2 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
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.
42 * Redistribution and use in source and binary forms, with or without
43 * modification, are permitted provided that the following conditions
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.
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
70 * $DragonFly: src/sys/kern/vfs_mount.c,v 1.37 2008/09/17 21:44:18 dillon Exp $
74 * External virtual filesystem routines
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>
84 #include <sys/vnode.h>
86 #include <sys/eventhandler.h>
87 #include <sys/kthread.h>
88 #include <sys/sysctl.h>
90 #include <machine/limits.h>
93 #include <sys/thread2.h>
94 #include <sys/sysref2.h>
95 #include <sys/mplock2.h>
98 #include <vm/vm_object.h>
100 struct mountscan_info {
101 TAILQ_ENTRY(mountscan_info) msi_entry;
103 struct mount *msi_node;
106 struct vmntvnodescan_info {
107 TAILQ_ENTRY(vmntvnodescan_info) entry;
115 static int vnlru_nowhere = 0;
116 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD,
118 "Number of times the vnlru process ran without success");
121 static struct lwkt_token mntid_token;
122 static struct mount dummymount;
124 /* note: mountlist exported to pstat */
125 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist);
126 static TAILQ_HEAD(,mountscan_info) mountscan_list;
127 static struct lwkt_token mountlist_token;
128 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list;
129 struct lwkt_token mntvnode_token;
131 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list);
134 * Called from vfsinit()
139 lwkt_token_init(&mountlist_token, "mntlist");
140 lwkt_token_init(&mntvnode_token, "mntvnode");
141 lwkt_token_init(&mntid_token, "mntid");
142 TAILQ_INIT(&mountscan_list);
143 TAILQ_INIT(&mntvnodescan_list);
144 mount_init(&dummymount);
145 dummymount.mnt_flag |= MNT_RDONLY;
146 dummymount.mnt_kern_flag |= MNTK_ALL_MPSAFE;
150 * Support function called with mntvnode_token held to remove a vnode
151 * from the mountlist. We must update any list scans which are in progress.
154 vremovevnodemnt(struct vnode *vp)
156 struct vmntvnodescan_info *info;
158 TAILQ_FOREACH(info, &mntvnodescan_list, entry) {
160 info->vp = TAILQ_NEXT(vp, v_nmntvnodes);
162 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
166 * Allocate a new vnode and associate it with a tag, mount point, and
169 * A VX locked and refd vnode is returned. The caller should setup the
170 * remaining fields and vx_put() or, if he wishes to leave a vref,
171 * vx_unlock() the vnode.
174 getnewvnode(enum vtagtype tag, struct mount *mp,
175 struct vnode **vpp, int lktimeout, int lkflags)
179 KKASSERT(mp != NULL);
181 vp = allocvnode(lktimeout, lkflags);
186 * By default the vnode is assigned the mount point's normal
189 vp->v_ops = &mp->mnt_vn_use_ops;
192 * Placing the vnode on the mount point's queue makes it visible.
193 * VNON prevents it from being messed with, however.
198 * A VX locked & refd vnode is returned.
205 * This function creates vnodes with special operations vectors. The
206 * mount point is optional.
208 * This routine is being phased out but is still used by vfs_conf to
209 * create vnodes for devices prior to the root mount (with mp == NULL).
212 getspecialvnode(enum vtagtype tag, struct mount *mp,
213 struct vop_ops **ops,
214 struct vnode **vpp, int lktimeout, int lkflags)
218 vp = allocvnode(lktimeout, lkflags);
227 * Placing the vnode on the mount point's queue makes it visible.
228 * VNON prevents it from being messed with, however.
233 * A VX locked & refd vnode is returned.
240 * Interlock against an unmount, return 0 on success, non-zero on failure.
242 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount
245 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits
246 * are used. A shared locked will be obtained and the filesystem will not
247 * be unmountable until the lock is released.
250 vfs_busy(struct mount *mp, int flags)
254 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
255 if (flags & LK_NOWAIT)
257 /* XXX not MP safe */
258 mp->mnt_kern_flag |= MNTK_MWAIT;
260 * Since all busy locks are shared except the exclusive
261 * lock granted when unmounting, the only place that a
262 * wakeup needs to be done is at the release of the
263 * exclusive lock at the end of dounmount.
265 tsleep((caddr_t)mp, 0, "vfs_busy", 0);
269 if (lockmgr(&mp->mnt_lock, lkflags))
270 panic("vfs_busy: unexpected lock failure");
275 * Free a busy filesystem.
278 vfs_unbusy(struct mount *mp)
280 lockmgr(&mp->mnt_lock, LK_RELEASE);
284 * Lookup a filesystem type, and if found allocate and initialize
285 * a mount structure for it.
287 * Devname is usually updated by mount(8) after booting.
290 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp)
292 struct vfsconf *vfsp;
295 if (fstypename == NULL)
298 vfsp = vfsconf_find_by_name(fstypename);
301 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO);
303 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0);
307 mp->mnt_op = vfsp->vfc_vfsops;
308 vfsp->vfc_refcount++;
309 mp->mnt_stat.f_type = vfsp->vfc_typenum;
310 mp->mnt_flag |= MNT_RDONLY;
311 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
312 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
313 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
319 * Basic mount structure initialization
322 mount_init(struct mount *mp)
324 lockinit(&mp->mnt_lock, "vfslock", 0, 0);
325 lwkt_token_init(&mp->mnt_token, "permnt");
327 TAILQ_INIT(&mp->mnt_nvnodelist);
328 TAILQ_INIT(&mp->mnt_reservedvnlist);
329 TAILQ_INIT(&mp->mnt_jlist);
330 mp->mnt_nvnodelistsize = 0;
332 mp->mnt_iosize_max = DFLTPHYS;
336 * Lookup a mount point by filesystem identifier.
339 vfs_getvfs(fsid_t *fsid)
343 lwkt_gettoken(&mountlist_token);
344 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
345 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
346 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
350 lwkt_reltoken(&mountlist_token);
355 * Get a new unique fsid. Try to make its val[0] unique, since this value
356 * will be used to create fake device numbers for stat(). Also try (but
357 * not so hard) make its val[0] unique mod 2^16, since some emulators only
358 * support 16-bit device numbers. We end up with unique val[0]'s for the
359 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
361 * Keep in mind that several mounts may be running in parallel. Starting
362 * the search one past where the previous search terminated is both a
363 * micro-optimization and a defense against returning the same fsid to
367 vfs_getnewfsid(struct mount *mp)
369 static u_int16_t mntid_base;
373 lwkt_gettoken(&mntid_token);
374 mtype = mp->mnt_vfc->vfc_typenum;
375 tfsid.val[1] = mtype;
376 mtype = (mtype & 0xFF) << 24;
378 tfsid.val[0] = makeudev(255,
379 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
381 if (vfs_getvfs(&tfsid) == NULL)
384 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
385 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
386 lwkt_reltoken(&mntid_token);
390 * Set the FSID for a new mount point to the template. Adjust
391 * the FSID to avoid collisions.
394 vfs_setfsid(struct mount *mp, fsid_t *template)
398 bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid));
400 if (vfs_getvfs(template) == NULL)
405 mp->mnt_stat.f_fsid = *template;
410 * This routine is called when we have too many vnodes. It attempts
411 * to free <count> vnodes and will potentially free vnodes that still
412 * have VM backing store (VM backing store is typically the cause
413 * of a vnode blowout so we want to do this). Therefore, this operation
414 * is not considered cheap.
416 * A number of conditions may prevent a vnode from being reclaimed.
417 * the buffer cache may have references on the vnode, a directory
418 * vnode may still have references due to the namei cache representing
419 * underlying files, or the vnode may be in active use. It is not
420 * desireable to reuse such vnodes. These conditions may cause the
421 * number of vnodes to reach some minimum value regardless of what
422 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
426 * This is a quick non-blocking check to determine if the vnode is a good
427 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is
428 * not a good candidate, 1 if it is.
431 vmightfree(struct vnode *vp, int page_count, int pass)
433 if (vp->v_flag & VRECLAIMED)
436 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
439 if (sysref_isactive(&vp->v_sysref))
441 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
445 * XXX horrible hack. Up to four passes will be taken. Each pass
446 * makes a larger set of vnodes eligible. For now what this really
447 * means is that we try to recycle files opened only once before
448 * recycling files opened multiple times.
450 switch(vp->v_flag & (VAGE0 | VAGE1)) {
470 * The vnode was found to be possibly vgone()able and the caller has locked it
471 * (thus the usecount should be 1 now). Determine if the vnode is actually
472 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode
473 * can be vgone()'d, 0 otherwise.
475 * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf
476 * in the namecache topology and (B) this vnode has buffer cache bufs.
477 * We cannot remove vnodes with non-leaf namecache associations. We do a
478 * tentitive leaf check prior to attempting to flush out any buffers but the
479 * 'real' test when all is said in done is that v_auxrefs must become 0 for
480 * the vnode to be freeable.
482 * We could theoretically just unconditionally flush when v_auxrefs != 0,
483 * but flushing data associated with non-leaf nodes (which are always
484 * directories), just throws it away for no benefit. It is the buffer
485 * cache's responsibility to choose buffers to recycle from the cached
486 * data point of view.
489 visleaf(struct vnode *vp)
491 struct namecache *ncp;
493 spin_lock(&vp->v_spin);
494 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
495 if (!TAILQ_EMPTY(&ncp->nc_list)) {
496 spin_unlock(&vp->v_spin);
500 spin_unlock(&vp->v_spin);
505 * Try to clean up the vnode to the point where it can be vgone()'d, returning
506 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike
507 * vmightfree() this routine may flush the vnode and block. Vnodes marked
508 * VFREE are still candidates for vgone()ing because they may hold namecache
509 * resources and could be blocking the namecache directory hierarchy (and
510 * related vnodes) from being freed.
513 vtrytomakegoneable(struct vnode *vp, int page_count)
515 if (vp->v_flag & VRECLAIMED)
517 if (vp->v_sysref.refcnt > 1)
519 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
521 if (vp->v_auxrefs && visleaf(vp)) {
522 vinvalbuf(vp, V_SAVE, 0, 0);
524 kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" :
525 "vrecycle: vp %p succeeded: %s\n"), vp,
526 (TAILQ_FIRST(&vp->v_namecache) ?
527 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
532 * This sequence may seem a little strange, but we need to optimize
533 * the critical path a bit. We can't recycle vnodes with other
534 * references and because we are trying to recycle an otherwise
535 * perfectly fine vnode we have to invalidate the namecache in a
536 * way that avoids possible deadlocks (since the vnode lock is being
537 * held here). Finally, we have to check for other references one
538 * last time in case something snuck in during the inval.
540 if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0)
542 if (cache_inval_vp_nonblock(vp))
544 return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0);
548 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try
549 * to avoid vnodes which have lots of resident pages (we are trying to free
550 * vnodes, not memory).
552 * This routine is a callback from the mountlist scan. The mount point
553 * in question will be busied.
555 * NOTE: The 1/10 reclamation also ensures that the inactive data set
556 * (the vnodes being recycled by the one-time use) does not degenerate
557 * into too-small a set. This is important because once a vnode is
558 * marked as not being one-time-use (VAGE0/VAGE1 both 0) that vnode
559 * will not be destroyed EXCEPT by this mechanism. VM pages can still
560 * be cleaned/freed by the pageout daemon.
563 vlrureclaim(struct mount *mp, void *data)
565 struct vnlru_info *info = data;
571 int trigger_mult = vnlru_nowhere;
574 * Calculate the trigger point for the resident pages check. The
575 * minimum trigger value is approximately the number of pages in
576 * the system divded by the number of vnodes. However, due to
577 * various other system memory overheads unrelated to data caching
578 * it is a good idea to double the trigger (at least).
580 * trigger_mult starts at 0. If the recycler is having problems
581 * finding enough freeable vnodes it will increase trigger_mult.
582 * This should not happen in normal operation, even on machines with
583 * low amounts of memory, but extraordinary memory use by the system
584 * verses the amount of cached data can trigger it.
586 usevnodes = desiredvnodes;
589 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
592 lwkt_gettoken(&mntvnode_token);
593 count = mp->mnt_nvnodelistsize / 10 + 1;
595 while (count && mp->mnt_syncer) {
597 * Next vnode. Use the special syncer vnode to placemark
598 * the LRU. This way the LRU code does not interfere with
601 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
602 TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes);
604 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp,
605 mp->mnt_syncer, v_nmntvnodes);
607 TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer,
609 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
617 * The VP will stick around while we hold mntvnode_token,
618 * at least until we block, so we can safely do an initial
619 * check, and then must check again after we lock the vnode.
621 if (vp->v_type == VNON || /* syncer or indeterminant */
622 !vmightfree(vp, trigger, info->pass) /* critical path opt */
629 * VX get the candidate vnode. If the VX get fails the
630 * vnode might still be on the mountlist. Our loop depends
631 * on us at least cycling the vnode to the end of the
634 if (vx_get_nonblock(vp) != 0) {
640 * Since we blocked locking the vp, make sure it is still
641 * a candidate for reclamation. That is, it has not already
642 * been reclaimed and only has our VX reference associated
645 if (vp->v_type == VNON || /* syncer or indeterminant */
646 (vp->v_flag & VRECLAIMED) ||
648 !vtrytomakegoneable(vp, trigger) /* critical path opt */
656 * All right, we are good, move the vp to the end of the
657 * mountlist and clean it out. The vget will have returned
658 * an error if the vnode was destroyed (VRECLAIMED set), so we
659 * do not have to check again. The vput() will move the
660 * vnode to the free list if the vgone() was successful.
662 KKASSERT(vp->v_mount == mp);
668 lwkt_reltoken(&mntvnode_token);
673 * Attempt to recycle vnodes in a context that is always safe to block.
674 * Calling vlrurecycle() from the bowels of file system code has some
675 * interesting deadlock problems.
677 static struct thread *vnlruthread;
678 static int vnlruproc_sig;
681 vnlru_proc_wait(void)
683 tsleep_interlock(&vnlruproc_sig, 0);
684 if (vnlruproc_sig == 0) {
685 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
688 tsleep(&vnlruproc_sig, PINTERLOCKED, "vlruwk", hz);
694 struct thread *td = curthread;
695 struct vnlru_info info;
698 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
705 kproc_suspend_loop();
708 * Try to free some vnodes if we have too many
710 if (numvnodes > desiredvnodes &&
711 freevnodes > desiredvnodes * 2 / 10) {
712 int count = numvnodes - desiredvnodes;
714 if (count > freevnodes / 100)
715 count = freevnodes / 100;
718 freesomevnodes(count);
722 * Nothing to do if most of our vnodes are already on
725 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
727 wakeup(&vnlruproc_sig);
728 tsleep(vnlruthread, 0, "vlruwt", hz);
734 * The pass iterates through the four combinations of
735 * VAGE0/VAGE1. We want to get rid of aged small files
740 while (done == 0 && info.pass < 4) {
741 done = mountlist_scan(vlrureclaim, &info,
747 * The vlrureclaim() call only processes 1/10 of the vnodes
748 * on each mount. If we couldn't find any repeat the loop
749 * at least enough times to cover all available vnodes before
750 * we start sleeping. Complain if the failure extends past
751 * 30 second, every 30 seconds.
755 if (vnlru_nowhere % 10 == 0)
756 tsleep(vnlruthread, 0, "vlrup", hz * 3);
757 if (vnlru_nowhere % 100 == 0)
758 kprintf("vnlru_proc: vnode recycler stopped working!\n");
759 if (vnlru_nowhere == 1000)
771 * MOUNTLIST FUNCTIONS
775 * mountlist_insert (MP SAFE)
777 * Add a new mount point to the mount list.
780 mountlist_insert(struct mount *mp, int how)
782 lwkt_gettoken(&mountlist_token);
783 if (how == MNTINS_FIRST)
784 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
786 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
787 lwkt_reltoken(&mountlist_token);
791 * mountlist_interlock (MP SAFE)
793 * Execute the specified interlock function with the mountlist token
794 * held. The function will be called in a serialized fashion verses
795 * other functions called through this mechanism.
798 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
802 lwkt_gettoken(&mountlist_token);
803 error = callback(mp);
804 lwkt_reltoken(&mountlist_token);
809 * mountlist_boot_getfirst (DURING BOOT ONLY)
811 * This function returns the first mount on the mountlist, which is
812 * expected to be the root mount. Since no interlocks are obtained
813 * this function is only safe to use during booting.
817 mountlist_boot_getfirst(void)
819 return(TAILQ_FIRST(&mountlist));
823 * mountlist_remove (MP SAFE)
825 * Remove a node from the mountlist. If this node is the next scan node
826 * for any active mountlist scans, the active mountlist scan will be
827 * adjusted to skip the node, thus allowing removals during mountlist
831 mountlist_remove(struct mount *mp)
833 struct mountscan_info *msi;
835 lwkt_gettoken(&mountlist_token);
836 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
837 if (msi->msi_node == mp) {
838 if (msi->msi_how & MNTSCAN_FORWARD)
839 msi->msi_node = TAILQ_NEXT(mp, mnt_list);
841 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
844 TAILQ_REMOVE(&mountlist, mp, mnt_list);
845 lwkt_reltoken(&mountlist_token);
849 * mountlist_scan (MP SAFE)
851 * Safely scan the mount points on the mount list. Unless otherwise
852 * specified each mount point will be busied prior to the callback and
853 * unbusied afterwords. The callback may safely remove any mount point
854 * without interfering with the scan. If the current callback
855 * mount is removed the scanner will not attempt to unbusy it.
857 * If a mount node cannot be busied it is silently skipped.
859 * The callback return value is aggregated and a total is returned. A return
860 * value of < 0 is not aggregated and will terminate the scan.
862 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction
863 * MNTSCAN_REVERSE - the mountlist is scanned in reverse
864 * MNTSCAN_NOBUSY - the scanner will make the callback without busying
868 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
870 struct mountscan_info info;
875 lwkt_gettoken(&mountlist_token);
878 info.msi_node = NULL; /* paranoia */
879 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
883 if (how & MNTSCAN_FORWARD) {
884 info.msi_node = TAILQ_FIRST(&mountlist);
885 while ((mp = info.msi_node) != NULL) {
886 if (how & MNTSCAN_NOBUSY) {
887 count = callback(mp, data);
888 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
889 count = callback(mp, data);
890 if (mp == info.msi_node)
898 if (mp == info.msi_node)
899 info.msi_node = TAILQ_NEXT(mp, mnt_list);
901 } else if (how & MNTSCAN_REVERSE) {
902 info.msi_node = TAILQ_LAST(&mountlist, mntlist);
903 while ((mp = info.msi_node) != NULL) {
904 if (how & MNTSCAN_NOBUSY) {
905 count = callback(mp, data);
906 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
907 count = callback(mp, data);
908 if (mp == info.msi_node)
916 if (mp == info.msi_node)
917 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
920 TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
921 lwkt_reltoken(&mountlist_token);
926 * MOUNT RELATED VNODE FUNCTIONS
929 static struct kproc_desc vnlru_kp = {
934 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
937 * Move a vnode from one mount queue to another.
942 insmntque(struct vnode *vp, struct mount *mp)
944 lwkt_gettoken(&mntvnode_token);
946 * Delete from old mount point vnode list, if on one.
948 if (vp->v_mount != NULL) {
949 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0,
950 ("bad mount point vnode list size"));
952 vp->v_mount->mnt_nvnodelistsize--;
955 * Insert into list of vnodes for the new mount point, if available.
956 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer.
958 if ((vp->v_mount = mp) == NULL) {
959 lwkt_reltoken(&mntvnode_token);
962 if (mp->mnt_syncer) {
963 TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes);
965 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
967 mp->mnt_nvnodelistsize++;
968 lwkt_reltoken(&mntvnode_token);
973 * Scan the vnodes under a mount point and issue appropriate callbacks.
975 * The fastfunc() callback is called with just the mountlist token held
976 * (no vnode lock). It may not block and the vnode may be undergoing
977 * modifications while the caller is processing it. The vnode will
978 * not be entirely destroyed, however, due to the fact that the mountlist
979 * token is held. A return value < 0 skips to the next vnode without calling
980 * the slowfunc(), a return value > 0 terminates the loop.
982 * The slowfunc() callback is called after the vnode has been successfully
983 * locked based on passed flags. The vnode is skipped if it gets rearranged
984 * or destroyed while blocking on the lock. A non-zero return value from
985 * the slow function terminates the loop. The slow function is allowed to
986 * arbitrarily block. The scanning code guarentees consistency of operation
987 * even if the slow function deletes or moves the node, or blocks and some
988 * other thread deletes or moves the node.
990 * NOTE: We hold vmobj_token to prevent a VM object from being destroyed
991 * out from under the fastfunc()'s vnode test. It will not prevent
992 * v_object from getting NULL'd out but it will ensure that the
993 * pointer (if we race) will remain stable.
999 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
1000 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
1003 struct vmntvnodescan_info info;
1006 int maxcount = mp->mnt_nvnodelistsize * 2;
1010 lwkt_gettoken(&mntvnode_token);
1011 lwkt_gettoken(&vmobj_token);
1014 * If asked to do one pass stop after iterating available vnodes.
1015 * Under heavy loads new vnodes can be added while we are scanning,
1016 * so this isn't perfect. Create a slop factor of 2x.
1018 if (flags & VMSC_ONEPASS)
1019 stopcount = mp->mnt_nvnodelistsize;
1021 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1022 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry);
1023 while ((vp = info.vp) != NULL) {
1024 if (--maxcount == 0) {
1025 kprintf("Warning: excessive fssync iteration\n");
1026 maxcount = mp->mnt_nvnodelistsize * 2;
1030 * Skip if visible but not ready, or special (e.g.
1033 if (vp->v_type == VNON)
1035 KKASSERT(vp->v_mount == mp);
1038 * Quick test. A negative return continues the loop without
1039 * calling the slow test. 0 continues onto the slow test.
1040 * A positive number aborts the loop.
1043 if ((r = fastfunc(mp, vp, data)) < 0) {
1052 * Get a vxlock on the vnode, retry if it has moved or isn't
1053 * in the mountlist where we expect it.
1058 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1060 error = vget(vp, LK_EXCLUSIVE);
1062 case VMSC_GETVP|VMSC_NOWAIT:
1063 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT);
1076 * Do not call the slow function if the vnode is
1077 * invalid or if it was ripped out from under us
1078 * while we (potentially) blocked.
1080 if (info.vp == vp && vp->v_type != VNON)
1081 r = slowfunc(mp, vp, data);
1086 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1088 case VMSC_GETVP|VMSC_NOWAIT:
1103 * Yield after some processing. Depending on the number
1104 * of vnodes, we might wind up running for a long time.
1105 * Because threads are not preemptable, time critical
1106 * userland processes might starve. Give them a chance
1109 if (++count == 10000) {
1110 /* We really want to yield a bit, so we simply sleep a tick */
1111 tsleep(mp, 0, "vnodescn", 1);
1116 * If doing one pass this decrements to zero. If it starts
1117 * at zero it is effectively unlimited for the purposes of
1120 if (--stopcount == 0)
1124 * Iterate. If the vnode was ripped out from under us
1125 * info.vp will already point to the next vnode, otherwise
1126 * we have to obtain the next valid vnode ourselves.
1129 info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
1131 TAILQ_REMOVE(&mntvnodescan_list, &info, entry);
1132 lwkt_reltoken(&vmobj_token);
1133 lwkt_reltoken(&mntvnode_token);
1138 * Remove any vnodes in the vnode table belonging to mount point mp.
1140 * If FORCECLOSE is not specified, there should not be any active ones,
1141 * return error if any are found (nb: this is a user error, not a
1142 * system error). If FORCECLOSE is specified, detach any active vnodes
1145 * If WRITECLOSE is set, only flush out regular file vnodes open for
1148 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
1150 * `rootrefs' specifies the base reference count for the root vnode
1151 * of this filesystem. The root vnode is considered busy if its
1152 * v_sysref.refcnt exceeds this value. On a successful return, vflush()
1153 * will call vrele() on the root vnode exactly rootrefs times.
1154 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
1158 static int busyprt = 0; /* print out busy vnodes */
1159 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
1162 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
1164 struct vflush_info {
1171 vflush(struct mount *mp, int rootrefs, int flags)
1173 struct thread *td = curthread; /* XXX */
1174 struct vnode *rootvp = NULL;
1176 struct vflush_info vflush_info;
1179 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1180 ("vflush: bad args"));
1182 * Get the filesystem root vnode. We can vput() it
1183 * immediately, since with rootrefs > 0, it won't go away.
1185 if ((error = VFS_ROOT(mp, &rootvp)) != 0) {
1186 if ((flags & FORCECLOSE) == 0)
1189 /* continue anyway */
1195 vflush_info.busy = 0;
1196 vflush_info.flags = flags;
1197 vflush_info.td = td;
1198 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1200 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1202 * If just the root vnode is busy, and if its refcount
1203 * is equal to `rootrefs', then go ahead and kill it.
1205 KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
1206 KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs"));
1207 if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) {
1209 vgone_vxlocked(rootvp);
1211 vflush_info.busy = 0;
1214 if (vflush_info.busy)
1216 for (; rootrefs > 0; rootrefs--)
1222 * The scan callback is made with an VX locked vnode.
1225 vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1227 struct vflush_info *info = data;
1231 * Skip over a vnodes marked VSYSTEM.
1233 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1238 * If WRITECLOSE is set, flush out unlinked but still open
1239 * files (even if open only for reading) and regular file
1240 * vnodes open for writing.
1242 if ((info->flags & WRITECLOSE) &&
1243 (vp->v_type == VNON ||
1244 (VOP_GETATTR(vp, &vattr) == 0 &&
1245 vattr.va_nlink > 0)) &&
1246 (vp->v_writecount == 0 || vp->v_type != VREG)) {
1251 * If we are the only holder (refcnt of 1) or the vnode is in
1252 * termination (refcnt < 0), we can vgone the vnode.
1254 if (vp->v_sysref.refcnt <= 1) {
1260 * If FORCECLOSE is set, forcibly destroy the vnode and then move
1261 * it to a dummymount structure so vop_*() functions don't deref
1264 if (info->flags & FORCECLOSE) {
1267 if (vp->v_mount == NULL)
1268 insmntque(vp, &dummymount);
1274 vprint("vflush: busy vnode", vp);
1281 add_bio_ops(struct bio_ops *ops)
1283 TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry);
1287 rem_bio_ops(struct bio_ops *ops)
1289 TAILQ_REMOVE(&bio_ops_list, ops, entry);
1293 * This calls the bio_ops io_sync function either for a mount point
1296 * WARNING: softdeps is weirdly coded and just isn't happy unless
1297 * io_sync is called with a NULL mount from the general syncing code.
1300 bio_ops_sync(struct mount *mp)
1302 struct bio_ops *ops;
1305 if ((ops = mp->mnt_bioops) != NULL)
1308 TAILQ_FOREACH(ops, &bio_ops_list, entry) {
1315 * Lookup a mount point by nch
1318 mount_get_by_nc(struct namecache *ncp)
1320 struct mount *mp = NULL;
1322 lwkt_gettoken(&mountlist_token);
1323 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
1324 if (ncp == mp->mnt_ncmountpt.ncp)
1327 lwkt_reltoken(&mountlist_token);