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])
130 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
132 LIST_HEAD(nchash_list, namecache);
135 struct nchash_list list;
136 struct spinlock spin;
139 static struct nchash_head *nchashtbl;
140 static struct namecache_list ncneglist;
141 static struct spinlock ncspin;
144 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server
145 * to create the namecache infrastructure leading to a dangling vnode.
147 * 0 Only errors are reported
148 * 1 Successes are reported
149 * 2 Successes + the whole directory scan is reported
150 * 3 Force the directory scan code run as if the parent vnode did not
151 * have a namecache record, even if it does have one.
153 static int ncvp_debug;
154 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0,
155 "Namecache debug level (0-3)");
157 static u_long nchash; /* size of hash table */
158 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
159 "Size of namecache hash table");
161 static int ncnegfactor = 16; /* ratio of negative entries */
162 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
163 "Ratio of namecache negative entries");
165 static int nclockwarn; /* warn on locked entries in ticks */
166 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0,
167 "Warn on locked namecache entries in ticks");
169 static int numdefered; /* number of cache entries allocated */
170 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0,
171 "Number of cache entries allocated");
173 static int ncposlimit; /* number of cache entries allocated */
174 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0,
175 "Number of cache entries allocated");
177 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode),
178 "sizeof(struct vnode)");
179 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache),
180 "sizeof(struct namecache)");
182 int cache_mpsafe = 1;
183 SYSCTL_INT(_vfs, OID_AUTO, cache_mpsafe, CTLFLAG_RW, &cache_mpsafe, 0, "");
185 static int cache_resolve_mp(struct mount *mp);
186 static struct vnode *cache_dvpref(struct namecache *ncp);
187 static void _cache_lock(struct namecache *ncp);
188 static void _cache_setunresolved(struct namecache *ncp);
189 static void _cache_cleanneg(int count);
190 static void _cache_cleanpos(int count);
191 static void _cache_cleandefered(void);
194 * The new name cache statistics
196 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
197 #define STATNODE(mode, name, var) \
198 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, "");
199 #define STATNODE_INT(mode, name, var) \
200 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, mode, var, 0, "");
201 static int numneg; STATNODE_INT(CTLFLAG_RD, numneg, &numneg);
202 static int numcache; STATNODE_INT(CTLFLAG_RD, numcache, &numcache);
203 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls);
204 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits);
205 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits);
206 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks);
207 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss);
208 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap);
209 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps);
210 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits);
211 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps);
212 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits);
214 struct nchstats nchstats[SMP_MAXCPU];
216 * Export VFS cache effectiveness statistics to user-land.
218 * The statistics are left for aggregation to user-land so
219 * neat things can be achieved, like observing per-CPU cache
223 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
225 struct globaldata *gd;
229 for (i = 0; i < ncpus; ++i) {
230 gd = globaldata_find(i);
231 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
232 sizeof(struct nchstats))))
238 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
239 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
241 static struct namecache *cache_zap(struct namecache *ncp, int nonblock);
244 * Namespace locking. The caller must already hold a reference to the
245 * namecache structure in order to lock/unlock it. This function prevents
246 * the namespace from being created or destroyed by accessors other then
249 * Note that holding a locked namecache structure prevents other threads
250 * from making namespace changes (e.g. deleting or creating), prevents
251 * vnode association state changes by other threads, and prevents the
252 * namecache entry from being resolved or unresolved by other threads.
254 * The lock owner has full authority to associate/disassociate vnodes
255 * and resolve/unresolve the locked ncp.
257 * The primary lock field is nc_exlocks. nc_locktd is set after the
258 * fact (when locking) or cleared prior to unlocking.
260 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed
261 * or recycled, but it does NOT help you if the vnode had already
262 * initiated a recyclement. If this is important, use cache_get()
263 * rather then cache_lock() (and deal with the differences in the
264 * way the refs counter is handled). Or, alternatively, make an
265 * unconditional call to cache_validate() or cache_resolve()
266 * after cache_lock() returns.
272 _cache_lock(struct namecache *ncp)
279 KKASSERT(ncp->nc_refs != 0);
284 count = ncp->nc_exlocks;
287 if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) {
289 * The vp associated with a locked ncp must
290 * be held to prevent it from being recycled.
292 * WARNING! If VRECLAIMED is set the vnode
293 * could already be in the middle of a recycle.
294 * Callers must use cache_vref() or
295 * cache_vget() on the locked ncp to
296 * validate the vp or set the cache entry
299 * NOTE! vhold() is allowed if we hold a
300 * lock on the ncp (which we do).
304 vhold(ncp->nc_vp); /* MPSAFE */
310 if (ncp->nc_locktd == td) {
311 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
318 tsleep_interlock(ncp, 0);
319 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
320 count | NC_EXLOCK_REQ) == 0) {
324 error = tsleep(ncp, PINTERLOCKED, "clock", nclockwarn);
325 if (error == EWOULDBLOCK) {
328 kprintf("[diagnostic] cache_lock: blocked "
331 kprintf(" \"%*.*s\"\n",
332 ncp->nc_nlen, ncp->nc_nlen,
338 kprintf("[diagnostic] cache_lock: unblocked %*.*s after "
340 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
341 (int)(ticks - didwarn) / hz);
346 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance,
347 * such as the case where one of its children is locked.
353 _cache_lock_nonblock(struct namecache *ncp)
361 count = ncp->nc_exlocks;
364 if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) {
366 * The vp associated with a locked ncp must
367 * be held to prevent it from being recycled.
369 * WARNING! If VRECLAIMED is set the vnode
370 * could already be in the middle of a recycle.
371 * Callers must use cache_vref() or
372 * cache_vget() on the locked ncp to
373 * validate the vp or set the cache entry
376 * NOTE! vhold() is allowed if we hold a
377 * lock on the ncp (which we do).
381 vhold(ncp->nc_vp); /* MPSAFE */
387 if (ncp->nc_locktd == td) {
388 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
403 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop).
405 * nc_locktd must be NULLed out prior to nc_exlocks getting cleared.
411 _cache_unlock(struct namecache *ncp)
413 thread_t td __debugvar = curthread;
416 KKASSERT(ncp->nc_refs >= 0);
417 KKASSERT(ncp->nc_exlocks > 0);
418 KKASSERT(ncp->nc_locktd == td);
420 count = ncp->nc_exlocks;
421 if ((count & ~NC_EXLOCK_REQ) == 1) {
422 ncp->nc_locktd = NULL;
427 if ((count & ~NC_EXLOCK_REQ) == 1) {
428 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 0)) {
429 if (count & NC_EXLOCK_REQ)
434 if (atomic_cmpset_int(&ncp->nc_exlocks, count,
439 count = ncp->nc_exlocks;
445 * cache_hold() and cache_drop() prevent the premature deletion of a
446 * namecache entry but do not prevent operations (such as zapping) on
447 * that namecache entry.
449 * This routine may only be called from outside this source module if
450 * nc_refs is already at least 1.
452 * This is a rare case where callers are allowed to hold a spinlock,
453 * so we can't ourselves.
459 _cache_hold(struct namecache *ncp)
461 atomic_add_int(&ncp->nc_refs, 1);
466 * Drop a cache entry, taking care to deal with races.
468 * For potential 1->0 transitions we must hold the ncp lock to safely
469 * test its flags. An unresolved entry with no children must be zapped
472 * The call to cache_zap() itself will handle all remaining races and
473 * will decrement the ncp's refs regardless. If we are resolved or
474 * have children nc_refs can safely be dropped to 0 without having to
477 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion.
479 * NOTE: cache_zap() may return a non-NULL referenced parent which must
480 * be dropped in a loop.
486 _cache_drop(struct namecache *ncp)
491 KKASSERT(ncp->nc_refs > 0);
495 if (_cache_lock_nonblock(ncp) == 0) {
496 ncp->nc_flag &= ~NCF_DEFEREDZAP;
497 if ((ncp->nc_flag & NCF_UNRESOLVED) &&
498 TAILQ_EMPTY(&ncp->nc_list)) {
499 ncp = cache_zap(ncp, 1);
502 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) {
509 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1))
517 * Link a new namecache entry to its parent and to the hash table. Be
518 * careful to avoid races if vhold() blocks in the future.
520 * Both ncp and par must be referenced and locked.
522 * NOTE: The hash table spinlock is likely held during this call, we
523 * can't do anything fancy.
528 _cache_link_parent(struct namecache *ncp, struct namecache *par,
529 struct nchash_head *nchpp)
531 KKASSERT(ncp->nc_parent == NULL);
532 ncp->nc_parent = par;
533 ncp->nc_head = nchpp;
536 * Set inheritance flags. Note that the parent flags may be
537 * stale due to getattr potentially not having been run yet
538 * (it gets run during nlookup()'s).
540 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE);
541 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE))
542 ncp->nc_flag |= NCF_SF_PNOCACHE;
543 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE))
544 ncp->nc_flag |= NCF_UF_PCACHE;
546 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash);
548 if (TAILQ_EMPTY(&par->nc_list)) {
549 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
551 * Any vp associated with an ncp which has children must
552 * be held to prevent it from being recycled.
557 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
562 * Remove the parent and hash associations from a namecache structure.
563 * If this is the last child of the parent the cache_drop(par) will
564 * attempt to recursively zap the parent.
566 * ncp must be locked. This routine will acquire a temporary lock on
567 * the parent as wlel as the appropriate hash chain.
572 _cache_unlink_parent(struct namecache *ncp)
574 struct namecache *par;
575 struct vnode *dropvp;
577 if ((par = ncp->nc_parent) != NULL) {
578 KKASSERT(ncp->nc_parent == par);
581 spin_lock(&ncp->nc_head->spin);
582 LIST_REMOVE(ncp, nc_hash);
583 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
585 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
587 spin_unlock(&ncp->nc_head->spin);
588 ncp->nc_parent = NULL;
594 * We can only safely vdrop with no spinlocks held.
602 * Allocate a new namecache structure. Most of the code does not require
603 * zero-termination of the string but it makes vop_compat_ncreate() easier.
607 static struct namecache *
608 cache_alloc(int nlen)
610 struct namecache *ncp;
612 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
614 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
616 ncp->nc_flag = NCF_UNRESOLVED;
617 ncp->nc_error = ENOTCONN; /* needs to be resolved */
620 TAILQ_INIT(&ncp->nc_list);
626 * Can only be called for the case where the ncp has never been
627 * associated with anything (so no spinlocks are needed).
632 _cache_free(struct namecache *ncp)
634 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
636 kfree(ncp->nc_name, M_VFSCACHE);
637 kfree(ncp, M_VFSCACHE);
644 cache_zero(struct nchandle *nch)
651 * Ref and deref a namecache structure.
653 * The caller must specify a stable ncp pointer, typically meaning the
654 * ncp is already referenced but this can also occur indirectly through
655 * e.g. holding a lock on a direct child.
657 * WARNING: Caller may hold an unrelated read spinlock, which means we can't
658 * use read spinlocks here.
663 cache_hold(struct nchandle *nch)
665 _cache_hold(nch->ncp);
666 atomic_add_int(&nch->mount->mnt_refs, 1);
671 * Create a copy of a namecache handle for an already-referenced
677 cache_copy(struct nchandle *nch, struct nchandle *target)
681 _cache_hold(target->ncp);
682 atomic_add_int(&nch->mount->mnt_refs, 1);
689 cache_changemount(struct nchandle *nch, struct mount *mp)
691 atomic_add_int(&nch->mount->mnt_refs, -1);
693 atomic_add_int(&nch->mount->mnt_refs, 1);
700 cache_drop(struct nchandle *nch)
702 atomic_add_int(&nch->mount->mnt_refs, -1);
703 _cache_drop(nch->ncp);
712 cache_lock(struct nchandle *nch)
714 _cache_lock(nch->ncp);
718 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller
719 * is responsible for checking both for validity on return as they
720 * may have become invalid.
722 * We have to deal with potential deadlocks here, just ping pong
723 * the lock until we get it (we will always block somewhere when
724 * looping so this is not cpu-intensive).
726 * which = 0 nch1 not locked, nch2 is locked
727 * which = 1 nch1 is locked, nch2 is not locked
730 cache_relock(struct nchandle *nch1, struct ucred *cred1,
731 struct nchandle *nch2, struct ucred *cred2)
739 if (cache_lock_nonblock(nch1) == 0) {
740 cache_resolve(nch1, cred1);
745 cache_resolve(nch1, cred1);
748 if (cache_lock_nonblock(nch2) == 0) {
749 cache_resolve(nch2, cred2);
754 cache_resolve(nch2, cred2);
764 cache_lock_nonblock(struct nchandle *nch)
766 return(_cache_lock_nonblock(nch->ncp));
774 cache_unlock(struct nchandle *nch)
776 _cache_unlock(nch->ncp);
780 * ref-and-lock, unlock-and-deref functions.
782 * This function is primarily used by nlookup. Even though cache_lock
783 * holds the vnode, it is possible that the vnode may have already
784 * initiated a recyclement.
786 * We want cache_get() to return a definitively usable vnode or a
787 * definitively unresolved ncp.
793 _cache_get(struct namecache *ncp)
797 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
798 _cache_setunresolved(ncp);
803 * This is a special form of _cache_lock() which only succeeds if
804 * it can get a pristine, non-recursive lock. The caller must have
805 * already ref'd the ncp.
807 * On success the ncp will be locked, on failure it will not. The
808 * ref count does not change either way.
810 * We want _cache_lock_special() (on success) to return a definitively
811 * usable vnode or a definitively unresolved ncp.
816 _cache_lock_special(struct namecache *ncp)
818 if (_cache_lock_nonblock(ncp) == 0) {
819 if ((ncp->nc_exlocks & ~NC_EXLOCK_REQ) == 1) {
820 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
821 _cache_setunresolved(ncp);
831 * NOTE: The same nchandle can be passed for both arguments.
836 cache_get(struct nchandle *nch, struct nchandle *target)
838 KKASSERT(nch->ncp->nc_refs > 0);
839 target->mount = nch->mount;
840 target->ncp = _cache_get(nch->ncp);
841 atomic_add_int(&target->mount->mnt_refs, 1);
849 _cache_put(struct namecache *ncp)
859 cache_put(struct nchandle *nch)
861 atomic_add_int(&nch->mount->mnt_refs, -1);
862 _cache_put(nch->ncp);
868 * Resolve an unresolved ncp by associating a vnode with it. If the
869 * vnode is NULL, a negative cache entry is created.
871 * The ncp should be locked on entry and will remain locked on return.
877 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp)
879 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
883 * Any vp associated with an ncp which has children must
884 * be held. Any vp associated with a locked ncp must be held.
886 if (!TAILQ_EMPTY(&ncp->nc_list))
888 spin_lock(&vp->v_spinlock);
890 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
891 spin_unlock(&vp->v_spinlock);
896 * Set auxiliary flags
900 ncp->nc_flag |= NCF_ISDIR;
903 ncp->nc_flag |= NCF_ISSYMLINK;
904 /* XXX cache the contents of the symlink */
909 atomic_add_int(&numcache, 1);
913 * When creating a negative cache hit we set the
914 * namecache_gen. A later resolve will clean out the
915 * negative cache hit if the mount point's namecache_gen
916 * has changed. Used by devfs, could also be used by
921 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
923 spin_unlock(&ncspin);
924 ncp->nc_error = ENOENT;
926 ncp->nc_namecache_gen = mp->mnt_namecache_gen;
928 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP);
935 cache_setvp(struct nchandle *nch, struct vnode *vp)
937 _cache_setvp(nch->mount, nch->ncp, vp);
944 cache_settimeout(struct nchandle *nch, int nticks)
946 struct namecache *ncp = nch->ncp;
948 if ((ncp->nc_timeout = ticks + nticks) == 0)
953 * Disassociate the vnode or negative-cache association and mark a
954 * namecache entry as unresolved again. Note that the ncp is still
955 * left in the hash table and still linked to its parent.
957 * The ncp should be locked and refd on entry and will remain locked and refd
960 * This routine is normally never called on a directory containing children.
961 * However, NFS often does just that in its rename() code as a cop-out to
962 * avoid complex namespace operations. This disconnects a directory vnode
963 * from its namecache and can cause the OLDAPI and NEWAPI to get out of
970 _cache_setunresolved(struct namecache *ncp)
974 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
975 ncp->nc_flag |= NCF_UNRESOLVED;
977 ncp->nc_error = ENOTCONN;
978 if ((vp = ncp->nc_vp) != NULL) {
979 atomic_add_int(&numcache, -1);
980 spin_lock(&vp->v_spinlock);
982 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
983 spin_unlock(&vp->v_spinlock);
986 * Any vp associated with an ncp with children is
987 * held by that ncp. Any vp associated with a locked
988 * ncp is held by that ncp. These conditions must be
989 * undone when the vp is cleared out from the ncp.
991 if (!TAILQ_EMPTY(&ncp->nc_list))
997 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
999 spin_unlock(&ncspin);
1001 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK);
1006 * The cache_nresolve() code calls this function to automatically
1007 * set a resolved cache element to unresolved if it has timed out
1008 * or if it is a negative cache hit and the mount point namecache_gen
1013 static __inline void
1014 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp)
1017 * Already in an unresolved state, nothing to do.
1019 if (ncp->nc_flag & NCF_UNRESOLVED)
1023 * Try to zap entries that have timed out. We have
1024 * to be careful here because locked leafs may depend
1025 * on the vnode remaining intact in a parent, so only
1026 * do this under very specific conditions.
1028 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 &&
1029 TAILQ_EMPTY(&ncp->nc_list)) {
1030 _cache_setunresolved(ncp);
1035 * If a resolved negative cache hit is invalid due to
1036 * the mount's namecache generation being bumped, zap it.
1038 if (ncp->nc_vp == NULL &&
1039 ncp->nc_namecache_gen != mp->mnt_namecache_gen) {
1040 _cache_setunresolved(ncp);
1049 cache_setunresolved(struct nchandle *nch)
1051 _cache_setunresolved(nch->ncp);
1055 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
1056 * looking for matches. This flag tells the lookup code when it must
1057 * check for a mount linkage and also prevents the directories in question
1058 * from being deleted or renamed.
1064 cache_clrmountpt_callback(struct mount *mp, void *data)
1066 struct nchandle *nch = data;
1068 if (mp->mnt_ncmounton.ncp == nch->ncp)
1070 if (mp->mnt_ncmountpt.ncp == nch->ncp)
1079 cache_clrmountpt(struct nchandle *nch)
1083 count = mountlist_scan(cache_clrmountpt_callback, nch,
1084 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1086 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
1090 * Invalidate portions of the namecache topology given a starting entry.
1091 * The passed ncp is set to an unresolved state and:
1093 * The passed ncp must be referencxed and locked. The routine may unlock
1094 * and relock ncp several times, and will recheck the children and loop
1095 * to catch races. When done the passed ncp will be returned with the
1096 * reference and lock intact.
1098 * CINV_DESTROY - Set a flag in the passed ncp entry indicating
1099 * that the physical underlying nodes have been
1100 * destroyed... as in deleted. For example, when
1101 * a directory is removed. This will cause record
1102 * lookups on the name to no longer be able to find
1103 * the record and tells the resolver to return failure
1104 * rather then trying to resolve through the parent.
1106 * The topology itself, including ncp->nc_name,
1109 * This only applies to the passed ncp, if CINV_CHILDREN
1110 * is specified the children are not flagged.
1112 * CINV_CHILDREN - Set all children (recursively) to an unresolved
1115 * Note that this will also have the side effect of
1116 * cleaning out any unreferenced nodes in the topology
1117 * from the leaves up as the recursion backs out.
1119 * Note that the topology for any referenced nodes remains intact, but
1120 * the nodes will be marked as having been destroyed and will be set
1121 * to an unresolved state.
1123 * It is possible for cache_inval() to race a cache_resolve(), meaning that
1124 * the namecache entry may not actually be invalidated on return if it was
1125 * revalidated while recursing down into its children. This code guarentees
1126 * that the node(s) will go through an invalidation cycle, but does not
1127 * guarentee that they will remain in an invalidated state.
1129 * Returns non-zero if a revalidation was detected during the invalidation
1130 * recursion, zero otherwise. Note that since only the original ncp is
1131 * locked the revalidation ultimately can only indicate that the original ncp
1132 * *MIGHT* no have been reresolved.
1134 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
1135 * have to avoid blowing out the kernel stack. We do this by saving the
1136 * deep namecache node and aborting the recursion, then re-recursing at that
1137 * node using a depth-first algorithm in order to allow multiple deep
1138 * recursions to chain through each other, then we restart the invalidation
1145 struct namecache *resume_ncp;
1149 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
1153 _cache_inval(struct namecache *ncp, int flags)
1155 struct cinvtrack track;
1156 struct namecache *ncp2;
1160 track.resume_ncp = NULL;
1163 r = _cache_inval_internal(ncp, flags, &track);
1164 if (track.resume_ncp == NULL)
1166 kprintf("Warning: deep namecache recursion at %s\n",
1169 while ((ncp2 = track.resume_ncp) != NULL) {
1170 track.resume_ncp = NULL;
1172 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
1182 cache_inval(struct nchandle *nch, int flags)
1184 return(_cache_inval(nch->ncp, flags));
1188 * Helper for _cache_inval(). The passed ncp is refd and locked and
1189 * remains that way on return, but may be unlocked/relocked multiple
1190 * times by the routine.
1193 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
1195 struct namecache *kid;
1196 struct namecache *nextkid;
1199 KKASSERT(ncp->nc_exlocks);
1201 _cache_setunresolved(ncp);
1202 if (flags & CINV_DESTROY)
1203 ncp->nc_flag |= NCF_DESTROYED;
1204 if ((flags & CINV_CHILDREN) &&
1205 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
1208 if (++track->depth > MAX_RECURSION_DEPTH) {
1209 track->resume_ncp = ncp;
1215 if (track->resume_ncp) {
1219 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
1220 _cache_hold(nextkid);
1221 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
1222 TAILQ_FIRST(&kid->nc_list)
1225 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
1236 * Someone could have gotten in there while ncp was unlocked,
1239 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1245 * Invalidate a vnode's namecache associations. To avoid races against
1246 * the resolver we do not invalidate a node which we previously invalidated
1247 * but which was then re-resolved while we were in the invalidation loop.
1249 * Returns non-zero if any namecache entries remain after the invalidation
1252 * NOTE: Unlike the namecache topology which guarentees that ncp's will not
1253 * be ripped out of the topology while held, the vnode's v_namecache
1254 * list has no such restriction. NCP's can be ripped out of the list
1255 * at virtually any time if not locked, even if held.
1257 * In addition, the v_namecache list itself must be locked via
1258 * the vnode's spinlock.
1263 cache_inval_vp(struct vnode *vp, int flags)
1265 struct namecache *ncp;
1266 struct namecache *next;
1269 spin_lock(&vp->v_spinlock);
1270 ncp = TAILQ_FIRST(&vp->v_namecache);
1274 /* loop entered with ncp held and vp spin-locked */
1275 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1277 spin_unlock(&vp->v_spinlock);
1279 if (ncp->nc_vp != vp) {
1280 kprintf("Warning: cache_inval_vp: race-A detected on "
1281 "%s\n", ncp->nc_name);
1287 _cache_inval(ncp, flags);
1288 _cache_put(ncp); /* also releases reference */
1290 spin_lock(&vp->v_spinlock);
1291 if (ncp && ncp->nc_vp != vp) {
1292 spin_unlock(&vp->v_spinlock);
1293 kprintf("Warning: cache_inval_vp: race-B detected on "
1294 "%s\n", ncp->nc_name);
1299 spin_unlock(&vp->v_spinlock);
1300 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1304 * This routine is used instead of the normal cache_inval_vp() when we
1305 * are trying to recycle otherwise good vnodes.
1307 * Return 0 on success, non-zero if not all namecache records could be
1308 * disassociated from the vnode (for various reasons).
1313 cache_inval_vp_nonblock(struct vnode *vp)
1315 struct namecache *ncp;
1316 struct namecache *next;
1318 spin_lock(&vp->v_spinlock);
1319 ncp = TAILQ_FIRST(&vp->v_namecache);
1323 /* loop entered with ncp held */
1324 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1326 spin_unlock(&vp->v_spinlock);
1327 if (_cache_lock_nonblock(ncp)) {
1333 if (ncp->nc_vp != vp) {
1334 kprintf("Warning: cache_inval_vp: race-A detected on "
1335 "%s\n", ncp->nc_name);
1341 _cache_inval(ncp, 0);
1342 _cache_put(ncp); /* also releases reference */
1344 spin_lock(&vp->v_spinlock);
1345 if (ncp && ncp->nc_vp != vp) {
1346 spin_unlock(&vp->v_spinlock);
1347 kprintf("Warning: cache_inval_vp: race-B detected on "
1348 "%s\n", ncp->nc_name);
1353 spin_unlock(&vp->v_spinlock);
1355 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1359 * The source ncp has been renamed to the target ncp. Both fncp and tncp
1360 * must be locked. The target ncp is destroyed (as a normal rename-over
1361 * would destroy the target file or directory).
1363 * Because there may be references to the source ncp we cannot copy its
1364 * contents to the target. Instead the source ncp is relinked as the target
1365 * and the target ncp is removed from the namecache topology.
1370 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
1372 struct namecache *fncp = fnch->ncp;
1373 struct namecache *tncp = tnch->ncp;
1374 struct namecache *tncp_par;
1375 struct nchash_head *nchpp;
1380 * Rename fncp (unlink)
1382 _cache_unlink_parent(fncp);
1383 oname = fncp->nc_name;
1384 fncp->nc_name = tncp->nc_name;
1385 fncp->nc_nlen = tncp->nc_nlen;
1386 tncp_par = tncp->nc_parent;
1387 _cache_hold(tncp_par);
1388 _cache_lock(tncp_par);
1391 * Rename fncp (relink)
1393 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT);
1394 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash);
1395 nchpp = NCHHASH(hash);
1397 spin_lock(&nchpp->spin);
1398 _cache_link_parent(fncp, tncp_par, nchpp);
1399 spin_unlock(&nchpp->spin);
1401 _cache_put(tncp_par);
1404 * Get rid of the overwritten tncp (unlink)
1406 _cache_setunresolved(tncp);
1407 _cache_unlink_parent(tncp);
1408 tncp->nc_name = NULL;
1412 kfree(oname, M_VFSCACHE);
1416 * vget the vnode associated with the namecache entry. Resolve the namecache
1417 * entry if necessary. The passed ncp must be referenced and locked.
1419 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked
1420 * (depending on the passed lk_type) will be returned in *vpp with an error
1421 * of 0, or NULL will be returned in *vpp with a non-0 error code. The
1422 * most typical error is ENOENT, meaning that the ncp represents a negative
1423 * cache hit and there is no vnode to retrieve, but other errors can occur
1426 * The vget() can race a reclaim. If this occurs we re-resolve the
1429 * There are numerous places in the kernel where vget() is called on a
1430 * vnode while one or more of its namecache entries is locked. Releasing
1431 * a vnode never deadlocks against locked namecache entries (the vnode
1432 * will not get recycled while referenced ncp's exist). This means we
1433 * can safely acquire the vnode. In fact, we MUST NOT release the ncp
1434 * lock when acquiring the vp lock or we might cause a deadlock.
1439 cache_vget(struct nchandle *nch, struct ucred *cred,
1440 int lk_type, struct vnode **vpp)
1442 struct namecache *ncp;
1447 KKASSERT(ncp->nc_locktd == curthread);
1450 if (ncp->nc_flag & NCF_UNRESOLVED)
1451 error = cache_resolve(nch, cred);
1455 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1456 error = vget(vp, lk_type);
1461 if (error == ENOENT) {
1462 kprintf("Warning: vnode reclaim race detected "
1463 "in cache_vget on %p (%s)\n",
1465 _cache_setunresolved(ncp);
1470 * Not a reclaim race, some other error.
1472 KKASSERT(ncp->nc_vp == vp);
1475 KKASSERT(ncp->nc_vp == vp);
1476 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1479 if (error == 0 && vp == NULL)
1486 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
1488 struct namecache *ncp;
1493 KKASSERT(ncp->nc_locktd == curthread);
1496 if (ncp->nc_flag & NCF_UNRESOLVED)
1497 error = cache_resolve(nch, cred);
1501 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1502 error = vget(vp, LK_SHARED);
1507 if (error == ENOENT) {
1508 kprintf("Warning: vnode reclaim race detected "
1509 "in cache_vget on %p (%s)\n",
1511 _cache_setunresolved(ncp);
1516 * Not a reclaim race, some other error.
1518 KKASSERT(ncp->nc_vp == vp);
1521 KKASSERT(ncp->nc_vp == vp);
1522 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1523 /* caller does not want a lock */
1527 if (error == 0 && vp == NULL)
1534 * Return a referenced vnode representing the parent directory of
1537 * Because the caller has locked the ncp it should not be possible for
1538 * the parent ncp to go away. However, the parent can unresolve its
1539 * dvp at any time so we must be able to acquire a lock on the parent
1540 * to safely access nc_vp.
1542 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock,
1543 * so use vhold()/vdrop() while holding the lock to prevent dvp from
1544 * getting destroyed.
1546 * MPSAFE - Note vhold() is allowed when dvp has 0 refs if we hold a
1547 * lock on the ncp in question..
1549 static struct vnode *
1550 cache_dvpref(struct namecache *ncp)
1552 struct namecache *par;
1556 if ((par = ncp->nc_parent) != NULL) {
1559 if ((par->nc_flag & NCF_UNRESOLVED) == 0) {
1560 if ((dvp = par->nc_vp) != NULL)
1565 if (vget(dvp, LK_SHARED) == 0) {
1568 /* return refd, unlocked dvp */
1580 * Convert a directory vnode to a namecache record without any other
1581 * knowledge of the topology. This ONLY works with directory vnodes and
1582 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the
1583 * returned ncp (if not NULL) will be held and unlocked.
1585 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1586 * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1587 * for dvp. This will fail only if the directory has been deleted out from
1590 * Callers must always check for a NULL return no matter the value of 'makeit'.
1592 * To avoid underflowing the kernel stack each recursive call increments
1593 * the makeit variable.
1596 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1597 struct vnode *dvp, char *fakename);
1598 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1599 struct vnode **saved_dvp);
1602 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
1603 struct nchandle *nch)
1605 struct vnode *saved_dvp;
1611 nch->mount = dvp->v_mount;
1616 * Handle the makeit == 0 degenerate case
1619 spin_lock(&dvp->v_spinlock);
1620 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1623 spin_unlock(&dvp->v_spinlock);
1627 * Loop until resolution, inside code will break out on error.
1631 * Break out if we successfully acquire a working ncp.
1633 spin_lock(&dvp->v_spinlock);
1634 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1637 spin_unlock(&dvp->v_spinlock);
1640 spin_unlock(&dvp->v_spinlock);
1643 * If dvp is the root of its filesystem it should already
1644 * have a namecache pointer associated with it as a side
1645 * effect of the mount, but it may have been disassociated.
1647 if (dvp->v_flag & VROOT) {
1648 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
1649 error = cache_resolve_mp(nch->mount);
1650 _cache_put(nch->ncp);
1652 kprintf("cache_fromdvp: resolve root of mount %p error %d",
1653 dvp->v_mount, error);
1657 kprintf(" failed\n");
1662 kprintf(" succeeded\n");
1667 * If we are recursed too deeply resort to an O(n^2)
1668 * algorithm to resolve the namecache topology. The
1669 * resolved pvp is left referenced in saved_dvp to
1670 * prevent the tree from being destroyed while we loop.
1673 error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1675 kprintf("lookupdotdot(longpath) failed %d "
1676 "dvp %p\n", error, dvp);
1684 * Get the parent directory and resolve its ncp.
1687 kfree(fakename, M_TEMP);
1690 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
1693 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
1699 * Reuse makeit as a recursion depth counter. On success
1700 * nch will be fully referenced.
1702 cache_fromdvp(pvp, cred, makeit + 1, nch);
1704 if (nch->ncp == NULL)
1708 * Do an inefficient scan of pvp (embodied by ncp) to look
1709 * for dvp. This will create a namecache record for dvp on
1710 * success. We loop up to recheck on success.
1712 * ncp and dvp are both held but not locked.
1714 error = cache_inefficient_scan(nch, cred, dvp, fakename);
1716 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1717 pvp, nch->ncp->nc_name, dvp);
1719 /* nch was NULLed out, reload mount */
1720 nch->mount = dvp->v_mount;
1724 kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
1725 pvp, nch->ncp->nc_name);
1728 /* nch was NULLed out, reload mount */
1729 nch->mount = dvp->v_mount;
1733 * If nch->ncp is non-NULL it will have been held already.
1736 kfree(fakename, M_TEMP);
1745 * Go up the chain of parent directories until we find something
1746 * we can resolve into the namecache. This is very inefficient.
1750 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1751 struct vnode **saved_dvp)
1753 struct nchandle nch;
1756 static time_t last_fromdvp_report;
1760 * Loop getting the parent directory vnode until we get something we
1761 * can resolve in the namecache.
1764 nch.mount = dvp->v_mount;
1770 kfree(fakename, M_TEMP);
1773 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
1780 spin_lock(&pvp->v_spinlock);
1781 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1782 _cache_hold(nch.ncp);
1783 spin_unlock(&pvp->v_spinlock);
1787 spin_unlock(&pvp->v_spinlock);
1788 if (pvp->v_flag & VROOT) {
1789 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
1790 error = cache_resolve_mp(nch.mount);
1791 _cache_unlock(nch.ncp);
1794 _cache_drop(nch.ncp);
1804 if (last_fromdvp_report != time_second) {
1805 last_fromdvp_report = time_second;
1806 kprintf("Warning: extremely inefficient path "
1807 "resolution on %s\n",
1810 error = cache_inefficient_scan(&nch, cred, dvp, fakename);
1813 * Hopefully dvp now has a namecache record associated with
1814 * it. Leave it referenced to prevent the kernel from
1815 * recycling the vnode. Otherwise extremely long directory
1816 * paths could result in endless recycling.
1821 _cache_drop(nch.ncp);
1824 kfree(fakename, M_TEMP);
1829 * Do an inefficient scan of the directory represented by ncp looking for
1830 * the directory vnode dvp. ncp must be held but not locked on entry and
1831 * will be held on return. dvp must be refd but not locked on entry and
1832 * will remain refd on return.
1834 * Why do this at all? Well, due to its stateless nature the NFS server
1835 * converts file handles directly to vnodes without necessarily going through
1836 * the namecache ops that would otherwise create the namecache topology
1837 * leading to the vnode. We could either (1) Change the namecache algorithms
1838 * to allow disconnect namecache records that are re-merged opportunistically,
1839 * or (2) Make the NFS server backtrack and scan to recover a connected
1840 * namecache topology in order to then be able to issue new API lookups.
1842 * It turns out that (1) is a huge mess. It takes a nice clean set of
1843 * namecache algorithms and introduces a lot of complication in every subsystem
1844 * that calls into the namecache to deal with the re-merge case, especially
1845 * since we are using the namecache to placehold negative lookups and the
1846 * vnode might not be immediately assigned. (2) is certainly far less
1847 * efficient then (1), but since we are only talking about directories here
1848 * (which are likely to remain cached), the case does not actually run all
1849 * that often and has the supreme advantage of not polluting the namecache
1852 * If a fakename is supplied just construct a namecache entry using the
1856 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1857 struct vnode *dvp, char *fakename)
1859 struct nlcomponent nlc;
1860 struct nchandle rncp;
1872 vat.va_blocksize = 0;
1873 if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1876 error = cache_vref(nch, cred, &pvp);
1881 kprintf("inefficient_scan: directory iosize %ld "
1882 "vattr fileid = %lld\n",
1884 (long long)vat.va_fileid);
1888 * Use the supplied fakename if not NULL. Fake names are typically
1889 * not in the actual filesystem hierarchy. This is used by HAMMER
1890 * to glue @@timestamp recursions together.
1893 nlc.nlc_nameptr = fakename;
1894 nlc.nlc_namelen = strlen(fakename);
1895 rncp = cache_nlookup(nch, &nlc);
1899 if ((blksize = vat.va_blocksize) == 0)
1900 blksize = DEV_BSIZE;
1901 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
1907 iov.iov_base = rbuf;
1908 iov.iov_len = blksize;
1911 uio.uio_resid = blksize;
1912 uio.uio_segflg = UIO_SYSSPACE;
1913 uio.uio_rw = UIO_READ;
1914 uio.uio_td = curthread;
1916 if (ncvp_debug >= 2)
1917 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1918 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1920 den = (struct dirent *)rbuf;
1921 bytes = blksize - uio.uio_resid;
1924 if (ncvp_debug >= 2) {
1925 kprintf("cache_inefficient_scan: %*.*s\n",
1926 den->d_namlen, den->d_namlen,
1929 if (den->d_type != DT_WHT &&
1930 den->d_ino == vat.va_fileid) {
1932 kprintf("cache_inefficient_scan: "
1933 "MATCHED inode %lld path %s/%*.*s\n",
1934 (long long)vat.va_fileid,
1936 den->d_namlen, den->d_namlen,
1939 nlc.nlc_nameptr = den->d_name;
1940 nlc.nlc_namelen = den->d_namlen;
1941 rncp = cache_nlookup(nch, &nlc);
1942 KKASSERT(rncp.ncp != NULL);
1945 bytes -= _DIRENT_DIRSIZ(den);
1946 den = _DIRENT_NEXT(den);
1948 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
1951 kfree(rbuf, M_TEMP);
1955 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
1956 _cache_setvp(rncp.mount, rncp.ncp, dvp);
1957 if (ncvp_debug >= 2) {
1958 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
1959 nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
1962 if (ncvp_debug >= 2) {
1963 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
1964 nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
1968 if (rncp.ncp->nc_vp == NULL)
1969 error = rncp.ncp->nc_error;
1971 * Release rncp after a successful nlookup. rncp was fully
1976 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
1977 dvp, nch->ncp->nc_name);
1984 * Zap a namecache entry. The ncp is unconditionally set to an unresolved
1985 * state, which disassociates it from its vnode or ncneglist.
1987 * Then, if there are no additional references to the ncp and no children,
1988 * the ncp is removed from the topology and destroyed.
1990 * References and/or children may exist if the ncp is in the middle of the
1991 * topology, preventing the ncp from being destroyed.
1993 * This function must be called with the ncp held and locked and will unlock
1994 * and drop it during zapping.
1996 * If nonblock is non-zero and the parent ncp cannot be locked we give up.
1997 * This case can occur in the cache_drop() path.
1999 * This function may returned a held (but NOT locked) parent node which the
2000 * caller must drop. We do this so _cache_drop() can loop, to avoid
2001 * blowing out the kernel stack.
2003 * WARNING! For MPSAFE operation this routine must acquire up to three
2004 * spin locks to be able to safely test nc_refs. Lock order is
2007 * hash spinlock if on hash list
2008 * parent spinlock if child of parent
2009 * (the ncp is unresolved so there is no vnode association)
2011 static struct namecache *
2012 cache_zap(struct namecache *ncp, int nonblock)
2014 struct namecache *par;
2015 struct vnode *dropvp;
2019 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
2021 _cache_setunresolved(ncp);
2024 * Try to scrap the entry and possibly tail-recurse on its parent.
2025 * We only scrap unref'd (other then our ref) unresolved entries,
2026 * we do not scrap 'live' entries.
2028 * Note that once the spinlocks are acquired if nc_refs == 1 no
2029 * other references are possible. If it isn't, however, we have
2030 * to decrement but also be sure to avoid a 1->0 transition.
2032 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
2033 KKASSERT(ncp->nc_refs > 0);
2036 * Acquire locks. Note that the parent can't go away while we hold
2039 if ((par = ncp->nc_parent) != NULL) {
2042 if (_cache_lock_nonblock(par) == 0)
2044 refs = ncp->nc_refs;
2045 ncp->nc_flag |= NCF_DEFEREDZAP;
2046 ++numdefered; /* MP race ok */
2047 if (atomic_cmpset_int(&ncp->nc_refs,
2059 spin_lock(&ncp->nc_head->spin);
2063 * If someone other then us has a ref or we have children
2064 * we cannot zap the entry. The 1->0 transition and any
2065 * further list operation is protected by the spinlocks
2066 * we have acquired but other transitions are not.
2069 refs = ncp->nc_refs;
2070 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list))
2072 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) {
2074 spin_unlock(&ncp->nc_head->spin);
2084 * We are the only ref and with the spinlocks held no further
2085 * refs can be acquired by others.
2087 * Remove us from the hash list and parent list. We have to
2088 * drop a ref on the parent's vp if the parent's list becomes
2093 struct nchash_head *nchpp = ncp->nc_head;
2095 KKASSERT(nchpp != NULL);
2096 LIST_REMOVE(ncp, nc_hash);
2097 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
2098 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
2099 dropvp = par->nc_vp;
2100 ncp->nc_head = NULL;
2101 ncp->nc_parent = NULL;
2102 spin_unlock(&nchpp->spin);
2105 KKASSERT(ncp->nc_head == NULL);
2109 * ncp should not have picked up any refs. Physically
2112 KKASSERT(ncp->nc_refs == 1);
2113 /* _cache_unlock(ncp) not required */
2114 ncp->nc_refs = -1; /* safety */
2116 kfree(ncp->nc_name, M_VFSCACHE);
2117 kfree(ncp, M_VFSCACHE);
2120 * Delayed drop (we had to release our spinlocks)
2122 * The refed parent (if not NULL) must be dropped. The
2123 * caller is responsible for looping.
2131 * Clean up dangling negative cache and defered-drop entries in the
2134 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t;
2136 static cache_hs_t neg_cache_hysteresis_state = CHI_LOW;
2137 static cache_hs_t pos_cache_hysteresis_state = CHI_LOW;
2140 cache_hysteresis(void)
2145 * Don't cache too many negative hits. We use hysteresis to reduce
2146 * the impact on the critical path.
2148 switch(neg_cache_hysteresis_state) {
2150 if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
2151 _cache_cleanneg(10);
2152 neg_cache_hysteresis_state = CHI_HIGH;
2156 if (numneg > MINNEG * 9 / 10 &&
2157 numneg * ncnegfactor * 9 / 10 > numcache
2159 _cache_cleanneg(10);
2161 neg_cache_hysteresis_state = CHI_LOW;
2167 * Don't cache too many positive hits. We use hysteresis to reduce
2168 * the impact on the critical path.
2170 * Excessive positive hits can accumulate due to large numbers of
2171 * hardlinks (the vnode cache will not prevent hl ncps from growing
2174 if ((poslimit = ncposlimit) == 0)
2175 poslimit = desiredvnodes * 2;
2177 switch(pos_cache_hysteresis_state) {
2179 if (numcache > poslimit && numcache > MINPOS) {
2180 _cache_cleanpos(10);
2181 pos_cache_hysteresis_state = CHI_HIGH;
2185 if (numcache > poslimit * 5 / 6 && numcache > MINPOS) {
2186 _cache_cleanpos(10);
2188 pos_cache_hysteresis_state = CHI_LOW;
2194 * Clean out dangling defered-zap ncps which could not
2195 * be cleanly dropped if too many build up. Note
2196 * that numdefered is not an exact number as such ncps
2197 * can be reused and the counter is not handled in a MP
2198 * safe manner by design.
2200 if (numdefered * ncnegfactor > numcache) {
2201 _cache_cleandefered();
2206 * NEW NAMECACHE LOOKUP API
2208 * Lookup an entry in the namecache. The passed par_nch must be referenced
2209 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp
2210 * is ALWAYS returned, eve if the supplied component is illegal.
2212 * The resulting namecache entry should be returned to the system with
2213 * cache_put() or cache_unlock() + cache_drop().
2215 * namecache locks are recursive but care must be taken to avoid lock order
2216 * reversals (hence why the passed par_nch must be unlocked). Locking
2217 * rules are to order for parent traversals, not for child traversals.
2219 * Nobody else will be able to manipulate the associated namespace (e.g.
2220 * create, delete, rename, rename-target) until the caller unlocks the
2223 * The returned entry will be in one of three states: positive hit (non-null
2224 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
2225 * Unresolved entries must be resolved through the filesystem to associate the
2226 * vnode and/or determine whether a positive or negative hit has occured.
2228 * It is not necessary to lock a directory in order to lock namespace under
2229 * that directory. In fact, it is explicitly not allowed to do that. A
2230 * directory is typically only locked when being created, renamed, or
2233 * The directory (par) may be unresolved, in which case any returned child
2234 * will likely also be marked unresolved. Likely but not guarenteed. Since
2235 * the filesystem lookup requires a resolved directory vnode the caller is
2236 * responsible for resolving the namecache chain top-down. This API
2237 * specifically allows whole chains to be created in an unresolved state.
2240 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
2242 struct nchandle nch;
2243 struct namecache *ncp;
2244 struct namecache *new_ncp;
2245 struct nchash_head *nchpp;
2253 mp = par_nch->mount;
2257 * This is a good time to call it, no ncp's are locked by
2263 * Try to locate an existing entry
2265 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2266 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2268 nchpp = NCHHASH(hash);
2270 spin_lock(&nchpp->spin);
2271 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2275 * Break out if we find a matching entry. Note that
2276 * UNRESOLVED entries may match, but DESTROYED entries
2279 if (ncp->nc_parent == par_nch->ncp &&
2280 ncp->nc_nlen == nlc->nlc_namelen &&
2281 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2282 (ncp->nc_flag & NCF_DESTROYED) == 0
2285 spin_unlock(&nchpp->spin);
2287 _cache_unlock(par_nch->ncp);
2290 if (_cache_lock_special(ncp) == 0) {
2291 _cache_auto_unresolve(mp, ncp);
2293 _cache_free(new_ncp);
2304 * We failed to locate an entry, create a new entry and add it to
2305 * the cache. The parent ncp must also be locked so we
2308 * We have to relookup after possibly blocking in kmalloc or
2309 * when locking par_nch.
2311 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2312 * mount case, in which case nc_name will be NULL.
2314 if (new_ncp == NULL) {
2315 spin_unlock(&nchpp->spin);
2316 new_ncp = cache_alloc(nlc->nlc_namelen);
2317 if (nlc->nlc_namelen) {
2318 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2320 new_ncp->nc_name[nlc->nlc_namelen] = 0;
2324 if (par_locked == 0) {
2325 spin_unlock(&nchpp->spin);
2326 _cache_lock(par_nch->ncp);
2332 * WARNING! We still hold the spinlock. We have to set the hash
2333 * table entry atomically.
2336 _cache_link_parent(ncp, par_nch->ncp, nchpp);
2337 spin_unlock(&nchpp->spin);
2338 _cache_unlock(par_nch->ncp);
2339 /* par_locked = 0 - not used */
2342 * stats and namecache size management
2344 if (ncp->nc_flag & NCF_UNRESOLVED)
2345 ++gd->gd_nchstats->ncs_miss;
2346 else if (ncp->nc_vp)
2347 ++gd->gd_nchstats->ncs_goodhits;
2349 ++gd->gd_nchstats->ncs_neghits;
2352 atomic_add_int(&nch.mount->mnt_refs, 1);
2357 * This is a non-blocking verison of cache_nlookup() used by
2358 * nfs_readdirplusrpc_uio(). It can fail for any reason and
2359 * will return nch.ncp == NULL in that case.
2362 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc)
2364 struct nchandle nch;
2365 struct namecache *ncp;
2366 struct namecache *new_ncp;
2367 struct nchash_head *nchpp;
2375 mp = par_nch->mount;
2379 * Try to locate an existing entry
2381 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2382 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2384 nchpp = NCHHASH(hash);
2386 spin_lock(&nchpp->spin);
2387 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2391 * Break out if we find a matching entry. Note that
2392 * UNRESOLVED entries may match, but DESTROYED entries
2395 if (ncp->nc_parent == par_nch->ncp &&
2396 ncp->nc_nlen == nlc->nlc_namelen &&
2397 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2398 (ncp->nc_flag & NCF_DESTROYED) == 0
2401 spin_unlock(&nchpp->spin);
2403 _cache_unlock(par_nch->ncp);
2406 if (_cache_lock_special(ncp) == 0) {
2407 _cache_auto_unresolve(mp, ncp);
2409 _cache_free(new_ncp);
2420 * We failed to locate an entry, create a new entry and add it to
2421 * the cache. The parent ncp must also be locked so we
2424 * We have to relookup after possibly blocking in kmalloc or
2425 * when locking par_nch.
2427 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2428 * mount case, in which case nc_name will be NULL.
2430 if (new_ncp == NULL) {
2431 spin_unlock(&nchpp->spin);
2432 new_ncp = cache_alloc(nlc->nlc_namelen);
2433 if (nlc->nlc_namelen) {
2434 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2436 new_ncp->nc_name[nlc->nlc_namelen] = 0;
2440 if (par_locked == 0) {
2441 spin_unlock(&nchpp->spin);
2442 if (_cache_lock_nonblock(par_nch->ncp) == 0) {
2450 * WARNING! We still hold the spinlock. We have to set the hash
2451 * table entry atomically.
2454 _cache_link_parent(ncp, par_nch->ncp, nchpp);
2455 spin_unlock(&nchpp->spin);
2456 _cache_unlock(par_nch->ncp);
2457 /* par_locked = 0 - not used */
2460 * stats and namecache size management
2462 if (ncp->nc_flag & NCF_UNRESOLVED)
2463 ++gd->gd_nchstats->ncs_miss;
2464 else if (ncp->nc_vp)
2465 ++gd->gd_nchstats->ncs_goodhits;
2467 ++gd->gd_nchstats->ncs_neghits;
2470 atomic_add_int(&nch.mount->mnt_refs, 1);
2474 _cache_free(new_ncp);
2483 * The namecache entry is marked as being used as a mount point.
2484 * Locate the mount if it is visible to the caller.
2486 struct findmount_info {
2487 struct mount *result;
2488 struct mount *nch_mount;
2489 struct namecache *nch_ncp;
2494 cache_findmount_callback(struct mount *mp, void *data)
2496 struct findmount_info *info = data;
2499 * Check the mount's mounted-on point against the passed nch.
2501 if (mp->mnt_ncmounton.mount == info->nch_mount &&
2502 mp->mnt_ncmounton.ncp == info->nch_ncp
2511 cache_findmount(struct nchandle *nch)
2513 struct findmount_info info;
2516 info.nch_mount = nch->mount;
2517 info.nch_ncp = nch->ncp;
2518 mountlist_scan(cache_findmount_callback, &info,
2519 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
2520 return(info.result);
2524 * Resolve an unresolved namecache entry, generally by looking it up.
2525 * The passed ncp must be locked and refd.
2527 * Theoretically since a vnode cannot be recycled while held, and since
2528 * the nc_parent chain holds its vnode as long as children exist, the
2529 * direct parent of the cache entry we are trying to resolve should
2530 * have a valid vnode. If not then generate an error that we can
2531 * determine is related to a resolver bug.
2533 * However, if a vnode was in the middle of a recyclement when the NCP
2534 * got locked, ncp->nc_vp might point to a vnode that is about to become
2535 * invalid. cache_resolve() handles this case by unresolving the entry
2536 * and then re-resolving it.
2538 * Note that successful resolution does not necessarily return an error
2539 * code of 0. If the ncp resolves to a negative cache hit then ENOENT
2545 cache_resolve(struct nchandle *nch, struct ucred *cred)
2547 struct namecache *par_tmp;
2548 struct namecache *par;
2549 struct namecache *ncp;
2550 struct nchandle nctmp;
2559 * If the ncp is already resolved we have nothing to do. However,
2560 * we do want to guarentee that a usable vnode is returned when
2561 * a vnode is present, so make sure it hasn't been reclaimed.
2563 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2564 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
2565 _cache_setunresolved(ncp);
2566 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
2567 return (ncp->nc_error);
2571 * Mount points need special handling because the parent does not
2572 * belong to the same filesystem as the ncp.
2574 if (ncp == mp->mnt_ncmountpt.ncp)
2575 return (cache_resolve_mp(mp));
2578 * We expect an unbroken chain of ncps to at least the mount point,
2579 * and even all the way to root (but this code doesn't have to go
2580 * past the mount point).
2582 if (ncp->nc_parent == NULL) {
2583 kprintf("EXDEV case 1 %p %*.*s\n", ncp,
2584 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2585 ncp->nc_error = EXDEV;
2586 return(ncp->nc_error);
2590 * The vp's of the parent directories in the chain are held via vhold()
2591 * due to the existance of the child, and should not disappear.
2592 * However, there are cases where they can disappear:
2594 * - due to filesystem I/O errors.
2595 * - due to NFS being stupid about tracking the namespace and
2596 * destroys the namespace for entire directories quite often.
2597 * - due to forced unmounts.
2598 * - due to an rmdir (parent will be marked DESTROYED)
2600 * When this occurs we have to track the chain backwards and resolve
2601 * it, looping until the resolver catches up to the current node. We
2602 * could recurse here but we might run ourselves out of kernel stack
2603 * so we do it in a more painful manner. This situation really should
2604 * not occur all that often, or if it does not have to go back too
2605 * many nodes to resolve the ncp.
2607 while ((dvp = cache_dvpref(ncp)) == NULL) {
2609 * This case can occur if a process is CD'd into a
2610 * directory which is then rmdir'd. If the parent is marked
2611 * destroyed there is no point trying to resolve it.
2613 if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
2615 par = ncp->nc_parent;
2618 while ((par_tmp = par->nc_parent) != NULL &&
2619 par_tmp->nc_vp == NULL) {
2620 _cache_hold(par_tmp);
2621 _cache_lock(par_tmp);
2625 if (par->nc_parent == NULL) {
2626 kprintf("EXDEV case 2 %*.*s\n",
2627 par->nc_nlen, par->nc_nlen, par->nc_name);
2631 kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
2632 par->nc_nlen, par->nc_nlen, par->nc_name);
2634 * The parent is not set in stone, ref and lock it to prevent
2635 * it from disappearing. Also note that due to renames it
2636 * is possible for our ncp to move and for par to no longer
2637 * be one of its parents. We resolve it anyway, the loop
2638 * will handle any moves.
2640 _cache_get(par); /* additional hold/lock */
2641 _cache_put(par); /* from earlier hold/lock */
2642 if (par == nch->mount->mnt_ncmountpt.ncp) {
2643 cache_resolve_mp(nch->mount);
2644 } else if ((dvp = cache_dvpref(par)) == NULL) {
2645 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
2649 if (par->nc_flag & NCF_UNRESOLVED) {
2652 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2656 if ((error = par->nc_error) != 0) {
2657 if (par->nc_error != EAGAIN) {
2658 kprintf("EXDEV case 3 %*.*s error %d\n",
2659 par->nc_nlen, par->nc_nlen, par->nc_name,
2664 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
2665 par, par->nc_nlen, par->nc_nlen, par->nc_name);
2672 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
2673 * ncp's and reattach them. If this occurs the original ncp is marked
2674 * EAGAIN to force a relookup.
2676 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
2677 * ncp must already be resolved.
2682 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2685 ncp->nc_error = EPERM;
2687 if (ncp->nc_error == EAGAIN) {
2688 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
2689 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2692 return(ncp->nc_error);
2696 * Resolve the ncp associated with a mount point. Such ncp's almost always
2697 * remain resolved and this routine is rarely called. NFS MPs tends to force
2698 * re-resolution more often due to its mac-truck-smash-the-namecache
2699 * method of tracking namespace changes.
2701 * The semantics for this call is that the passed ncp must be locked on
2702 * entry and will be locked on return. However, if we actually have to
2703 * resolve the mount point we temporarily unlock the entry in order to
2704 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of
2705 * the unlock we have to recheck the flags after we relock.
2708 cache_resolve_mp(struct mount *mp)
2710 struct namecache *ncp = mp->mnt_ncmountpt.ncp;
2714 KKASSERT(mp != NULL);
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);
2726 if (ncp->nc_flag & NCF_UNRESOLVED) {
2728 while (vfs_busy(mp, 0))
2730 error = VFS_ROOT(mp, &vp);
2734 * recheck the ncp state after relocking.
2736 if (ncp->nc_flag & NCF_UNRESOLVED) {
2737 ncp->nc_error = error;
2739 _cache_setvp(mp, ncp, vp);
2742 kprintf("[diagnostic] cache_resolve_mp: failed"
2743 " to resolve mount %p err=%d ncp=%p\n",
2745 _cache_setvp(mp, ncp, NULL);
2747 } else if (error == 0) {
2752 return(ncp->nc_error);
2756 * Clean out negative cache entries when too many have accumulated.
2761 _cache_cleanneg(int count)
2763 struct namecache *ncp;
2766 * Attempt to clean out the specified number of negative cache
2771 ncp = TAILQ_FIRST(&ncneglist);
2773 spin_unlock(&ncspin);
2776 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
2777 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
2779 spin_unlock(&ncspin);
2780 if (_cache_lock_special(ncp) == 0) {
2781 ncp = cache_zap(ncp, 1);
2792 * Clean out positive cache entries when too many have accumulated.
2797 _cache_cleanpos(int count)
2799 static volatile int rover;
2800 struct nchash_head *nchpp;
2801 struct namecache *ncp;
2805 * Attempt to clean out the specified number of negative cache
2809 rover_copy = ++rover; /* MPSAFEENOUGH */
2810 nchpp = NCHHASH(rover_copy);
2812 spin_lock(&nchpp->spin);
2813 ncp = LIST_FIRST(&nchpp->list);
2816 spin_unlock(&nchpp->spin);
2819 if (_cache_lock_special(ncp) == 0) {
2820 ncp = cache_zap(ncp, 1);
2832 * This is a kitchen sink function to clean out ncps which we
2833 * tried to zap from cache_drop() but failed because we were
2834 * unable to acquire the parent lock.
2836 * Such entries can also be removed via cache_inval_vp(), such
2837 * as when unmounting.
2842 _cache_cleandefered(void)
2844 struct nchash_head *nchpp;
2845 struct namecache *ncp;
2846 struct namecache dummy;
2850 bzero(&dummy, sizeof(dummy));
2851 dummy.nc_flag = NCF_DESTROYED;
2853 for (i = 0; i <= nchash; ++i) {
2854 nchpp = &nchashtbl[i];
2856 spin_lock(&nchpp->spin);
2857 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash);
2859 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) {
2860 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0)
2862 LIST_REMOVE(&dummy, nc_hash);
2863 LIST_INSERT_AFTER(ncp, &dummy, nc_hash);
2865 spin_unlock(&nchpp->spin);
2866 if (_cache_lock_nonblock(ncp) == 0) {
2867 ncp->nc_flag &= ~NCF_DEFEREDZAP;
2871 spin_lock(&nchpp->spin);
2874 LIST_REMOVE(&dummy, nc_hash);
2875 spin_unlock(&nchpp->spin);
2880 * Name cache initialization, from vfsinit() when we are booting
2888 /* initialise per-cpu namecache effectiveness statistics. */
2889 for (i = 0; i < ncpus; ++i) {
2890 gd = globaldata_find(i);
2891 gd->gd_nchstats = &nchstats[i];
2893 TAILQ_INIT(&ncneglist);
2895 nchashtbl = hashinit_ext(desiredvnodes / 2,
2896 sizeof(struct nchash_head),
2897 M_VFSCACHE, &nchash);
2898 for (i = 0; i <= (int)nchash; ++i) {
2899 LIST_INIT(&nchashtbl[i].list);
2900 spin_init(&nchashtbl[i].spin);
2902 nclockwarn = 5 * hz;
2906 * Called from start_init() to bootstrap the root filesystem. Returns
2907 * a referenced, unlocked namecache record.
2910 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
2912 nch->ncp = cache_alloc(0);
2914 atomic_add_int(&mp->mnt_refs, 1);
2916 _cache_setvp(nch->mount, nch->ncp, vp);
2920 * vfs_cache_setroot()
2922 * Create an association between the root of our namecache and
2923 * the root vnode. This routine may be called several times during
2926 * If the caller intends to save the returned namecache pointer somewhere
2927 * it must cache_hold() it.
2930 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
2933 struct nchandle onch;
2941 cache_zero(&rootnch);
2949 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache
2950 * topology and is being removed as quickly as possible. The new VOP_N*()
2951 * API calls are required to make specific adjustments using the supplied
2952 * ncp pointers rather then just bogusly purging random vnodes.
2954 * Invalidate all namecache entries to a particular vnode as well as
2955 * any direct children of that vnode in the namecache. This is a
2956 * 'catch all' purge used by filesystems that do not know any better.
2958 * Note that the linkage between the vnode and its namecache entries will
2959 * be removed, but the namecache entries themselves might stay put due to
2960 * active references from elsewhere in the system or due to the existance of
2961 * the children. The namecache topology is left intact even if we do not
2962 * know what the vnode association is. Such entries will be marked
2966 cache_purge(struct vnode *vp)
2968 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
2972 * Flush all entries referencing a particular filesystem.
2974 * Since we need to check it anyway, we will flush all the invalid
2975 * entries at the same time.
2980 cache_purgevfs(struct mount *mp)
2982 struct nchash_head *nchpp;
2983 struct namecache *ncp, *nnp;
2986 * Scan hash tables for applicable entries.
2988 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
2989 spin_lock_wr(&nchpp->spin); XXX
2990 ncp = LIST_FIRST(&nchpp->list);
2994 nnp = LIST_NEXT(ncp, nc_hash);
2997 if (ncp->nc_mount == mp) {
2999 ncp = cache_zap(ncp, 0);
3007 spin_unlock_wr(&nchpp->spin); XXX
3013 static int disablecwd;
3014 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
3017 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls);
3018 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1);
3019 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2);
3020 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3);
3021 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4);
3022 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound);
3028 sys___getcwd(struct __getcwd_args *uap)
3038 buflen = uap->buflen;
3041 if (buflen > MAXPATHLEN)
3042 buflen = MAXPATHLEN;
3044 buf = kmalloc(buflen, M_TEMP, M_WAITOK);
3046 bp = kern_getcwd(buf, buflen, &error);
3049 error = copyout(bp, uap->buf, strlen(bp) + 1);
3055 kern_getcwd(char *buf, size_t buflen, int *error)
3057 struct proc *p = curproc;
3059 int i, slash_prefixed;
3060 struct filedesc *fdp;
3061 struct nchandle nch;
3062 struct namecache *ncp;
3071 nch = fdp->fd_ncdir;
3076 while (ncp && (ncp != fdp->fd_nrdir.ncp ||
3077 nch.mount != fdp->fd_nrdir.mount)
3080 * While traversing upwards if we encounter the root
3081 * of the current mount we have to skip to the mount point
3082 * in the underlying filesystem.
3084 if (ncp == nch.mount->mnt_ncmountpt.ncp) {
3085 nch = nch.mount->mnt_ncmounton;
3094 * Prepend the path segment
3096 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3103 *--bp = ncp->nc_name[i];
3115 * Go up a directory. This isn't a mount point so we don't
3116 * have to check again.
3118 while ((nch.ncp = ncp->nc_parent) != NULL) {
3120 if (nch.ncp != ncp->nc_parent) {
3124 _cache_hold(nch.ncp);
3137 if (!slash_prefixed) {
3155 * Thus begins the fullpath magic.
3157 * The passed nchp is referenced but not locked.
3160 #define STATNODE(name) \
3161 static u_int name; \
3162 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "")
3164 static int disablefullpath;
3165 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
3166 &disablefullpath, 0,
3167 "Disable fullpath lookups");
3169 STATNODE(numfullpathcalls);
3170 STATNODE(numfullpathfail1);
3171 STATNODE(numfullpathfail2);
3172 STATNODE(numfullpathfail3);
3173 STATNODE(numfullpathfail4);
3174 STATNODE(numfullpathfound);
3177 cache_fullpath(struct proc *p, struct nchandle *nchp,
3178 char **retbuf, char **freebuf, int guess)
3180 struct nchandle fd_nrdir;
3181 struct nchandle nch;
3182 struct namecache *ncp;
3183 struct mount *mp, *new_mp;
3189 atomic_add_int(&numfullpathcalls, -1);
3194 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
3195 bp = buf + MAXPATHLEN - 1;
3198 fd_nrdir = p->p_fd->fd_nrdir;
3208 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) {
3212 * If we are asked to guess the upwards path, we do so whenever
3213 * we encounter an ncp marked as a mountpoint. We try to find
3214 * the actual mountpoint by finding the mountpoint with this ncp.
3216 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) {
3217 new_mp = mount_get_by_nc(ncp);
3220 * While traversing upwards if we encounter the root
3221 * of the current mount we have to skip to the mount point.
3223 if (ncp == mp->mnt_ncmountpt.ncp) {
3227 nch = new_mp->mnt_ncmounton;
3237 * Prepend the path segment
3239 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3246 *--bp = ncp->nc_name[i];
3258 * Go up a directory. This isn't a mount point so we don't
3259 * have to check again.
3261 * We can only safely access nc_parent with ncp held locked.
3263 while ((nch.ncp = ncp->nc_parent) != NULL) {
3265 if (nch.ncp != ncp->nc_parent) {
3269 _cache_hold(nch.ncp);
3283 if (!slash_prefixed) {
3303 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf, int guess)
3305 struct namecache *ncp;
3306 struct nchandle nch;
3309 atomic_add_int(&numfullpathcalls, 1);
3310 if (disablefullpath)
3316 /* vn is NULL, client wants us to use p->p_textvp */
3318 if ((vn = p->p_textvp) == NULL)
3321 spin_lock(&vn->v_spinlock);
3322 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
3327 spin_unlock(&vn->v_spinlock);
3331 spin_unlock(&vn->v_spinlock);
3333 atomic_add_int(&numfullpathcalls, -1);
3335 nch.mount = vn->v_mount;
3336 error = cache_fullpath(p, &nch, retbuf, freebuf, guess);