2 * Copyright (c) 2003,2004,2009 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
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
72 #include <sys/sysctl.h>
73 #include <sys/mount.h>
74 #include <sys/vnode.h>
75 #include <sys/malloc.h>
76 #include <sys/sysproto.h>
77 #include <sys/spinlock.h>
79 #include <sys/namei.h>
80 #include <sys/nlookup.h>
81 #include <sys/filedesc.h>
82 #include <sys/fnv_hash.h>
83 #include <sys/globaldata.h>
84 #include <sys/kern_syscall.h>
85 #include <sys/dirent.h>
88 #include <sys/sysref2.h>
89 #include <sys/spinlock2.h>
90 #include <sys/mplock2.h>
92 #define MAX_RECURSION_DEPTH 64
95 * Random lookups in the cache are accomplished with a hash table using
96 * a hash key of (nc_src_vp, name). Each hash chain has its own spin lock.
98 * Negative entries may exist and correspond to resolved namecache
99 * structures where nc_vp is NULL. In a negative entry, NCF_WHITEOUT
100 * will be set if the entry corresponds to a whited-out directory entry
101 * (verses simply not finding the entry at all). ncneglist is locked
102 * with a global spinlock (ncspin).
106 * (1) A ncp must be referenced before it can be locked.
108 * (2) A ncp must be locked in order to modify it.
110 * (3) ncp locks are always ordered child -> parent. That may seem
111 * backwards but forward scans use the hash table and thus can hold
112 * the parent unlocked when traversing downward.
114 * This allows insert/rename/delete/dot-dot and other operations
115 * to use ncp->nc_parent links.
117 * This also prevents a locked up e.g. NFS node from creating a
118 * chain reaction all the way back to the root vnode / namecache.
120 * (4) parent linkages require both the parent and child to be locked.
124 * Structures associated with name cacheing.
126 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash])
129 #define NCMOUNT_NUMCACHE 1009 /* prime number */
131 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
133 LIST_HEAD(nchash_list, namecache);
136 struct nchash_list list;
137 struct spinlock spin;
140 struct ncmount_cache {
141 struct spinlock spin;
142 struct namecache *ncp;
146 static struct nchash_head *nchashtbl;
147 static struct namecache_list ncneglist;
148 static struct spinlock ncspin;
149 static struct ncmount_cache ncmount_cache[NCMOUNT_NUMCACHE];
152 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server
153 * to create the namecache infrastructure leading to a dangling vnode.
155 * 0 Only errors are reported
156 * 1 Successes are reported
157 * 2 Successes + the whole directory scan is reported
158 * 3 Force the directory scan code run as if the parent vnode did not
159 * have a namecache record, even if it does have one.
161 static int ncvp_debug;
162 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0,
163 "Namecache debug level (0-3)");
165 static u_long nchash; /* size of hash table */
166 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
167 "Size of namecache hash table");
169 static int ncnegfactor = 16; /* ratio of negative entries */
170 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
171 "Ratio of namecache negative entries");
173 static int nclockwarn; /* warn on locked entries in ticks */
174 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0,
175 "Warn on locked namecache entries in ticks");
177 static int numdefered; /* number of cache entries allocated */
178 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0,
179 "Number of cache entries allocated");
181 static int ncposlimit; /* number of cache entries allocated */
182 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0,
183 "Number of cache entries allocated");
185 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode),
186 "sizeof(struct vnode)");
187 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache),
188 "sizeof(struct namecache)");
190 static int ncmount_cache_enable = 1;
191 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW,
192 &ncmount_cache_enable, 0, "mount point cache");
193 static long ncmount_cache_hit;
194 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_hit, CTLFLAG_RW,
195 &ncmount_cache_hit, 0, "mpcache hits");
196 static long ncmount_cache_miss;
197 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_miss, CTLFLAG_RW,
198 &ncmount_cache_miss, 0, "mpcache misses");
199 static long ncmount_cache_overwrite;
200 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_overwrite, CTLFLAG_RW,
201 &ncmount_cache_overwrite, 0, "mpcache entry overwrites");
203 static int cache_resolve_mp(struct mount *mp);
204 static struct vnode *cache_dvpref(struct namecache *ncp);
205 static void _cache_lock(struct namecache *ncp);
206 static void _cache_setunresolved(struct namecache *ncp);
207 static void _cache_cleanneg(int count);
208 static void _cache_cleanpos(int count);
209 static void _cache_cleandefered(void);
210 static void _cache_unlink(struct namecache *ncp);
213 * The new name cache statistics
215 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
217 SYSCTL_INT(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
218 "Number of negative namecache entries");
220 SYSCTL_INT(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
221 "Number of namecaches entries");
222 static u_long numcalls;
223 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcalls, CTLFLAG_RD, &numcalls, 0,
224 "Number of namecache lookups");
225 static u_long numchecks;
226 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numchecks, CTLFLAG_RD, &numchecks, 0,
227 "Number of checked entries in namecache lookups");
229 struct nchstats nchstats[SMP_MAXCPU];
231 * Export VFS cache effectiveness statistics to user-land.
233 * The statistics are left for aggregation to user-land so
234 * neat things can be achieved, like observing per-CPU cache
238 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
240 struct globaldata *gd;
244 for (i = 0; i < ncpus; ++i) {
245 gd = globaldata_find(i);
246 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
247 sizeof(struct nchstats))))
253 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
254 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
256 static struct namecache *cache_zap(struct namecache *ncp, int nonblock);
259 * Namespace locking. The caller must already hold a reference to the
260 * namecache structure in order to lock/unlock it. This function prevents
261 * the namespace from being created or destroyed by accessors other then
264 * Note that holding a locked namecache structure prevents other threads
265 * from making namespace changes (e.g. deleting or creating), prevents
266 * vnode association state changes by other threads, and prevents the
267 * namecache entry from being resolved or unresolved by other threads.
269 * The lock owner has full authority to associate/disassociate vnodes
270 * and resolve/unresolve the locked ncp.
272 * The primary lock field is nc_exlocks. nc_locktd is set after the
273 * fact (when locking) or cleared prior to unlocking.
275 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed
276 * or recycled, but it does NOT help you if the vnode had already
277 * initiated a recyclement. If this is important, use cache_get()
278 * rather then cache_lock() (and deal with the differences in the
279 * way the refs counter is handled). Or, alternatively, make an
280 * unconditional call to cache_validate() or cache_resolve()
281 * after cache_lock() returns.
287 _cache_lock(struct namecache *ncp)
294 KKASSERT(ncp->nc_refs != 0);
299 count = ncp->nc_exlocks;
302 if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) {
304 * The vp associated with a locked ncp must
305 * be held to prevent it from being recycled.
307 * WARNING! If VRECLAIMED is set the vnode
308 * could already be in the middle of a recycle.
309 * Callers must use cache_vref() or
310 * cache_vget() on the locked ncp to
311 * validate the vp or set the cache entry
314 * NOTE! vhold() is allowed if we hold a
315 * lock on the ncp (which we do).
319 vhold(ncp->nc_vp); /* MPSAFE */
325 if (ncp->nc_locktd == td) {
326 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
333 tsleep_interlock(ncp, 0);
334 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
335 count | NC_EXLOCK_REQ) == 0) {
339 error = tsleep(ncp, PINTERLOCKED, "clock", nclockwarn);
340 if (error == EWOULDBLOCK) {
343 kprintf("[diagnostic] cache_lock: blocked "
346 kprintf(" \"%*.*s\"\n",
347 ncp->nc_nlen, ncp->nc_nlen,
353 kprintf("[diagnostic] cache_lock: unblocked %*.*s after "
355 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
356 (int)(ticks - didwarn) / hz);
361 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance,
362 * such as the case where one of its children is locked.
368 _cache_lock_nonblock(struct namecache *ncp)
376 count = ncp->nc_exlocks;
379 if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) {
381 * The vp associated with a locked ncp must
382 * be held to prevent it from being recycled.
384 * WARNING! If VRECLAIMED is set the vnode
385 * could already be in the middle of a recycle.
386 * Callers must use cache_vref() or
387 * cache_vget() on the locked ncp to
388 * validate the vp or set the cache entry
391 * NOTE! vhold() is allowed if we hold a
392 * lock on the ncp (which we do).
396 vhold(ncp->nc_vp); /* MPSAFE */
402 if (ncp->nc_locktd == td) {
403 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
418 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop).
420 * nc_locktd must be NULLed out prior to nc_exlocks getting cleared.
426 _cache_unlock(struct namecache *ncp)
428 thread_t td __debugvar = curthread;
431 KKASSERT(ncp->nc_refs >= 0);
432 KKASSERT(ncp->nc_exlocks > 0);
433 KKASSERT(ncp->nc_locktd == td);
435 count = ncp->nc_exlocks;
436 if ((count & ~NC_EXLOCK_REQ) == 1) {
437 ncp->nc_locktd = NULL;
442 if ((count & ~NC_EXLOCK_REQ) == 1) {
443 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 0)) {
444 if (count & NC_EXLOCK_REQ)
449 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
454 count = ncp->nc_exlocks;
460 * cache_hold() and cache_drop() prevent the premature deletion of a
461 * namecache entry but do not prevent operations (such as zapping) on
462 * that namecache entry.
464 * This routine may only be called from outside this source module if
465 * nc_refs is already at least 1.
467 * This is a rare case where callers are allowed to hold a spinlock,
468 * so we can't ourselves.
474 _cache_hold(struct namecache *ncp)
476 atomic_add_int(&ncp->nc_refs, 1);
481 * Drop a cache entry, taking care to deal with races.
483 * For potential 1->0 transitions we must hold the ncp lock to safely
484 * test its flags. An unresolved entry with no children must be zapped
487 * The call to cache_zap() itself will handle all remaining races and
488 * will decrement the ncp's refs regardless. If we are resolved or
489 * have children nc_refs can safely be dropped to 0 without having to
492 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion.
494 * NOTE: cache_zap() may return a non-NULL referenced parent which must
495 * be dropped in a loop.
501 _cache_drop(struct namecache *ncp)
506 KKASSERT(ncp->nc_refs > 0);
510 if (_cache_lock_nonblock(ncp) == 0) {
511 ncp->nc_flag &= ~NCF_DEFEREDZAP;
512 if ((ncp->nc_flag & NCF_UNRESOLVED) &&
513 TAILQ_EMPTY(&ncp->nc_list)) {
514 ncp = cache_zap(ncp, 1);
517 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) {
524 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1))
532 * Link a new namecache entry to its parent and to the hash table. Be
533 * careful to avoid races if vhold() blocks in the future.
535 * Both ncp and par must be referenced and locked.
537 * NOTE: The hash table spinlock is likely held during this call, we
538 * can't do anything fancy.
543 _cache_link_parent(struct namecache *ncp, struct namecache *par,
544 struct nchash_head *nchpp)
546 KKASSERT(ncp->nc_parent == NULL);
547 ncp->nc_parent = par;
548 ncp->nc_head = nchpp;
551 * Set inheritance flags. Note that the parent flags may be
552 * stale due to getattr potentially not having been run yet
553 * (it gets run during nlookup()'s).
555 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE);
556 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE))
557 ncp->nc_flag |= NCF_SF_PNOCACHE;
558 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE))
559 ncp->nc_flag |= NCF_UF_PCACHE;
561 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash);
563 if (TAILQ_EMPTY(&par->nc_list)) {
564 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
566 * Any vp associated with an ncp which has children must
567 * be held to prevent it from being recycled.
572 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
577 * Remove the parent and hash associations from a namecache structure.
578 * If this is the last child of the parent the cache_drop(par) will
579 * attempt to recursively zap the parent.
581 * ncp must be locked. This routine will acquire a temporary lock on
582 * the parent as wlel as the appropriate hash chain.
587 _cache_unlink_parent(struct namecache *ncp)
589 struct namecache *par;
590 struct vnode *dropvp;
592 if ((par = ncp->nc_parent) != NULL) {
593 KKASSERT(ncp->nc_parent == par);
596 spin_lock(&ncp->nc_head->spin);
597 LIST_REMOVE(ncp, nc_hash);
598 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
600 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
602 spin_unlock(&ncp->nc_head->spin);
603 ncp->nc_parent = NULL;
609 * We can only safely vdrop with no spinlocks held.
617 * Allocate a new namecache structure. Most of the code does not require
618 * zero-termination of the string but it makes vop_compat_ncreate() easier.
622 static struct namecache *
623 cache_alloc(int nlen)
625 struct namecache *ncp;
627 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
629 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
631 ncp->nc_flag = NCF_UNRESOLVED;
632 ncp->nc_error = ENOTCONN; /* needs to be resolved */
635 TAILQ_INIT(&ncp->nc_list);
641 * Can only be called for the case where the ncp has never been
642 * associated with anything (so no spinlocks are needed).
647 _cache_free(struct namecache *ncp)
649 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
651 kfree(ncp->nc_name, M_VFSCACHE);
652 kfree(ncp, M_VFSCACHE);
659 cache_zero(struct nchandle *nch)
666 * Ref and deref a namecache structure.
668 * The caller must specify a stable ncp pointer, typically meaning the
669 * ncp is already referenced but this can also occur indirectly through
670 * e.g. holding a lock on a direct child.
672 * WARNING: Caller may hold an unrelated read spinlock, which means we can't
673 * use read spinlocks here.
678 cache_hold(struct nchandle *nch)
680 _cache_hold(nch->ncp);
681 atomic_add_int(&nch->mount->mnt_refs, 1);
686 * Create a copy of a namecache handle for an already-referenced
692 cache_copy(struct nchandle *nch, struct nchandle *target)
696 _cache_hold(target->ncp);
697 atomic_add_int(&nch->mount->mnt_refs, 1);
704 cache_changemount(struct nchandle *nch, struct mount *mp)
706 atomic_add_int(&nch->mount->mnt_refs, -1);
708 atomic_add_int(&nch->mount->mnt_refs, 1);
715 cache_drop(struct nchandle *nch)
717 atomic_add_int(&nch->mount->mnt_refs, -1);
718 _cache_drop(nch->ncp);
727 cache_lock(struct nchandle *nch)
729 _cache_lock(nch->ncp);
733 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller
734 * is responsible for checking both for validity on return as they
735 * may have become invalid.
737 * We have to deal with potential deadlocks here, just ping pong
738 * the lock until we get it (we will always block somewhere when
739 * looping so this is not cpu-intensive).
741 * which = 0 nch1 not locked, nch2 is locked
742 * which = 1 nch1 is locked, nch2 is not locked
745 cache_relock(struct nchandle *nch1, struct ucred *cred1,
746 struct nchandle *nch2, struct ucred *cred2)
754 if (cache_lock_nonblock(nch1) == 0) {
755 cache_resolve(nch1, cred1);
760 cache_resolve(nch1, cred1);
763 if (cache_lock_nonblock(nch2) == 0) {
764 cache_resolve(nch2, cred2);
769 cache_resolve(nch2, cred2);
779 cache_lock_nonblock(struct nchandle *nch)
781 return(_cache_lock_nonblock(nch->ncp));
789 cache_unlock(struct nchandle *nch)
791 _cache_unlock(nch->ncp);
795 * ref-and-lock, unlock-and-deref functions.
797 * This function is primarily used by nlookup. Even though cache_lock
798 * holds the vnode, it is possible that the vnode may have already
799 * initiated a recyclement.
801 * We want cache_get() to return a definitively usable vnode or a
802 * definitively unresolved ncp.
808 _cache_get(struct namecache *ncp)
812 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
813 _cache_setunresolved(ncp);
818 * This is a special form of _cache_lock() which only succeeds if
819 * it can get a pristine, non-recursive lock. The caller must have
820 * already ref'd the ncp.
822 * On success the ncp will be locked, on failure it will not. The
823 * ref count does not change either way.
825 * We want _cache_lock_special() (on success) to return a definitively
826 * usable vnode or a definitively unresolved ncp.
831 _cache_lock_special(struct namecache *ncp)
833 if (_cache_lock_nonblock(ncp) == 0) {
834 if ((ncp->nc_exlocks & ~NC_EXLOCK_REQ) == 1) {
835 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
836 _cache_setunresolved(ncp);
846 * NOTE: The same nchandle can be passed for both arguments.
851 cache_get(struct nchandle *nch, struct nchandle *target)
853 KKASSERT(nch->ncp->nc_refs > 0);
854 target->mount = nch->mount;
855 target->ncp = _cache_get(nch->ncp);
856 atomic_add_int(&target->mount->mnt_refs, 1);
864 _cache_put(struct namecache *ncp)
874 cache_put(struct nchandle *nch)
876 atomic_add_int(&nch->mount->mnt_refs, -1);
877 _cache_put(nch->ncp);
883 * Resolve an unresolved ncp by associating a vnode with it. If the
884 * vnode is NULL, a negative cache entry is created.
886 * The ncp should be locked on entry and will remain locked on return.
892 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp)
894 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
898 * Any vp associated with an ncp which has children must
899 * be held. Any vp associated with a locked ncp must be held.
901 if (!TAILQ_EMPTY(&ncp->nc_list))
903 spin_lock(&vp->v_spin);
905 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
906 spin_unlock(&vp->v_spin);
911 * Set auxiliary flags
915 ncp->nc_flag |= NCF_ISDIR;
918 ncp->nc_flag |= NCF_ISSYMLINK;
919 /* XXX cache the contents of the symlink */
924 atomic_add_int(&numcache, 1);
926 /* XXX: this is a hack to work-around the lack of a real pfs vfs
929 if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0)
933 * When creating a negative cache hit we set the
934 * namecache_gen. A later resolve will clean out the
935 * negative cache hit if the mount point's namecache_gen
936 * has changed. Used by devfs, could also be used by
941 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
943 spin_unlock(&ncspin);
944 ncp->nc_error = ENOENT;
946 VFS_NCPGEN_SET(mp, ncp);
948 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP);
955 cache_setvp(struct nchandle *nch, struct vnode *vp)
957 _cache_setvp(nch->mount, nch->ncp, vp);
964 cache_settimeout(struct nchandle *nch, int nticks)
966 struct namecache *ncp = nch->ncp;
968 if ((ncp->nc_timeout = ticks + nticks) == 0)
973 * Disassociate the vnode or negative-cache association and mark a
974 * namecache entry as unresolved again. Note that the ncp is still
975 * left in the hash table and still linked to its parent.
977 * The ncp should be locked and refd on entry and will remain locked and refd
980 * This routine is normally never called on a directory containing children.
981 * However, NFS often does just that in its rename() code as a cop-out to
982 * avoid complex namespace operations. This disconnects a directory vnode
983 * from its namecache and can cause the OLDAPI and NEWAPI to get out of
990 _cache_setunresolved(struct namecache *ncp)
994 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
995 ncp->nc_flag |= NCF_UNRESOLVED;
997 ncp->nc_error = ENOTCONN;
998 if ((vp = ncp->nc_vp) != NULL) {
999 atomic_add_int(&numcache, -1);
1000 spin_lock(&vp->v_spin);
1002 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
1003 spin_unlock(&vp->v_spin);
1006 * Any vp associated with an ncp with children is
1007 * held by that ncp. Any vp associated with a locked
1008 * ncp is held by that ncp. These conditions must be
1009 * undone when the vp is cleared out from the ncp.
1011 if (!TAILQ_EMPTY(&ncp->nc_list))
1013 if (ncp->nc_exlocks)
1017 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1019 spin_unlock(&ncspin);
1021 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK);
1026 * The cache_nresolve() code calls this function to automatically
1027 * set a resolved cache element to unresolved if it has timed out
1028 * or if it is a negative cache hit and the mount point namecache_gen
1033 static __inline void
1034 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp)
1037 * Already in an unresolved state, nothing to do.
1039 if (ncp->nc_flag & NCF_UNRESOLVED)
1043 * Try to zap entries that have timed out. We have
1044 * to be careful here because locked leafs may depend
1045 * on the vnode remaining intact in a parent, so only
1046 * do this under very specific conditions.
1048 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 &&
1049 TAILQ_EMPTY(&ncp->nc_list)) {
1050 _cache_setunresolved(ncp);
1055 * If a resolved negative cache hit is invalid due to
1056 * the mount's namecache generation being bumped, zap it.
1058 if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) {
1059 _cache_setunresolved(ncp);
1068 cache_setunresolved(struct nchandle *nch)
1070 _cache_setunresolved(nch->ncp);
1074 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
1075 * looking for matches. This flag tells the lookup code when it must
1076 * check for a mount linkage and also prevents the directories in question
1077 * from being deleted or renamed.
1083 cache_clrmountpt_callback(struct mount *mp, void *data)
1085 struct nchandle *nch = data;
1087 if (mp->mnt_ncmounton.ncp == nch->ncp)
1089 if (mp->mnt_ncmountpt.ncp == nch->ncp)
1098 cache_clrmountpt(struct nchandle *nch)
1102 count = mountlist_scan(cache_clrmountpt_callback, nch,
1103 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1105 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
1109 * Invalidate portions of the namecache topology given a starting entry.
1110 * The passed ncp is set to an unresolved state and:
1112 * The passed ncp must be referencxed and locked. The routine may unlock
1113 * and relock ncp several times, and will recheck the children and loop
1114 * to catch races. When done the passed ncp will be returned with the
1115 * reference and lock intact.
1117 * CINV_DESTROY - Set a flag in the passed ncp entry indicating
1118 * that the physical underlying nodes have been
1119 * destroyed... as in deleted. For example, when
1120 * a directory is removed. This will cause record
1121 * lookups on the name to no longer be able to find
1122 * the record and tells the resolver to return failure
1123 * rather then trying to resolve through the parent.
1125 * The topology itself, including ncp->nc_name,
1128 * This only applies to the passed ncp, if CINV_CHILDREN
1129 * is specified the children are not flagged.
1131 * CINV_CHILDREN - Set all children (recursively) to an unresolved
1134 * Note that this will also have the side effect of
1135 * cleaning out any unreferenced nodes in the topology
1136 * from the leaves up as the recursion backs out.
1138 * Note that the topology for any referenced nodes remains intact, but
1139 * the nodes will be marked as having been destroyed and will be set
1140 * to an unresolved state.
1142 * It is possible for cache_inval() to race a cache_resolve(), meaning that
1143 * the namecache entry may not actually be invalidated on return if it was
1144 * revalidated while recursing down into its children. This code guarentees
1145 * that the node(s) will go through an invalidation cycle, but does not
1146 * guarentee that they will remain in an invalidated state.
1148 * Returns non-zero if a revalidation was detected during the invalidation
1149 * recursion, zero otherwise. Note that since only the original ncp is
1150 * locked the revalidation ultimately can only indicate that the original ncp
1151 * *MIGHT* no have been reresolved.
1153 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
1154 * have to avoid blowing out the kernel stack. We do this by saving the
1155 * deep namecache node and aborting the recursion, then re-recursing at that
1156 * node using a depth-first algorithm in order to allow multiple deep
1157 * recursions to chain through each other, then we restart the invalidation
1164 struct namecache *resume_ncp;
1168 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
1172 _cache_inval(struct namecache *ncp, int flags)
1174 struct cinvtrack track;
1175 struct namecache *ncp2;
1179 track.resume_ncp = NULL;
1182 r = _cache_inval_internal(ncp, flags, &track);
1183 if (track.resume_ncp == NULL)
1185 kprintf("Warning: deep namecache recursion at %s\n",
1188 while ((ncp2 = track.resume_ncp) != NULL) {
1189 track.resume_ncp = NULL;
1191 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
1201 cache_inval(struct nchandle *nch, int flags)
1203 return(_cache_inval(nch->ncp, flags));
1207 * Helper for _cache_inval(). The passed ncp is refd and locked and
1208 * remains that way on return, but may be unlocked/relocked multiple
1209 * times by the routine.
1212 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
1214 struct namecache *kid;
1215 struct namecache *nextkid;
1218 KKASSERT(ncp->nc_exlocks);
1220 _cache_setunresolved(ncp);
1221 if (flags & CINV_DESTROY)
1222 ncp->nc_flag |= NCF_DESTROYED;
1223 if ((flags & CINV_CHILDREN) &&
1224 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
1227 if (++track->depth > MAX_RECURSION_DEPTH) {
1228 track->resume_ncp = ncp;
1234 if (track->resume_ncp) {
1238 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
1239 _cache_hold(nextkid);
1240 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
1241 TAILQ_FIRST(&kid->nc_list)
1244 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
1255 * Someone could have gotten in there while ncp was unlocked,
1258 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1264 * Invalidate a vnode's namecache associations. To avoid races against
1265 * the resolver we do not invalidate a node which we previously invalidated
1266 * but which was then re-resolved while we were in the invalidation loop.
1268 * Returns non-zero if any namecache entries remain after the invalidation
1271 * NOTE: Unlike the namecache topology which guarentees that ncp's will not
1272 * be ripped out of the topology while held, the vnode's v_namecache
1273 * list has no such restriction. NCP's can be ripped out of the list
1274 * at virtually any time if not locked, even if held.
1276 * In addition, the v_namecache list itself must be locked via
1277 * the vnode's spinlock.
1282 cache_inval_vp(struct vnode *vp, int flags)
1284 struct namecache *ncp;
1285 struct namecache *next;
1288 spin_lock(&vp->v_spin);
1289 ncp = TAILQ_FIRST(&vp->v_namecache);
1293 /* loop entered with ncp held and vp spin-locked */
1294 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1296 spin_unlock(&vp->v_spin);
1298 if (ncp->nc_vp != vp) {
1299 kprintf("Warning: cache_inval_vp: race-A detected on "
1300 "%s\n", ncp->nc_name);
1306 _cache_inval(ncp, flags);
1307 _cache_put(ncp); /* also releases reference */
1309 spin_lock(&vp->v_spin);
1310 if (ncp && ncp->nc_vp != vp) {
1311 spin_unlock(&vp->v_spin);
1312 kprintf("Warning: cache_inval_vp: race-B detected on "
1313 "%s\n", ncp->nc_name);
1318 spin_unlock(&vp->v_spin);
1319 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1323 * This routine is used instead of the normal cache_inval_vp() when we
1324 * are trying to recycle otherwise good vnodes.
1326 * Return 0 on success, non-zero if not all namecache records could be
1327 * disassociated from the vnode (for various reasons).
1332 cache_inval_vp_nonblock(struct vnode *vp)
1334 struct namecache *ncp;
1335 struct namecache *next;
1337 spin_lock(&vp->v_spin);
1338 ncp = TAILQ_FIRST(&vp->v_namecache);
1342 /* loop entered with ncp held */
1343 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1345 spin_unlock(&vp->v_spin);
1346 if (_cache_lock_nonblock(ncp)) {
1352 if (ncp->nc_vp != vp) {
1353 kprintf("Warning: cache_inval_vp: race-A detected on "
1354 "%s\n", ncp->nc_name);
1360 _cache_inval(ncp, 0);
1361 _cache_put(ncp); /* also releases reference */
1363 spin_lock(&vp->v_spin);
1364 if (ncp && ncp->nc_vp != vp) {
1365 spin_unlock(&vp->v_spin);
1366 kprintf("Warning: cache_inval_vp: race-B detected on "
1367 "%s\n", ncp->nc_name);
1372 spin_unlock(&vp->v_spin);
1374 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1378 * The source ncp has been renamed to the target ncp. Both fncp and tncp
1379 * must be locked. The target ncp is destroyed (as a normal rename-over
1380 * would destroy the target file or directory).
1382 * Because there may be references to the source ncp we cannot copy its
1383 * contents to the target. Instead the source ncp is relinked as the target
1384 * and the target ncp is removed from the namecache topology.
1389 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
1391 struct namecache *fncp = fnch->ncp;
1392 struct namecache *tncp = tnch->ncp;
1393 struct namecache *tncp_par;
1394 struct nchash_head *nchpp;
1399 if (tncp->nc_nlen) {
1400 nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK);
1401 bcopy(tncp->nc_name, nname, tncp->nc_nlen);
1402 nname[tncp->nc_nlen] = 0;
1408 * Rename fncp (unlink)
1410 _cache_unlink_parent(fncp);
1411 oname = fncp->nc_name;
1412 fncp->nc_name = nname;
1413 fncp->nc_nlen = tncp->nc_nlen;
1415 kfree(oname, M_VFSCACHE);
1417 tncp_par = tncp->nc_parent;
1418 _cache_hold(tncp_par);
1419 _cache_lock(tncp_par);
1422 * Rename fncp (relink)
1424 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT);
1425 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash);
1426 nchpp = NCHHASH(hash);
1428 spin_lock(&nchpp->spin);
1429 _cache_link_parent(fncp, tncp_par, nchpp);
1430 spin_unlock(&nchpp->spin);
1432 _cache_put(tncp_par);
1435 * Get rid of the overwritten tncp (unlink)
1437 _cache_unlink(tncp);
1441 * Perform actions consistent with unlinking a file. The passed-in ncp
1444 * The ncp is marked DESTROYED so it no longer shows up in searches,
1445 * and will be physically deleted when the vnode goes away.
1447 * If the related vnode has no refs then we cycle it through vget()/vput()
1448 * to (possibly if we don't have a ref race) trigger a deactivation,
1449 * allowing the VFS to trivially detect and recycle the deleted vnode
1450 * via VOP_INACTIVE().
1452 * NOTE: _cache_rename() will automatically call _cache_unlink() on the
1456 cache_unlink(struct nchandle *nch)
1458 _cache_unlink(nch->ncp);
1462 _cache_unlink(struct namecache *ncp)
1467 * Causes lookups to fail and allows another ncp with the same
1468 * name to be created under ncp->nc_parent.
1470 ncp->nc_flag |= NCF_DESTROYED;
1473 * Attempt to trigger a deactivation.
1475 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1476 (vp = ncp->nc_vp) != NULL &&
1477 !sysref_isactive(&vp->v_sysref)) {
1478 if (vget(vp, LK_SHARED) == 0)
1484 * vget the vnode associated with the namecache entry. Resolve the namecache
1485 * entry if necessary. The passed ncp must be referenced and locked.
1487 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked
1488 * (depending on the passed lk_type) will be returned in *vpp with an error
1489 * of 0, or NULL will be returned in *vpp with a non-0 error code. The
1490 * most typical error is ENOENT, meaning that the ncp represents a negative
1491 * cache hit and there is no vnode to retrieve, but other errors can occur
1494 * The vget() can race a reclaim. If this occurs we re-resolve the
1497 * There are numerous places in the kernel where vget() is called on a
1498 * vnode while one or more of its namecache entries is locked. Releasing
1499 * a vnode never deadlocks against locked namecache entries (the vnode
1500 * will not get recycled while referenced ncp's exist). This means we
1501 * can safely acquire the vnode. In fact, we MUST NOT release the ncp
1502 * lock when acquiring the vp lock or we might cause a deadlock.
1507 cache_vget(struct nchandle *nch, struct ucred *cred,
1508 int lk_type, struct vnode **vpp)
1510 struct namecache *ncp;
1515 KKASSERT(ncp->nc_locktd == curthread);
1518 if (ncp->nc_flag & NCF_UNRESOLVED)
1519 error = cache_resolve(nch, cred);
1523 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1524 error = vget(vp, lk_type);
1529 if (error == ENOENT) {
1530 kprintf("Warning: vnode reclaim race detected "
1531 "in cache_vget on %p (%s)\n",
1533 _cache_setunresolved(ncp);
1538 * Not a reclaim race, some other error.
1540 KKASSERT(ncp->nc_vp == vp);
1543 KKASSERT(ncp->nc_vp == vp);
1544 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1547 if (error == 0 && vp == NULL)
1554 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
1556 struct namecache *ncp;
1561 KKASSERT(ncp->nc_locktd == curthread);
1564 if (ncp->nc_flag & NCF_UNRESOLVED)
1565 error = cache_resolve(nch, cred);
1569 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1570 error = vget(vp, LK_SHARED);
1575 if (error == ENOENT) {
1576 kprintf("Warning: vnode reclaim race detected "
1577 "in cache_vget on %p (%s)\n",
1579 _cache_setunresolved(ncp);
1584 * Not a reclaim race, some other error.
1586 KKASSERT(ncp->nc_vp == vp);
1589 KKASSERT(ncp->nc_vp == vp);
1590 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1591 /* caller does not want a lock */
1595 if (error == 0 && vp == NULL)
1602 * Return a referenced vnode representing the parent directory of
1605 * Because the caller has locked the ncp it should not be possible for
1606 * the parent ncp to go away. However, the parent can unresolve its
1607 * dvp at any time so we must be able to acquire a lock on the parent
1608 * to safely access nc_vp.
1610 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock,
1611 * so use vhold()/vdrop() while holding the lock to prevent dvp from
1612 * getting destroyed.
1614 * MPSAFE - Note vhold() is allowed when dvp has 0 refs if we hold a
1615 * lock on the ncp in question..
1617 static struct vnode *
1618 cache_dvpref(struct namecache *ncp)
1620 struct namecache *par;
1624 if ((par = ncp->nc_parent) != NULL) {
1627 if ((par->nc_flag & NCF_UNRESOLVED) == 0) {
1628 if ((dvp = par->nc_vp) != NULL)
1633 if (vget(dvp, LK_SHARED) == 0) {
1636 /* return refd, unlocked dvp */
1648 * Convert a directory vnode to a namecache record without any other
1649 * knowledge of the topology. This ONLY works with directory vnodes and
1650 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the
1651 * returned ncp (if not NULL) will be held and unlocked.
1653 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1654 * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1655 * for dvp. This will fail only if the directory has been deleted out from
1658 * Callers must always check for a NULL return no matter the value of 'makeit'.
1660 * To avoid underflowing the kernel stack each recursive call increments
1661 * the makeit variable.
1664 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1665 struct vnode *dvp, char *fakename);
1666 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1667 struct vnode **saved_dvp);
1670 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
1671 struct nchandle *nch)
1673 struct vnode *saved_dvp;
1679 nch->mount = dvp->v_mount;
1684 * Handle the makeit == 0 degenerate case
1687 spin_lock(&dvp->v_spin);
1688 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1691 spin_unlock(&dvp->v_spin);
1695 * Loop until resolution, inside code will break out on error.
1699 * Break out if we successfully acquire a working ncp.
1701 spin_lock(&dvp->v_spin);
1702 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1705 spin_unlock(&dvp->v_spin);
1708 spin_unlock(&dvp->v_spin);
1711 * If dvp is the root of its filesystem it should already
1712 * have a namecache pointer associated with it as a side
1713 * effect of the mount, but it may have been disassociated.
1715 if (dvp->v_flag & VROOT) {
1716 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
1717 error = cache_resolve_mp(nch->mount);
1718 _cache_put(nch->ncp);
1720 kprintf("cache_fromdvp: resolve root of mount %p error %d",
1721 dvp->v_mount, error);
1725 kprintf(" failed\n");
1730 kprintf(" succeeded\n");
1735 * If we are recursed too deeply resort to an O(n^2)
1736 * algorithm to resolve the namecache topology. The
1737 * resolved pvp is left referenced in saved_dvp to
1738 * prevent the tree from being destroyed while we loop.
1741 error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1743 kprintf("lookupdotdot(longpath) failed %d "
1744 "dvp %p\n", error, dvp);
1752 * Get the parent directory and resolve its ncp.
1755 kfree(fakename, M_TEMP);
1758 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
1761 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
1767 * Reuse makeit as a recursion depth counter. On success
1768 * nch will be fully referenced.
1770 cache_fromdvp(pvp, cred, makeit + 1, nch);
1772 if (nch->ncp == NULL)
1776 * Do an inefficient scan of pvp (embodied by ncp) to look
1777 * for dvp. This will create a namecache record for dvp on
1778 * success. We loop up to recheck on success.
1780 * ncp and dvp are both held but not locked.
1782 error = cache_inefficient_scan(nch, cred, dvp, fakename);
1784 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1785 pvp, nch->ncp->nc_name, dvp);
1787 /* nch was NULLed out, reload mount */
1788 nch->mount = dvp->v_mount;
1792 kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
1793 pvp, nch->ncp->nc_name);
1796 /* nch was NULLed out, reload mount */
1797 nch->mount = dvp->v_mount;
1801 * If nch->ncp is non-NULL it will have been held already.
1804 kfree(fakename, M_TEMP);
1813 * Go up the chain of parent directories until we find something
1814 * we can resolve into the namecache. This is very inefficient.
1818 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1819 struct vnode **saved_dvp)
1821 struct nchandle nch;
1824 static time_t last_fromdvp_report;
1828 * Loop getting the parent directory vnode until we get something we
1829 * can resolve in the namecache.
1832 nch.mount = dvp->v_mount;
1838 kfree(fakename, M_TEMP);
1841 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
1848 spin_lock(&pvp->v_spin);
1849 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1850 _cache_hold(nch.ncp);
1851 spin_unlock(&pvp->v_spin);
1855 spin_unlock(&pvp->v_spin);
1856 if (pvp->v_flag & VROOT) {
1857 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
1858 error = cache_resolve_mp(nch.mount);
1859 _cache_unlock(nch.ncp);
1862 _cache_drop(nch.ncp);
1872 if (last_fromdvp_report != time_second) {
1873 last_fromdvp_report = time_second;
1874 kprintf("Warning: extremely inefficient path "
1875 "resolution on %s\n",
1878 error = cache_inefficient_scan(&nch, cred, dvp, fakename);
1881 * Hopefully dvp now has a namecache record associated with
1882 * it. Leave it referenced to prevent the kernel from
1883 * recycling the vnode. Otherwise extremely long directory
1884 * paths could result in endless recycling.
1889 _cache_drop(nch.ncp);
1892 kfree(fakename, M_TEMP);
1897 * Do an inefficient scan of the directory represented by ncp looking for
1898 * the directory vnode dvp. ncp must be held but not locked on entry and
1899 * will be held on return. dvp must be refd but not locked on entry and
1900 * will remain refd on return.
1902 * Why do this at all? Well, due to its stateless nature the NFS server
1903 * converts file handles directly to vnodes without necessarily going through
1904 * the namecache ops that would otherwise create the namecache topology
1905 * leading to the vnode. We could either (1) Change the namecache algorithms
1906 * to allow disconnect namecache records that are re-merged opportunistically,
1907 * or (2) Make the NFS server backtrack and scan to recover a connected
1908 * namecache topology in order to then be able to issue new API lookups.
1910 * It turns out that (1) is a huge mess. It takes a nice clean set of
1911 * namecache algorithms and introduces a lot of complication in every subsystem
1912 * that calls into the namecache to deal with the re-merge case, especially
1913 * since we are using the namecache to placehold negative lookups and the
1914 * vnode might not be immediately assigned. (2) is certainly far less
1915 * efficient then (1), but since we are only talking about directories here
1916 * (which are likely to remain cached), the case does not actually run all
1917 * that often and has the supreme advantage of not polluting the namecache
1920 * If a fakename is supplied just construct a namecache entry using the
1924 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1925 struct vnode *dvp, char *fakename)
1927 struct nlcomponent nlc;
1928 struct nchandle rncp;
1940 vat.va_blocksize = 0;
1941 if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1944 error = cache_vref(nch, cred, &pvp);
1949 kprintf("inefficient_scan: directory iosize %ld "
1950 "vattr fileid = %lld\n",
1952 (long long)vat.va_fileid);
1956 * Use the supplied fakename if not NULL. Fake names are typically
1957 * not in the actual filesystem hierarchy. This is used by HAMMER
1958 * to glue @@timestamp recursions together.
1961 nlc.nlc_nameptr = fakename;
1962 nlc.nlc_namelen = strlen(fakename);
1963 rncp = cache_nlookup(nch, &nlc);
1967 if ((blksize = vat.va_blocksize) == 0)
1968 blksize = DEV_BSIZE;
1969 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
1975 iov.iov_base = rbuf;
1976 iov.iov_len = blksize;
1979 uio.uio_resid = blksize;
1980 uio.uio_segflg = UIO_SYSSPACE;
1981 uio.uio_rw = UIO_READ;
1982 uio.uio_td = curthread;
1984 if (ncvp_debug >= 2)
1985 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1986 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1988 den = (struct dirent *)rbuf;
1989 bytes = blksize - uio.uio_resid;
1992 if (ncvp_debug >= 2) {
1993 kprintf("cache_inefficient_scan: %*.*s\n",
1994 den->d_namlen, den->d_namlen,
1997 if (den->d_type != DT_WHT &&
1998 den->d_ino == vat.va_fileid) {
2000 kprintf("cache_inefficient_scan: "
2001 "MATCHED inode %lld path %s/%*.*s\n",
2002 (long long)vat.va_fileid,
2004 den->d_namlen, den->d_namlen,
2007 nlc.nlc_nameptr = den->d_name;
2008 nlc.nlc_namelen = den->d_namlen;
2009 rncp = cache_nlookup(nch, &nlc);
2010 KKASSERT(rncp.ncp != NULL);
2013 bytes -= _DIRENT_DIRSIZ(den);
2014 den = _DIRENT_NEXT(den);
2016 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
2019 kfree(rbuf, M_TEMP);
2023 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
2024 _cache_setvp(rncp.mount, rncp.ncp, dvp);
2025 if (ncvp_debug >= 2) {
2026 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
2027 nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
2030 if (ncvp_debug >= 2) {
2031 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
2032 nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
2036 if (rncp.ncp->nc_vp == NULL)
2037 error = rncp.ncp->nc_error;
2039 * Release rncp after a successful nlookup. rncp was fully
2044 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
2045 dvp, nch->ncp->nc_name);
2052 * Zap a namecache entry. The ncp is unconditionally set to an unresolved
2053 * state, which disassociates it from its vnode or ncneglist.
2055 * Then, if there are no additional references to the ncp and no children,
2056 * the ncp is removed from the topology and destroyed.
2058 * References and/or children may exist if the ncp is in the middle of the
2059 * topology, preventing the ncp from being destroyed.
2061 * This function must be called with the ncp held and locked and will unlock
2062 * and drop it during zapping.
2064 * If nonblock is non-zero and the parent ncp cannot be locked we give up.
2065 * This case can occur in the cache_drop() path.
2067 * This function may returned a held (but NOT locked) parent node which the
2068 * caller must drop. We do this so _cache_drop() can loop, to avoid
2069 * blowing out the kernel stack.
2071 * WARNING! For MPSAFE operation this routine must acquire up to three
2072 * spin locks to be able to safely test nc_refs. Lock order is
2075 * hash spinlock if on hash list
2076 * parent spinlock if child of parent
2077 * (the ncp is unresolved so there is no vnode association)
2079 static struct namecache *
2080 cache_zap(struct namecache *ncp, int nonblock)
2082 struct namecache *par;
2083 struct vnode *dropvp;
2087 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
2089 _cache_setunresolved(ncp);
2092 * Try to scrap the entry and possibly tail-recurse on its parent.
2093 * We only scrap unref'd (other then our ref) unresolved entries,
2094 * we do not scrap 'live' entries.
2096 * Note that once the spinlocks are acquired if nc_refs == 1 no
2097 * other references are possible. If it isn't, however, we have
2098 * to decrement but also be sure to avoid a 1->0 transition.
2100 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
2101 KKASSERT(ncp->nc_refs > 0);
2104 * Acquire locks. Note that the parent can't go away while we hold
2107 if ((par = ncp->nc_parent) != NULL) {
2110 if (_cache_lock_nonblock(par) == 0)
2112 refs = ncp->nc_refs;
2113 ncp->nc_flag |= NCF_DEFEREDZAP;
2114 ++numdefered; /* MP race ok */
2115 if (atomic_cmpset_int(&ncp->nc_refs,
2127 spin_lock(&ncp->nc_head->spin);
2131 * If someone other then us has a ref or we have children
2132 * we cannot zap the entry. The 1->0 transition and any
2133 * further list operation is protected by the spinlocks
2134 * we have acquired but other transitions are not.
2137 refs = ncp->nc_refs;
2138 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list))
2140 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) {
2142 spin_unlock(&ncp->nc_head->spin);
2152 * We are the only ref and with the spinlocks held no further
2153 * refs can be acquired by others.
2155 * Remove us from the hash list and parent list. We have to
2156 * drop a ref on the parent's vp if the parent's list becomes
2161 struct nchash_head *nchpp = ncp->nc_head;
2163 KKASSERT(nchpp != NULL);
2164 LIST_REMOVE(ncp, nc_hash);
2165 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
2166 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
2167 dropvp = par->nc_vp;
2168 ncp->nc_head = NULL;
2169 ncp->nc_parent = NULL;
2170 spin_unlock(&nchpp->spin);
2173 KKASSERT(ncp->nc_head == NULL);
2177 * ncp should not have picked up any refs. Physically
2180 KKASSERT(ncp->nc_refs == 1);
2181 /* _cache_unlock(ncp) not required */
2182 ncp->nc_refs = -1; /* safety */
2184 kfree(ncp->nc_name, M_VFSCACHE);
2185 kfree(ncp, M_VFSCACHE);
2188 * Delayed drop (we had to release our spinlocks)
2190 * The refed parent (if not NULL) must be dropped. The
2191 * caller is responsible for looping.
2199 * Clean up dangling negative cache and defered-drop entries in the
2202 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t;
2204 static cache_hs_t neg_cache_hysteresis_state = CHI_LOW;
2205 static cache_hs_t pos_cache_hysteresis_state = CHI_LOW;
2208 cache_hysteresis(void)
2213 * Don't cache too many negative hits. We use hysteresis to reduce
2214 * the impact on the critical path.
2216 switch(neg_cache_hysteresis_state) {
2218 if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
2219 _cache_cleanneg(10);
2220 neg_cache_hysteresis_state = CHI_HIGH;
2224 if (numneg > MINNEG * 9 / 10 &&
2225 numneg * ncnegfactor * 9 / 10 > numcache
2227 _cache_cleanneg(10);
2229 neg_cache_hysteresis_state = CHI_LOW;
2235 * Don't cache too many positive hits. We use hysteresis to reduce
2236 * the impact on the critical path.
2238 * Excessive positive hits can accumulate due to large numbers of
2239 * hardlinks (the vnode cache will not prevent hl ncps from growing
2242 if ((poslimit = ncposlimit) == 0)
2243 poslimit = desiredvnodes * 2;
2245 switch(pos_cache_hysteresis_state) {
2247 if (numcache > poslimit && numcache > MINPOS) {
2248 _cache_cleanpos(10);
2249 pos_cache_hysteresis_state = CHI_HIGH;
2253 if (numcache > poslimit * 5 / 6 && numcache > MINPOS) {
2254 _cache_cleanpos(10);
2256 pos_cache_hysteresis_state = CHI_LOW;
2262 * Clean out dangling defered-zap ncps which could not
2263 * be cleanly dropped if too many build up. Note
2264 * that numdefered is not an exact number as such ncps
2265 * can be reused and the counter is not handled in a MP
2266 * safe manner by design.
2268 if (numdefered * ncnegfactor > numcache) {
2269 _cache_cleandefered();
2274 * NEW NAMECACHE LOOKUP API
2276 * Lookup an entry in the namecache. The passed par_nch must be referenced
2277 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp
2278 * is ALWAYS returned, eve if the supplied component is illegal.
2280 * The resulting namecache entry should be returned to the system with
2281 * cache_put() or cache_unlock() + cache_drop().
2283 * namecache locks are recursive but care must be taken to avoid lock order
2284 * reversals (hence why the passed par_nch must be unlocked). Locking
2285 * rules are to order for parent traversals, not for child traversals.
2287 * Nobody else will be able to manipulate the associated namespace (e.g.
2288 * create, delete, rename, rename-target) until the caller unlocks the
2291 * The returned entry will be in one of three states: positive hit (non-null
2292 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
2293 * Unresolved entries must be resolved through the filesystem to associate the
2294 * vnode and/or determine whether a positive or negative hit has occured.
2296 * It is not necessary to lock a directory in order to lock namespace under
2297 * that directory. In fact, it is explicitly not allowed to do that. A
2298 * directory is typically only locked when being created, renamed, or
2301 * The directory (par) may be unresolved, in which case any returned child
2302 * will likely also be marked unresolved. Likely but not guarenteed. Since
2303 * the filesystem lookup requires a resolved directory vnode the caller is
2304 * responsible for resolving the namecache chain top-down. This API
2305 * specifically allows whole chains to be created in an unresolved state.
2308 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
2310 struct nchandle nch;
2311 struct namecache *ncp;
2312 struct namecache *new_ncp;
2313 struct nchash_head *nchpp;
2321 mp = par_nch->mount;
2325 * This is a good time to call it, no ncp's are locked by
2331 * Try to locate an existing entry
2333 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2334 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2336 nchpp = NCHHASH(hash);
2338 spin_lock(&nchpp->spin);
2339 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2343 * Break out if we find a matching entry. Note that
2344 * UNRESOLVED entries may match, but DESTROYED entries
2347 if (ncp->nc_parent == par_nch->ncp &&
2348 ncp->nc_nlen == nlc->nlc_namelen &&
2349 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2350 (ncp->nc_flag & NCF_DESTROYED) == 0
2353 spin_unlock(&nchpp->spin);
2355 _cache_unlock(par_nch->ncp);
2358 if (_cache_lock_special(ncp) == 0) {
2359 _cache_auto_unresolve(mp, ncp);
2361 _cache_free(new_ncp);
2372 * We failed to locate an entry, create a new entry and add it to
2373 * the cache. The parent ncp must also be locked so we
2376 * We have to relookup after possibly blocking in kmalloc or
2377 * when locking par_nch.
2379 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2380 * mount case, in which case nc_name will be NULL.
2382 if (new_ncp == NULL) {
2383 spin_unlock(&nchpp->spin);
2384 new_ncp = cache_alloc(nlc->nlc_namelen);
2385 if (nlc->nlc_namelen) {
2386 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2388 new_ncp->nc_name[nlc->nlc_namelen] = 0;
2392 if (par_locked == 0) {
2393 spin_unlock(&nchpp->spin);
2394 _cache_lock(par_nch->ncp);
2400 * WARNING! We still hold the spinlock. We have to set the hash
2401 * table entry atomically.
2404 _cache_link_parent(ncp, par_nch->ncp, nchpp);
2405 spin_unlock(&nchpp->spin);
2406 _cache_unlock(par_nch->ncp);
2407 /* par_locked = 0 - not used */
2410 * stats and namecache size management
2412 if (ncp->nc_flag & NCF_UNRESOLVED)
2413 ++gd->gd_nchstats->ncs_miss;
2414 else if (ncp->nc_vp)
2415 ++gd->gd_nchstats->ncs_goodhits;
2417 ++gd->gd_nchstats->ncs_neghits;
2420 atomic_add_int(&nch.mount->mnt_refs, 1);
2425 * This is a non-blocking verison of cache_nlookup() used by
2426 * nfs_readdirplusrpc_uio(). It can fail for any reason and
2427 * will return nch.ncp == NULL in that case.
2430 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc)
2432 struct nchandle nch;
2433 struct namecache *ncp;
2434 struct namecache *new_ncp;
2435 struct nchash_head *nchpp;
2443 mp = par_nch->mount;
2447 * Try to locate an existing entry
2449 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2450 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2452 nchpp = NCHHASH(hash);
2454 spin_lock(&nchpp->spin);
2455 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2459 * Break out if we find a matching entry. Note that
2460 * UNRESOLVED entries may match, but DESTROYED entries
2463 if (ncp->nc_parent == par_nch->ncp &&
2464 ncp->nc_nlen == nlc->nlc_namelen &&
2465 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2466 (ncp->nc_flag & NCF_DESTROYED) == 0
2469 spin_unlock(&nchpp->spin);
2471 _cache_unlock(par_nch->ncp);
2474 if (_cache_lock_special(ncp) == 0) {
2475 _cache_auto_unresolve(mp, ncp);
2477 _cache_free(new_ncp);
2488 * We failed to locate an entry, create a new entry and add it to
2489 * the cache. The parent ncp must also be locked so we
2492 * We have to relookup after possibly blocking in kmalloc or
2493 * when locking par_nch.
2495 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2496 * mount case, in which case nc_name will be NULL.
2498 if (new_ncp == NULL) {
2499 spin_unlock(&nchpp->spin);
2500 new_ncp = cache_alloc(nlc->nlc_namelen);
2501 if (nlc->nlc_namelen) {
2502 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2504 new_ncp->nc_name[nlc->nlc_namelen] = 0;
2508 if (par_locked == 0) {
2509 spin_unlock(&nchpp->spin);
2510 if (_cache_lock_nonblock(par_nch->ncp) == 0) {
2518 * WARNING! We still hold the spinlock. We have to set the hash
2519 * table entry atomically.
2522 _cache_link_parent(ncp, par_nch->ncp, nchpp);
2523 spin_unlock(&nchpp->spin);
2524 _cache_unlock(par_nch->ncp);
2525 /* par_locked = 0 - not used */
2528 * stats and namecache size management
2530 if (ncp->nc_flag & NCF_UNRESOLVED)
2531 ++gd->gd_nchstats->ncs_miss;
2532 else if (ncp->nc_vp)
2533 ++gd->gd_nchstats->ncs_goodhits;
2535 ++gd->gd_nchstats->ncs_neghits;
2538 atomic_add_int(&nch.mount->mnt_refs, 1);
2542 _cache_free(new_ncp);
2551 * The namecache entry is marked as being used as a mount point.
2552 * Locate the mount if it is visible to the caller. The DragonFly
2553 * mount system allows arbitrary loops in the topology and disentangles
2554 * those loops by matching against (mp, ncp) rather than just (ncp).
2555 * This means any given ncp can dive any number of mounts, depending
2556 * on the relative mount (e.g. nullfs) the caller is at in the topology.
2558 * We use a very simple frontend cache to reduce SMP conflicts,
2559 * which we have to do because the mountlist scan needs an exclusive
2560 * lock around its ripout info list. Not to mention that there might
2561 * be a lot of mounts.
2563 struct findmount_info {
2564 struct mount *result;
2565 struct mount *nch_mount;
2566 struct namecache *nch_ncp;
2570 struct ncmount_cache *
2571 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp)
2575 hash = ((int)(intptr_t)mp / sizeof(*mp)) ^
2576 ((int)(intptr_t)ncp / sizeof(*ncp));
2577 hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE;
2578 return (&ncmount_cache[hash]);
2583 cache_findmount_callback(struct mount *mp, void *data)
2585 struct findmount_info *info = data;
2588 * Check the mount's mounted-on point against the passed nch.
2590 if (mp->mnt_ncmounton.mount == info->nch_mount &&
2591 mp->mnt_ncmounton.ncp == info->nch_ncp
2594 atomic_add_int(&mp->mnt_refs, 1);
2601 cache_findmount(struct nchandle *nch)
2603 struct findmount_info info;
2604 struct ncmount_cache *ncc;
2610 if (ncmount_cache_enable == 0) {
2614 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
2615 if (ncc->ncp == nch->ncp) {
2616 spin_lock_shared(&ncc->spin);
2617 if (ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) {
2618 if (mp->mnt_ncmounton.mount == nch->mount &&
2619 mp->mnt_ncmounton.ncp == nch->ncp) {
2620 atomic_add_int(&mp->mnt_refs, 1);
2621 spin_unlock_shared(&ncc->spin);
2622 ++ncmount_cache_hit;
2626 spin_unlock_shared(&ncc->spin);
2634 info.nch_mount = nch->mount;
2635 info.nch_ncp = nch->ncp;
2636 mountlist_scan(cache_findmount_callback, &info,
2637 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
2639 if (info.result && ncc) {
2640 spin_lock(&ncc->spin);
2641 if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) {
2643 atomic_add_int(&ncc->mp->mnt_refs, -1);
2644 atomic_add_int(&info.result->mnt_refs, 1);
2645 ncc->mp = info.result;
2646 ncc->ncp = nch->ncp;
2647 spin_unlock(&ncc->spin);
2648 ++ncmount_cache_overwrite;
2650 spin_unlock(&ncc->spin);
2652 ++ncmount_cache_miss;
2654 return(info.result);
2658 cache_dropmount(struct mount *mp)
2660 atomic_add_int(&mp->mnt_refs, -1);
2664 cache_unmounting(struct mount *mp)
2666 struct nchandle *nch = &mp->mnt_ncmounton;
2667 struct ncmount_cache *ncc;
2669 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
2670 if (ncc->ncp == nch->ncp && ncc->mp == mp) {
2671 spin_lock(&ncc->spin);
2672 if (ncc->ncp == nch->ncp && ncc->mp == mp) {
2673 atomic_add_int(&mp->mnt_refs, -1);
2677 spin_unlock(&ncc->spin);
2682 * Resolve an unresolved namecache entry, generally by looking it up.
2683 * The passed ncp must be locked and refd.
2685 * Theoretically since a vnode cannot be recycled while held, and since
2686 * the nc_parent chain holds its vnode as long as children exist, the
2687 * direct parent of the cache entry we are trying to resolve should
2688 * have a valid vnode. If not then generate an error that we can
2689 * determine is related to a resolver bug.
2691 * However, if a vnode was in the middle of a recyclement when the NCP
2692 * got locked, ncp->nc_vp might point to a vnode that is about to become
2693 * invalid. cache_resolve() handles this case by unresolving the entry
2694 * and then re-resolving it.
2696 * Note that successful resolution does not necessarily return an error
2697 * code of 0. If the ncp resolves to a negative cache hit then ENOENT
2703 cache_resolve(struct nchandle *nch, struct ucred *cred)
2705 struct namecache *par_tmp;
2706 struct namecache *par;
2707 struct namecache *ncp;
2708 struct nchandle nctmp;
2717 * If the ncp is already resolved we have nothing to do. However,
2718 * we do want to guarentee that a usable vnode is returned when
2719 * a vnode is present, so make sure it hasn't been reclaimed.
2721 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2722 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
2723 _cache_setunresolved(ncp);
2724 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
2725 return (ncp->nc_error);
2729 * If the ncp was destroyed it will never resolve again. This
2730 * can basically only happen when someone is chdir'd into an
2731 * empty directory which is then rmdir'd. We want to catch this
2732 * here and not dive the VFS because the VFS might actually
2733 * have a way to re-resolve the disconnected ncp, which will
2734 * result in inconsistencies in the cdir/nch for proc->p_fd.
2736 if (ncp->nc_flag & NCF_DESTROYED) {
2737 kprintf("Warning: cache_resolve: ncp '%s' was unlinked\n",
2743 * Mount points need special handling because the parent does not
2744 * belong to the same filesystem as the ncp.
2746 if (ncp == mp->mnt_ncmountpt.ncp)
2747 return (cache_resolve_mp(mp));
2750 * We expect an unbroken chain of ncps to at least the mount point,
2751 * and even all the way to root (but this code doesn't have to go
2752 * past the mount point).
2754 if (ncp->nc_parent == NULL) {
2755 kprintf("EXDEV case 1 %p %*.*s\n", ncp,
2756 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2757 ncp->nc_error = EXDEV;
2758 return(ncp->nc_error);
2762 * The vp's of the parent directories in the chain are held via vhold()
2763 * due to the existance of the child, and should not disappear.
2764 * However, there are cases where they can disappear:
2766 * - due to filesystem I/O errors.
2767 * - due to NFS being stupid about tracking the namespace and
2768 * destroys the namespace for entire directories quite often.
2769 * - due to forced unmounts.
2770 * - due to an rmdir (parent will be marked DESTROYED)
2772 * When this occurs we have to track the chain backwards and resolve
2773 * it, looping until the resolver catches up to the current node. We
2774 * could recurse here but we might run ourselves out of kernel stack
2775 * so we do it in a more painful manner. This situation really should
2776 * not occur all that often, or if it does not have to go back too
2777 * many nodes to resolve the ncp.
2779 while ((dvp = cache_dvpref(ncp)) == NULL) {
2781 * This case can occur if a process is CD'd into a
2782 * directory which is then rmdir'd. If the parent is marked
2783 * destroyed there is no point trying to resolve it.
2785 if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
2787 par = ncp->nc_parent;
2790 while ((par_tmp = par->nc_parent) != NULL &&
2791 par_tmp->nc_vp == NULL) {
2792 _cache_hold(par_tmp);
2793 _cache_lock(par_tmp);
2797 if (par->nc_parent == NULL) {
2798 kprintf("EXDEV case 2 %*.*s\n",
2799 par->nc_nlen, par->nc_nlen, par->nc_name);
2803 kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
2804 par->nc_nlen, par->nc_nlen, par->nc_name);
2806 * The parent is not set in stone, ref and lock it to prevent
2807 * it from disappearing. Also note that due to renames it
2808 * is possible for our ncp to move and for par to no longer
2809 * be one of its parents. We resolve it anyway, the loop
2810 * will handle any moves.
2812 _cache_get(par); /* additional hold/lock */
2813 _cache_put(par); /* from earlier hold/lock */
2814 if (par == nch->mount->mnt_ncmountpt.ncp) {
2815 cache_resolve_mp(nch->mount);
2816 } else if ((dvp = cache_dvpref(par)) == NULL) {
2817 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
2821 if (par->nc_flag & NCF_UNRESOLVED) {
2824 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2828 if ((error = par->nc_error) != 0) {
2829 if (par->nc_error != EAGAIN) {
2830 kprintf("EXDEV case 3 %*.*s error %d\n",
2831 par->nc_nlen, par->nc_nlen, par->nc_name,
2836 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
2837 par, par->nc_nlen, par->nc_nlen, par->nc_name);
2844 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
2845 * ncp's and reattach them. If this occurs the original ncp is marked
2846 * EAGAIN to force a relookup.
2848 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
2849 * ncp must already be resolved.
2854 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2857 ncp->nc_error = EPERM;
2859 if (ncp->nc_error == EAGAIN) {
2860 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
2861 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2864 return(ncp->nc_error);
2868 * Resolve the ncp associated with a mount point. Such ncp's almost always
2869 * remain resolved and this routine is rarely called. NFS MPs tends to force
2870 * re-resolution more often due to its mac-truck-smash-the-namecache
2871 * method of tracking namespace changes.
2873 * The semantics for this call is that the passed ncp must be locked on
2874 * entry and will be locked on return. However, if we actually have to
2875 * resolve the mount point we temporarily unlock the entry in order to
2876 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of
2877 * the unlock we have to recheck the flags after we relock.
2880 cache_resolve_mp(struct mount *mp)
2882 struct namecache *ncp = mp->mnt_ncmountpt.ncp;
2886 KKASSERT(mp != NULL);
2889 * If the ncp is already resolved we have nothing to do. However,
2890 * we do want to guarentee that a usable vnode is returned when
2891 * a vnode is present, so make sure it hasn't been reclaimed.
2893 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2894 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
2895 _cache_setunresolved(ncp);
2898 if (ncp->nc_flag & NCF_UNRESOLVED) {
2900 while (vfs_busy(mp, 0))
2902 error = VFS_ROOT(mp, &vp);
2906 * recheck the ncp state after relocking.
2908 if (ncp->nc_flag & NCF_UNRESOLVED) {
2909 ncp->nc_error = error;
2911 _cache_setvp(mp, ncp, vp);
2914 kprintf("[diagnostic] cache_resolve_mp: failed"
2915 " to resolve mount %p err=%d ncp=%p\n",
2917 _cache_setvp(mp, ncp, NULL);
2919 } else if (error == 0) {
2924 return(ncp->nc_error);
2928 * Clean out negative cache entries when too many have accumulated.
2933 _cache_cleanneg(int count)
2935 struct namecache *ncp;
2938 * Attempt to clean out the specified number of negative cache
2943 ncp = TAILQ_FIRST(&ncneglist);
2945 spin_unlock(&ncspin);
2948 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
2949 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
2951 spin_unlock(&ncspin);
2954 * This can race, so we must re-check that the ncp
2955 * is on the ncneglist after successfully locking it.
2957 if (_cache_lock_special(ncp) == 0) {
2958 if (ncp->nc_vp == NULL &&
2959 (ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2960 ncp = cache_zap(ncp, 1);
2964 kprintf("cache_cleanneg: race avoided\n");
2975 * Clean out positive cache entries when too many have accumulated.
2980 _cache_cleanpos(int count)
2982 static volatile int rover;
2983 struct nchash_head *nchpp;
2984 struct namecache *ncp;
2988 * Attempt to clean out the specified number of negative cache
2992 rover_copy = ++rover; /* MPSAFEENOUGH */
2994 nchpp = NCHHASH(rover_copy);
2996 spin_lock(&nchpp->spin);
2997 ncp = LIST_FIRST(&nchpp->list);
3000 spin_unlock(&nchpp->spin);
3003 if (_cache_lock_special(ncp) == 0) {
3004 ncp = cache_zap(ncp, 1);
3016 * This is a kitchen sink function to clean out ncps which we
3017 * tried to zap from cache_drop() but failed because we were
3018 * unable to acquire the parent lock.
3020 * Such entries can also be removed via cache_inval_vp(), such
3021 * as when unmounting.
3026 _cache_cleandefered(void)
3028 struct nchash_head *nchpp;
3029 struct namecache *ncp;
3030 struct namecache dummy;
3034 bzero(&dummy, sizeof(dummy));
3035 dummy.nc_flag = NCF_DESTROYED;
3037 for (i = 0; i <= nchash; ++i) {
3038 nchpp = &nchashtbl[i];
3040 spin_lock(&nchpp->spin);
3041 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash);
3043 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) {
3044 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0)
3046 LIST_REMOVE(&dummy, nc_hash);
3047 LIST_INSERT_AFTER(ncp, &dummy, nc_hash);
3049 spin_unlock(&nchpp->spin);
3050 if (_cache_lock_nonblock(ncp) == 0) {
3051 ncp->nc_flag &= ~NCF_DEFEREDZAP;
3055 spin_lock(&nchpp->spin);
3058 LIST_REMOVE(&dummy, nc_hash);
3059 spin_unlock(&nchpp->spin);
3064 * Name cache initialization, from vfsinit() when we are booting
3072 /* initialise per-cpu namecache effectiveness statistics. */
3073 for (i = 0; i < ncpus; ++i) {
3074 gd = globaldata_find(i);
3075 gd->gd_nchstats = &nchstats[i];
3077 TAILQ_INIT(&ncneglist);
3079 nchashtbl = hashinit_ext(desiredvnodes / 2,
3080 sizeof(struct nchash_head),
3081 M_VFSCACHE, &nchash);
3082 for (i = 0; i <= (int)nchash; ++i) {
3083 LIST_INIT(&nchashtbl[i].list);
3084 spin_init(&nchashtbl[i].spin);
3086 for (i = 0; i < NCMOUNT_NUMCACHE; ++i)
3087 spin_init(&ncmount_cache[i].spin);
3088 nclockwarn = 5 * hz;
3092 * Called from start_init() to bootstrap the root filesystem. Returns
3093 * a referenced, unlocked namecache record.
3096 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
3098 nch->ncp = cache_alloc(0);
3100 atomic_add_int(&mp->mnt_refs, 1);
3102 _cache_setvp(nch->mount, nch->ncp, vp);
3106 * vfs_cache_setroot()
3108 * Create an association between the root of our namecache and
3109 * the root vnode. This routine may be called several times during
3112 * If the caller intends to save the returned namecache pointer somewhere
3113 * it must cache_hold() it.
3116 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
3119 struct nchandle onch;
3127 cache_zero(&rootnch);
3135 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache
3136 * topology and is being removed as quickly as possible. The new VOP_N*()
3137 * API calls are required to make specific adjustments using the supplied
3138 * ncp pointers rather then just bogusly purging random vnodes.
3140 * Invalidate all namecache entries to a particular vnode as well as
3141 * any direct children of that vnode in the namecache. This is a
3142 * 'catch all' purge used by filesystems that do not know any better.
3144 * Note that the linkage between the vnode and its namecache entries will
3145 * be removed, but the namecache entries themselves might stay put due to
3146 * active references from elsewhere in the system or due to the existance of
3147 * the children. The namecache topology is left intact even if we do not
3148 * know what the vnode association is. Such entries will be marked
3152 cache_purge(struct vnode *vp)
3154 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
3158 * Flush all entries referencing a particular filesystem.
3160 * Since we need to check it anyway, we will flush all the invalid
3161 * entries at the same time.
3166 cache_purgevfs(struct mount *mp)
3168 struct nchash_head *nchpp;
3169 struct namecache *ncp, *nnp;
3172 * Scan hash tables for applicable entries.
3174 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
3175 spin_lock_wr(&nchpp->spin); XXX
3176 ncp = LIST_FIRST(&nchpp->list);
3180 nnp = LIST_NEXT(ncp, nc_hash);
3183 if (ncp->nc_mount == mp) {
3185 ncp = cache_zap(ncp, 0);
3193 spin_unlock_wr(&nchpp->spin); XXX
3199 static int disablecwd;
3200 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
3203 static u_long numcwdcalls;
3204 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0,
3205 "Number of current directory resolution calls");
3206 static u_long numcwdfailnf;
3207 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0,
3208 "Number of current directory failures due to lack of file");
3209 static u_long numcwdfailsz;
3210 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0,
3211 "Number of current directory failures due to large result");
3212 static u_long numcwdfound;
3213 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0,
3214 "Number of current directory resolution successes");
3220 sys___getcwd(struct __getcwd_args *uap)
3230 buflen = uap->buflen;
3233 if (buflen > MAXPATHLEN)
3234 buflen = MAXPATHLEN;
3236 buf = kmalloc(buflen, M_TEMP, M_WAITOK);
3237 bp = kern_getcwd(buf, buflen, &error);
3239 error = copyout(bp, uap->buf, strlen(bp) + 1);
3245 kern_getcwd(char *buf, size_t buflen, int *error)
3247 struct proc *p = curproc;
3249 int i, slash_prefixed;
3250 struct filedesc *fdp;
3251 struct nchandle nch;
3252 struct namecache *ncp;
3261 nch = fdp->fd_ncdir;
3266 while (ncp && (ncp != fdp->fd_nrdir.ncp ||
3267 nch.mount != fdp->fd_nrdir.mount)
3270 * While traversing upwards if we encounter the root
3271 * of the current mount we have to skip to the mount point
3272 * in the underlying filesystem.
3274 if (ncp == nch.mount->mnt_ncmountpt.ncp) {
3275 nch = nch.mount->mnt_ncmounton;
3284 * Prepend the path segment
3286 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3293 *--bp = ncp->nc_name[i];
3305 * Go up a directory. This isn't a mount point so we don't
3306 * have to check again.
3308 while ((nch.ncp = ncp->nc_parent) != NULL) {
3310 if (nch.ncp != ncp->nc_parent) {
3314 _cache_hold(nch.ncp);
3327 if (!slash_prefixed) {
3345 * Thus begins the fullpath magic.
3347 * The passed nchp is referenced but not locked.
3349 static int disablefullpath;
3350 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
3351 &disablefullpath, 0,
3352 "Disable fullpath lookups");
3354 static u_int numfullpathcalls;
3355 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD,
3356 &numfullpathcalls, 0,
3357 "Number of full path resolutions in progress");
3358 static u_int numfullpathfailnf;
3359 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD,
3360 &numfullpathfailnf, 0,
3361 "Number of full path resolution failures due to lack of file");
3362 static u_int numfullpathfailsz;
3363 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD,
3364 &numfullpathfailsz, 0,
3365 "Number of full path resolution failures due to insufficient memory");
3366 static u_int numfullpathfound;
3367 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD,
3368 &numfullpathfound, 0,
3369 "Number of full path resolution successes");
3372 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase,
3373 char **retbuf, char **freebuf, int guess)
3375 struct nchandle fd_nrdir;
3376 struct nchandle nch;
3377 struct namecache *ncp;
3378 struct mount *mp, *new_mp;
3384 atomic_add_int(&numfullpathcalls, -1);
3389 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
3390 bp = buf + MAXPATHLEN - 1;
3393 fd_nrdir = *nchbase;
3395 fd_nrdir = p->p_fd->fd_nrdir;
3405 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) {
3409 * If we are asked to guess the upwards path, we do so whenever
3410 * we encounter an ncp marked as a mountpoint. We try to find
3411 * the actual mountpoint by finding the mountpoint with this
3414 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) {
3415 new_mp = mount_get_by_nc(ncp);
3418 * While traversing upwards if we encounter the root
3419 * of the current mount we have to skip to the mount point.
3421 if (ncp == mp->mnt_ncmountpt.ncp) {
3425 nch = new_mp->mnt_ncmounton;
3435 * Prepend the path segment
3437 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3439 numfullpathfailsz++;
3444 *--bp = ncp->nc_name[i];
3447 numfullpathfailsz++;
3456 * Go up a directory. This isn't a mount point so we don't
3457 * have to check again.
3459 * We can only safely access nc_parent with ncp held locked.
3461 while ((nch.ncp = ncp->nc_parent) != NULL) {
3463 if (nch.ncp != ncp->nc_parent) {
3467 _cache_hold(nch.ncp);
3475 numfullpathfailnf++;
3481 if (!slash_prefixed) {
3483 numfullpathfailsz++;
3501 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf,
3504 struct namecache *ncp;
3505 struct nchandle nch;
3509 atomic_add_int(&numfullpathcalls, 1);
3510 if (disablefullpath)
3516 /* vn is NULL, client wants us to use p->p_textvp */
3518 if ((vn = p->p_textvp) == NULL)
3521 spin_lock(&vn->v_spin);
3522 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
3527 spin_unlock(&vn->v_spin);
3531 spin_unlock(&vn->v_spin);
3533 atomic_add_int(&numfullpathcalls, -1);
3535 nch.mount = vn->v_mount;
3536 error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess);