xref: /freebsd/sys/kern/vfs_subr.c (revision 71fe318b852b8dfb3e799cb12ef184750f7f8eac)
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 #include "opt_mac.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/conf.h>
53 #include <sys/eventhandler.h>
54 #include <sys/extattr.h>
55 #include <sys/fcntl.h>
56 #include <sys/kernel.h>
57 #include <sys/kthread.h>
58 #include <sys/mac.h>
59 #include <sys/malloc.h>
60 #include <sys/mount.h>
61 #include <sys/namei.h>
62 #include <sys/stat.h>
63 #include <sys/sysctl.h>
64 #include <sys/syslog.h>
65 #include <sys/vmmeter.h>
66 #include <sys/vnode.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_page.h>
74 #include <vm/uma.h>
75 
76 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
77 
78 static void	addalias(struct vnode *vp, dev_t nvp_rdev);
79 static void	insmntque(struct vnode *vp, struct mount *mp);
80 static void	vclean(struct vnode *vp, int flags, struct thread *td);
81 static void	vlruvp(struct vnode *vp);
82 static int	flushbuflist(struct buf *blist, int flags, struct vnode *vp,
83 		    int slpflag, int slptimeo, int *errorp);
84 static int	vcanrecycle(struct vnode *vp, struct mount **vnmpp);
85 
86 
87 /*
88  * Number of vnodes in existence.  Increased whenever getnewvnode()
89  * allocates a new vnode, never decreased.
90  */
91 static unsigned long	numvnodes;
92 
93 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
94 
95 /*
96  * Conversion tables for conversion from vnode types to inode formats
97  * and back.
98  */
99 enum vtype iftovt_tab[16] = {
100 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
101 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
102 };
103 int vttoif_tab[9] = {
104 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
105 	S_IFSOCK, S_IFIFO, S_IFMT,
106 };
107 
108 /*
109  * List of vnodes that are ready for recycling.
110  */
111 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
112 
113 /*
114  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
115  * getnewvnode() will return a newly allocated vnode.
116  */
117 static u_long wantfreevnodes = 25;
118 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
119 /* Number of vnodes in the free list. */
120 static u_long freevnodes;
121 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
122 
123 /*
124  * Various variables used for debugging the new implementation of
125  * reassignbuf().
126  * XXX these are probably of (very) limited utility now.
127  */
128 static int reassignbufcalls;
129 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
130 static int nameileafonly;
131 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
132 
133 #ifdef ENABLE_VFS_IOOPT
134 /* See NOTES for a description of this setting. */
135 int vfs_ioopt;
136 SYSCTL_INT(_vfs, OID_AUTO, ioopt, CTLFLAG_RW, &vfs_ioopt, 0, "");
137 #endif
138 
139 /*
140  * Cache for the mount type id assigned to NFS.  This is used for
141  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
142  */
143 int	nfs_mount_type = -1;
144 
145 /* To keep more than one thread at a time from running vfs_getnewfsid */
146 static struct mtx mntid_mtx;
147 
148 /*
149  * Lock for any access to the following:
150  *	vnode_free_list
151  *	numvnodes
152  *	freevnodes
153  */
154 static struct mtx vnode_free_list_mtx;
155 
156 /*
157  * For any iteration/modification of dev->si_hlist (linked through
158  * v_specnext)
159  */
160 static struct mtx spechash_mtx;
161 
162 /* Publicly exported FS */
163 struct nfs_public nfs_pub;
164 
165 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
166 static uma_zone_t vnode_zone;
167 static uma_zone_t vnodepoll_zone;
168 
169 /* Set to 1 to print out reclaim of active vnodes */
170 int	prtactive;
171 
172 /*
173  * The workitem queue.
174  *
175  * It is useful to delay writes of file data and filesystem metadata
176  * for tens of seconds so that quickly created and deleted files need
177  * not waste disk bandwidth being created and removed. To realize this,
178  * we append vnodes to a "workitem" queue. When running with a soft
179  * updates implementation, most pending metadata dependencies should
180  * not wait for more than a few seconds. Thus, mounted on block devices
181  * are delayed only about a half the time that file data is delayed.
182  * Similarly, directory updates are more critical, so are only delayed
183  * about a third the time that file data is delayed. Thus, there are
184  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
185  * one each second (driven off the filesystem syncer process). The
186  * syncer_delayno variable indicates the next queue that is to be processed.
187  * Items that need to be processed soon are placed in this queue:
188  *
189  *	syncer_workitem_pending[syncer_delayno]
190  *
191  * A delay of fifteen seconds is done by placing the request fifteen
192  * entries later in the queue:
193  *
194  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
195  *
196  */
197 static int syncer_delayno;
198 static long syncer_mask;
199 LIST_HEAD(synclist, vnode);
200 static struct synclist *syncer_workitem_pending;
201 /*
202  * The sync_mtx protects:
203  *	vp->v_synclist
204  *	syncer_delayno
205  *	syncer_workitem_pending
206  *	rushjob
207  */
208 static struct mtx sync_mtx;
209 
210 #define SYNCER_MAXDELAY		32
211 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
212 static int syncdelay = 30;		/* max time to delay syncing data */
213 static int filedelay = 30;		/* time to delay syncing files */
214 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
215 static int dirdelay = 29;		/* time to delay syncing directories */
216 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
217 static int metadelay = 28;		/* time to delay syncing metadata */
218 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
219 static int rushjob;		/* number of slots to run ASAP */
220 static int stat_rush_requests;	/* number of times I/O speeded up */
221 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
222 
223 /*
224  * Number of vnodes we want to exist at any one time.  This is mostly used
225  * to size hash tables in vnode-related code.  It is normally not used in
226  * getnewvnode(), as wantfreevnodes is normally nonzero.)
227  *
228  * XXX desiredvnodes is historical cruft and should not exist.
229  */
230 int desiredvnodes;
231 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
232     &desiredvnodes, 0, "Maximum number of vnodes");
233 static int minvnodes;
234 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
235     &minvnodes, 0, "Minimum number of vnodes");
236 static int vnlru_nowhere;
237 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0,
238     "Number of times the vnlru process ran without success");
239 
240 /* Hook for calling soft updates */
241 int (*softdep_process_worklist_hook)(struct mount *);
242 
243 /*
244  * This only exists to supress warnings from unlocked specfs accesses.  It is
245  * no longer ok to have an unlocked VFS.
246  */
247 #define IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
248 
249 /* Print lock violations */
250 int vfs_badlock_print = 1;
251 
252 /* Panic on violation */
253 int vfs_badlock_panic = 1;
254 
255 /* Check for interlock across VOPs */
256 int vfs_badlock_mutex = 1;
257 
258 static void
259 vfs_badlock(char *msg, char *str, struct vnode *vp)
260 {
261 	if (vfs_badlock_print)
262 		printf("%s: %p %s\n", str, vp, msg);
263 	if (vfs_badlock_panic)
264 		Debugger("Lock violation.\n");
265 }
266 
267 void
268 assert_vi_unlocked(struct vnode *vp, char *str)
269 {
270 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
271 		vfs_badlock("interlock is locked but should not be", str, vp);
272 }
273 
274 void
275 assert_vi_locked(struct vnode *vp, char *str)
276 {
277 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
278 		vfs_badlock("interlock is not locked but should be", str, vp);
279 }
280 
281 void
282 assert_vop_locked(struct vnode *vp, char *str)
283 {
284 	if (vp && !IGNORE_LOCK(vp) && !VOP_ISLOCKED(vp, NULL))
285 		vfs_badlock("is not locked but should be", str, vp);
286 }
287 
288 void
289 assert_vop_unlocked(struct vnode *vp, char *str)
290 {
291 	if (vp && !IGNORE_LOCK(vp) &&
292 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
293 		vfs_badlock("is locked but should not be", str, vp);
294 }
295 
296 void
297 assert_vop_elocked(struct vnode *vp, char *str)
298 {
299 	if (vp && !IGNORE_LOCK(vp) &&
300 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
301 		vfs_badlock("is not exclusive locked but should be", str, vp);
302 }
303 
304 void
305 assert_vop_elocked_other(struct vnode *vp, char *str)
306 {
307 	if (vp && !IGNORE_LOCK(vp) &&
308 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
309 		vfs_badlock("is not exclusive locked by another thread",
310 		    str, vp);
311 }
312 
313 void
314 assert_vop_slocked(struct vnode *vp, char *str)
315 {
316 	if (vp && !IGNORE_LOCK(vp) &&
317 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
318 		vfs_badlock("is not locked shared but should be", str, vp);
319 }
320 
321 void
322 vop_rename_pre(void *ap)
323 {
324 	struct vop_rename_args *a = ap;
325 
326 	if (a->a_tvp)
327 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
328 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
329 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
330 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
331 
332 	/* Check the source (from) */
333 	if (a->a_tdvp != a->a_fdvp)
334 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked.\n");
335 	if (a->a_tvp != a->a_fvp)
336 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked.\n");
337 
338 	/* Check the target */
339 	if (a->a_tvp)
340 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked.\n");
341 
342 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked.\n");
343 }
344 
345 void
346 vop_strategy_pre(void *ap)
347 {
348 	struct vop_strategy_args *a = ap;
349 	struct buf *bp;
350 
351 	bp = a->a_bp;
352 
353 	/*
354 	 * Cluster ops lock their component buffers but not the IO container.
355 	 */
356 	if ((bp->b_flags & B_CLUSTER) != 0)
357 		return;
358 
359 	if (BUF_REFCNT(bp) < 1) {
360 		if (vfs_badlock_print)
361 			printf("VOP_STRATEGY: bp is not locked but should be.\n");
362 		if (vfs_badlock_panic)
363 			Debugger("Lock violation.\n");
364 	}
365 }
366 
367 void
368 vop_lookup_pre(void *ap)
369 {
370 	struct vop_lookup_args *a = ap;
371 	struct vnode *dvp;
372 
373 	dvp = a->a_dvp;
374 
375 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
376 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
377 }
378 
379 void
380 vop_lookup_post(void *ap, int rc)
381 {
382 	struct vop_lookup_args *a = ap;
383 	struct componentname *cnp;
384 	struct vnode *dvp;
385 	struct vnode *vp;
386 	int flags;
387 
388 	dvp = a->a_dvp;
389 	cnp = a->a_cnp;
390 	vp = *(a->a_vpp);
391 	flags = cnp->cn_flags;
392 
393 
394 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
395 	/*
396 	 * If this is the last path component for this lookup and LOCPARENT
397 	 * is set, OR if there is an error the directory has to be locked.
398 	 */
399 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
400 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
401 	else if (rc != 0)
402 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
403 	else if (dvp != vp)
404 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
405 
406 	if (flags & PDIRUNLOCK)
407 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
408 }
409 
410 void
411 vop_unlock_pre(void *ap)
412 {
413 	struct vop_unlock_args *a = ap;
414 
415 	if (a->a_flags & LK_INTERLOCK)
416 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
417 
418 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
419 }
420 
421 void
422 vop_unlock_post(void *ap, int rc)
423 {
424 	struct vop_unlock_args *a = ap;
425 
426 	if (a->a_flags & LK_INTERLOCK)
427 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
428 }
429 
430 void
431 vop_lock_pre(void *ap)
432 {
433 	struct vop_lock_args *a = ap;
434 
435 	if ((a->a_flags & LK_INTERLOCK) == 0)
436 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
437 	else
438 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
439 }
440 
441 void
442 vop_lock_post(void *ap, int rc)
443 {
444 	struct vop_lock_args *a;
445 
446 	a = ap;
447 
448 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
449 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
450 }
451 
452 void
453 v_addpollinfo(struct vnode *vp)
454 {
455 	vp->v_pollinfo = uma_zalloc(vnodepoll_zone, M_WAITOK);
456 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
457 }
458 
459 /*
460  * Initialize the vnode management data structures.
461  */
462 static void
463 vntblinit(void *dummy __unused)
464 {
465 
466 	desiredvnodes = maxproc + cnt.v_page_count / 4;
467 	minvnodes = desiredvnodes / 4;
468 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
469 	mtx_init(&mntvnode_mtx, "mntvnode", NULL, MTX_DEF);
470 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
471 	mtx_init(&spechash_mtx, "spechash", NULL, MTX_DEF);
472 	TAILQ_INIT(&vnode_free_list);
473 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
474 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
475 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
476 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
477 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
478 	/*
479 	 * Initialize the filesystem syncer.
480 	 */
481 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
482 		&syncer_mask);
483 	syncer_maxdelay = syncer_mask + 1;
484 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
485 }
486 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
487 
488 
489 /*
490  * Mark a mount point as busy. Used to synchronize access and to delay
491  * unmounting. Interlock is not released on failure.
492  */
493 int
494 vfs_busy(mp, flags, interlkp, td)
495 	struct mount *mp;
496 	int flags;
497 	struct mtx *interlkp;
498 	struct thread *td;
499 {
500 	int lkflags;
501 
502 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
503 		if (flags & LK_NOWAIT)
504 			return (ENOENT);
505 		mp->mnt_kern_flag |= MNTK_MWAIT;
506 		/*
507 		 * Since all busy locks are shared except the exclusive
508 		 * lock granted when unmounting, the only place that a
509 		 * wakeup needs to be done is at the release of the
510 		 * exclusive lock at the end of dounmount.
511 		 */
512 		msleep(mp, interlkp, PVFS, "vfs_busy", 0);
513 		return (ENOENT);
514 	}
515 	lkflags = LK_SHARED | LK_NOPAUSE;
516 	if (interlkp)
517 		lkflags |= LK_INTERLOCK;
518 	if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td))
519 		panic("vfs_busy: unexpected lock failure");
520 	return (0);
521 }
522 
523 /*
524  * Free a busy filesystem.
525  */
526 void
527 vfs_unbusy(mp, td)
528 	struct mount *mp;
529 	struct thread *td;
530 {
531 
532 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
533 }
534 
535 /*
536  * Lookup a mount point by filesystem identifier.
537  */
538 struct mount *
539 vfs_getvfs(fsid)
540 	fsid_t *fsid;
541 {
542 	register struct mount *mp;
543 
544 	mtx_lock(&mountlist_mtx);
545 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
546 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
547 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
548 			mtx_unlock(&mountlist_mtx);
549 			return (mp);
550 		}
551 	}
552 	mtx_unlock(&mountlist_mtx);
553 	return ((struct mount *) 0);
554 }
555 
556 /*
557  * Get a new unique fsid.  Try to make its val[0] unique, since this value
558  * will be used to create fake device numbers for stat().  Also try (but
559  * not so hard) make its val[0] unique mod 2^16, since some emulators only
560  * support 16-bit device numbers.  We end up with unique val[0]'s for the
561  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
562  *
563  * Keep in mind that several mounts may be running in parallel.  Starting
564  * the search one past where the previous search terminated is both a
565  * micro-optimization and a defense against returning the same fsid to
566  * different mounts.
567  */
568 void
569 vfs_getnewfsid(mp)
570 	struct mount *mp;
571 {
572 	static u_int16_t mntid_base;
573 	fsid_t tfsid;
574 	int mtype;
575 
576 	mtx_lock(&mntid_mtx);
577 	mtype = mp->mnt_vfc->vfc_typenum;
578 	tfsid.val[1] = mtype;
579 	mtype = (mtype & 0xFF) << 24;
580 	for (;;) {
581 		tfsid.val[0] = makeudev(255,
582 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
583 		mntid_base++;
584 		if (vfs_getvfs(&tfsid) == NULL)
585 			break;
586 	}
587 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
588 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
589 	mtx_unlock(&mntid_mtx);
590 }
591 
592 /*
593  * Knob to control the precision of file timestamps:
594  *
595  *   0 = seconds only; nanoseconds zeroed.
596  *   1 = seconds and nanoseconds, accurate within 1/HZ.
597  *   2 = seconds and nanoseconds, truncated to microseconds.
598  * >=3 = seconds and nanoseconds, maximum precision.
599  */
600 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
601 
602 static int timestamp_precision = TSP_SEC;
603 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
604     &timestamp_precision, 0, "");
605 
606 /*
607  * Get a current timestamp.
608  */
609 void
610 vfs_timestamp(tsp)
611 	struct timespec *tsp;
612 {
613 	struct timeval tv;
614 
615 	switch (timestamp_precision) {
616 	case TSP_SEC:
617 		tsp->tv_sec = time_second;
618 		tsp->tv_nsec = 0;
619 		break;
620 	case TSP_HZ:
621 		getnanotime(tsp);
622 		break;
623 	case TSP_USEC:
624 		microtime(&tv);
625 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
626 		break;
627 	case TSP_NSEC:
628 	default:
629 		nanotime(tsp);
630 		break;
631 	}
632 }
633 
634 /*
635  * Set vnode attributes to VNOVAL
636  */
637 void
638 vattr_null(vap)
639 	register struct vattr *vap;
640 {
641 
642 	vap->va_type = VNON;
643 	vap->va_size = VNOVAL;
644 	vap->va_bytes = VNOVAL;
645 	vap->va_mode = VNOVAL;
646 	vap->va_nlink = VNOVAL;
647 	vap->va_uid = VNOVAL;
648 	vap->va_gid = VNOVAL;
649 	vap->va_fsid = VNOVAL;
650 	vap->va_fileid = VNOVAL;
651 	vap->va_blocksize = VNOVAL;
652 	vap->va_rdev = VNOVAL;
653 	vap->va_atime.tv_sec = VNOVAL;
654 	vap->va_atime.tv_nsec = VNOVAL;
655 	vap->va_mtime.tv_sec = VNOVAL;
656 	vap->va_mtime.tv_nsec = VNOVAL;
657 	vap->va_ctime.tv_sec = VNOVAL;
658 	vap->va_ctime.tv_nsec = VNOVAL;
659 	vap->va_birthtime.tv_sec = VNOVAL;
660 	vap->va_birthtime.tv_nsec = VNOVAL;
661 	vap->va_flags = VNOVAL;
662 	vap->va_gen = VNOVAL;
663 	vap->va_vaflags = 0;
664 }
665 
666 /*
667  * This routine is called when we have too many vnodes.  It attempts
668  * to free <count> vnodes and will potentially free vnodes that still
669  * have VM backing store (VM backing store is typically the cause
670  * of a vnode blowout so we want to do this).  Therefore, this operation
671  * is not considered cheap.
672  *
673  * A number of conditions may prevent a vnode from being reclaimed.
674  * the buffer cache may have references on the vnode, a directory
675  * vnode may still have references due to the namei cache representing
676  * underlying files, or the vnode may be in active use.   It is not
677  * desireable to reuse such vnodes.  These conditions may cause the
678  * number of vnodes to reach some minimum value regardless of what
679  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
680  */
681 static int
682 vlrureclaim(struct mount *mp, int count)
683 {
684 	struct vnode *vp;
685 	int done;
686 	int trigger;
687 	int usevnodes;
688 
689 	/*
690 	 * Calculate the trigger point, don't allow user
691 	 * screwups to blow us up.   This prevents us from
692 	 * recycling vnodes with lots of resident pages.  We
693 	 * aren't trying to free memory, we are trying to
694 	 * free vnodes.
695 	 */
696 	usevnodes = desiredvnodes;
697 	if (usevnodes <= 0)
698 		usevnodes = 1;
699 	trigger = cnt.v_page_count * 2 / usevnodes;
700 
701 	done = 0;
702 	mtx_lock(&mntvnode_mtx);
703 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
704 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
705 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
706 
707 		if (vp->v_type != VNON &&
708 		    vp->v_type != VBAD &&
709 		    VI_TRYLOCK(vp)) {
710 			if (VMIGHTFREE(vp) &&           /* critical path opt */
711 			    (vp->v_object == NULL ||
712 			    vp->v_object->resident_page_count < trigger)) {
713 				mtx_unlock(&mntvnode_mtx);
714 				vgonel(vp, curthread);
715 				done++;
716 				mtx_lock(&mntvnode_mtx);
717 			} else
718 				VI_UNLOCK(vp);
719 		}
720 		--count;
721 	}
722 	mtx_unlock(&mntvnode_mtx);
723 	return done;
724 }
725 
726 /*
727  * Attempt to recycle vnodes in a context that is always safe to block.
728  * Calling vlrurecycle() from the bowels of filesystem code has some
729  * interesting deadlock problems.
730  */
731 static struct proc *vnlruproc;
732 static int vnlruproc_sig;
733 
734 static void
735 vnlru_proc(void)
736 {
737 	struct mount *mp, *nmp;
738 	int s;
739 	int done;
740 	struct proc *p = vnlruproc;
741 	struct thread *td = FIRST_THREAD_IN_PROC(p);	/* XXXKSE */
742 
743 	mtx_lock(&Giant);
744 
745 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
746 	    SHUTDOWN_PRI_FIRST);
747 
748 	s = splbio();
749 	for (;;) {
750 		kthread_suspend_check(p);
751 		mtx_lock(&vnode_free_list_mtx);
752 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
753 			mtx_unlock(&vnode_free_list_mtx);
754 			vnlruproc_sig = 0;
755 			tsleep(vnlruproc, PVFS, "vlruwt", 0);
756 			continue;
757 		}
758 		mtx_unlock(&vnode_free_list_mtx);
759 		done = 0;
760 		mtx_lock(&mountlist_mtx);
761 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
762 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
763 				nmp = TAILQ_NEXT(mp, mnt_list);
764 				continue;
765 			}
766 			done += vlrureclaim(mp, 10);
767 			mtx_lock(&mountlist_mtx);
768 			nmp = TAILQ_NEXT(mp, mnt_list);
769 			vfs_unbusy(mp, td);
770 		}
771 		mtx_unlock(&mountlist_mtx);
772 		if (done == 0) {
773 #if 0
774 			/* These messages are temporary debugging aids */
775 			if (vnlru_nowhere < 5)
776 				printf("vnlru process getting nowhere..\n");
777 			else if (vnlru_nowhere == 5)
778 				printf("vnlru process messages stopped.\n");
779 #endif
780 			vnlru_nowhere++;
781 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
782 		}
783 	}
784 	splx(s);
785 }
786 
787 static struct kproc_desc vnlru_kp = {
788 	"vnlru",
789 	vnlru_proc,
790 	&vnlruproc
791 };
792 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
793 
794 
795 /*
796  * Routines having to do with the management of the vnode table.
797  */
798 
799 /*
800  * Check to see if a free vnode can be recycled. If it can,
801  * return it locked with the vn lock, but not interlock. Also
802  * get the vn_start_write lock. Otherwise indicate the error.
803  */
804 static int
805 vcanrecycle(struct vnode *vp, struct mount **vnmpp)
806 {
807 	struct thread *td = curthread;
808 	vm_object_t object;
809 	int error;
810 
811 	/* Don't recycle if we can't get the interlock */
812 	if (!VI_TRYLOCK(vp))
813 		return (EWOULDBLOCK);
814 
815 	/* We should be able to immediately acquire this */
816 	/* XXX This looks like it should panic if it fails */
817 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td) != 0) {
818 		if (VOP_ISLOCKED(vp, td))
819 			panic("vcanrecycle: locked vnode");
820 		return (EWOULDBLOCK);
821 	}
822 
823 	/*
824 	 * Don't recycle if its filesystem is being suspended.
825 	 */
826 	if (vn_start_write(vp, vnmpp, V_NOWAIT) != 0) {
827 		error = EBUSY;
828 		goto done;
829 	}
830 
831 	/*
832 	 * Don't recycle if we still have cached pages.
833 	 */
834 	if (VOP_GETVOBJECT(vp, &object) == 0 &&
835 	     (object->resident_page_count ||
836 	      object->ref_count)) {
837 		error = EBUSY;
838 		goto done;
839 	}
840 	if (LIST_FIRST(&vp->v_cache_src)) {
841 		/*
842 		 * note: nameileafonly sysctl is temporary,
843 		 * for debugging only, and will eventually be
844 		 * removed.
845 		 */
846 		if (nameileafonly > 0) {
847 			/*
848 			 * Do not reuse namei-cached directory
849 			 * vnodes that have cached
850 			 * subdirectories.
851 			 */
852 			if (cache_leaf_test(vp) < 0) {
853 				error = EISDIR;
854 				goto done;
855 			}
856 		} else if (nameileafonly < 0 ||
857 			    vmiodirenable == 0) {
858 			/*
859 			 * Do not reuse namei-cached directory
860 			 * vnodes if nameileafonly is -1 or
861 			 * if VMIO backing for directories is
862 			 * turned off (otherwise we reuse them
863 			 * too quickly).
864 			 */
865 			error = EBUSY;
866 			goto done;
867 		}
868 	}
869 	return (0);
870 done:
871 	VOP_UNLOCK(vp, 0, td);
872 	return (error);
873 }
874 
875 /*
876  * Return the next vnode from the free list.
877  */
878 int
879 getnewvnode(tag, mp, vops, vpp)
880 	const char *tag;
881 	struct mount *mp;
882 	vop_t **vops;
883 	struct vnode **vpp;
884 {
885 	int s;
886 	struct thread *td = curthread;	/* XXX */
887 	struct vnode *vp = NULL;
888 	struct mount *vnmp;
889 
890 	s = splbio();
891 	mtx_lock(&vnode_free_list_mtx);
892 
893 	/*
894 	 * Try to reuse vnodes if we hit the max.  This situation only
895 	 * occurs in certain large-memory (2G+) situations.  We cannot
896 	 * attempt to directly reclaim vnodes due to nasty recursion
897 	 * problems.
898 	 */
899 	if (vnlruproc_sig == 0 && numvnodes - freevnodes > desiredvnodes) {
900 		vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
901 		wakeup(vnlruproc);
902 	}
903 
904 	/*
905 	 * Attempt to reuse a vnode already on the free list, allocating
906 	 * a new vnode if we can't find one or if we have not reached a
907 	 * good minimum for good LRU performance.
908 	 */
909 
910 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
911 		int error;
912 		int count;
913 
914 		for (count = 0; count < freevnodes; count++) {
915 			vp = TAILQ_FIRST(&vnode_free_list);
916 
917 			KASSERT(vp->v_usecount == 0,
918 			    ("getnewvnode: free vnode isn't"));
919 
920 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
921 			/*
922 			 * We have to drop the free list mtx to avoid lock
923 			 * order reversals with interlock.
924 			 */
925 			mtx_unlock(&vnode_free_list_mtx);
926 			error = vcanrecycle(vp, &vnmp);
927 			mtx_lock(&vnode_free_list_mtx);
928 			if (error == 0)
929 				break;
930 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
931 			vp = NULL;
932 		}
933 	}
934 	if (vp) {
935 		freevnodes--;
936 		mtx_unlock(&vnode_free_list_mtx);
937 
938 		cache_purge(vp);
939 		VI_LOCK(vp);
940 		vp->v_iflag |= VI_DOOMED;
941 		vp->v_iflag &= ~VI_FREE;
942 		if (vp->v_type != VBAD) {
943 			VOP_UNLOCK(vp, 0, td);
944 			vgonel(vp, td);
945 			VI_LOCK(vp);
946 		} else {
947 			VOP_UNLOCK(vp, 0, td);
948 		}
949 		vn_finished_write(vnmp);
950 
951 #ifdef INVARIANTS
952 		{
953 			if (vp->v_data)
954 				panic("cleaned vnode isn't");
955 			if (vp->v_numoutput)
956 				panic("Clean vnode has pending I/O's");
957 			if (vp->v_writecount != 0)
958 				panic("Non-zero write count");
959 		}
960 #endif
961 		if (vp->v_pollinfo) {
962 			mtx_destroy(&vp->v_pollinfo->vpi_lock);
963 			uma_zfree(vnodepoll_zone, vp->v_pollinfo);
964 		}
965 		vp->v_pollinfo = NULL;
966 #ifdef MAC
967 		mac_destroy_vnode(vp);
968 #endif
969 		vp->v_iflag = 0;
970 		vp->v_vflag = 0;
971 		vp->v_lastw = 0;
972 		vp->v_lasta = 0;
973 		vp->v_cstart = 0;
974 		vp->v_clen = 0;
975 		vp->v_socket = 0;
976 		lockdestroy(vp->v_vnlock);
977 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
978 		KASSERT(vp->v_cleanblkroot == NULL, ("cleanblkroot not NULL"));
979 		KASSERT(vp->v_dirtyblkroot == NULL, ("dirtyblkroot not NULL"));
980 	} else {
981 		numvnodes++;
982 		mtx_unlock(&vnode_free_list_mtx);
983 
984 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
985 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
986 		VI_LOCK(vp);
987 		vp->v_dd = vp;
988 		vp->v_vnlock = &vp->v_lock;
989 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
990 		cache_purge(vp);
991 		LIST_INIT(&vp->v_cache_src);
992 		TAILQ_INIT(&vp->v_cache_dst);
993 	}
994 
995 	TAILQ_INIT(&vp->v_cleanblkhd);
996 	TAILQ_INIT(&vp->v_dirtyblkhd);
997 	vp->v_type = VNON;
998 	vp->v_tag = tag;
999 	vp->v_op = vops;
1000 	*vpp = vp;
1001 	vp->v_usecount = 1;
1002 	vp->v_data = 0;
1003 	vp->v_cachedid = -1;
1004 	VI_UNLOCK(vp);
1005 #ifdef MAC
1006 	mac_init_vnode(vp);
1007 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1008 		mac_associate_vnode_singlelabel(mp, vp);
1009 #endif
1010 	insmntque(vp, mp);
1011 
1012 	return (0);
1013 }
1014 
1015 /*
1016  * Move a vnode from one mount queue to another.
1017  */
1018 static void
1019 insmntque(vp, mp)
1020 	register struct vnode *vp;
1021 	register struct mount *mp;
1022 {
1023 
1024 	mtx_lock(&mntvnode_mtx);
1025 	/*
1026 	 * Delete from old mount point vnode list, if on one.
1027 	 */
1028 	if (vp->v_mount != NULL)
1029 		TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
1030 	/*
1031 	 * Insert into list of vnodes for the new mount point, if available.
1032 	 */
1033 	if ((vp->v_mount = mp) == NULL) {
1034 		mtx_unlock(&mntvnode_mtx);
1035 		return;
1036 	}
1037 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1038 	mtx_unlock(&mntvnode_mtx);
1039 }
1040 
1041 /*
1042  * Update outstanding I/O count and do wakeup if requested.
1043  */
1044 void
1045 vwakeup(bp)
1046 	register struct buf *bp;
1047 {
1048 	register struct vnode *vp;
1049 
1050 	bp->b_flags &= ~B_WRITEINPROG;
1051 	if ((vp = bp->b_vp)) {
1052 		VI_LOCK(vp);
1053 		vp->v_numoutput--;
1054 		if (vp->v_numoutput < 0)
1055 			panic("vwakeup: neg numoutput");
1056 		if ((vp->v_numoutput == 0) && (vp->v_iflag & VI_BWAIT)) {
1057 			vp->v_iflag &= ~VI_BWAIT;
1058 			wakeup(&vp->v_numoutput);
1059 		}
1060 		VI_UNLOCK(vp);
1061 	}
1062 }
1063 
1064 /*
1065  * Flush out and invalidate all buffers associated with a vnode.
1066  * Called with the underlying object locked.
1067  */
1068 int
1069 vinvalbuf(vp, flags, cred, td, slpflag, slptimeo)
1070 	struct vnode *vp;
1071 	int flags;
1072 	struct ucred *cred;
1073 	struct thread *td;
1074 	int slpflag, slptimeo;
1075 {
1076 	struct buf *blist;
1077 	int s, error;
1078 	vm_object_t object;
1079 
1080 	GIANT_REQUIRED;
1081 
1082 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1083 
1084 	VI_LOCK(vp);
1085 	if (flags & V_SAVE) {
1086 		s = splbio();
1087 		while (vp->v_numoutput) {
1088 			vp->v_iflag |= VI_BWAIT;
1089 			error = msleep(&vp->v_numoutput, VI_MTX(vp),
1090 			    slpflag | (PRIBIO + 1), "vinvlbuf", slptimeo);
1091 			if (error) {
1092 				VI_UNLOCK(vp);
1093 				splx(s);
1094 				return (error);
1095 			}
1096 		}
1097 		if (!TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1098 			splx(s);
1099 			VI_UNLOCK(vp);
1100 			if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)) != 0)
1101 				return (error);
1102 			/*
1103 			 * XXX We could save a lock/unlock if this was only
1104 			 * enabled under INVARIANTS
1105 			 */
1106 			VI_LOCK(vp);
1107 			s = splbio();
1108 			if (vp->v_numoutput > 0 ||
1109 			    !TAILQ_EMPTY(&vp->v_dirtyblkhd))
1110 				panic("vinvalbuf: dirty bufs");
1111 		}
1112 		splx(s);
1113 	}
1114 	s = splbio();
1115 	/*
1116 	 * If you alter this loop please notice that interlock is dropped and
1117 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1118 	 * no race conditions occur from this.
1119 	 */
1120 	for (error = 0;;) {
1121 		if ((blist = TAILQ_FIRST(&vp->v_cleanblkhd)) != 0 &&
1122 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
1123 			if (error)
1124 				break;
1125 			continue;
1126 		}
1127 		if ((blist = TAILQ_FIRST(&vp->v_dirtyblkhd)) != 0 &&
1128 		    flushbuflist(blist, flags, vp, slpflag, slptimeo, &error)) {
1129 			if (error)
1130 				break;
1131 			continue;
1132 		}
1133 		break;
1134 	}
1135 	if (error) {
1136 		splx(s);
1137 		VI_UNLOCK(vp);
1138 		return (error);
1139 	}
1140 
1141 	/*
1142 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1143 	 * have write I/O in-progress but if there is a VM object then the
1144 	 * VM object can also have read-I/O in-progress.
1145 	 */
1146 	do {
1147 		while (vp->v_numoutput > 0) {
1148 			vp->v_iflag |= VI_BWAIT;
1149 			msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vnvlbv", 0);
1150 		}
1151 		VI_UNLOCK(vp);
1152 		if (VOP_GETVOBJECT(vp, &object) == 0) {
1153 			while (object->paging_in_progress)
1154 			vm_object_pip_sleep(object, "vnvlbx");
1155 		}
1156 		VI_LOCK(vp);
1157 	} while (vp->v_numoutput > 0);
1158 	VI_UNLOCK(vp);
1159 
1160 	splx(s);
1161 
1162 	/*
1163 	 * Destroy the copy in the VM cache, too.
1164 	 */
1165 	if (VOP_GETVOBJECT(vp, &object) == 0) {
1166 		vm_object_page_remove(object, 0, 0,
1167 			(flags & V_SAVE) ? TRUE : FALSE);
1168 	}
1169 
1170 #ifdef INVARIANTS
1171 	VI_LOCK(vp);
1172 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1173 	    (!TAILQ_EMPTY(&vp->v_dirtyblkhd) ||
1174 	     !TAILQ_EMPTY(&vp->v_cleanblkhd)))
1175 		panic("vinvalbuf: flush failed");
1176 	VI_UNLOCK(vp);
1177 #endif
1178 	return (0);
1179 }
1180 
1181 /*
1182  * Flush out buffers on the specified list.
1183  *
1184  */
1185 static int
1186 flushbuflist(blist, flags, vp, slpflag, slptimeo, errorp)
1187 	struct buf *blist;
1188 	int flags;
1189 	struct vnode *vp;
1190 	int slpflag, slptimeo;
1191 	int *errorp;
1192 {
1193 	struct buf *bp, *nbp;
1194 	int found, error;
1195 
1196 	ASSERT_VI_LOCKED(vp, "flushbuflist");
1197 
1198 	for (found = 0, bp = blist; bp; bp = nbp) {
1199 		nbp = TAILQ_NEXT(bp, b_vnbufs);
1200 		VI_UNLOCK(vp);
1201 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1202 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1203 			VI_LOCK(vp);
1204 			continue;
1205 		}
1206 		found += 1;
1207 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1208 			error = BUF_TIMELOCK(bp,
1209 			    LK_EXCLUSIVE | LK_SLEEPFAIL,
1210 			    "flushbuf", slpflag, slptimeo);
1211 			if (error != ENOLCK)
1212 				*errorp = error;
1213 			goto done;
1214 		}
1215 		/*
1216 		 * XXX Since there are no node locks for NFS, I
1217 		 * believe there is a slight chance that a delayed
1218 		 * write will occur while sleeping just above, so
1219 		 * check for it.  Note that vfs_bio_awrite expects
1220 		 * buffers to reside on a queue, while BUF_WRITE and
1221 		 * brelse do not.
1222 		 */
1223 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1224 			(flags & V_SAVE)) {
1225 
1226 			if (bp->b_vp == vp) {
1227 				if (bp->b_flags & B_CLUSTEROK) {
1228 					BUF_UNLOCK(bp);
1229 					vfs_bio_awrite(bp);
1230 				} else {
1231 					bremfree(bp);
1232 					bp->b_flags |= B_ASYNC;
1233 					BUF_WRITE(bp);
1234 				}
1235 			} else {
1236 				bremfree(bp);
1237 				(void) BUF_WRITE(bp);
1238 			}
1239 			goto done;
1240 		}
1241 		bremfree(bp);
1242 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1243 		bp->b_flags &= ~B_ASYNC;
1244 		brelse(bp);
1245 		VI_LOCK(vp);
1246 	}
1247 	return (found);
1248 done:
1249 	VI_LOCK(vp);
1250 	return (found);
1251 }
1252 
1253 /*
1254  * Truncate a file's buffer and pages to a specified length.  This
1255  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1256  * sync activity.
1257  */
1258 int
1259 vtruncbuf(vp, cred, td, length, blksize)
1260 	register struct vnode *vp;
1261 	struct ucred *cred;
1262 	struct thread *td;
1263 	off_t length;
1264 	int blksize;
1265 {
1266 	register struct buf *bp;
1267 	struct buf *nbp;
1268 	int s, anyfreed;
1269 	int trunclbn;
1270 
1271 	/*
1272 	 * Round up to the *next* lbn.
1273 	 */
1274 	trunclbn = (length + blksize - 1) / blksize;
1275 
1276 	s = splbio();
1277 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1278 restart:
1279 	VI_LOCK(vp);
1280 	anyfreed = 1;
1281 	for (;anyfreed;) {
1282 		anyfreed = 0;
1283 		for (bp = TAILQ_FIRST(&vp->v_cleanblkhd); bp; bp = nbp) {
1284 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1285 			VI_UNLOCK(vp);
1286 			if (bp->b_lblkno >= trunclbn) {
1287 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1288 					BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL);
1289 					goto restart;
1290 				} else {
1291 					bremfree(bp);
1292 					bp->b_flags |= (B_INVAL | B_RELBUF);
1293 					bp->b_flags &= ~B_ASYNC;
1294 					brelse(bp);
1295 					anyfreed = 1;
1296 				}
1297 				if (nbp &&
1298 				    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1299 				    (nbp->b_vp != vp) ||
1300 				    (nbp->b_flags & B_DELWRI))) {
1301 					goto restart;
1302 				}
1303 			}
1304 			VI_LOCK(vp);
1305 		}
1306 
1307 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1308 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1309 			VI_UNLOCK(vp);
1310 			if (bp->b_lblkno >= trunclbn) {
1311 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1312 					BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL);
1313 					goto restart;
1314 				} else {
1315 					bremfree(bp);
1316 					bp->b_flags |= (B_INVAL | B_RELBUF);
1317 					bp->b_flags &= ~B_ASYNC;
1318 					brelse(bp);
1319 					anyfreed = 1;
1320 				}
1321 				if (nbp &&
1322 				    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1323 				    (nbp->b_vp != vp) ||
1324 				    (nbp->b_flags & B_DELWRI) == 0)) {
1325 					goto restart;
1326 				}
1327 			}
1328 			VI_LOCK(vp);
1329 		}
1330 	}
1331 
1332 	if (length > 0) {
1333 restartsync:
1334 		for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
1335 			nbp = TAILQ_NEXT(bp, b_vnbufs);
1336 			VI_UNLOCK(vp);
1337 			if ((bp->b_flags & B_DELWRI) && (bp->b_lblkno < 0)) {
1338 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
1339 					BUF_LOCK(bp, LK_EXCLUSIVE|LK_SLEEPFAIL);
1340 					goto restart;
1341 				} else {
1342 					bremfree(bp);
1343 					if (bp->b_vp == vp) {
1344 						bp->b_flags |= B_ASYNC;
1345 					} else {
1346 						bp->b_flags &= ~B_ASYNC;
1347 					}
1348 					BUF_WRITE(bp);
1349 				}
1350 				VI_LOCK(vp);
1351 				goto restartsync;
1352 			}
1353 			VI_LOCK(vp);
1354 		}
1355 	}
1356 
1357 	while (vp->v_numoutput > 0) {
1358 		vp->v_iflag |= VI_BWAIT;
1359 		msleep(&vp->v_numoutput, VI_MTX(vp), PVM, "vbtrunc", 0);
1360 	}
1361 	VI_UNLOCK(vp);
1362 	splx(s);
1363 
1364 	vnode_pager_setsize(vp, length);
1365 
1366 	return (0);
1367 }
1368 
1369 /*
1370  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1371  * 		 a vnode.
1372  *
1373  *	NOTE: We have to deal with the special case of a background bitmap
1374  *	buffer, a situation where two buffers will have the same logical
1375  *	block offset.  We want (1) only the foreground buffer to be accessed
1376  *	in a lookup and (2) must differentiate between the foreground and
1377  *	background buffer in the splay tree algorithm because the splay
1378  *	tree cannot normally handle multiple entities with the same 'index'.
1379  *	We accomplish this by adding differentiating flags to the splay tree's
1380  *	numerical domain.
1381  */
1382 static
1383 struct buf *
1384 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1385 {
1386 	struct buf dummy;
1387 	struct buf *lefttreemax, *righttreemin, *y;
1388 
1389 	if (root == NULL)
1390 		return (NULL);
1391 	lefttreemax = righttreemin = &dummy;
1392 	for (;;) {
1393 		if (lblkno < root->b_lblkno ||
1394 		    (lblkno == root->b_lblkno &&
1395 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1396 			if ((y = root->b_left) == NULL)
1397 				break;
1398 			if (lblkno < y->b_lblkno) {
1399 				/* Rotate right. */
1400 				root->b_left = y->b_right;
1401 				y->b_right = root;
1402 				root = y;
1403 				if ((y = root->b_left) == NULL)
1404 					break;
1405 			}
1406 			/* Link into the new root's right tree. */
1407 			righttreemin->b_left = root;
1408 			righttreemin = root;
1409 		} else if (lblkno > root->b_lblkno ||
1410 		    (lblkno == root->b_lblkno &&
1411 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1412 			if ((y = root->b_right) == NULL)
1413 				break;
1414 			if (lblkno > y->b_lblkno) {
1415 				/* Rotate left. */
1416 				root->b_right = y->b_left;
1417 				y->b_left = root;
1418 				root = y;
1419 				if ((y = root->b_right) == NULL)
1420 					break;
1421 			}
1422 			/* Link into the new root's left tree. */
1423 			lefttreemax->b_right = root;
1424 			lefttreemax = root;
1425 		} else {
1426 			break;
1427 		}
1428 		root = y;
1429 	}
1430 	/* Assemble the new root. */
1431 	lefttreemax->b_right = root->b_left;
1432 	righttreemin->b_left = root->b_right;
1433 	root->b_left = dummy.b_right;
1434 	root->b_right = dummy.b_left;
1435 	return (root);
1436 }
1437 
1438 static
1439 void
1440 buf_vlist_remove(struct buf *bp)
1441 {
1442 	struct vnode *vp = bp->b_vp;
1443 	struct buf *root;
1444 
1445 	ASSERT_VI_LOCKED(vp, "buf_vlist_remove");
1446 	if (bp->b_xflags & BX_VNDIRTY) {
1447 		if (bp != vp->v_dirtyblkroot) {
1448 			root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1449 			KASSERT(root == bp, ("splay lookup failed during dirty remove"));
1450 		}
1451 		if (bp->b_left == NULL) {
1452 			root = bp->b_right;
1453 		} else {
1454 			root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1455 			root->b_right = bp->b_right;
1456 		}
1457 		vp->v_dirtyblkroot = root;
1458 		TAILQ_REMOVE(&vp->v_dirtyblkhd, bp, b_vnbufs);
1459 	} else {
1460 		/* KASSERT(bp->b_xflags & BX_VNCLEAN, ("bp wasn't clean")); */
1461 		if (bp != vp->v_cleanblkroot) {
1462 			root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1463 			KASSERT(root == bp, ("splay lookup failed during clean remove"));
1464 		}
1465 		if (bp->b_left == NULL) {
1466 			root = bp->b_right;
1467 		} else {
1468 			root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1469 			root->b_right = bp->b_right;
1470 		}
1471 		vp->v_cleanblkroot = root;
1472 		TAILQ_REMOVE(&vp->v_cleanblkhd, bp, b_vnbufs);
1473 	}
1474 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1475 }
1476 
1477 /*
1478  * Add the buffer to the sorted clean or dirty block list using a
1479  * splay tree algorithm.
1480  *
1481  * NOTE: xflags is passed as a constant, optimizing this inline function!
1482  */
1483 static
1484 void
1485 buf_vlist_add(struct buf *bp, struct vnode *vp, b_xflags_t xflags)
1486 {
1487 	struct buf *root;
1488 
1489 	ASSERT_VI_LOCKED(vp, "buf_vlist_add");
1490 	bp->b_xflags |= xflags;
1491 	if (xflags & BX_VNDIRTY) {
1492 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_dirtyblkroot);
1493 		if (root == NULL) {
1494 			bp->b_left = NULL;
1495 			bp->b_right = NULL;
1496 			TAILQ_INSERT_TAIL(&vp->v_dirtyblkhd, bp, b_vnbufs);
1497 		} else if (bp->b_lblkno < root->b_lblkno ||
1498 		    (bp->b_lblkno == root->b_lblkno &&
1499 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1500 			bp->b_left = root->b_left;
1501 			bp->b_right = root;
1502 			root->b_left = NULL;
1503 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1504 		} else {
1505 			bp->b_right = root->b_right;
1506 			bp->b_left = root;
1507 			root->b_right = NULL;
1508 			TAILQ_INSERT_AFTER(&vp->v_dirtyblkhd,
1509 			    root, bp, b_vnbufs);
1510 		}
1511 		vp->v_dirtyblkroot = bp;
1512 	} else {
1513 		/* KASSERT(xflags & BX_VNCLEAN, ("xflags not clean")); */
1514 		root = buf_splay(bp->b_lblkno, bp->b_xflags, vp->v_cleanblkroot);
1515 		if (root == NULL) {
1516 			bp->b_left = NULL;
1517 			bp->b_right = NULL;
1518 			TAILQ_INSERT_TAIL(&vp->v_cleanblkhd, bp, b_vnbufs);
1519 		} else if (bp->b_lblkno < root->b_lblkno ||
1520 		    (bp->b_lblkno == root->b_lblkno &&
1521 		    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1522 			bp->b_left = root->b_left;
1523 			bp->b_right = root;
1524 			root->b_left = NULL;
1525 			TAILQ_INSERT_BEFORE(root, bp, b_vnbufs);
1526 		} else {
1527 			bp->b_right = root->b_right;
1528 			bp->b_left = root;
1529 			root->b_right = NULL;
1530 			TAILQ_INSERT_AFTER(&vp->v_cleanblkhd,
1531 			    root, bp, b_vnbufs);
1532 		}
1533 		vp->v_cleanblkroot = bp;
1534 	}
1535 }
1536 
1537 #ifndef USE_BUFHASH
1538 
1539 /*
1540  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1541  * shadow buffers used in background bitmap writes.
1542  *
1543  * This code isn't quite efficient as it could be because we are maintaining
1544  * two sorted lists and do not know which list the block resides in.
1545  */
1546 struct buf *
1547 gbincore(struct vnode *vp, daddr_t lblkno)
1548 {
1549 	struct buf *bp;
1550 
1551 	GIANT_REQUIRED;
1552 
1553 	ASSERT_VI_LOCKED(vp, "gbincore");
1554 	bp = vp->v_cleanblkroot = buf_splay(lblkno, 0, vp->v_cleanblkroot);
1555 	if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1556 		return(bp);
1557 	bp = vp->v_dirtyblkroot = buf_splay(lblkno, 0, vp->v_dirtyblkroot);
1558 	if (bp && bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1559 		return(bp);
1560 	return(NULL);
1561 }
1562 
1563 #endif
1564 
1565 /*
1566  * Associate a buffer with a vnode.
1567  */
1568 void
1569 bgetvp(vp, bp)
1570 	register struct vnode *vp;
1571 	register struct buf *bp;
1572 {
1573 	int s;
1574 
1575 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1576 
1577 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1578 	    ("bgetvp: bp already attached! %p", bp));
1579 
1580 	VI_LOCK(vp);
1581 	vholdl(vp);
1582 	bp->b_vp = vp;
1583 	bp->b_dev = vn_todev(vp);
1584 	/*
1585 	 * Insert onto list for new vnode.
1586 	 */
1587 	s = splbio();
1588 	buf_vlist_add(bp, vp, BX_VNCLEAN);
1589 	splx(s);
1590 	VI_UNLOCK(vp);
1591 }
1592 
1593 /*
1594  * Disassociate a buffer from a vnode.
1595  */
1596 void
1597 brelvp(bp)
1598 	register struct buf *bp;
1599 {
1600 	struct vnode *vp;
1601 	int s;
1602 
1603 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1604 
1605 	/*
1606 	 * Delete from old vnode list, if on one.
1607 	 */
1608 	vp = bp->b_vp;
1609 	s = splbio();
1610 	VI_LOCK(vp);
1611 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1612 		buf_vlist_remove(bp);
1613 	if ((vp->v_iflag & VI_ONWORKLST) && TAILQ_EMPTY(&vp->v_dirtyblkhd)) {
1614 		vp->v_iflag &= ~VI_ONWORKLST;
1615 		mtx_lock(&sync_mtx);
1616 		LIST_REMOVE(vp, v_synclist);
1617 		mtx_unlock(&sync_mtx);
1618 	}
1619 	vdropl(vp);
1620 	VI_UNLOCK(vp);
1621 	bp->b_vp = (struct vnode *) 0;
1622 	if (bp->b_object)
1623 		bp->b_object = NULL;
1624 	splx(s);
1625 }
1626 
1627 /*
1628  * Add an item to the syncer work queue.
1629  */
1630 static void
1631 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
1632 {
1633 	int s, slot;
1634 
1635 	s = splbio();
1636 	ASSERT_VI_LOCKED(vp, "vn_syncer_add_to_worklist");
1637 
1638 	mtx_lock(&sync_mtx);
1639 	if (vp->v_iflag & VI_ONWORKLST)
1640 		LIST_REMOVE(vp, v_synclist);
1641 	else
1642 		vp->v_iflag |= VI_ONWORKLST;
1643 
1644 	if (delay > syncer_maxdelay - 2)
1645 		delay = syncer_maxdelay - 2;
1646 	slot = (syncer_delayno + delay) & syncer_mask;
1647 
1648 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
1649 	mtx_unlock(&sync_mtx);
1650 
1651 	splx(s);
1652 }
1653 
1654 struct  proc *updateproc;
1655 static void sched_sync(void);
1656 static struct kproc_desc up_kp = {
1657 	"syncer",
1658 	sched_sync,
1659 	&updateproc
1660 };
1661 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1662 
1663 /*
1664  * System filesystem synchronizer daemon.
1665  */
1666 static void
1667 sched_sync(void)
1668 {
1669 	struct synclist *slp;
1670 	struct vnode *vp;
1671 	struct mount *mp;
1672 	long starttime;
1673 	int s;
1674 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);  /* XXXKSE */
1675 
1676 	mtx_lock(&Giant);
1677 
1678 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, td->td_proc,
1679 	    SHUTDOWN_PRI_LAST);
1680 
1681 	for (;;) {
1682 		kthread_suspend_check(td->td_proc);
1683 
1684 		starttime = time_second;
1685 
1686 		/*
1687 		 * Push files whose dirty time has expired.  Be careful
1688 		 * of interrupt race on slp queue.
1689 		 */
1690 		s = splbio();
1691 		mtx_lock(&sync_mtx);
1692 		slp = &syncer_workitem_pending[syncer_delayno];
1693 		syncer_delayno += 1;
1694 		if (syncer_delayno == syncer_maxdelay)
1695 			syncer_delayno = 0;
1696 		splx(s);
1697 
1698 		while ((vp = LIST_FIRST(slp)) != NULL) {
1699 			mtx_unlock(&sync_mtx);
1700 			if (VOP_ISLOCKED(vp, NULL) == 0 &&
1701 			    vn_start_write(vp, &mp, V_NOWAIT) == 0) {
1702 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1703 				(void) VOP_FSYNC(vp, td->td_ucred, MNT_LAZY, td);
1704 				VOP_UNLOCK(vp, 0, td);
1705 				vn_finished_write(mp);
1706 			}
1707 			s = splbio();
1708 			mtx_lock(&sync_mtx);
1709 			if (LIST_FIRST(slp) == vp) {
1710 				mtx_unlock(&sync_mtx);
1711 				/*
1712 				 * Note: VFS vnodes can remain on the
1713 				 * worklist too with no dirty blocks, but
1714 				 * since sync_fsync() moves it to a different
1715 				 * slot we are safe.
1716 				 */
1717 				VI_LOCK(vp);
1718 				if (TAILQ_EMPTY(&vp->v_dirtyblkhd) &&
1719 				    !vn_isdisk(vp, NULL)) {
1720 					panic("sched_sync: fsync failed "
1721 					      "vp %p tag %s", vp, vp->v_tag);
1722 				}
1723 				/*
1724 				 * Put us back on the worklist.  The worklist
1725 				 * routine will remove us from our current
1726 				 * position and then add us back in at a later
1727 				 * position.
1728 				 */
1729 				vn_syncer_add_to_worklist(vp, syncdelay);
1730 				VI_UNLOCK(vp);
1731 				mtx_lock(&sync_mtx);
1732 			}
1733 			splx(s);
1734 		}
1735 		mtx_unlock(&sync_mtx);
1736 
1737 		/*
1738 		 * Do soft update processing.
1739 		 */
1740 		if (softdep_process_worklist_hook != NULL)
1741 			(*softdep_process_worklist_hook)(NULL);
1742 
1743 		/*
1744 		 * The variable rushjob allows the kernel to speed up the
1745 		 * processing of the filesystem syncer process. A rushjob
1746 		 * value of N tells the filesystem syncer to process the next
1747 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1748 		 * is used by the soft update code to speed up the filesystem
1749 		 * syncer process when the incore state is getting so far
1750 		 * ahead of the disk that the kernel memory pool is being
1751 		 * threatened with exhaustion.
1752 		 */
1753 		mtx_lock(&sync_mtx);
1754 		if (rushjob > 0) {
1755 			rushjob -= 1;
1756 			mtx_unlock(&sync_mtx);
1757 			continue;
1758 		}
1759 		mtx_unlock(&sync_mtx);
1760 		/*
1761 		 * If it has taken us less than a second to process the
1762 		 * current work, then wait. Otherwise start right over
1763 		 * again. We can still lose time if any single round
1764 		 * takes more than two seconds, but it does not really
1765 		 * matter as we are just trying to generally pace the
1766 		 * filesystem activity.
1767 		 */
1768 		if (time_second == starttime)
1769 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1770 	}
1771 }
1772 
1773 /*
1774  * Request the syncer daemon to speed up its work.
1775  * We never push it to speed up more than half of its
1776  * normal turn time, otherwise it could take over the cpu.
1777  * XXXKSE  only one update?
1778  */
1779 int
1780 speedup_syncer()
1781 {
1782 	struct thread *td;
1783 	int ret = 0;
1784 
1785 	td = FIRST_THREAD_IN_PROC(updateproc);
1786 	mtx_lock_spin(&sched_lock);
1787 	if (td->td_wchan == &lbolt) {
1788 		unsleep(td);
1789 		TD_CLR_SLEEPING(td);
1790 		setrunnable(td);
1791 	}
1792 	mtx_unlock_spin(&sched_lock);
1793 	mtx_lock(&sync_mtx);
1794 	if (rushjob < syncdelay / 2) {
1795 		rushjob += 1;
1796 		stat_rush_requests += 1;
1797 		ret = 1;
1798 	}
1799 	mtx_unlock(&sync_mtx);
1800 	return (ret);
1801 }
1802 
1803 /*
1804  * Associate a p-buffer with a vnode.
1805  *
1806  * Also sets B_PAGING flag to indicate that vnode is not fully associated
1807  * with the buffer.  i.e. the bp has not been linked into the vnode or
1808  * ref-counted.
1809  */
1810 void
1811 pbgetvp(vp, bp)
1812 	register struct vnode *vp;
1813 	register struct buf *bp;
1814 {
1815 
1816 	KASSERT(bp->b_vp == NULL, ("pbgetvp: not free"));
1817 
1818 	bp->b_vp = vp;
1819 	bp->b_flags |= B_PAGING;
1820 	bp->b_dev = vn_todev(vp);
1821 }
1822 
1823 /*
1824  * Disassociate a p-buffer from a vnode.
1825  */
1826 void
1827 pbrelvp(bp)
1828 	register struct buf *bp;
1829 {
1830 
1831 	KASSERT(bp->b_vp != NULL, ("pbrelvp: NULL"));
1832 
1833 	/* XXX REMOVE ME */
1834 	VI_LOCK(bp->b_vp);
1835 	if (TAILQ_NEXT(bp, b_vnbufs) != NULL) {
1836 		panic(
1837 		    "relpbuf(): b_vp was probably reassignbuf()d %p %x",
1838 		    bp,
1839 		    (int)bp->b_flags
1840 		);
1841 	}
1842 	VI_UNLOCK(bp->b_vp);
1843 	bp->b_vp = (struct vnode *) 0;
1844 	bp->b_flags &= ~B_PAGING;
1845 }
1846 
1847 /*
1848  * Reassign a buffer from one vnode to another.
1849  * Used to assign file specific control information
1850  * (indirect blocks) to the vnode to which they belong.
1851  */
1852 void
1853 reassignbuf(bp, newvp)
1854 	register struct buf *bp;
1855 	register struct vnode *newvp;
1856 {
1857 	int delay;
1858 	int s;
1859 
1860 	if (newvp == NULL) {
1861 		printf("reassignbuf: NULL");
1862 		return;
1863 	}
1864 	++reassignbufcalls;
1865 
1866 	/*
1867 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1868 	 * is not fully linked in.
1869 	 */
1870 	if (bp->b_flags & B_PAGING)
1871 		panic("cannot reassign paging buffer");
1872 
1873 	s = splbio();
1874 	/*
1875 	 * Delete from old vnode list, if on one.
1876 	 */
1877 	VI_LOCK(bp->b_vp);
1878 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) {
1879 		buf_vlist_remove(bp);
1880 		if (bp->b_vp != newvp) {
1881 			vdropl(bp->b_vp);
1882 			bp->b_vp = NULL;	/* for clarification */
1883 		}
1884 	}
1885 	VI_UNLOCK(bp->b_vp);
1886 	/*
1887 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1888 	 * of clean buffers.
1889 	 */
1890 	VI_LOCK(newvp);
1891 	if (bp->b_flags & B_DELWRI) {
1892 		if ((newvp->v_iflag & VI_ONWORKLST) == 0) {
1893 			switch (newvp->v_type) {
1894 			case VDIR:
1895 				delay = dirdelay;
1896 				break;
1897 			case VCHR:
1898 				if (newvp->v_rdev->si_mountpoint != NULL) {
1899 					delay = metadelay;
1900 					break;
1901 				}
1902 				/* FALLTHROUGH */
1903 			default:
1904 				delay = filedelay;
1905 			}
1906 			vn_syncer_add_to_worklist(newvp, delay);
1907 		}
1908 		buf_vlist_add(bp, newvp, BX_VNDIRTY);
1909 	} else {
1910 		buf_vlist_add(bp, newvp, BX_VNCLEAN);
1911 
1912 		if ((newvp->v_iflag & VI_ONWORKLST) &&
1913 		    TAILQ_EMPTY(&newvp->v_dirtyblkhd)) {
1914 			mtx_lock(&sync_mtx);
1915 			LIST_REMOVE(newvp, v_synclist);
1916 			mtx_unlock(&sync_mtx);
1917 			newvp->v_iflag &= ~VI_ONWORKLST;
1918 		}
1919 	}
1920 	if (bp->b_vp != newvp) {
1921 		bp->b_vp = newvp;
1922 		vholdl(bp->b_vp);
1923 	}
1924 	VI_UNLOCK(newvp);
1925 	splx(s);
1926 }
1927 
1928 /*
1929  * Create a vnode for a device.
1930  * Used for mounting the root filesystem.
1931  */
1932 int
1933 bdevvp(dev, vpp)
1934 	dev_t dev;
1935 	struct vnode **vpp;
1936 {
1937 	register struct vnode *vp;
1938 	struct vnode *nvp;
1939 	int error;
1940 
1941 	if (dev == NODEV) {
1942 		*vpp = NULLVP;
1943 		return (ENXIO);
1944 	}
1945 	if (vfinddev(dev, VCHR, vpp))
1946 		return (0);
1947 	error = getnewvnode("none", (struct mount *)0, spec_vnodeop_p, &nvp);
1948 	if (error) {
1949 		*vpp = NULLVP;
1950 		return (error);
1951 	}
1952 	vp = nvp;
1953 	vp->v_type = VCHR;
1954 	addalias(vp, dev);
1955 	*vpp = vp;
1956 	return (0);
1957 }
1958 
1959 static void
1960 v_incr_usecount(struct vnode *vp, int delta)
1961 {
1962 	vp->v_usecount += delta;
1963 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1964 		mtx_lock(&spechash_mtx);
1965 		vp->v_rdev->si_usecount += delta;
1966 		mtx_unlock(&spechash_mtx);
1967 	}
1968 }
1969 
1970 /*
1971  * Add vnode to the alias list hung off the dev_t.
1972  *
1973  * The reason for this gunk is that multiple vnodes can reference
1974  * the same physical device, so checking vp->v_usecount to see
1975  * how many users there are is inadequate; the v_usecount for
1976  * the vnodes need to be accumulated.  vcount() does that.
1977  */
1978 struct vnode *
1979 addaliasu(nvp, nvp_rdev)
1980 	struct vnode *nvp;
1981 	udev_t nvp_rdev;
1982 {
1983 	struct vnode *ovp;
1984 	vop_t **ops;
1985 	dev_t dev;
1986 
1987 	if (nvp->v_type == VBLK)
1988 		return (nvp);
1989 	if (nvp->v_type != VCHR)
1990 		panic("addaliasu on non-special vnode");
1991 	dev = udev2dev(nvp_rdev, 0);
1992 	/*
1993 	 * Check to see if we have a bdevvp vnode with no associated
1994 	 * filesystem. If so, we want to associate the filesystem of
1995 	 * the new newly instigated vnode with the bdevvp vnode and
1996 	 * discard the newly created vnode rather than leaving the
1997 	 * bdevvp vnode lying around with no associated filesystem.
1998 	 */
1999 	if (vfinddev(dev, nvp->v_type, &ovp) == 0 || ovp->v_data != NULL) {
2000 		addalias(nvp, dev);
2001 		return (nvp);
2002 	}
2003 	/*
2004 	 * Discard unneeded vnode, but save its node specific data.
2005 	 * Note that if there is a lock, it is carried over in the
2006 	 * node specific data to the replacement vnode.
2007 	 */
2008 	vref(ovp);
2009 	ovp->v_data = nvp->v_data;
2010 	ovp->v_tag = nvp->v_tag;
2011 	nvp->v_data = NULL;
2012 	lockdestroy(ovp->v_vnlock);
2013 	lockinit(ovp->v_vnlock, PVFS, nvp->v_vnlock->lk_wmesg,
2014 	    nvp->v_vnlock->lk_timo, nvp->v_vnlock->lk_flags & LK_EXTFLG_MASK);
2015 	ops = ovp->v_op;
2016 	ovp->v_op = nvp->v_op;
2017 	if (VOP_ISLOCKED(nvp, curthread)) {
2018 		VOP_UNLOCK(nvp, 0, curthread);
2019 		vn_lock(ovp, LK_EXCLUSIVE | LK_RETRY, curthread);
2020 	}
2021 	nvp->v_op = ops;
2022 	insmntque(ovp, nvp->v_mount);
2023 	vrele(nvp);
2024 	vgone(nvp);
2025 	return (ovp);
2026 }
2027 
2028 /* This is a local helper function that do the same as addaliasu, but for a
2029  * dev_t instead of an udev_t. */
2030 static void
2031 addalias(nvp, dev)
2032 	struct vnode *nvp;
2033 	dev_t dev;
2034 {
2035 
2036 	KASSERT(nvp->v_type == VCHR, ("addalias on non-special vnode"));
2037 	nvp->v_rdev = dev;
2038 	VI_LOCK(nvp);
2039 	mtx_lock(&spechash_mtx);
2040 	SLIST_INSERT_HEAD(&dev->si_hlist, nvp, v_specnext);
2041 	dev->si_usecount += nvp->v_usecount;
2042 	mtx_unlock(&spechash_mtx);
2043 	VI_UNLOCK(nvp);
2044 }
2045 
2046 /*
2047  * Grab a particular vnode from the free list, increment its
2048  * reference count and lock it. The vnode lock bit is set if the
2049  * vnode is being eliminated in vgone. The process is awakened
2050  * when the transition is completed, and an error returned to
2051  * indicate that the vnode is no longer usable (possibly having
2052  * been changed to a new filesystem type).
2053  */
2054 int
2055 vget(vp, flags, td)
2056 	register struct vnode *vp;
2057 	int flags;
2058 	struct thread *td;
2059 {
2060 	int error;
2061 
2062 	/*
2063 	 * If the vnode is in the process of being cleaned out for
2064 	 * another use, we wait for the cleaning to finish and then
2065 	 * return failure. Cleaning is determined by checking that
2066 	 * the VI_XLOCK flag is set.
2067 	 */
2068 	if ((flags & LK_INTERLOCK) == 0)
2069 		VI_LOCK(vp);
2070 	if (vp->v_iflag & VI_XLOCK && vp->v_vxproc != curthread) {
2071 		vp->v_iflag |= VI_XWANT;
2072 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0);
2073 		return (ENOENT);
2074 	}
2075 
2076 	v_incr_usecount(vp, 1);
2077 
2078 	if (VSHOULDBUSY(vp))
2079 		vbusy(vp);
2080 	if (flags & LK_TYPE_MASK) {
2081 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2082 			/*
2083 			 * must expand vrele here because we do not want
2084 			 * to call VOP_INACTIVE if the reference count
2085 			 * drops back to zero since it was never really
2086 			 * active. We must remove it from the free list
2087 			 * before sleeping so that multiple processes do
2088 			 * not try to recycle it.
2089 			 */
2090 			VI_LOCK(vp);
2091 			v_incr_usecount(vp, -1);
2092 			if (VSHOULDFREE(vp))
2093 				vfree(vp);
2094 			else
2095 				vlruvp(vp);
2096 			VI_UNLOCK(vp);
2097 		}
2098 		return (error);
2099 	}
2100 	VI_UNLOCK(vp);
2101 	return (0);
2102 }
2103 
2104 /*
2105  * Increase the reference count of a vnode.
2106  */
2107 void
2108 vref(struct vnode *vp)
2109 {
2110 	VI_LOCK(vp);
2111 	v_incr_usecount(vp, 1);
2112 	VI_UNLOCK(vp);
2113 }
2114 
2115 /*
2116  * Return reference count of a vnode.
2117  *
2118  * The results of this call are only guaranteed when some mechanism other
2119  * than the VI lock is used to stop other processes from gaining references
2120  * to the vnode.  This may be the case if the caller holds the only reference.
2121  * This is also useful when stale data is acceptable as race conditions may
2122  * be accounted for by some other means.
2123  */
2124 int
2125 vrefcnt(struct vnode *vp)
2126 {
2127 	int usecnt;
2128 
2129 	VI_LOCK(vp);
2130 	usecnt = vp->v_usecount;
2131 	VI_UNLOCK(vp);
2132 
2133 	return (usecnt);
2134 }
2135 
2136 
2137 /*
2138  * Vnode put/release.
2139  * If count drops to zero, call inactive routine and return to freelist.
2140  */
2141 void
2142 vrele(vp)
2143 	struct vnode *vp;
2144 {
2145 	struct thread *td = curthread;	/* XXX */
2146 
2147 	KASSERT(vp != NULL, ("vrele: null vp"));
2148 
2149 	VI_LOCK(vp);
2150 
2151 	/* Skip this v_writecount check if we're going to panic below. */
2152 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2153 	    ("vrele: missed vn_close"));
2154 
2155 	if (vp->v_usecount > 1) {
2156 
2157 		v_incr_usecount(vp, -1);
2158 		VI_UNLOCK(vp);
2159 
2160 		return;
2161 	}
2162 
2163 	if (vp->v_usecount == 1) {
2164 		v_incr_usecount(vp, -1);
2165 		/*
2166 		 * We must call VOP_INACTIVE with the node locked.
2167 		 * If we are doing a vput, the node is already locked,
2168 		 * but, in the case of vrele, we must explicitly lock
2169 		 * the vnode before calling VOP_INACTIVE.
2170 		 */
2171 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0)
2172 			VOP_INACTIVE(vp, td);
2173 		VI_LOCK(vp);
2174 		if (VSHOULDFREE(vp))
2175 			vfree(vp);
2176 		else
2177 			vlruvp(vp);
2178 		VI_UNLOCK(vp);
2179 
2180 	} else {
2181 #ifdef DIAGNOSTIC
2182 		vprint("vrele: negative ref count", vp);
2183 #endif
2184 		VI_UNLOCK(vp);
2185 		panic("vrele: negative ref cnt");
2186 	}
2187 }
2188 
2189 /*
2190  * Release an already locked vnode.  This give the same effects as
2191  * unlock+vrele(), but takes less time and avoids releasing and
2192  * re-aquiring the lock (as vrele() aquires the lock internally.)
2193  */
2194 void
2195 vput(vp)
2196 	struct vnode *vp;
2197 {
2198 	struct thread *td = curthread;	/* XXX */
2199 
2200 	GIANT_REQUIRED;
2201 
2202 	KASSERT(vp != NULL, ("vput: null vp"));
2203 	VI_LOCK(vp);
2204 	/* Skip this v_writecount check if we're going to panic below. */
2205 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
2206 	    ("vput: missed vn_close"));
2207 
2208 	if (vp->v_usecount > 1) {
2209 		v_incr_usecount(vp, -1);
2210 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
2211 		return;
2212 	}
2213 
2214 	if (vp->v_usecount == 1) {
2215 		v_incr_usecount(vp, -1);
2216 		/*
2217 		 * We must call VOP_INACTIVE with the node locked.
2218 		 * If we are doing a vput, the node is already locked,
2219 		 * so we just need to release the vnode mutex.
2220 		 */
2221 		VI_UNLOCK(vp);
2222 		VOP_INACTIVE(vp, td);
2223 		VI_LOCK(vp);
2224 		if (VSHOULDFREE(vp))
2225 			vfree(vp);
2226 		else
2227 			vlruvp(vp);
2228 		VI_UNLOCK(vp);
2229 
2230 	} else {
2231 #ifdef DIAGNOSTIC
2232 		vprint("vput: negative ref count", vp);
2233 #endif
2234 		panic("vput: negative ref cnt");
2235 	}
2236 }
2237 
2238 /*
2239  * Somebody doesn't want the vnode recycled.
2240  */
2241 void
2242 vhold(struct vnode *vp)
2243 {
2244 	VI_LOCK(vp);
2245 	vholdl(vp);
2246 	VI_UNLOCK(vp);
2247 }
2248 
2249 void
2250 vholdl(vp)
2251 	register struct vnode *vp;
2252 {
2253 	int s;
2254 
2255 	s = splbio();
2256 	vp->v_holdcnt++;
2257 	if (VSHOULDBUSY(vp))
2258 		vbusy(vp);
2259 	splx(s);
2260 }
2261 
2262 /*
2263  * Note that there is one less who cares about this vnode.  vdrop() is the
2264  * opposite of vhold().
2265  */
2266 void
2267 vdrop(struct vnode *vp)
2268 {
2269 	VI_LOCK(vp);
2270 	vdropl(vp);
2271 	VI_UNLOCK(vp);
2272 }
2273 
2274 void
2275 vdropl(vp)
2276 	register struct vnode *vp;
2277 {
2278 	int s;
2279 
2280 	s = splbio();
2281 	if (vp->v_holdcnt <= 0)
2282 		panic("vdrop: holdcnt");
2283 	vp->v_holdcnt--;
2284 	if (VSHOULDFREE(vp))
2285 		vfree(vp);
2286 	else
2287 		vlruvp(vp);
2288 	splx(s);
2289 }
2290 
2291 /*
2292  * Remove any vnodes in the vnode table belonging to mount point mp.
2293  *
2294  * If FORCECLOSE is not specified, there should not be any active ones,
2295  * return error if any are found (nb: this is a user error, not a
2296  * system error). If FORCECLOSE is specified, detach any active vnodes
2297  * that are found.
2298  *
2299  * If WRITECLOSE is set, only flush out regular file vnodes open for
2300  * writing.
2301  *
2302  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2303  *
2304  * `rootrefs' specifies the base reference count for the root vnode
2305  * of this filesystem. The root vnode is considered busy if its
2306  * v_usecount exceeds this value. On a successful return, vflush()
2307  * will call vrele() on the root vnode exactly rootrefs times.
2308  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2309  * be zero.
2310  */
2311 #ifdef DIAGNOSTIC
2312 static int busyprt = 0;		/* print out busy vnodes */
2313 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2314 #endif
2315 
2316 int
2317 vflush(mp, rootrefs, flags)
2318 	struct mount *mp;
2319 	int rootrefs;
2320 	int flags;
2321 {
2322 	struct thread *td = curthread;	/* XXX */
2323 	struct vnode *vp, *nvp, *rootvp = NULL;
2324 	struct vattr vattr;
2325 	int busy = 0, error;
2326 
2327 	if (rootrefs > 0) {
2328 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2329 		    ("vflush: bad args"));
2330 		/*
2331 		 * Get the filesystem root vnode. We can vput() it
2332 		 * immediately, since with rootrefs > 0, it won't go away.
2333 		 */
2334 		if ((error = VFS_ROOT(mp, &rootvp)) != 0)
2335 			return (error);
2336 		vput(rootvp);
2337 
2338 	}
2339 	mtx_lock(&mntvnode_mtx);
2340 loop:
2341 	for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp; vp = nvp) {
2342 		/*
2343 		 * Make sure this vnode wasn't reclaimed in getnewvnode().
2344 		 * Start over if it has (it won't be on the list anymore).
2345 		 */
2346 		if (vp->v_mount != mp)
2347 			goto loop;
2348 		nvp = TAILQ_NEXT(vp, v_nmntvnodes);
2349 
2350 		VI_LOCK(vp);
2351 		mtx_unlock(&mntvnode_mtx);
2352 		vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, td);
2353 		/*
2354 		 * Skip over a vnodes marked VV_SYSTEM.
2355 		 */
2356 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2357 			VOP_UNLOCK(vp, 0, td);
2358 			mtx_lock(&mntvnode_mtx);
2359 			continue;
2360 		}
2361 		/*
2362 		 * If WRITECLOSE is set, flush out unlinked but still open
2363 		 * files (even if open only for reading) and regular file
2364 		 * vnodes open for writing.
2365 		 */
2366 		if (flags & WRITECLOSE) {
2367 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2368 			VI_LOCK(vp);
2369 
2370 			if ((vp->v_type == VNON ||
2371 			    (error == 0 && vattr.va_nlink > 0)) &&
2372 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2373 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2374 				mtx_lock(&mntvnode_mtx);
2375 				continue;
2376 			}
2377 		} else
2378 			VI_LOCK(vp);
2379 
2380 		VOP_UNLOCK(vp, 0, td);
2381 
2382 		/*
2383 		 * With v_usecount == 0, all we need to do is clear out the
2384 		 * vnode data structures and we are done.
2385 		 */
2386 		if (vp->v_usecount == 0) {
2387 			vgonel(vp, td);
2388 			mtx_lock(&mntvnode_mtx);
2389 			continue;
2390 		}
2391 
2392 		/*
2393 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2394 		 * or character devices, revert to an anonymous device. For
2395 		 * all other files, just kill them.
2396 		 */
2397 		if (flags & FORCECLOSE) {
2398 			if (vp->v_type != VCHR) {
2399 				vgonel(vp, td);
2400 			} else {
2401 				vclean(vp, 0, td);
2402 				VI_UNLOCK(vp);
2403 				vp->v_op = spec_vnodeop_p;
2404 				insmntque(vp, (struct mount *) 0);
2405 			}
2406 			mtx_lock(&mntvnode_mtx);
2407 			continue;
2408 		}
2409 #ifdef DIAGNOSTIC
2410 		if (busyprt)
2411 			vprint("vflush: busy vnode", vp);
2412 #endif
2413 		VI_UNLOCK(vp);
2414 		mtx_lock(&mntvnode_mtx);
2415 		busy++;
2416 	}
2417 	mtx_unlock(&mntvnode_mtx);
2418 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2419 		/*
2420 		 * If just the root vnode is busy, and if its refcount
2421 		 * is equal to `rootrefs', then go ahead and kill it.
2422 		 */
2423 		VI_LOCK(rootvp);
2424 		KASSERT(busy > 0, ("vflush: not busy"));
2425 		KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs"));
2426 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2427 			vgonel(rootvp, td);
2428 			busy = 0;
2429 		} else
2430 			VI_UNLOCK(rootvp);
2431 	}
2432 	if (busy)
2433 		return (EBUSY);
2434 	for (; rootrefs > 0; rootrefs--)
2435 		vrele(rootvp);
2436 	return (0);
2437 }
2438 
2439 /*
2440  * This moves a now (likely recyclable) vnode to the end of the
2441  * mountlist.  XXX However, it is temporarily disabled until we
2442  * can clean up ffs_sync() and friends, which have loop restart
2443  * conditions which this code causes to operate O(N^2).
2444  */
2445 static void
2446 vlruvp(struct vnode *vp)
2447 {
2448 #if 0
2449 	struct mount *mp;
2450 
2451 	if ((mp = vp->v_mount) != NULL) {
2452 		mtx_lock(&mntvnode_mtx);
2453 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2454 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2455 		mtx_unlock(&mntvnode_mtx);
2456 	}
2457 #endif
2458 }
2459 
2460 /*
2461  * Disassociate the underlying filesystem from a vnode.
2462  */
2463 static void
2464 vclean(vp, flags, td)
2465 	struct vnode *vp;
2466 	int flags;
2467 	struct thread *td;
2468 {
2469 	int active;
2470 
2471 	ASSERT_VI_LOCKED(vp, "vclean");
2472 	/*
2473 	 * Check to see if the vnode is in use. If so we have to reference it
2474 	 * before we clean it out so that its count cannot fall to zero and
2475 	 * generate a race against ourselves to recycle it.
2476 	 */
2477 	if ((active = vp->v_usecount))
2478 		v_incr_usecount(vp, 1);
2479 
2480 	/*
2481 	 * Prevent the vnode from being recycled or brought into use while we
2482 	 * clean it out.
2483 	 */
2484 	if (vp->v_iflag & VI_XLOCK)
2485 		panic("vclean: deadlock");
2486 	vp->v_iflag |= VI_XLOCK;
2487 	vp->v_vxproc = curthread;
2488 	/*
2489 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2490 	 * have the object locked while it cleans it out. The VOP_LOCK
2491 	 * ensures that the VOP_INACTIVE routine is done with its work.
2492 	 * For active vnodes, it ensures that no other activity can
2493 	 * occur while the underlying object is being cleaned out.
2494 	 */
2495 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2496 
2497 	/*
2498 	 * Clean out any buffers associated with the vnode.
2499 	 * If the flush fails, just toss the buffers.
2500 	 */
2501 	if (flags & DOCLOSE) {
2502 		struct buf *bp;
2503 		VI_LOCK(vp);
2504 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
2505 		VI_UNLOCK(vp);
2506 		if (bp != NULL)
2507 			(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2508 		if (vinvalbuf(vp, V_SAVE, NOCRED, td, 0, 0) != 0)
2509 			vinvalbuf(vp, 0, NOCRED, td, 0, 0);
2510 	}
2511 
2512 	VOP_DESTROYVOBJECT(vp);
2513 
2514 	/*
2515 	 * Any other processes trying to obtain this lock must first
2516 	 * wait for VXLOCK to clear, then call the new lock operation.
2517 	 */
2518 	VOP_UNLOCK(vp, 0, td);
2519 
2520 	/*
2521 	 * If purging an active vnode, it must be closed and
2522 	 * deactivated before being reclaimed. Note that the
2523 	 * VOP_INACTIVE will unlock the vnode.
2524 	 */
2525 	if (active) {
2526 		if (flags & DOCLOSE)
2527 			VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2528 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2529 			panic("vclean: cannot relock.");
2530 		VOP_INACTIVE(vp, td);
2531 	}
2532 
2533 	/*
2534 	 * Reclaim the vnode.
2535 	 */
2536 	if (VOP_RECLAIM(vp, td))
2537 		panic("vclean: cannot reclaim");
2538 
2539 	if (active) {
2540 		/*
2541 		 * Inline copy of vrele() since VOP_INACTIVE
2542 		 * has already been called.
2543 		 */
2544 		VI_LOCK(vp);
2545 		v_incr_usecount(vp, -1);
2546 		if (vp->v_usecount <= 0) {
2547 #ifdef DIAGNOSTIC
2548 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2549 				vprint("vclean: bad ref count", vp);
2550 				panic("vclean: ref cnt");
2551 			}
2552 #endif
2553 			vfree(vp);
2554 		}
2555 		VI_UNLOCK(vp);
2556 	}
2557 
2558 	cache_purge(vp);
2559 	VI_LOCK(vp);
2560 	if (VSHOULDFREE(vp))
2561 		vfree(vp);
2562 
2563 	/*
2564 	 * Done with purge, reset to the standard lock and
2565 	 * notify sleepers of the grim news.
2566 	 */
2567 	vp->v_vnlock = &vp->v_lock;
2568 	vp->v_op = dead_vnodeop_p;
2569 	if (vp->v_pollinfo != NULL)
2570 		vn_pollgone(vp);
2571 	vp->v_tag = "none";
2572 	vp->v_iflag &= ~VI_XLOCK;
2573 	vp->v_vxproc = NULL;
2574 	if (vp->v_iflag & VI_XWANT) {
2575 		vp->v_iflag &= ~VI_XWANT;
2576 		wakeup(vp);
2577 	}
2578 }
2579 
2580 /*
2581  * Eliminate all activity associated with the requested vnode
2582  * and with all vnodes aliased to the requested vnode.
2583  */
2584 int
2585 vop_revoke(ap)
2586 	struct vop_revoke_args /* {
2587 		struct vnode *a_vp;
2588 		int a_flags;
2589 	} */ *ap;
2590 {
2591 	struct vnode *vp, *vq;
2592 	dev_t dev;
2593 
2594 	KASSERT((ap->a_flags & REVOKEALL) != 0, ("vop_revoke"));
2595 
2596 	vp = ap->a_vp;
2597 	VI_LOCK(vp);
2598 	/*
2599 	 * If a vgone (or vclean) is already in progress,
2600 	 * wait until it is done and return.
2601 	 */
2602 	if (vp->v_iflag & VI_XLOCK) {
2603 		vp->v_iflag |= VI_XWANT;
2604 		msleep(vp, VI_MTX(vp), PINOD | PDROP,
2605 		    "vop_revokeall", 0);
2606 		return (0);
2607 	}
2608 	VI_UNLOCK(vp);
2609 	dev = vp->v_rdev;
2610 	for (;;) {
2611 		mtx_lock(&spechash_mtx);
2612 		vq = SLIST_FIRST(&dev->si_hlist);
2613 		mtx_unlock(&spechash_mtx);
2614 		if (!vq)
2615 			break;
2616 		vgone(vq);
2617 	}
2618 	return (0);
2619 }
2620 
2621 /*
2622  * Recycle an unused vnode to the front of the free list.
2623  * Release the passed interlock if the vnode will be recycled.
2624  */
2625 int
2626 vrecycle(vp, inter_lkp, td)
2627 	struct vnode *vp;
2628 	struct mtx *inter_lkp;
2629 	struct thread *td;
2630 {
2631 
2632 	VI_LOCK(vp);
2633 	if (vp->v_usecount == 0) {
2634 		if (inter_lkp) {
2635 			mtx_unlock(inter_lkp);
2636 		}
2637 		vgonel(vp, td);
2638 		return (1);
2639 	}
2640 	VI_UNLOCK(vp);
2641 	return (0);
2642 }
2643 
2644 /*
2645  * Eliminate all activity associated with a vnode
2646  * in preparation for reuse.
2647  */
2648 void
2649 vgone(vp)
2650 	register struct vnode *vp;
2651 {
2652 	struct thread *td = curthread;	/* XXX */
2653 
2654 	VI_LOCK(vp);
2655 	vgonel(vp, td);
2656 }
2657 
2658 /*
2659  * vgone, with the vp interlock held.
2660  */
2661 void
2662 vgonel(vp, td)
2663 	struct vnode *vp;
2664 	struct thread *td;
2665 {
2666 	int s;
2667 
2668 	/*
2669 	 * If a vgone (or vclean) is already in progress,
2670 	 * wait until it is done and return.
2671 	 */
2672 	ASSERT_VI_LOCKED(vp, "vgonel");
2673 	if (vp->v_iflag & VI_XLOCK) {
2674 		vp->v_iflag |= VI_XWANT;
2675 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0);
2676 		return;
2677 	}
2678 
2679 	/*
2680 	 * Clean out the filesystem specific data.
2681 	 */
2682 	vclean(vp, DOCLOSE, td);
2683 	VI_UNLOCK(vp);
2684 
2685 	/*
2686 	 * Delete from old mount point vnode list, if on one.
2687 	 */
2688 	if (vp->v_mount != NULL)
2689 		insmntque(vp, (struct mount *)0);
2690 	/*
2691 	 * If special device, remove it from special device alias list
2692 	 * if it is on one.
2693 	 */
2694 	if (vp->v_type == VCHR && vp->v_rdev != NULL && vp->v_rdev != NODEV) {
2695 		VI_LOCK(vp);
2696 		mtx_lock(&spechash_mtx);
2697 		SLIST_REMOVE(&vp->v_rdev->si_hlist, vp, vnode, v_specnext);
2698 		vp->v_rdev->si_usecount -= vp->v_usecount;
2699 		mtx_unlock(&spechash_mtx);
2700 		VI_UNLOCK(vp);
2701 		vp->v_rdev = NULL;
2702 	}
2703 
2704 	/*
2705 	 * If it is on the freelist and not already at the head,
2706 	 * move it to the head of the list. The test of the
2707 	 * VDOOMED flag and the reference count of zero is because
2708 	 * it will be removed from the free list by getnewvnode,
2709 	 * but will not have its reference count incremented until
2710 	 * after calling vgone. If the reference count were
2711 	 * incremented first, vgone would (incorrectly) try to
2712 	 * close the previous instance of the underlying object.
2713 	 */
2714 	VI_LOCK(vp);
2715 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2716 		s = splbio();
2717 		mtx_lock(&vnode_free_list_mtx);
2718 		if (vp->v_iflag & VI_FREE) {
2719 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2720 		} else {
2721 			vp->v_iflag |= VI_FREE;
2722 			freevnodes++;
2723 		}
2724 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2725 		mtx_unlock(&vnode_free_list_mtx);
2726 		splx(s);
2727 	}
2728 
2729 	vp->v_type = VBAD;
2730 	VI_UNLOCK(vp);
2731 }
2732 
2733 /*
2734  * Lookup a vnode by device number.
2735  */
2736 int
2737 vfinddev(dev, type, vpp)
2738 	dev_t dev;
2739 	enum vtype type;
2740 	struct vnode **vpp;
2741 {
2742 	struct vnode *vp;
2743 
2744 	mtx_lock(&spechash_mtx);
2745 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2746 		if (type == vp->v_type) {
2747 			*vpp = vp;
2748 			mtx_unlock(&spechash_mtx);
2749 			return (1);
2750 		}
2751 	}
2752 	mtx_unlock(&spechash_mtx);
2753 	return (0);
2754 }
2755 
2756 /*
2757  * Calculate the total number of references to a special device.
2758  */
2759 int
2760 vcount(vp)
2761 	struct vnode *vp;
2762 {
2763 	int count;
2764 
2765 	mtx_lock(&spechash_mtx);
2766 	count = vp->v_rdev->si_usecount;
2767 	mtx_unlock(&spechash_mtx);
2768 	return (count);
2769 }
2770 
2771 /*
2772  * Same as above, but using the dev_t as argument
2773  */
2774 int
2775 count_dev(dev)
2776 	dev_t dev;
2777 {
2778 	struct vnode *vp;
2779 
2780 	vp = SLIST_FIRST(&dev->si_hlist);
2781 	if (vp == NULL)
2782 		return (0);
2783 	return(vcount(vp));
2784 }
2785 
2786 /*
2787  * Print out a description of a vnode.
2788  */
2789 static char *typename[] =
2790 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2791 
2792 void
2793 vprint(label, vp)
2794 	char *label;
2795 	struct vnode *vp;
2796 {
2797 	char buf[96];
2798 
2799 	if (label != NULL)
2800 		printf("%s: %p: ", label, (void *)vp);
2801 	else
2802 		printf("%p: ", (void *)vp);
2803 	printf("tag %s, type %s, usecount %d, writecount %d, refcount %d,",
2804 	    vp->v_tag, typename[vp->v_type], vp->v_usecount,
2805 	    vp->v_writecount, vp->v_holdcnt);
2806 	buf[0] = '\0';
2807 	if (vp->v_vflag & VV_ROOT)
2808 		strcat(buf, "|VV_ROOT");
2809 	if (vp->v_vflag & VV_TEXT)
2810 		strcat(buf, "|VV_TEXT");
2811 	if (vp->v_vflag & VV_SYSTEM)
2812 		strcat(buf, "|VV_SYSTEM");
2813 	if (vp->v_iflag & VI_XLOCK)
2814 		strcat(buf, "|VI_XLOCK");
2815 	if (vp->v_iflag & VI_XWANT)
2816 		strcat(buf, "|VI_XWANT");
2817 	if (vp->v_iflag & VI_BWAIT)
2818 		strcat(buf, "|VI_BWAIT");
2819 	if (vp->v_iflag & VI_DOOMED)
2820 		strcat(buf, "|VI_DOOMED");
2821 	if (vp->v_iflag & VI_FREE)
2822 		strcat(buf, "|VI_FREE");
2823 	if (vp->v_vflag & VV_OBJBUF)
2824 		strcat(buf, "|VV_OBJBUF");
2825 	if (buf[0] != '\0')
2826 		printf(" flags (%s),", &buf[1]);
2827 	lockmgr_printinfo(vp->v_vnlock);
2828 	printf("\n");
2829 	if (vp->v_data != NULL) {
2830 		printf("\t");
2831 		VOP_PRINT(vp);
2832 	}
2833 }
2834 
2835 #ifdef DDB
2836 #include <ddb/ddb.h>
2837 /*
2838  * List all of the locked vnodes in the system.
2839  * Called when debugging the kernel.
2840  */
2841 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2842 {
2843 	struct thread *td = curthread;	/* XXX */
2844 	struct mount *mp, *nmp;
2845 	struct vnode *vp;
2846 
2847 	printf("Locked vnodes\n");
2848 	mtx_lock(&mountlist_mtx);
2849 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2850 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
2851 			nmp = TAILQ_NEXT(mp, mnt_list);
2852 			continue;
2853 		}
2854 		mtx_lock(&mntvnode_mtx);
2855 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2856 			if (VOP_ISLOCKED(vp, NULL))
2857 				vprint((char *)0, vp);
2858 		}
2859 		mtx_unlock(&mntvnode_mtx);
2860 		mtx_lock(&mountlist_mtx);
2861 		nmp = TAILQ_NEXT(mp, mnt_list);
2862 		vfs_unbusy(mp, td);
2863 	}
2864 	mtx_unlock(&mountlist_mtx);
2865 }
2866 #endif
2867 
2868 /*
2869  * Fill in a struct xvfsconf based on a struct vfsconf.
2870  */
2871 static void
2872 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2873 {
2874 
2875 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2876 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2877 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2878 	xvfsp->vfc_flags = vfsp->vfc_flags;
2879 	/*
2880 	 * These are unused in userland, we keep them
2881 	 * to not break binary compatibility.
2882 	 */
2883 	xvfsp->vfc_vfsops = NULL;
2884 	xvfsp->vfc_next = NULL;
2885 }
2886 
2887 static int
2888 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2889 {
2890 	struct vfsconf *vfsp;
2891 	struct xvfsconf *xvfsp;
2892 	int cnt, error, i;
2893 
2894 	cnt = 0;
2895 	for (vfsp = vfsconf; vfsp != NULL; vfsp = vfsp->vfc_next)
2896 		cnt++;
2897 	xvfsp = malloc(sizeof(struct xvfsconf) * cnt, M_TEMP, M_WAITOK);
2898 	/*
2899 	 * Handle the race that we will have here when struct vfsconf
2900 	 * will be locked down by using both cnt and checking vfc_next
2901 	 * against NULL to determine the end of the loop.  The race will
2902 	 * happen because we will have to unlock before calling malloc().
2903 	 * We are protected by Giant for now.
2904 	 */
2905 	i = 0;
2906 	for (vfsp = vfsconf; vfsp != NULL && i < cnt; vfsp = vfsp->vfc_next) {
2907 		vfsconf2x(vfsp, xvfsp + i);
2908 		i++;
2909 	}
2910 	error = SYSCTL_OUT(req, xvfsp, sizeof(struct xvfsconf) * i);
2911 	free(xvfsp, M_TEMP);
2912 	return (error);
2913 }
2914 
2915 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2916     "S,xvfsconf", "List of all configured filesystems");
2917 
2918 /*
2919  * Top level filesystem related information gathering.
2920  */
2921 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2922 
2923 static int
2924 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2925 {
2926 	int *name = (int *)arg1 - 1;	/* XXX */
2927 	u_int namelen = arg2 + 1;	/* XXX */
2928 	struct vfsconf *vfsp;
2929 	struct xvfsconf xvfsp;
2930 
2931 	printf("WARNING: userland calling deprecated sysctl, "
2932 	    "please rebuild world\n");
2933 
2934 #if 1 || defined(COMPAT_PRELITE2)
2935 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2936 	if (namelen == 1)
2937 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2938 #endif
2939 
2940 	switch (name[1]) {
2941 	case VFS_MAXTYPENUM:
2942 		if (namelen != 2)
2943 			return (ENOTDIR);
2944 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2945 	case VFS_CONF:
2946 		if (namelen != 3)
2947 			return (ENOTDIR);	/* overloaded */
2948 		for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next)
2949 			if (vfsp->vfc_typenum == name[2])
2950 				break;
2951 		if (vfsp == NULL)
2952 			return (EOPNOTSUPP);
2953 		vfsconf2x(vfsp, &xvfsp);
2954 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2955 	}
2956 	return (EOPNOTSUPP);
2957 }
2958 
2959 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl,
2960 	"Generic filesystem");
2961 
2962 #if 1 || defined(COMPAT_PRELITE2)
2963 
2964 static int
2965 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2966 {
2967 	int error;
2968 	struct vfsconf *vfsp;
2969 	struct ovfsconf ovfs;
2970 
2971 	for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) {
2972 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2973 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2974 		ovfs.vfc_index = vfsp->vfc_typenum;
2975 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2976 		ovfs.vfc_flags = vfsp->vfc_flags;
2977 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2978 		if (error)
2979 			return error;
2980 	}
2981 	return 0;
2982 }
2983 
2984 #endif /* 1 || COMPAT_PRELITE2 */
2985 
2986 #define KINFO_VNODESLOP		10
2987 /*
2988  * Dump vnode list (via sysctl).
2989  */
2990 /* ARGSUSED */
2991 static int
2992 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2993 {
2994 	struct xvnode *xvn;
2995 	struct thread *td = req->td;
2996 	struct mount *mp;
2997 	struct vnode *vp;
2998 	int error, len, n;
2999 
3000 	/*
3001 	 * Stale numvnodes access is not fatal here.
3002 	 */
3003 	req->lock = 0;
3004 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3005 	if (!req->oldptr)
3006 		/* Make an estimate */
3007 		return (SYSCTL_OUT(req, 0, len));
3008 
3009 	sysctl_wire_old_buffer(req, 0);
3010 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3011 	n = 0;
3012 	mtx_lock(&mountlist_mtx);
3013 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3014 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
3015 			continue;
3016 		mtx_lock(&mntvnode_mtx);
3017 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3018 			if (n == len)
3019 				break;
3020 			vref(vp);
3021 			xvn[n].xv_size = sizeof *xvn;
3022 			xvn[n].xv_vnode = vp;
3023 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3024 			XV_COPY(usecount);
3025 			XV_COPY(writecount);
3026 			XV_COPY(holdcnt);
3027 			XV_COPY(id);
3028 			XV_COPY(mount);
3029 			XV_COPY(numoutput);
3030 			XV_COPY(type);
3031 #undef XV_COPY
3032 			xvn[n].xv_flag = vp->v_vflag;
3033 
3034 			switch (vp->v_type) {
3035 			case VREG:
3036 			case VDIR:
3037 			case VLNK:
3038 				xvn[n].xv_dev = vp->v_cachedfs;
3039 				xvn[n].xv_ino = vp->v_cachedid;
3040 				break;
3041 			case VBLK:
3042 			case VCHR:
3043 				if (vp->v_rdev == NULL) {
3044 					vrele(vp);
3045 					continue;
3046 				}
3047 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3048 				break;
3049 			case VSOCK:
3050 				xvn[n].xv_socket = vp->v_socket;
3051 				break;
3052 			case VFIFO:
3053 				xvn[n].xv_fifo = vp->v_fifoinfo;
3054 				break;
3055 			case VNON:
3056 			case VBAD:
3057 			default:
3058 				/* shouldn't happen? */
3059 				vrele(vp);
3060 				continue;
3061 			}
3062 			vrele(vp);
3063 			++n;
3064 		}
3065 		mtx_unlock(&mntvnode_mtx);
3066 		mtx_lock(&mountlist_mtx);
3067 		vfs_unbusy(mp, td);
3068 		if (n == len)
3069 			break;
3070 	}
3071 	mtx_unlock(&mountlist_mtx);
3072 
3073 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3074 	free(xvn, M_TEMP);
3075 	return (error);
3076 }
3077 
3078 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
3079 	0, 0, sysctl_vnode, "S,xvnode", "");
3080 
3081 /*
3082  * Check to see if a filesystem is mounted on a block device.
3083  */
3084 int
3085 vfs_mountedon(vp)
3086 	struct vnode *vp;
3087 {
3088 
3089 	if (vp->v_rdev->si_mountpoint != NULL)
3090 		return (EBUSY);
3091 	return (0);
3092 }
3093 
3094 /*
3095  * Unmount all filesystems. The list is traversed in reverse order
3096  * of mounting to avoid dependencies.
3097  */
3098 void
3099 vfs_unmountall()
3100 {
3101 	struct mount *mp;
3102 	struct thread *td;
3103 	int error;
3104 
3105 	if (curthread != NULL)
3106 		td = curthread;
3107 	else
3108 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
3109 	/*
3110 	 * Since this only runs when rebooting, it is not interlocked.
3111 	 */
3112 	while(!TAILQ_EMPTY(&mountlist)) {
3113 		mp = TAILQ_LAST(&mountlist, mntlist);
3114 		error = dounmount(mp, MNT_FORCE, td);
3115 		if (error) {
3116 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
3117 			printf("unmount of %s failed (",
3118 			    mp->mnt_stat.f_mntonname);
3119 			if (error == EBUSY)
3120 				printf("BUSY)\n");
3121 			else
3122 				printf("%d)\n", error);
3123 		} else {
3124 			/* The unmount has removed mp from the mountlist */
3125 		}
3126 	}
3127 }
3128 
3129 /*
3130  * perform msync on all vnodes under a mount point
3131  * the mount point must be locked.
3132  */
3133 void
3134 vfs_msync(struct mount *mp, int flags)
3135 {
3136 	struct vnode *vp, *nvp;
3137 	struct vm_object *obj;
3138 	int tries;
3139 
3140 	GIANT_REQUIRED;
3141 
3142 	tries = 5;
3143 	mtx_lock(&mntvnode_mtx);
3144 loop:
3145 	for (vp = TAILQ_FIRST(&mp->mnt_nvnodelist); vp != NULL; vp = nvp) {
3146 		if (vp->v_mount != mp) {
3147 			if (--tries > 0)
3148 				goto loop;
3149 			break;
3150 		}
3151 		nvp = TAILQ_NEXT(vp, v_nmntvnodes);
3152 
3153 		VI_LOCK(vp);
3154 		if (vp->v_iflag & VI_XLOCK) {	/* XXX: what if MNT_WAIT? */
3155 			VI_UNLOCK(vp);
3156 			continue;
3157 		}
3158 
3159 		if ((vp->v_iflag & VI_OBJDIRTY) &&
3160 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
3161 			mtx_unlock(&mntvnode_mtx);
3162 			if (!vget(vp,
3163 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3164 			    curthread)) {
3165 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3166 					vput(vp);
3167 					mtx_lock(&mntvnode_mtx);
3168 					continue;
3169 				}
3170 
3171 				if (VOP_GETVOBJECT(vp, &obj) == 0) {
3172 					vm_object_page_clean(obj, 0, 0,
3173 					    flags == MNT_WAIT ?
3174 					    OBJPC_SYNC : OBJPC_NOSYNC);
3175 				}
3176 				vput(vp);
3177 			}
3178 			mtx_lock(&mntvnode_mtx);
3179 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
3180 				if (--tries > 0)
3181 					goto loop;
3182 				break;
3183 			}
3184 		} else
3185 			VI_UNLOCK(vp);
3186 	}
3187 	mtx_unlock(&mntvnode_mtx);
3188 }
3189 
3190 /*
3191  * Create the VM object needed for VMIO and mmap support.  This
3192  * is done for all VREG files in the system.  Some filesystems might
3193  * afford the additional metadata buffering capability of the
3194  * VMIO code by making the device node be VMIO mode also.
3195  *
3196  * vp must be locked when vfs_object_create is called.
3197  */
3198 int
3199 vfs_object_create(vp, td, cred)
3200 	struct vnode *vp;
3201 	struct thread *td;
3202 	struct ucred *cred;
3203 {
3204 	GIANT_REQUIRED;
3205 	return (VOP_CREATEVOBJECT(vp, cred, td));
3206 }
3207 
3208 /*
3209  * Mark a vnode as free, putting it up for recycling.
3210  */
3211 void
3212 vfree(vp)
3213 	struct vnode *vp;
3214 {
3215 	int s;
3216 
3217 	ASSERT_VI_LOCKED(vp, "vfree");
3218 	s = splbio();
3219 	mtx_lock(&vnode_free_list_mtx);
3220 	KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free"));
3221 	if (vp->v_iflag & VI_AGE) {
3222 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
3223 	} else {
3224 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
3225 	}
3226 	freevnodes++;
3227 	mtx_unlock(&vnode_free_list_mtx);
3228 	vp->v_iflag &= ~VI_AGE;
3229 	vp->v_iflag |= VI_FREE;
3230 	splx(s);
3231 }
3232 
3233 /*
3234  * Opposite of vfree() - mark a vnode as in use.
3235  */
3236 void
3237 vbusy(vp)
3238 	struct vnode *vp;
3239 {
3240 	int s;
3241 
3242 	s = splbio();
3243 	ASSERT_VI_LOCKED(vp, "vbusy");
3244 	KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free"));
3245 
3246 	mtx_lock(&vnode_free_list_mtx);
3247 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
3248 	freevnodes--;
3249 	mtx_unlock(&vnode_free_list_mtx);
3250 
3251 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
3252 	splx(s);
3253 }
3254 
3255 /*
3256  * Record a process's interest in events which might happen to
3257  * a vnode.  Because poll uses the historic select-style interface
3258  * internally, this routine serves as both the ``check for any
3259  * pending events'' and the ``record my interest in future events''
3260  * functions.  (These are done together, while the lock is held,
3261  * to avoid race conditions.)
3262  */
3263 int
3264 vn_pollrecord(vp, td, events)
3265 	struct vnode *vp;
3266 	struct thread *td;
3267 	short events;
3268 {
3269 
3270 	if (vp->v_pollinfo == NULL)
3271 		v_addpollinfo(vp);
3272 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3273 	if (vp->v_pollinfo->vpi_revents & events) {
3274 		/*
3275 		 * This leaves events we are not interested
3276 		 * in available for the other process which
3277 		 * which presumably had requested them
3278 		 * (otherwise they would never have been
3279 		 * recorded).
3280 		 */
3281 		events &= vp->v_pollinfo->vpi_revents;
3282 		vp->v_pollinfo->vpi_revents &= ~events;
3283 
3284 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3285 		return events;
3286 	}
3287 	vp->v_pollinfo->vpi_events |= events;
3288 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3289 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3290 	return 0;
3291 }
3292 
3293 /*
3294  * Note the occurrence of an event.  If the VN_POLLEVENT macro is used,
3295  * it is possible for us to miss an event due to race conditions, but
3296  * that condition is expected to be rare, so for the moment it is the
3297  * preferred interface.
3298  */
3299 void
3300 vn_pollevent(vp, events)
3301 	struct vnode *vp;
3302 	short events;
3303 {
3304 
3305 	if (vp->v_pollinfo == NULL)
3306 		v_addpollinfo(vp);
3307 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3308 	if (vp->v_pollinfo->vpi_events & events) {
3309 		/*
3310 		 * We clear vpi_events so that we don't
3311 		 * call selwakeup() twice if two events are
3312 		 * posted before the polling process(es) is
3313 		 * awakened.  This also ensures that we take at
3314 		 * most one selwakeup() if the polling process
3315 		 * is no longer interested.  However, it does
3316 		 * mean that only one event can be noticed at
3317 		 * a time.  (Perhaps we should only clear those
3318 		 * event bits which we note?) XXX
3319 		 */
3320 		vp->v_pollinfo->vpi_events = 0;	/* &= ~events ??? */
3321 		vp->v_pollinfo->vpi_revents |= events;
3322 		selwakeup(&vp->v_pollinfo->vpi_selinfo);
3323 	}
3324 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3325 }
3326 
3327 /*
3328  * Wake up anyone polling on vp because it is being revoked.
3329  * This depends on dead_poll() returning POLLHUP for correct
3330  * behavior.
3331  */
3332 void
3333 vn_pollgone(vp)
3334 	struct vnode *vp;
3335 {
3336 
3337 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3338 	VN_KNOTE(vp, NOTE_REVOKE);
3339 	if (vp->v_pollinfo->vpi_events) {
3340 		vp->v_pollinfo->vpi_events = 0;
3341 		selwakeup(&vp->v_pollinfo->vpi_selinfo);
3342 	}
3343 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3344 }
3345 
3346 
3347 
3348 /*
3349  * Routine to create and manage a filesystem syncer vnode.
3350  */
3351 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3352 static int	sync_fsync(struct  vop_fsync_args *);
3353 static int	sync_inactive(struct  vop_inactive_args *);
3354 static int	sync_reclaim(struct  vop_reclaim_args *);
3355 static int	sync_print(struct vop_print_args *);
3356 
3357 static vop_t **sync_vnodeop_p;
3358 static struct vnodeopv_entry_desc sync_vnodeop_entries[] = {
3359 	{ &vop_default_desc,	(vop_t *) vop_eopnotsupp },
3360 	{ &vop_close_desc,	(vop_t *) sync_close },		/* close */
3361 	{ &vop_fsync_desc,	(vop_t *) sync_fsync },		/* fsync */
3362 	{ &vop_inactive_desc,	(vop_t *) sync_inactive },	/* inactive */
3363 	{ &vop_reclaim_desc,	(vop_t *) sync_reclaim },	/* reclaim */
3364 	{ &vop_lock_desc,	(vop_t *) vop_stdlock },	/* lock */
3365 	{ &vop_unlock_desc,	(vop_t *) vop_stdunlock },	/* unlock */
3366 	{ &vop_print_desc,	(vop_t *) sync_print },		/* print */
3367 	{ &vop_islocked_desc,	(vop_t *) vop_stdislocked },	/* islocked */
3368 	{ NULL, NULL }
3369 };
3370 static struct vnodeopv_desc sync_vnodeop_opv_desc =
3371 	{ &sync_vnodeop_p, sync_vnodeop_entries };
3372 
3373 VNODEOP_SET(sync_vnodeop_opv_desc);
3374 
3375 /*
3376  * Create a new filesystem syncer vnode for the specified mount point.
3377  */
3378 int
3379 vfs_allocate_syncvnode(mp)
3380 	struct mount *mp;
3381 {
3382 	struct vnode *vp;
3383 	static long start, incr, next;
3384 	int error;
3385 
3386 	/* Allocate a new vnode */
3387 	if ((error = getnewvnode("vfs", mp, sync_vnodeop_p, &vp)) != 0) {
3388 		mp->mnt_syncer = NULL;
3389 		return (error);
3390 	}
3391 	vp->v_type = VNON;
3392 	/*
3393 	 * Place the vnode onto the syncer worklist. We attempt to
3394 	 * scatter them about on the list so that they will go off
3395 	 * at evenly distributed times even if all the filesystems
3396 	 * are mounted at once.
3397 	 */
3398 	next += incr;
3399 	if (next == 0 || next > syncer_maxdelay) {
3400 		start /= 2;
3401 		incr /= 2;
3402 		if (start == 0) {
3403 			start = syncer_maxdelay / 2;
3404 			incr = syncer_maxdelay;
3405 		}
3406 		next = start;
3407 	}
3408 	VI_LOCK(vp);
3409 	vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0);
3410 	VI_UNLOCK(vp);
3411 	mp->mnt_syncer = vp;
3412 	return (0);
3413 }
3414 
3415 /*
3416  * Do a lazy sync of the filesystem.
3417  */
3418 static int
3419 sync_fsync(ap)
3420 	struct vop_fsync_args /* {
3421 		struct vnode *a_vp;
3422 		struct ucred *a_cred;
3423 		int a_waitfor;
3424 		struct thread *a_td;
3425 	} */ *ap;
3426 {
3427 	struct vnode *syncvp = ap->a_vp;
3428 	struct mount *mp = syncvp->v_mount;
3429 	struct thread *td = ap->a_td;
3430 	int error, asyncflag;
3431 
3432 	/*
3433 	 * We only need to do something if this is a lazy evaluation.
3434 	 */
3435 	if (ap->a_waitfor != MNT_LAZY)
3436 		return (0);
3437 
3438 	/*
3439 	 * Move ourselves to the back of the sync list.
3440 	 */
3441 	VI_LOCK(syncvp);
3442 	vn_syncer_add_to_worklist(syncvp, syncdelay);
3443 	VI_UNLOCK(syncvp);
3444 
3445 	/*
3446 	 * Walk the list of vnodes pushing all that are dirty and
3447 	 * not already on the sync list.
3448 	 */
3449 	mtx_lock(&mountlist_mtx);
3450 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3451 		mtx_unlock(&mountlist_mtx);
3452 		return (0);
3453 	}
3454 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3455 		vfs_unbusy(mp, td);
3456 		return (0);
3457 	}
3458 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3459 	mp->mnt_flag &= ~MNT_ASYNC;
3460 	vfs_msync(mp, MNT_NOWAIT);
3461 	error = VFS_SYNC(mp, MNT_LAZY, ap->a_cred, td);
3462 	if (asyncflag)
3463 		mp->mnt_flag |= MNT_ASYNC;
3464 	vn_finished_write(mp);
3465 	vfs_unbusy(mp, td);
3466 	return (error);
3467 }
3468 
3469 /*
3470  * The syncer vnode is no referenced.
3471  */
3472 static int
3473 sync_inactive(ap)
3474 	struct vop_inactive_args /* {
3475 		struct vnode *a_vp;
3476 		struct thread *a_td;
3477 	} */ *ap;
3478 {
3479 
3480 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3481 	vgone(ap->a_vp);
3482 	return (0);
3483 }
3484 
3485 /*
3486  * The syncer vnode is no longer needed and is being decommissioned.
3487  *
3488  * Modifications to the worklist must be protected at splbio().
3489  */
3490 static int
3491 sync_reclaim(ap)
3492 	struct vop_reclaim_args /* {
3493 		struct vnode *a_vp;
3494 	} */ *ap;
3495 {
3496 	struct vnode *vp = ap->a_vp;
3497 	int s;
3498 
3499 	s = splbio();
3500 	vp->v_mount->mnt_syncer = NULL;
3501 	VI_LOCK(vp);
3502 	if (vp->v_iflag & VI_ONWORKLST) {
3503 		mtx_lock(&sync_mtx);
3504 		LIST_REMOVE(vp, v_synclist);
3505 		mtx_unlock(&sync_mtx);
3506 		vp->v_iflag &= ~VI_ONWORKLST;
3507 	}
3508 	VI_UNLOCK(vp);
3509 	splx(s);
3510 
3511 	return (0);
3512 }
3513 
3514 /*
3515  * Print out a syncer vnode.
3516  */
3517 static int
3518 sync_print(ap)
3519 	struct vop_print_args /* {
3520 		struct vnode *a_vp;
3521 	} */ *ap;
3522 {
3523 	struct vnode *vp = ap->a_vp;
3524 
3525 	printf("syncer vnode");
3526 	if (vp->v_vnlock != NULL)
3527 		lockmgr_printinfo(vp->v_vnlock);
3528 	printf("\n");
3529 	return (0);
3530 }
3531 
3532 /*
3533  * extract the dev_t from a VCHR
3534  */
3535 dev_t
3536 vn_todev(vp)
3537 	struct vnode *vp;
3538 {
3539 	if (vp->v_type != VCHR)
3540 		return (NODEV);
3541 	return (vp->v_rdev);
3542 }
3543 
3544 /*
3545  * Check if vnode represents a disk device
3546  */
3547 int
3548 vn_isdisk(vp, errp)
3549 	struct vnode *vp;
3550 	int *errp;
3551 {
3552 	struct cdevsw *cdevsw;
3553 
3554 	if (vp->v_type != VCHR) {
3555 		if (errp != NULL)
3556 			*errp = ENOTBLK;
3557 		return (0);
3558 	}
3559 	if (vp->v_rdev == NULL) {
3560 		if (errp != NULL)
3561 			*errp = ENXIO;
3562 		return (0);
3563 	}
3564 	cdevsw = devsw(vp->v_rdev);
3565 	if (cdevsw == NULL) {
3566 		if (errp != NULL)
3567 			*errp = ENXIO;
3568 		return (0);
3569 	}
3570 	if (!(cdevsw->d_flags & D_DISK)) {
3571 		if (errp != NULL)
3572 			*errp = ENOTBLK;
3573 		return (0);
3574 	}
3575 	if (errp != NULL)
3576 		*errp = 0;
3577 	return (1);
3578 }
3579 
3580 /*
3581  * Free data allocated by namei(); see namei(9) for details.
3582  */
3583 void
3584 NDFREE(ndp, flags)
3585      struct nameidata *ndp;
3586      const uint flags;
3587 {
3588 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3589 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3590 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3591 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3592 	}
3593 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3594 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3595 	    ndp->ni_dvp != ndp->ni_vp)
3596 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3597 	if (!(flags & NDF_NO_DVP_RELE) &&
3598 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3599 		vrele(ndp->ni_dvp);
3600 		ndp->ni_dvp = NULL;
3601 	}
3602 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3603 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3604 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3605 	if (!(flags & NDF_NO_VP_RELE) &&
3606 	    ndp->ni_vp) {
3607 		vrele(ndp->ni_vp);
3608 		ndp->ni_vp = NULL;
3609 	}
3610 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3611 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3612 		vrele(ndp->ni_startdir);
3613 		ndp->ni_startdir = NULL;
3614 	}
3615 }
3616 
3617 /*
3618  * Common filesystem object access control check routine.  Accepts a
3619  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3620  * and optional call-by-reference privused argument allowing vaccess()
3621  * to indicate to the caller whether privilege was used to satisfy the
3622  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3623  */
3624 int
3625 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3626 	enum vtype type;
3627 	mode_t file_mode;
3628 	uid_t file_uid;
3629 	gid_t file_gid;
3630 	mode_t acc_mode;
3631 	struct ucred *cred;
3632 	int *privused;
3633 {
3634 	mode_t dac_granted;
3635 #ifdef CAPABILITIES
3636 	mode_t cap_granted;
3637 #endif
3638 
3639 	/*
3640 	 * Look for a normal, non-privileged way to access the file/directory
3641 	 * as requested.  If it exists, go with that.
3642 	 */
3643 
3644 	if (privused != NULL)
3645 		*privused = 0;
3646 
3647 	dac_granted = 0;
3648 
3649 	/* Check the owner. */
3650 	if (cred->cr_uid == file_uid) {
3651 		dac_granted |= VADMIN;
3652 		if (file_mode & S_IXUSR)
3653 			dac_granted |= VEXEC;
3654 		if (file_mode & S_IRUSR)
3655 			dac_granted |= VREAD;
3656 		if (file_mode & S_IWUSR)
3657 			dac_granted |= (VWRITE | VAPPEND);
3658 
3659 		if ((acc_mode & dac_granted) == acc_mode)
3660 			return (0);
3661 
3662 		goto privcheck;
3663 	}
3664 
3665 	/* Otherwise, check the groups (first match) */
3666 	if (groupmember(file_gid, cred)) {
3667 		if (file_mode & S_IXGRP)
3668 			dac_granted |= VEXEC;
3669 		if (file_mode & S_IRGRP)
3670 			dac_granted |= VREAD;
3671 		if (file_mode & S_IWGRP)
3672 			dac_granted |= (VWRITE | VAPPEND);
3673 
3674 		if ((acc_mode & dac_granted) == acc_mode)
3675 			return (0);
3676 
3677 		goto privcheck;
3678 	}
3679 
3680 	/* Otherwise, check everyone else. */
3681 	if (file_mode & S_IXOTH)
3682 		dac_granted |= VEXEC;
3683 	if (file_mode & S_IROTH)
3684 		dac_granted |= VREAD;
3685 	if (file_mode & S_IWOTH)
3686 		dac_granted |= (VWRITE | VAPPEND);
3687 	if ((acc_mode & dac_granted) == acc_mode)
3688 		return (0);
3689 
3690 privcheck:
3691 	if (!suser_cred(cred, PRISON_ROOT)) {
3692 		/* XXX audit: privilege used */
3693 		if (privused != NULL)
3694 			*privused = 1;
3695 		return (0);
3696 	}
3697 
3698 #ifdef CAPABILITIES
3699 	/*
3700 	 * Build a capability mask to determine if the set of capabilities
3701 	 * satisfies the requirements when combined with the granted mask
3702 	 * from above.
3703 	 * For each capability, if the capability is required, bitwise
3704 	 * or the request type onto the cap_granted mask.
3705 	 */
3706 	cap_granted = 0;
3707 
3708 	if (type == VDIR) {
3709 		/*
3710 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3711 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3712 		 */
3713 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3714 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT))
3715 			cap_granted |= VEXEC;
3716 	} else {
3717 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3718 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, PRISON_ROOT))
3719 			cap_granted |= VEXEC;
3720 	}
3721 
3722 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3723 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, PRISON_ROOT))
3724 		cap_granted |= VREAD;
3725 
3726 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3727 	    !cap_check(cred, NULL, CAP_DAC_WRITE, PRISON_ROOT))
3728 		cap_granted |= (VWRITE | VAPPEND);
3729 
3730 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3731 	    !cap_check(cred, NULL, CAP_FOWNER, PRISON_ROOT))
3732 		cap_granted |= VADMIN;
3733 
3734 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3735 		/* XXX audit: privilege used */
3736 		if (privused != NULL)
3737 			*privused = 1;
3738 		return (0);
3739 	}
3740 #endif
3741 
3742 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3743 }
3744 
3745 /*
3746  * Credential check based on process requesting service, and per-attribute
3747  * permissions.
3748  */
3749 int
3750 extattr_check_cred(struct vnode *vp, int attrnamespace,
3751     struct ucred *cred, struct thread *td, int access)
3752 {
3753 
3754 	/*
3755 	 * Kernel-invoked always succeeds.
3756 	 */
3757 	if (cred == NOCRED)
3758 		return (0);
3759 
3760 	/*
3761 	 * Do not allow privileged processes in jail to directly
3762 	 * manipulate system attributes.
3763 	 *
3764 	 * XXX What capability should apply here?
3765 	 * Probably CAP_SYS_SETFFLAG.
3766 	 */
3767 	switch (attrnamespace) {
3768 	case EXTATTR_NAMESPACE_SYSTEM:
3769 		/* Potentially should be: return (EPERM); */
3770 		return (suser_cred(cred, 0));
3771 	case EXTATTR_NAMESPACE_USER:
3772 		return (VOP_ACCESS(vp, access, cred, td));
3773 	default:
3774 		return (EPERM);
3775 	}
3776 }
3777