooting, it is not interlocked.
2695 	 */
2696 	while(!TAILQ_EMPTY(&mountlist)) {
2697 		mp = TAILQ_LAST(&mountlist, mntlist);
2698 		error = dounmount(mp, MNT_FORCE, td);
2699 		if (error) {
2700 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2701 			printf("unmount of %s failed (",
2702 			    mp->mnt_stat.f_mntonname);
2703 			if (error == EBUSY)
2704 				printf("BUSY)\n");
2705 			else
2706 				printf("%d)\n", error);
2707 		} else {
2708 			/* The unmount has removed mp from the mountlist */
2709 		}
2710 	}
2711 }
2712 
2713 /*
2714  * perform msync on all vnodes under a mount point
2715  * the mount point must be locked.
2716  */
2717 void
2718 vfs_msync(struct mount *mp, int flags)
2719 {
2720 	struct vnode *vp, *nvp;
2721 	struct vm_object *obj;
2722 	int tries;
2723 
2724 	GIANT_REQUIRED;
2725 
2726 	tries = 5;
2727 	mtx_lock(&mntvnode_mtx);
2728 loop:
2729 	for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) {
2730 		if (vp->v_mount != mp) {
2731 			if (--tries > 0)
2732 				goto loop;
2733 			break;
2734 		}
2735 		nvp = TAILQ_NEXT(vp, v_nmntvnodes);
2736 
2737 		if (vp->v_flag & VXLOCK)	/* XXX: what if MNT_WAIT? */
2738 			continue;
2739 
2740 		if (vp->v_flag & VNOSYNC)	/* unlinked, skip it */
2741 			continue;
2742 
2743 		if ((vp->v_flag & VOBJDIRTY) &&
2744 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2745 			mtx_unlock(&mntvnode_mtx);
2746 			if (!vget(vp,
2747 			    LK_EXCLUSIVE | LK_RETRY | LK_NOOBJ, curthread)) {
2748 				if (VOP_GETVOBJECT(vp, &obj) == 0) {
2749 					vm_object_page_clean(obj, 0, 0,
2750 					    flags == MNT_WAIT ?
2751 					    OBJPC_SYNC : OBJPC_NOSYNC);
2752 				}
2753 				vput(vp);
2754 			}
2755 			mtx_lock(&mntvnode_mtx);
2756 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2757 				if (--tries > 0)
2758 					goto loop;
2759 				break;
2760 			}
2761 		}
2762 	}
2763 	mtx_unlock(&mntvnode_mtx);
2764 }
2765 
2766 /*
2767  * Create the VM object needed for VMIO and mmap support.  This
2768  * is done for all VREG files in the system.  Some filesystems might
2769  * afford the additional metadata buffering capability of the
2770  * VMIO code by making the device node be VMIO mode also.
2771  *
2772  * vp must be locked when vfs_object_create is called.
2773  */
2774 int
2775 vfs_object_create(vp, td, cred)
2776 	struct vnode *vp;
2777 	struct thread *td;
2778 	struct ucred *cred;
2779 {
2780 	GIANT_REQUIRED;
2781 	return (VOP_CREATEVOBJECT(vp, cred, td));
2782 }
2783 
2784 /*
2785  * Mark a vnode as free, putting it up for recycling.
2786  */
2787 void
2788 vfree(vp)
2789 	struct vnode *vp;
2790 {
2791 	int s;
2792 
2793 	s = splbio();
2794 	mtx_lock(&vnode_free_list_mtx);
2795 	KASSERT((vp->v_flag & VFREE) == 0, ("vnode already free"));
2796 	if (vp->v_flag & VAGE) {
2797 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2798 	} else {
2799 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2800 	}
2801 	freevnodes++;
2802 	mtx_unlock(&vnode_free_list_mtx);
2803 	vp->v_flag &= ~VAGE;
2804 	vp->v_flag |= VFREE;
2805 	splx(s);
2806 }
2807 
2808 /*
2809  * Opposite of vfree() - mark a vnode as in use.
2810  */
2811 void
2812 vbusy(vp)
2813 	struct vnode *vp;
2814 {
2815 	int s;
2816 
2817 	s = splbio();
2818 	mtx_lock(&vnode_free_list_mtx);
2819 	KASSERT((vp->v_flag & VFREE) != 0, ("vnode not free"));
2820 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2821 	freevnodes--;
2822 	mtx_unlock(&vnode_free_list_mtx);
2823 	vp->v_flag &= ~(VFREE|VAGE);
2824 	splx(s);
2825 }
2826 
2827 /*
2828  * Record a process's interest in events which might happen to
2829  * a vnode.  Because poll uses the historic select-style interface
2830  * internally, this routine serves as both the ``check for any
2831  * pending events'' and the ``record my interest in future events''
2832  * functions.  (These are done together, while the lock is held,
2833  * to avoid race conditions.)
2834  */
2835 int
2836 vn_pollrecord(vp, td, events)
2837 	struct vnode *vp;
2838 	struct thread *td;
2839 	short events;
2840 {
2841 
2842 	if (vp->v_pollinfo == NULL)
2843 		v_addpollinfo(vp);
2844 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2845 	if (vp->v_pollinfo->vpi_revents & events) {
2846 		/*
2847 		 * This leaves events we are not interested
2848 		 * in available for the other process which
2849 		 * which presumably had requested them
2850 		 * (otherwise they would never have been
2851 		 * recorded).
2852 		 */
2853 		events &= vp->v_pollinfo->vpi_revents;
2854 		vp->v_pollinfo->vpi_revents &= ~events;
2855 
2856 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2857 		return events;
2858 	}
2859 	vp->v_pollinfo->vpi_events |= events;
2860 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2861 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2862 	return 0;
2863 }
2864 
2865 /*
2866  * Note the occurrence of an event.  If the VN_POLLEVENT macro is used,
2867  * it is possible for us to miss an event due to race conditions, but
2868  * that condition is expected to be rare, so for the moment it is the
2869  * preferred interface.
2870  */
2871 void
2872 vn_pollevent(vp, events)
2873 	struct vnode *vp;
2874 	short events;
2875 {
2876 
2877 	if (vp->v_pollinfo == NULL)
2878 		v_addpollinfo(vp);
2879 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2880 	if (vp->v_pollinfo->vpi_events & events) {
2881 		/*
2882 		 * We clear vpi_events so that we don't
2883 		 * call selwakeup() twice if two events are
2884 		 * posted before the polling process(es) is
2885 		 * awakened.  This also ensures that we take at
2886 		 * most one selwakeup() if the polling process
2887 		 * is no longer interested.  However, it does
2888 		 * mean that only one event can be noticed at
2889 		 * a time.  (Perhaps we should only clear those
2890 		 * event bits which we note?) XXX
2891 		 */
2892 		vp->v_pollinfo->vpi_events = 0;	/* &= ~events ??? */
2893 		vp->v_pollinfo->vpi_revents |= events;
2894 		selwakeup(&vp->v_pollinfo->vpi_selinfo);
2895 	}
2896 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2897 }
2898 
2899 /*
2900  * Wake up anyone polling on vp because it is being revoked.
2901  * This depends on dead_poll() returning POLLHUP for correct
2902  * behavior.
2903  */
2904 void
2905 vn_pollgone(vp)
2906 	struct vnode *vp;
2907 {
2908 
2909 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2910 	VN_KNOTE(vp, NOTE_REVOKE);
2911 	if (vp->v_pollinfo->vpi_events) {
2912 		vp->v_pollinfo->vpi_events = 0;
2913 		selwakeup(&vp->v_pollinfo->vpi_selinfo);
2914 	}
2915 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2916 }
2917 
2918 
2919 
2920 /*
2921  * Routine to create and manage a filesystem syncer vnode.
2922  */
2923 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2924 static int	sync_fsync(struct  vop_fsync_args *);
2925 static int	sync_inactive(struct  vop_inactive_args *);
2926 static int	sync_reclaim(struct  vop_reclaim_args *);
2927 static int	sync_print(struct vop_print_args *);
2928 
2929 static vop_t **sync_vnodeop_p;
2930 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = {
2931 	{ &vop_default_desc,	(vop_t *) vop_eopnotsupp },
2932 	{ &vop_close_desc,	(vop_t *) sync_close },		/* close */
2933 	{ &vop_fsync_desc,	(vop_t *) sync_fsync },		/* fsync */
2934 	{ &vop_inactive_desc,	(vop_t *) sync_inactive },	/* inactive */
2935 	{ &vop_reclaim_desc,	(vop_t *) sync_reclaim },	/* reclaim */
2936 	{ &vop_lock_desc,	(vop_t *) vop_stdlock },	/* lock */
2937 	{ &vop_unlock_desc,	(vop_t *) vop_stdunlock },	/* unlock */
2938 	{ &vop_print_desc,	(vop_t *) sync_print },		/* print */
2939 	{ &vop_islocked_desc,	(vop_t *) vop_stdislocked },	/* islocked */
2940 	{ NULL, NULL }
2941 };
2942 static struct vnodeopv_desc sync_vnodeop_opv_desc =
2943 	{ &sync_vnodeop_p, sync_vnodeop_entries };
2944 
2945 VNODEOP_SET(sync_vnodeop_opv_desc);
2946 
2947 /*
2948  * Create a new filesystem syncer vnode for the specified mount point.
2949  */
2950 int
2951 vfs_allocate_syncvnode(mp)
2952 	struct mount *mp;
2953 {
2954 	struct vnode *vp;
2955 	static long start, incr, next;
2956 	int error;
2957 
2958 	/* Allocate a new vnode */
2959 	if ((error = getnewvnode(VT_VFS, mp, sync_vnodeop_p, &vp)) != 0) {
2960 		mp->mnt_syncer = NULL;
2961 		return (error);
2962 	}
2963 	vp->v_type = VNON;
2964 	/*
2965 	 * Place the vnode onto the syncer worklist. We attempt to
2966 	 * scatter them about on the list so that they will go off
2967 	 * at evenly distributed times even if all the filesystems
2968 	 * are mounted at once.
2969 	 */
2970 	next += incr;
2971 	if (next == 0 || next > syncer_maxdelay) {
2972 		start /= 2;
2973 		incr /= 2;
2974 		if (start == 0) {
2975 			start = syncer_maxdelay / 2;
2976 			incr = syncer_maxdelay;
2977 		}
2978 		next = start;
2979 	}
2980 	vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0);
2981 	mp->mnt_syncer = vp;
2982 	return (0);
2983 }
2984 
2985 /*
2986  * Do a lazy sync of the filesystem.
2987  */
2988 static int
2989 sync_fsync(ap)
2990 	struct vop_fsync_args /* {
2991 		struct vnode *a_vp;
2992 		struct ucred *a_cred;
2993 		int a_waitfor;
2994 		struct thread *a_td;
2995 	} */ *ap;
2996 {
2997 	struct vnode *syncvp = ap->a_vp;
2998 	struct mount *mp = syncvp->v_mount;
2999 	struct thread *td = ap->a_td;
3000 	int asyncflag;
3001 
3002 	/*
3003 	 * We only need to do something if this is a lazy evaluation.
3004 	 */
3005 	if (ap->a_waitfor != MNT_LAZY)
3006 		return (0);
3007 
3008 	/*
3009 	 * Move ourselves to the back of the sync list.
3010 	 */
3011 	vn_syncer_add_to_worklist(syncvp, syncdelay);
3012 
3013 	/*
3014 	 * Walk the list of vnodes pushing all that are dirty and
3015 	 * not already on the sync list.
3016 	 */
3017 	mtx_lock(&mountlist_mtx);
3018 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3019 		mtx_unlock(&mountlist_mtx);
3020 		return (0);
3021 	}
3022 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3023 		vfs_unbusy(mp, td);
3024 		return (0);
3025 	}
3026 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3027 	mp->mnt_flag &= ~MNT_ASYNC;
3028 	vfs_msync(mp, MNT_NOWAIT);
3029 	VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td);
3030 	if (asyncflag)
3031 		mp->mnt_flag |= MNT_ASYNC;
3032 	vn_finished_write(mp);
3033 	vfs_unbusy(mp, td);
3034 	return (0);
3035 }
3036 
3037 /*
3038  * The syncer vnode is no referenced.
3039  */
3040 static int
3041 sync_inactive(ap)
3042 	struct vop_inactive_args /* {
3043 		struct vnode *a_vp;
3044 		struct thread *a_td;
3045 	} */ *ap;
3046 {
3047 
3048 	vgone(ap->a_vp);
3049 	return (0);
3050 }
3051 
3052 /*
3053  * The syncer vnode is no longer needed and is being decommissioned.
3054  *
3055  * Modifications to the worklist must be protected at splbio().
3056  */
3057 static int
3058 sync_reclaim(ap)
3059 	struct vop_reclaim_args /* {
3060 		struct vnode *a_vp;
3061 	} */ *ap;
3062 {
3063 	struct vnode *vp = ap->a_vp;
3064 	int s;
3065 
3066 	s = splbio();
3067 	vp->v_mount->mnt_syncer = NULL;
3068 	if (vp->v_flag & VONWORKLST) {
3069 		LIST_REMOVE(vp, v_synclist);
3070 		vp->v_flag &= ~VONWORKLST;
3071 	}
3072 	splx(s);
3073 
3074 	return (0);
3075 }
3076 
3077 /*
3078  * Print out a syncer vnode.
3079  */
3080 static int
3081 sync_print(ap)
3082 	struct vop_print_args /* {
3083 		struct vnode *a_vp;
3084 	} */ *ap;
3085 {
3086 	struct vnode *vp = ap->a_vp;
3087 
3088 	printf("syncer vnode");
3089 	if (vp->v_vnlock != NULL)
3090 		lockmgr_printinfo(vp->v_vnlock);
3091 	printf("\n");
3092 	return (0);
3093 }
3094 
3095 /*
3096  * extract the dev_t from a VCHR
3097  */
3098 dev_t
3099 vn_todev(vp)
3100 	struct vnode *vp;
3101 {
3102 	if (vp->v_type != VCHR)
3103 		return (NODEV);
3104 	return (vp->v_rdev);
3105 }
3106 
3107 /*
3108  * Check if vnode represents a disk device
3109  */
3110 int
3111 vn_isdisk(vp, errp)
3112 	struct vnode *vp;
3113 	int *errp;
3114 {
3115 	struct cdevsw *cdevsw;
3116 
3117 	if (vp->v_type != VCHR) {
3118 		if (errp != NULL)
3119 			*errp = ENOTBLK;
3120 		return (0);
3121 	}
3122 	if (vp->v_rdev == NULL) {
3123 		if (errp != NULL)
3124 			*errp = ENXIO;
3125 		return (0);
3126 	}
3127 	cdevsw = devsw(vp->v_rdev);
3128 	if (cdevsw == NULL) {
3129 		if (errp != NULL)
3130 			*errp = ENXIO;
3131 		return (0);
3132 	}
3133 	if (!(cdevsw->d_flags & D_DISK)) {
3134 		if (errp != NULL)
3135 			*errp = ENOTBLK;
3136 		return (0);
3137 	}
3138 	if (errp != NULL)
3139 		*errp = 0;
3140 	return (1);
3141 }
3142 
3143 /*
3144  * Free data allocated by namei(); see namei(9) for details.
3145  */
3146 void
3147 NDFREE(ndp, flags)
3148      struct nameidata *ndp;
3149      const uint flags;
3150 {
3151 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3152 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3153 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3154 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3155 	}
3156 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3157 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3158 	    ndp->ni_dvp != ndp->ni_vp)
3159 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3160 	if (!(flags & NDF_NO_DVP_RELE) &&
3161 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3162 		vrele(ndp->ni_dvp);
3163 		ndp->ni_dvp = NULL;
3164 	}
3165 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3166 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3167 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3168 	if (!(flags & NDF_NO_VP_RELE) &&
3169 	    ndp->ni_vp) {
3170 		vrele(ndp->ni_vp);
3171 		ndp->ni_vp = NULL;
3172 	}
3173 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3174 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3175 		vrele(ndp->ni_startdir);
3176 		ndp->ni_startdir = NULL;
3177 	}
3178 }
3179 
3180 /*
3181  * Common filesystem object access control check routine.  Accepts a
3182  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3183  * and optional call-by-reference privused argument allowing vaccess()
3184  * to indicate to the caller whether privilege was used to satisfy the
3185  * request.  Returns 0 on success, or an errno on failure.
3186  */
3187 int
3188 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3189 	enum vtype type;
3190 	mode_t file_mode;
3191 	uid_t file_uid;
3192 	gid_t file_gid;
3193 	mode_t acc_mode;
3194 	struct ucred *cred;
3195 	int *privused;
3196 {
3197 	mode_t dac_granted;
3198 #ifdef CAPABILITIES
3199 	mode_t cap_granted;
3200 #endif
3201 
3202 	/*
3203 	 * Look for a normal, non-privileged way to access the file/directory
3204 	 * as requested.  If it exists, go with that.
3205 	 */
3206 
3207 	if (privused != NULL)
3208 		*privused = 0;
3209 
3210 	dac_granted = 0;
3211 
3212 	/* Check the owner. */
3213 	if (cred->cr_uid == file_uid) {
3214 		dac_granted |= VADMIN;
3215 		if (file_mode & S_IXUSR)
3216 			dac_granted |= VEXEC;
3217 		if (file_mode & S_IRUSR)
3218 			dac_granted |= VREAD;
3219 		if (file_mode & S_IWUSR)
3220 			dac_granted |= VWRITE;
3221 
3222 		if ((acc_mode & dac_granted) == acc_mode)
3223 			return (0);
3224 
3225 		goto privcheck;
3226 	}
3227 
3228 	/* Otherwise, check the groups (first match) */
3229 	if (groupmember(file_gid, cred)) {
3230 		if (file_mode & S_IXGRP)
3231 			dac_granted |= VEXEC;
3232 		if (file_mode & S_IRGRP)
3233 			dac_granted |= VREAD;
3234 		if (file_mode & S_IWGRP)
3235 			dac_granted |= VWRITE;
3236 
3237 		if ((acc_mode & dac_granted) == acc_mode)
3238 			return (0);
3239 
3240 		goto privcheck;
3241 	}
3242 
3243 	/* Otherwise, check everyone else. */
3244 	if (file_mode & S_IXOTH)
3245 		dac_granted |= VEXEC;
3246 	if (file_mode & S_IROTH)
3247 		dac_granted |= VREAD;
3248 	if (file_mode & S_IWOTH)
3249 		dac_granted |= VWRITE;
3250 	if ((acc_mode & dac_granted) == acc_mode)
3251 		return (0);
3252 
3253 privcheck:
3254 	if (!suser_cred(cred, PRISON_ROOT)) {
3255 		/* XXX audit: privilege used */
3256 		if (privused != NULL)
3257 			*privused = 1;
3258 		return (0);
3259 	}
3260 
3261 #ifdef CAPABILITIES
3262 	/*
3263 	 * Build a capability mask to determine if the set of capabilities
3264 	 * satisfies the requirements when combined with the granted mask
3265 	 * from above.
3266 	 * For each capability, if the capability is required, bitwise
3267 	 * or the request type onto the cap_granted mask.
3268 	 */
3269 	cap_granted = 0;
3270 
3271 	if (type == VDIR) {
3272 		/*
3273 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3274 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3275 		 */
3276 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3277 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT))
3278 			cap_granted |= VEXEC;
3279 	} else {
3280 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3281 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT))
3282 			cap_granted |= VEXEC;
3283 	}
3284 
3285 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3286 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT))
3287 		cap_granted |= VREAD;
3288 
3289 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3290 	    !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT))
3291 		cap_granted |= VWRITE;
3292 
3293 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3294 	    !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT))
3295 		cap_granted |= VADMIN;
3296 
3297 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3298 		/* XXX audit: privilege used */
3299 		if (privused != NULL)
3300 			*privused = 1;
3301 		return (0);
3302 	}
3303 #endif
3304 
3305 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3306 }
3307