2 * Copyright (c) 2003,2004 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, 1995
35 * The Regents of the University of California. All rights reserved.
37 * This code is derived from software contributed to Berkeley by
38 * Poul-Henning Kamp of the FreeBSD Project.
40 * Redistribution and use in source and binary forms, with or without
41 * modification, are permitted provided that the following conditions
43 * 1. Redistributions of source code must retain the above copyright
44 * notice, this list of conditions and the following disclaimer.
45 * 2. Redistributions in binary form must reproduce the above copyright
46 * notice, this list of conditions and the following disclaimer in the
47 * documentation and/or other materials provided with the distribution.
48 * 3. All advertising materials mentioning features or use of this software
49 * must display the following acknowledgement:
50 * This product includes software developed by the University of
51 * California, Berkeley and its contributors.
52 * 4. Neither the name of the University nor the names of its contributors
53 * may be used to endorse or promote products derived from this software
54 * without specific prior written permission.
56 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
68 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95
69 * $FreeBSD: src/sys/kern/vfs_cache.c,v 1.42.2.6 2001/10/05 20:07:03 dillon Exp $
70 * $DragonFly: src/sys/kern/vfs_cache.c,v 1.70 2006/06/02 04:59:52 dillon Exp $
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mount.h>
78 #include <sys/vnode.h>
79 #include <sys/malloc.h>
80 #include <sys/sysproto.h>
82 #include <sys/namei.h>
83 #include <sys/nlookup.h>
84 #include <sys/filedesc.h>
85 #include <sys/fnv_hash.h>
86 #include <sys/globaldata.h>
87 #include <sys/kern_syscall.h>
88 #include <sys/dirent.h>
92 * Random lookups in the cache are accomplished with a hash table using
93 * a hash key of (nc_src_vp, name).
95 * Negative entries may exist and correspond to structures where nc_vp
96 * is NULL. In a negative entry, NCF_WHITEOUT will be set if the entry
97 * corresponds to a whited-out directory entry (verses simply not finding the
100 * Upon reaching the last segment of a path, if the reference is for DELETE,
101 * or NOCACHE is set (rewrite), and the name is located in the cache, it
106 * Structures associated with name cacheing.
108 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash])
111 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
113 static LIST_HEAD(nchashhead, namecache) *nchashtbl; /* Hash Table */
114 static struct namecache_list ncneglist; /* instead of vnode */
117 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server
118 * to create the namecache infrastructure leading to a dangling vnode.
120 * 0 Only errors are reported
121 * 1 Successes are reported
122 * 2 Successes + the whole directory scan is reported
123 * 3 Force the directory scan code run as if the parent vnode did not
124 * have a namecache record, even if it does have one.
126 static int ncvp_debug;
127 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, "");
129 static u_long nchash; /* size of hash table */
130 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, "");
132 static u_long ncnegfactor = 16; /* ratio of negative entries */
133 SYSCTL_ULONG(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, "");
135 static int nclockwarn; /* warn on locked entries in ticks */
136 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, "");
138 static u_long numneg; /* number of cache entries allocated */
139 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, "");
141 static u_long numcache; /* number of cache entries allocated */
142 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, "");
144 static u_long numunres; /* number of unresolved entries */
145 SYSCTL_ULONG(_debug, OID_AUTO, numunres, CTLFLAG_RD, &numunres, 0, "");
147 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), "");
148 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), "");
150 static int cache_resolve_mp(struct namecache *ncp);
151 static void cache_rehash(struct namecache *ncp);
154 * The new name cache statistics
156 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
157 #define STATNODE(mode, name, var) \
158 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, "");
159 STATNODE(CTLFLAG_RD, numneg, &numneg);
160 STATNODE(CTLFLAG_RD, numcache, &numcache);
161 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls);
162 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits);
163 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits);
164 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks);
165 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss);
166 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap);
167 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps);
168 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits);
169 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps);
170 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits);
172 struct nchstats nchstats[SMP_MAXCPU];
174 * Export VFS cache effectiveness statistics to user-land.
176 * The statistics are left for aggregation to user-land so
177 * neat things can be achieved, like observing per-CPU cache
181 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
183 struct globaldata *gd;
187 for (i = 0; i < ncpus; ++i) {
188 gd = globaldata_find(i);
189 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
190 sizeof(struct nchstats))))
196 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
197 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
199 static void cache_zap(struct namecache *ncp);
202 * cache_hold() and cache_drop() prevent the premature deletion of a
203 * namecache entry but do not prevent operations (such as zapping) on
204 * that namecache entry.
206 * This routine may only be called from outside this source module if
207 * nc_refs is already at least 1.
209 * This is a rare case where callers are allowed to hold a spinlock,
210 * so we can't ourselves.
214 _cache_hold(struct namecache *ncp)
216 atomic_add_int(&ncp->nc_refs, 1);
221 * When dropping an entry, if only one ref remains and the entry has not
222 * been resolved, zap it. Since the one reference is being dropped the
223 * entry had better not be locked.
227 _cache_drop(struct namecache *ncp)
229 KKASSERT(ncp->nc_refs > 0);
230 if (ncp->nc_refs == 1 &&
231 (ncp->nc_flag & NCF_UNRESOLVED) &&
232 TAILQ_EMPTY(&ncp->nc_list)
234 KKASSERT(ncp->nc_exlocks == 0);
238 atomic_subtract_int(&ncp->nc_refs, 1);
243 * Link a new namecache entry to its parent. Be careful to avoid races
244 * if vhold() blocks in the future.
246 * If we are creating a child under an oldapi parent we must mark the
247 * child as being an oldapi entry as well.
250 cache_link_parent(struct namecache *ncp, struct namecache *par)
252 KKASSERT(ncp->nc_parent == NULL);
253 ncp->nc_parent = par;
254 if (TAILQ_EMPTY(&par->nc_list)) {
255 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
257 * Any vp associated with an ncp which has children must
258 * be held to prevent it from being recycled.
263 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
268 * Remove the parent association from a namecache structure. If this is
269 * the last child of the parent the cache_drop(par) will attempt to
270 * recursively zap the parent.
273 cache_unlink_parent(struct namecache *ncp)
275 struct namecache *par;
277 if ((par = ncp->nc_parent) != NULL) {
278 ncp->nc_parent = NULL;
279 par = cache_hold(par);
280 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
281 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
288 * Allocate a new namecache structure. Most of the code does not require
289 * zero-termination of the string but it makes vop_compat_ncreate() easier.
291 static struct namecache *
292 cache_alloc(int nlen)
294 struct namecache *ncp;
296 ncp = malloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
298 ncp->nc_name = malloc(nlen + 1, M_VFSCACHE, M_WAITOK);
300 ncp->nc_flag = NCF_UNRESOLVED;
301 ncp->nc_error = ENOTCONN; /* needs to be resolved */
305 * Construct a fake FSMID based on the time of day and a 32 bit
306 * roller for uniqueness. This is used to generate a useful
307 * FSMID for filesystems which do not support it.
309 ncp->nc_fsmid = cache_getnewfsmid();
310 TAILQ_INIT(&ncp->nc_list);
316 cache_free(struct namecache *ncp)
318 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
320 free(ncp->nc_name, M_VFSCACHE);
321 free(ncp, M_VFSCACHE);
325 * Ref and deref a namecache structure.
327 * Warning: caller may hold an unrelated read spinlock, which means we can't
328 * use read spinlocks here.
331 cache_hold(struct namecache *ncp)
333 return(_cache_hold(ncp));
337 cache_drop(struct namecache *ncp)
343 * Namespace locking. The caller must already hold a reference to the
344 * namecache structure in order to lock/unlock it. This function prevents
345 * the namespace from being created or destroyed by accessors other then
348 * Note that holding a locked namecache structure prevents other threads
349 * from making namespace changes (e.g. deleting or creating), prevents
350 * vnode association state changes by other threads, and prevents the
351 * namecache entry from being resolved or unresolved by other threads.
353 * The lock owner has full authority to associate/disassociate vnodes
354 * and resolve/unresolve the locked ncp.
356 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed
357 * or recycled, but it does NOT help you if the vnode had already initiated
358 * a recyclement. If this is important, use cache_get() rather then
359 * cache_lock() (and deal with the differences in the way the refs counter
360 * is handled). Or, alternatively, make an unconditional call to
361 * cache_validate() or cache_resolve() after cache_lock() returns.
364 cache_lock(struct namecache *ncp)
369 KKASSERT(ncp->nc_refs != 0);
374 if (ncp->nc_exlocks == 0) {
378 * The vp associated with a locked ncp must be held
379 * to prevent it from being recycled (which would
380 * cause the ncp to become unresolved).
382 * WARNING! If VRECLAIMED is set the vnode could
383 * already be in the middle of a recycle. Callers
384 * should not assume that nc_vp is usable when
385 * not NULL. cache_vref() or cache_vget() must be
388 * XXX loop on race for later MPSAFE work.
394 if (ncp->nc_locktd == td) {
398 ncp->nc_flag |= NCF_LOCKREQ;
399 if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) {
403 printf("[diagnostic] cache_lock: blocked on %p", ncp);
404 if ((ncp->nc_flag & NCF_MOUNTPT) && ncp->nc_mount)
405 printf(" [MOUNTFROM %s]\n", ncp->nc_mount->mnt_stat.f_mntfromname);
407 printf(" \"%*.*s\"\n",
408 ncp->nc_nlen, ncp->nc_nlen,
414 printf("[diagnostic] cache_lock: unblocked %*.*s\n",
415 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
420 cache_lock_nonblock(struct namecache *ncp)
424 KKASSERT(ncp->nc_refs != 0);
426 if (ncp->nc_exlocks == 0) {
430 * The vp associated with a locked ncp must be held
431 * to prevent it from being recycled (which would
432 * cause the ncp to become unresolved).
434 * WARNING! If VRECLAIMED is set the vnode could
435 * already be in the middle of a recycle. Callers
436 * should not assume that nc_vp is usable when
437 * not NULL. cache_vref() or cache_vget() must be
440 * XXX loop on race for later MPSAFE work.
451 cache_unlock(struct namecache *ncp)
453 thread_t td = curthread;
455 KKASSERT(ncp->nc_refs > 0);
456 KKASSERT(ncp->nc_exlocks > 0);
457 KKASSERT(ncp->nc_locktd == td);
458 if (--ncp->nc_exlocks == 0) {
461 ncp->nc_locktd = NULL;
462 if (ncp->nc_flag & NCF_LOCKREQ) {
463 ncp->nc_flag &= ~NCF_LOCKREQ;
470 * ref-and-lock, unlock-and-deref functions.
472 * This function is primarily used by nlookup. Even though cache_lock
473 * holds the vnode, it is possible that the vnode may have already
474 * initiated a recyclement. We want cache_get() to return a definitively
475 * usable vnode or a definitively unresolved ncp.
478 cache_get(struct namecache *ncp)
482 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
483 cache_setunresolved(ncp);
488 cache_get_nonblock(struct namecache *ncp)
491 if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) {
494 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
495 cache_setunresolved(ncp);
502 cache_put(struct namecache *ncp)
509 * Resolve an unresolved ncp by associating a vnode with it. If the
510 * vnode is NULL, a negative cache entry is created.
512 * The ncp should be locked on entry and will remain locked on return.
515 cache_setvp(struct namecache *ncp, struct vnode *vp)
517 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
521 * Any vp associated with an ncp which has children must
522 * be held. Any vp associated with a locked ncp must be held.
524 if (!TAILQ_EMPTY(&ncp->nc_list))
526 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
531 * Set auxillary flags
535 ncp->nc_flag |= NCF_ISDIR;
538 ncp->nc_flag |= NCF_ISSYMLINK;
539 /* XXX cache the contents of the symlink */
547 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
549 ncp->nc_error = ENOENT;
551 ncp->nc_flag &= ~NCF_UNRESOLVED;
555 cache_settimeout(struct namecache *ncp, int nticks)
557 if ((ncp->nc_timeout = ticks + nticks) == 0)
562 * Disassociate the vnode or negative-cache association and mark a
563 * namecache entry as unresolved again. Note that the ncp is still
564 * left in the hash table and still linked to its parent.
566 * The ncp should be locked and refd on entry and will remain locked and refd
569 * This routine is normally never called on a directory containing children.
570 * However, NFS often does just that in its rename() code as a cop-out to
571 * avoid complex namespace operations. This disconnects a directory vnode
572 * from its namecache and can cause the OLDAPI and NEWAPI to get out of
575 * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as
576 * in a create, properly propogates flag up the chain.
579 cache_setunresolved(struct namecache *ncp)
583 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
584 ncp->nc_flag |= NCF_UNRESOLVED;
586 ncp->nc_error = ENOTCONN;
588 if ((vp = ncp->nc_vp) != NULL) {
591 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
594 * Any vp associated with an ncp with children is
595 * held by that ncp. Any vp associated with a locked
596 * ncp is held by that ncp. These conditions must be
597 * undone when the vp is cleared out from the ncp.
599 if (ncp->nc_flag & NCF_FSMID)
601 if (!TAILQ_EMPTY(&ncp->nc_list))
606 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
609 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK|
615 * Invalidate portions of the namecache topology given a starting entry.
616 * The passed ncp is set to an unresolved state and:
618 * The passed ncp must be locked.
620 * CINV_DESTROY - Set a flag in the passed ncp entry indicating
621 * that the physical underlying nodes have been
622 * destroyed... as in deleted. For example, when
623 * a directory is removed. This will cause record
624 * lookups on the name to no longer be able to find
625 * the record and tells the resolver to return failure
626 * rather then trying to resolve through the parent.
628 * The topology itself, including ncp->nc_name,
631 * This only applies to the passed ncp, if CINV_CHILDREN
632 * is specified the children are not flagged.
634 * CINV_CHILDREN - Set all children (recursively) to an unresolved
637 * Note that this will also have the side effect of
638 * cleaning out any unreferenced nodes in the topology
639 * from the leaves up as the recursion backs out.
641 * Note that the topology for any referenced nodes remains intact.
643 * It is possible for cache_inval() to race a cache_resolve(), meaning that
644 * the namecache entry may not actually be invalidated on return if it was
645 * revalidated while recursing down into its children. This code guarentees
646 * that the node(s) will go through an invalidation cycle, but does not
647 * guarentee that they will remain in an invalidated state.
649 * Returns non-zero if a revalidation was detected during the invalidation
650 * recursion, zero otherwise. Note that since only the original ncp is
651 * locked the revalidation ultimately can only indicate that the original ncp
652 * *MIGHT* no have been reresolved.
655 cache_inval(struct namecache *ncp, int flags)
657 struct namecache *kid;
658 struct namecache *nextkid;
661 KKASSERT(ncp->nc_exlocks);
663 cache_setunresolved(ncp);
664 if (flags & CINV_DESTROY)
665 ncp->nc_flag |= NCF_DESTROYED;
667 if ((flags & CINV_CHILDREN) &&
668 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
673 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
675 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
676 TAILQ_FIRST(&kid->nc_list)
679 rcnt += cache_inval(kid, flags & ~CINV_DESTROY);
689 * Someone could have gotten in there while ncp was unlocked,
692 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
698 * Invalidate a vnode's namecache associations. To avoid races against
699 * the resolver we do not invalidate a node which we previously invalidated
700 * but which was then re-resolved while we were in the invalidation loop.
702 * Returns non-zero if any namecache entries remain after the invalidation
705 * NOTE: unlike the namecache topology which guarentees that ncp's will not
706 * be ripped out of the topology while held, the vnode's v_namecache list
707 * has no such restriction. NCP's can be ripped out of the list at virtually
708 * any time if not locked, even if held.
711 cache_inval_vp(struct vnode *vp, int flags)
713 struct namecache *ncp;
714 struct namecache *next;
717 ncp = TAILQ_FIRST(&vp->v_namecache);
721 /* loop entered with ncp held */
722 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
725 if (ncp->nc_vp != vp) {
726 printf("Warning: cache_inval_vp: race-A detected on "
727 "%s\n", ncp->nc_name);
733 cache_inval(ncp, flags);
734 cache_put(ncp); /* also releases reference */
736 if (ncp && ncp->nc_vp != vp) {
737 printf("Warning: cache_inval_vp: race-B detected on "
738 "%s\n", ncp->nc_name);
743 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
747 * The source ncp has been renamed to the target ncp. Both fncp and tncp
748 * must be locked. Both will be set to unresolved, any children of tncp
749 * will be disconnected (the prior contents of the target is assumed to be
750 * destroyed by the rename operation, e.g. renaming over an empty directory),
751 * and all children of fncp will be moved to tncp.
753 * XXX the disconnection could pose a problem, check code paths to make
754 * sure any code that blocks can handle the parent being changed out from
755 * under it. Maybe we should lock the children (watch out for deadlocks) ?
757 * After we return the caller has the option of calling cache_setvp() if
758 * the vnode of the new target ncp is known.
760 * Any process CD'd into any of the children will no longer be able to ".."
761 * back out. An rm -rf can cause this situation to occur.
764 cache_rename(struct namecache *fncp, struct namecache *tncp)
766 struct namecache *scan;
769 cache_setunresolved(fncp);
770 cache_setunresolved(tncp);
771 while (cache_inval(tncp, CINV_CHILDREN) != 0) {
772 if (didwarn++ % 10 == 0) {
773 printf("Warning: cache_rename: race during "
775 fncp->nc_name, tncp->nc_name);
777 tsleep(tncp, 0, "mvrace", hz / 10);
778 cache_setunresolved(tncp);
780 while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) {
782 cache_unlink_parent(scan);
783 cache_link_parent(scan, tncp);
784 if (scan->nc_flag & NCF_HASHED)
791 * vget the vnode associated with the namecache entry. Resolve the namecache
792 * entry if necessary and deal with namecache/vp races. The passed ncp must
793 * be referenced and may be locked. The ncp's ref/locking state is not
794 * effected by this call.
796 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked
797 * (depending on the passed lk_type) will be returned in *vpp with an error
798 * of 0, or NULL will be returned in *vpp with a non-0 error code. The
799 * most typical error is ENOENT, meaning that the ncp represents a negative
800 * cache hit and there is no vnode to retrieve, but other errors can occur
803 * The main race we have to deal with are namecache zaps. The ncp itself
804 * will not disappear since it is referenced, and it turns out that the
805 * validity of the vp pointer can be checked simply by rechecking the
806 * contents of ncp->nc_vp.
809 cache_vget(struct namecache *ncp, struct ucred *cred,
810 int lk_type, struct vnode **vpp)
817 if (ncp->nc_flag & NCF_UNRESOLVED) {
819 error = cache_resolve(ncp, cred);
824 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
826 * Accessing the vnode from the namecache is a bit
827 * dangerous. Because there are no refs on the vnode, it
828 * could be in the middle of a reclaim.
830 if (vp->v_flag & VRECLAIMED) {
831 printf("Warning: vnode reclaim race detected in cache_vget on %p (%s)\n", vp, ncp->nc_name);
833 cache_setunresolved(ncp);
837 error = vget(vp, lk_type);
839 if (vp != ncp->nc_vp)
842 } else if (vp != ncp->nc_vp) {
845 } else if (vp->v_flag & VRECLAIMED) {
846 panic("vget succeeded on a VRECLAIMED node! vp %p", vp);
849 if (error == 0 && vp == NULL)
856 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp)
863 if (ncp->nc_flag & NCF_UNRESOLVED) {
865 error = cache_resolve(ncp, cred);
870 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
872 * Since we did not obtain any locks, a cache zap
873 * race can occur here if the vnode is in the middle
874 * of being reclaimed and has not yet been able to
875 * clean out its cache node. If that case occurs,
876 * we must lock and unresolve the cache, then loop
879 if (vp->v_flag & VRECLAIMED) {
880 printf("Warning: vnode reclaim race detected on cache_vref %p (%s)\n", vp, ncp->nc_name);
882 cache_setunresolved(ncp);
888 if (error == 0 && vp == NULL)
895 * Recursively set the FSMID update flag for namecache nodes leading
896 * to root. This will cause the next getattr or reclaim to increment the
897 * fsmid and mark the inode for lazy updating.
899 * Stop recursing when we hit a node whos NCF_FSMID flag is already set.
900 * This makes FSMIDs work in an Einsteinian fashion - where the observation
901 * effects the result. In this case a program monitoring a higher level
902 * node will have detected some prior change and started its scan (clearing
903 * NCF_FSMID in higher level nodes), but since it has not yet observed the
904 * node where we find NCF_FSMID still set, we can safely make the related
905 * modification without interfering with the theorized program.
907 * This also means that FSMIDs cannot represent time-domain quantities
908 * in a hierarchical sense. But the main reason for doing it this way
909 * is to reduce the amount of recursion that occurs in the critical path
910 * when e.g. a program is writing to a file that sits deep in a directory
914 cache_update_fsmid(struct namecache *ncp)
917 struct namecache *scan;
920 * Warning: even if we get a non-NULL vp it could still be in the
921 * middle of a recyclement. Don't do anything fancy, just set
924 if ((vp = ncp->nc_vp) != NULL) {
925 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
926 for (scan = ncp; scan; scan = scan->nc_parent) {
927 if (scan->nc_flag & NCF_FSMID)
929 scan->nc_flag |= NCF_FSMID;
933 while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) {
934 ncp->nc_flag |= NCF_FSMID;
935 ncp = ncp->nc_parent;
941 cache_update_fsmid_vp(struct vnode *vp)
943 struct namecache *ncp;
944 struct namecache *scan;
946 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
947 for (scan = ncp; scan; scan = scan->nc_parent) {
948 if (scan->nc_flag & NCF_FSMID)
950 scan->nc_flag |= NCF_FSMID;
956 * If getattr is called on a vnode (e.g. a stat call), the filesystem
957 * may call this routine to determine if the namecache has the hierarchical
958 * change flag set, requiring the fsmid to be updated.
960 * Since 0 indicates no support, make sure the filesystem fsmid is at least
964 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid)
966 struct namecache *ncp;
969 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
970 if (ncp->nc_flag & NCF_FSMID) {
971 ncp->nc_flag &= ~NCF_FSMID;
983 * Convert a directory vnode to a namecache record without any other
984 * knowledge of the topology. This ONLY works with directory vnodes and
985 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the
986 * returned ncp (if not NULL) will be held and unlocked.
988 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
989 * If 'makeit' is 1 we attempt to track-down and create the namecache topology
990 * for dvp. This will fail only if the directory has been deleted out from
993 * Callers must always check for a NULL return no matter the value of 'makeit'.
995 * To avoid underflowing the kernel stack each recursive call increments
996 * the makeit variable.
999 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1001 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1002 struct vnode **saved_dvp);
1005 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit)
1007 struct namecache *ncp;
1008 struct vnode *saved_dvp;
1016 * Temporary debugging code to force the directory scanning code
1019 if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) {
1020 ncp = TAILQ_FIRST(&dvp->v_namecache);
1021 printf("cache_fromdvp: forcing %s\n", ncp->nc_name);
1026 * Loop until resolution, inside code will break out on error.
1028 while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) {
1031 * If dvp is the root of its filesystem it should already
1032 * have a namecache pointer associated with it as a side
1033 * effect of the mount, but it may have been disassociated.
1035 if (dvp->v_flag & VROOT) {
1036 ncp = cache_get(dvp->v_mount->mnt_ncp);
1037 error = cache_resolve_mp(ncp);
1040 printf("cache_fromdvp: resolve root of mount %p error %d",
1041 dvp->v_mount, error);
1045 printf(" failed\n");
1050 printf(" succeeded\n");
1055 * If we are recursed too deeply resort to an O(n^2)
1056 * algorithm to resolve the namecache topology. The
1057 * resolved pvp is left referenced in saved_dvp to
1058 * prevent the tree from being destroyed while we loop.
1061 error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1063 printf("lookupdotdot(longpath) failed %d "
1064 "dvp %p\n", error, dvp);
1071 * Get the parent directory and resolve its ncp.
1073 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1075 printf("lookupdotdot failed %d dvp %p\n", error, dvp);
1081 * Reuse makeit as a recursion depth counter.
1083 ncp = cache_fromdvp(pvp, cred, makeit + 1);
1089 * Do an inefficient scan of pvp (embodied by ncp) to look
1090 * for dvp. This will create a namecache record for dvp on
1091 * success. We loop up to recheck on success.
1093 * ncp and dvp are both held but not locked.
1095 error = cache_inefficient_scan(ncp, cred, dvp);
1098 printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1099 pvp, ncp->nc_name, dvp);
1104 printf("cache_fromdvp: scan %p (%s) succeeded\n",
1116 * Go up the chain of parent directories until we find something
1117 * we can resolve into the namecache. This is very inefficient.
1121 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1122 struct vnode **saved_dvp)
1124 struct namecache *ncp;
1127 static time_t last_fromdvp_report;
1130 * Loop getting the parent directory vnode until we get something we
1131 * can resolve in the namecache.
1135 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1141 if ((ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1146 if (pvp->v_flag & VROOT) {
1147 ncp = cache_get(pvp->v_mount->mnt_ncp);
1148 error = cache_resolve_mp(ncp);
1161 if (last_fromdvp_report != time_second) {
1162 last_fromdvp_report = time_second;
1163 printf("Warning: extremely inefficient path resolution on %s\n",
1166 error = cache_inefficient_scan(ncp, cred, dvp);
1169 * Hopefully dvp now has a namecache record associated with it.
1170 * Leave it referenced to prevent the kernel from recycling the
1171 * vnode. Otherwise extremely long directory paths could result
1172 * in endless recycling.
1182 * Do an inefficient scan of the directory represented by ncp looking for
1183 * the directory vnode dvp. ncp must be held but not locked on entry and
1184 * will be held on return. dvp must be refd but not locked on entry and
1185 * will remain refd on return.
1187 * Why do this at all? Well, due to its stateless nature the NFS server
1188 * converts file handles directly to vnodes without necessarily going through
1189 * the namecache ops that would otherwise create the namecache topology
1190 * leading to the vnode. We could either (1) Change the namecache algorithms
1191 * to allow disconnect namecache records that are re-merged opportunistically,
1192 * or (2) Make the NFS server backtrack and scan to recover a connected
1193 * namecache topology in order to then be able to issue new API lookups.
1195 * It turns out that (1) is a huge mess. It takes a nice clean set of
1196 * namecache algorithms and introduces a lot of complication in every subsystem
1197 * that calls into the namecache to deal with the re-merge case, especially
1198 * since we are using the namecache to placehold negative lookups and the
1199 * vnode might not be immediately assigned. (2) is certainly far less
1200 * efficient then (1), but since we are only talking about directories here
1201 * (which are likely to remain cached), the case does not actually run all
1202 * that often and has the supreme advantage of not polluting the namecache
1206 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1209 struct nlcomponent nlc;
1210 struct namecache *rncp;
1222 vat.va_blocksize = 0;
1223 if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1225 if ((error = cache_vget(ncp, cred, LK_SHARED, &pvp)) != 0)
1228 printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid);
1229 if ((blksize = vat.va_blocksize) == 0)
1230 blksize = DEV_BSIZE;
1231 rbuf = malloc(blksize, M_TEMP, M_WAITOK);
1237 iov.iov_base = rbuf;
1238 iov.iov_len = blksize;
1241 uio.uio_resid = blksize;
1242 uio.uio_segflg = UIO_SYSSPACE;
1243 uio.uio_rw = UIO_READ;
1244 uio.uio_td = curthread;
1246 if (ncvp_debug >= 2)
1247 printf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1248 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1250 den = (struct dirent *)rbuf;
1251 bytes = blksize - uio.uio_resid;
1254 if (ncvp_debug >= 2) {
1255 printf("cache_inefficient_scan: %*.*s\n",
1256 den->d_namlen, den->d_namlen,
1259 if (den->d_type != DT_WHT &&
1260 den->d_ino == vat.va_fileid) {
1262 printf("cache_inefficient_scan: "
1263 "MATCHED inode %ld path %s/%*.*s\n",
1264 vat.va_fileid, ncp->nc_name,
1265 den->d_namlen, den->d_namlen,
1268 nlc.nlc_nameptr = den->d_name;
1269 nlc.nlc_namelen = den->d_namlen;
1271 rncp = cache_nlookup(ncp, &nlc);
1272 KKASSERT(rncp != NULL);
1275 bytes -= _DIRENT_DIRSIZ(den);
1276 den = _DIRENT_NEXT(den);
1278 if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
1283 if (rncp->nc_flag & NCF_UNRESOLVED) {
1284 cache_setvp(rncp, dvp);
1285 if (ncvp_debug >= 2) {
1286 printf("cache_inefficient_scan: setvp %s/%s = %p\n",
1287 ncp->nc_name, rncp->nc_name, dvp);
1290 if (ncvp_debug >= 2) {
1291 printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
1292 ncp->nc_name, rncp->nc_name, dvp,
1296 if (rncp->nc_vp == NULL)
1297 error = rncp->nc_error;
1300 printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
1310 * Zap a namecache entry. The ncp is unconditionally set to an unresolved
1311 * state, which disassociates it from its vnode or ncneglist.
1313 * Then, if there are no additional references to the ncp and no children,
1314 * the ncp is removed from the topology and destroyed. This function will
1315 * also run through the nc_parent chain and destroy parent ncps if possible.
1316 * As a side benefit, it turns out the only conditions that allow running
1317 * up the chain are also the conditions to ensure no deadlock will occur.
1319 * References and/or children may exist if the ncp is in the middle of the
1320 * topology, preventing the ncp from being destroyed.
1322 * This function must be called with the ncp held and locked and will unlock
1323 * and drop it during zapping.
1326 cache_zap(struct namecache *ncp)
1328 struct namecache *par;
1331 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
1333 cache_setunresolved(ncp);
1336 * Try to scrap the entry and possibly tail-recurse on its parent.
1337 * We only scrap unref'd (other then our ref) unresolved entries,
1338 * we do not scrap 'live' entries.
1340 while (ncp->nc_flag & NCF_UNRESOLVED) {
1342 * Someone other then us has a ref, stop.
1344 if (ncp->nc_refs > 1)
1348 * We have children, stop.
1350 if (!TAILQ_EMPTY(&ncp->nc_list))
1354 * Remove ncp from the topology: hash table and parent linkage.
1356 if (ncp->nc_flag & NCF_HASHED) {
1357 ncp->nc_flag &= ~NCF_HASHED;
1358 LIST_REMOVE(ncp, nc_hash);
1360 if ((par = ncp->nc_parent) != NULL) {
1361 par = cache_hold(par);
1362 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
1363 ncp->nc_parent = NULL;
1364 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
1369 * ncp should not have picked up any refs. Physically
1372 KKASSERT(ncp->nc_refs == 1);
1374 /* cache_unlock(ncp) not required */
1375 ncp->nc_refs = -1; /* safety */
1377 free(ncp->nc_name, M_VFSCACHE);
1378 free(ncp, M_VFSCACHE);
1381 * Loop on the parent (it may be NULL). Only bother looping
1382 * if the parent has a single ref (ours), which also means
1383 * we can lock it trivially.
1388 if (ncp->nc_refs != 1) {
1392 KKASSERT(par->nc_exlocks == 0);
1397 atomic_subtract_int(&ncp->nc_refs, 1);
1400 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW;
1404 cache_hysteresis(void)
1407 * Don't cache too many negative hits. We use hysteresis to reduce
1408 * the impact on the critical path.
1410 switch(cache_hysteresis_state) {
1412 if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
1414 cache_hysteresis_state = CHI_HIGH;
1418 if (numneg > MINNEG * 9 / 10 &&
1419 numneg * ncnegfactor * 9 / 10 > numcache
1423 cache_hysteresis_state = CHI_LOW;
1430 * NEW NAMECACHE LOOKUP API
1432 * Lookup an entry in the cache. A locked, referenced, non-NULL
1433 * entry is *always* returned, even if the supplied component is illegal.
1434 * The resulting namecache entry should be returned to the system with
1435 * cache_put() or cache_unlock() + cache_drop().
1437 * namecache locks are recursive but care must be taken to avoid lock order
1440 * Nobody else will be able to manipulate the associated namespace (e.g.
1441 * create, delete, rename, rename-target) until the caller unlocks the
1444 * The returned entry will be in one of three states: positive hit (non-null
1445 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
1446 * Unresolved entries must be resolved through the filesystem to associate the
1447 * vnode and/or determine whether a positive or negative hit has occured.
1449 * It is not necessary to lock a directory in order to lock namespace under
1450 * that directory. In fact, it is explicitly not allowed to do that. A
1451 * directory is typically only locked when being created, renamed, or
1454 * The directory (par) may be unresolved, in which case any returned child
1455 * will likely also be marked unresolved. Likely but not guarenteed. Since
1456 * the filesystem lookup requires a resolved directory vnode the caller is
1457 * responsible for resolving the namecache chain top-down. This API
1458 * specifically allows whole chains to be created in an unresolved state.
1461 cache_nlookup(struct namecache *par, struct nlcomponent *nlc)
1463 struct namecache *ncp;
1464 struct namecache *new_ncp;
1465 struct nchashhead *nchpp;
1473 * Try to locate an existing entry
1475 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
1476 hash = fnv_32_buf(&par, sizeof(par), hash);
1479 LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1483 * Zap entries that have timed out.
1485 if (ncp->nc_timeout &&
1486 (int)(ncp->nc_timeout - ticks) < 0 &&
1487 (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1488 ncp->nc_exlocks == 0
1490 cache_zap(cache_get(ncp));
1495 * Break out if we find a matching entry. Note that
1496 * UNRESOLVED entries may match, but DESTROYED entries
1499 if (ncp->nc_parent == par &&
1500 ncp->nc_nlen == nlc->nlc_namelen &&
1501 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
1502 (ncp->nc_flag & NCF_DESTROYED) == 0
1504 if (cache_get_nonblock(ncp) == 0) {
1506 cache_free(new_ncp);
1516 * We failed to locate an entry, create a new entry and add it to
1517 * the cache. We have to relookup after possibly blocking in
1520 if (new_ncp == NULL) {
1521 new_ncp = cache_alloc(nlc->nlc_namelen);
1528 * Initialize as a new UNRESOLVED entry, lock (non-blocking),
1529 * and link to the parent. The mount point is usually inherited
1530 * from the parent unless this is a special case such as a mount
1531 * point where nlc_namelen is 0. The caller is responsible for
1532 * setting nc_mount in that case. If nlc_namelen is 0 nc_name will
1535 if (nlc->nlc_namelen) {
1536 bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen);
1537 ncp->nc_name[nlc->nlc_namelen] = 0;
1538 ncp->nc_mount = par->nc_mount;
1540 nchpp = NCHHASH(hash);
1541 LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1542 ncp->nc_flag |= NCF_HASHED;
1543 cache_link_parent(ncp, par);
1546 * stats and namecache size management
1548 if (ncp->nc_flag & NCF_UNRESOLVED)
1549 ++gd->gd_nchstats->ncs_miss;
1550 else if (ncp->nc_vp)
1551 ++gd->gd_nchstats->ncs_goodhits;
1553 ++gd->gd_nchstats->ncs_neghits;
1559 * Given a locked ncp, validate that the vnode, if present, is actually
1560 * usable. If it is not usable set the ncp to an unresolved state.
1563 cache_validate(struct namecache *ncp)
1565 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1566 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1567 cache_setunresolved(ncp);
1572 * Resolve an unresolved namecache entry, generally by looking it up.
1573 * The passed ncp must be locked and refd.
1575 * Theoretically since a vnode cannot be recycled while held, and since
1576 * the nc_parent chain holds its vnode as long as children exist, the
1577 * direct parent of the cache entry we are trying to resolve should
1578 * have a valid vnode. If not then generate an error that we can
1579 * determine is related to a resolver bug.
1581 * However, if a vnode was in the middle of a recyclement when the NCP
1582 * got locked, ncp->nc_vp might point to a vnode that is about to become
1583 * invalid. cache_resolve() handles this case by unresolving the entry
1584 * and then re-resolving it.
1586 * Note that successful resolution does not necessarily return an error
1587 * code of 0. If the ncp resolves to a negative cache hit then ENOENT
1591 cache_resolve(struct namecache *ncp, struct ucred *cred)
1593 struct namecache *par;
1598 * If the ncp is already resolved we have nothing to do. However,
1599 * we do want to guarentee that a usable vnode is returned when
1600 * a vnode is present, so make sure it hasn't been reclaimed.
1602 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1603 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1604 cache_setunresolved(ncp);
1605 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1606 return (ncp->nc_error);
1610 * Mount points need special handling because the parent does not
1611 * belong to the same filesystem as the ncp.
1613 if (ncp->nc_flag & NCF_MOUNTPT)
1614 return (cache_resolve_mp(ncp));
1617 * We expect an unbroken chain of ncps to at least the mount point,
1618 * and even all the way to root (but this code doesn't have to go
1619 * past the mount point).
1621 if (ncp->nc_parent == NULL) {
1622 printf("EXDEV case 1 %p %*.*s\n", ncp,
1623 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1624 ncp->nc_error = EXDEV;
1625 return(ncp->nc_error);
1629 * The vp's of the parent directories in the chain are held via vhold()
1630 * due to the existance of the child, and should not disappear.
1631 * However, there are cases where they can disappear:
1633 * - due to filesystem I/O errors.
1634 * - due to NFS being stupid about tracking the namespace and
1635 * destroys the namespace for entire directories quite often.
1636 * - due to forced unmounts.
1637 * - due to an rmdir (parent will be marked DESTROYED)
1639 * When this occurs we have to track the chain backwards and resolve
1640 * it, looping until the resolver catches up to the current node. We
1641 * could recurse here but we might run ourselves out of kernel stack
1642 * so we do it in a more painful manner. This situation really should
1643 * not occur all that often, or if it does not have to go back too
1644 * many nodes to resolve the ncp.
1646 while (ncp->nc_parent->nc_vp == NULL) {
1648 * This case can occur if a process is CD'd into a
1649 * directory which is then rmdir'd. If the parent is marked
1650 * destroyed there is no point trying to resolve it.
1652 if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
1655 par = ncp->nc_parent;
1656 while (par->nc_parent && par->nc_parent->nc_vp == NULL)
1657 par = par->nc_parent;
1658 if (par->nc_parent == NULL) {
1659 printf("EXDEV case 2 %*.*s\n",
1660 par->nc_nlen, par->nc_nlen, par->nc_name);
1663 printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
1664 par->nc_nlen, par->nc_nlen, par->nc_name);
1666 * The parent is not set in stone, ref and lock it to prevent
1667 * it from disappearing. Also note that due to renames it
1668 * is possible for our ncp to move and for par to no longer
1669 * be one of its parents. We resolve it anyway, the loop
1670 * will handle any moves.
1673 if (par->nc_flag & NCF_MOUNTPT) {
1674 cache_resolve_mp(par);
1675 } else if (par->nc_parent->nc_vp == NULL) {
1676 printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
1679 } else if (par->nc_flag & NCF_UNRESOLVED) {
1680 par->nc_error = VOP_NRESOLVE(par, cred);
1682 if ((error = par->nc_error) != 0) {
1683 if (par->nc_error != EAGAIN) {
1684 printf("EXDEV case 3 %*.*s error %d\n",
1685 par->nc_nlen, par->nc_nlen, par->nc_name,
1690 printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
1691 par, par->nc_nlen, par->nc_nlen, par->nc_name);
1698 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
1699 * ncp's and reattach them. If this occurs the original ncp is marked
1700 * EAGAIN to force a relookup.
1702 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
1703 * ncp must already be resolved.
1705 KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0);
1706 ncp->nc_error = VOP_NRESOLVE(ncp, cred);
1707 /*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/
1708 if (ncp->nc_error == EAGAIN) {
1709 printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
1710 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1713 return(ncp->nc_error);
1717 * Resolve the ncp associated with a mount point. Such ncp's almost always
1718 * remain resolved and this routine is rarely called. NFS MPs tends to force
1719 * re-resolution more often due to its mac-truck-smash-the-namecache
1720 * method of tracking namespace changes.
1722 * The semantics for this call is that the passed ncp must be locked on
1723 * entry and will be locked on return. However, if we actually have to
1724 * resolve the mount point we temporarily unlock the entry in order to
1725 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of
1726 * the unlock we have to recheck the flags after we relock.
1729 cache_resolve_mp(struct namecache *ncp)
1732 struct mount *mp = ncp->nc_mount;
1735 KKASSERT(mp != NULL);
1738 * If the ncp is already resolved we have nothing to do. However,
1739 * we do want to guarentee that a usable vnode is returned when
1740 * a vnode is present, so make sure it hasn't been reclaimed.
1742 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1743 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1744 cache_setunresolved(ncp);
1747 if (ncp->nc_flag & NCF_UNRESOLVED) {
1749 while (vfs_busy(mp, 0))
1751 error = VFS_ROOT(mp, &vp);
1755 * recheck the ncp state after relocking.
1757 if (ncp->nc_flag & NCF_UNRESOLVED) {
1758 ncp->nc_error = error;
1760 cache_setvp(ncp, vp);
1763 printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp);
1764 cache_setvp(ncp, NULL);
1766 } else if (error == 0) {
1771 return(ncp->nc_error);
1775 cache_cleanneg(int count)
1777 struct namecache *ncp;
1780 * Automode from the vnlru proc - clean out 10% of the negative cache
1784 count = numneg / 10 + 1;
1787 * Attempt to clean out the specified number of negative cache
1791 ncp = TAILQ_FIRST(&ncneglist);
1793 KKASSERT(numneg == 0);
1796 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1797 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
1798 if (cache_get_nonblock(ncp) == 0)
1805 * Rehash a ncp. Rehashing is typically required if the name changes (should
1806 * not generally occur) or the parent link changes. This function will
1807 * unhash the ncp if the ncp is no longer hashable.
1810 cache_rehash(struct namecache *ncp)
1812 struct nchashhead *nchpp;
1815 if (ncp->nc_flag & NCF_HASHED) {
1816 ncp->nc_flag &= ~NCF_HASHED;
1817 LIST_REMOVE(ncp, nc_hash);
1819 if (ncp->nc_nlen && ncp->nc_parent) {
1820 hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT);
1821 hash = fnv_32_buf(&ncp->nc_parent,
1822 sizeof(ncp->nc_parent), hash);
1823 nchpp = NCHHASH(hash);
1824 LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1825 ncp->nc_flag |= NCF_HASHED;
1830 * Name cache initialization, from vfsinit() when we are booting
1838 /* initialise per-cpu namecache effectiveness statistics. */
1839 for (i = 0; i < ncpus; ++i) {
1840 gd = globaldata_find(i);
1841 gd->gd_nchstats = &nchstats[i];
1843 TAILQ_INIT(&ncneglist);
1844 nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash);
1845 nclockwarn = 1 * hz;
1849 * Called from start_init() to bootstrap the root filesystem. Returns
1850 * a referenced, unlocked namecache record.
1853 cache_allocroot(struct mount *mp, struct vnode *vp)
1855 struct namecache *ncp = cache_alloc(0);
1857 ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT;
1859 cache_setvp(ncp, vp);
1864 * vfs_cache_setroot()
1866 * Create an association between the root of our namecache and
1867 * the root vnode. This routine may be called several times during
1870 * If the caller intends to save the returned namecache pointer somewhere
1871 * it must cache_hold() it.
1874 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp)
1877 struct namecache *oncp;
1891 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache
1892 * topology and is being removed as quickly as possible. The new VOP_N*()
1893 * API calls are required to make specific adjustments using the supplied
1894 * ncp pointers rather then just bogusly purging random vnodes.
1896 * Invalidate all namecache entries to a particular vnode as well as
1897 * any direct children of that vnode in the namecache. This is a
1898 * 'catch all' purge used by filesystems that do not know any better.
1900 * Note that the linkage between the vnode and its namecache entries will
1901 * be removed, but the namecache entries themselves might stay put due to
1902 * active references from elsewhere in the system or due to the existance of
1903 * the children. The namecache topology is left intact even if we do not
1904 * know what the vnode association is. Such entries will be marked
1908 cache_purge(struct vnode *vp)
1910 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
1914 * Flush all entries referencing a particular filesystem.
1916 * Since we need to check it anyway, we will flush all the invalid
1917 * entries at the same time.
1920 cache_purgevfs(struct mount *mp)
1922 struct nchashhead *nchpp;
1923 struct namecache *ncp, *nnp;
1926 * Scan hash tables for applicable entries.
1928 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
1929 ncp = LIST_FIRST(nchpp);
1933 nnp = LIST_NEXT(ncp, nc_hash);
1936 if (ncp->nc_mount == mp) {
1948 * Create a new (theoretically) unique fsmid
1951 cache_getnewfsmid(void)
1953 static int fsmid_roller;
1957 fsmid = ((int64_t)time_second << 32) |
1958 (fsmid_roller & 0x7FFFFFFF);
1963 static int disablecwd;
1964 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, "");
1966 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls);
1967 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1);
1968 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2);
1969 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3);
1970 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4);
1971 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound);
1974 __getcwd(struct __getcwd_args *uap)
1984 buflen = uap->buflen;
1987 if (buflen > MAXPATHLEN)
1988 buflen = MAXPATHLEN;
1990 buf = malloc(buflen, M_TEMP, M_WAITOK);
1991 bp = kern_getcwd(buf, buflen, &error);
1993 error = copyout(bp, uap->buf, strlen(bp) + 1);
1999 kern_getcwd(char *buf, size_t buflen, int *error)
2001 struct proc *p = curproc;
2003 int i, slash_prefixed;
2004 struct filedesc *fdp;
2005 struct namecache *ncp;
2014 ncp = fdp->fd_ncdir;
2015 while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2016 if (ncp->nc_flag & NCF_MOUNTPT) {
2017 if (ncp->nc_mount == NULL) {
2018 *error = EBADF; /* forced unmount? */
2021 ncp = ncp->nc_parent;
2024 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2030 *--bp = ncp->nc_name[i];
2039 ncp = ncp->nc_parent;
2046 if (!slash_prefixed) {
2060 * Thus begins the fullpath magic.
2064 #define STATNODE(name) \
2065 static u_int name; \
2066 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "")
2068 static int disablefullpath;
2069 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
2070 &disablefullpath, 0, "");
2072 STATNODE(numfullpathcalls);
2073 STATNODE(numfullpathfail1);
2074 STATNODE(numfullpathfail2);
2075 STATNODE(numfullpathfail3);
2076 STATNODE(numfullpathfail4);
2077 STATNODE(numfullpathfound);
2080 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf)
2083 int i, slash_prefixed;
2084 struct namecache *fd_nrdir;
2088 buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2089 bp = buf + MAXPATHLEN - 1;
2092 fd_nrdir = p->p_fd->fd_nrdir;
2096 while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2097 if (ncp->nc_flag & NCF_MOUNTPT) {
2098 if (ncp->nc_mount == NULL) {
2102 ncp = ncp->nc_parent;
2105 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2111 *--bp = ncp->nc_name[i];
2120 ncp = ncp->nc_parent;
2127 if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) {
2128 bp = buf + MAXPATHLEN - 1;
2132 if (!slash_prefixed) {
2148 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf)
2150 struct namecache *ncp;
2153 if (disablefullpath)
2159 /* vn is NULL, client wants us to use p->p_textvp */
2161 if ((vn = p->p_textvp) == NULL)
2164 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
2172 return(cache_fullpath(p, ncp, retbuf, freebuf));