xref: /dflybsd-src/sys/kern/vfs_mount.c (revision 15b85273287111e7f177850e47c2bd7a1f804ca1)
1 /*
2  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1989, 1993
35  *	The Regents of the University of California.  All rights reserved.
36  * (c) UNIX System Laboratories, Inc.
37  * All or some portions of this file are derived from material licensed
38  * to the University of California by American Telephone and Telegraph
39  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
40  * the permission of UNIX System Laboratories, Inc.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. All advertising materials mentioning features or use of this software
51  *    must display the following acknowledgement:
52  *	This product includes software developed by the University of
53  *	California, Berkeley and its contributors.
54  * 4. Neither the name of the University nor the names of its contributors
55  *    may be used to endorse or promote products derived from this software
56  *    without specific prior written permission.
57  *
58  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
59  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
60  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
61  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
62  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
63  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
64  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
65  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
66  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
67  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
68  * SUCH DAMAGE.
69  *
70  * $DragonFly: src/sys/kern/vfs_mount.c,v 1.37 2008/09/17 21:44:18 dillon Exp $
71  */
72 
73 /*
74  * External virtual filesystem routines
75  */
76 #include "opt_ddb.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/kernel.h>
81 #include <sys/malloc.h>
82 #include <sys/mount.h>
83 #include <sys/proc.h>
84 #include <sys/vnode.h>
85 #include <sys/buf.h>
86 #include <sys/eventhandler.h>
87 #include <sys/kthread.h>
88 #include <sys/sysctl.h>
89 
90 #include <machine/limits.h>
91 
92 #include <sys/buf2.h>
93 #include <sys/thread2.h>
94 #include <sys/sysref2.h>
95 
96 #include <vm/vm.h>
97 #include <vm/vm_object.h>
98 
99 struct mountscan_info {
100 	TAILQ_ENTRY(mountscan_info) msi_entry;
101 	int msi_how;
102 	struct mount *msi_node;
103 };
104 
105 struct vmntvnodescan_info {
106 	TAILQ_ENTRY(vmntvnodescan_info) entry;
107 	struct vnode *vp;
108 };
109 
110 static int vnlru_nowhere = 0;
111 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD,
112 	    &vnlru_nowhere, 0,
113 	    "Number of times the vnlru process ran without success");
114 
115 
116 static struct lwkt_token mntid_token;
117 
118 /* note: mountlist exported to pstat */
119 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist);
120 static TAILQ_HEAD(,mountscan_info) mountscan_list;
121 static struct lwkt_token mountlist_token;
122 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list;
123 struct lwkt_token mntvnode_token;
124 
125 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list);
126 
127 /*
128  * Called from vfsinit()
129  */
130 void
131 vfs_mount_init(void)
132 {
133 	lwkt_token_init(&mountlist_token);
134 	lwkt_token_init(&mntvnode_token);
135 	lwkt_token_init(&mntid_token);
136 	TAILQ_INIT(&mountscan_list);
137 	TAILQ_INIT(&mntvnodescan_list);
138 }
139 
140 /*
141  * Support function called with mntvnode_token held to remove a vnode
142  * from the mountlist.  We must update any list scans which are in progress.
143  */
144 static void
145 vremovevnodemnt(struct vnode *vp)
146 {
147         struct vmntvnodescan_info *info;
148 
149 	TAILQ_FOREACH(info, &mntvnodescan_list, entry) {
150 		if (info->vp == vp)
151 			info->vp = TAILQ_NEXT(vp, v_nmntvnodes);
152 	}
153 	TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
154 }
155 
156 /*
157  * Allocate a new vnode and associate it with a tag, mount point, and
158  * operations vector.
159  *
160  * A VX locked and refd vnode is returned.  The caller should setup the
161  * remaining fields and vx_put() or, if he wishes to leave a vref,
162  * vx_unlock() the vnode.
163  */
164 int
165 getnewvnode(enum vtagtype tag, struct mount *mp,
166 		struct vnode **vpp, int lktimeout, int lkflags)
167 {
168 	struct vnode *vp;
169 
170 	KKASSERT(mp != NULL);
171 
172 	vp = allocvnode(lktimeout, lkflags);
173 	vp->v_tag = tag;
174 	vp->v_data = NULL;
175 
176 	/*
177 	 * By default the vnode is assigned the mount point's normal
178 	 * operations vector.
179 	 */
180 	vp->v_ops = &mp->mnt_vn_use_ops;
181 
182 	/*
183 	 * Placing the vnode on the mount point's queue makes it visible.
184 	 * VNON prevents it from being messed with, however.
185 	 */
186 	insmntque(vp, mp);
187 
188 	/*
189 	 * A VX locked & refd vnode is returned.
190 	 */
191 	*vpp = vp;
192 	return (0);
193 }
194 
195 /*
196  * This function creates vnodes with special operations vectors.  The
197  * mount point is optional.
198  *
199  * This routine is being phased out.
200  */
201 int
202 getspecialvnode(enum vtagtype tag, struct mount *mp,
203 		struct vop_ops **ops,
204 		struct vnode **vpp, int lktimeout, int lkflags)
205 {
206 	struct vnode *vp;
207 
208 	vp = allocvnode(lktimeout, lkflags);
209 	vp->v_tag = tag;
210 	vp->v_data = NULL;
211 	vp->v_ops = ops;
212 
213 	/*
214 	 * Placing the vnode on the mount point's queue makes it visible.
215 	 * VNON prevents it from being messed with, however.
216 	 */
217 	insmntque(vp, mp);
218 
219 	/*
220 	 * A VX locked & refd vnode is returned.
221 	 */
222 	*vpp = vp;
223 	return (0);
224 }
225 
226 /*
227  * Interlock against an unmount, return 0 on success, non-zero on failure.
228  *
229  * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount
230  * is in-progress.
231  *
232  * If no unmount is in-progress LK_NOWAIT is ignored.  No other flag bits
233  * are used.  A shared locked will be obtained and the filesystem will not
234  * be unmountable until the lock is released.
235  */
236 int
237 vfs_busy(struct mount *mp, int flags)
238 {
239 	int lkflags;
240 
241 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
242 		if (flags & LK_NOWAIT)
243 			return (ENOENT);
244 		/* XXX not MP safe */
245 		mp->mnt_kern_flag |= MNTK_MWAIT;
246 		/*
247 		 * Since all busy locks are shared except the exclusive
248 		 * lock granted when unmounting, the only place that a
249 		 * wakeup needs to be done is at the release of the
250 		 * exclusive lock at the end of dounmount.
251 		 */
252 		tsleep((caddr_t)mp, 0, "vfs_busy", 0);
253 		return (ENOENT);
254 	}
255 	lkflags = LK_SHARED;
256 	if (lockmgr(&mp->mnt_lock, lkflags))
257 		panic("vfs_busy: unexpected lock failure");
258 	return (0);
259 }
260 
261 /*
262  * Free a busy filesystem.
263  */
264 void
265 vfs_unbusy(struct mount *mp)
266 {
267 	lockmgr(&mp->mnt_lock, LK_RELEASE);
268 }
269 
270 /*
271  * Lookup a filesystem type, and if found allocate and initialize
272  * a mount structure for it.
273  *
274  * Devname is usually updated by mount(8) after booting.
275  */
276 int
277 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp)
278 {
279 	struct vfsconf *vfsp;
280 	struct mount *mp;
281 
282 	if (fstypename == NULL)
283 		return (ENODEV);
284 	for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) {
285 		if (!strcmp(vfsp->vfc_name, fstypename))
286 			break;
287 	}
288 	if (vfsp == NULL)
289 		return (ENODEV);
290 	mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO);
291 	lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0);
292 	vfs_busy(mp, LK_NOWAIT);
293 	TAILQ_INIT(&mp->mnt_nvnodelist);
294 	TAILQ_INIT(&mp->mnt_reservedvnlist);
295 	TAILQ_INIT(&mp->mnt_jlist);
296 	mp->mnt_nvnodelistsize = 0;
297 	mp->mnt_vfc = vfsp;
298 	mp->mnt_op = vfsp->vfc_vfsops;
299 	mp->mnt_flag = MNT_RDONLY;
300 	vfsp->vfc_refcount++;
301 	mp->mnt_iosize_max = DFLTPHYS;
302 	mp->mnt_stat.f_type = vfsp->vfc_typenum;
303 	mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
304 	strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
305 	copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
306 	*mpp = mp;
307 	return (0);
308 }
309 
310 /*
311  * Lookup a mount point by filesystem identifier.
312  */
313 struct mount *
314 vfs_getvfs(fsid_t *fsid)
315 {
316 	struct mount *mp;
317 	lwkt_tokref ilock;
318 
319 	lwkt_gettoken(&ilock, &mountlist_token);
320 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
321 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
322 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
323 			break;
324 		}
325 	}
326 	lwkt_reltoken(&ilock);
327 	return (mp);
328 }
329 
330 /*
331  * Get a new unique fsid.  Try to make its val[0] unique, since this value
332  * will be used to create fake device numbers for stat().  Also try (but
333  * not so hard) make its val[0] unique mod 2^16, since some emulators only
334  * support 16-bit device numbers.  We end up with unique val[0]'s for the
335  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
336  *
337  * Keep in mind that several mounts may be running in parallel.  Starting
338  * the search one past where the previous search terminated is both a
339  * micro-optimization and a defense against returning the same fsid to
340  * different mounts.
341  */
342 void
343 vfs_getnewfsid(struct mount *mp)
344 {
345 	static u_int16_t mntid_base;
346 	lwkt_tokref ilock;
347 	fsid_t tfsid;
348 	int mtype;
349 
350 	lwkt_gettoken(&ilock, &mntid_token);
351 	mtype = mp->mnt_vfc->vfc_typenum;
352 	tfsid.val[1] = mtype;
353 	mtype = (mtype & 0xFF) << 24;
354 	for (;;) {
355 		tfsid.val[0] = makeudev(255,
356 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
357 		mntid_base++;
358 		if (vfs_getvfs(&tfsid) == NULL)
359 			break;
360 	}
361 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
362 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
363 	lwkt_reltoken(&ilock);
364 }
365 
366 /*
367  * Set the FSID for a new mount point to the template.  Adjust
368  * the FSID to avoid collisions.
369  */
370 int
371 vfs_setfsid(struct mount *mp, fsid_t *template)
372 {
373 	int didmunge = 0;
374 
375 	bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid));
376 	for (;;) {
377 		if (vfs_getvfs(template) == NULL)
378 			break;
379 		didmunge = 1;
380 		++template->val[1];
381 	}
382 	mp->mnt_stat.f_fsid = *template;
383 	return(didmunge);
384 }
385 
386 /*
387  * This routine is called when we have too many vnodes.  It attempts
388  * to free <count> vnodes and will potentially free vnodes that still
389  * have VM backing store (VM backing store is typically the cause
390  * of a vnode blowout so we want to do this).  Therefore, this operation
391  * is not considered cheap.
392  *
393  * A number of conditions may prevent a vnode from being reclaimed.
394  * the buffer cache may have references on the vnode, a directory
395  * vnode may still have references due to the namei cache representing
396  * underlying files, or the vnode may be in active use.   It is not
397  * desireable to reuse such vnodes.  These conditions may cause the
398  * number of vnodes to reach some minimum value regardless of what
399  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
400  */
401 
402 /*
403  * This is a quick non-blocking check to determine if the vnode is a good
404  * candidate for being (eventually) vgone()'d.  Returns 0 if the vnode is
405  * not a good candidate, 1 if it is.
406  */
407 static __inline int
408 vmightfree(struct vnode *vp, int page_count)
409 {
410 	if (vp->v_flag & VRECLAIMED)
411 		return (0);
412 #if 0
413 	if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
414 		return (0);
415 #endif
416 	if (sysref_isactive(&vp->v_sysref))
417 		return (0);
418 	if (vp->v_object && vp->v_object->resident_page_count >= page_count)
419 		return (0);
420 	return (1);
421 }
422 
423 /*
424  * The vnode was found to be possibly vgone()able and the caller has locked it
425  * (thus the usecount should be 1 now).  Determine if the vnode is actually
426  * vgone()able, doing some cleanups in the process.  Returns 1 if the vnode
427  * can be vgone()'d, 0 otherwise.
428  *
429  * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf
430  * in the namecache topology and (B) this vnode has buffer cache bufs.
431  * We cannot remove vnodes with non-leaf namecache associations.  We do a
432  * tentitive leaf check prior to attempting to flush out any buffers but the
433  * 'real' test when all is said in done is that v_auxrefs must become 0 for
434  * the vnode to be freeable.
435  *
436  * We could theoretically just unconditionally flush when v_auxrefs != 0,
437  * but flushing data associated with non-leaf nodes (which are always
438  * directories), just throws it away for no benefit.  It is the buffer
439  * cache's responsibility to choose buffers to recycle from the cached
440  * data point of view.
441  */
442 static int
443 visleaf(struct vnode *vp)
444 {
445 	struct namecache *ncp;
446 
447 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
448 		if (!TAILQ_EMPTY(&ncp->nc_list))
449 			return(0);
450 	}
451 	return(1);
452 }
453 
454 /*
455  * Try to clean up the vnode to the point where it can be vgone()'d, returning
456  * 0 if it cannot be vgone()'d (or already has been), 1 if it can.  Unlike
457  * vmightfree() this routine may flush the vnode and block.  Vnodes marked
458  * VFREE are still candidates for vgone()ing because they may hold namecache
459  * resources and could be blocking the namecache directory hierarchy (and
460  * related vnodes) from being freed.
461  */
462 static int
463 vtrytomakegoneable(struct vnode *vp, int page_count)
464 {
465 	if (vp->v_flag & VRECLAIMED)
466 		return (0);
467 	if (vp->v_sysref.refcnt > 1)
468 		return (0);
469 	if (vp->v_object && vp->v_object->resident_page_count >= page_count)
470 		return (0);
471 	if (vp->v_auxrefs && visleaf(vp)) {
472 		vinvalbuf(vp, V_SAVE, 0, 0);
473 #if 0	/* DEBUG */
474 		kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" :
475 			"vrecycle: vp %p succeeded: %s\n"), vp,
476 			(TAILQ_FIRST(&vp->v_namecache) ?
477 			    TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
478 #endif
479 	}
480 
481 	/*
482 	 * This sequence may seem a little strange, but we need to optimize
483 	 * the critical path a bit.  We can't recycle vnodes with other
484 	 * references and because we are trying to recycle an otherwise
485 	 * perfectly fine vnode we have to invalidate the namecache in a
486 	 * way that avoids possible deadlocks (since the vnode lock is being
487 	 * held here).  Finally, we have to check for other references one
488 	 * last time in case something snuck in during the inval.
489 	 */
490 	if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0)
491 		return (0);
492 	if (cache_inval_vp_nonblock(vp))
493 		return (0);
494 	return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0);
495 }
496 
497 /*
498  * Reclaim up to 1/10 of the vnodes associated with a mount point.  Try
499  * to avoid vnodes which have lots of resident pages (we are trying to free
500  * vnodes, not memory).
501  *
502  * This routine is a callback from the mountlist scan.  The mount point
503  * in question will be busied.
504  */
505 static int
506 vlrureclaim(struct mount *mp, void *data)
507 {
508 	struct vnode *vp;
509 	lwkt_tokref ilock;
510 	int done;
511 	int trigger;
512 	int usevnodes;
513 	int count;
514 	int trigger_mult = vnlru_nowhere;
515 
516 	/*
517 	 * Calculate the trigger point for the resident pages check.  The
518 	 * minimum trigger value is approximately the number of pages in
519 	 * the system divded by the number of vnodes.  However, due to
520 	 * various other system memory overheads unrelated to data caching
521 	 * it is a good idea to double the trigger (at least).
522 	 *
523 	 * trigger_mult starts at 0.  If the recycler is having problems
524 	 * finding enough freeable vnodes it will increase trigger_mult.
525 	 * This should not happen in normal operation, even on machines with
526 	 * low amounts of memory, but extraordinary memory use by the system
527 	 * verses the amount of cached data can trigger it.
528 	 */
529 	usevnodes = desiredvnodes;
530 	if (usevnodes <= 0)
531 		usevnodes = 1;
532 	trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
533 
534 	done = 0;
535 	lwkt_gettoken(&ilock, &mntvnode_token);
536 	count = mp->mnt_nvnodelistsize / 10 + 1;
537 	while (count && mp->mnt_syncer) {
538 		/*
539 		 * Next vnode.  Use the special syncer vnode to placemark
540 		 * the LRU.  This way the LRU code does not interfere with
541 		 * vmntvnodescan().
542 		 */
543 		vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
544 		TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes);
545 		if (vp) {
546 			TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp,
547 					   mp->mnt_syncer, v_nmntvnodes);
548 		} else {
549 			TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer,
550 					  v_nmntvnodes);
551 			vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
552 			if (vp == NULL)
553 				break;
554 		}
555 
556 		/*
557 		 * __VNODESCAN__
558 		 *
559 		 * The VP will stick around while we hold mntvnode_token,
560 		 * at least until we block, so we can safely do an initial
561 		 * check, and then must check again after we lock the vnode.
562 		 */
563 		if (vp->v_type == VNON ||	/* syncer or indeterminant */
564 		    !vmightfree(vp, trigger)	/* critical path opt */
565 		) {
566 			--count;
567 			continue;
568 		}
569 
570 		/*
571 		 * VX get the candidate vnode.  If the VX get fails the
572 		 * vnode might still be on the mountlist.  Our loop depends
573 		 * on us at least cycling the vnode to the end of the
574 		 * mountlist.
575 		 */
576 		if (vx_get_nonblock(vp) != 0) {
577 			--count;
578 			continue;
579 		}
580 
581 		/*
582 		 * Since we blocked locking the vp, make sure it is still
583 		 * a candidate for reclamation.  That is, it has not already
584 		 * been reclaimed and only has our VX reference associated
585 		 * with it.
586 		 */
587 		if (vp->v_type == VNON ||	/* syncer or indeterminant */
588 		    (vp->v_flag & VRECLAIMED) ||
589 		    vp->v_mount != mp ||
590 		    !vtrytomakegoneable(vp, trigger)	/* critical path opt */
591 		) {
592 			--count;
593 			vx_put(vp);
594 			continue;
595 		}
596 
597 		/*
598 		 * All right, we are good, move the vp to the end of the
599 		 * mountlist and clean it out.  The vget will have returned
600 		 * an error if the vnode was destroyed (VRECLAIMED set), so we
601 		 * do not have to check again.  The vput() will move the
602 		 * vnode to the free list if the vgone() was successful.
603 		 */
604 		KKASSERT(vp->v_mount == mp);
605 		vgone_vxlocked(vp);
606 		vx_put(vp);
607 		++done;
608 		--count;
609 	}
610 	lwkt_reltoken(&ilock);
611 	return (done);
612 }
613 
614 /*
615  * Attempt to recycle vnodes in a context that is always safe to block.
616  * Calling vlrurecycle() from the bowels of file system code has some
617  * interesting deadlock problems.
618  */
619 static struct thread *vnlruthread;
620 static int vnlruproc_sig;
621 
622 void
623 vnlru_proc_wait(void)
624 {
625 	if (vnlruproc_sig == 0) {
626 		vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
627 		wakeup(vnlruthread);
628 	}
629 	tsleep(&vnlruproc_sig, 0, "vlruwk", hz);
630 }
631 
632 static void
633 vnlru_proc(void)
634 {
635 	struct thread *td = curthread;
636 	int done;
637 
638 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
639 	    SHUTDOWN_PRI_FIRST);
640 
641 	crit_enter();
642 	for (;;) {
643 		kproc_suspend_loop();
644 
645 		/*
646 		 * Try to free some vnodes if we have too many
647 		 */
648 		if (numvnodes > desiredvnodes &&
649 		    freevnodes > desiredvnodes * 2 / 10) {
650 			int count = numvnodes - desiredvnodes;
651 
652 			if (count > freevnodes / 100)
653 				count = freevnodes / 100;
654 			if (count < 5)
655 				count = 5;
656 			freesomevnodes(count);
657 		}
658 
659 		/*
660 		 * Nothing to do if most of our vnodes are already on
661 		 * the free list.
662 		 */
663 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
664 			vnlruproc_sig = 0;
665 			wakeup(&vnlruproc_sig);
666 			tsleep(td, 0, "vlruwt", hz);
667 			continue;
668 		}
669 		cache_cleanneg(0);
670 		done = mountlist_scan(vlrureclaim, NULL, MNTSCAN_FORWARD);
671 
672 		/*
673 		 * The vlrureclaim() call only processes 1/10 of the vnodes
674 		 * on each mount.  If we couldn't find any repeat the loop
675 		 * at least enough times to cover all available vnodes before
676 		 * we start sleeping.  Complain if the failure extends past
677 		 * 30 second, every 30 seconds.
678 		 */
679 		if (done == 0) {
680 			++vnlru_nowhere;
681 			if (vnlru_nowhere % 10 == 0)
682 				tsleep(td, 0, "vlrup", hz * 3);
683 			if (vnlru_nowhere % 100 == 0)
684 				kprintf("vnlru_proc: vnode recycler stopped working!\n");
685 			if (vnlru_nowhere == 1000)
686 				vnlru_nowhere = 900;
687 		} else {
688 			vnlru_nowhere = 0;
689 		}
690 	}
691 	crit_exit();
692 }
693 
694 /*
695  * MOUNTLIST FUNCTIONS
696  */
697 
698 /*
699  * mountlist_insert (MP SAFE)
700  *
701  * Add a new mount point to the mount list.
702  */
703 void
704 mountlist_insert(struct mount *mp, int how)
705 {
706 	lwkt_tokref ilock;
707 
708 	lwkt_gettoken(&ilock, &mountlist_token);
709 	if (how == MNTINS_FIRST)
710 	    TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
711 	else
712 	    TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
713 	lwkt_reltoken(&ilock);
714 }
715 
716 /*
717  * mountlist_interlock (MP SAFE)
718  *
719  * Execute the specified interlock function with the mountlist token
720  * held.  The function will be called in a serialized fashion verses
721  * other functions called through this mechanism.
722  */
723 int
724 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
725 {
726 	lwkt_tokref ilock;
727 	int error;
728 
729 	lwkt_gettoken(&ilock, &mountlist_token);
730 	error = callback(mp);
731 	lwkt_reltoken(&ilock);
732 	return (error);
733 }
734 
735 /*
736  * mountlist_boot_getfirst (DURING BOOT ONLY)
737  *
738  * This function returns the first mount on the mountlist, which is
739  * expected to be the root mount.  Since no interlocks are obtained
740  * this function is only safe to use during booting.
741  */
742 
743 struct mount *
744 mountlist_boot_getfirst(void)
745 {
746 	return(TAILQ_FIRST(&mountlist));
747 }
748 
749 /*
750  * mountlist_remove (MP SAFE)
751  *
752  * Remove a node from the mountlist.  If this node is the next scan node
753  * for any active mountlist scans, the active mountlist scan will be
754  * adjusted to skip the node, thus allowing removals during mountlist
755  * scans.
756  */
757 void
758 mountlist_remove(struct mount *mp)
759 {
760 	struct mountscan_info *msi;
761 	lwkt_tokref ilock;
762 
763 	lwkt_gettoken(&ilock, &mountlist_token);
764 	TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
765 		if (msi->msi_node == mp) {
766 			if (msi->msi_how & MNTSCAN_FORWARD)
767 				msi->msi_node = TAILQ_NEXT(mp, mnt_list);
768 			else
769 				msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
770 		}
771 	}
772 	TAILQ_REMOVE(&mountlist, mp, mnt_list);
773 	lwkt_reltoken(&ilock);
774 }
775 
776 /*
777  * mountlist_scan (MP SAFE)
778  *
779  * Safely scan the mount points on the mount list.  Unless otherwise
780  * specified each mount point will be busied prior to the callback and
781  * unbusied afterwords.  The callback may safely remove any mount point
782  * without interfering with the scan.  If the current callback
783  * mount is removed the scanner will not attempt to unbusy it.
784  *
785  * If a mount node cannot be busied it is silently skipped.
786  *
787  * The callback return value is aggregated and a total is returned.  A return
788  * value of < 0 is not aggregated and will terminate the scan.
789  *
790  * MNTSCAN_FORWARD	- the mountlist is scanned in the forward direction
791  * MNTSCAN_REVERSE	- the mountlist is scanned in reverse
792  * MNTSCAN_NOBUSY	- the scanner will make the callback without busying
793  *			  the mount node.
794  */
795 int
796 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
797 {
798 	struct mountscan_info info;
799 	lwkt_tokref ilock;
800 	struct mount *mp;
801 	thread_t td;
802 	int count;
803 	int res;
804 
805 	lwkt_gettoken(&ilock, &mountlist_token);
806 
807 	info.msi_how = how;
808 	info.msi_node = NULL;	/* paranoia */
809 	TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
810 
811 	res = 0;
812 	td = curthread;
813 
814 	if (how & MNTSCAN_FORWARD) {
815 		info.msi_node = TAILQ_FIRST(&mountlist);
816 		while ((mp = info.msi_node) != NULL) {
817 			if (how & MNTSCAN_NOBUSY) {
818 				count = callback(mp, data);
819 			} else if (vfs_busy(mp, LK_NOWAIT) == 0) {
820 				count = callback(mp, data);
821 				if (mp == info.msi_node)
822 					vfs_unbusy(mp);
823 			} else {
824 				count = 0;
825 			}
826 			if (count < 0)
827 				break;
828 			res += count;
829 			if (mp == info.msi_node)
830 				info.msi_node = TAILQ_NEXT(mp, mnt_list);
831 		}
832 	} else if (how & MNTSCAN_REVERSE) {
833 		info.msi_node = TAILQ_LAST(&mountlist, mntlist);
834 		while ((mp = info.msi_node) != NULL) {
835 			if (how & MNTSCAN_NOBUSY) {
836 				count = callback(mp, data);
837 			} else if (vfs_busy(mp, LK_NOWAIT) == 0) {
838 				count = callback(mp, data);
839 				if (mp == info.msi_node)
840 					vfs_unbusy(mp);
841 			} else {
842 				count = 0;
843 			}
844 			if (count < 0)
845 				break;
846 			res += count;
847 			if (mp == info.msi_node)
848 				info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
849 		}
850 	}
851 	TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
852 	lwkt_reltoken(&ilock);
853 	return(res);
854 }
855 
856 /*
857  * MOUNT RELATED VNODE FUNCTIONS
858  */
859 
860 static struct kproc_desc vnlru_kp = {
861 	"vnlru",
862 	vnlru_proc,
863 	&vnlruthread
864 };
865 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
866 
867 /*
868  * Move a vnode from one mount queue to another.
869  */
870 void
871 insmntque(struct vnode *vp, struct mount *mp)
872 {
873 	lwkt_tokref ilock;
874 
875 	lwkt_gettoken(&ilock, &mntvnode_token);
876 	/*
877 	 * Delete from old mount point vnode list, if on one.
878 	 */
879 	if (vp->v_mount != NULL) {
880 		KASSERT(vp->v_mount->mnt_nvnodelistsize > 0,
881 			("bad mount point vnode list size"));
882 		vremovevnodemnt(vp);
883 		vp->v_mount->mnt_nvnodelistsize--;
884 	}
885 	/*
886 	 * Insert into list of vnodes for the new mount point, if available.
887 	 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer.
888 	 */
889 	if ((vp->v_mount = mp) == NULL) {
890 		lwkt_reltoken(&ilock);
891 		return;
892 	}
893 	if (mp->mnt_syncer) {
894 		TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes);
895 	} else {
896 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
897 	}
898 	mp->mnt_nvnodelistsize++;
899 	lwkt_reltoken(&ilock);
900 }
901 
902 
903 /*
904  * Scan the vnodes under a mount point and issue appropriate callbacks.
905  *
906  * The fastfunc() callback is called with just the mountlist token held
907  * (no vnode lock).  It may not block and the vnode may be undergoing
908  * modifications while the caller is processing it.  The vnode will
909  * not be entirely destroyed, however, due to the fact that the mountlist
910  * token is held.  A return value < 0 skips to the next vnode without calling
911  * the slowfunc(), a return value > 0 terminates the loop.
912  *
913  * The slowfunc() callback is called after the vnode has been successfully
914  * locked based on passed flags.  The vnode is skipped if it gets rearranged
915  * or destroyed while blocking on the lock.  A non-zero return value from
916  * the slow function terminates the loop.  The slow function is allowed to
917  * arbitrarily block.  The scanning code guarentees consistency of operation
918  * even if the slow function deletes or moves the node, or blocks and some
919  * other thread deletes or moves the node.
920  */
921 int
922 vmntvnodescan(
923     struct mount *mp,
924     int flags,
925     int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
926     int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
927     void *data
928 ) {
929 	struct vmntvnodescan_info info;
930 	lwkt_tokref ilock;
931 	struct vnode *vp;
932 	int r = 0;
933 	int maxcount = 1000000;
934 	int stopcount = 0;
935 	int count = 0;
936 
937 	lwkt_gettoken(&ilock, &mntvnode_token);
938 
939 	/*
940 	 * If asked to do one pass stop after iterating available vnodes.
941 	 * Under heavy loads new vnodes can be added while we are scanning,
942 	 * so this isn't perfect.  Create a slop factor of 2x.
943 	 */
944 	if (flags & VMSC_ONEPASS)
945 		stopcount = mp->mnt_nvnodelistsize * 2;
946 
947 	info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
948 	TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry);
949 	while ((vp = info.vp) != NULL) {
950 		if (--maxcount == 0)
951 			panic("maxcount reached during vmntvnodescan");
952 
953 		/*
954 		 * Skip if visible but not ready, or special (e.g.
955 		 * mp->mnt_syncer)
956 		 */
957 		if (vp->v_type == VNON)
958 			goto next;
959 		KKASSERT(vp->v_mount == mp);
960 
961 		/*
962 		 * Quick test.  A negative return continues the loop without
963 		 * calling the slow test.  0 continues onto the slow test.
964 		 * A positive number aborts the loop.
965 		 */
966 		if (fastfunc) {
967 			if ((r = fastfunc(mp, vp, data)) < 0) {
968 				r = 0;
969 				goto next;
970 			}
971 			if (r)
972 				break;
973 		}
974 
975 		/*
976 		 * Get a vxlock on the vnode, retry if it has moved or isn't
977 		 * in the mountlist where we expect it.
978 		 */
979 		if (slowfunc) {
980 			int error;
981 
982 			switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
983 			case VMSC_GETVP:
984 				error = vget(vp, LK_EXCLUSIVE);
985 				break;
986 			case VMSC_GETVP|VMSC_NOWAIT:
987 				error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT);
988 				break;
989 			case VMSC_GETVX:
990 				vx_get(vp);
991 				error = 0;
992 				break;
993 			default:
994 				error = 0;
995 				break;
996 			}
997 			if (error)
998 				goto next;
999 			/*
1000 			 * Do not call the slow function if the vnode is
1001 			 * invalid or if it was ripped out from under us
1002 			 * while we (potentially) blocked.
1003 			 */
1004 			if (info.vp == vp && vp->v_type != VNON)
1005 				r = slowfunc(mp, vp, data);
1006 
1007 			/*
1008 			 * Cleanup
1009 			 */
1010 			switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1011 			case VMSC_GETVP:
1012 			case VMSC_GETVP|VMSC_NOWAIT:
1013 				vput(vp);
1014 				break;
1015 			case VMSC_GETVX:
1016 				vx_put(vp);
1017 				break;
1018 			default:
1019 				break;
1020 			}
1021 			if (r != 0)
1022 				break;
1023 		}
1024 
1025 next:
1026 		/*
1027 		 * Yield after some processing.  Depending on the number
1028 		 * of vnodes, we might wind up running for a long time.
1029 		 * Because threads are not preemptable, time critical
1030 		 * userland processes might starve.  Give them a chance
1031 		 * now and then.
1032 		 */
1033 		if (++count == 10000) {
1034 			/* We really want to yield a bit, so we simply sleep a tick */
1035 			tsleep(mp, 0, "vnodescn", 1);
1036 			count = 0;
1037 		}
1038 
1039 		/*
1040 		 * If doing one pass this decrements to zero.  If it starts
1041 		 * at zero it is effectively unlimited for the purposes of
1042 		 * this loop.
1043 		 */
1044 		if (--stopcount == 0)
1045 			break;
1046 
1047 		/*
1048 		 * Iterate.  If the vnode was ripped out from under us
1049 		 * info.vp will already point to the next vnode, otherwise
1050 		 * we have to obtain the next valid vnode ourselves.
1051 		 */
1052 		if (info.vp == vp)
1053 			info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
1054 	}
1055 	TAILQ_REMOVE(&mntvnodescan_list, &info, entry);
1056 	lwkt_reltoken(&ilock);
1057 	return(r);
1058 }
1059 
1060 /*
1061  * Remove any vnodes in the vnode table belonging to mount point mp.
1062  *
1063  * If FORCECLOSE is not specified, there should not be any active ones,
1064  * return error if any are found (nb: this is a user error, not a
1065  * system error). If FORCECLOSE is specified, detach any active vnodes
1066  * that are found.
1067  *
1068  * If WRITECLOSE is set, only flush out regular file vnodes open for
1069  * writing.
1070  *
1071  * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
1072  *
1073  * `rootrefs' specifies the base reference count for the root vnode
1074  * of this filesystem. The root vnode is considered busy if its
1075  * v_sysref.refcnt exceeds this value. On a successful return, vflush()
1076  * will call vrele() on the root vnode exactly rootrefs times.
1077  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
1078  * be zero.
1079  */
1080 #ifdef DIAGNOSTIC
1081 static int busyprt = 0;		/* print out busy vnodes */
1082 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
1083 #endif
1084 
1085 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
1086 
1087 struct vflush_info {
1088 	int flags;
1089 	int busy;
1090 	thread_t td;
1091 };
1092 
1093 int
1094 vflush(struct mount *mp, int rootrefs, int flags)
1095 {
1096 	struct thread *td = curthread;	/* XXX */
1097 	struct vnode *rootvp = NULL;
1098 	int error;
1099 	struct vflush_info vflush_info;
1100 
1101 	if (rootrefs > 0) {
1102 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1103 		    ("vflush: bad args"));
1104 		/*
1105 		 * Get the filesystem root vnode. We can vput() it
1106 		 * immediately, since with rootrefs > 0, it won't go away.
1107 		 */
1108 		if ((error = VFS_ROOT(mp, &rootvp)) != 0)
1109 			return (error);
1110 		vput(rootvp);
1111 	}
1112 
1113 	vflush_info.busy = 0;
1114 	vflush_info.flags = flags;
1115 	vflush_info.td = td;
1116 	vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1117 
1118 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1119 		/*
1120 		 * If just the root vnode is busy, and if its refcount
1121 		 * is equal to `rootrefs', then go ahead and kill it.
1122 		 */
1123 		KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
1124 		KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs"));
1125 		if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) {
1126 			vx_lock(rootvp);
1127 			vgone_vxlocked(rootvp);
1128 			vx_unlock(rootvp);
1129 			vflush_info.busy = 0;
1130 		}
1131 	}
1132 	if (vflush_info.busy)
1133 		return (EBUSY);
1134 	for (; rootrefs > 0; rootrefs--)
1135 		vrele(rootvp);
1136 	return (0);
1137 }
1138 
1139 /*
1140  * The scan callback is made with an VX locked vnode.
1141  */
1142 static int
1143 vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1144 {
1145 	struct vflush_info *info = data;
1146 	struct vattr vattr;
1147 
1148 	/*
1149 	 * Skip over a vnodes marked VSYSTEM.
1150 	 */
1151 	if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1152 		return(0);
1153 	}
1154 
1155 	/*
1156 	 * If WRITECLOSE is set, flush out unlinked but still open
1157 	 * files (even if open only for reading) and regular file
1158 	 * vnodes open for writing.
1159 	 */
1160 	if ((info->flags & WRITECLOSE) &&
1161 	    (vp->v_type == VNON ||
1162 	    (VOP_GETATTR(vp, &vattr) == 0 &&
1163 	    vattr.va_nlink > 0)) &&
1164 	    (vp->v_writecount == 0 || vp->v_type != VREG)) {
1165 		return(0);
1166 	}
1167 
1168 	/*
1169 	 * If we are the only holder (refcnt of 1) or the vnode is in
1170 	 * termination (refcnt < 0), we can vgone the vnode.
1171 	 */
1172 	if (vp->v_sysref.refcnt <= 1) {
1173 		vgone_vxlocked(vp);
1174 		return(0);
1175 	}
1176 
1177 	/*
1178 	 * If FORCECLOSE is set, forcibly close the vnode. For block
1179 	 * or character devices, revert to an anonymous device. For
1180 	 * all other files, just kill them.
1181 	 */
1182 	if (info->flags & FORCECLOSE) {
1183 		if (vp->v_type != VBLK && vp->v_type != VCHR) {
1184 			vgone_vxlocked(vp);
1185 		} else {
1186 			vclean_vxlocked(vp, 0);
1187 			vp->v_ops = &spec_vnode_vops_p;
1188 			insmntque(vp, NULL);
1189 		}
1190 		return(0);
1191 	}
1192 #ifdef DIAGNOSTIC
1193 	if (busyprt)
1194 		vprint("vflush: busy vnode", vp);
1195 #endif
1196 	++info->busy;
1197 	return(0);
1198 }
1199 
1200 void
1201 add_bio_ops(struct bio_ops *ops)
1202 {
1203 	TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry);
1204 }
1205 
1206 void
1207 rem_bio_ops(struct bio_ops *ops)
1208 {
1209 	TAILQ_REMOVE(&bio_ops_list, ops, entry);
1210 }
1211 
1212 /*
1213  * This calls the bio_ops io_sync function either for a mount point
1214  * or generally.
1215  *
1216  * WARNING: softdeps is weirdly coded and just isn't happy unless
1217  * io_sync is called with a NULL mount from the general syncing code.
1218  */
1219 void
1220 bio_ops_sync(struct mount *mp)
1221 {
1222 	struct bio_ops *ops;
1223 
1224 	if (mp) {
1225 		if ((ops = mp->mnt_bioops) != NULL)
1226 			ops->io_sync(mp);
1227 	} else {
1228 		TAILQ_FOREACH(ops, &bio_ops_list, entry) {
1229 			ops->io_sync(NULL);
1230 		}
1231 	}
1232 }
1233 
1234