xref: /freebsd/sys/kern/vfs_subr.c (revision 050570efa79efcc9cf5adeb545f1a679c8dc377b)
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 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/bio.h>
49 #include <sys/buf.h>
50 #include <sys/condvar.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/lockf.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/namei.h>
66 #include <sys/priv.h>
67 #include <sys/reboot.h>
68 #include <sys/sched.h>
69 #include <sys/sleepqueue.h>
70 #include <sys/stat.h>
71 #include <sys/sysctl.h>
72 #include <sys/syslog.h>
73 #include <sys/vmmeter.h>
74 #include <sys/vnode.h>
75 
76 #include <machine/stdarg.h>
77 
78 #include <security/mac/mac_framework.h>
79 
80 #include <vm/vm.h>
81 #include <vm/vm_object.h>
82 #include <vm/vm_extern.h>
83 #include <vm/pmap.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_kern.h>
87 #include <vm/uma.h>
88 
89 #ifdef DDB
90 #include <ddb/ddb.h>
91 #endif
92 
93 #define	WI_MPSAFEQ	0
94 #define	WI_GIANTQ	1
95 
96 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure");
97 
98 static void	delmntque(struct vnode *vp);
99 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
100 		    int slpflag, int slptimeo);
101 static void	syncer_shutdown(void *arg, int howto);
102 static int	vtryrecycle(struct vnode *vp);
103 static void	vbusy(struct vnode *vp);
104 static void	vinactive(struct vnode *, struct thread *);
105 static void	v_incr_usecount(struct vnode *);
106 static void	v_decr_usecount(struct vnode *);
107 static void	v_decr_useonly(struct vnode *);
108 static void	v_upgrade_usecount(struct vnode *);
109 static void	vfree(struct vnode *);
110 static void	vnlru_free(int);
111 static void	vgonel(struct vnode *);
112 static void	vfs_knllock(void *arg);
113 static void	vfs_knlunlock(void *arg);
114 static void	vfs_knl_assert_locked(void *arg);
115 static void	vfs_knl_assert_unlocked(void *arg);
116 static void	destroy_vpollinfo(struct vpollinfo *vi);
117 
118 /*
119  * Number of vnodes in existence.  Increased whenever getnewvnode()
120  * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed
121  * vnode.
122  */
123 static unsigned long	numvnodes;
124 
125 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
126     "Number of vnodes in existence");
127 
128 /*
129  * Conversion tables for conversion from vnode types to inode formats
130  * and back.
131  */
132 enum vtype iftovt_tab[16] = {
133 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
134 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
135 };
136 int vttoif_tab[10] = {
137 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
138 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
139 };
140 
141 /*
142  * List of vnodes that are ready for recycling.
143  */
144 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
145 
146 /*
147  * Free vnode target.  Free vnodes may simply be files which have been stat'd
148  * but not read.  This is somewhat common, and a small cache of such files
149  * should be kept to avoid recreation costs.
150  */
151 static u_long wantfreevnodes;
152 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
153 /* Number of vnodes in the free list. */
154 static u_long freevnodes;
155 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0,
156     "Number of vnodes in the free list");
157 
158 static int vlru_allow_cache_src;
159 SYSCTL_INT(_vfs, OID_AUTO, vlru_allow_cache_src, CTLFLAG_RW,
160     &vlru_allow_cache_src, 0, "Allow vlru to reclaim source vnode");
161 
162 /*
163  * Various variables used for debugging the new implementation of
164  * reassignbuf().
165  * XXX these are probably of (very) limited utility now.
166  */
167 static int reassignbufcalls;
168 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
169     "Number of calls to reassignbuf");
170 
171 /*
172  * Cache for the mount type id assigned to NFS.  This is used for
173  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
174  */
175 int	nfs_mount_type = -1;
176 
177 /* To keep more than one thread at a time from running vfs_getnewfsid */
178 static struct mtx mntid_mtx;
179 
180 /*
181  * Lock for any access to the following:
182  *	vnode_free_list
183  *	numvnodes
184  *	freevnodes
185  */
186 static struct mtx vnode_free_list_mtx;
187 
188 /* Publicly exported FS */
189 struct nfs_public nfs_pub;
190 
191 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
192 static uma_zone_t vnode_zone;
193 static uma_zone_t vnodepoll_zone;
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[2];
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 static struct cv sync_wakeup;
236 
237 #define SYNCER_MAXDELAY		32
238 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
239 static int syncdelay = 30;		/* max time to delay syncing data */
240 static int filedelay = 30;		/* time to delay syncing files */
241 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
242     "Time to delay syncing files (in seconds)");
243 static int dirdelay = 29;		/* time to delay syncing directories */
244 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
245     "Time to delay syncing directories (in seconds)");
246 static int metadelay = 28;		/* time to delay syncing metadata */
247 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
248     "Time to delay syncing metadata (in seconds)");
249 static int rushjob;		/* number of slots to run ASAP */
250 static int stat_rush_requests;	/* number of times I/O speeded up */
251 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
252     "Number of times I/O speeded up (rush requests)");
253 
254 /*
255  * When shutting down the syncer, run it at four times normal speed.
256  */
257 #define SYNCER_SHUTDOWN_SPEEDUP		4
258 static int sync_vnode_count;
259 static int syncer_worklist_len;
260 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
261     syncer_state;
262 
263 /*
264  * Number of vnodes we want to exist at any one time.  This is mostly used
265  * to size hash tables in vnode-related code.  It is normally not used in
266  * getnewvnode(), as wantfreevnodes is normally nonzero.)
267  *
268  * XXX desiredvnodes is historical cruft and should not exist.
269  */
270 int desiredvnodes;
271 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
272     &desiredvnodes, 0, "Maximum number of vnodes");
273 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
274     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
275 static int vnlru_nowhere;
276 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
277     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
278 
279 /*
280  * Macros to control when a vnode is freed and recycled.  All require
281  * the vnode interlock.
282  */
283 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
284 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
285 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
286 
287 
288 /*
289  * Initialize the vnode management data structures.
290  *
291  * Reevaluate the following cap on the number of vnodes after the physical
292  * memory size exceeds 512GB.  In the limit, as the physical memory size
293  * grows, the ratio of physical pages to vnodes approaches sixteen to one.
294  */
295 #ifndef	MAXVNODES_MAX
296 #define	MAXVNODES_MAX	(512 * (1024 * 1024 * 1024 / (int)PAGE_SIZE / 16))
297 #endif
298 static void
299 vntblinit(void *dummy __unused)
300 {
301 	int physvnodes, virtvnodes;
302 
303 	/*
304 	 * Desiredvnodes is a function of the physical memory size and the
305 	 * kernel's heap size.  Generally speaking, it scales with the
306 	 * physical memory size.  The ratio of desiredvnodes to physical pages
307 	 * is one to four until desiredvnodes exceeds 98,304.  Thereafter, the
308 	 * marginal ratio of desiredvnodes to physical pages is one to
309 	 * sixteen.  However, desiredvnodes is limited by the kernel's heap
310 	 * size.  The memory required by desiredvnodes vnodes and vm objects
311 	 * may not exceed one seventh of the kernel's heap size.
312 	 */
313 	physvnodes = maxproc + cnt.v_page_count / 16 + 3 * min(98304 * 4,
314 	    cnt.v_page_count) / 16;
315 	virtvnodes = vm_kmem_size / (7 * (sizeof(struct vm_object) +
316 	    sizeof(struct vnode)));
317 	desiredvnodes = min(physvnodes, virtvnodes);
318 	if (desiredvnodes > MAXVNODES_MAX) {
319 		if (bootverbose)
320 			printf("Reducing kern.maxvnodes %d -> %d\n",
321 			    desiredvnodes, MAXVNODES_MAX);
322 		desiredvnodes = MAXVNODES_MAX;
323 	}
324 	wantfreevnodes = desiredvnodes / 4;
325 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
326 	TAILQ_INIT(&vnode_free_list);
327 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
328 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
329 	    NULL, NULL, UMA_ALIGN_PTR, 0);
330 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
331 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
332 	/*
333 	 * Initialize the filesystem syncer.
334 	 */
335 	syncer_workitem_pending[WI_MPSAFEQ] = hashinit(syncer_maxdelay, M_VNODE,
336 	    &syncer_mask);
337 	syncer_workitem_pending[WI_GIANTQ] = hashinit(syncer_maxdelay, M_VNODE,
338 	    &syncer_mask);
339 	syncer_maxdelay = syncer_mask + 1;
340 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
341 	cv_init(&sync_wakeup, "syncer");
342 }
343 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
344 
345 
346 /*
347  * Mark a mount point as busy. Used to synchronize access and to delay
348  * unmounting. Eventually, mountlist_mtx is not released on failure.
349  */
350 int
351 vfs_busy(struct mount *mp, int flags)
352 {
353 
354 	MPASS((flags & ~MBF_MASK) == 0);
355 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
356 
357 	MNT_ILOCK(mp);
358 	MNT_REF(mp);
359 	/*
360 	 * If mount point is currenly being unmounted, sleep until the
361 	 * mount point fate is decided.  If thread doing the unmounting fails,
362 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
363 	 * that this mount point has survived the unmount attempt and vfs_busy
364 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
365 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
366 	 * about to be really destroyed.  vfs_busy needs to release its
367 	 * reference on the mount point in this case and return with ENOENT,
368 	 * telling the caller that mount mount it tried to busy is no longer
369 	 * valid.
370 	 */
371 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
372 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
373 			MNT_REL(mp);
374 			MNT_IUNLOCK(mp);
375 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
376 			    __func__);
377 			return (ENOENT);
378 		}
379 		if (flags & MBF_MNTLSTLOCK)
380 			mtx_unlock(&mountlist_mtx);
381 		mp->mnt_kern_flag |= MNTK_MWAIT;
382 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
383 		if (flags & MBF_MNTLSTLOCK)
384 			mtx_lock(&mountlist_mtx);
385 	}
386 	if (flags & MBF_MNTLSTLOCK)
387 		mtx_unlock(&mountlist_mtx);
388 	mp->mnt_lockref++;
389 	MNT_IUNLOCK(mp);
390 	return (0);
391 }
392 
393 /*
394  * Free a busy filesystem.
395  */
396 void
397 vfs_unbusy(struct mount *mp)
398 {
399 
400 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
401 	MNT_ILOCK(mp);
402 	MNT_REL(mp);
403 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
404 	mp->mnt_lockref--;
405 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
406 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
407 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
408 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
409 		wakeup(&mp->mnt_lockref);
410 	}
411 	MNT_IUNLOCK(mp);
412 }
413 
414 /*
415  * Lookup a mount point by filesystem identifier.
416  */
417 struct mount *
418 vfs_getvfs(fsid_t *fsid)
419 {
420 	struct mount *mp;
421 
422 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
423 	mtx_lock(&mountlist_mtx);
424 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
425 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
426 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
427 			vfs_ref(mp);
428 			mtx_unlock(&mountlist_mtx);
429 			return (mp);
430 		}
431 	}
432 	mtx_unlock(&mountlist_mtx);
433 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
434 	return ((struct mount *) 0);
435 }
436 
437 /*
438  * Lookup a mount point by filesystem identifier, busying it before
439  * returning.
440  */
441 struct mount *
442 vfs_busyfs(fsid_t *fsid)
443 {
444 	struct mount *mp;
445 	int error;
446 
447 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
448 	mtx_lock(&mountlist_mtx);
449 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
450 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
451 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
452 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
453 			if (error) {
454 				mtx_unlock(&mountlist_mtx);
455 				return (NULL);
456 			}
457 			return (mp);
458 		}
459 	}
460 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
461 	mtx_unlock(&mountlist_mtx);
462 	return ((struct mount *) 0);
463 }
464 
465 /*
466  * Check if a user can access privileged mount options.
467  */
468 int
469 vfs_suser(struct mount *mp, struct thread *td)
470 {
471 	int error;
472 
473 	/*
474 	 * If the thread is jailed, but this is not a jail-friendly file
475 	 * system, deny immediately.
476 	 */
477 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
478 		return (EPERM);
479 
480 	/*
481 	 * If the file system was mounted outside the jail of the calling
482 	 * thread, deny immediately.
483 	 */
484 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
485 		return (EPERM);
486 
487 	/*
488 	 * If file system supports delegated administration, we don't check
489 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
490 	 * by the file system itself.
491 	 * If this is not the user that did original mount, we check for
492 	 * the PRIV_VFS_MOUNT_OWNER privilege.
493 	 */
494 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
495 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
496 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
497 			return (error);
498 	}
499 	return (0);
500 }
501 
502 /*
503  * Get a new unique fsid.  Try to make its val[0] unique, since this value
504  * will be used to create fake device numbers for stat().  Also try (but
505  * not so hard) make its val[0] unique mod 2^16, since some emulators only
506  * support 16-bit device numbers.  We end up with unique val[0]'s for the
507  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
508  *
509  * Keep in mind that several mounts may be running in parallel.  Starting
510  * the search one past where the previous search terminated is both a
511  * micro-optimization and a defense against returning the same fsid to
512  * different mounts.
513  */
514 void
515 vfs_getnewfsid(struct mount *mp)
516 {
517 	static uint16_t mntid_base;
518 	struct mount *nmp;
519 	fsid_t tfsid;
520 	int mtype;
521 
522 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
523 	mtx_lock(&mntid_mtx);
524 	mtype = mp->mnt_vfc->vfc_typenum;
525 	tfsid.val[1] = mtype;
526 	mtype = (mtype & 0xFF) << 24;
527 	for (;;) {
528 		tfsid.val[0] = makedev(255,
529 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
530 		mntid_base++;
531 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
532 			break;
533 		vfs_rel(nmp);
534 	}
535 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
536 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
537 	mtx_unlock(&mntid_mtx);
538 }
539 
540 /*
541  * Knob to control the precision of file timestamps:
542  *
543  *   0 = seconds only; nanoseconds zeroed.
544  *   1 = seconds and nanoseconds, accurate within 1/HZ.
545  *   2 = seconds and nanoseconds, truncated to microseconds.
546  * >=3 = seconds and nanoseconds, maximum precision.
547  */
548 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
549 
550 static int timestamp_precision = TSP_SEC;
551 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
552     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
553     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, "
554     "3+: sec + ns (max. precision))");
555 
556 /*
557  * Get a current timestamp.
558  */
559 void
560 vfs_timestamp(struct timespec *tsp)
561 {
562 	struct timeval tv;
563 
564 	switch (timestamp_precision) {
565 	case TSP_SEC:
566 		tsp->tv_sec = time_second;
567 		tsp->tv_nsec = 0;
568 		break;
569 	case TSP_HZ:
570 		getnanotime(tsp);
571 		break;
572 	case TSP_USEC:
573 		microtime(&tv);
574 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
575 		break;
576 	case TSP_NSEC:
577 	default:
578 		nanotime(tsp);
579 		break;
580 	}
581 }
582 
583 /*
584  * Set vnode attributes to VNOVAL
585  */
586 void
587 vattr_null(struct vattr *vap)
588 {
589 
590 	vap->va_type = VNON;
591 	vap->va_size = VNOVAL;
592 	vap->va_bytes = VNOVAL;
593 	vap->va_mode = VNOVAL;
594 	vap->va_nlink = VNOVAL;
595 	vap->va_uid = VNOVAL;
596 	vap->va_gid = VNOVAL;
597 	vap->va_fsid = VNOVAL;
598 	vap->va_fileid = VNOVAL;
599 	vap->va_blocksize = VNOVAL;
600 	vap->va_rdev = VNOVAL;
601 	vap->va_atime.tv_sec = VNOVAL;
602 	vap->va_atime.tv_nsec = VNOVAL;
603 	vap->va_mtime.tv_sec = VNOVAL;
604 	vap->va_mtime.tv_nsec = VNOVAL;
605 	vap->va_ctime.tv_sec = VNOVAL;
606 	vap->va_ctime.tv_nsec = VNOVAL;
607 	vap->va_birthtime.tv_sec = VNOVAL;
608 	vap->va_birthtime.tv_nsec = VNOVAL;
609 	vap->va_flags = VNOVAL;
610 	vap->va_gen = VNOVAL;
611 	vap->va_vaflags = 0;
612 }
613 
614 /*
615  * This routine is called when we have too many vnodes.  It attempts
616  * to free <count> vnodes and will potentially free vnodes that still
617  * have VM backing store (VM backing store is typically the cause
618  * of a vnode blowout so we want to do this).  Therefore, this operation
619  * is not considered cheap.
620  *
621  * A number of conditions may prevent a vnode from being reclaimed.
622  * the buffer cache may have references on the vnode, a directory
623  * vnode may still have references due to the namei cache representing
624  * underlying files, or the vnode may be in active use.   It is not
625  * desireable to reuse such vnodes.  These conditions may cause the
626  * number of vnodes to reach some minimum value regardless of what
627  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
628  */
629 static int
630 vlrureclaim(struct mount *mp)
631 {
632 	struct vnode *vp;
633 	int done;
634 	int trigger;
635 	int usevnodes;
636 	int count;
637 
638 	/*
639 	 * Calculate the trigger point, don't allow user
640 	 * screwups to blow us up.   This prevents us from
641 	 * recycling vnodes with lots of resident pages.  We
642 	 * aren't trying to free memory, we are trying to
643 	 * free vnodes.
644 	 */
645 	usevnodes = desiredvnodes;
646 	if (usevnodes <= 0)
647 		usevnodes = 1;
648 	trigger = cnt.v_page_count * 2 / usevnodes;
649 	done = 0;
650 	vn_start_write(NULL, &mp, V_WAIT);
651 	MNT_ILOCK(mp);
652 	count = mp->mnt_nvnodelistsize / 10 + 1;
653 	while (count != 0) {
654 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
655 		while (vp != NULL && vp->v_type == VMARKER)
656 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
657 		if (vp == NULL)
658 			break;
659 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
660 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
661 		--count;
662 		if (!VI_TRYLOCK(vp))
663 			goto next_iter;
664 		/*
665 		 * If it's been deconstructed already, it's still
666 		 * referenced, or it exceeds the trigger, skip it.
667 		 */
668 		if (vp->v_usecount ||
669 		    (!vlru_allow_cache_src &&
670 			!LIST_EMPTY(&(vp)->v_cache_src)) ||
671 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
672 		    vp->v_object->resident_page_count > trigger)) {
673 			VI_UNLOCK(vp);
674 			goto next_iter;
675 		}
676 		MNT_IUNLOCK(mp);
677 		vholdl(vp);
678 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
679 			vdrop(vp);
680 			goto next_iter_mntunlocked;
681 		}
682 		VI_LOCK(vp);
683 		/*
684 		 * v_usecount may have been bumped after VOP_LOCK() dropped
685 		 * the vnode interlock and before it was locked again.
686 		 *
687 		 * It is not necessary to recheck VI_DOOMED because it can
688 		 * only be set by another thread that holds both the vnode
689 		 * lock and vnode interlock.  If another thread has the
690 		 * vnode lock before we get to VOP_LOCK() and obtains the
691 		 * vnode interlock after VOP_LOCK() drops the vnode
692 		 * interlock, the other thread will be unable to drop the
693 		 * vnode lock before our VOP_LOCK() call fails.
694 		 */
695 		if (vp->v_usecount ||
696 		    (!vlru_allow_cache_src &&
697 			!LIST_EMPTY(&(vp)->v_cache_src)) ||
698 		    (vp->v_object != NULL &&
699 		    vp->v_object->resident_page_count > trigger)) {
700 			VOP_UNLOCK(vp, LK_INTERLOCK);
701 			goto next_iter_mntunlocked;
702 		}
703 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
704 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
705 		vgonel(vp);
706 		VOP_UNLOCK(vp, 0);
707 		vdropl(vp);
708 		done++;
709 next_iter_mntunlocked:
710 		if ((count % 256) != 0)
711 			goto relock_mnt;
712 		goto yield;
713 next_iter:
714 		if ((count % 256) != 0)
715 			continue;
716 		MNT_IUNLOCK(mp);
717 yield:
718 		uio_yield();
719 relock_mnt:
720 		MNT_ILOCK(mp);
721 	}
722 	MNT_IUNLOCK(mp);
723 	vn_finished_write(mp);
724 	return done;
725 }
726 
727 /*
728  * Attempt to keep the free list at wantfreevnodes length.
729  */
730 static void
731 vnlru_free(int count)
732 {
733 	struct vnode *vp;
734 	int vfslocked;
735 
736 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
737 	for (; count > 0; count--) {
738 		vp = TAILQ_FIRST(&vnode_free_list);
739 		/*
740 		 * The list can be modified while the free_list_mtx
741 		 * has been dropped and vp could be NULL here.
742 		 */
743 		if (!vp)
744 			break;
745 		VNASSERT(vp->v_op != NULL, vp,
746 		    ("vnlru_free: vnode already reclaimed."));
747 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
748 		/*
749 		 * Don't recycle if we can't get the interlock.
750 		 */
751 		if (!VI_TRYLOCK(vp)) {
752 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
753 			continue;
754 		}
755 		VNASSERT(VCANRECYCLE(vp), vp,
756 		    ("vp inconsistent on freelist"));
757 		freevnodes--;
758 		vp->v_iflag &= ~VI_FREE;
759 		vholdl(vp);
760 		mtx_unlock(&vnode_free_list_mtx);
761 		VI_UNLOCK(vp);
762 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
763 		vtryrecycle(vp);
764 		VFS_UNLOCK_GIANT(vfslocked);
765 		/*
766 		 * If the recycled succeeded this vdrop will actually free
767 		 * the vnode.  If not it will simply place it back on
768 		 * the free list.
769 		 */
770 		vdrop(vp);
771 		mtx_lock(&vnode_free_list_mtx);
772 	}
773 }
774 /*
775  * Attempt to recycle vnodes in a context that is always safe to block.
776  * Calling vlrurecycle() from the bowels of filesystem code has some
777  * interesting deadlock problems.
778  */
779 static struct proc *vnlruproc;
780 static int vnlruproc_sig;
781 
782 static void
783 vnlru_proc(void)
784 {
785 	struct mount *mp, *nmp;
786 	int done, vfslocked;
787 	struct proc *p = vnlruproc;
788 
789 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
790 	    SHUTDOWN_PRI_FIRST);
791 
792 	for (;;) {
793 		kproc_suspend_check(p);
794 		mtx_lock(&vnode_free_list_mtx);
795 		if (freevnodes > wantfreevnodes)
796 			vnlru_free(freevnodes - wantfreevnodes);
797 		if (numvnodes <= desiredvnodes * 9 / 10) {
798 			vnlruproc_sig = 0;
799 			wakeup(&vnlruproc_sig);
800 			msleep(vnlruproc, &vnode_free_list_mtx,
801 			    PVFS|PDROP, "vlruwt", hz);
802 			continue;
803 		}
804 		mtx_unlock(&vnode_free_list_mtx);
805 		done = 0;
806 		mtx_lock(&mountlist_mtx);
807 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
808 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
809 				nmp = TAILQ_NEXT(mp, mnt_list);
810 				continue;
811 			}
812 			vfslocked = VFS_LOCK_GIANT(mp);
813 			done += vlrureclaim(mp);
814 			VFS_UNLOCK_GIANT(vfslocked);
815 			mtx_lock(&mountlist_mtx);
816 			nmp = TAILQ_NEXT(mp, mnt_list);
817 			vfs_unbusy(mp);
818 		}
819 		mtx_unlock(&mountlist_mtx);
820 		if (done == 0) {
821 #if 0
822 			/* These messages are temporary debugging aids */
823 			if (vnlru_nowhere < 5)
824 				printf("vnlru process getting nowhere..\n");
825 			else if (vnlru_nowhere == 5)
826 				printf("vnlru process messages stopped.\n");
827 #endif
828 			vnlru_nowhere++;
829 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
830 		} else
831 			uio_yield();
832 	}
833 }
834 
835 static struct kproc_desc vnlru_kp = {
836 	"vnlru",
837 	vnlru_proc,
838 	&vnlruproc
839 };
840 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
841     &vnlru_kp);
842 
843 /*
844  * Routines having to do with the management of the vnode table.
845  */
846 
847 void
848 vdestroy(struct vnode *vp)
849 {
850 	struct bufobj *bo;
851 
852 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
853 	mtx_lock(&vnode_free_list_mtx);
854 	numvnodes--;
855 	mtx_unlock(&vnode_free_list_mtx);
856 	bo = &vp->v_bufobj;
857 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
858 	    ("cleaned vnode still on the free list."));
859 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
860 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
861 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
862 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
863 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
864 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
865 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
866 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
867 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
868 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
869 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
870 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
871 	VI_UNLOCK(vp);
872 #ifdef MAC
873 	mac_vnode_destroy(vp);
874 #endif
875 	if (vp->v_pollinfo != NULL)
876 		destroy_vpollinfo(vp->v_pollinfo);
877 #ifdef INVARIANTS
878 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
879 	vp->v_op = NULL;
880 #endif
881 	lockdestroy(vp->v_vnlock);
882 	mtx_destroy(&vp->v_interlock);
883 	mtx_destroy(BO_MTX(bo));
884 	uma_zfree(vnode_zone, vp);
885 }
886 
887 /*
888  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
889  * before we actually vgone().  This function must be called with the vnode
890  * held to prevent the vnode from being returned to the free list midway
891  * through vgone().
892  */
893 static int
894 vtryrecycle(struct vnode *vp)
895 {
896 	struct mount *vnmp;
897 
898 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
899 	VNASSERT(vp->v_holdcnt, vp,
900 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
901 	/*
902 	 * This vnode may found and locked via some other list, if so we
903 	 * can't recycle it yet.
904 	 */
905 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
906 		CTR2(KTR_VFS,
907 		    "%s: impossible to recycle, vp %p lock is already held",
908 		    __func__, vp);
909 		return (EWOULDBLOCK);
910 	}
911 	/*
912 	 * Don't recycle if its filesystem is being suspended.
913 	 */
914 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
915 		VOP_UNLOCK(vp, 0);
916 		CTR2(KTR_VFS,
917 		    "%s: impossible to recycle, cannot start the write for %p",
918 		    __func__, vp);
919 		return (EBUSY);
920 	}
921 	/*
922 	 * If we got this far, we need to acquire the interlock and see if
923 	 * anyone picked up this vnode from another list.  If not, we will
924 	 * mark it with DOOMED via vgonel() so that anyone who does find it
925 	 * will skip over it.
926 	 */
927 	VI_LOCK(vp);
928 	if (vp->v_usecount) {
929 		VOP_UNLOCK(vp, LK_INTERLOCK);
930 		vn_finished_write(vnmp);
931 		CTR2(KTR_VFS,
932 		    "%s: impossible to recycle, %p is already referenced",
933 		    __func__, vp);
934 		return (EBUSY);
935 	}
936 	if ((vp->v_iflag & VI_DOOMED) == 0)
937 		vgonel(vp);
938 	VOP_UNLOCK(vp, LK_INTERLOCK);
939 	vn_finished_write(vnmp);
940 	return (0);
941 }
942 
943 /*
944  * Return the next vnode from the free list.
945  */
946 int
947 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
948     struct vnode **vpp)
949 {
950 	struct vnode *vp = NULL;
951 	struct bufobj *bo;
952 
953 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
954 	mtx_lock(&vnode_free_list_mtx);
955 	/*
956 	 * Lend our context to reclaim vnodes if they've exceeded the max.
957 	 */
958 	if (freevnodes > wantfreevnodes)
959 		vnlru_free(1);
960 	/*
961 	 * Wait for available vnodes.
962 	 */
963 	if (numvnodes > desiredvnodes) {
964 		if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) {
965 			/*
966 			 * File system is beeing suspended, we cannot risk a
967 			 * deadlock here, so allocate new vnode anyway.
968 			 */
969 			if (freevnodes > wantfreevnodes)
970 				vnlru_free(freevnodes - wantfreevnodes);
971 			goto alloc;
972 		}
973 		if (vnlruproc_sig == 0) {
974 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
975 			wakeup(vnlruproc);
976 		}
977 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
978 		    "vlruwk", hz);
979 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
980 		if (numvnodes > desiredvnodes) {
981 			mtx_unlock(&vnode_free_list_mtx);
982 			return (ENFILE);
983 		}
984 #endif
985 	}
986 alloc:
987 	numvnodes++;
988 	mtx_unlock(&vnode_free_list_mtx);
989 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
990 	/*
991 	 * Setup locks.
992 	 */
993 	vp->v_vnlock = &vp->v_lock;
994 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
995 	/*
996 	 * By default, don't allow shared locks unless filesystems
997 	 * opt-in.
998 	 */
999 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
1000 	/*
1001 	 * Initialize bufobj.
1002 	 */
1003 	bo = &vp->v_bufobj;
1004 	bo->__bo_vnode = vp;
1005 	mtx_init(BO_MTX(bo), "bufobj interlock", NULL, MTX_DEF);
1006 	bo->bo_ops = &buf_ops_bio;
1007 	bo->bo_private = vp;
1008 	TAILQ_INIT(&bo->bo_clean.bv_hd);
1009 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
1010 	/*
1011 	 * Initialize namecache.
1012 	 */
1013 	LIST_INIT(&vp->v_cache_src);
1014 	TAILQ_INIT(&vp->v_cache_dst);
1015 	/*
1016 	 * Finalize various vnode identity bits.
1017 	 */
1018 	vp->v_type = VNON;
1019 	vp->v_tag = tag;
1020 	vp->v_op = vops;
1021 	v_incr_usecount(vp);
1022 	vp->v_data = 0;
1023 #ifdef MAC
1024 	mac_vnode_init(vp);
1025 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1026 		mac_vnode_associate_singlelabel(mp, vp);
1027 	else if (mp == NULL && vops != &dead_vnodeops)
1028 		printf("NULL mp in getnewvnode()\n");
1029 #endif
1030 	if (mp != NULL) {
1031 		bo->bo_bsize = mp->mnt_stat.f_iosize;
1032 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1033 			vp->v_vflag |= VV_NOKNOTE;
1034 	}
1035 
1036 	*vpp = vp;
1037 	return (0);
1038 }
1039 
1040 /*
1041  * Delete from old mount point vnode list, if on one.
1042  */
1043 static void
1044 delmntque(struct vnode *vp)
1045 {
1046 	struct mount *mp;
1047 
1048 	mp = vp->v_mount;
1049 	if (mp == NULL)
1050 		return;
1051 	MNT_ILOCK(mp);
1052 	vp->v_mount = NULL;
1053 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1054 		("bad mount point vnode list size"));
1055 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1056 	mp->mnt_nvnodelistsize--;
1057 	MNT_REL(mp);
1058 	MNT_IUNLOCK(mp);
1059 }
1060 
1061 static void
1062 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1063 {
1064 
1065 	vp->v_data = NULL;
1066 	vp->v_op = &dead_vnodeops;
1067 	/* XXX non mp-safe fs may still call insmntque with vnode
1068 	   unlocked */
1069 	if (!VOP_ISLOCKED(vp))
1070 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1071 	vgone(vp);
1072 	vput(vp);
1073 }
1074 
1075 /*
1076  * Insert into list of vnodes for the new mount point, if available.
1077  */
1078 int
1079 insmntque1(struct vnode *vp, struct mount *mp,
1080 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1081 {
1082 	int locked;
1083 
1084 	KASSERT(vp->v_mount == NULL,
1085 		("insmntque: vnode already on per mount vnode list"));
1086 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1087 #ifdef DEBUG_VFS_LOCKS
1088 	if (!VFS_NEEDSGIANT(mp))
1089 		ASSERT_VOP_ELOCKED(vp,
1090 		    "insmntque: mp-safe fs and non-locked vp");
1091 #endif
1092 	MNT_ILOCK(mp);
1093 	if ((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1094 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1095 	     mp->mnt_nvnodelistsize == 0)) {
1096 		locked = VOP_ISLOCKED(vp);
1097 		if (!locked || (locked == LK_EXCLUSIVE &&
1098 		     (vp->v_vflag & VV_FORCEINSMQ) == 0)) {
1099 			MNT_IUNLOCK(mp);
1100 			if (dtr != NULL)
1101 				dtr(vp, dtr_arg);
1102 			return (EBUSY);
1103 		}
1104 	}
1105 	vp->v_mount = mp;
1106 	MNT_REF(mp);
1107 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1108 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1109 		("neg mount point vnode list size"));
1110 	mp->mnt_nvnodelistsize++;
1111 	MNT_IUNLOCK(mp);
1112 	return (0);
1113 }
1114 
1115 int
1116 insmntque(struct vnode *vp, struct mount *mp)
1117 {
1118 
1119 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1120 }
1121 
1122 /*
1123  * Flush out and invalidate all buffers associated with a bufobj
1124  * Called with the underlying object locked.
1125  */
1126 int
1127 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1128 {
1129 	int error;
1130 
1131 	BO_LOCK(bo);
1132 	if (flags & V_SAVE) {
1133 		error = bufobj_wwait(bo, slpflag, slptimeo);
1134 		if (error) {
1135 			BO_UNLOCK(bo);
1136 			return (error);
1137 		}
1138 		if (bo->bo_dirty.bv_cnt > 0) {
1139 			BO_UNLOCK(bo);
1140 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1141 				return (error);
1142 			/*
1143 			 * XXX We could save a lock/unlock if this was only
1144 			 * enabled under INVARIANTS
1145 			 */
1146 			BO_LOCK(bo);
1147 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1148 				panic("vinvalbuf: dirty bufs");
1149 		}
1150 	}
1151 	/*
1152 	 * If you alter this loop please notice that interlock is dropped and
1153 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1154 	 * no race conditions occur from this.
1155 	 */
1156 	do {
1157 		error = flushbuflist(&bo->bo_clean,
1158 		    flags, bo, slpflag, slptimeo);
1159 		if (error == 0)
1160 			error = flushbuflist(&bo->bo_dirty,
1161 			    flags, bo, slpflag, slptimeo);
1162 		if (error != 0 && error != EAGAIN) {
1163 			BO_UNLOCK(bo);
1164 			return (error);
1165 		}
1166 	} while (error != 0);
1167 
1168 	/*
1169 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1170 	 * have write I/O in-progress but if there is a VM object then the
1171 	 * VM object can also have read-I/O in-progress.
1172 	 */
1173 	do {
1174 		bufobj_wwait(bo, 0, 0);
1175 		BO_UNLOCK(bo);
1176 		if (bo->bo_object != NULL) {
1177 			VM_OBJECT_LOCK(bo->bo_object);
1178 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1179 			VM_OBJECT_UNLOCK(bo->bo_object);
1180 		}
1181 		BO_LOCK(bo);
1182 	} while (bo->bo_numoutput > 0);
1183 	BO_UNLOCK(bo);
1184 
1185 	/*
1186 	 * Destroy the copy in the VM cache, too.
1187 	 */
1188 	if (bo->bo_object != NULL && (flags & (V_ALT | V_NORMAL)) == 0) {
1189 		VM_OBJECT_LOCK(bo->bo_object);
1190 		vm_object_page_remove(bo->bo_object, 0, 0,
1191 			(flags & V_SAVE) ? TRUE : FALSE);
1192 		VM_OBJECT_UNLOCK(bo->bo_object);
1193 	}
1194 
1195 #ifdef INVARIANTS
1196 	BO_LOCK(bo);
1197 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1198 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1199 		panic("vinvalbuf: flush failed");
1200 	BO_UNLOCK(bo);
1201 #endif
1202 	return (0);
1203 }
1204 
1205 /*
1206  * Flush out and invalidate all buffers associated with a vnode.
1207  * Called with the underlying object locked.
1208  */
1209 int
1210 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1211 {
1212 
1213 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1214 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1215 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1216 }
1217 
1218 /*
1219  * Flush out buffers on the specified list.
1220  *
1221  */
1222 static int
1223 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1224     int slptimeo)
1225 {
1226 	struct buf *bp, *nbp;
1227 	int retval, error;
1228 	daddr_t lblkno;
1229 	b_xflags_t xflags;
1230 
1231 	ASSERT_BO_LOCKED(bo);
1232 
1233 	retval = 0;
1234 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1235 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1236 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1237 			continue;
1238 		}
1239 		lblkno = 0;
1240 		xflags = 0;
1241 		if (nbp != NULL) {
1242 			lblkno = nbp->b_lblkno;
1243 			xflags = nbp->b_xflags &
1244 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1245 		}
1246 		retval = EAGAIN;
1247 		error = BUF_TIMELOCK(bp,
1248 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1249 		    "flushbuf", slpflag, slptimeo);
1250 		if (error) {
1251 			BO_LOCK(bo);
1252 			return (error != ENOLCK ? error : EAGAIN);
1253 		}
1254 		KASSERT(bp->b_bufobj == bo,
1255 		    ("bp %p wrong b_bufobj %p should be %p",
1256 		    bp, bp->b_bufobj, bo));
1257 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1258 			BUF_UNLOCK(bp);
1259 			BO_LOCK(bo);
1260 			return (EAGAIN);
1261 		}
1262 		/*
1263 		 * XXX Since there are no node locks for NFS, I
1264 		 * believe there is a slight chance that a delayed
1265 		 * write will occur while sleeping just above, so
1266 		 * check for it.
1267 		 */
1268 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1269 		    (flags & V_SAVE)) {
1270 			BO_LOCK(bo);
1271 			bremfree(bp);
1272 			BO_UNLOCK(bo);
1273 			bp->b_flags |= B_ASYNC;
1274 			bwrite(bp);
1275 			BO_LOCK(bo);
1276 			return (EAGAIN);	/* XXX: why not loop ? */
1277 		}
1278 		BO_LOCK(bo);
1279 		bremfree(bp);
1280 		BO_UNLOCK(bo);
1281 		bp->b_flags |= (B_INVAL | B_RELBUF);
1282 		bp->b_flags &= ~B_ASYNC;
1283 		brelse(bp);
1284 		BO_LOCK(bo);
1285 		if (nbp != NULL &&
1286 		    (nbp->b_bufobj != bo ||
1287 		     nbp->b_lblkno != lblkno ||
1288 		     (nbp->b_xflags &
1289 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1290 			break;			/* nbp invalid */
1291 	}
1292 	return (retval);
1293 }
1294 
1295 /*
1296  * Truncate a file's buffer and pages to a specified length.  This
1297  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1298  * sync activity.
1299  */
1300 int
1301 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1302     off_t length, int blksize)
1303 {
1304 	struct buf *bp, *nbp;
1305 	int anyfreed;
1306 	int trunclbn;
1307 	struct bufobj *bo;
1308 
1309 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1310 	    vp, cred, blksize, (uintmax_t)length);
1311 
1312 	/*
1313 	 * Round up to the *next* lbn.
1314 	 */
1315 	trunclbn = (length + blksize - 1) / blksize;
1316 
1317 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1318 restart:
1319 	bo = &vp->v_bufobj;
1320 	BO_LOCK(bo);
1321 	anyfreed = 1;
1322 	for (;anyfreed;) {
1323 		anyfreed = 0;
1324 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1325 			if (bp->b_lblkno < trunclbn)
1326 				continue;
1327 			if (BUF_LOCK(bp,
1328 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1329 			    BO_MTX(bo)) == ENOLCK)
1330 				goto restart;
1331 
1332 			BO_LOCK(bo);
1333 			bremfree(bp);
1334 			BO_UNLOCK(bo);
1335 			bp->b_flags |= (B_INVAL | B_RELBUF);
1336 			bp->b_flags &= ~B_ASYNC;
1337 			brelse(bp);
1338 			anyfreed = 1;
1339 
1340 			if (nbp != NULL &&
1341 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1342 			    (nbp->b_vp != vp) ||
1343 			    (nbp->b_flags & B_DELWRI))) {
1344 				goto restart;
1345 			}
1346 			BO_LOCK(bo);
1347 		}
1348 
1349 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1350 			if (bp->b_lblkno < trunclbn)
1351 				continue;
1352 			if (BUF_LOCK(bp,
1353 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1354 			    BO_MTX(bo)) == ENOLCK)
1355 				goto restart;
1356 			BO_LOCK(bo);
1357 			bremfree(bp);
1358 			BO_UNLOCK(bo);
1359 			bp->b_flags |= (B_INVAL | B_RELBUF);
1360 			bp->b_flags &= ~B_ASYNC;
1361 			brelse(bp);
1362 			anyfreed = 1;
1363 			if (nbp != NULL &&
1364 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1365 			    (nbp->b_vp != vp) ||
1366 			    (nbp->b_flags & B_DELWRI) == 0)) {
1367 				goto restart;
1368 			}
1369 			BO_LOCK(bo);
1370 		}
1371 	}
1372 
1373 	if (length > 0) {
1374 restartsync:
1375 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1376 			if (bp->b_lblkno > 0)
1377 				continue;
1378 			/*
1379 			 * Since we hold the vnode lock this should only
1380 			 * fail if we're racing with the buf daemon.
1381 			 */
1382 			if (BUF_LOCK(bp,
1383 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1384 			    BO_MTX(bo)) == ENOLCK) {
1385 				goto restart;
1386 			}
1387 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1388 			    ("buf(%p) on dirty queue without DELWRI", bp));
1389 
1390 			BO_LOCK(bo);
1391 			bremfree(bp);
1392 			BO_UNLOCK(bo);
1393 			bawrite(bp);
1394 			BO_LOCK(bo);
1395 			goto restartsync;
1396 		}
1397 	}
1398 
1399 	bufobj_wwait(bo, 0, 0);
1400 	BO_UNLOCK(bo);
1401 	vnode_pager_setsize(vp, length);
1402 
1403 	return (0);
1404 }
1405 
1406 /*
1407  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1408  *		 a vnode.
1409  *
1410  *	NOTE: We have to deal with the special case of a background bitmap
1411  *	buffer, a situation where two buffers will have the same logical
1412  *	block offset.  We want (1) only the foreground buffer to be accessed
1413  *	in a lookup and (2) must differentiate between the foreground and
1414  *	background buffer in the splay tree algorithm because the splay
1415  *	tree cannot normally handle multiple entities with the same 'index'.
1416  *	We accomplish this by adding differentiating flags to the splay tree's
1417  *	numerical domain.
1418  */
1419 static
1420 struct buf *
1421 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1422 {
1423 	struct buf dummy;
1424 	struct buf *lefttreemax, *righttreemin, *y;
1425 
1426 	if (root == NULL)
1427 		return (NULL);
1428 	lefttreemax = righttreemin = &dummy;
1429 	for (;;) {
1430 		if (lblkno < root->b_lblkno ||
1431 		    (lblkno == root->b_lblkno &&
1432 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1433 			if ((y = root->b_left) == NULL)
1434 				break;
1435 			if (lblkno < y->b_lblkno) {
1436 				/* Rotate right. */
1437 				root->b_left = y->b_right;
1438 				y->b_right = root;
1439 				root = y;
1440 				if ((y = root->b_left) == NULL)
1441 					break;
1442 			}
1443 			/* Link into the new root's right tree. */
1444 			righttreemin->b_left = root;
1445 			righttreemin = root;
1446 		} else if (lblkno > root->b_lblkno ||
1447 		    (lblkno == root->b_lblkno &&
1448 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1449 			if ((y = root->b_right) == NULL)
1450 				break;
1451 			if (lblkno > y->b_lblkno) {
1452 				/* Rotate left. */
1453 				root->b_right = y->b_left;
1454 				y->b_left = root;
1455 				root = y;
1456 				if ((y = root->b_right) == NULL)
1457 					break;
1458 			}
1459 			/* Link into the new root's left tree. */
1460 			lefttreemax->b_right = root;
1461 			lefttreemax = root;
1462 		} else {
1463 			break;
1464 		}
1465 		root = y;
1466 	}
1467 	/* Assemble the new root. */
1468 	lefttreemax->b_right = root->b_left;
1469 	righttreemin->b_left = root->b_right;
1470 	root->b_left = dummy.b_right;
1471 	root->b_right = dummy.b_left;
1472 	return (root);
1473 }
1474 
1475 static void
1476 buf_vlist_remove(struct buf *bp)
1477 {
1478 	struct buf *root;
1479 	struct bufv *bv;
1480 
1481 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1482 	ASSERT_BO_LOCKED(bp->b_bufobj);
1483 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1484 	    (BX_VNDIRTY|BX_VNCLEAN),
1485 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1486 	if (bp->b_xflags & BX_VNDIRTY)
1487 		bv = &bp->b_bufobj->bo_dirty;
1488 	else
1489 		bv = &bp->b_bufobj->bo_clean;
1490 	if (bp != bv->bv_root) {
1491 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1492 		KASSERT(root == bp, ("splay lookup failed in remove"));
1493 	}
1494 	if (bp->b_left == NULL) {
1495 		root = bp->b_right;
1496 	} else {
1497 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1498 		root->b_right = bp->b_right;
1499 	}
1500 	bv->bv_root = root;
1501 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1502 	bv->bv_cnt--;
1503 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1504 }
1505 
1506 /*
1507  * Add the buffer to the sorted clean or dirty block list using a
1508  * splay tree algorithm.
1509  *
1510  * NOTE: xflags is passed as a constant, optimizing this inline function!
1511  */
1512 static void
1513 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1514 {
1515 	struct buf *root;
1516 	struct bufv *bv;
1517 
1518 	ASSERT_BO_LOCKED(bo);
1519 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1520 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1521 	bp->b_xflags |= xflags;
1522 	if (xflags & BX_VNDIRTY)
1523 		bv = &bo->bo_dirty;
1524 	else
1525 		bv = &bo->bo_clean;
1526 
1527 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1528 	if (root == NULL) {
1529 		bp->b_left = NULL;
1530 		bp->b_right = NULL;
1531 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1532 	} else if (bp->b_lblkno < root->b_lblkno ||
1533 	    (bp->b_lblkno == root->b_lblkno &&
1534 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1535 		bp->b_left = root->b_left;
1536 		bp->b_right = root;
1537 		root->b_left = NULL;
1538 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1539 	} else {
1540 		bp->b_right = root->b_right;
1541 		bp->b_left = root;
1542 		root->b_right = NULL;
1543 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1544 	}
1545 	bv->bv_cnt++;
1546 	bv->bv_root = bp;
1547 }
1548 
1549 /*
1550  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1551  * shadow buffers used in background bitmap writes.
1552  *
1553  * This code isn't quite efficient as it could be because we are maintaining
1554  * two sorted lists and do not know which list the block resides in.
1555  *
1556  * During a "make buildworld" the desired buffer is found at one of
1557  * the roots more than 60% of the time.  Thus, checking both roots
1558  * before performing either splay eliminates unnecessary splays on the
1559  * first tree splayed.
1560  */
1561 struct buf *
1562 gbincore(struct bufobj *bo, daddr_t lblkno)
1563 {
1564 	struct buf *bp;
1565 
1566 	ASSERT_BO_LOCKED(bo);
1567 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1568 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1569 		return (bp);
1570 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1571 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1572 		return (bp);
1573 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1574 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1575 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1576 			return (bp);
1577 	}
1578 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1579 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1580 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1581 			return (bp);
1582 	}
1583 	return (NULL);
1584 }
1585 
1586 /*
1587  * Associate a buffer with a vnode.
1588  */
1589 void
1590 bgetvp(struct vnode *vp, struct buf *bp)
1591 {
1592 	struct bufobj *bo;
1593 
1594 	bo = &vp->v_bufobj;
1595 	ASSERT_BO_LOCKED(bo);
1596 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1597 
1598 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1599 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1600 	    ("bgetvp: bp already attached! %p", bp));
1601 
1602 	vhold(vp);
1603 	if (VFS_NEEDSGIANT(vp->v_mount) || bo->bo_flag & BO_NEEDSGIANT)
1604 		bp->b_flags |= B_NEEDSGIANT;
1605 	bp->b_vp = vp;
1606 	bp->b_bufobj = bo;
1607 	/*
1608 	 * Insert onto list for new vnode.
1609 	 */
1610 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1611 }
1612 
1613 /*
1614  * Disassociate a buffer from a vnode.
1615  */
1616 void
1617 brelvp(struct buf *bp)
1618 {
1619 	struct bufobj *bo;
1620 	struct vnode *vp;
1621 
1622 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1623 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1624 
1625 	/*
1626 	 * Delete from old vnode list, if on one.
1627 	 */
1628 	vp = bp->b_vp;		/* XXX */
1629 	bo = bp->b_bufobj;
1630 	BO_LOCK(bo);
1631 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1632 		buf_vlist_remove(bp);
1633 	else
1634 		panic("brelvp: Buffer %p not on queue.", bp);
1635 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1636 		bo->bo_flag &= ~BO_ONWORKLST;
1637 		mtx_lock(&sync_mtx);
1638 		LIST_REMOVE(bo, bo_synclist);
1639 		syncer_worklist_len--;
1640 		mtx_unlock(&sync_mtx);
1641 	}
1642 	bp->b_flags &= ~B_NEEDSGIANT;
1643 	bp->b_vp = NULL;
1644 	bp->b_bufobj = NULL;
1645 	BO_UNLOCK(bo);
1646 	vdrop(vp);
1647 }
1648 
1649 /*
1650  * Add an item to the syncer work queue.
1651  */
1652 static void
1653 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1654 {
1655 	int queue, slot;
1656 
1657 	ASSERT_BO_LOCKED(bo);
1658 
1659 	mtx_lock(&sync_mtx);
1660 	if (bo->bo_flag & BO_ONWORKLST)
1661 		LIST_REMOVE(bo, bo_synclist);
1662 	else {
1663 		bo->bo_flag |= BO_ONWORKLST;
1664 		syncer_worklist_len++;
1665 	}
1666 
1667 	if (delay > syncer_maxdelay - 2)
1668 		delay = syncer_maxdelay - 2;
1669 	slot = (syncer_delayno + delay) & syncer_mask;
1670 
1671 	queue = VFS_NEEDSGIANT(bo->__bo_vnode->v_mount) ? WI_GIANTQ :
1672 	    WI_MPSAFEQ;
1673 	LIST_INSERT_HEAD(&syncer_workitem_pending[queue][slot], bo,
1674 	    bo_synclist);
1675 	mtx_unlock(&sync_mtx);
1676 }
1677 
1678 static int
1679 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1680 {
1681 	int error, len;
1682 
1683 	mtx_lock(&sync_mtx);
1684 	len = syncer_worklist_len - sync_vnode_count;
1685 	mtx_unlock(&sync_mtx);
1686 	error = SYSCTL_OUT(req, &len, sizeof(len));
1687 	return (error);
1688 }
1689 
1690 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1691     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1692 
1693 static struct proc *updateproc;
1694 static void sched_sync(void);
1695 static struct kproc_desc up_kp = {
1696 	"syncer",
1697 	sched_sync,
1698 	&updateproc
1699 };
1700 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
1701 
1702 static int
1703 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
1704 {
1705 	struct vnode *vp;
1706 	struct mount *mp;
1707 
1708 	*bo = LIST_FIRST(slp);
1709 	if (*bo == NULL)
1710 		return (0);
1711 	vp = (*bo)->__bo_vnode;	/* XXX */
1712 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
1713 		return (1);
1714 	/*
1715 	 * We use vhold in case the vnode does not
1716 	 * successfully sync.  vhold prevents the vnode from
1717 	 * going away when we unlock the sync_mtx so that
1718 	 * we can acquire the vnode interlock.
1719 	 */
1720 	vholdl(vp);
1721 	mtx_unlock(&sync_mtx);
1722 	VI_UNLOCK(vp);
1723 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1724 		vdrop(vp);
1725 		mtx_lock(&sync_mtx);
1726 		return (*bo == LIST_FIRST(slp));
1727 	}
1728 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1729 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1730 	VOP_UNLOCK(vp, 0);
1731 	vn_finished_write(mp);
1732 	BO_LOCK(*bo);
1733 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
1734 		/*
1735 		 * Put us back on the worklist.  The worklist
1736 		 * routine will remove us from our current
1737 		 * position and then add us back in at a later
1738 		 * position.
1739 		 */
1740 		vn_syncer_add_to_worklist(*bo, syncdelay);
1741 	}
1742 	BO_UNLOCK(*bo);
1743 	vdrop(vp);
1744 	mtx_lock(&sync_mtx);
1745 	return (0);
1746 }
1747 
1748 /*
1749  * System filesystem synchronizer daemon.
1750  */
1751 static void
1752 sched_sync(void)
1753 {
1754 	struct synclist *gnext, *next;
1755 	struct synclist *gslp, *slp;
1756 	struct bufobj *bo;
1757 	long starttime;
1758 	struct thread *td = curthread;
1759 	int last_work_seen;
1760 	int net_worklist_len;
1761 	int syncer_final_iter;
1762 	int first_printf;
1763 	int error;
1764 
1765 	last_work_seen = 0;
1766 	syncer_final_iter = 0;
1767 	first_printf = 1;
1768 	syncer_state = SYNCER_RUNNING;
1769 	starttime = time_uptime;
1770 	td->td_pflags |= TDP_NORUNNINGBUF;
1771 
1772 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1773 	    SHUTDOWN_PRI_LAST);
1774 
1775 	mtx_lock(&sync_mtx);
1776 	for (;;) {
1777 		if (syncer_state == SYNCER_FINAL_DELAY &&
1778 		    syncer_final_iter == 0) {
1779 			mtx_unlock(&sync_mtx);
1780 			kproc_suspend_check(td->td_proc);
1781 			mtx_lock(&sync_mtx);
1782 		}
1783 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1784 		if (syncer_state != SYNCER_RUNNING &&
1785 		    starttime != time_uptime) {
1786 			if (first_printf) {
1787 				printf("\nSyncing disks, vnodes remaining...");
1788 				first_printf = 0;
1789 			}
1790 			printf("%d ", net_worklist_len);
1791 		}
1792 		starttime = time_uptime;
1793 
1794 		/*
1795 		 * Push files whose dirty time has expired.  Be careful
1796 		 * of interrupt race on slp queue.
1797 		 *
1798 		 * Skip over empty worklist slots when shutting down.
1799 		 */
1800 		do {
1801 			slp = &syncer_workitem_pending[WI_MPSAFEQ][syncer_delayno];
1802 			gslp = &syncer_workitem_pending[WI_GIANTQ][syncer_delayno];
1803 			syncer_delayno += 1;
1804 			if (syncer_delayno == syncer_maxdelay)
1805 				syncer_delayno = 0;
1806 			next = &syncer_workitem_pending[WI_MPSAFEQ][syncer_delayno];
1807 			gnext = &syncer_workitem_pending[WI_GIANTQ][syncer_delayno];
1808 			/*
1809 			 * If the worklist has wrapped since the
1810 			 * it was emptied of all but syncer vnodes,
1811 			 * switch to the FINAL_DELAY state and run
1812 			 * for one more second.
1813 			 */
1814 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1815 			    net_worklist_len == 0 &&
1816 			    last_work_seen == syncer_delayno) {
1817 				syncer_state = SYNCER_FINAL_DELAY;
1818 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1819 			}
1820 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1821 		    LIST_EMPTY(gslp) && syncer_worklist_len > 0);
1822 
1823 		/*
1824 		 * Keep track of the last time there was anything
1825 		 * on the worklist other than syncer vnodes.
1826 		 * Return to the SHUTTING_DOWN state if any
1827 		 * new work appears.
1828 		 */
1829 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1830 			last_work_seen = syncer_delayno;
1831 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1832 			syncer_state = SYNCER_SHUTTING_DOWN;
1833 		while (!LIST_EMPTY(slp)) {
1834 			error = sync_vnode(slp, &bo, td);
1835 			if (error == 1) {
1836 				LIST_REMOVE(bo, bo_synclist);
1837 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1838 				continue;
1839 			}
1840 		}
1841 		if (!LIST_EMPTY(gslp)) {
1842 			mtx_unlock(&sync_mtx);
1843 			mtx_lock(&Giant);
1844 			mtx_lock(&sync_mtx);
1845 			while (!LIST_EMPTY(gslp)) {
1846 				error = sync_vnode(gslp, &bo, td);
1847 				if (error == 1) {
1848 					LIST_REMOVE(bo, bo_synclist);
1849 					LIST_INSERT_HEAD(gnext, bo,
1850 					    bo_synclist);
1851 					continue;
1852 				}
1853 			}
1854 			mtx_unlock(&Giant);
1855 		}
1856 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1857 			syncer_final_iter--;
1858 		/*
1859 		 * The variable rushjob allows the kernel to speed up the
1860 		 * processing of the filesystem syncer process. A rushjob
1861 		 * value of N tells the filesystem syncer to process the next
1862 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1863 		 * is used by the soft update code to speed up the filesystem
1864 		 * syncer process when the incore state is getting so far
1865 		 * ahead of the disk that the kernel memory pool is being
1866 		 * threatened with exhaustion.
1867 		 */
1868 		if (rushjob > 0) {
1869 			rushjob -= 1;
1870 			continue;
1871 		}
1872 		/*
1873 		 * Just sleep for a short period of time between
1874 		 * iterations when shutting down to allow some I/O
1875 		 * to happen.
1876 		 *
1877 		 * If it has taken us less than a second to process the
1878 		 * current work, then wait. Otherwise start right over
1879 		 * again. We can still lose time if any single round
1880 		 * takes more than two seconds, but it does not really
1881 		 * matter as we are just trying to generally pace the
1882 		 * filesystem activity.
1883 		 */
1884 		if (syncer_state != SYNCER_RUNNING ||
1885 		    time_uptime == starttime) {
1886 			thread_lock(td);
1887 			sched_prio(td, PPAUSE);
1888 			thread_unlock(td);
1889 		}
1890 		if (syncer_state != SYNCER_RUNNING)
1891 			cv_timedwait(&sync_wakeup, &sync_mtx,
1892 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1893 		else if (time_uptime == starttime)
1894 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
1895 	}
1896 }
1897 
1898 /*
1899  * Request the syncer daemon to speed up its work.
1900  * We never push it to speed up more than half of its
1901  * normal turn time, otherwise it could take over the cpu.
1902  */
1903 int
1904 speedup_syncer(void)
1905 {
1906 	int ret = 0;
1907 
1908 	mtx_lock(&sync_mtx);
1909 	if (rushjob < syncdelay / 2) {
1910 		rushjob += 1;
1911 		stat_rush_requests += 1;
1912 		ret = 1;
1913 	}
1914 	mtx_unlock(&sync_mtx);
1915 	cv_broadcast(&sync_wakeup);
1916 	return (ret);
1917 }
1918 
1919 /*
1920  * Tell the syncer to speed up its work and run though its work
1921  * list several times, then tell it to shut down.
1922  */
1923 static void
1924 syncer_shutdown(void *arg, int howto)
1925 {
1926 
1927 	if (howto & RB_NOSYNC)
1928 		return;
1929 	mtx_lock(&sync_mtx);
1930 	syncer_state = SYNCER_SHUTTING_DOWN;
1931 	rushjob = 0;
1932 	mtx_unlock(&sync_mtx);
1933 	cv_broadcast(&sync_wakeup);
1934 	kproc_shutdown(arg, howto);
1935 }
1936 
1937 /*
1938  * Reassign a buffer from one vnode to another.
1939  * Used to assign file specific control information
1940  * (indirect blocks) to the vnode to which they belong.
1941  */
1942 void
1943 reassignbuf(struct buf *bp)
1944 {
1945 	struct vnode *vp;
1946 	struct bufobj *bo;
1947 	int delay;
1948 #ifdef INVARIANTS
1949 	struct bufv *bv;
1950 #endif
1951 
1952 	vp = bp->b_vp;
1953 	bo = bp->b_bufobj;
1954 	++reassignbufcalls;
1955 
1956 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1957 	    bp, bp->b_vp, bp->b_flags);
1958 	/*
1959 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1960 	 * is not fully linked in.
1961 	 */
1962 	if (bp->b_flags & B_PAGING)
1963 		panic("cannot reassign paging buffer");
1964 
1965 	/*
1966 	 * Delete from old vnode list, if on one.
1967 	 */
1968 	BO_LOCK(bo);
1969 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1970 		buf_vlist_remove(bp);
1971 	else
1972 		panic("reassignbuf: Buffer %p not on queue.", bp);
1973 	/*
1974 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1975 	 * of clean buffers.
1976 	 */
1977 	if (bp->b_flags & B_DELWRI) {
1978 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1979 			switch (vp->v_type) {
1980 			case VDIR:
1981 				delay = dirdelay;
1982 				break;
1983 			case VCHR:
1984 				delay = metadelay;
1985 				break;
1986 			default:
1987 				delay = filedelay;
1988 			}
1989 			vn_syncer_add_to_worklist(bo, delay);
1990 		}
1991 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1992 	} else {
1993 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1994 
1995 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1996 			mtx_lock(&sync_mtx);
1997 			LIST_REMOVE(bo, bo_synclist);
1998 			syncer_worklist_len--;
1999 			mtx_unlock(&sync_mtx);
2000 			bo->bo_flag &= ~BO_ONWORKLST;
2001 		}
2002 	}
2003 #ifdef INVARIANTS
2004 	bv = &bo->bo_clean;
2005 	bp = TAILQ_FIRST(&bv->bv_hd);
2006 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2007 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2008 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2009 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2010 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2011 	bv = &bo->bo_dirty;
2012 	bp = TAILQ_FIRST(&bv->bv_hd);
2013 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2014 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2015 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2016 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2017 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2018 #endif
2019 	BO_UNLOCK(bo);
2020 }
2021 
2022 /*
2023  * Increment the use and hold counts on the vnode, taking care to reference
2024  * the driver's usecount if this is a chardev.  The vholdl() will remove
2025  * the vnode from the free list if it is presently free.  Requires the
2026  * vnode interlock and returns with it held.
2027  */
2028 static void
2029 v_incr_usecount(struct vnode *vp)
2030 {
2031 
2032 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2033 	vp->v_usecount++;
2034 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2035 		dev_lock();
2036 		vp->v_rdev->si_usecount++;
2037 		dev_unlock();
2038 	}
2039 	vholdl(vp);
2040 }
2041 
2042 /*
2043  * Turn a holdcnt into a use+holdcnt such that only one call to
2044  * v_decr_usecount is needed.
2045  */
2046 static void
2047 v_upgrade_usecount(struct vnode *vp)
2048 {
2049 
2050 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2051 	vp->v_usecount++;
2052 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2053 		dev_lock();
2054 		vp->v_rdev->si_usecount++;
2055 		dev_unlock();
2056 	}
2057 }
2058 
2059 /*
2060  * Decrement the vnode use and hold count along with the driver's usecount
2061  * if this is a chardev.  The vdropl() below releases the vnode interlock
2062  * as it may free the vnode.
2063  */
2064 static void
2065 v_decr_usecount(struct vnode *vp)
2066 {
2067 
2068 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2069 	VNASSERT(vp->v_usecount > 0, vp,
2070 	    ("v_decr_usecount: negative usecount"));
2071 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2072 	vp->v_usecount--;
2073 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2074 		dev_lock();
2075 		vp->v_rdev->si_usecount--;
2076 		dev_unlock();
2077 	}
2078 	vdropl(vp);
2079 }
2080 
2081 /*
2082  * Decrement only the use count and driver use count.  This is intended to
2083  * be paired with a follow on vdropl() to release the remaining hold count.
2084  * In this way we may vgone() a vnode with a 0 usecount without risk of
2085  * having it end up on a free list because the hold count is kept above 0.
2086  */
2087 static void
2088 v_decr_useonly(struct vnode *vp)
2089 {
2090 
2091 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2092 	VNASSERT(vp->v_usecount > 0, vp,
2093 	    ("v_decr_useonly: negative usecount"));
2094 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2095 	vp->v_usecount--;
2096 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2097 		dev_lock();
2098 		vp->v_rdev->si_usecount--;
2099 		dev_unlock();
2100 	}
2101 }
2102 
2103 /*
2104  * Grab a particular vnode from the free list, increment its
2105  * reference count and lock it.  VI_DOOMED is set if the vnode
2106  * is being destroyed.  Only callers who specify LK_RETRY will
2107  * see doomed vnodes.  If inactive processing was delayed in
2108  * vput try to do it here.
2109  */
2110 int
2111 vget(struct vnode *vp, int flags, struct thread *td)
2112 {
2113 	int error;
2114 
2115 	error = 0;
2116 	VFS_ASSERT_GIANT(vp->v_mount);
2117 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2118 	    ("vget: invalid lock operation"));
2119 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2120 
2121 	if ((flags & LK_INTERLOCK) == 0)
2122 		VI_LOCK(vp);
2123 	vholdl(vp);
2124 	if ((error = vn_lock(vp, flags | LK_INTERLOCK)) != 0) {
2125 		vdrop(vp);
2126 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2127 		    vp);
2128 		return (error);
2129 	}
2130 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2131 		panic("vget: vn_lock failed to return ENOENT\n");
2132 	VI_LOCK(vp);
2133 	/* Upgrade our holdcnt to a usecount. */
2134 	v_upgrade_usecount(vp);
2135 	/*
2136 	 * We don't guarantee that any particular close will
2137 	 * trigger inactive processing so just make a best effort
2138 	 * here at preventing a reference to a removed file.  If
2139 	 * we don't succeed no harm is done.
2140 	 */
2141 	if (vp->v_iflag & VI_OWEINACT) {
2142 		if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2143 		    (flags & LK_NOWAIT) == 0)
2144 			vinactive(vp, td);
2145 		vp->v_iflag &= ~VI_OWEINACT;
2146 	}
2147 	VI_UNLOCK(vp);
2148 	return (0);
2149 }
2150 
2151 /*
2152  * Increase the reference count of a vnode.
2153  */
2154 void
2155 vref(struct vnode *vp)
2156 {
2157 
2158 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2159 	VI_LOCK(vp);
2160 	v_incr_usecount(vp);
2161 	VI_UNLOCK(vp);
2162 }
2163 
2164 /*
2165  * Return reference count of a vnode.
2166  *
2167  * The results of this call are only guaranteed when some mechanism other
2168  * than the VI lock is used to stop other processes from gaining references
2169  * to the vnode.  This may be the case if the caller holds the only reference.
2170  * This is also useful when stale data is acceptable as race conditions may
2171  * be accounted for by some other means.
2172  */
2173 int
2174 vrefcnt(struct vnode *vp)
2175 {
2176 	int usecnt;
2177 
2178 	VI_LOCK(vp);
2179 	usecnt = vp->v_usecount;
2180 	VI_UNLOCK(vp);
2181 
2182 	return (usecnt);
2183 }
2184 
2185 #define	VPUTX_VRELE	1
2186 #define	VPUTX_VPUT	2
2187 #define	VPUTX_VUNREF	3
2188 
2189 static void
2190 vputx(struct vnode *vp, int func)
2191 {
2192 	int error;
2193 
2194 	KASSERT(vp != NULL, ("vputx: null vp"));
2195 	if (func == VPUTX_VUNREF)
2196 		ASSERT_VOP_LOCKED(vp, "vunref");
2197 	else if (func == VPUTX_VPUT)
2198 		ASSERT_VOP_LOCKED(vp, "vput");
2199 	else
2200 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2201 	VFS_ASSERT_GIANT(vp->v_mount);
2202 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2203 	VI_LOCK(vp);
2204 
2205 	/* Skip this v_writecount check if we're going to panic below. */
2206 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2207 	    ("vputx: missed vn_close"));
2208 	error = 0;
2209 
2210 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2211 	    vp->v_usecount == 1)) {
2212 		if (func == VPUTX_VPUT)
2213 			VOP_UNLOCK(vp, 0);
2214 		v_decr_usecount(vp);
2215 		return;
2216 	}
2217 
2218 	if (vp->v_usecount != 1) {
2219 		vprint("vputx: negative ref count", vp);
2220 		panic("vputx: negative ref cnt");
2221 	}
2222 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2223 	/*
2224 	 * We want to hold the vnode until the inactive finishes to
2225 	 * prevent vgone() races.  We drop the use count here and the
2226 	 * hold count below when we're done.
2227 	 */
2228 	v_decr_useonly(vp);
2229 	/*
2230 	 * We must call VOP_INACTIVE with the node locked. Mark
2231 	 * as VI_DOINGINACT to avoid recursion.
2232 	 */
2233 	vp->v_iflag |= VI_OWEINACT;
2234 	switch (func) {
2235 	case VPUTX_VRELE:
2236 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2237 		VI_LOCK(vp);
2238 		break;
2239 	case VPUTX_VPUT:
2240 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2241 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2242 			    LK_NOWAIT);
2243 			VI_LOCK(vp);
2244 		}
2245 		break;
2246 	case VPUTX_VUNREF:
2247 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
2248 			error = EBUSY;
2249 		break;
2250 	}
2251 	if (vp->v_usecount > 0)
2252 		vp->v_iflag &= ~VI_OWEINACT;
2253 	if (error == 0) {
2254 		if (vp->v_iflag & VI_OWEINACT)
2255 			vinactive(vp, curthread);
2256 		if (func != VPUTX_VUNREF)
2257 			VOP_UNLOCK(vp, 0);
2258 	}
2259 	vdropl(vp);
2260 }
2261 
2262 /*
2263  * Vnode put/release.
2264  * If count drops to zero, call inactive routine and return to freelist.
2265  */
2266 void
2267 vrele(struct vnode *vp)
2268 {
2269 
2270 	vputx(vp, VPUTX_VRELE);
2271 }
2272 
2273 /*
2274  * Release an already locked vnode.  This give the same effects as
2275  * unlock+vrele(), but takes less time and avoids releasing and
2276  * re-aquiring the lock (as vrele() acquires the lock internally.)
2277  */
2278 void
2279 vput(struct vnode *vp)
2280 {
2281 
2282 	vputx(vp, VPUTX_VPUT);
2283 }
2284 
2285 /*
2286  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2287  */
2288 void
2289 vunref(struct vnode *vp)
2290 {
2291 
2292 	vputx(vp, VPUTX_VUNREF);
2293 }
2294 
2295 /*
2296  * Somebody doesn't want the vnode recycled.
2297  */
2298 void
2299 vhold(struct vnode *vp)
2300 {
2301 
2302 	VI_LOCK(vp);
2303 	vholdl(vp);
2304 	VI_UNLOCK(vp);
2305 }
2306 
2307 void
2308 vholdl(struct vnode *vp)
2309 {
2310 
2311 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2312 	vp->v_holdcnt++;
2313 	if (VSHOULDBUSY(vp))
2314 		vbusy(vp);
2315 }
2316 
2317 /*
2318  * Note that there is one less who cares about this vnode.  vdrop() is the
2319  * opposite of vhold().
2320  */
2321 void
2322 vdrop(struct vnode *vp)
2323 {
2324 
2325 	VI_LOCK(vp);
2326 	vdropl(vp);
2327 }
2328 
2329 /*
2330  * Drop the hold count of the vnode.  If this is the last reference to
2331  * the vnode we will free it if it has been vgone'd otherwise it is
2332  * placed on the free list.
2333  */
2334 void
2335 vdropl(struct vnode *vp)
2336 {
2337 
2338 	ASSERT_VI_LOCKED(vp, "vdropl");
2339 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2340 	if (vp->v_holdcnt <= 0)
2341 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2342 	vp->v_holdcnt--;
2343 	if (vp->v_holdcnt == 0) {
2344 		if (vp->v_iflag & VI_DOOMED) {
2345 			CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__,
2346 			    vp);
2347 			vdestroy(vp);
2348 			return;
2349 		} else
2350 			vfree(vp);
2351 	}
2352 	VI_UNLOCK(vp);
2353 }
2354 
2355 /*
2356  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2357  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2358  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2359  * failed lock upgrade.
2360  */
2361 static void
2362 vinactive(struct vnode *vp, struct thread *td)
2363 {
2364 
2365 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2366 	ASSERT_VI_LOCKED(vp, "vinactive");
2367 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2368 	    ("vinactive: recursed on VI_DOINGINACT"));
2369 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2370 	vp->v_iflag |= VI_DOINGINACT;
2371 	vp->v_iflag &= ~VI_OWEINACT;
2372 	VI_UNLOCK(vp);
2373 	VOP_INACTIVE(vp, td);
2374 	VI_LOCK(vp);
2375 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2376 	    ("vinactive: lost VI_DOINGINACT"));
2377 	vp->v_iflag &= ~VI_DOINGINACT;
2378 }
2379 
2380 /*
2381  * Remove any vnodes in the vnode table belonging to mount point mp.
2382  *
2383  * If FORCECLOSE is not specified, there should not be any active ones,
2384  * return error if any are found (nb: this is a user error, not a
2385  * system error). If FORCECLOSE is specified, detach any active vnodes
2386  * that are found.
2387  *
2388  * If WRITECLOSE is set, only flush out regular file vnodes open for
2389  * writing.
2390  *
2391  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2392  *
2393  * `rootrefs' specifies the base reference count for the root vnode
2394  * of this filesystem. The root vnode is considered busy if its
2395  * v_usecount exceeds this value. On a successful return, vflush(, td)
2396  * will call vrele() on the root vnode exactly rootrefs times.
2397  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2398  * be zero.
2399  */
2400 #ifdef DIAGNOSTIC
2401 static int busyprt = 0;		/* print out busy vnodes */
2402 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2403 #endif
2404 
2405 int
2406 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2407 {
2408 	struct vnode *vp, *mvp, *rootvp = NULL;
2409 	struct vattr vattr;
2410 	int busy = 0, error;
2411 
2412 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2413 	    rootrefs, flags);
2414 	if (rootrefs > 0) {
2415 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2416 		    ("vflush: bad args"));
2417 		/*
2418 		 * Get the filesystem root vnode. We can vput() it
2419 		 * immediately, since with rootrefs > 0, it won't go away.
2420 		 */
2421 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2422 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2423 			    __func__, error);
2424 			return (error);
2425 		}
2426 		vput(rootvp);
2427 	}
2428 	MNT_ILOCK(mp);
2429 loop:
2430 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2431 		VI_LOCK(vp);
2432 		vholdl(vp);
2433 		MNT_IUNLOCK(mp);
2434 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2435 		if (error) {
2436 			vdrop(vp);
2437 			MNT_ILOCK(mp);
2438 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2439 			goto loop;
2440 		}
2441 		/*
2442 		 * Skip over a vnodes marked VV_SYSTEM.
2443 		 */
2444 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2445 			VOP_UNLOCK(vp, 0);
2446 			vdrop(vp);
2447 			MNT_ILOCK(mp);
2448 			continue;
2449 		}
2450 		/*
2451 		 * If WRITECLOSE is set, flush out unlinked but still open
2452 		 * files (even if open only for reading) and regular file
2453 		 * vnodes open for writing.
2454 		 */
2455 		if (flags & WRITECLOSE) {
2456 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
2457 			VI_LOCK(vp);
2458 
2459 			if ((vp->v_type == VNON ||
2460 			    (error == 0 && vattr.va_nlink > 0)) &&
2461 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2462 				VOP_UNLOCK(vp, 0);
2463 				vdropl(vp);
2464 				MNT_ILOCK(mp);
2465 				continue;
2466 			}
2467 		} else
2468 			VI_LOCK(vp);
2469 		/*
2470 		 * With v_usecount == 0, all we need to do is clear out the
2471 		 * vnode data structures and we are done.
2472 		 *
2473 		 * If FORCECLOSE is set, forcibly close the vnode.
2474 		 */
2475 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2476 			VNASSERT(vp->v_usecount == 0 ||
2477 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2478 			    ("device VNODE %p is FORCECLOSED", vp));
2479 			vgonel(vp);
2480 		} else {
2481 			busy++;
2482 #ifdef DIAGNOSTIC
2483 			if (busyprt)
2484 				vprint("vflush: busy vnode", vp);
2485 #endif
2486 		}
2487 		VOP_UNLOCK(vp, 0);
2488 		vdropl(vp);
2489 		MNT_ILOCK(mp);
2490 	}
2491 	MNT_IUNLOCK(mp);
2492 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2493 		/*
2494 		 * If just the root vnode is busy, and if its refcount
2495 		 * is equal to `rootrefs', then go ahead and kill it.
2496 		 */
2497 		VI_LOCK(rootvp);
2498 		KASSERT(busy > 0, ("vflush: not busy"));
2499 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2500 		    ("vflush: usecount %d < rootrefs %d",
2501 		     rootvp->v_usecount, rootrefs));
2502 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2503 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
2504 			vgone(rootvp);
2505 			VOP_UNLOCK(rootvp, 0);
2506 			busy = 0;
2507 		} else
2508 			VI_UNLOCK(rootvp);
2509 	}
2510 	if (busy) {
2511 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
2512 		    busy);
2513 		return (EBUSY);
2514 	}
2515 	for (; rootrefs > 0; rootrefs--)
2516 		vrele(rootvp);
2517 	return (0);
2518 }
2519 
2520 /*
2521  * Recycle an unused vnode to the front of the free list.
2522  */
2523 int
2524 vrecycle(struct vnode *vp, struct thread *td)
2525 {
2526 	int recycled;
2527 
2528 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
2529 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2530 	recycled = 0;
2531 	VI_LOCK(vp);
2532 	if (vp->v_usecount == 0) {
2533 		recycled = 1;
2534 		vgonel(vp);
2535 	}
2536 	VI_UNLOCK(vp);
2537 	return (recycled);
2538 }
2539 
2540 /*
2541  * Eliminate all activity associated with a vnode
2542  * in preparation for reuse.
2543  */
2544 void
2545 vgone(struct vnode *vp)
2546 {
2547 	VI_LOCK(vp);
2548 	vgonel(vp);
2549 	VI_UNLOCK(vp);
2550 }
2551 
2552 /*
2553  * vgone, with the vp interlock held.
2554  */
2555 void
2556 vgonel(struct vnode *vp)
2557 {
2558 	struct thread *td;
2559 	int oweinact;
2560 	int active;
2561 	struct mount *mp;
2562 
2563 	ASSERT_VOP_ELOCKED(vp, "vgonel");
2564 	ASSERT_VI_LOCKED(vp, "vgonel");
2565 	VNASSERT(vp->v_holdcnt, vp,
2566 	    ("vgonel: vp %p has no reference.", vp));
2567 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2568 	td = curthread;
2569 
2570 	/*
2571 	 * Don't vgonel if we're already doomed.
2572 	 */
2573 	if (vp->v_iflag & VI_DOOMED)
2574 		return;
2575 	vp->v_iflag |= VI_DOOMED;
2576 	/*
2577 	 * Check to see if the vnode is in use.  If so, we have to call
2578 	 * VOP_CLOSE() and VOP_INACTIVE().
2579 	 */
2580 	active = vp->v_usecount;
2581 	oweinact = (vp->v_iflag & VI_OWEINACT);
2582 	VI_UNLOCK(vp);
2583 	/*
2584 	 * Clean out any buffers associated with the vnode.
2585 	 * If the flush fails, just toss the buffers.
2586 	 */
2587 	mp = NULL;
2588 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2589 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
2590 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0)
2591 		vinvalbuf(vp, 0, 0, 0);
2592 
2593 	/*
2594 	 * If purging an active vnode, it must be closed and
2595 	 * deactivated before being reclaimed.
2596 	 */
2597 	if (active)
2598 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2599 	if (oweinact || active) {
2600 		VI_LOCK(vp);
2601 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2602 			vinactive(vp, td);
2603 		VI_UNLOCK(vp);
2604 	}
2605 	/*
2606 	 * Reclaim the vnode.
2607 	 */
2608 	if (VOP_RECLAIM(vp, td))
2609 		panic("vgone: cannot reclaim");
2610 	if (mp != NULL)
2611 		vn_finished_secondary_write(mp);
2612 	VNASSERT(vp->v_object == NULL, vp,
2613 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2614 	/*
2615 	 * Clear the advisory locks and wake up waiting threads.
2616 	 */
2617 	(void)VOP_ADVLOCKPURGE(vp);
2618 	/*
2619 	 * Delete from old mount point vnode list.
2620 	 */
2621 	delmntque(vp);
2622 	cache_purge(vp);
2623 	/*
2624 	 * Done with purge, reset to the standard lock and invalidate
2625 	 * the vnode.
2626 	 */
2627 	VI_LOCK(vp);
2628 	vp->v_vnlock = &vp->v_lock;
2629 	vp->v_op = &dead_vnodeops;
2630 	vp->v_tag = "none";
2631 	vp->v_type = VBAD;
2632 }
2633 
2634 /*
2635  * Calculate the total number of references to a special device.
2636  */
2637 int
2638 vcount(struct vnode *vp)
2639 {
2640 	int count;
2641 
2642 	dev_lock();
2643 	count = vp->v_rdev->si_usecount;
2644 	dev_unlock();
2645 	return (count);
2646 }
2647 
2648 /*
2649  * Same as above, but using the struct cdev *as argument
2650  */
2651 int
2652 count_dev(struct cdev *dev)
2653 {
2654 	int count;
2655 
2656 	dev_lock();
2657 	count = dev->si_usecount;
2658 	dev_unlock();
2659 	return(count);
2660 }
2661 
2662 /*
2663  * Print out a description of a vnode.
2664  */
2665 static char *typename[] =
2666 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2667  "VMARKER"};
2668 
2669 void
2670 vn_printf(struct vnode *vp, const char *fmt, ...)
2671 {
2672 	va_list ap;
2673 	char buf[256], buf2[16];
2674 	u_long flags;
2675 
2676 	va_start(ap, fmt);
2677 	vprintf(fmt, ap);
2678 	va_end(ap);
2679 	printf("%p: ", (void *)vp);
2680 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2681 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2682 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2683 	buf[0] = '\0';
2684 	buf[1] = '\0';
2685 	if (vp->v_vflag & VV_ROOT)
2686 		strlcat(buf, "|VV_ROOT", sizeof(buf));
2687 	if (vp->v_vflag & VV_ISTTY)
2688 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
2689 	if (vp->v_vflag & VV_NOSYNC)
2690 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
2691 	if (vp->v_vflag & VV_CACHEDLABEL)
2692 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
2693 	if (vp->v_vflag & VV_TEXT)
2694 		strlcat(buf, "|VV_TEXT", sizeof(buf));
2695 	if (vp->v_vflag & VV_COPYONWRITE)
2696 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
2697 	if (vp->v_vflag & VV_SYSTEM)
2698 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
2699 	if (vp->v_vflag & VV_PROCDEP)
2700 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
2701 	if (vp->v_vflag & VV_NOKNOTE)
2702 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
2703 	if (vp->v_vflag & VV_DELETED)
2704 		strlcat(buf, "|VV_DELETED", sizeof(buf));
2705 	if (vp->v_vflag & VV_MD)
2706 		strlcat(buf, "|VV_MD", sizeof(buf));
2707 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC |
2708 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
2709 	    VV_NOKNOTE | VV_DELETED | VV_MD);
2710 	if (flags != 0) {
2711 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
2712 		strlcat(buf, buf2, sizeof(buf));
2713 	}
2714 	if (vp->v_iflag & VI_MOUNT)
2715 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
2716 	if (vp->v_iflag & VI_AGE)
2717 		strlcat(buf, "|VI_AGE", sizeof(buf));
2718 	if (vp->v_iflag & VI_DOOMED)
2719 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
2720 	if (vp->v_iflag & VI_FREE)
2721 		strlcat(buf, "|VI_FREE", sizeof(buf));
2722 	if (vp->v_iflag & VI_DOINGINACT)
2723 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
2724 	if (vp->v_iflag & VI_OWEINACT)
2725 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
2726 	flags = vp->v_iflag & ~(VI_MOUNT | VI_AGE | VI_DOOMED | VI_FREE |
2727 	    VI_DOINGINACT | VI_OWEINACT);
2728 	if (flags != 0) {
2729 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
2730 		strlcat(buf, buf2, sizeof(buf));
2731 	}
2732 	printf("    flags (%s)\n", buf + 1);
2733 	if (mtx_owned(VI_MTX(vp)))
2734 		printf(" VI_LOCKed");
2735 	if (vp->v_object != NULL)
2736 		printf("    v_object %p ref %d pages %d\n",
2737 		    vp->v_object, vp->v_object->ref_count,
2738 		    vp->v_object->resident_page_count);
2739 	printf("    ");
2740 	lockmgr_printinfo(vp->v_vnlock);
2741 	if (vp->v_data != NULL)
2742 		VOP_PRINT(vp);
2743 }
2744 
2745 #ifdef DDB
2746 /*
2747  * List all of the locked vnodes in the system.
2748  * Called when debugging the kernel.
2749  */
2750 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2751 {
2752 	struct mount *mp, *nmp;
2753 	struct vnode *vp;
2754 
2755 	/*
2756 	 * Note: because this is DDB, we can't obey the locking semantics
2757 	 * for these structures, which means we could catch an inconsistent
2758 	 * state and dereference a nasty pointer.  Not much to be done
2759 	 * about that.
2760 	 */
2761 	db_printf("Locked vnodes\n");
2762 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2763 		nmp = TAILQ_NEXT(mp, mnt_list);
2764 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2765 			if (vp->v_type != VMARKER &&
2766 			    VOP_ISLOCKED(vp))
2767 				vprint("", vp);
2768 		}
2769 		nmp = TAILQ_NEXT(mp, mnt_list);
2770 	}
2771 }
2772 
2773 /*
2774  * Show details about the given vnode.
2775  */
2776 DB_SHOW_COMMAND(vnode, db_show_vnode)
2777 {
2778 	struct vnode *vp;
2779 
2780 	if (!have_addr)
2781 		return;
2782 	vp = (struct vnode *)addr;
2783 	vn_printf(vp, "vnode ");
2784 }
2785 
2786 /*
2787  * Show details about the given mount point.
2788  */
2789 DB_SHOW_COMMAND(mount, db_show_mount)
2790 {
2791 	struct mount *mp;
2792 	struct vfsopt *opt;
2793 	struct statfs *sp;
2794 	struct vnode *vp;
2795 	char buf[512];
2796 	u_int flags;
2797 
2798 	if (!have_addr) {
2799 		/* No address given, print short info about all mount points. */
2800 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2801 			db_printf("%p %s on %s (%s)\n", mp,
2802 			    mp->mnt_stat.f_mntfromname,
2803 			    mp->mnt_stat.f_mntonname,
2804 			    mp->mnt_stat.f_fstypename);
2805 			if (db_pager_quit)
2806 				break;
2807 		}
2808 		db_printf("\nMore info: show mount <addr>\n");
2809 		return;
2810 	}
2811 
2812 	mp = (struct mount *)addr;
2813 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
2814 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
2815 
2816 	buf[0] = '\0';
2817 	flags = mp->mnt_flag;
2818 #define	MNT_FLAG(flag)	do {						\
2819 	if (flags & (flag)) {						\
2820 		if (buf[0] != '\0')					\
2821 			strlcat(buf, ", ", sizeof(buf));		\
2822 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
2823 		flags &= ~(flag);					\
2824 	}								\
2825 } while (0)
2826 	MNT_FLAG(MNT_RDONLY);
2827 	MNT_FLAG(MNT_SYNCHRONOUS);
2828 	MNT_FLAG(MNT_NOEXEC);
2829 	MNT_FLAG(MNT_NOSUID);
2830 	MNT_FLAG(MNT_UNION);
2831 	MNT_FLAG(MNT_ASYNC);
2832 	MNT_FLAG(MNT_SUIDDIR);
2833 	MNT_FLAG(MNT_SOFTDEP);
2834 	MNT_FLAG(MNT_NOSYMFOLLOW);
2835 	MNT_FLAG(MNT_GJOURNAL);
2836 	MNT_FLAG(MNT_MULTILABEL);
2837 	MNT_FLAG(MNT_ACLS);
2838 	MNT_FLAG(MNT_NOATIME);
2839 	MNT_FLAG(MNT_NOCLUSTERR);
2840 	MNT_FLAG(MNT_NOCLUSTERW);
2841 	MNT_FLAG(MNT_NFS4ACLS);
2842 	MNT_FLAG(MNT_EXRDONLY);
2843 	MNT_FLAG(MNT_EXPORTED);
2844 	MNT_FLAG(MNT_DEFEXPORTED);
2845 	MNT_FLAG(MNT_EXPORTANON);
2846 	MNT_FLAG(MNT_EXKERB);
2847 	MNT_FLAG(MNT_EXPUBLIC);
2848 	MNT_FLAG(MNT_LOCAL);
2849 	MNT_FLAG(MNT_QUOTA);
2850 	MNT_FLAG(MNT_ROOTFS);
2851 	MNT_FLAG(MNT_USER);
2852 	MNT_FLAG(MNT_IGNORE);
2853 	MNT_FLAG(MNT_UPDATE);
2854 	MNT_FLAG(MNT_DELEXPORT);
2855 	MNT_FLAG(MNT_RELOAD);
2856 	MNT_FLAG(MNT_FORCE);
2857 	MNT_FLAG(MNT_SNAPSHOT);
2858 	MNT_FLAG(MNT_BYFSID);
2859 	MNT_FLAG(MNT_SOFTDEP);
2860 #undef MNT_FLAG
2861 	if (flags != 0) {
2862 		if (buf[0] != '\0')
2863 			strlcat(buf, ", ", sizeof(buf));
2864 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
2865 		    "0x%08x", flags);
2866 	}
2867 	db_printf("    mnt_flag = %s\n", buf);
2868 
2869 	buf[0] = '\0';
2870 	flags = mp->mnt_kern_flag;
2871 #define	MNT_KERN_FLAG(flag)	do {					\
2872 	if (flags & (flag)) {						\
2873 		if (buf[0] != '\0')					\
2874 			strlcat(buf, ", ", sizeof(buf));		\
2875 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
2876 		flags &= ~(flag);					\
2877 	}								\
2878 } while (0)
2879 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
2880 	MNT_KERN_FLAG(MNTK_ASYNC);
2881 	MNT_KERN_FLAG(MNTK_SOFTDEP);
2882 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
2883 	MNT_KERN_FLAG(MNTK_DRAINING);
2884 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
2885 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
2886 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
2887 	MNT_KERN_FLAG(MNTK_SUJ);
2888 	MNT_KERN_FLAG(MNTK_UNMOUNT);
2889 	MNT_KERN_FLAG(MNTK_MWAIT);
2890 	MNT_KERN_FLAG(MNTK_SUSPEND);
2891 	MNT_KERN_FLAG(MNTK_SUSPEND2);
2892 	MNT_KERN_FLAG(MNTK_SUSPENDED);
2893 	MNT_KERN_FLAG(MNTK_MPSAFE);
2894 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
2895 	MNT_KERN_FLAG(MNTK_NOKNOTE);
2896 #undef MNT_KERN_FLAG
2897 	if (flags != 0) {
2898 		if (buf[0] != '\0')
2899 			strlcat(buf, ", ", sizeof(buf));
2900 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
2901 		    "0x%08x", flags);
2902 	}
2903 	db_printf("    mnt_kern_flag = %s\n", buf);
2904 
2905 	db_printf("    mnt_opt = ");
2906 	opt = TAILQ_FIRST(mp->mnt_opt);
2907 	if (opt != NULL) {
2908 		db_printf("%s", opt->name);
2909 		opt = TAILQ_NEXT(opt, link);
2910 		while (opt != NULL) {
2911 			db_printf(", %s", opt->name);
2912 			opt = TAILQ_NEXT(opt, link);
2913 		}
2914 	}
2915 	db_printf("\n");
2916 
2917 	sp = &mp->mnt_stat;
2918 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
2919 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
2920 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
2921 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
2922 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
2923 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
2924 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
2925 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
2926 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
2927 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
2928 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
2929 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
2930 
2931 	db_printf("    mnt_cred = { uid=%u ruid=%u",
2932 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
2933 	if (jailed(mp->mnt_cred))
2934 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
2935 	db_printf(" }\n");
2936 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
2937 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
2938 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
2939 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
2940 	db_printf("    mnt_noasync = %u\n", mp->mnt_noasync);
2941 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
2942 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
2943 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
2944 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
2945 	db_printf("    mnt_secondary_accwrites = %d\n",
2946 	    mp->mnt_secondary_accwrites);
2947 	db_printf("    mnt_gjprovider = %s\n",
2948 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
2949 	db_printf("\n");
2950 
2951 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2952 		if (vp->v_type != VMARKER) {
2953 			vn_printf(vp, "vnode ");
2954 			if (db_pager_quit)
2955 				break;
2956 		}
2957 	}
2958 }
2959 #endif	/* DDB */
2960 
2961 /*
2962  * Fill in a struct xvfsconf based on a struct vfsconf.
2963  */
2964 static void
2965 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2966 {
2967 
2968 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2969 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2970 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2971 	xvfsp->vfc_flags = vfsp->vfc_flags;
2972 	/*
2973 	 * These are unused in userland, we keep them
2974 	 * to not break binary compatibility.
2975 	 */
2976 	xvfsp->vfc_vfsops = NULL;
2977 	xvfsp->vfc_next = NULL;
2978 }
2979 
2980 /*
2981  * Top level filesystem related information gathering.
2982  */
2983 static int
2984 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2985 {
2986 	struct vfsconf *vfsp;
2987 	struct xvfsconf xvfsp;
2988 	int error;
2989 
2990 	error = 0;
2991 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2992 		bzero(&xvfsp, sizeof(xvfsp));
2993 		vfsconf2x(vfsp, &xvfsp);
2994 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2995 		if (error)
2996 			break;
2997 	}
2998 	return (error);
2999 }
3000 
3001 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD,
3002     NULL, 0, sysctl_vfs_conflist,
3003     "S,xvfsconf", "List of all configured filesystems");
3004 
3005 #ifndef BURN_BRIDGES
3006 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3007 
3008 static int
3009 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3010 {
3011 	int *name = (int *)arg1 - 1;	/* XXX */
3012 	u_int namelen = arg2 + 1;	/* XXX */
3013 	struct vfsconf *vfsp;
3014 	struct xvfsconf xvfsp;
3015 
3016 	printf("WARNING: userland calling deprecated sysctl, "
3017 	    "please rebuild world\n");
3018 
3019 #if 1 || defined(COMPAT_PRELITE2)
3020 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3021 	if (namelen == 1)
3022 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3023 #endif
3024 
3025 	switch (name[1]) {
3026 	case VFS_MAXTYPENUM:
3027 		if (namelen != 2)
3028 			return (ENOTDIR);
3029 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3030 	case VFS_CONF:
3031 		if (namelen != 3)
3032 			return (ENOTDIR);	/* overloaded */
3033 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
3034 			if (vfsp->vfc_typenum == name[2])
3035 				break;
3036 		if (vfsp == NULL)
3037 			return (EOPNOTSUPP);
3038 		bzero(&xvfsp, sizeof(xvfsp));
3039 		vfsconf2x(vfsp, &xvfsp);
3040 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3041 	}
3042 	return (EOPNOTSUPP);
3043 }
3044 
3045 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
3046     vfs_sysctl, "Generic filesystem");
3047 
3048 #if 1 || defined(COMPAT_PRELITE2)
3049 
3050 static int
3051 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3052 {
3053 	int error;
3054 	struct vfsconf *vfsp;
3055 	struct ovfsconf ovfs;
3056 
3057 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3058 		bzero(&ovfs, sizeof(ovfs));
3059 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3060 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3061 		ovfs.vfc_index = vfsp->vfc_typenum;
3062 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3063 		ovfs.vfc_flags = vfsp->vfc_flags;
3064 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3065 		if (error)
3066 			return error;
3067 	}
3068 	return 0;
3069 }
3070 
3071 #endif /* 1 || COMPAT_PRELITE2 */
3072 #endif /* !BURN_BRIDGES */
3073 
3074 #define KINFO_VNODESLOP		10
3075 #ifdef notyet
3076 /*
3077  * Dump vnode list (via sysctl).
3078  */
3079 /* ARGSUSED */
3080 static int
3081 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3082 {
3083 	struct xvnode *xvn;
3084 	struct mount *mp;
3085 	struct vnode *vp;
3086 	int error, len, n;
3087 
3088 	/*
3089 	 * Stale numvnodes access is not fatal here.
3090 	 */
3091 	req->lock = 0;
3092 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3093 	if (!req->oldptr)
3094 		/* Make an estimate */
3095 		return (SYSCTL_OUT(req, 0, len));
3096 
3097 	error = sysctl_wire_old_buffer(req, 0);
3098 	if (error != 0)
3099 		return (error);
3100 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3101 	n = 0;
3102 	mtx_lock(&mountlist_mtx);
3103 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3104 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3105 			continue;
3106 		MNT_ILOCK(mp);
3107 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3108 			if (n == len)
3109 				break;
3110 			vref(vp);
3111 			xvn[n].xv_size = sizeof *xvn;
3112 			xvn[n].xv_vnode = vp;
3113 			xvn[n].xv_id = 0;	/* XXX compat */
3114 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3115 			XV_COPY(usecount);
3116 			XV_COPY(writecount);
3117 			XV_COPY(holdcnt);
3118 			XV_COPY(mount);
3119 			XV_COPY(numoutput);
3120 			XV_COPY(type);
3121 #undef XV_COPY
3122 			xvn[n].xv_flag = vp->v_vflag;
3123 
3124 			switch (vp->v_type) {
3125 			case VREG:
3126 			case VDIR:
3127 			case VLNK:
3128 				break;
3129 			case VBLK:
3130 			case VCHR:
3131 				if (vp->v_rdev == NULL) {
3132 					vrele(vp);
3133 					continue;
3134 				}
3135 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3136 				break;
3137 			case VSOCK:
3138 				xvn[n].xv_socket = vp->v_socket;
3139 				break;
3140 			case VFIFO:
3141 				xvn[n].xv_fifo = vp->v_fifoinfo;
3142 				break;
3143 			case VNON:
3144 			case VBAD:
3145 			default:
3146 				/* shouldn't happen? */
3147 				vrele(vp);
3148 				continue;
3149 			}
3150 			vrele(vp);
3151 			++n;
3152 		}
3153 		MNT_IUNLOCK(mp);
3154 		mtx_lock(&mountlist_mtx);
3155 		vfs_unbusy(mp);
3156 		if (n == len)
3157 			break;
3158 	}
3159 	mtx_unlock(&mountlist_mtx);
3160 
3161 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3162 	free(xvn, M_TEMP);
3163 	return (error);
3164 }
3165 
3166 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
3167     0, 0, sysctl_vnode, "S,xvnode", "");
3168 #endif
3169 
3170 /*
3171  * Unmount all filesystems. The list is traversed in reverse order
3172  * of mounting to avoid dependencies.
3173  */
3174 void
3175 vfs_unmountall(void)
3176 {
3177 	struct mount *mp;
3178 	struct thread *td;
3179 	int error;
3180 
3181 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
3182 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3183 	td = curthread;
3184 
3185 	/*
3186 	 * Since this only runs when rebooting, it is not interlocked.
3187 	 */
3188 	while(!TAILQ_EMPTY(&mountlist)) {
3189 		mp = TAILQ_LAST(&mountlist, mntlist);
3190 		error = dounmount(mp, MNT_FORCE, td);
3191 		if (error) {
3192 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
3193 			/*
3194 			 * XXX: Due to the way in which we mount the root
3195 			 * file system off of devfs, devfs will generate a
3196 			 * "busy" warning when we try to unmount it before
3197 			 * the root.  Don't print a warning as a result in
3198 			 * order to avoid false positive errors that may
3199 			 * cause needless upset.
3200 			 */
3201 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
3202 				printf("unmount of %s failed (",
3203 				    mp->mnt_stat.f_mntonname);
3204 				if (error == EBUSY)
3205 					printf("BUSY)\n");
3206 				else
3207 					printf("%d)\n", error);
3208 			}
3209 		} else {
3210 			/* The unmount has removed mp from the mountlist */
3211 		}
3212 	}
3213 }
3214 
3215 /*
3216  * perform msync on all vnodes under a mount point
3217  * the mount point must be locked.
3218  */
3219 void
3220 vfs_msync(struct mount *mp, int flags)
3221 {
3222 	struct vnode *vp, *mvp;
3223 	struct vm_object *obj;
3224 
3225 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3226 	MNT_ILOCK(mp);
3227 	MNT_VNODE_FOREACH(vp, mp, mvp) {
3228 		VI_LOCK(vp);
3229 		obj = vp->v_object;
3230 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3231 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3232 			MNT_IUNLOCK(mp);
3233 			if (!vget(vp,
3234 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3235 			    curthread)) {
3236 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3237 					vput(vp);
3238 					MNT_ILOCK(mp);
3239 					continue;
3240 				}
3241 
3242 				obj = vp->v_object;
3243 				if (obj != NULL) {
3244 					VM_OBJECT_LOCK(obj);
3245 					vm_object_page_clean(obj, 0, 0,
3246 					    flags == MNT_WAIT ?
3247 					    OBJPC_SYNC : OBJPC_NOSYNC);
3248 					VM_OBJECT_UNLOCK(obj);
3249 				}
3250 				vput(vp);
3251 			}
3252 			MNT_ILOCK(mp);
3253 		} else
3254 			VI_UNLOCK(vp);
3255 	}
3256 	MNT_IUNLOCK(mp);
3257 }
3258 
3259 /*
3260  * Mark a vnode as free, putting it up for recycling.
3261  */
3262 static void
3263 vfree(struct vnode *vp)
3264 {
3265 
3266 	ASSERT_VI_LOCKED(vp, "vfree");
3267 	mtx_lock(&vnode_free_list_mtx);
3268 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
3269 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
3270 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
3271 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
3272 	    ("vfree: Freeing doomed vnode"));
3273 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3274 	if (vp->v_iflag & VI_AGE) {
3275 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
3276 	} else {
3277 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
3278 	}
3279 	freevnodes++;
3280 	vp->v_iflag &= ~VI_AGE;
3281 	vp->v_iflag |= VI_FREE;
3282 	mtx_unlock(&vnode_free_list_mtx);
3283 }
3284 
3285 /*
3286  * Opposite of vfree() - mark a vnode as in use.
3287  */
3288 static void
3289 vbusy(struct vnode *vp)
3290 {
3291 	ASSERT_VI_LOCKED(vp, "vbusy");
3292 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
3293 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
3294 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3295 
3296 	mtx_lock(&vnode_free_list_mtx);
3297 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
3298 	freevnodes--;
3299 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
3300 	mtx_unlock(&vnode_free_list_mtx);
3301 }
3302 
3303 static void
3304 destroy_vpollinfo(struct vpollinfo *vi)
3305 {
3306 	knlist_destroy(&vi->vpi_selinfo.si_note);
3307 	mtx_destroy(&vi->vpi_lock);
3308 	uma_zfree(vnodepoll_zone, vi);
3309 }
3310 
3311 /*
3312  * Initalize per-vnode helper structure to hold poll-related state.
3313  */
3314 void
3315 v_addpollinfo(struct vnode *vp)
3316 {
3317 	struct vpollinfo *vi;
3318 
3319 	if (vp->v_pollinfo != NULL)
3320 		return;
3321 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
3322 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3323 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3324 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3325 	VI_LOCK(vp);
3326 	if (vp->v_pollinfo != NULL) {
3327 		VI_UNLOCK(vp);
3328 		destroy_vpollinfo(vi);
3329 		return;
3330 	}
3331 	vp->v_pollinfo = vi;
3332 	VI_UNLOCK(vp);
3333 }
3334 
3335 /*
3336  * Record a process's interest in events which might happen to
3337  * a vnode.  Because poll uses the historic select-style interface
3338  * internally, this routine serves as both the ``check for any
3339  * pending events'' and the ``record my interest in future events''
3340  * functions.  (These are done together, while the lock is held,
3341  * to avoid race conditions.)
3342  */
3343 int
3344 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3345 {
3346 
3347 	v_addpollinfo(vp);
3348 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3349 	if (vp->v_pollinfo->vpi_revents & events) {
3350 		/*
3351 		 * This leaves events we are not interested
3352 		 * in available for the other process which
3353 		 * which presumably had requested them
3354 		 * (otherwise they would never have been
3355 		 * recorded).
3356 		 */
3357 		events &= vp->v_pollinfo->vpi_revents;
3358 		vp->v_pollinfo->vpi_revents &= ~events;
3359 
3360 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3361 		return (events);
3362 	}
3363 	vp->v_pollinfo->vpi_events |= events;
3364 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3365 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3366 	return (0);
3367 }
3368 
3369 /*
3370  * Routine to create and manage a filesystem syncer vnode.
3371  */
3372 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3373 static int	sync_fsync(struct  vop_fsync_args *);
3374 static int	sync_inactive(struct  vop_inactive_args *);
3375 static int	sync_reclaim(struct  vop_reclaim_args *);
3376 
3377 static struct vop_vector sync_vnodeops = {
3378 	.vop_bypass =	VOP_EOPNOTSUPP,
3379 	.vop_close =	sync_close,		/* close */
3380 	.vop_fsync =	sync_fsync,		/* fsync */
3381 	.vop_inactive =	sync_inactive,	/* inactive */
3382 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3383 	.vop_lock1 =	vop_stdlock,	/* lock */
3384 	.vop_unlock =	vop_stdunlock,	/* unlock */
3385 	.vop_islocked =	vop_stdislocked,	/* islocked */
3386 };
3387 
3388 /*
3389  * Create a new filesystem syncer vnode for the specified mount point.
3390  */
3391 void
3392 vfs_allocate_syncvnode(struct mount *mp)
3393 {
3394 	struct vnode *vp;
3395 	struct bufobj *bo;
3396 	static long start, incr, next;
3397 	int error;
3398 
3399 	/* Allocate a new vnode */
3400 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
3401 	if (error != 0)
3402 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
3403 	vp->v_type = VNON;
3404 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3405 	vp->v_vflag |= VV_FORCEINSMQ;
3406 	error = insmntque(vp, mp);
3407 	if (error != 0)
3408 		panic("vfs_allocate_syncvnode: insmntque() failed");
3409 	vp->v_vflag &= ~VV_FORCEINSMQ;
3410 	VOP_UNLOCK(vp, 0);
3411 	/*
3412 	 * Place the vnode onto the syncer worklist. We attempt to
3413 	 * scatter them about on the list so that they will go off
3414 	 * at evenly distributed times even if all the filesystems
3415 	 * are mounted at once.
3416 	 */
3417 	next += incr;
3418 	if (next == 0 || next > syncer_maxdelay) {
3419 		start /= 2;
3420 		incr /= 2;
3421 		if (start == 0) {
3422 			start = syncer_maxdelay / 2;
3423 			incr = syncer_maxdelay;
3424 		}
3425 		next = start;
3426 	}
3427 	bo = &vp->v_bufobj;
3428 	BO_LOCK(bo);
3429 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
3430 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3431 	mtx_lock(&sync_mtx);
3432 	sync_vnode_count++;
3433 	if (mp->mnt_syncer == NULL) {
3434 		mp->mnt_syncer = vp;
3435 		vp = NULL;
3436 	}
3437 	mtx_unlock(&sync_mtx);
3438 	BO_UNLOCK(bo);
3439 	if (vp != NULL) {
3440 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3441 		vgone(vp);
3442 		vput(vp);
3443 	}
3444 }
3445 
3446 void
3447 vfs_deallocate_syncvnode(struct mount *mp)
3448 {
3449 	struct vnode *vp;
3450 
3451 	mtx_lock(&sync_mtx);
3452 	vp = mp->mnt_syncer;
3453 	if (vp != NULL)
3454 		mp->mnt_syncer = NULL;
3455 	mtx_unlock(&sync_mtx);
3456 	if (vp != NULL)
3457 		vrele(vp);
3458 }
3459 
3460 /*
3461  * Do a lazy sync of the filesystem.
3462  */
3463 static int
3464 sync_fsync(struct vop_fsync_args *ap)
3465 {
3466 	struct vnode *syncvp = ap->a_vp;
3467 	struct mount *mp = syncvp->v_mount;
3468 	int error;
3469 	struct bufobj *bo;
3470 
3471 	/*
3472 	 * We only need to do something if this is a lazy evaluation.
3473 	 */
3474 	if (ap->a_waitfor != MNT_LAZY)
3475 		return (0);
3476 
3477 	/*
3478 	 * Move ourselves to the back of the sync list.
3479 	 */
3480 	bo = &syncvp->v_bufobj;
3481 	BO_LOCK(bo);
3482 	vn_syncer_add_to_worklist(bo, syncdelay);
3483 	BO_UNLOCK(bo);
3484 
3485 	/*
3486 	 * Walk the list of vnodes pushing all that are dirty and
3487 	 * not already on the sync list.
3488 	 */
3489 	mtx_lock(&mountlist_mtx);
3490 	if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) != 0) {
3491 		mtx_unlock(&mountlist_mtx);
3492 		return (0);
3493 	}
3494 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3495 		vfs_unbusy(mp);
3496 		return (0);
3497 	}
3498 	MNT_ILOCK(mp);
3499 	mp->mnt_noasync++;
3500 	mp->mnt_kern_flag &= ~MNTK_ASYNC;
3501 	MNT_IUNLOCK(mp);
3502 	vfs_msync(mp, MNT_NOWAIT);
3503 	error = VFS_SYNC(mp, MNT_LAZY);
3504 	MNT_ILOCK(mp);
3505 	mp->mnt_noasync--;
3506 	if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0)
3507 		mp->mnt_kern_flag |= MNTK_ASYNC;
3508 	MNT_IUNLOCK(mp);
3509 	vn_finished_write(mp);
3510 	vfs_unbusy(mp);
3511 	return (error);
3512 }
3513 
3514 /*
3515  * The syncer vnode is no referenced.
3516  */
3517 static int
3518 sync_inactive(struct vop_inactive_args *ap)
3519 {
3520 
3521 	vgone(ap->a_vp);
3522 	return (0);
3523 }
3524 
3525 /*
3526  * The syncer vnode is no longer needed and is being decommissioned.
3527  *
3528  * Modifications to the worklist must be protected by sync_mtx.
3529  */
3530 static int
3531 sync_reclaim(struct vop_reclaim_args *ap)
3532 {
3533 	struct vnode *vp = ap->a_vp;
3534 	struct bufobj *bo;
3535 
3536 	bo = &vp->v_bufobj;
3537 	BO_LOCK(bo);
3538 	mtx_lock(&sync_mtx);
3539 	if (vp->v_mount->mnt_syncer == vp)
3540 		vp->v_mount->mnt_syncer = NULL;
3541 	if (bo->bo_flag & BO_ONWORKLST) {
3542 		LIST_REMOVE(bo, bo_synclist);
3543 		syncer_worklist_len--;
3544 		sync_vnode_count--;
3545 		bo->bo_flag &= ~BO_ONWORKLST;
3546 	}
3547 	mtx_unlock(&sync_mtx);
3548 	BO_UNLOCK(bo);
3549 
3550 	return (0);
3551 }
3552 
3553 /*
3554  * Check if vnode represents a disk device
3555  */
3556 int
3557 vn_isdisk(struct vnode *vp, int *errp)
3558 {
3559 	int error;
3560 
3561 	error = 0;
3562 	dev_lock();
3563 	if (vp->v_type != VCHR)
3564 		error = ENOTBLK;
3565 	else if (vp->v_rdev == NULL)
3566 		error = ENXIO;
3567 	else if (vp->v_rdev->si_devsw == NULL)
3568 		error = ENXIO;
3569 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3570 		error = ENOTBLK;
3571 	dev_unlock();
3572 	if (errp != NULL)
3573 		*errp = error;
3574 	return (error == 0);
3575 }
3576 
3577 /*
3578  * Common filesystem object access control check routine.  Accepts a
3579  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3580  * and optional call-by-reference privused argument allowing vaccess()
3581  * to indicate to the caller whether privilege was used to satisfy the
3582  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3583  *
3584  * The ifdef'd CAPABILITIES version is here for reference, but is not
3585  * actually used.
3586  */
3587 int
3588 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3589     accmode_t accmode, struct ucred *cred, int *privused)
3590 {
3591 	accmode_t dac_granted;
3592 	accmode_t priv_granted;
3593 
3594 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
3595 	    ("invalid bit in accmode"));
3596 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
3597 	    ("VAPPEND without VWRITE"));
3598 
3599 	/*
3600 	 * Look for a normal, non-privileged way to access the file/directory
3601 	 * as requested.  If it exists, go with that.
3602 	 */
3603 
3604 	if (privused != NULL)
3605 		*privused = 0;
3606 
3607 	dac_granted = 0;
3608 
3609 	/* Check the owner. */
3610 	if (cred->cr_uid == file_uid) {
3611 		dac_granted |= VADMIN;
3612 		if (file_mode & S_IXUSR)
3613 			dac_granted |= VEXEC;
3614 		if (file_mode & S_IRUSR)
3615 			dac_granted |= VREAD;
3616 		if (file_mode & S_IWUSR)
3617 			dac_granted |= (VWRITE | VAPPEND);
3618 
3619 		if ((accmode & dac_granted) == accmode)
3620 			return (0);
3621 
3622 		goto privcheck;
3623 	}
3624 
3625 	/* Otherwise, check the groups (first match) */
3626 	if (groupmember(file_gid, cred)) {
3627 		if (file_mode & S_IXGRP)
3628 			dac_granted |= VEXEC;
3629 		if (file_mode & S_IRGRP)
3630 			dac_granted |= VREAD;
3631 		if (file_mode & S_IWGRP)
3632 			dac_granted |= (VWRITE | VAPPEND);
3633 
3634 		if ((accmode & dac_granted) == accmode)
3635 			return (0);
3636 
3637 		goto privcheck;
3638 	}
3639 
3640 	/* Otherwise, check everyone else. */
3641 	if (file_mode & S_IXOTH)
3642 		dac_granted |= VEXEC;
3643 	if (file_mode & S_IROTH)
3644 		dac_granted |= VREAD;
3645 	if (file_mode & S_IWOTH)
3646 		dac_granted |= (VWRITE | VAPPEND);
3647 	if ((accmode & dac_granted) == accmode)
3648 		return (0);
3649 
3650 privcheck:
3651 	/*
3652 	 * Build a privilege mask to determine if the set of privileges
3653 	 * satisfies the requirements when combined with the granted mask
3654 	 * from above.  For each privilege, if the privilege is required,
3655 	 * bitwise or the request type onto the priv_granted mask.
3656 	 */
3657 	priv_granted = 0;
3658 
3659 	if (type == VDIR) {
3660 		/*
3661 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
3662 		 * requests, instead of PRIV_VFS_EXEC.
3663 		 */
3664 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3665 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
3666 			priv_granted |= VEXEC;
3667 	} else {
3668 		/*
3669 		 * Ensure that at least one execute bit is on. Otherwise,
3670 		 * a privileged user will always succeed, and we don't want
3671 		 * this to happen unless the file really is executable.
3672 		 */
3673 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3674 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
3675 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
3676 			priv_granted |= VEXEC;
3677 	}
3678 
3679 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
3680 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
3681 		priv_granted |= VREAD;
3682 
3683 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3684 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
3685 		priv_granted |= (VWRITE | VAPPEND);
3686 
3687 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3688 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
3689 		priv_granted |= VADMIN;
3690 
3691 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
3692 		/* XXX audit: privilege used */
3693 		if (privused != NULL)
3694 			*privused = 1;
3695 		return (0);
3696 	}
3697 
3698 	return ((accmode & VADMIN) ? EPERM : EACCES);
3699 }
3700 
3701 /*
3702  * Credential check based on process requesting service, and per-attribute
3703  * permissions.
3704  */
3705 int
3706 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3707     struct thread *td, accmode_t accmode)
3708 {
3709 
3710 	/*
3711 	 * Kernel-invoked always succeeds.
3712 	 */
3713 	if (cred == NOCRED)
3714 		return (0);
3715 
3716 	/*
3717 	 * Do not allow privileged processes in jail to directly manipulate
3718 	 * system attributes.
3719 	 */
3720 	switch (attrnamespace) {
3721 	case EXTATTR_NAMESPACE_SYSTEM:
3722 		/* Potentially should be: return (EPERM); */
3723 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
3724 	case EXTATTR_NAMESPACE_USER:
3725 		return (VOP_ACCESS(vp, accmode, cred, td));
3726 	default:
3727 		return (EPERM);
3728 	}
3729 }
3730 
3731 #ifdef DEBUG_VFS_LOCKS
3732 /*
3733  * This only exists to supress warnings from unlocked specfs accesses.  It is
3734  * no longer ok to have an unlocked VFS.
3735  */
3736 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
3737 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
3738 
3739 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3740 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
3741     "Drop into debugger on lock violation");
3742 
3743 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3744 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
3745     0, "Check for interlock across VOPs");
3746 
3747 int vfs_badlock_print = 1;	/* Print lock violations. */
3748 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
3749     0, "Print lock violations");
3750 
3751 #ifdef KDB
3752 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3753 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
3754     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
3755 #endif
3756 
3757 static void
3758 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3759 {
3760 
3761 #ifdef KDB
3762 	if (vfs_badlock_backtrace)
3763 		kdb_backtrace();
3764 #endif
3765 	if (vfs_badlock_print)
3766 		printf("%s: %p %s\n", str, (void *)vp, msg);
3767 	if (vfs_badlock_ddb)
3768 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
3769 }
3770 
3771 void
3772 assert_vi_locked(struct vnode *vp, const char *str)
3773 {
3774 
3775 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3776 		vfs_badlock("interlock is not locked but should be", str, vp);
3777 }
3778 
3779 void
3780 assert_vi_unlocked(struct vnode *vp, const char *str)
3781 {
3782 
3783 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3784 		vfs_badlock("interlock is locked but should not be", str, vp);
3785 }
3786 
3787 void
3788 assert_vop_locked(struct vnode *vp, const char *str)
3789 {
3790 
3791 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == 0)
3792 		vfs_badlock("is not locked but should be", str, vp);
3793 }
3794 
3795 void
3796 assert_vop_unlocked(struct vnode *vp, const char *str)
3797 {
3798 
3799 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
3800 		vfs_badlock("is locked but should not be", str, vp);
3801 }
3802 
3803 void
3804 assert_vop_elocked(struct vnode *vp, const char *str)
3805 {
3806 
3807 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
3808 		vfs_badlock("is not exclusive locked but should be", str, vp);
3809 }
3810 
3811 #if 0
3812 void
3813 assert_vop_elocked_other(struct vnode *vp, const char *str)
3814 {
3815 
3816 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
3817 		vfs_badlock("is not exclusive locked by another thread",
3818 		    str, vp);
3819 }
3820 
3821 void
3822 assert_vop_slocked(struct vnode *vp, const char *str)
3823 {
3824 
3825 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
3826 		vfs_badlock("is not locked shared but should be", str, vp);
3827 }
3828 #endif /* 0 */
3829 #endif /* DEBUG_VFS_LOCKS */
3830 
3831 void
3832 vop_rename_fail(struct vop_rename_args *ap)
3833 {
3834 
3835 	if (ap->a_tvp != NULL)
3836 		vput(ap->a_tvp);
3837 	if (ap->a_tdvp == ap->a_tvp)
3838 		vrele(ap->a_tdvp);
3839 	else
3840 		vput(ap->a_tdvp);
3841 	vrele(ap->a_fdvp);
3842 	vrele(ap->a_fvp);
3843 }
3844 
3845 void
3846 vop_rename_pre(void *ap)
3847 {
3848 	struct vop_rename_args *a = ap;
3849 
3850 #ifdef DEBUG_VFS_LOCKS
3851 	if (a->a_tvp)
3852 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3853 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3854 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3855 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3856 
3857 	/* Check the source (from). */
3858 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
3859 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
3860 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3861 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
3862 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
3863 
3864 	/* Check the target. */
3865 	if (a->a_tvp)
3866 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3867 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3868 #endif
3869 	if (a->a_tdvp != a->a_fdvp)
3870 		vhold(a->a_fdvp);
3871 	if (a->a_tvp != a->a_fvp)
3872 		vhold(a->a_fvp);
3873 	vhold(a->a_tdvp);
3874 	if (a->a_tvp)
3875 		vhold(a->a_tvp);
3876 }
3877 
3878 void
3879 vop_strategy_pre(void *ap)
3880 {
3881 #ifdef DEBUG_VFS_LOCKS
3882 	struct vop_strategy_args *a;
3883 	struct buf *bp;
3884 
3885 	a = ap;
3886 	bp = a->a_bp;
3887 
3888 	/*
3889 	 * Cluster ops lock their component buffers but not the IO container.
3890 	 */
3891 	if ((bp->b_flags & B_CLUSTER) != 0)
3892 		return;
3893 
3894 	if (!BUF_ISLOCKED(bp)) {
3895 		if (vfs_badlock_print)
3896 			printf(
3897 			    "VOP_STRATEGY: bp is not locked but should be\n");
3898 		if (vfs_badlock_ddb)
3899 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
3900 	}
3901 #endif
3902 }
3903 
3904 void
3905 vop_lookup_pre(void *ap)
3906 {
3907 #ifdef DEBUG_VFS_LOCKS
3908 	struct vop_lookup_args *a;
3909 	struct vnode *dvp;
3910 
3911 	a = ap;
3912 	dvp = a->a_dvp;
3913 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3914 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3915 #endif
3916 }
3917 
3918 void
3919 vop_lookup_post(void *ap, int rc)
3920 {
3921 #ifdef DEBUG_VFS_LOCKS
3922 	struct vop_lookup_args *a;
3923 	struct vnode *dvp;
3924 	struct vnode *vp;
3925 
3926 	a = ap;
3927 	dvp = a->a_dvp;
3928 	vp = *(a->a_vpp);
3929 
3930 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3931 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3932 
3933 	if (!rc)
3934 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3935 #endif
3936 }
3937 
3938 void
3939 vop_lock_pre(void *ap)
3940 {
3941 #ifdef DEBUG_VFS_LOCKS
3942 	struct vop_lock1_args *a = ap;
3943 
3944 	if ((a->a_flags & LK_INTERLOCK) == 0)
3945 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3946 	else
3947 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3948 #endif
3949 }
3950 
3951 void
3952 vop_lock_post(void *ap, int rc)
3953 {
3954 #ifdef DEBUG_VFS_LOCKS
3955 	struct vop_lock1_args *a = ap;
3956 
3957 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3958 	if (rc == 0)
3959 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3960 #endif
3961 }
3962 
3963 void
3964 vop_unlock_pre(void *ap)
3965 {
3966 #ifdef DEBUG_VFS_LOCKS
3967 	struct vop_unlock_args *a = ap;
3968 
3969 	if (a->a_flags & LK_INTERLOCK)
3970 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3971 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3972 #endif
3973 }
3974 
3975 void
3976 vop_unlock_post(void *ap, int rc)
3977 {
3978 #ifdef DEBUG_VFS_LOCKS
3979 	struct vop_unlock_args *a = ap;
3980 
3981 	if (a->a_flags & LK_INTERLOCK)
3982 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3983 #endif
3984 }
3985 
3986 void
3987 vop_create_post(void *ap, int rc)
3988 {
3989 	struct vop_create_args *a = ap;
3990 
3991 	if (!rc)
3992 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3993 }
3994 
3995 void
3996 vop_link_post(void *ap, int rc)
3997 {
3998 	struct vop_link_args *a = ap;
3999 
4000 	if (!rc) {
4001 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4002 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4003 	}
4004 }
4005 
4006 void
4007 vop_mkdir_post(void *ap, int rc)
4008 {
4009 	struct vop_mkdir_args *a = ap;
4010 
4011 	if (!rc)
4012 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4013 }
4014 
4015 void
4016 vop_mknod_post(void *ap, int rc)
4017 {
4018 	struct vop_mknod_args *a = ap;
4019 
4020 	if (!rc)
4021 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4022 }
4023 
4024 void
4025 vop_remove_post(void *ap, int rc)
4026 {
4027 	struct vop_remove_args *a = ap;
4028 
4029 	if (!rc) {
4030 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4031 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4032 	}
4033 }
4034 
4035 void
4036 vop_rename_post(void *ap, int rc)
4037 {
4038 	struct vop_rename_args *a = ap;
4039 
4040 	if (!rc) {
4041 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
4042 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
4043 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4044 		if (a->a_tvp)
4045 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4046 	}
4047 	if (a->a_tdvp != a->a_fdvp)
4048 		vdrop(a->a_fdvp);
4049 	if (a->a_tvp != a->a_fvp)
4050 		vdrop(a->a_fvp);
4051 	vdrop(a->a_tdvp);
4052 	if (a->a_tvp)
4053 		vdrop(a->a_tvp);
4054 }
4055 
4056 void
4057 vop_rmdir_post(void *ap, int rc)
4058 {
4059 	struct vop_rmdir_args *a = ap;
4060 
4061 	if (!rc) {
4062 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4063 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4064 	}
4065 }
4066 
4067 void
4068 vop_setattr_post(void *ap, int rc)
4069 {
4070 	struct vop_setattr_args *a = ap;
4071 
4072 	if (!rc)
4073 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4074 }
4075 
4076 void
4077 vop_symlink_post(void *ap, int rc)
4078 {
4079 	struct vop_symlink_args *a = ap;
4080 
4081 	if (!rc)
4082 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4083 }
4084 
4085 static struct knlist fs_knlist;
4086 
4087 static void
4088 vfs_event_init(void *arg)
4089 {
4090 	knlist_init_mtx(&fs_knlist, NULL);
4091 }
4092 /* XXX - correct order? */
4093 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4094 
4095 void
4096 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4097 {
4098 
4099 	KNOTE_UNLOCKED(&fs_knlist, event);
4100 }
4101 
4102 static int	filt_fsattach(struct knote *kn);
4103 static void	filt_fsdetach(struct knote *kn);
4104 static int	filt_fsevent(struct knote *kn, long hint);
4105 
4106 struct filterops fs_filtops = {
4107 	.f_isfd = 0,
4108 	.f_attach = filt_fsattach,
4109 	.f_detach = filt_fsdetach,
4110 	.f_event = filt_fsevent
4111 };
4112 
4113 static int
4114 filt_fsattach(struct knote *kn)
4115 {
4116 
4117 	kn->kn_flags |= EV_CLEAR;
4118 	knlist_add(&fs_knlist, kn, 0);
4119 	return (0);
4120 }
4121 
4122 static void
4123 filt_fsdetach(struct knote *kn)
4124 {
4125 
4126 	knlist_remove(&fs_knlist, kn, 0);
4127 }
4128 
4129 static int
4130 filt_fsevent(struct knote *kn, long hint)
4131 {
4132 
4133 	kn->kn_fflags |= hint;
4134 	return (kn->kn_fflags != 0);
4135 }
4136 
4137 static int
4138 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4139 {
4140 	struct vfsidctl vc;
4141 	int error;
4142 	struct mount *mp;
4143 
4144 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4145 	if (error)
4146 		return (error);
4147 	if (vc.vc_vers != VFS_CTL_VERS1)
4148 		return (EINVAL);
4149 	mp = vfs_getvfs(&vc.vc_fsid);
4150 	if (mp == NULL)
4151 		return (ENOENT);
4152 	/* ensure that a specific sysctl goes to the right filesystem. */
4153 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4154 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4155 		vfs_rel(mp);
4156 		return (EINVAL);
4157 	}
4158 	VCTLTOREQ(&vc, req);
4159 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4160 	vfs_rel(mp);
4161 	return (error);
4162 }
4163 
4164 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4165     NULL, 0, sysctl_vfs_ctl, "",
4166     "Sysctl by fsid");
4167 
4168 /*
4169  * Function to initialize a va_filerev field sensibly.
4170  * XXX: Wouldn't a random number make a lot more sense ??
4171  */
4172 u_quad_t
4173 init_va_filerev(void)
4174 {
4175 	struct bintime bt;
4176 
4177 	getbinuptime(&bt);
4178 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4179 }
4180 
4181 static int	filt_vfsread(struct knote *kn, long hint);
4182 static int	filt_vfswrite(struct knote *kn, long hint);
4183 static int	filt_vfsvnode(struct knote *kn, long hint);
4184 static void	filt_vfsdetach(struct knote *kn);
4185 static struct filterops vfsread_filtops = {
4186 	.f_isfd = 1,
4187 	.f_detach = filt_vfsdetach,
4188 	.f_event = filt_vfsread
4189 };
4190 static struct filterops vfswrite_filtops = {
4191 	.f_isfd = 1,
4192 	.f_detach = filt_vfsdetach,
4193 	.f_event = filt_vfswrite
4194 };
4195 static struct filterops vfsvnode_filtops = {
4196 	.f_isfd = 1,
4197 	.f_detach = filt_vfsdetach,
4198 	.f_event = filt_vfsvnode
4199 };
4200 
4201 static void
4202 vfs_knllock(void *arg)
4203 {
4204 	struct vnode *vp = arg;
4205 
4206 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4207 }
4208 
4209 static void
4210 vfs_knlunlock(void *arg)
4211 {
4212 	struct vnode *vp = arg;
4213 
4214 	VOP_UNLOCK(vp, 0);
4215 }
4216 
4217 static void
4218 vfs_knl_assert_locked(void *arg)
4219 {
4220 #ifdef DEBUG_VFS_LOCKS
4221 	struct vnode *vp = arg;
4222 
4223 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4224 #endif
4225 }
4226 
4227 static void
4228 vfs_knl_assert_unlocked(void *arg)
4229 {
4230 #ifdef DEBUG_VFS_LOCKS
4231 	struct vnode *vp = arg;
4232 
4233 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4234 #endif
4235 }
4236 
4237 int
4238 vfs_kqfilter(struct vop_kqfilter_args *ap)
4239 {
4240 	struct vnode *vp = ap->a_vp;
4241 	struct knote *kn = ap->a_kn;
4242 	struct knlist *knl;
4243 
4244 	switch (kn->kn_filter) {
4245 	case EVFILT_READ:
4246 		kn->kn_fop = &vfsread_filtops;
4247 		break;
4248 	case EVFILT_WRITE:
4249 		kn->kn_fop = &vfswrite_filtops;
4250 		break;
4251 	case EVFILT_VNODE:
4252 		kn->kn_fop = &vfsvnode_filtops;
4253 		break;
4254 	default:
4255 		return (EINVAL);
4256 	}
4257 
4258 	kn->kn_hook = (caddr_t)vp;
4259 
4260 	v_addpollinfo(vp);
4261 	if (vp->v_pollinfo == NULL)
4262 		return (ENOMEM);
4263 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4264 	knlist_add(knl, kn, 0);
4265 
4266 	return (0);
4267 }
4268 
4269 /*
4270  * Detach knote from vnode
4271  */
4272 static void
4273 filt_vfsdetach(struct knote *kn)
4274 {
4275 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4276 
4277 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4278 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4279 }
4280 
4281 /*ARGSUSED*/
4282 static int
4283 filt_vfsread(struct knote *kn, long hint)
4284 {
4285 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4286 	struct vattr va;
4287 	int res;
4288 
4289 	/*
4290 	 * filesystem is gone, so set the EOF flag and schedule
4291 	 * the knote for deletion.
4292 	 */
4293 	if (hint == NOTE_REVOKE) {
4294 		VI_LOCK(vp);
4295 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4296 		VI_UNLOCK(vp);
4297 		return (1);
4298 	}
4299 
4300 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4301 		return (0);
4302 
4303 	VI_LOCK(vp);
4304 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4305 	res = (kn->kn_data != 0);
4306 	VI_UNLOCK(vp);
4307 	return (res);
4308 }
4309 
4310 /*ARGSUSED*/
4311 static int
4312 filt_vfswrite(struct knote *kn, long hint)
4313 {
4314 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4315 
4316 	VI_LOCK(vp);
4317 
4318 	/*
4319 	 * filesystem is gone, so set the EOF flag and schedule
4320 	 * the knote for deletion.
4321 	 */
4322 	if (hint == NOTE_REVOKE)
4323 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4324 
4325 	kn->kn_data = 0;
4326 	VI_UNLOCK(vp);
4327 	return (1);
4328 }
4329 
4330 static int
4331 filt_vfsvnode(struct knote *kn, long hint)
4332 {
4333 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4334 	int res;
4335 
4336 	VI_LOCK(vp);
4337 	if (kn->kn_sfflags & hint)
4338 		kn->kn_fflags |= hint;
4339 	if (hint == NOTE_REVOKE) {
4340 		kn->kn_flags |= EV_EOF;
4341 		VI_UNLOCK(vp);
4342 		return (1);
4343 	}
4344 	res = (kn->kn_fflags != 0);
4345 	VI_UNLOCK(vp);
4346 	return (res);
4347 }
4348 
4349 int
4350 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
4351 {
4352 	int error;
4353 
4354 	if (dp->d_reclen > ap->a_uio->uio_resid)
4355 		return (ENAMETOOLONG);
4356 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
4357 	if (error) {
4358 		if (ap->a_ncookies != NULL) {
4359 			if (ap->a_cookies != NULL)
4360 				free(ap->a_cookies, M_TEMP);
4361 			ap->a_cookies = NULL;
4362 			*ap->a_ncookies = 0;
4363 		}
4364 		return (error);
4365 	}
4366 	if (ap->a_ncookies == NULL)
4367 		return (0);
4368 
4369 	KASSERT(ap->a_cookies,
4370 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
4371 
4372 	*ap->a_cookies = realloc(*ap->a_cookies,
4373 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
4374 	(*ap->a_cookies)[*ap->a_ncookies] = off;
4375 	return (0);
4376 }
4377 
4378 /*
4379  * Mark for update the access time of the file if the filesystem
4380  * supports VOP_MARKATIME.  This functionality is used by execve and
4381  * mmap, so we want to avoid the I/O implied by directly setting
4382  * va_atime for the sake of efficiency.
4383  */
4384 void
4385 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
4386 {
4387 	struct mount *mp;
4388 
4389 	mp = vp->v_mount;
4390 	VFS_ASSERT_GIANT(mp);
4391 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
4392 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
4393 		(void)VOP_MARKATIME(vp);
4394 }
4395 
4396 /*
4397  * The purpose of this routine is to remove granularity from accmode_t,
4398  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
4399  * VADMIN and VAPPEND.
4400  *
4401  * If it returns 0, the caller is supposed to continue with the usual
4402  * access checks using 'accmode' as modified by this routine.  If it
4403  * returns nonzero value, the caller is supposed to return that value
4404  * as errno.
4405  *
4406  * Note that after this routine runs, accmode may be zero.
4407  */
4408 int
4409 vfs_unixify_accmode(accmode_t *accmode)
4410 {
4411 	/*
4412 	 * There is no way to specify explicit "deny" rule using
4413 	 * file mode or POSIX.1e ACLs.
4414 	 */
4415 	if (*accmode & VEXPLICIT_DENY) {
4416 		*accmode = 0;
4417 		return (0);
4418 	}
4419 
4420 	/*
4421 	 * None of these can be translated into usual access bits.
4422 	 * Also, the common case for NFSv4 ACLs is to not contain
4423 	 * either of these bits. Caller should check for VWRITE
4424 	 * on the containing directory instead.
4425 	 */
4426 	if (*accmode & (VDELETE_CHILD | VDELETE))
4427 		return (EPERM);
4428 
4429 	if (*accmode & VADMIN_PERMS) {
4430 		*accmode &= ~VADMIN_PERMS;
4431 		*accmode |= VADMIN;
4432 	}
4433 
4434 	/*
4435 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
4436 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
4437 	 */
4438 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
4439 
4440 	return (0);
4441 }
4442