xref: /freebsd/sys/kern/vfs_subr.c (revision 94942af266ac119ede0ca836f9aa5a5ac0582938)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ddb.h"
45 #include "opt_mac.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/conf.h>
52 #include <sys/dirent.h>
53 #include <sys/event.h>
54 #include <sys/eventhandler.h>
55 #include <sys/extattr.h>
56 #include <sys/file.h>
57 #include <sys/fcntl.h>
58 #include <sys/jail.h>
59 #include <sys/kdb.h>
60 #include <sys/kernel.h>
61 #include <sys/kthread.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/namei.h>
65 #include <sys/priv.h>
66 #include <sys/reboot.h>
67 #include <sys/sleepqueue.h>
68 #include <sys/stat.h>
69 #include <sys/sysctl.h>
70 #include <sys/syslog.h>
71 #include <sys/vmmeter.h>
72 #include <sys/vnode.h>
73 
74 #include <machine/stdarg.h>
75 
76 #include <security/mac/mac_framework.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_extern.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_kern.h>
85 #include <vm/uma.h>
86 
87 #ifdef DDB
88 #include <ddb/ddb.h>
89 #endif
90 
91 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure");
92 
93 static void	delmntque(struct vnode *vp);
94 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
95 		    int slpflag, int slptimeo);
96 static void	syncer_shutdown(void *arg, int howto);
97 static int	vtryrecycle(struct vnode *vp);
98 static void	vbusy(struct vnode *vp);
99 static void	vinactive(struct vnode *, struct thread *);
100 static void	v_incr_usecount(struct vnode *);
101 static void	v_decr_usecount(struct vnode *);
102 static void	v_decr_useonly(struct vnode *);
103 static void	v_upgrade_usecount(struct vnode *);
104 static void	vfree(struct vnode *);
105 static void	vnlru_free(int);
106 static void	vdestroy(struct vnode *);
107 static void	vgonel(struct vnode *);
108 static void	vfs_knllock(void *arg);
109 static void	vfs_knlunlock(void *arg);
110 static int	vfs_knllocked(void *arg);
111 
112 
113 /*
114  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
115  * build.  Without mpsafevm the buffer cache can not run Giant free.
116  */
117 int mpsafe_vfs = 1;
118 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
119 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
120     "MPSAFE VFS");
121 
122 /*
123  * Number of vnodes in existence.  Increased whenever getnewvnode()
124  * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed
125  * vnode.
126  */
127 static unsigned long	numvnodes;
128 
129 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
130 
131 /*
132  * Conversion tables for conversion from vnode types to inode formats
133  * and back.
134  */
135 enum vtype iftovt_tab[16] = {
136 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
137 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
138 };
139 int vttoif_tab[10] = {
140 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
141 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
142 };
143 
144 /*
145  * List of vnodes that are ready for recycling.
146  */
147 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
148 
149 /*
150  * Free vnode target.  Free vnodes may simply be files which have been stat'd
151  * but not read.  This is somewhat common, and a small cache of such files
152  * should be kept to avoid recreation costs.
153  */
154 static u_long wantfreevnodes;
155 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
156 /* Number of vnodes in the free list. */
157 static u_long freevnodes;
158 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
159 
160 /*
161  * Various variables used for debugging the new implementation of
162  * reassignbuf().
163  * XXX these are probably of (very) limited utility now.
164  */
165 static int reassignbufcalls;
166 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
167 
168 /*
169  * Cache for the mount type id assigned to NFS.  This is used for
170  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
171  */
172 int	nfs_mount_type = -1;
173 
174 /* To keep more than one thread at a time from running vfs_getnewfsid */
175 static struct mtx mntid_mtx;
176 
177 /*
178  * Lock for any access to the following:
179  *	vnode_free_list
180  *	numvnodes
181  *	freevnodes
182  */
183 static struct mtx vnode_free_list_mtx;
184 
185 /* Publicly exported FS */
186 struct nfs_public nfs_pub;
187 
188 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
189 static uma_zone_t vnode_zone;
190 static uma_zone_t vnodepoll_zone;
191 
192 /* Set to 1 to print out reclaim of active vnodes */
193 int	prtactive;
194 
195 /*
196  * The workitem queue.
197  *
198  * It is useful to delay writes of file data and filesystem metadata
199  * for tens of seconds so that quickly created and deleted files need
200  * not waste disk bandwidth being created and removed. To realize this,
201  * we append vnodes to a "workitem" queue. When running with a soft
202  * updates implementation, most pending metadata dependencies should
203  * not wait for more than a few seconds. Thus, mounted on block devices
204  * are delayed only about a half the time that file data is delayed.
205  * Similarly, directory updates are more critical, so are only delayed
206  * about a third the time that file data is delayed. Thus, there are
207  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
208  * one each second (driven off the filesystem syncer process). The
209  * syncer_delayno variable indicates the next queue that is to be processed.
210  * Items that need to be processed soon are placed in this queue:
211  *
212  *	syncer_workitem_pending[syncer_delayno]
213  *
214  * A delay of fifteen seconds is done by placing the request fifteen
215  * entries later in the queue:
216  *
217  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
218  *
219  */
220 static int syncer_delayno;
221 static long syncer_mask;
222 LIST_HEAD(synclist, bufobj);
223 static struct synclist *syncer_workitem_pending;
224 /*
225  * The sync_mtx protects:
226  *	bo->bo_synclist
227  *	sync_vnode_count
228  *	syncer_delayno
229  *	syncer_state
230  *	syncer_workitem_pending
231  *	syncer_worklist_len
232  *	rushjob
233  */
234 static struct mtx sync_mtx;
235 
236 #define SYNCER_MAXDELAY		32
237 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
238 static int syncdelay = 30;		/* max time to delay syncing data */
239 static int filedelay = 30;		/* time to delay syncing files */
240 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
241 static int dirdelay = 29;		/* time to delay syncing directories */
242 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
243 static int metadelay = 28;		/* time to delay syncing metadata */
244 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
245 static int rushjob;		/* number of slots to run ASAP */
246 static int stat_rush_requests;	/* number of times I/O speeded up */
247 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
248 
249 /*
250  * When shutting down the syncer, run it at four times normal speed.
251  */
252 #define SYNCER_SHUTDOWN_SPEEDUP		4
253 static int sync_vnode_count;
254 static int syncer_worklist_len;
255 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
256     syncer_state;
257 
258 /*
259  * Number of vnodes we want to exist at any one time.  This is mostly used
260  * to size hash tables in vnode-related code.  It is normally not used in
261  * getnewvnode(), as wantfreevnodes is normally nonzero.)
262  *
263  * XXX desiredvnodes is historical cruft and should not exist.
264  */
265 int desiredvnodes;
266 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
267     &desiredvnodes, 0, "Maximum number of vnodes");
268 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
269     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
270 static int vnlru_nowhere;
271 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
272     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
273 
274 /*
275  * Macros to control when a vnode is freed and recycled.  All require
276  * the vnode interlock.
277  */
278 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
279 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
280 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
281 
282 
283 /*
284  * Initialize the vnode management data structures.
285  */
286 #ifndef	MAXVNODES_MAX
287 #define	MAXVNODES_MAX	100000
288 #endif
289 static void
290 vntblinit(void *dummy __unused)
291 {
292 
293 	/*
294 	 * Desiredvnodes is a function of the physical memory size and
295 	 * the kernel's heap size.  Specifically, desiredvnodes scales
296 	 * in proportion to the physical memory size until two fifths
297 	 * of the kernel's heap size is consumed by vnodes and vm
298 	 * objects.
299 	 */
300 	desiredvnodes = min(maxproc + VMCNT_GET(page_count) / 4, 2 *
301 	    vm_kmem_size / (5 * (sizeof(struct vm_object) +
302 	    sizeof(struct vnode))));
303 	if (desiredvnodes > MAXVNODES_MAX) {
304 		if (bootverbose)
305 			printf("Reducing kern.maxvnodes %d -> %d\n",
306 			    desiredvnodes, MAXVNODES_MAX);
307 		desiredvnodes = MAXVNODES_MAX;
308 	}
309 	wantfreevnodes = desiredvnodes / 4;
310 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
311 	TAILQ_INIT(&vnode_free_list);
312 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
313 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
314 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
316 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
317 	/*
318 	 * Initialize the filesystem syncer.
319 	 */
320 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
321 		&syncer_mask);
322 	syncer_maxdelay = syncer_mask + 1;
323 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
324 }
325 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
326 
327 
328 /*
329  * Mark a mount point as busy. Used to synchronize access and to delay
330  * unmounting. Interlock is not released on failure.
331  */
332 int
333 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp,
334     struct thread *td)
335 {
336 	int lkflags;
337 
338 	MNT_ILOCK(mp);
339 	MNT_REF(mp);
340 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
341 		if (flags & LK_NOWAIT) {
342 			MNT_REL(mp);
343 			MNT_IUNLOCK(mp);
344 			return (ENOENT);
345 		}
346 		if (interlkp)
347 			mtx_unlock(interlkp);
348 		mp->mnt_kern_flag |= MNTK_MWAIT;
349 		/*
350 		 * Since all busy locks are shared except the exclusive
351 		 * lock granted when unmounting, the only place that a
352 		 * wakeup needs to be done is at the release of the
353 		 * exclusive lock at the end of dounmount.
354 		 */
355 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
356 		MNT_REL(mp);
357 		MNT_IUNLOCK(mp);
358 		if (interlkp)
359 			mtx_lock(interlkp);
360 		return (ENOENT);
361 	}
362 	if (interlkp)
363 		mtx_unlock(interlkp);
364 	lkflags = LK_SHARED | LK_INTERLOCK;
365 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
366 		panic("vfs_busy: unexpected lock failure");
367 	return (0);
368 }
369 
370 /*
371  * Free a busy filesystem.
372  */
373 void
374 vfs_unbusy(struct mount *mp, struct thread *td)
375 {
376 
377 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
378 	vfs_rel(mp);
379 }
380 
381 /*
382  * Lookup a mount point by filesystem identifier.
383  */
384 struct mount *
385 vfs_getvfs(fsid_t *fsid)
386 {
387 	struct mount *mp;
388 
389 	mtx_lock(&mountlist_mtx);
390 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
391 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
392 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
393 			vfs_ref(mp);
394 			mtx_unlock(&mountlist_mtx);
395 			return (mp);
396 		}
397 	}
398 	mtx_unlock(&mountlist_mtx);
399 	return ((struct mount *) 0);
400 }
401 
402 /*
403  * Check if a user can access privileged mount options.
404  */
405 int
406 vfs_suser(struct mount *mp, struct thread *td)
407 {
408 	int error;
409 
410 	/*
411 	 * If the thread is jailed, but this is not a jail-friendly file
412 	 * system, deny immediately.
413 	 */
414 	if (jailed(td->td_ucred) && !(mp->mnt_vfc->vfc_flags & VFCF_JAIL))
415 		return (EPERM);
416 
417 	/*
418 	 * If the file system was mounted outside a jail and a jailed thread
419 	 * tries to access it, deny immediately.
420 	 */
421 	if (!jailed(mp->mnt_cred) && jailed(td->td_ucred))
422 		return (EPERM);
423 
424 	/*
425 	 * If the file system was mounted inside different jail that the jail of
426 	 * the calling thread, deny immediately.
427 	 */
428 	if (jailed(mp->mnt_cred) && jailed(td->td_ucred) &&
429 	    mp->mnt_cred->cr_prison != td->td_ucred->cr_prison) {
430 		return (EPERM);
431 	}
432 
433 	if ((mp->mnt_flag & MNT_USER) == 0 ||
434 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
435 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
436 			return (error);
437 	}
438 	return (0);
439 }
440 
441 /*
442  * Get a new unique fsid.  Try to make its val[0] unique, since this value
443  * will be used to create fake device numbers for stat().  Also try (but
444  * not so hard) make its val[0] unique mod 2^16, since some emulators only
445  * support 16-bit device numbers.  We end up with unique val[0]'s for the
446  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
447  *
448  * Keep in mind that several mounts may be running in parallel.  Starting
449  * the search one past where the previous search terminated is both a
450  * micro-optimization and a defense against returning the same fsid to
451  * different mounts.
452  */
453 void
454 vfs_getnewfsid(struct mount *mp)
455 {
456 	static u_int16_t mntid_base;
457 	struct mount *nmp;
458 	fsid_t tfsid;
459 	int mtype;
460 
461 	mtx_lock(&mntid_mtx);
462 	mtype = mp->mnt_vfc->vfc_typenum;
463 	tfsid.val[1] = mtype;
464 	mtype = (mtype & 0xFF) << 24;
465 	for (;;) {
466 		tfsid.val[0] = makedev(255,
467 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
468 		mntid_base++;
469 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
470 			break;
471 		vfs_rel(nmp);
472 	}
473 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
474 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
475 	mtx_unlock(&mntid_mtx);
476 }
477 
478 /*
479  * Knob to control the precision of file timestamps:
480  *
481  *   0 = seconds only; nanoseconds zeroed.
482  *   1 = seconds and nanoseconds, accurate within 1/HZ.
483  *   2 = seconds and nanoseconds, truncated to microseconds.
484  * >=3 = seconds and nanoseconds, maximum precision.
485  */
486 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
487 
488 static int timestamp_precision = TSP_SEC;
489 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
490     &timestamp_precision, 0, "");
491 
492 /*
493  * Get a current timestamp.
494  */
495 void
496 vfs_timestamp(struct timespec *tsp)
497 {
498 	struct timeval tv;
499 
500 	switch (timestamp_precision) {
501 	case TSP_SEC:
502 		tsp->tv_sec = time_second;
503 		tsp->tv_nsec = 0;
504 		break;
505 	case TSP_HZ:
506 		getnanotime(tsp);
507 		break;
508 	case TSP_USEC:
509 		microtime(&tv);
510 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
511 		break;
512 	case TSP_NSEC:
513 	default:
514 		nanotime(tsp);
515 		break;
516 	}
517 }
518 
519 /*
520  * Set vnode attributes to VNOVAL
521  */
522 void
523 vattr_null(struct vattr *vap)
524 {
525 
526 	vap->va_type = VNON;
527 	vap->va_size = VNOVAL;
528 	vap->va_bytes = VNOVAL;
529 	vap->va_mode = VNOVAL;
530 	vap->va_nlink = VNOVAL;
531 	vap->va_uid = VNOVAL;
532 	vap->va_gid = VNOVAL;
533 	vap->va_fsid = VNOVAL;
534 	vap->va_fileid = VNOVAL;
535 	vap->va_blocksize = VNOVAL;
536 	vap->va_rdev = VNOVAL;
537 	vap->va_atime.tv_sec = VNOVAL;
538 	vap->va_atime.tv_nsec = VNOVAL;
539 	vap->va_mtime.tv_sec = VNOVAL;
540 	vap->va_mtime.tv_nsec = VNOVAL;
541 	vap->va_ctime.tv_sec = VNOVAL;
542 	vap->va_ctime.tv_nsec = VNOVAL;
543 	vap->va_birthtime.tv_sec = VNOVAL;
544 	vap->va_birthtime.tv_nsec = VNOVAL;
545 	vap->va_flags = VNOVAL;
546 	vap->va_gen = VNOVAL;
547 	vap->va_vaflags = 0;
548 }
549 
550 /*
551  * This routine is called when we have too many vnodes.  It attempts
552  * to free <count> vnodes and will potentially free vnodes that still
553  * have VM backing store (VM backing store is typically the cause
554  * of a vnode blowout so we want to do this).  Therefore, this operation
555  * is not considered cheap.
556  *
557  * A number of conditions may prevent a vnode from being reclaimed.
558  * the buffer cache may have references on the vnode, a directory
559  * vnode may still have references due to the namei cache representing
560  * underlying files, or the vnode may be in active use.   It is not
561  * desireable to reuse such vnodes.  These conditions may cause the
562  * number of vnodes to reach some minimum value regardless of what
563  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
564  */
565 static int
566 vlrureclaim(struct mount *mp)
567 {
568 	struct thread *td;
569 	struct vnode *vp;
570 	int done;
571 	int trigger;
572 	int usevnodes;
573 	int count;
574 
575 	/*
576 	 * Calculate the trigger point, don't allow user
577 	 * screwups to blow us up.   This prevents us from
578 	 * recycling vnodes with lots of resident pages.  We
579 	 * aren't trying to free memory, we are trying to
580 	 * free vnodes.
581 	 */
582 	usevnodes = desiredvnodes;
583 	if (usevnodes <= 0)
584 		usevnodes = 1;
585 	trigger = VMCNT_GET(page_count) * 2 / usevnodes;
586 	done = 0;
587 	td = curthread;
588 	vn_start_write(NULL, &mp, V_WAIT);
589 	MNT_ILOCK(mp);
590 	count = mp->mnt_nvnodelistsize / 10 + 1;
591 	while (count != 0) {
592 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
593 		while (vp != NULL && vp->v_type == VMARKER)
594 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
595 		if (vp == NULL)
596 			break;
597 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
598 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
599 		--count;
600 		if (!VI_TRYLOCK(vp))
601 			goto next_iter;
602 		/*
603 		 * If it's been deconstructed already, it's still
604 		 * referenced, or it exceeds the trigger, skip it.
605 		 */
606 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
607 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
608 		    vp->v_object->resident_page_count > trigger)) {
609 			VI_UNLOCK(vp);
610 			goto next_iter;
611 		}
612 		MNT_IUNLOCK(mp);
613 		vholdl(vp);
614 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
615 			vdrop(vp);
616 			goto next_iter_mntunlocked;
617 		}
618 		VI_LOCK(vp);
619 		/*
620 		 * v_usecount may have been bumped after VOP_LOCK() dropped
621 		 * the vnode interlock and before it was locked again.
622 		 *
623 		 * It is not necessary to recheck VI_DOOMED because it can
624 		 * only be set by another thread that holds both the vnode
625 		 * lock and vnode interlock.  If another thread has the
626 		 * vnode lock before we get to VOP_LOCK() and obtains the
627 		 * vnode interlock after VOP_LOCK() drops the vnode
628 		 * interlock, the other thread will be unable to drop the
629 		 * vnode lock before our VOP_LOCK() call fails.
630 		 */
631 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
632 		    (vp->v_object != NULL &&
633 		    vp->v_object->resident_page_count > trigger)) {
634 			VOP_UNLOCK(vp, LK_INTERLOCK, td);
635 			goto next_iter_mntunlocked;
636 		}
637 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
638 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
639 		vgonel(vp);
640 		VOP_UNLOCK(vp, 0, td);
641 		vdropl(vp);
642 		done++;
643 next_iter_mntunlocked:
644 		if ((count % 256) != 0)
645 			goto relock_mnt;
646 		goto yield;
647 next_iter:
648 		if ((count % 256) != 0)
649 			continue;
650 		MNT_IUNLOCK(mp);
651 yield:
652 		uio_yield();
653 relock_mnt:
654 		MNT_ILOCK(mp);
655 	}
656 	MNT_IUNLOCK(mp);
657 	vn_finished_write(mp);
658 	return done;
659 }
660 
661 /*
662  * Attempt to keep the free list at wantfreevnodes length.
663  */
664 static void
665 vnlru_free(int count)
666 {
667 	struct vnode *vp;
668 	int vfslocked;
669 
670 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
671 	for (; count > 0; count--) {
672 		vp = TAILQ_FIRST(&vnode_free_list);
673 		/*
674 		 * The list can be modified while the free_list_mtx
675 		 * has been dropped and vp could be NULL here.
676 		 */
677 		if (!vp)
678 			break;
679 		VNASSERT(vp->v_op != NULL, vp,
680 		    ("vnlru_free: vnode already reclaimed."));
681 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
682 		/*
683 		 * Don't recycle if we can't get the interlock.
684 		 */
685 		if (!VI_TRYLOCK(vp)) {
686 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
687 			continue;
688 		}
689 		VNASSERT(VCANRECYCLE(vp), vp,
690 		    ("vp inconsistent on freelist"));
691 		freevnodes--;
692 		vp->v_iflag &= ~VI_FREE;
693 		vholdl(vp);
694 		mtx_unlock(&vnode_free_list_mtx);
695 		VI_UNLOCK(vp);
696 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
697 		vtryrecycle(vp);
698 		VFS_UNLOCK_GIANT(vfslocked);
699 		/*
700 		 * If the recycled succeeded this vdrop will actually free
701 		 * the vnode.  If not it will simply place it back on
702 		 * the free list.
703 		 */
704 		vdrop(vp);
705 		mtx_lock(&vnode_free_list_mtx);
706 	}
707 }
708 /*
709  * Attempt to recycle vnodes in a context that is always safe to block.
710  * Calling vlrurecycle() from the bowels of filesystem code has some
711  * interesting deadlock problems.
712  */
713 static struct proc *vnlruproc;
714 static int vnlruproc_sig;
715 
716 static void
717 vnlru_proc(void)
718 {
719 	struct mount *mp, *nmp;
720 	int done;
721 	struct proc *p = vnlruproc;
722 	struct thread *td = FIRST_THREAD_IN_PROC(p);
723 
724 	mtx_lock(&Giant);
725 
726 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
727 	    SHUTDOWN_PRI_FIRST);
728 
729 	for (;;) {
730 		kthread_suspend_check(p);
731 		mtx_lock(&vnode_free_list_mtx);
732 		if (freevnodes > wantfreevnodes)
733 			vnlru_free(freevnodes - wantfreevnodes);
734 		if (numvnodes <= desiredvnodes * 9 / 10) {
735 			vnlruproc_sig = 0;
736 			wakeup(&vnlruproc_sig);
737 			msleep(vnlruproc, &vnode_free_list_mtx,
738 			    PVFS|PDROP, "vlruwt", hz);
739 			continue;
740 		}
741 		mtx_unlock(&vnode_free_list_mtx);
742 		done = 0;
743 		mtx_lock(&mountlist_mtx);
744 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
745 			int vfsunlocked;
746 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
747 				nmp = TAILQ_NEXT(mp, mnt_list);
748 				continue;
749 			}
750 			if (!VFS_NEEDSGIANT(mp)) {
751 				mtx_unlock(&Giant);
752 				vfsunlocked = 1;
753 			} else
754 				vfsunlocked = 0;
755 			done += vlrureclaim(mp);
756 			if (vfsunlocked)
757 				mtx_lock(&Giant);
758 			mtx_lock(&mountlist_mtx);
759 			nmp = TAILQ_NEXT(mp, mnt_list);
760 			vfs_unbusy(mp, td);
761 		}
762 		mtx_unlock(&mountlist_mtx);
763 		if (done == 0) {
764 			EVENTHANDLER_INVOKE(vfs_lowvnodes, desiredvnodes / 10);
765 #if 0
766 			/* These messages are temporary debugging aids */
767 			if (vnlru_nowhere < 5)
768 				printf("vnlru process getting nowhere..\n");
769 			else if (vnlru_nowhere == 5)
770 				printf("vnlru process messages stopped.\n");
771 #endif
772 			vnlru_nowhere++;
773 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
774 		} else
775 			uio_yield();
776 	}
777 }
778 
779 static struct kproc_desc vnlru_kp = {
780 	"vnlru",
781 	vnlru_proc,
782 	&vnlruproc
783 };
784 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
785 
786 /*
787  * Routines having to do with the management of the vnode table.
788  */
789 
790 static void
791 vdestroy(struct vnode *vp)
792 {
793 	struct bufobj *bo;
794 
795 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
796 	mtx_lock(&vnode_free_list_mtx);
797 	numvnodes--;
798 	mtx_unlock(&vnode_free_list_mtx);
799 	bo = &vp->v_bufobj;
800 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
801 	    ("cleaned vnode still on the free list."));
802 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
803 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
804 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
805 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
806 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
807 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
808 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
809 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
810 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
811 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
812 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
813 	VI_UNLOCK(vp);
814 #ifdef MAC
815 	mac_destroy_vnode(vp);
816 #endif
817 	if (vp->v_pollinfo != NULL) {
818 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
819 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
820 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
821 	}
822 #ifdef INVARIANTS
823 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
824 	vp->v_op = NULL;
825 #endif
826 	lockdestroy(vp->v_vnlock);
827 	mtx_destroy(&vp->v_interlock);
828 	uma_zfree(vnode_zone, vp);
829 }
830 
831 /*
832  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
833  * before we actually vgone().  This function must be called with the vnode
834  * held to prevent the vnode from being returned to the free list midway
835  * through vgone().
836  */
837 static int
838 vtryrecycle(struct vnode *vp)
839 {
840 	struct thread *td = curthread;
841 	struct mount *vnmp;
842 
843 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
844 	VNASSERT(vp->v_holdcnt, vp,
845 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
846 	/*
847 	 * This vnode may found and locked via some other list, if so we
848 	 * can't recycle it yet.
849 	 */
850 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
851 		return (EWOULDBLOCK);
852 	/*
853 	 * Don't recycle if its filesystem is being suspended.
854 	 */
855 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
856 		VOP_UNLOCK(vp, 0, td);
857 		return (EBUSY);
858 	}
859 	/*
860 	 * If we got this far, we need to acquire the interlock and see if
861 	 * anyone picked up this vnode from another list.  If not, we will
862 	 * mark it with DOOMED via vgonel() so that anyone who does find it
863 	 * will skip over it.
864 	 */
865 	VI_LOCK(vp);
866 	if (vp->v_usecount) {
867 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
868 		vn_finished_write(vnmp);
869 		return (EBUSY);
870 	}
871 	if ((vp->v_iflag & VI_DOOMED) == 0)
872 		vgonel(vp);
873 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
874 	vn_finished_write(vnmp);
875 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
876 	return (0);
877 }
878 
879 /*
880  * Return the next vnode from the free list.
881  */
882 int
883 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
884     struct vnode **vpp)
885 {
886 	struct vnode *vp = NULL;
887 	struct bufobj *bo;
888 
889 	mtx_lock(&vnode_free_list_mtx);
890 	/*
891 	 * Lend our context to reclaim vnodes if they've exceeded the max.
892 	 */
893 	if (freevnodes > wantfreevnodes)
894 		vnlru_free(1);
895 	/*
896 	 * Wait for available vnodes.
897 	 */
898 	if (numvnodes > desiredvnodes) {
899 		if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) {
900 			/*
901 			 * File system is beeing suspended, we cannot risk a
902 			 * deadlock here, so allocate new vnode anyway.
903 			 */
904 			if (freevnodes > wantfreevnodes)
905 				vnlru_free(freevnodes - wantfreevnodes);
906 			goto alloc;
907 		}
908 		if (vnlruproc_sig == 0) {
909 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
910 			wakeup(vnlruproc);
911 		}
912 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
913 		    "vlruwk", hz);
914 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
915 		if (numvnodes > desiredvnodes) {
916 			mtx_unlock(&vnode_free_list_mtx);
917 			return (ENFILE);
918 		}
919 #endif
920 	}
921 alloc:
922 	numvnodes++;
923 	mtx_unlock(&vnode_free_list_mtx);
924 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
925 	/*
926 	 * Setup locks.
927 	 */
928 	vp->v_vnlock = &vp->v_lock;
929 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
930 	/*
931 	 * By default, don't allow shared locks unless filesystems
932 	 * opt-in.
933 	 */
934 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
935 	/*
936 	 * Initialize bufobj.
937 	 */
938 	bo = &vp->v_bufobj;
939 	bo->__bo_vnode = vp;
940 	bo->bo_mtx = &vp->v_interlock;
941 	bo->bo_ops = &buf_ops_bio;
942 	bo->bo_private = vp;
943 	TAILQ_INIT(&bo->bo_clean.bv_hd);
944 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
945 	/*
946 	 * Initialize namecache.
947 	 */
948 	LIST_INIT(&vp->v_cache_src);
949 	TAILQ_INIT(&vp->v_cache_dst);
950 	/*
951 	 * Finalize various vnode identity bits.
952 	 */
953 	vp->v_type = VNON;
954 	vp->v_tag = tag;
955 	vp->v_op = vops;
956 	v_incr_usecount(vp);
957 	vp->v_data = 0;
958 #ifdef MAC
959 	mac_init_vnode(vp);
960 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
961 		mac_associate_vnode_singlelabel(mp, vp);
962 	else if (mp == NULL)
963 		printf("NULL mp in getnewvnode()\n");
964 #endif
965 	if (mp != NULL) {
966 		bo->bo_bsize = mp->mnt_stat.f_iosize;
967 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
968 			vp->v_vflag |= VV_NOKNOTE;
969 	}
970 
971 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
972 	*vpp = vp;
973 	return (0);
974 }
975 
976 /*
977  * Delete from old mount point vnode list, if on one.
978  */
979 static void
980 delmntque(struct vnode *vp)
981 {
982 	struct mount *mp;
983 
984 	mp = vp->v_mount;
985 	if (mp == NULL)
986 		return;
987 	MNT_ILOCK(mp);
988 	vp->v_mount = NULL;
989 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
990 		("bad mount point vnode list size"));
991 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
992 	mp->mnt_nvnodelistsize--;
993 	MNT_REL(mp);
994 	MNT_IUNLOCK(mp);
995 }
996 
997 static void
998 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
999 {
1000 	struct thread *td;
1001 
1002 	td = curthread; /* XXX ? */
1003 	vp->v_data = NULL;
1004 	vp->v_op = &dead_vnodeops;
1005 	/* XXX non mp-safe fs may still call insmntque with vnode
1006 	   unlocked */
1007 	if (!VOP_ISLOCKED(vp, td))
1008 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1009 	vgone(vp);
1010 	vput(vp);
1011 }
1012 
1013 /*
1014  * Insert into list of vnodes for the new mount point, if available.
1015  */
1016 int
1017 insmntque1(struct vnode *vp, struct mount *mp,
1018 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1019 {
1020 
1021 	KASSERT(vp->v_mount == NULL,
1022 		("insmntque: vnode already on per mount vnode list"));
1023 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1024 	MNT_ILOCK(mp);
1025 	if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1026 	    mp->mnt_nvnodelistsize == 0) {
1027 		MNT_IUNLOCK(mp);
1028 		if (dtr != NULL)
1029 			dtr(vp, dtr_arg);
1030 		return (EBUSY);
1031 	}
1032 	vp->v_mount = mp;
1033 	MNT_REF(mp);
1034 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1035 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1036 		("neg mount point vnode list size"));
1037 	mp->mnt_nvnodelistsize++;
1038 	MNT_IUNLOCK(mp);
1039 	return (0);
1040 }
1041 
1042 int
1043 insmntque(struct vnode *vp, struct mount *mp)
1044 {
1045 
1046 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1047 }
1048 
1049 /*
1050  * Flush out and invalidate all buffers associated with a bufobj
1051  * Called with the underlying object locked.
1052  */
1053 int
1054 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag,
1055     int slptimeo)
1056 {
1057 	int error;
1058 
1059 	BO_LOCK(bo);
1060 	if (flags & V_SAVE) {
1061 		error = bufobj_wwait(bo, slpflag, slptimeo);
1062 		if (error) {
1063 			BO_UNLOCK(bo);
1064 			return (error);
1065 		}
1066 		if (bo->bo_dirty.bv_cnt > 0) {
1067 			BO_UNLOCK(bo);
1068 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
1069 				return (error);
1070 			/*
1071 			 * XXX We could save a lock/unlock if this was only
1072 			 * enabled under INVARIANTS
1073 			 */
1074 			BO_LOCK(bo);
1075 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1076 				panic("vinvalbuf: dirty bufs");
1077 		}
1078 	}
1079 	/*
1080 	 * If you alter this loop please notice that interlock is dropped and
1081 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1082 	 * no race conditions occur from this.
1083 	 */
1084 	do {
1085 		error = flushbuflist(&bo->bo_clean,
1086 		    flags, bo, slpflag, slptimeo);
1087 		if (error == 0)
1088 			error = flushbuflist(&bo->bo_dirty,
1089 			    flags, bo, slpflag, slptimeo);
1090 		if (error != 0 && error != EAGAIN) {
1091 			BO_UNLOCK(bo);
1092 			return (error);
1093 		}
1094 	} while (error != 0);
1095 
1096 	/*
1097 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1098 	 * have write I/O in-progress but if there is a VM object then the
1099 	 * VM object can also have read-I/O in-progress.
1100 	 */
1101 	do {
1102 		bufobj_wwait(bo, 0, 0);
1103 		BO_UNLOCK(bo);
1104 		if (bo->bo_object != NULL) {
1105 			VM_OBJECT_LOCK(bo->bo_object);
1106 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1107 			VM_OBJECT_UNLOCK(bo->bo_object);
1108 		}
1109 		BO_LOCK(bo);
1110 	} while (bo->bo_numoutput > 0);
1111 	BO_UNLOCK(bo);
1112 
1113 	/*
1114 	 * Destroy the copy in the VM cache, too.
1115 	 */
1116 	if (bo->bo_object != NULL) {
1117 		VM_OBJECT_LOCK(bo->bo_object);
1118 		vm_object_page_remove(bo->bo_object, 0, 0,
1119 			(flags & V_SAVE) ? TRUE : FALSE);
1120 		VM_OBJECT_UNLOCK(bo->bo_object);
1121 	}
1122 
1123 #ifdef INVARIANTS
1124 	BO_LOCK(bo);
1125 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1126 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1127 		panic("vinvalbuf: flush failed");
1128 	BO_UNLOCK(bo);
1129 #endif
1130 	return (0);
1131 }
1132 
1133 /*
1134  * Flush out and invalidate all buffers associated with a vnode.
1135  * Called with the underlying object locked.
1136  */
1137 int
1138 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag,
1139     int slptimeo)
1140 {
1141 
1142 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1143 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1144 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1145 }
1146 
1147 /*
1148  * Flush out buffers on the specified list.
1149  *
1150  */
1151 static int
1152 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1153     int slptimeo)
1154 {
1155 	struct buf *bp, *nbp;
1156 	int retval, error;
1157 	daddr_t lblkno;
1158 	b_xflags_t xflags;
1159 
1160 	ASSERT_BO_LOCKED(bo);
1161 
1162 	retval = 0;
1163 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1164 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1165 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1166 			continue;
1167 		}
1168 		lblkno = 0;
1169 		xflags = 0;
1170 		if (nbp != NULL) {
1171 			lblkno = nbp->b_lblkno;
1172 			xflags = nbp->b_xflags &
1173 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1174 		}
1175 		retval = EAGAIN;
1176 		error = BUF_TIMELOCK(bp,
1177 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1178 		    "flushbuf", slpflag, slptimeo);
1179 		if (error) {
1180 			BO_LOCK(bo);
1181 			return (error != ENOLCK ? error : EAGAIN);
1182 		}
1183 		KASSERT(bp->b_bufobj == bo,
1184 		    ("bp %p wrong b_bufobj %p should be %p",
1185 		    bp, bp->b_bufobj, bo));
1186 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1187 			BUF_UNLOCK(bp);
1188 			BO_LOCK(bo);
1189 			return (EAGAIN);
1190 		}
1191 		/*
1192 		 * XXX Since there are no node locks for NFS, I
1193 		 * believe there is a slight chance that a delayed
1194 		 * write will occur while sleeping just above, so
1195 		 * check for it.
1196 		 */
1197 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1198 		    (flags & V_SAVE)) {
1199 			bremfree(bp);
1200 			bp->b_flags |= B_ASYNC;
1201 			bwrite(bp);
1202 			BO_LOCK(bo);
1203 			return (EAGAIN);	/* XXX: why not loop ? */
1204 		}
1205 		bremfree(bp);
1206 		bp->b_flags |= (B_INVAL | B_RELBUF);
1207 		bp->b_flags &= ~B_ASYNC;
1208 		brelse(bp);
1209 		BO_LOCK(bo);
1210 		if (nbp != NULL &&
1211 		    (nbp->b_bufobj != bo ||
1212 		     nbp->b_lblkno != lblkno ||
1213 		     (nbp->b_xflags &
1214 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1215 			break;			/* nbp invalid */
1216 	}
1217 	return (retval);
1218 }
1219 
1220 /*
1221  * Truncate a file's buffer and pages to a specified length.  This
1222  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1223  * sync activity.
1224  */
1225 int
1226 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1227     off_t length, int blksize)
1228 {
1229 	struct buf *bp, *nbp;
1230 	int anyfreed;
1231 	int trunclbn;
1232 	struct bufobj *bo;
1233 
1234 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1235 	/*
1236 	 * Round up to the *next* lbn.
1237 	 */
1238 	trunclbn = (length + blksize - 1) / blksize;
1239 
1240 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1241 restart:
1242 	VI_LOCK(vp);
1243 	bo = &vp->v_bufobj;
1244 	anyfreed = 1;
1245 	for (;anyfreed;) {
1246 		anyfreed = 0;
1247 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1248 			if (bp->b_lblkno < trunclbn)
1249 				continue;
1250 			if (BUF_LOCK(bp,
1251 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1252 			    VI_MTX(vp)) == ENOLCK)
1253 				goto restart;
1254 
1255 			bremfree(bp);
1256 			bp->b_flags |= (B_INVAL | B_RELBUF);
1257 			bp->b_flags &= ~B_ASYNC;
1258 			brelse(bp);
1259 			anyfreed = 1;
1260 
1261 			if (nbp != NULL &&
1262 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1263 			    (nbp->b_vp != vp) ||
1264 			    (nbp->b_flags & B_DELWRI))) {
1265 				goto restart;
1266 			}
1267 			VI_LOCK(vp);
1268 		}
1269 
1270 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1271 			if (bp->b_lblkno < trunclbn)
1272 				continue;
1273 			if (BUF_LOCK(bp,
1274 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1275 			    VI_MTX(vp)) == ENOLCK)
1276 				goto restart;
1277 			bremfree(bp);
1278 			bp->b_flags |= (B_INVAL | B_RELBUF);
1279 			bp->b_flags &= ~B_ASYNC;
1280 			brelse(bp);
1281 			anyfreed = 1;
1282 			if (nbp != NULL &&
1283 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1284 			    (nbp->b_vp != vp) ||
1285 			    (nbp->b_flags & B_DELWRI) == 0)) {
1286 				goto restart;
1287 			}
1288 			VI_LOCK(vp);
1289 		}
1290 	}
1291 
1292 	if (length > 0) {
1293 restartsync:
1294 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1295 			if (bp->b_lblkno > 0)
1296 				continue;
1297 			/*
1298 			 * Since we hold the vnode lock this should only
1299 			 * fail if we're racing with the buf daemon.
1300 			 */
1301 			if (BUF_LOCK(bp,
1302 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1303 			    VI_MTX(vp)) == ENOLCK) {
1304 				goto restart;
1305 			}
1306 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1307 			    ("buf(%p) on dirty queue without DELWRI", bp));
1308 
1309 			bremfree(bp);
1310 			bawrite(bp);
1311 			VI_LOCK(vp);
1312 			goto restartsync;
1313 		}
1314 	}
1315 
1316 	bufobj_wwait(bo, 0, 0);
1317 	VI_UNLOCK(vp);
1318 	vnode_pager_setsize(vp, length);
1319 
1320 	return (0);
1321 }
1322 
1323 /*
1324  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1325  * 		 a vnode.
1326  *
1327  *	NOTE: We have to deal with the special case of a background bitmap
1328  *	buffer, a situation where two buffers will have the same logical
1329  *	block offset.  We want (1) only the foreground buffer to be accessed
1330  *	in a lookup and (2) must differentiate between the foreground and
1331  *	background buffer in the splay tree algorithm because the splay
1332  *	tree cannot normally handle multiple entities with the same 'index'.
1333  *	We accomplish this by adding differentiating flags to the splay tree's
1334  *	numerical domain.
1335  */
1336 static
1337 struct buf *
1338 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1339 {
1340 	struct buf dummy;
1341 	struct buf *lefttreemax, *righttreemin, *y;
1342 
1343 	if (root == NULL)
1344 		return (NULL);
1345 	lefttreemax = righttreemin = &dummy;
1346 	for (;;) {
1347 		if (lblkno < root->b_lblkno ||
1348 		    (lblkno == root->b_lblkno &&
1349 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1350 			if ((y = root->b_left) == NULL)
1351 				break;
1352 			if (lblkno < y->b_lblkno) {
1353 				/* Rotate right. */
1354 				root->b_left = y->b_right;
1355 				y->b_right = root;
1356 				root = y;
1357 				if ((y = root->b_left) == NULL)
1358 					break;
1359 			}
1360 			/* Link into the new root's right tree. */
1361 			righttreemin->b_left = root;
1362 			righttreemin = root;
1363 		} else if (lblkno > root->b_lblkno ||
1364 		    (lblkno == root->b_lblkno &&
1365 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1366 			if ((y = root->b_right) == NULL)
1367 				break;
1368 			if (lblkno > y->b_lblkno) {
1369 				/* Rotate left. */
1370 				root->b_right = y->b_left;
1371 				y->b_left = root;
1372 				root = y;
1373 				if ((y = root->b_right) == NULL)
1374 					break;
1375 			}
1376 			/* Link into the new root's left tree. */
1377 			lefttreemax->b_right = root;
1378 			lefttreemax = root;
1379 		} else {
1380 			break;
1381 		}
1382 		root = y;
1383 	}
1384 	/* Assemble the new root. */
1385 	lefttreemax->b_right = root->b_left;
1386 	righttreemin->b_left = root->b_right;
1387 	root->b_left = dummy.b_right;
1388 	root->b_right = dummy.b_left;
1389 	return (root);
1390 }
1391 
1392 static void
1393 buf_vlist_remove(struct buf *bp)
1394 {
1395 	struct buf *root;
1396 	struct bufv *bv;
1397 
1398 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1399 	ASSERT_BO_LOCKED(bp->b_bufobj);
1400 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1401 	    (BX_VNDIRTY|BX_VNCLEAN),
1402 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1403 	if (bp->b_xflags & BX_VNDIRTY)
1404 		bv = &bp->b_bufobj->bo_dirty;
1405 	else
1406 		bv = &bp->b_bufobj->bo_clean;
1407 	if (bp != bv->bv_root) {
1408 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1409 		KASSERT(root == bp, ("splay lookup failed in remove"));
1410 	}
1411 	if (bp->b_left == NULL) {
1412 		root = bp->b_right;
1413 	} else {
1414 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1415 		root->b_right = bp->b_right;
1416 	}
1417 	bv->bv_root = root;
1418 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1419 	bv->bv_cnt--;
1420 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1421 }
1422 
1423 /*
1424  * Add the buffer to the sorted clean or dirty block list using a
1425  * splay tree algorithm.
1426  *
1427  * NOTE: xflags is passed as a constant, optimizing this inline function!
1428  */
1429 static void
1430 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1431 {
1432 	struct buf *root;
1433 	struct bufv *bv;
1434 
1435 	ASSERT_BO_LOCKED(bo);
1436 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1437 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1438 	bp->b_xflags |= xflags;
1439 	if (xflags & BX_VNDIRTY)
1440 		bv = &bo->bo_dirty;
1441 	else
1442 		bv = &bo->bo_clean;
1443 
1444 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1445 	if (root == NULL) {
1446 		bp->b_left = NULL;
1447 		bp->b_right = NULL;
1448 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1449 	} else if (bp->b_lblkno < root->b_lblkno ||
1450 	    (bp->b_lblkno == root->b_lblkno &&
1451 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1452 		bp->b_left = root->b_left;
1453 		bp->b_right = root;
1454 		root->b_left = NULL;
1455 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1456 	} else {
1457 		bp->b_right = root->b_right;
1458 		bp->b_left = root;
1459 		root->b_right = NULL;
1460 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1461 	}
1462 	bv->bv_cnt++;
1463 	bv->bv_root = bp;
1464 }
1465 
1466 /*
1467  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1468  * shadow buffers used in background bitmap writes.
1469  *
1470  * This code isn't quite efficient as it could be because we are maintaining
1471  * two sorted lists and do not know which list the block resides in.
1472  *
1473  * During a "make buildworld" the desired buffer is found at one of
1474  * the roots more than 60% of the time.  Thus, checking both roots
1475  * before performing either splay eliminates unnecessary splays on the
1476  * first tree splayed.
1477  */
1478 struct buf *
1479 gbincore(struct bufobj *bo, daddr_t lblkno)
1480 {
1481 	struct buf *bp;
1482 
1483 	ASSERT_BO_LOCKED(bo);
1484 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1485 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1486 		return (bp);
1487 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1488 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1489 		return (bp);
1490 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1491 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1492 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1493 			return (bp);
1494 	}
1495 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1496 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1497 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1498 			return (bp);
1499 	}
1500 	return (NULL);
1501 }
1502 
1503 /*
1504  * Associate a buffer with a vnode.
1505  */
1506 void
1507 bgetvp(struct vnode *vp, struct buf *bp)
1508 {
1509 
1510 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1511 
1512 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1513 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1514 	    ("bgetvp: bp already attached! %p", bp));
1515 
1516 	ASSERT_VI_LOCKED(vp, "bgetvp");
1517 	vholdl(vp);
1518 	if (VFS_NEEDSGIANT(vp->v_mount) ||
1519 	    vp->v_bufobj.bo_flag & BO_NEEDSGIANT)
1520 		bp->b_flags |= B_NEEDSGIANT;
1521 	bp->b_vp = vp;
1522 	bp->b_bufobj = &vp->v_bufobj;
1523 	/*
1524 	 * Insert onto list for new vnode.
1525 	 */
1526 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1527 }
1528 
1529 /*
1530  * Disassociate a buffer from a vnode.
1531  */
1532 void
1533 brelvp(struct buf *bp)
1534 {
1535 	struct bufobj *bo;
1536 	struct vnode *vp;
1537 
1538 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1539 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1540 
1541 	/*
1542 	 * Delete from old vnode list, if on one.
1543 	 */
1544 	vp = bp->b_vp;		/* XXX */
1545 	bo = bp->b_bufobj;
1546 	BO_LOCK(bo);
1547 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1548 		buf_vlist_remove(bp);
1549 	else
1550 		panic("brelvp: Buffer %p not on queue.", bp);
1551 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1552 		bo->bo_flag &= ~BO_ONWORKLST;
1553 		mtx_lock(&sync_mtx);
1554 		LIST_REMOVE(bo, bo_synclist);
1555 		syncer_worklist_len--;
1556 		mtx_unlock(&sync_mtx);
1557 	}
1558 	bp->b_flags &= ~B_NEEDSGIANT;
1559 	bp->b_vp = NULL;
1560 	bp->b_bufobj = NULL;
1561 	vdropl(vp);
1562 }
1563 
1564 /*
1565  * Add an item to the syncer work queue.
1566  */
1567 static void
1568 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1569 {
1570 	int slot;
1571 
1572 	ASSERT_BO_LOCKED(bo);
1573 
1574 	mtx_lock(&sync_mtx);
1575 	if (bo->bo_flag & BO_ONWORKLST)
1576 		LIST_REMOVE(bo, bo_synclist);
1577 	else {
1578 		bo->bo_flag |= BO_ONWORKLST;
1579 		syncer_worklist_len++;
1580 	}
1581 
1582 	if (delay > syncer_maxdelay - 2)
1583 		delay = syncer_maxdelay - 2;
1584 	slot = (syncer_delayno + delay) & syncer_mask;
1585 
1586 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1587 	mtx_unlock(&sync_mtx);
1588 }
1589 
1590 static int
1591 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1592 {
1593 	int error, len;
1594 
1595 	mtx_lock(&sync_mtx);
1596 	len = syncer_worklist_len - sync_vnode_count;
1597 	mtx_unlock(&sync_mtx);
1598 	error = SYSCTL_OUT(req, &len, sizeof(len));
1599 	return (error);
1600 }
1601 
1602 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1603     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1604 
1605 static struct proc *updateproc;
1606 static void sched_sync(void);
1607 static struct kproc_desc up_kp = {
1608 	"syncer",
1609 	sched_sync,
1610 	&updateproc
1611 };
1612 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1613 
1614 static int
1615 sync_vnode(struct bufobj *bo, struct thread *td)
1616 {
1617 	struct vnode *vp;
1618 	struct mount *mp;
1619 
1620 	vp = bo->__bo_vnode;	/* XXX */
1621 	if (VOP_ISLOCKED(vp, NULL) != 0)
1622 		return (1);
1623 	if (VI_TRYLOCK(vp) == 0)
1624 		return (1);
1625 	/*
1626 	 * We use vhold in case the vnode does not
1627 	 * successfully sync.  vhold prevents the vnode from
1628 	 * going away when we unlock the sync_mtx so that
1629 	 * we can acquire the vnode interlock.
1630 	 */
1631 	vholdl(vp);
1632 	mtx_unlock(&sync_mtx);
1633 	VI_UNLOCK(vp);
1634 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1635 		vdrop(vp);
1636 		mtx_lock(&sync_mtx);
1637 		return (1);
1638 	}
1639 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1640 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1641 	VOP_UNLOCK(vp, 0, td);
1642 	vn_finished_write(mp);
1643 	VI_LOCK(vp);
1644 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1645 		/*
1646 		 * Put us back on the worklist.  The worklist
1647 		 * routine will remove us from our current
1648 		 * position and then add us back in at a later
1649 		 * position.
1650 		 */
1651 		vn_syncer_add_to_worklist(bo, syncdelay);
1652 	}
1653 	vdropl(vp);
1654 	mtx_lock(&sync_mtx);
1655 	return (0);
1656 }
1657 
1658 /*
1659  * System filesystem synchronizer daemon.
1660  */
1661 static void
1662 sched_sync(void)
1663 {
1664 	struct synclist *next;
1665 	struct synclist *slp;
1666 	struct bufobj *bo;
1667 	long starttime;
1668 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1669 	static int dummychan;
1670 	int last_work_seen;
1671 	int net_worklist_len;
1672 	int syncer_final_iter;
1673 	int first_printf;
1674 	int error;
1675 
1676 	mtx_lock(&Giant);
1677 	last_work_seen = 0;
1678 	syncer_final_iter = 0;
1679 	first_printf = 1;
1680 	syncer_state = SYNCER_RUNNING;
1681 	starttime = time_uptime;
1682 	td->td_pflags |= TDP_NORUNNINGBUF;
1683 
1684 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1685 	    SHUTDOWN_PRI_LAST);
1686 
1687 	mtx_lock(&sync_mtx);
1688 	for (;;) {
1689 		if (syncer_state == SYNCER_FINAL_DELAY &&
1690 		    syncer_final_iter == 0) {
1691 			mtx_unlock(&sync_mtx);
1692 			kthread_suspend_check(td->td_proc);
1693 			mtx_lock(&sync_mtx);
1694 		}
1695 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1696 		if (syncer_state != SYNCER_RUNNING &&
1697 		    starttime != time_uptime) {
1698 			if (first_printf) {
1699 				printf("\nSyncing disks, vnodes remaining...");
1700 				first_printf = 0;
1701 			}
1702 			printf("%d ", net_worklist_len);
1703 		}
1704 		starttime = time_uptime;
1705 
1706 		/*
1707 		 * Push files whose dirty time has expired.  Be careful
1708 		 * of interrupt race on slp queue.
1709 		 *
1710 		 * Skip over empty worklist slots when shutting down.
1711 		 */
1712 		do {
1713 			slp = &syncer_workitem_pending[syncer_delayno];
1714 			syncer_delayno += 1;
1715 			if (syncer_delayno == syncer_maxdelay)
1716 				syncer_delayno = 0;
1717 			next = &syncer_workitem_pending[syncer_delayno];
1718 			/*
1719 			 * If the worklist has wrapped since the
1720 			 * it was emptied of all but syncer vnodes,
1721 			 * switch to the FINAL_DELAY state and run
1722 			 * for one more second.
1723 			 */
1724 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1725 			    net_worklist_len == 0 &&
1726 			    last_work_seen == syncer_delayno) {
1727 				syncer_state = SYNCER_FINAL_DELAY;
1728 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1729 			}
1730 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1731 		    syncer_worklist_len > 0);
1732 
1733 		/*
1734 		 * Keep track of the last time there was anything
1735 		 * on the worklist other than syncer vnodes.
1736 		 * Return to the SHUTTING_DOWN state if any
1737 		 * new work appears.
1738 		 */
1739 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1740 			last_work_seen = syncer_delayno;
1741 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1742 			syncer_state = SYNCER_SHUTTING_DOWN;
1743 		while ((bo = LIST_FIRST(slp)) != NULL) {
1744 			error = sync_vnode(bo, td);
1745 			if (error == 1) {
1746 				LIST_REMOVE(bo, bo_synclist);
1747 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1748 				continue;
1749 			}
1750 		}
1751 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1752 			syncer_final_iter--;
1753 		/*
1754 		 * The variable rushjob allows the kernel to speed up the
1755 		 * processing of the filesystem syncer process. A rushjob
1756 		 * value of N tells the filesystem syncer to process the next
1757 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1758 		 * is used by the soft update code to speed up the filesystem
1759 		 * syncer process when the incore state is getting so far
1760 		 * ahead of the disk that the kernel memory pool is being
1761 		 * threatened with exhaustion.
1762 		 */
1763 		if (rushjob > 0) {
1764 			rushjob -= 1;
1765 			continue;
1766 		}
1767 		/*
1768 		 * Just sleep for a short period of time between
1769 		 * iterations when shutting down to allow some I/O
1770 		 * to happen.
1771 		 *
1772 		 * If it has taken us less than a second to process the
1773 		 * current work, then wait. Otherwise start right over
1774 		 * again. We can still lose time if any single round
1775 		 * takes more than two seconds, but it does not really
1776 		 * matter as we are just trying to generally pace the
1777 		 * filesystem activity.
1778 		 */
1779 		if (syncer_state != SYNCER_RUNNING)
1780 			msleep(&dummychan, &sync_mtx, PPAUSE, "syncfnl",
1781 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1782 		else if (time_uptime == starttime)
1783 			msleep(&lbolt, &sync_mtx, PPAUSE, "syncer", 0);
1784 	}
1785 }
1786 
1787 /*
1788  * Request the syncer daemon to speed up its work.
1789  * We never push it to speed up more than half of its
1790  * normal turn time, otherwise it could take over the cpu.
1791  */
1792 int
1793 speedup_syncer(void)
1794 {
1795 	struct thread *td;
1796 	int ret = 0;
1797 
1798 	td = FIRST_THREAD_IN_PROC(updateproc);
1799 	sleepq_remove(td, &lbolt);
1800 	mtx_lock(&sync_mtx);
1801 	if (rushjob < syncdelay / 2) {
1802 		rushjob += 1;
1803 		stat_rush_requests += 1;
1804 		ret = 1;
1805 	}
1806 	mtx_unlock(&sync_mtx);
1807 	return (ret);
1808 }
1809 
1810 /*
1811  * Tell the syncer to speed up its work and run though its work
1812  * list several times, then tell it to shut down.
1813  */
1814 static void
1815 syncer_shutdown(void *arg, int howto)
1816 {
1817 	struct thread *td;
1818 
1819 	if (howto & RB_NOSYNC)
1820 		return;
1821 	td = FIRST_THREAD_IN_PROC(updateproc);
1822 	sleepq_remove(td, &lbolt);
1823 	mtx_lock(&sync_mtx);
1824 	syncer_state = SYNCER_SHUTTING_DOWN;
1825 	rushjob = 0;
1826 	mtx_unlock(&sync_mtx);
1827 	kproc_shutdown(arg, howto);
1828 }
1829 
1830 /*
1831  * Reassign a buffer from one vnode to another.
1832  * Used to assign file specific control information
1833  * (indirect blocks) to the vnode to which they belong.
1834  */
1835 void
1836 reassignbuf(struct buf *bp)
1837 {
1838 	struct vnode *vp;
1839 	struct bufobj *bo;
1840 	int delay;
1841 #ifdef INVARIANTS
1842 	struct bufv *bv;
1843 #endif
1844 
1845 	vp = bp->b_vp;
1846 	bo = bp->b_bufobj;
1847 	++reassignbufcalls;
1848 
1849 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1850 	    bp, bp->b_vp, bp->b_flags);
1851 	/*
1852 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1853 	 * is not fully linked in.
1854 	 */
1855 	if (bp->b_flags & B_PAGING)
1856 		panic("cannot reassign paging buffer");
1857 
1858 	/*
1859 	 * Delete from old vnode list, if on one.
1860 	 */
1861 	VI_LOCK(vp);
1862 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1863 		buf_vlist_remove(bp);
1864 	else
1865 		panic("reassignbuf: Buffer %p not on queue.", bp);
1866 	/*
1867 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1868 	 * of clean buffers.
1869 	 */
1870 	if (bp->b_flags & B_DELWRI) {
1871 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1872 			switch (vp->v_type) {
1873 			case VDIR:
1874 				delay = dirdelay;
1875 				break;
1876 			case VCHR:
1877 				delay = metadelay;
1878 				break;
1879 			default:
1880 				delay = filedelay;
1881 			}
1882 			vn_syncer_add_to_worklist(bo, delay);
1883 		}
1884 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1885 	} else {
1886 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1887 
1888 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1889 			mtx_lock(&sync_mtx);
1890 			LIST_REMOVE(bo, bo_synclist);
1891 			syncer_worklist_len--;
1892 			mtx_unlock(&sync_mtx);
1893 			bo->bo_flag &= ~BO_ONWORKLST;
1894 		}
1895 	}
1896 #ifdef INVARIANTS
1897 	bv = &bo->bo_clean;
1898 	bp = TAILQ_FIRST(&bv->bv_hd);
1899 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1900 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1901 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1902 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1903 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1904 	bv = &bo->bo_dirty;
1905 	bp = TAILQ_FIRST(&bv->bv_hd);
1906 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1907 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1908 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1909 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1910 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1911 #endif
1912 	VI_UNLOCK(vp);
1913 }
1914 
1915 /*
1916  * Increment the use and hold counts on the vnode, taking care to reference
1917  * the driver's usecount if this is a chardev.  The vholdl() will remove
1918  * the vnode from the free list if it is presently free.  Requires the
1919  * vnode interlock and returns with it held.
1920  */
1921 static void
1922 v_incr_usecount(struct vnode *vp)
1923 {
1924 
1925 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1926 	    vp, vp->v_holdcnt, vp->v_usecount);
1927 	vp->v_usecount++;
1928 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1929 		dev_lock();
1930 		vp->v_rdev->si_usecount++;
1931 		dev_unlock();
1932 	}
1933 	vholdl(vp);
1934 }
1935 
1936 /*
1937  * Turn a holdcnt into a use+holdcnt such that only one call to
1938  * v_decr_usecount is needed.
1939  */
1940 static void
1941 v_upgrade_usecount(struct vnode *vp)
1942 {
1943 
1944 	CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n",
1945 	    vp, vp->v_holdcnt, vp->v_usecount);
1946 	vp->v_usecount++;
1947 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1948 		dev_lock();
1949 		vp->v_rdev->si_usecount++;
1950 		dev_unlock();
1951 	}
1952 }
1953 
1954 /*
1955  * Decrement the vnode use and hold count along with the driver's usecount
1956  * if this is a chardev.  The vdropl() below releases the vnode interlock
1957  * as it may free the vnode.
1958  */
1959 static void
1960 v_decr_usecount(struct vnode *vp)
1961 {
1962 
1963 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1964 	    vp, vp->v_holdcnt, vp->v_usecount);
1965 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1966 	VNASSERT(vp->v_usecount > 0, vp,
1967 	    ("v_decr_usecount: negative usecount"));
1968 	vp->v_usecount--;
1969 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1970 		dev_lock();
1971 		vp->v_rdev->si_usecount--;
1972 		dev_unlock();
1973 	}
1974 	vdropl(vp);
1975 }
1976 
1977 /*
1978  * Decrement only the use count and driver use count.  This is intended to
1979  * be paired with a follow on vdropl() to release the remaining hold count.
1980  * In this way we may vgone() a vnode with a 0 usecount without risk of
1981  * having it end up on a free list because the hold count is kept above 0.
1982  */
1983 static void
1984 v_decr_useonly(struct vnode *vp)
1985 {
1986 
1987 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1988 	    vp, vp->v_holdcnt, vp->v_usecount);
1989 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1990 	VNASSERT(vp->v_usecount > 0, vp,
1991 	    ("v_decr_useonly: negative usecount"));
1992 	vp->v_usecount--;
1993 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1994 		dev_lock();
1995 		vp->v_rdev->si_usecount--;
1996 		dev_unlock();
1997 	}
1998 }
1999 
2000 /*
2001  * Grab a particular vnode from the free list, increment its
2002  * reference count and lock it. The vnode lock bit is set if the
2003  * vnode is being eliminated in vgone. The process is awakened
2004  * when the transition is completed, and an error returned to
2005  * indicate that the vnode is no longer usable (possibly having
2006  * been changed to a new filesystem type).
2007  */
2008 int
2009 vget(struct vnode *vp, int flags, struct thread *td)
2010 {
2011 	int oweinact;
2012 	int oldflags;
2013 	int error;
2014 
2015 	error = 0;
2016 	oldflags = flags;
2017 	oweinact = 0;
2018 	VFS_ASSERT_GIANT(vp->v_mount);
2019 	if ((flags & LK_INTERLOCK) == 0)
2020 		VI_LOCK(vp);
2021 	/*
2022 	 * If the inactive call was deferred because vput() was called
2023 	 * with a shared lock, we have to do it here before another thread
2024 	 * gets a reference to data that should be dead.
2025 	 */
2026 	if (vp->v_iflag & VI_OWEINACT) {
2027 		if (flags & LK_NOWAIT) {
2028 			VI_UNLOCK(vp);
2029 			return (EBUSY);
2030 		}
2031 		flags &= ~LK_TYPE_MASK;
2032 		flags |= LK_EXCLUSIVE;
2033 		oweinact = 1;
2034 	}
2035 	vholdl(vp);
2036 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2037 		vdrop(vp);
2038 		return (error);
2039 	}
2040 	VI_LOCK(vp);
2041 	/* Upgrade our holdcnt to a usecount. */
2042 	v_upgrade_usecount(vp);
2043 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2044 		panic("vget: vn_lock failed to return ENOENT\n");
2045 	if (oweinact) {
2046 		if (vp->v_iflag & VI_OWEINACT)
2047 			vinactive(vp, td);
2048 		VI_UNLOCK(vp);
2049 		if ((oldflags & LK_TYPE_MASK) == 0)
2050 			VOP_UNLOCK(vp, 0, td);
2051 	} else
2052 		VI_UNLOCK(vp);
2053 	return (0);
2054 }
2055 
2056 /*
2057  * Increase the reference count of a vnode.
2058  */
2059 void
2060 vref(struct vnode *vp)
2061 {
2062 
2063 	VI_LOCK(vp);
2064 	v_incr_usecount(vp);
2065 	VI_UNLOCK(vp);
2066 }
2067 
2068 /*
2069  * Return reference count of a vnode.
2070  *
2071  * The results of this call are only guaranteed when some mechanism other
2072  * than the VI lock is used to stop other processes from gaining references
2073  * to the vnode.  This may be the case if the caller holds the only reference.
2074  * This is also useful when stale data is acceptable as race conditions may
2075  * be accounted for by some other means.
2076  */
2077 int
2078 vrefcnt(struct vnode *vp)
2079 {
2080 	int usecnt;
2081 
2082 	VI_LOCK(vp);
2083 	usecnt = vp->v_usecount;
2084 	VI_UNLOCK(vp);
2085 
2086 	return (usecnt);
2087 }
2088 
2089 
2090 /*
2091  * Vnode put/release.
2092  * If count drops to zero, call inactive routine and return to freelist.
2093  */
2094 void
2095 vrele(struct vnode *vp)
2096 {
2097 	struct thread *td = curthread;	/* XXX */
2098 
2099 	KASSERT(vp != NULL, ("vrele: null vp"));
2100 	VFS_ASSERT_GIANT(vp->v_mount);
2101 
2102 	VI_LOCK(vp);
2103 
2104 	/* Skip this v_writecount check if we're going to panic below. */
2105 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2106 	    ("vrele: missed vn_close"));
2107 
2108 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2109 	    vp->v_usecount == 1)) {
2110 		v_decr_usecount(vp);
2111 		return;
2112 	}
2113 	if (vp->v_usecount != 1) {
2114 #ifdef DIAGNOSTIC
2115 		vprint("vrele: negative ref count", vp);
2116 #endif
2117 		VI_UNLOCK(vp);
2118 		panic("vrele: negative ref cnt");
2119 	}
2120 	/*
2121 	 * We want to hold the vnode until the inactive finishes to
2122 	 * prevent vgone() races.  We drop the use count here and the
2123 	 * hold count below when we're done.
2124 	 */
2125 	v_decr_useonly(vp);
2126 	/*
2127 	 * We must call VOP_INACTIVE with the node locked. Mark
2128 	 * as VI_DOINGINACT to avoid recursion.
2129 	 */
2130 	vp->v_iflag |= VI_OWEINACT;
2131 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2132 		VI_LOCK(vp);
2133 		if (vp->v_usecount > 0)
2134 			vp->v_iflag &= ~VI_OWEINACT;
2135 		if (vp->v_iflag & VI_OWEINACT)
2136 			vinactive(vp, td);
2137 		VOP_UNLOCK(vp, 0, td);
2138 	} else {
2139 		VI_LOCK(vp);
2140 		if (vp->v_usecount > 0)
2141 			vp->v_iflag &= ~VI_OWEINACT;
2142 	}
2143 	vdropl(vp);
2144 }
2145 
2146 /*
2147  * Release an already locked vnode.  This give the same effects as
2148  * unlock+vrele(), but takes less time and avoids releasing and
2149  * re-aquiring the lock (as vrele() aquires the lock internally.)
2150  */
2151 void
2152 vput(struct vnode *vp)
2153 {
2154 	struct thread *td = curthread;	/* XXX */
2155 	int error;
2156 
2157 	KASSERT(vp != NULL, ("vput: null vp"));
2158 	ASSERT_VOP_LOCKED(vp, "vput");
2159 	VFS_ASSERT_GIANT(vp->v_mount);
2160 	VI_LOCK(vp);
2161 	/* Skip this v_writecount check if we're going to panic below. */
2162 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2163 	    ("vput: missed vn_close"));
2164 	error = 0;
2165 
2166 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2167 	    vp->v_usecount == 1)) {
2168 		VOP_UNLOCK(vp, 0, td);
2169 		v_decr_usecount(vp);
2170 		return;
2171 	}
2172 
2173 	if (vp->v_usecount != 1) {
2174 #ifdef DIAGNOSTIC
2175 		vprint("vput: negative ref count", vp);
2176 #endif
2177 		panic("vput: negative ref cnt");
2178 	}
2179 	/*
2180 	 * We want to hold the vnode until the inactive finishes to
2181 	 * prevent vgone() races.  We drop the use count here and the
2182 	 * hold count below when we're done.
2183 	 */
2184 	v_decr_useonly(vp);
2185 	vp->v_iflag |= VI_OWEINACT;
2186 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2187 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2188 		VI_LOCK(vp);
2189 		if (error) {
2190 			if (vp->v_usecount > 0)
2191 				vp->v_iflag &= ~VI_OWEINACT;
2192 			goto done;
2193 		}
2194 	}
2195 	if (vp->v_usecount > 0)
2196 		vp->v_iflag &= ~VI_OWEINACT;
2197 	if (vp->v_iflag & VI_OWEINACT)
2198 		vinactive(vp, td);
2199 	VOP_UNLOCK(vp, 0, td);
2200 done:
2201 	vdropl(vp);
2202 }
2203 
2204 /*
2205  * Somebody doesn't want the vnode recycled.
2206  */
2207 void
2208 vhold(struct vnode *vp)
2209 {
2210 
2211 	VI_LOCK(vp);
2212 	vholdl(vp);
2213 	VI_UNLOCK(vp);
2214 }
2215 
2216 void
2217 vholdl(struct vnode *vp)
2218 {
2219 
2220 	vp->v_holdcnt++;
2221 	if (VSHOULDBUSY(vp))
2222 		vbusy(vp);
2223 }
2224 
2225 /*
2226  * Note that there is one less who cares about this vnode.  vdrop() is the
2227  * opposite of vhold().
2228  */
2229 void
2230 vdrop(struct vnode *vp)
2231 {
2232 
2233 	VI_LOCK(vp);
2234 	vdropl(vp);
2235 }
2236 
2237 /*
2238  * Drop the hold count of the vnode.  If this is the last reference to
2239  * the vnode we will free it if it has been vgone'd otherwise it is
2240  * placed on the free list.
2241  */
2242 void
2243 vdropl(struct vnode *vp)
2244 {
2245 
2246 	ASSERT_VI_LOCKED(vp, "vdropl");
2247 	if (vp->v_holdcnt <= 0)
2248 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2249 	vp->v_holdcnt--;
2250 	if (vp->v_holdcnt == 0) {
2251 		if (vp->v_iflag & VI_DOOMED) {
2252 			vdestroy(vp);
2253 			return;
2254 		} else
2255 			vfree(vp);
2256 	}
2257 	VI_UNLOCK(vp);
2258 }
2259 
2260 /*
2261  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2262  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2263  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2264  * failed lock upgrade.
2265  */
2266 static void
2267 vinactive(struct vnode *vp, struct thread *td)
2268 {
2269 
2270 	ASSERT_VOP_LOCKED(vp, "vinactive");
2271 	ASSERT_VI_LOCKED(vp, "vinactive");
2272 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2273 	    ("vinactive: recursed on VI_DOINGINACT"));
2274 	vp->v_iflag |= VI_DOINGINACT;
2275 	vp->v_iflag &= ~VI_OWEINACT;
2276 	VI_UNLOCK(vp);
2277 	VOP_INACTIVE(vp, td);
2278 	VI_LOCK(vp);
2279 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2280 	    ("vinactive: lost VI_DOINGINACT"));
2281 	vp->v_iflag &= ~VI_DOINGINACT;
2282 }
2283 
2284 /*
2285  * Remove any vnodes in the vnode table belonging to mount point mp.
2286  *
2287  * If FORCECLOSE is not specified, there should not be any active ones,
2288  * return error if any are found (nb: this is a user error, not a
2289  * system error). If FORCECLOSE is specified, detach any active vnodes
2290  * that are found.
2291  *
2292  * If WRITECLOSE is set, only flush out regular file vnodes open for
2293  * writing.
2294  *
2295  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2296  *
2297  * `rootrefs' specifies the base reference count for the root vnode
2298  * of this filesystem. The root vnode is considered busy if its
2299  * v_usecount exceeds this value. On a successful return, vflush(, td)
2300  * will call vrele() on the root vnode exactly rootrefs times.
2301  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2302  * be zero.
2303  */
2304 #ifdef DIAGNOSTIC
2305 static int busyprt = 0;		/* print out busy vnodes */
2306 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2307 #endif
2308 
2309 int
2310 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td)
2311 {
2312 	struct vnode *vp, *mvp, *rootvp = NULL;
2313 	struct vattr vattr;
2314 	int busy = 0, error;
2315 
2316 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2317 	if (rootrefs > 0) {
2318 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2319 		    ("vflush: bad args"));
2320 		/*
2321 		 * Get the filesystem root vnode. We can vput() it
2322 		 * immediately, since with rootrefs > 0, it won't go away.
2323 		 */
2324 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2325 			return (error);
2326 		vput(rootvp);
2327 
2328 	}
2329 	MNT_ILOCK(mp);
2330 loop:
2331 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2332 
2333 		VI_LOCK(vp);
2334 		vholdl(vp);
2335 		MNT_IUNLOCK(mp);
2336 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2337 		if (error) {
2338 			vdrop(vp);
2339 			MNT_ILOCK(mp);
2340 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2341 			goto loop;
2342 		}
2343 		/*
2344 		 * Skip over a vnodes marked VV_SYSTEM.
2345 		 */
2346 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2347 			VOP_UNLOCK(vp, 0, td);
2348 			vdrop(vp);
2349 			MNT_ILOCK(mp);
2350 			continue;
2351 		}
2352 		/*
2353 		 * If WRITECLOSE is set, flush out unlinked but still open
2354 		 * files (even if open only for reading) and regular file
2355 		 * vnodes open for writing.
2356 		 */
2357 		if (flags & WRITECLOSE) {
2358 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2359 			VI_LOCK(vp);
2360 
2361 			if ((vp->v_type == VNON ||
2362 			    (error == 0 && vattr.va_nlink > 0)) &&
2363 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2364 				VOP_UNLOCK(vp, 0, td);
2365 				vdropl(vp);
2366 				MNT_ILOCK(mp);
2367 				continue;
2368 			}
2369 		} else
2370 			VI_LOCK(vp);
2371 		/*
2372 		 * With v_usecount == 0, all we need to do is clear out the
2373 		 * vnode data structures and we are done.
2374 		 *
2375 		 * If FORCECLOSE is set, forcibly close the vnode.
2376 		 */
2377 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2378 			VNASSERT(vp->v_usecount == 0 ||
2379 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2380 			    ("device VNODE %p is FORCECLOSED", vp));
2381 			vgonel(vp);
2382 		} else {
2383 			busy++;
2384 #ifdef DIAGNOSTIC
2385 			if (busyprt)
2386 				vprint("vflush: busy vnode", vp);
2387 #endif
2388 		}
2389 		VOP_UNLOCK(vp, 0, td);
2390 		vdropl(vp);
2391 		MNT_ILOCK(mp);
2392 	}
2393 	MNT_IUNLOCK(mp);
2394 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2395 		/*
2396 		 * If just the root vnode is busy, and if its refcount
2397 		 * is equal to `rootrefs', then go ahead and kill it.
2398 		 */
2399 		VI_LOCK(rootvp);
2400 		KASSERT(busy > 0, ("vflush: not busy"));
2401 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2402 		    ("vflush: usecount %d < rootrefs %d",
2403 		     rootvp->v_usecount, rootrefs));
2404 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2405 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2406 			vgone(rootvp);
2407 			VOP_UNLOCK(rootvp, 0, td);
2408 			busy = 0;
2409 		} else
2410 			VI_UNLOCK(rootvp);
2411 	}
2412 	if (busy)
2413 		return (EBUSY);
2414 	for (; rootrefs > 0; rootrefs--)
2415 		vrele(rootvp);
2416 	return (0);
2417 }
2418 
2419 /*
2420  * Recycle an unused vnode to the front of the free list.
2421  */
2422 int
2423 vrecycle(struct vnode *vp, struct thread *td)
2424 {
2425 	int recycled;
2426 
2427 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2428 	recycled = 0;
2429 	VI_LOCK(vp);
2430 	if (vp->v_usecount == 0) {
2431 		recycled = 1;
2432 		vgonel(vp);
2433 	}
2434 	VI_UNLOCK(vp);
2435 	return (recycled);
2436 }
2437 
2438 /*
2439  * Eliminate all activity associated with a vnode
2440  * in preparation for reuse.
2441  */
2442 void
2443 vgone(struct vnode *vp)
2444 {
2445 	VI_LOCK(vp);
2446 	vgonel(vp);
2447 	VI_UNLOCK(vp);
2448 }
2449 
2450 /*
2451  * vgone, with the vp interlock held.
2452  */
2453 void
2454 vgonel(struct vnode *vp)
2455 {
2456 	struct thread *td;
2457 	int oweinact;
2458 	int active;
2459 	struct mount *mp;
2460 
2461 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2462 	ASSERT_VOP_LOCKED(vp, "vgonel");
2463 	ASSERT_VI_LOCKED(vp, "vgonel");
2464 	VNASSERT(vp->v_holdcnt, vp,
2465 	    ("vgonel: vp %p has no reference.", vp));
2466 	td = curthread;
2467 
2468 	/*
2469 	 * Don't vgonel if we're already doomed.
2470 	 */
2471 	if (vp->v_iflag & VI_DOOMED)
2472 		return;
2473 	vp->v_iflag |= VI_DOOMED;
2474 	/*
2475 	 * Check to see if the vnode is in use.  If so, we have to call
2476 	 * VOP_CLOSE() and VOP_INACTIVE().
2477 	 */
2478 	active = vp->v_usecount;
2479 	oweinact = (vp->v_iflag & VI_OWEINACT);
2480 	VI_UNLOCK(vp);
2481 	/*
2482 	 * Clean out any buffers associated with the vnode.
2483 	 * If the flush fails, just toss the buffers.
2484 	 */
2485 	mp = NULL;
2486 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2487 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
2488 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2489 		vinvalbuf(vp, 0, td, 0, 0);
2490 
2491 	/*
2492 	 * If purging an active vnode, it must be closed and
2493 	 * deactivated before being reclaimed.
2494 	 */
2495 	if (active)
2496 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2497 	if (oweinact || active) {
2498 		VI_LOCK(vp);
2499 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2500 			vinactive(vp, td);
2501 		VI_UNLOCK(vp);
2502 	}
2503 	/*
2504 	 * Reclaim the vnode.
2505 	 */
2506 	if (VOP_RECLAIM(vp, td))
2507 		panic("vgone: cannot reclaim");
2508 	if (mp != NULL)
2509 		vn_finished_secondary_write(mp);
2510 	VNASSERT(vp->v_object == NULL, vp,
2511 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2512 	/*
2513 	 * Delete from old mount point vnode list.
2514 	 */
2515 	delmntque(vp);
2516 	cache_purge(vp);
2517 	/*
2518 	 * Done with purge, reset to the standard lock and invalidate
2519 	 * the vnode.
2520 	 */
2521 	VI_LOCK(vp);
2522 	vp->v_vnlock = &vp->v_lock;
2523 	vp->v_op = &dead_vnodeops;
2524 	vp->v_tag = "none";
2525 	vp->v_type = VBAD;
2526 }
2527 
2528 /*
2529  * Calculate the total number of references to a special device.
2530  */
2531 int
2532 vcount(struct vnode *vp)
2533 {
2534 	int count;
2535 
2536 	dev_lock();
2537 	count = vp->v_rdev->si_usecount;
2538 	dev_unlock();
2539 	return (count);
2540 }
2541 
2542 /*
2543  * Same as above, but using the struct cdev *as argument
2544  */
2545 int
2546 count_dev(struct cdev *dev)
2547 {
2548 	int count;
2549 
2550 	dev_lock();
2551 	count = dev->si_usecount;
2552 	dev_unlock();
2553 	return(count);
2554 }
2555 
2556 /*
2557  * Print out a description of a vnode.
2558  */
2559 static char *typename[] =
2560 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2561  "VMARKER"};
2562 
2563 void
2564 vn_printf(struct vnode *vp, const char *fmt, ...)
2565 {
2566 	va_list ap;
2567 	char buf[96];
2568 
2569 	va_start(ap, fmt);
2570 	vprintf(fmt, ap);
2571 	va_end(ap);
2572 	printf("%p: ", (void *)vp);
2573 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2574 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2575 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2576 	buf[0] = '\0';
2577 	buf[1] = '\0';
2578 	if (vp->v_vflag & VV_ROOT)
2579 		strcat(buf, "|VV_ROOT");
2580 	if (vp->v_vflag & VV_TEXT)
2581 		strcat(buf, "|VV_TEXT");
2582 	if (vp->v_vflag & VV_SYSTEM)
2583 		strcat(buf, "|VV_SYSTEM");
2584 	if (vp->v_vflag & VV_DELETED)
2585 		strcat(buf, "|VV_DELETED");
2586 	if (vp->v_iflag & VI_DOOMED)
2587 		strcat(buf, "|VI_DOOMED");
2588 	if (vp->v_iflag & VI_FREE)
2589 		strcat(buf, "|VI_FREE");
2590 	printf("    flags (%s)\n", buf + 1);
2591 	if (mtx_owned(VI_MTX(vp)))
2592 		printf(" VI_LOCKed");
2593 	if (vp->v_object != NULL)
2594 		printf("    v_object %p ref %d pages %d\n",
2595 		    vp->v_object, vp->v_object->ref_count,
2596 		    vp->v_object->resident_page_count);
2597 	printf("    ");
2598 	lockmgr_printinfo(vp->v_vnlock);
2599 	printf("\n");
2600 	if (vp->v_data != NULL)
2601 		VOP_PRINT(vp);
2602 }
2603 
2604 #ifdef DDB
2605 /*
2606  * List all of the locked vnodes in the system.
2607  * Called when debugging the kernel.
2608  */
2609 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2610 {
2611 	struct mount *mp, *nmp;
2612 	struct vnode *vp;
2613 
2614 	/*
2615 	 * Note: because this is DDB, we can't obey the locking semantics
2616 	 * for these structures, which means we could catch an inconsistent
2617 	 * state and dereference a nasty pointer.  Not much to be done
2618 	 * about that.
2619 	 */
2620 	printf("Locked vnodes\n");
2621 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2622 		nmp = TAILQ_NEXT(mp, mnt_list);
2623 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2624 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL))
2625 				vprint("", vp);
2626 		}
2627 		nmp = TAILQ_NEXT(mp, mnt_list);
2628 	}
2629 }
2630 
2631 /*
2632  * Show details about the given vnode.
2633  */
2634 DB_SHOW_COMMAND(vnode, db_show_vnode)
2635 {
2636 	struct vnode *vp;
2637 
2638 	if (!have_addr)
2639 		return;
2640 	vp = (struct vnode *)addr;
2641 	vn_printf(vp, "vnode ");
2642 }
2643 #endif	/* DDB */
2644 
2645 /*
2646  * Fill in a struct xvfsconf based on a struct vfsconf.
2647  */
2648 static void
2649 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2650 {
2651 
2652 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2653 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2654 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2655 	xvfsp->vfc_flags = vfsp->vfc_flags;
2656 	/*
2657 	 * These are unused in userland, we keep them
2658 	 * to not break binary compatibility.
2659 	 */
2660 	xvfsp->vfc_vfsops = NULL;
2661 	xvfsp->vfc_next = NULL;
2662 }
2663 
2664 /*
2665  * Top level filesystem related information gathering.
2666  */
2667 static int
2668 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2669 {
2670 	struct vfsconf *vfsp;
2671 	struct xvfsconf xvfsp;
2672 	int error;
2673 
2674 	error = 0;
2675 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2676 		bzero(&xvfsp, sizeof(xvfsp));
2677 		vfsconf2x(vfsp, &xvfsp);
2678 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2679 		if (error)
2680 			break;
2681 	}
2682 	return (error);
2683 }
2684 
2685 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2686     "S,xvfsconf", "List of all configured filesystems");
2687 
2688 #ifndef BURN_BRIDGES
2689 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2690 
2691 static int
2692 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2693 {
2694 	int *name = (int *)arg1 - 1;	/* XXX */
2695 	u_int namelen = arg2 + 1;	/* XXX */
2696 	struct vfsconf *vfsp;
2697 	struct xvfsconf xvfsp;
2698 
2699 	printf("WARNING: userland calling deprecated sysctl, "
2700 	    "please rebuild world\n");
2701 
2702 #if 1 || defined(COMPAT_PRELITE2)
2703 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2704 	if (namelen == 1)
2705 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2706 #endif
2707 
2708 	switch (name[1]) {
2709 	case VFS_MAXTYPENUM:
2710 		if (namelen != 2)
2711 			return (ENOTDIR);
2712 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2713 	case VFS_CONF:
2714 		if (namelen != 3)
2715 			return (ENOTDIR);	/* overloaded */
2716 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2717 			if (vfsp->vfc_typenum == name[2])
2718 				break;
2719 		if (vfsp == NULL)
2720 			return (EOPNOTSUPP);
2721 		bzero(&xvfsp, sizeof(xvfsp));
2722 		vfsconf2x(vfsp, &xvfsp);
2723 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2724 	}
2725 	return (EOPNOTSUPP);
2726 }
2727 
2728 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2729 	vfs_sysctl, "Generic filesystem");
2730 
2731 #if 1 || defined(COMPAT_PRELITE2)
2732 
2733 static int
2734 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2735 {
2736 	int error;
2737 	struct vfsconf *vfsp;
2738 	struct ovfsconf ovfs;
2739 
2740 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2741 		bzero(&ovfs, sizeof(ovfs));
2742 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2743 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2744 		ovfs.vfc_index = vfsp->vfc_typenum;
2745 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2746 		ovfs.vfc_flags = vfsp->vfc_flags;
2747 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2748 		if (error)
2749 			return error;
2750 	}
2751 	return 0;
2752 }
2753 
2754 #endif /* 1 || COMPAT_PRELITE2 */
2755 #endif /* !BURN_BRIDGES */
2756 
2757 #define KINFO_VNODESLOP		10
2758 #ifdef notyet
2759 /*
2760  * Dump vnode list (via sysctl).
2761  */
2762 /* ARGSUSED */
2763 static int
2764 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2765 {
2766 	struct xvnode *xvn;
2767 	struct thread *td = req->td;
2768 	struct mount *mp;
2769 	struct vnode *vp;
2770 	int error, len, n;
2771 
2772 	/*
2773 	 * Stale numvnodes access is not fatal here.
2774 	 */
2775 	req->lock = 0;
2776 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2777 	if (!req->oldptr)
2778 		/* Make an estimate */
2779 		return (SYSCTL_OUT(req, 0, len));
2780 
2781 	error = sysctl_wire_old_buffer(req, 0);
2782 	if (error != 0)
2783 		return (error);
2784 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2785 	n = 0;
2786 	mtx_lock(&mountlist_mtx);
2787 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2788 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2789 			continue;
2790 		MNT_ILOCK(mp);
2791 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2792 			if (n == len)
2793 				break;
2794 			vref(vp);
2795 			xvn[n].xv_size = sizeof *xvn;
2796 			xvn[n].xv_vnode = vp;
2797 			xvn[n].xv_id = 0;	/* XXX compat */
2798 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2799 			XV_COPY(usecount);
2800 			XV_COPY(writecount);
2801 			XV_COPY(holdcnt);
2802 			XV_COPY(mount);
2803 			XV_COPY(numoutput);
2804 			XV_COPY(type);
2805 #undef XV_COPY
2806 			xvn[n].xv_flag = vp->v_vflag;
2807 
2808 			switch (vp->v_type) {
2809 			case VREG:
2810 			case VDIR:
2811 			case VLNK:
2812 				break;
2813 			case VBLK:
2814 			case VCHR:
2815 				if (vp->v_rdev == NULL) {
2816 					vrele(vp);
2817 					continue;
2818 				}
2819 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2820 				break;
2821 			case VSOCK:
2822 				xvn[n].xv_socket = vp->v_socket;
2823 				break;
2824 			case VFIFO:
2825 				xvn[n].xv_fifo = vp->v_fifoinfo;
2826 				break;
2827 			case VNON:
2828 			case VBAD:
2829 			default:
2830 				/* shouldn't happen? */
2831 				vrele(vp);
2832 				continue;
2833 			}
2834 			vrele(vp);
2835 			++n;
2836 		}
2837 		MNT_IUNLOCK(mp);
2838 		mtx_lock(&mountlist_mtx);
2839 		vfs_unbusy(mp, td);
2840 		if (n == len)
2841 			break;
2842 	}
2843 	mtx_unlock(&mountlist_mtx);
2844 
2845 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2846 	free(xvn, M_TEMP);
2847 	return (error);
2848 }
2849 
2850 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2851 	0, 0, sysctl_vnode, "S,xvnode", "");
2852 #endif
2853 
2854 /*
2855  * Unmount all filesystems. The list is traversed in reverse order
2856  * of mounting to avoid dependencies.
2857  */
2858 void
2859 vfs_unmountall(void)
2860 {
2861 	struct mount *mp;
2862 	struct thread *td;
2863 	int error;
2864 
2865 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2866 	td = curthread;
2867 	/*
2868 	 * Since this only runs when rebooting, it is not interlocked.
2869 	 */
2870 	while(!TAILQ_EMPTY(&mountlist)) {
2871 		mp = TAILQ_LAST(&mountlist, mntlist);
2872 		error = dounmount(mp, MNT_FORCE, td);
2873 		if (error) {
2874 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2875 			/*
2876 			 * XXX: Due to the way in which we mount the root
2877 			 * file system off of devfs, devfs will generate a
2878 			 * "busy" warning when we try to unmount it before
2879 			 * the root.  Don't print a warning as a result in
2880 			 * order to avoid false positive errors that may
2881 			 * cause needless upset.
2882 			 */
2883 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2884 				printf("unmount of %s failed (",
2885 				    mp->mnt_stat.f_mntonname);
2886 				if (error == EBUSY)
2887 					printf("BUSY)\n");
2888 				else
2889 					printf("%d)\n", error);
2890 			}
2891 		} else {
2892 			/* The unmount has removed mp from the mountlist */
2893 		}
2894 	}
2895 }
2896 
2897 /*
2898  * perform msync on all vnodes under a mount point
2899  * the mount point must be locked.
2900  */
2901 void
2902 vfs_msync(struct mount *mp, int flags)
2903 {
2904 	struct vnode *vp, *mvp;
2905 	struct vm_object *obj;
2906 
2907 	MNT_ILOCK(mp);
2908 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2909 		VI_LOCK(vp);
2910 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2911 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2912 			MNT_IUNLOCK(mp);
2913 			if (!vget(vp,
2914 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2915 			    curthread)) {
2916 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2917 					vput(vp);
2918 					MNT_ILOCK(mp);
2919 					continue;
2920 				}
2921 
2922 				obj = vp->v_object;
2923 				if (obj != NULL) {
2924 					VM_OBJECT_LOCK(obj);
2925 					vm_object_page_clean(obj, 0, 0,
2926 					    flags == MNT_WAIT ?
2927 					    OBJPC_SYNC : OBJPC_NOSYNC);
2928 					VM_OBJECT_UNLOCK(obj);
2929 				}
2930 				vput(vp);
2931 			}
2932 			MNT_ILOCK(mp);
2933 		} else
2934 			VI_UNLOCK(vp);
2935 	}
2936 	MNT_IUNLOCK(mp);
2937 }
2938 
2939 /*
2940  * Mark a vnode as free, putting it up for recycling.
2941  */
2942 static void
2943 vfree(struct vnode *vp)
2944 {
2945 
2946 	CTR1(KTR_VFS, "vfree vp %p", vp);
2947 	ASSERT_VI_LOCKED(vp, "vfree");
2948 	mtx_lock(&vnode_free_list_mtx);
2949 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2950 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2951 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2952 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2953 	    ("vfree: Freeing doomed vnode"));
2954 	if (vp->v_iflag & VI_AGE) {
2955 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2956 	} else {
2957 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2958 	}
2959 	freevnodes++;
2960 	vp->v_iflag &= ~VI_AGE;
2961 	vp->v_iflag |= VI_FREE;
2962 	mtx_unlock(&vnode_free_list_mtx);
2963 }
2964 
2965 /*
2966  * Opposite of vfree() - mark a vnode as in use.
2967  */
2968 static void
2969 vbusy(struct vnode *vp)
2970 {
2971 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2972 	ASSERT_VI_LOCKED(vp, "vbusy");
2973 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2974 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2975 
2976 	mtx_lock(&vnode_free_list_mtx);
2977 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2978 	freevnodes--;
2979 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2980 	mtx_unlock(&vnode_free_list_mtx);
2981 }
2982 
2983 /*
2984  * Initalize per-vnode helper structure to hold poll-related state.
2985  */
2986 void
2987 v_addpollinfo(struct vnode *vp)
2988 {
2989 	struct vpollinfo *vi;
2990 
2991 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2992 	if (vp->v_pollinfo != NULL) {
2993 		uma_zfree(vnodepoll_zone, vi);
2994 		return;
2995 	}
2996 	vp->v_pollinfo = vi;
2997 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2998 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
2999 	    vfs_knlunlock, vfs_knllocked);
3000 }
3001 
3002 /*
3003  * Record a process's interest in events which might happen to
3004  * a vnode.  Because poll uses the historic select-style interface
3005  * internally, this routine serves as both the ``check for any
3006  * pending events'' and the ``record my interest in future events''
3007  * functions.  (These are done together, while the lock is held,
3008  * to avoid race conditions.)
3009  */
3010 int
3011 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3012 {
3013 
3014 	if (vp->v_pollinfo == NULL)
3015 		v_addpollinfo(vp);
3016 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3017 	if (vp->v_pollinfo->vpi_revents & events) {
3018 		/*
3019 		 * This leaves events we are not interested
3020 		 * in available for the other process which
3021 		 * which presumably had requested them
3022 		 * (otherwise they would never have been
3023 		 * recorded).
3024 		 */
3025 		events &= vp->v_pollinfo->vpi_revents;
3026 		vp->v_pollinfo->vpi_revents &= ~events;
3027 
3028 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3029 		return events;
3030 	}
3031 	vp->v_pollinfo->vpi_events |= events;
3032 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3033 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3034 	return 0;
3035 }
3036 
3037 /*
3038  * Routine to create and manage a filesystem syncer vnode.
3039  */
3040 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3041 static int	sync_fsync(struct  vop_fsync_args *);
3042 static int	sync_inactive(struct  vop_inactive_args *);
3043 static int	sync_reclaim(struct  vop_reclaim_args *);
3044 
3045 static struct vop_vector sync_vnodeops = {
3046 	.vop_bypass =	VOP_EOPNOTSUPP,
3047 	.vop_close =	sync_close,		/* close */
3048 	.vop_fsync =	sync_fsync,		/* fsync */
3049 	.vop_inactive =	sync_inactive,	/* inactive */
3050 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3051 	.vop_lock1 =	vop_stdlock,	/* lock */
3052 	.vop_unlock =	vop_stdunlock,	/* unlock */
3053 	.vop_islocked =	vop_stdislocked,	/* islocked */
3054 };
3055 
3056 /*
3057  * Create a new filesystem syncer vnode for the specified mount point.
3058  */
3059 int
3060 vfs_allocate_syncvnode(struct mount *mp)
3061 {
3062 	struct vnode *vp;
3063 	static long start, incr, next;
3064 	int error;
3065 
3066 	/* Allocate a new vnode */
3067 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
3068 		mp->mnt_syncer = NULL;
3069 		return (error);
3070 	}
3071 	vp->v_type = VNON;
3072 	error = insmntque(vp, mp);
3073 	if (error != 0)
3074 		panic("vfs_allocate_syncvnode: insmntque failed");
3075 	/*
3076 	 * Place the vnode onto the syncer worklist. We attempt to
3077 	 * scatter them about on the list so that they will go off
3078 	 * at evenly distributed times even if all the filesystems
3079 	 * are mounted at once.
3080 	 */
3081 	next += incr;
3082 	if (next == 0 || next > syncer_maxdelay) {
3083 		start /= 2;
3084 		incr /= 2;
3085 		if (start == 0) {
3086 			start = syncer_maxdelay / 2;
3087 			incr = syncer_maxdelay;
3088 		}
3089 		next = start;
3090 	}
3091 	VI_LOCK(vp);
3092 	vn_syncer_add_to_worklist(&vp->v_bufobj,
3093 	    syncdelay > 0 ? next % syncdelay : 0);
3094 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3095 	mtx_lock(&sync_mtx);
3096 	sync_vnode_count++;
3097 	mtx_unlock(&sync_mtx);
3098 	VI_UNLOCK(vp);
3099 	mp->mnt_syncer = vp;
3100 	return (0);
3101 }
3102 
3103 /*
3104  * Do a lazy sync of the filesystem.
3105  */
3106 static int
3107 sync_fsync(struct vop_fsync_args *ap)
3108 {
3109 	struct vnode *syncvp = ap->a_vp;
3110 	struct mount *mp = syncvp->v_mount;
3111 	struct thread *td = ap->a_td;
3112 	int error;
3113 	struct bufobj *bo;
3114 
3115 	/*
3116 	 * We only need to do something if this is a lazy evaluation.
3117 	 */
3118 	if (ap->a_waitfor != MNT_LAZY)
3119 		return (0);
3120 
3121 	/*
3122 	 * Move ourselves to the back of the sync list.
3123 	 */
3124 	bo = &syncvp->v_bufobj;
3125 	BO_LOCK(bo);
3126 	vn_syncer_add_to_worklist(bo, syncdelay);
3127 	BO_UNLOCK(bo);
3128 
3129 	/*
3130 	 * Walk the list of vnodes pushing all that are dirty and
3131 	 * not already on the sync list.
3132 	 */
3133 	mtx_lock(&mountlist_mtx);
3134 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3135 		mtx_unlock(&mountlist_mtx);
3136 		return (0);
3137 	}
3138 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3139 		vfs_unbusy(mp, td);
3140 		return (0);
3141 	}
3142 	MNT_ILOCK(mp);
3143 	mp->mnt_noasync++;
3144 	mp->mnt_kern_flag &= ~MNTK_ASYNC;
3145 	MNT_IUNLOCK(mp);
3146 	vfs_msync(mp, MNT_NOWAIT);
3147 	error = VFS_SYNC(mp, MNT_LAZY, td);
3148 	MNT_ILOCK(mp);
3149 	mp->mnt_noasync--;
3150 	if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0)
3151 		mp->mnt_kern_flag |= MNTK_ASYNC;
3152 	MNT_IUNLOCK(mp);
3153 	vn_finished_write(mp);
3154 	vfs_unbusy(mp, td);
3155 	return (error);
3156 }
3157 
3158 /*
3159  * The syncer vnode is no referenced.
3160  */
3161 static int
3162 sync_inactive(struct vop_inactive_args *ap)
3163 {
3164 
3165 	vgone(ap->a_vp);
3166 	return (0);
3167 }
3168 
3169 /*
3170  * The syncer vnode is no longer needed and is being decommissioned.
3171  *
3172  * Modifications to the worklist must be protected by sync_mtx.
3173  */
3174 static int
3175 sync_reclaim(struct vop_reclaim_args *ap)
3176 {
3177 	struct vnode *vp = ap->a_vp;
3178 	struct bufobj *bo;
3179 
3180 	VI_LOCK(vp);
3181 	bo = &vp->v_bufobj;
3182 	vp->v_mount->mnt_syncer = NULL;
3183 	if (bo->bo_flag & BO_ONWORKLST) {
3184 		mtx_lock(&sync_mtx);
3185 		LIST_REMOVE(bo, bo_synclist);
3186 		syncer_worklist_len--;
3187 		sync_vnode_count--;
3188 		mtx_unlock(&sync_mtx);
3189 		bo->bo_flag &= ~BO_ONWORKLST;
3190 	}
3191 	VI_UNLOCK(vp);
3192 
3193 	return (0);
3194 }
3195 
3196 /*
3197  * Check if vnode represents a disk device
3198  */
3199 int
3200 vn_isdisk(struct vnode *vp, int *errp)
3201 {
3202 	int error;
3203 
3204 	error = 0;
3205 	dev_lock();
3206 	if (vp->v_type != VCHR)
3207 		error = ENOTBLK;
3208 	else if (vp->v_rdev == NULL)
3209 		error = ENXIO;
3210 	else if (vp->v_rdev->si_devsw == NULL)
3211 		error = ENXIO;
3212 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3213 		error = ENOTBLK;
3214 	dev_unlock();
3215 	if (errp != NULL)
3216 		*errp = error;
3217 	return (error == 0);
3218 }
3219 
3220 /*
3221  * Common filesystem object access control check routine.  Accepts a
3222  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3223  * and optional call-by-reference privused argument allowing vaccess()
3224  * to indicate to the caller whether privilege was used to satisfy the
3225  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3226  *
3227  * The ifdef'd CAPABILITIES version is here for reference, but is not
3228  * actually used.
3229  */
3230 int
3231 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3232     mode_t acc_mode, struct ucred *cred, int *privused)
3233 {
3234 	mode_t dac_granted;
3235 	mode_t priv_granted;
3236 
3237 	/*
3238 	 * Look for a normal, non-privileged way to access the file/directory
3239 	 * as requested.  If it exists, go with that.
3240 	 */
3241 
3242 	if (privused != NULL)
3243 		*privused = 0;
3244 
3245 	dac_granted = 0;
3246 
3247 	/* Check the owner. */
3248 	if (cred->cr_uid == file_uid) {
3249 		dac_granted |= VADMIN;
3250 		if (file_mode & S_IXUSR)
3251 			dac_granted |= VEXEC;
3252 		if (file_mode & S_IRUSR)
3253 			dac_granted |= VREAD;
3254 		if (file_mode & S_IWUSR)
3255 			dac_granted |= (VWRITE | VAPPEND);
3256 
3257 		if ((acc_mode & dac_granted) == acc_mode)
3258 			return (0);
3259 
3260 		goto privcheck;
3261 	}
3262 
3263 	/* Otherwise, check the groups (first match) */
3264 	if (groupmember(file_gid, cred)) {
3265 		if (file_mode & S_IXGRP)
3266 			dac_granted |= VEXEC;
3267 		if (file_mode & S_IRGRP)
3268 			dac_granted |= VREAD;
3269 		if (file_mode & S_IWGRP)
3270 			dac_granted |= (VWRITE | VAPPEND);
3271 
3272 		if ((acc_mode & dac_granted) == acc_mode)
3273 			return (0);
3274 
3275 		goto privcheck;
3276 	}
3277 
3278 	/* Otherwise, check everyone else. */
3279 	if (file_mode & S_IXOTH)
3280 		dac_granted |= VEXEC;
3281 	if (file_mode & S_IROTH)
3282 		dac_granted |= VREAD;
3283 	if (file_mode & S_IWOTH)
3284 		dac_granted |= (VWRITE | VAPPEND);
3285 	if ((acc_mode & dac_granted) == acc_mode)
3286 		return (0);
3287 
3288 privcheck:
3289 	/*
3290 	 * Build a privilege mask to determine if the set of privileges
3291 	 * satisfies the requirements when combined with the granted mask
3292 	 * from above.  For each privilege, if the privilege is required,
3293 	 * bitwise or the request type onto the priv_granted mask.
3294 	 */
3295 	priv_granted = 0;
3296 
3297 	if (type == VDIR) {
3298 		/*
3299 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
3300 		 * requests, instead of PRIV_VFS_EXEC.
3301 		 */
3302 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3303 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, SUSER_ALLOWJAIL))
3304 			priv_granted |= VEXEC;
3305 	} else {
3306 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3307 		    !priv_check_cred(cred, PRIV_VFS_EXEC, SUSER_ALLOWJAIL))
3308 			priv_granted |= VEXEC;
3309 	}
3310 
3311 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3312 	    !priv_check_cred(cred, PRIV_VFS_READ, SUSER_ALLOWJAIL))
3313 		priv_granted |= VREAD;
3314 
3315 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3316 	    !priv_check_cred(cred, PRIV_VFS_WRITE, SUSER_ALLOWJAIL))
3317 		priv_granted |= (VWRITE | VAPPEND);
3318 
3319 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3320 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, SUSER_ALLOWJAIL))
3321 		priv_granted |= VADMIN;
3322 
3323 	if ((acc_mode & (priv_granted | dac_granted)) == acc_mode) {
3324 		/* XXX audit: privilege used */
3325 		if (privused != NULL)
3326 			*privused = 1;
3327 		return (0);
3328 	}
3329 
3330 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3331 }
3332 
3333 /*
3334  * Credential check based on process requesting service, and per-attribute
3335  * permissions.
3336  */
3337 int
3338 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3339     struct thread *td, int access)
3340 {
3341 
3342 	/*
3343 	 * Kernel-invoked always succeeds.
3344 	 */
3345 	if (cred == NOCRED)
3346 		return (0);
3347 
3348 	/*
3349 	 * Do not allow privileged processes in jail to directly manipulate
3350 	 * system attributes.
3351 	 */
3352 	switch (attrnamespace) {
3353 	case EXTATTR_NAMESPACE_SYSTEM:
3354 		/* Potentially should be: return (EPERM); */
3355 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
3356 	case EXTATTR_NAMESPACE_USER:
3357 		return (VOP_ACCESS(vp, access, cred, td));
3358 	default:
3359 		return (EPERM);
3360 	}
3361 }
3362 
3363 #ifdef DEBUG_VFS_LOCKS
3364 /*
3365  * This only exists to supress warnings from unlocked specfs accesses.  It is
3366  * no longer ok to have an unlocked VFS.
3367  */
3368 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3369 
3370 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3371 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3372 
3373 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3374 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3375 
3376 int vfs_badlock_print = 1;	/* Print lock violations. */
3377 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3378 
3379 #ifdef KDB
3380 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3381 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3382 #endif
3383 
3384 static void
3385 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3386 {
3387 
3388 #ifdef KDB
3389 	if (vfs_badlock_backtrace)
3390 		kdb_backtrace();
3391 #endif
3392 	if (vfs_badlock_print)
3393 		printf("%s: %p %s\n", str, (void *)vp, msg);
3394 	if (vfs_badlock_ddb)
3395 		kdb_enter("lock violation");
3396 }
3397 
3398 void
3399 assert_vi_locked(struct vnode *vp, const char *str)
3400 {
3401 
3402 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3403 		vfs_badlock("interlock is not locked but should be", str, vp);
3404 }
3405 
3406 void
3407 assert_vi_unlocked(struct vnode *vp, const char *str)
3408 {
3409 
3410 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3411 		vfs_badlock("interlock is locked but should not be", str, vp);
3412 }
3413 
3414 void
3415 assert_vop_locked(struct vnode *vp, const char *str)
3416 {
3417 
3418 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3419 		vfs_badlock("is not locked but should be", str, vp);
3420 }
3421 
3422 void
3423 assert_vop_unlocked(struct vnode *vp, const char *str)
3424 {
3425 
3426 	if (vp && !IGNORE_LOCK(vp) &&
3427 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3428 		vfs_badlock("is locked but should not be", str, vp);
3429 }
3430 
3431 void
3432 assert_vop_elocked(struct vnode *vp, const char *str)
3433 {
3434 
3435 	if (vp && !IGNORE_LOCK(vp) &&
3436 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3437 		vfs_badlock("is not exclusive locked but should be", str, vp);
3438 }
3439 
3440 #if 0
3441 void
3442 assert_vop_elocked_other(struct vnode *vp, const char *str)
3443 {
3444 
3445 	if (vp && !IGNORE_LOCK(vp) &&
3446 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3447 		vfs_badlock("is not exclusive locked by another thread",
3448 		    str, vp);
3449 }
3450 
3451 void
3452 assert_vop_slocked(struct vnode *vp, const char *str)
3453 {
3454 
3455 	if (vp && !IGNORE_LOCK(vp) &&
3456 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3457 		vfs_badlock("is not locked shared but should be", str, vp);
3458 }
3459 #endif /* 0 */
3460 #endif /* DEBUG_VFS_LOCKS */
3461 
3462 void
3463 vop_rename_pre(void *ap)
3464 {
3465 	struct vop_rename_args *a = ap;
3466 
3467 #ifdef DEBUG_VFS_LOCKS
3468 	if (a->a_tvp)
3469 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3470 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3471 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3472 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3473 
3474 	/* Check the source (from). */
3475 	if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp)
3476 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3477 	if (a->a_tvp != a->a_fvp)
3478 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
3479 
3480 	/* Check the target. */
3481 	if (a->a_tvp)
3482 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3483 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3484 #endif
3485 	if (a->a_tdvp != a->a_fdvp)
3486 		vhold(a->a_fdvp);
3487 	if (a->a_tvp != a->a_fvp)
3488 		vhold(a->a_fvp);
3489 	vhold(a->a_tdvp);
3490 	if (a->a_tvp)
3491 		vhold(a->a_tvp);
3492 }
3493 
3494 void
3495 vop_strategy_pre(void *ap)
3496 {
3497 #ifdef DEBUG_VFS_LOCKS
3498 	struct vop_strategy_args *a;
3499 	struct buf *bp;
3500 
3501 	a = ap;
3502 	bp = a->a_bp;
3503 
3504 	/*
3505 	 * Cluster ops lock their component buffers but not the IO container.
3506 	 */
3507 	if ((bp->b_flags & B_CLUSTER) != 0)
3508 		return;
3509 
3510 	if (BUF_REFCNT(bp) < 1) {
3511 		if (vfs_badlock_print)
3512 			printf(
3513 			    "VOP_STRATEGY: bp is not locked but should be\n");
3514 		if (vfs_badlock_ddb)
3515 			kdb_enter("lock violation");
3516 	}
3517 #endif
3518 }
3519 
3520 void
3521 vop_lookup_pre(void *ap)
3522 {
3523 #ifdef DEBUG_VFS_LOCKS
3524 	struct vop_lookup_args *a;
3525 	struct vnode *dvp;
3526 
3527 	a = ap;
3528 	dvp = a->a_dvp;
3529 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3530 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3531 #endif
3532 }
3533 
3534 void
3535 vop_lookup_post(void *ap, int rc)
3536 {
3537 #ifdef DEBUG_VFS_LOCKS
3538 	struct vop_lookup_args *a;
3539 	struct vnode *dvp;
3540 	struct vnode *vp;
3541 
3542 	a = ap;
3543 	dvp = a->a_dvp;
3544 	vp = *(a->a_vpp);
3545 
3546 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3547 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3548 
3549 	if (!rc)
3550 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3551 #endif
3552 }
3553 
3554 void
3555 vop_lock_pre(void *ap)
3556 {
3557 #ifdef DEBUG_VFS_LOCKS
3558 	struct vop_lock1_args *a = ap;
3559 
3560 	if ((a->a_flags & LK_INTERLOCK) == 0)
3561 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3562 	else
3563 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3564 #endif
3565 }
3566 
3567 void
3568 vop_lock_post(void *ap, int rc)
3569 {
3570 #ifdef DEBUG_VFS_LOCKS
3571 	struct vop_lock1_args *a = ap;
3572 
3573 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3574 	if (rc == 0)
3575 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3576 #endif
3577 }
3578 
3579 void
3580 vop_unlock_pre(void *ap)
3581 {
3582 #ifdef DEBUG_VFS_LOCKS
3583 	struct vop_unlock_args *a = ap;
3584 
3585 	if (a->a_flags & LK_INTERLOCK)
3586 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3587 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3588 #endif
3589 }
3590 
3591 void
3592 vop_unlock_post(void *ap, int rc)
3593 {
3594 #ifdef DEBUG_VFS_LOCKS
3595 	struct vop_unlock_args *a = ap;
3596 
3597 	if (a->a_flags & LK_INTERLOCK)
3598 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3599 #endif
3600 }
3601 
3602 void
3603 vop_create_post(void *ap, int rc)
3604 {
3605 	struct vop_create_args *a = ap;
3606 
3607 	if (!rc)
3608 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3609 }
3610 
3611 void
3612 vop_link_post(void *ap, int rc)
3613 {
3614 	struct vop_link_args *a = ap;
3615 
3616 	if (!rc) {
3617 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3618 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3619 	}
3620 }
3621 
3622 void
3623 vop_mkdir_post(void *ap, int rc)
3624 {
3625 	struct vop_mkdir_args *a = ap;
3626 
3627 	if (!rc)
3628 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3629 }
3630 
3631 void
3632 vop_mknod_post(void *ap, int rc)
3633 {
3634 	struct vop_mknod_args *a = ap;
3635 
3636 	if (!rc)
3637 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3638 }
3639 
3640 void
3641 vop_remove_post(void *ap, int rc)
3642 {
3643 	struct vop_remove_args *a = ap;
3644 
3645 	if (!rc) {
3646 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3647 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3648 	}
3649 }
3650 
3651 void
3652 vop_rename_post(void *ap, int rc)
3653 {
3654 	struct vop_rename_args *a = ap;
3655 
3656 	if (!rc) {
3657 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3658 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3659 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3660 		if (a->a_tvp)
3661 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3662 	}
3663 	if (a->a_tdvp != a->a_fdvp)
3664 		vdrop(a->a_fdvp);
3665 	if (a->a_tvp != a->a_fvp)
3666 		vdrop(a->a_fvp);
3667 	vdrop(a->a_tdvp);
3668 	if (a->a_tvp)
3669 		vdrop(a->a_tvp);
3670 }
3671 
3672 void
3673 vop_rmdir_post(void *ap, int rc)
3674 {
3675 	struct vop_rmdir_args *a = ap;
3676 
3677 	if (!rc) {
3678 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3679 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3680 	}
3681 }
3682 
3683 void
3684 vop_setattr_post(void *ap, int rc)
3685 {
3686 	struct vop_setattr_args *a = ap;
3687 
3688 	if (!rc)
3689 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3690 }
3691 
3692 void
3693 vop_symlink_post(void *ap, int rc)
3694 {
3695 	struct vop_symlink_args *a = ap;
3696 
3697 	if (!rc)
3698 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3699 }
3700 
3701 static struct knlist fs_knlist;
3702 
3703 static void
3704 vfs_event_init(void *arg)
3705 {
3706 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3707 }
3708 /* XXX - correct order? */
3709 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3710 
3711 void
3712 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3713 {
3714 
3715 	KNOTE_UNLOCKED(&fs_knlist, event);
3716 }
3717 
3718 static int	filt_fsattach(struct knote *kn);
3719 static void	filt_fsdetach(struct knote *kn);
3720 static int	filt_fsevent(struct knote *kn, long hint);
3721 
3722 struct filterops fs_filtops =
3723 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3724 
3725 static int
3726 filt_fsattach(struct knote *kn)
3727 {
3728 
3729 	kn->kn_flags |= EV_CLEAR;
3730 	knlist_add(&fs_knlist, kn, 0);
3731 	return (0);
3732 }
3733 
3734 static void
3735 filt_fsdetach(struct knote *kn)
3736 {
3737 
3738 	knlist_remove(&fs_knlist, kn, 0);
3739 }
3740 
3741 static int
3742 filt_fsevent(struct knote *kn, long hint)
3743 {
3744 
3745 	kn->kn_fflags |= hint;
3746 	return (kn->kn_fflags != 0);
3747 }
3748 
3749 static int
3750 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3751 {
3752 	struct vfsidctl vc;
3753 	int error;
3754 	struct mount *mp;
3755 
3756 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3757 	if (error)
3758 		return (error);
3759 	if (vc.vc_vers != VFS_CTL_VERS1)
3760 		return (EINVAL);
3761 	mp = vfs_getvfs(&vc.vc_fsid);
3762 	if (mp == NULL)
3763 		return (ENOENT);
3764 	/* ensure that a specific sysctl goes to the right filesystem. */
3765 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3766 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3767 		vfs_rel(mp);
3768 		return (EINVAL);
3769 	}
3770 	VCTLTOREQ(&vc, req);
3771 	error = VFS_SYSCTL(mp, vc.vc_op, req);
3772 	vfs_rel(mp);
3773 	return (error);
3774 }
3775 
3776 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "",
3777     "Sysctl by fsid");
3778 
3779 /*
3780  * Function to initialize a va_filerev field sensibly.
3781  * XXX: Wouldn't a random number make a lot more sense ??
3782  */
3783 u_quad_t
3784 init_va_filerev(void)
3785 {
3786 	struct bintime bt;
3787 
3788 	getbinuptime(&bt);
3789 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3790 }
3791 
3792 static int	filt_vfsread(struct knote *kn, long hint);
3793 static int	filt_vfswrite(struct knote *kn, long hint);
3794 static int	filt_vfsvnode(struct knote *kn, long hint);
3795 static void	filt_vfsdetach(struct knote *kn);
3796 static struct filterops vfsread_filtops =
3797 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3798 static struct filterops vfswrite_filtops =
3799 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3800 static struct filterops vfsvnode_filtops =
3801 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3802 
3803 static void
3804 vfs_knllock(void *arg)
3805 {
3806 	struct vnode *vp = arg;
3807 
3808 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
3809 }
3810 
3811 static void
3812 vfs_knlunlock(void *arg)
3813 {
3814 	struct vnode *vp = arg;
3815 
3816 	VOP_UNLOCK(vp, 0, curthread);
3817 }
3818 
3819 static int
3820 vfs_knllocked(void *arg)
3821 {
3822 	struct vnode *vp = arg;
3823 
3824 	return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE);
3825 }
3826 
3827 int
3828 vfs_kqfilter(struct vop_kqfilter_args *ap)
3829 {
3830 	struct vnode *vp = ap->a_vp;
3831 	struct knote *kn = ap->a_kn;
3832 	struct knlist *knl;
3833 
3834 	switch (kn->kn_filter) {
3835 	case EVFILT_READ:
3836 		kn->kn_fop = &vfsread_filtops;
3837 		break;
3838 	case EVFILT_WRITE:
3839 		kn->kn_fop = &vfswrite_filtops;
3840 		break;
3841 	case EVFILT_VNODE:
3842 		kn->kn_fop = &vfsvnode_filtops;
3843 		break;
3844 	default:
3845 		return (EINVAL);
3846 	}
3847 
3848 	kn->kn_hook = (caddr_t)vp;
3849 
3850 	if (vp->v_pollinfo == NULL)
3851 		v_addpollinfo(vp);
3852 	if (vp->v_pollinfo == NULL)
3853 		return (ENOMEM);
3854 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3855 	knlist_add(knl, kn, 0);
3856 
3857 	return (0);
3858 }
3859 
3860 /*
3861  * Detach knote from vnode
3862  */
3863 static void
3864 filt_vfsdetach(struct knote *kn)
3865 {
3866 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3867 
3868 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3869 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3870 }
3871 
3872 /*ARGSUSED*/
3873 static int
3874 filt_vfsread(struct knote *kn, long hint)
3875 {
3876 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3877 	struct vattr va;
3878 
3879 	/*
3880 	 * filesystem is gone, so set the EOF flag and schedule
3881 	 * the knote for deletion.
3882 	 */
3883 	if (hint == NOTE_REVOKE) {
3884 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3885 		return (1);
3886 	}
3887 
3888 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3889 		return (0);
3890 
3891 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3892 	return (kn->kn_data != 0);
3893 }
3894 
3895 /*ARGSUSED*/
3896 static int
3897 filt_vfswrite(struct knote *kn, long hint)
3898 {
3899 	/*
3900 	 * filesystem is gone, so set the EOF flag and schedule
3901 	 * the knote for deletion.
3902 	 */
3903 	if (hint == NOTE_REVOKE)
3904 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3905 
3906 	kn->kn_data = 0;
3907 	return (1);
3908 }
3909 
3910 static int
3911 filt_vfsvnode(struct knote *kn, long hint)
3912 {
3913 	if (kn->kn_sfflags & hint)
3914 		kn->kn_fflags |= hint;
3915 	if (hint == NOTE_REVOKE) {
3916 		kn->kn_flags |= EV_EOF;
3917 		return (1);
3918 	}
3919 	return (kn->kn_fflags != 0);
3920 }
3921 
3922 int
3923 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3924 {
3925 	int error;
3926 
3927 	if (dp->d_reclen > ap->a_uio->uio_resid)
3928 		return (ENAMETOOLONG);
3929 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3930 	if (error) {
3931 		if (ap->a_ncookies != NULL) {
3932 			if (ap->a_cookies != NULL)
3933 				free(ap->a_cookies, M_TEMP);
3934 			ap->a_cookies = NULL;
3935 			*ap->a_ncookies = 0;
3936 		}
3937 		return (error);
3938 	}
3939 	if (ap->a_ncookies == NULL)
3940 		return (0);
3941 
3942 	KASSERT(ap->a_cookies,
3943 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
3944 
3945 	*ap->a_cookies = realloc(*ap->a_cookies,
3946 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
3947 	(*ap->a_cookies)[*ap->a_ncookies] = off;
3948 	return (0);
3949 }
3950 
3951 /*
3952  * Mark for update the access time of the file if the filesystem
3953  * supports VA_MARK_ATIME.  This functionality is used by execve
3954  * and mmap, so we want to avoid the synchronous I/O implied by
3955  * directly setting va_atime for the sake of efficiency.
3956  */
3957 void
3958 vfs_mark_atime(struct vnode *vp, struct thread *td)
3959 {
3960 	struct vattr atimeattr;
3961 
3962 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
3963 		VATTR_NULL(&atimeattr);
3964 		atimeattr.va_vaflags |= VA_MARK_ATIME;
3965 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
3966 	}
3967 }
3968