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. Neither the name of the University nor the names of its contributors
49 * may be used to endorse or promote products derived from this software
50 * without specific prior written permission.
52 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
53 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
54 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
55 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
56 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
57 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
58 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
59 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
60 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
61 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/kernel.h>
68 #include <sys/sysctl.h>
69 #include <sys/mount.h>
70 #include <sys/vnode.h>
71 #include <sys/malloc.h>
72 #include <sys/sysproto.h>
73 #include <sys/spinlock.h>
75 #include <sys/namei.h>
76 #include <sys/nlookup.h>
77 #include <sys/filedesc.h>
78 #include <sys/fnv_hash.h>
79 #include <sys/globaldata.h>
80 #include <sys/kern_syscall.h>
81 #include <sys/dirent.h>
84 #include <sys/sysref2.h>
85 #include <sys/spinlock2.h>
86 #include <sys/mplock2.h>
88 #define MAX_RECURSION_DEPTH 64
91 * Random lookups in the cache are accomplished with a hash table using
92 * a hash key of (nc_src_vp, name). Each hash chain has its own spin lock.
94 * Negative entries may exist and correspond to resolved namecache
95 * structures where nc_vp is NULL. In a negative entry, NCF_WHITEOUT
96 * will be set if the entry corresponds to a whited-out directory entry
97 * (verses simply not finding the entry at all). ncneglist is locked
98 * with a global spinlock (ncspin).
102 * (1) A ncp must be referenced before it can be locked.
104 * (2) A ncp must be locked in order to modify it.
106 * (3) ncp locks are always ordered child -> parent. That may seem
107 * backwards but forward scans use the hash table and thus can hold
108 * the parent unlocked when traversing downward.
110 * This allows insert/rename/delete/dot-dot and other operations
111 * to use ncp->nc_parent links.
113 * This also prevents a locked up e.g. NFS node from creating a
114 * chain reaction all the way back to the root vnode / namecache.
116 * (4) parent linkages require both the parent and child to be locked.
120 * Structures associated with name cacheing.
122 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash])
125 #define NCMOUNT_NUMCACHE 1009 /* prime number */
127 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
129 LIST_HEAD(nchash_list, namecache);
132 struct nchash_list list;
133 struct spinlock spin;
136 struct ncmount_cache {
137 struct spinlock spin;
138 struct namecache *ncp;
140 int isneg; /* if != 0 mp is originator and not target */
143 static struct nchash_head *nchashtbl;
144 static struct namecache_list ncneglist;
145 static struct spinlock ncspin;
146 static struct ncmount_cache ncmount_cache[NCMOUNT_NUMCACHE];
149 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server
150 * to create the namecache infrastructure leading to a dangling vnode.
152 * 0 Only errors are reported
153 * 1 Successes are reported
154 * 2 Successes + the whole directory scan is reported
155 * 3 Force the directory scan code run as if the parent vnode did not
156 * have a namecache record, even if it does have one.
158 static int ncvp_debug;
159 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0,
160 "Namecache debug level (0-3)");
162 static u_long nchash; /* size of hash table */
163 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
164 "Size of namecache hash table");
166 static int ncnegflush = 10; /* burst for negative flush */
167 SYSCTL_INT(_debug, OID_AUTO, ncnegflush, CTLFLAG_RW, &ncnegflush, 0,
168 "Batch flush negative entries");
170 static int ncposflush = 10; /* burst for positive flush */
171 SYSCTL_INT(_debug, OID_AUTO, ncposflush, CTLFLAG_RW, &ncposflush, 0,
172 "Batch flush positive entries");
174 static int ncnegfactor = 16; /* ratio of negative entries */
175 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
176 "Ratio of namecache negative entries");
178 static int nclockwarn; /* warn on locked entries in ticks */
179 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0,
180 "Warn on locked namecache entries in ticks");
182 static int numdefered; /* number of cache entries allocated */
183 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0,
184 "Number of cache entries allocated");
186 static int ncposlimit; /* number of cache entries allocated */
187 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0,
188 "Number of cache entries allocated");
190 static int ncp_shared_lock_disable = 0;
191 SYSCTL_INT(_debug, OID_AUTO, ncp_shared_lock_disable, CTLFLAG_RW,
192 &ncp_shared_lock_disable, 0, "Disable shared namecache locks");
194 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode),
195 "sizeof(struct vnode)");
196 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache),
197 "sizeof(struct namecache)");
199 static int ncmount_cache_enable = 1;
200 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW,
201 &ncmount_cache_enable, 0, "mount point cache");
202 static long ncmount_cache_hit;
203 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_hit, CTLFLAG_RW,
204 &ncmount_cache_hit, 0, "mpcache hits");
205 static long ncmount_cache_miss;
206 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_miss, CTLFLAG_RW,
207 &ncmount_cache_miss, 0, "mpcache misses");
208 static long ncmount_cache_overwrite;
209 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_overwrite, CTLFLAG_RW,
210 &ncmount_cache_overwrite, 0, "mpcache entry overwrites");
212 static int cache_resolve_mp(struct mount *mp);
213 static struct vnode *cache_dvpref(struct namecache *ncp);
214 static void _cache_lock(struct namecache *ncp);
215 static void _cache_setunresolved(struct namecache *ncp);
216 static void _cache_cleanneg(int count);
217 static void _cache_cleanpos(int count);
218 static void _cache_cleandefered(void);
219 static void _cache_unlink(struct namecache *ncp);
222 * The new name cache statistics
224 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
226 SYSCTL_INT(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
227 "Number of negative namecache entries");
229 SYSCTL_INT(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
230 "Number of namecaches entries");
231 static u_long numcalls;
232 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcalls, CTLFLAG_RD, &numcalls, 0,
233 "Number of namecache lookups");
234 static u_long numchecks;
235 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numchecks, CTLFLAG_RD, &numchecks, 0,
236 "Number of checked entries in namecache lookups");
238 struct nchstats nchstats[SMP_MAXCPU];
240 * Export VFS cache effectiveness statistics to user-land.
242 * The statistics are left for aggregation to user-land so
243 * neat things can be achieved, like observing per-CPU cache
247 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
249 struct globaldata *gd;
253 for (i = 0; i < ncpus; ++i) {
254 gd = globaldata_find(i);
255 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
256 sizeof(struct nchstats))))
262 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
263 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
265 static struct namecache *cache_zap(struct namecache *ncp, int nonblock);
268 * Namespace locking. The caller must already hold a reference to the
269 * namecache structure in order to lock/unlock it. This function prevents
270 * the namespace from being created or destroyed by accessors other then
273 * Note that holding a locked namecache structure prevents other threads
274 * from making namespace changes (e.g. deleting or creating), prevents
275 * vnode association state changes by other threads, and prevents the
276 * namecache entry from being resolved or unresolved by other threads.
278 * An exclusive lock owner has full authority to associate/disassociate
279 * vnodes and resolve/unresolve the locked ncp.
281 * A shared lock owner only has authority to acquire the underlying vnode,
284 * The primary lock field is nc_lockstatus. nc_locktd is set after the
285 * fact (when locking) or cleared prior to unlocking.
287 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed
288 * or recycled, but it does NOT help you if the vnode had already
289 * initiated a recyclement. If this is important, use cache_get()
290 * rather then cache_lock() (and deal with the differences in the
291 * way the refs counter is handled). Or, alternatively, make an
292 * unconditional call to cache_validate() or cache_resolve()
293 * after cache_lock() returns.
297 _cache_lock(struct namecache *ncp)
305 KKASSERT(ncp->nc_refs != 0);
311 count = ncp->nc_lockstatus;
314 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) {
315 if (atomic_cmpset_int(&ncp->nc_lockstatus,
318 * The vp associated with a locked ncp must
319 * be held to prevent it from being recycled.
321 * WARNING! If VRECLAIMED is set the vnode
322 * could already be in the middle of a recycle.
323 * Callers must use cache_vref() or
324 * cache_vget() on the locked ncp to
325 * validate the vp or set the cache entry
328 * NOTE! vhold() is allowed if we hold a
329 * lock on the ncp (which we do).
339 if (ncp->nc_locktd == td) {
340 KKASSERT((count & NC_SHLOCK_FLAG) == 0);
341 if (atomic_cmpset_int(&ncp->nc_lockstatus,
348 tsleep_interlock(&ncp->nc_locktd, 0);
349 if (atomic_cmpset_int(&ncp->nc_lockstatus, count,
350 count | NC_EXLOCK_REQ) == 0) {
356 error = tsleep(&ncp->nc_locktd, PINTERLOCKED,
357 "clock", nclockwarn);
358 if (error == EWOULDBLOCK) {
361 kprintf("[diagnostic] cache_lock: "
362 "blocked on %p %08x",
364 kprintf(" \"%*.*s\"\n",
365 ncp->nc_nlen, ncp->nc_nlen,
372 kprintf("[diagnostic] cache_lock: unblocked %*.*s after "
374 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
375 (int)(ticks + (hz / 2) - begticks) / hz);
380 * The shared lock works similarly to the exclusive lock except
381 * nc_locktd is left NULL and we need an interlock (VHOLD) to
382 * prevent vhold() races, since the moment our cmpset_int succeeds
383 * another cpu can come in and get its own shared lock.
385 * A critical section is needed to prevent interruption during the
390 _cache_lock_shared(struct namecache *ncp)
395 u_int optreq = NC_EXLOCK_REQ;
397 KKASSERT(ncp->nc_refs != 0);
401 count = ncp->nc_lockstatus;
404 if ((count & ~NC_SHLOCK_REQ) == 0) {
406 if (atomic_cmpset_int(&ncp->nc_lockstatus,
408 (count + 1) | NC_SHLOCK_FLAG |
411 * The vp associated with a locked ncp must
412 * be held to prevent it from being recycled.
414 * WARNING! If VRECLAIMED is set the vnode
415 * could already be in the middle of a recycle.
416 * Callers must use cache_vref() or
417 * cache_vget() on the locked ncp to
418 * validate the vp or set the cache entry
421 * NOTE! vhold() is allowed if we hold a
422 * lock on the ncp (which we do).
426 atomic_clear_int(&ncp->nc_lockstatus,
437 * If already held shared we can just bump the count, but
438 * only allow this if nobody is trying to get the lock
439 * exclusively. If we are blocking too long ignore excl
440 * requests (which can race/deadlock us).
442 * VHOLD is a bit of a hack. Even though we successfully
443 * added another shared ref, the cpu that got the first
444 * shared ref might not yet have held the vnode.
446 if ((count & (optreq|NC_SHLOCK_FLAG)) == NC_SHLOCK_FLAG) {
447 KKASSERT((count & ~(NC_EXLOCK_REQ |
449 NC_SHLOCK_FLAG)) > 0);
450 if (atomic_cmpset_int(&ncp->nc_lockstatus,
452 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD)
458 tsleep_interlock(ncp, 0);
459 if (atomic_cmpset_int(&ncp->nc_lockstatus, count,
460 count | NC_SHLOCK_REQ) == 0) {
464 error = tsleep(ncp, PINTERLOCKED, "clocksh", nclockwarn);
465 if (error == EWOULDBLOCK) {
469 kprintf("[diagnostic] cache_lock_shared: "
470 "blocked on %p %08x",
472 kprintf(" \"%*.*s\"\n",
473 ncp->nc_nlen, ncp->nc_nlen,
480 kprintf("[diagnostic] cache_lock_shared: "
481 "unblocked %*.*s after %d secs\n",
482 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
483 (int)(ticks - didwarn) / hz);
488 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance,
489 * such as the case where one of its children is locked.
493 _cache_lock_nonblock(struct namecache *ncp)
501 count = ncp->nc_lockstatus;
503 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) {
504 if (atomic_cmpset_int(&ncp->nc_lockstatus,
507 * The vp associated with a locked ncp must
508 * be held to prevent it from being recycled.
510 * WARNING! If VRECLAIMED is set the vnode
511 * could already be in the middle of a recycle.
512 * Callers must use cache_vref() or
513 * cache_vget() on the locked ncp to
514 * validate the vp or set the cache entry
517 * NOTE! vhold() is allowed if we hold a
518 * lock on the ncp (which we do).
528 if (ncp->nc_locktd == td) {
529 if (atomic_cmpset_int(&ncp->nc_lockstatus,
542 * The shared lock works similarly to the exclusive lock except
543 * nc_locktd is left NULL and we need an interlock (VHOLD) to
544 * prevent vhold() races, since the moment our cmpset_int succeeds
545 * another cpu can come in and get its own shared lock.
547 * A critical section is needed to prevent interruption during the
552 _cache_lock_shared_nonblock(struct namecache *ncp)
557 count = ncp->nc_lockstatus;
559 if ((count & ~NC_SHLOCK_REQ) == 0) {
561 if (atomic_cmpset_int(&ncp->nc_lockstatus,
563 (count + 1) | NC_SHLOCK_FLAG |
566 * The vp associated with a locked ncp must
567 * be held to prevent it from being recycled.
569 * WARNING! If VRECLAIMED is set the vnode
570 * could already be in the middle of a recycle.
571 * Callers must use cache_vref() or
572 * cache_vget() on the locked ncp to
573 * validate the vp or set the cache entry
576 * NOTE! vhold() is allowed if we hold a
577 * lock on the ncp (which we do).
581 atomic_clear_int(&ncp->nc_lockstatus,
592 * If already held shared we can just bump the count, but
593 * only allow this if nobody is trying to get the lock
596 * VHOLD is a bit of a hack. Even though we successfully
597 * added another shared ref, the cpu that got the first
598 * shared ref might not yet have held the vnode.
600 if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) ==
602 KKASSERT((count & ~(NC_EXLOCK_REQ |
604 NC_SHLOCK_FLAG)) > 0);
605 if (atomic_cmpset_int(&ncp->nc_lockstatus,
607 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD)
621 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop).
623 * nc_locktd must be NULLed out prior to nc_lockstatus getting cleared.
627 _cache_unlock(struct namecache *ncp)
629 thread_t td __debugvar = curthread;
632 struct vnode *dropvp;
634 KKASSERT(ncp->nc_refs >= 0);
635 KKASSERT((ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) > 0);
636 KKASSERT((ncp->nc_lockstatus & NC_SHLOCK_FLAG) || ncp->nc_locktd == td);
638 count = ncp->nc_lockstatus;
642 * Clear nc_locktd prior to the atomic op (excl lock only)
644 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1)
645 ncp->nc_locktd = NULL;
650 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ|NC_SHLOCK_FLAG)) == 1) {
652 if (count & NC_EXLOCK_REQ)
653 ncount = count & NC_SHLOCK_REQ; /* cnt->0 */
657 if (atomic_cmpset_int(&ncp->nc_lockstatus,
659 if (count & NC_EXLOCK_REQ)
660 wakeup(&ncp->nc_locktd);
661 else if (count & NC_SHLOCK_REQ)
667 KKASSERT((count & NC_SHLOCK_VHOLD) == 0);
668 KKASSERT((count & ~(NC_EXLOCK_REQ |
670 NC_SHLOCK_FLAG)) > 1);
671 if (atomic_cmpset_int(&ncp->nc_lockstatus,
676 count = ncp->nc_lockstatus;
681 * Don't actually drop the vp until we successfully clean out
682 * the lock, otherwise we may race another shared lock.
690 _cache_lockstatus(struct namecache *ncp)
692 if (ncp->nc_locktd == curthread)
693 return(LK_EXCLUSIVE);
694 if (ncp->nc_lockstatus & NC_SHLOCK_FLAG)
700 * cache_hold() and cache_drop() prevent the premature deletion of a
701 * namecache entry but do not prevent operations (such as zapping) on
702 * that namecache entry.
704 * This routine may only be called from outside this source module if
705 * nc_refs is already at least 1.
707 * This is a rare case where callers are allowed to hold a spinlock,
708 * so we can't ourselves.
712 _cache_hold(struct namecache *ncp)
714 atomic_add_int(&ncp->nc_refs, 1);
719 * Drop a cache entry, taking care to deal with races.
721 * For potential 1->0 transitions we must hold the ncp lock to safely
722 * test its flags. An unresolved entry with no children must be zapped
725 * The call to cache_zap() itself will handle all remaining races and
726 * will decrement the ncp's refs regardless. If we are resolved or
727 * have children nc_refs can safely be dropped to 0 without having to
730 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion.
732 * NOTE: cache_zap() may return a non-NULL referenced parent which must
733 * be dropped in a loop.
737 _cache_drop(struct namecache *ncp)
742 KKASSERT(ncp->nc_refs > 0);
746 if (_cache_lock_nonblock(ncp) == 0) {
747 ncp->nc_flag &= ~NCF_DEFEREDZAP;
748 if ((ncp->nc_flag & NCF_UNRESOLVED) &&
749 TAILQ_EMPTY(&ncp->nc_list)) {
750 ncp = cache_zap(ncp, 1);
753 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) {
760 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1))
768 * Link a new namecache entry to its parent and to the hash table. Be
769 * careful to avoid races if vhold() blocks in the future.
771 * Both ncp and par must be referenced and locked.
773 * NOTE: The hash table spinlock is held during this call, we can't do
777 _cache_link_parent(struct namecache *ncp, struct namecache *par,
778 struct nchash_head *nchpp)
780 KKASSERT(ncp->nc_parent == NULL);
781 ncp->nc_parent = par;
782 ncp->nc_head = nchpp;
785 * Set inheritance flags. Note that the parent flags may be
786 * stale due to getattr potentially not having been run yet
787 * (it gets run during nlookup()'s).
789 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE);
790 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE))
791 ncp->nc_flag |= NCF_SF_PNOCACHE;
792 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE))
793 ncp->nc_flag |= NCF_UF_PCACHE;
795 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash);
797 if (TAILQ_EMPTY(&par->nc_list)) {
798 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
800 * Any vp associated with an ncp which has children must
801 * be held to prevent it from being recycled.
806 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
811 * Remove the parent and hash associations from a namecache structure.
812 * If this is the last child of the parent the cache_drop(par) will
813 * attempt to recursively zap the parent.
815 * ncp must be locked. This routine will acquire a temporary lock on
816 * the parent as wlel as the appropriate hash chain.
819 _cache_unlink_parent(struct namecache *ncp)
821 struct namecache *par;
822 struct vnode *dropvp;
824 if ((par = ncp->nc_parent) != NULL) {
825 KKASSERT(ncp->nc_parent == par);
828 spin_lock(&ncp->nc_head->spin);
829 LIST_REMOVE(ncp, nc_hash);
830 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
832 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
834 spin_unlock(&ncp->nc_head->spin);
835 ncp->nc_parent = NULL;
841 * We can only safely vdrop with no spinlocks held.
849 * Allocate a new namecache structure. Most of the code does not require
850 * zero-termination of the string but it makes vop_compat_ncreate() easier.
852 static struct namecache *
853 cache_alloc(int nlen)
855 struct namecache *ncp;
857 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
859 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
861 ncp->nc_flag = NCF_UNRESOLVED;
862 ncp->nc_error = ENOTCONN; /* needs to be resolved */
865 TAILQ_INIT(&ncp->nc_list);
871 * Can only be called for the case where the ncp has never been
872 * associated with anything (so no spinlocks are needed).
875 _cache_free(struct namecache *ncp)
877 KKASSERT(ncp->nc_refs == 1 && ncp->nc_lockstatus == 1);
879 kfree(ncp->nc_name, M_VFSCACHE);
880 kfree(ncp, M_VFSCACHE);
884 * [re]initialize a nchandle.
887 cache_zero(struct nchandle *nch)
894 * Ref and deref a namecache structure.
896 * The caller must specify a stable ncp pointer, typically meaning the
897 * ncp is already referenced but this can also occur indirectly through
898 * e.g. holding a lock on a direct child.
900 * WARNING: Caller may hold an unrelated read spinlock, which means we can't
901 * use read spinlocks here.
906 cache_hold(struct nchandle *nch)
908 _cache_hold(nch->ncp);
909 atomic_add_int(&nch->mount->mnt_refs, 1);
914 * Create a copy of a namecache handle for an already-referenced
920 cache_copy(struct nchandle *nch, struct nchandle *target)
924 _cache_hold(target->ncp);
925 atomic_add_int(&nch->mount->mnt_refs, 1);
932 cache_changemount(struct nchandle *nch, struct mount *mp)
934 atomic_add_int(&nch->mount->mnt_refs, -1);
936 atomic_add_int(&nch->mount->mnt_refs, 1);
940 cache_drop(struct nchandle *nch)
942 atomic_add_int(&nch->mount->mnt_refs, -1);
943 _cache_drop(nch->ncp);
949 cache_lockstatus(struct nchandle *nch)
951 return(_cache_lockstatus(nch->ncp));
955 cache_lock(struct nchandle *nch)
957 _cache_lock(nch->ncp);
961 cache_lock_maybe_shared(struct nchandle *nch, int excl)
963 struct namecache *ncp = nch->ncp;
965 if (ncp_shared_lock_disable || excl ||
966 (ncp->nc_flag & NCF_UNRESOLVED)) {
969 _cache_lock_shared(ncp);
970 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
971 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) {
983 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller
984 * is responsible for checking both for validity on return as they
985 * may have become invalid.
987 * We have to deal with potential deadlocks here, just ping pong
988 * the lock until we get it (we will always block somewhere when
989 * looping so this is not cpu-intensive).
991 * which = 0 nch1 not locked, nch2 is locked
992 * which = 1 nch1 is locked, nch2 is not locked
995 cache_relock(struct nchandle *nch1, struct ucred *cred1,
996 struct nchandle *nch2, struct ucred *cred2)
1004 if (cache_lock_nonblock(nch1) == 0) {
1005 cache_resolve(nch1, cred1);
1010 cache_resolve(nch1, cred1);
1013 if (cache_lock_nonblock(nch2) == 0) {
1014 cache_resolve(nch2, cred2);
1019 cache_resolve(nch2, cred2);
1026 cache_lock_nonblock(struct nchandle *nch)
1028 return(_cache_lock_nonblock(nch->ncp));
1032 cache_unlock(struct nchandle *nch)
1034 _cache_unlock(nch->ncp);
1038 * ref-and-lock, unlock-and-deref functions.
1040 * This function is primarily used by nlookup. Even though cache_lock
1041 * holds the vnode, it is possible that the vnode may have already
1042 * initiated a recyclement.
1044 * We want cache_get() to return a definitively usable vnode or a
1045 * definitively unresolved ncp.
1049 _cache_get(struct namecache *ncp)
1053 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1054 _cache_setunresolved(ncp);
1059 * Attempt to obtain a shared lock on the ncp. A shared lock will only
1060 * be obtained if the ncp is resolved and the vnode (if not ENOENT) is
1061 * valid. Otherwise an exclusive lock will be acquired instead.
1065 _cache_get_maybe_shared(struct namecache *ncp, int excl)
1067 if (ncp_shared_lock_disable || excl ||
1068 (ncp->nc_flag & NCF_UNRESOLVED)) {
1069 return(_cache_get(ncp));
1072 _cache_lock_shared(ncp);
1073 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1074 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) {
1076 ncp = _cache_get(ncp);
1081 ncp = _cache_get(ncp);
1088 * This is a special form of _cache_lock() which only succeeds if
1089 * it can get a pristine, non-recursive lock. The caller must have
1090 * already ref'd the ncp.
1092 * On success the ncp will be locked, on failure it will not. The
1093 * ref count does not change either way.
1095 * We want _cache_lock_special() (on success) to return a definitively
1096 * usable vnode or a definitively unresolved ncp.
1099 _cache_lock_special(struct namecache *ncp)
1101 if (_cache_lock_nonblock(ncp) == 0) {
1102 if ((ncp->nc_lockstatus &
1103 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) {
1104 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1105 _cache_setunresolved(ncp);
1110 return(EWOULDBLOCK);
1114 * This function tries to get a shared lock but will back-off to an exclusive
1117 * (1) Some other thread is trying to obtain an exclusive lock
1118 * (to prevent the exclusive requester from getting livelocked out
1119 * by many shared locks).
1121 * (2) The current thread already owns an exclusive lock (to avoid
1124 * WARNING! On machines with lots of cores we really want to try hard to
1125 * get a shared lock or concurrent path lookups can chain-react
1126 * into a very high-latency exclusive lock.
1129 _cache_lock_shared_special(struct namecache *ncp)
1131 if (_cache_lock_shared_nonblock(ncp) == 0) {
1132 if ((ncp->nc_lockstatus &
1133 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == (NC_SHLOCK_FLAG | 1)) {
1134 if (ncp->nc_vp == NULL ||
1135 (ncp->nc_vp->v_flag & VRECLAIMED) == 0) {
1140 return(EWOULDBLOCK);
1142 if (ncp->nc_locktd == curthread) {
1146 _cache_lock_shared(ncp);
1152 * NOTE: The same nchandle can be passed for both arguments.
1155 cache_get(struct nchandle *nch, struct nchandle *target)
1157 KKASSERT(nch->ncp->nc_refs > 0);
1158 target->mount = nch->mount;
1159 target->ncp = _cache_get(nch->ncp);
1160 atomic_add_int(&target->mount->mnt_refs, 1);
1164 cache_get_maybe_shared(struct nchandle *nch, struct nchandle *target, int excl)
1166 KKASSERT(nch->ncp->nc_refs > 0);
1167 target->mount = nch->mount;
1168 target->ncp = _cache_get_maybe_shared(nch->ncp, excl);
1169 atomic_add_int(&target->mount->mnt_refs, 1);
1177 _cache_put(struct namecache *ncp)
1187 cache_put(struct nchandle *nch)
1189 atomic_add_int(&nch->mount->mnt_refs, -1);
1190 _cache_put(nch->ncp);
1196 * Resolve an unresolved ncp by associating a vnode with it. If the
1197 * vnode is NULL, a negative cache entry is created.
1199 * The ncp should be locked on entry and will remain locked on return.
1203 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp)
1205 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
1206 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
1210 * Any vp associated with an ncp which has children must
1211 * be held. Any vp associated with a locked ncp must be held.
1213 if (!TAILQ_EMPTY(&ncp->nc_list))
1215 spin_lock(&vp->v_spin);
1217 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
1218 spin_unlock(&vp->v_spin);
1219 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ))
1223 * Set auxiliary flags
1225 switch(vp->v_type) {
1227 ncp->nc_flag |= NCF_ISDIR;
1230 ncp->nc_flag |= NCF_ISSYMLINK;
1231 /* XXX cache the contents of the symlink */
1236 atomic_add_int(&numcache, 1);
1238 /* XXX: this is a hack to work-around the lack of a real pfs vfs
1241 if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0)
1245 * When creating a negative cache hit we set the
1246 * namecache_gen. A later resolve will clean out the
1247 * negative cache hit if the mount point's namecache_gen
1248 * has changed. Used by devfs, could also be used by
1253 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
1255 spin_unlock(&ncspin);
1256 ncp->nc_error = ENOENT;
1258 VFS_NCPGEN_SET(mp, ncp);
1260 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP);
1267 cache_setvp(struct nchandle *nch, struct vnode *vp)
1269 _cache_setvp(nch->mount, nch->ncp, vp);
1276 cache_settimeout(struct nchandle *nch, int nticks)
1278 struct namecache *ncp = nch->ncp;
1280 if ((ncp->nc_timeout = ticks + nticks) == 0)
1281 ncp->nc_timeout = 1;
1285 * Disassociate the vnode or negative-cache association and mark a
1286 * namecache entry as unresolved again. Note that the ncp is still
1287 * left in the hash table and still linked to its parent.
1289 * The ncp should be locked and refd on entry and will remain locked and refd
1292 * This routine is normally never called on a directory containing children.
1293 * However, NFS often does just that in its rename() code as a cop-out to
1294 * avoid complex namespace operations. This disconnects a directory vnode
1295 * from its namecache and can cause the OLDAPI and NEWAPI to get out of
1301 _cache_setunresolved(struct namecache *ncp)
1305 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1306 ncp->nc_flag |= NCF_UNRESOLVED;
1307 ncp->nc_timeout = 0;
1308 ncp->nc_error = ENOTCONN;
1309 if ((vp = ncp->nc_vp) != NULL) {
1310 atomic_add_int(&numcache, -1);
1311 spin_lock(&vp->v_spin);
1313 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
1314 spin_unlock(&vp->v_spin);
1317 * Any vp associated with an ncp with children is
1318 * held by that ncp. Any vp associated with a locked
1319 * ncp is held by that ncp. These conditions must be
1320 * undone when the vp is cleared out from the ncp.
1322 if (!TAILQ_EMPTY(&ncp->nc_list))
1324 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ))
1328 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1330 spin_unlock(&ncspin);
1332 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK);
1337 * The cache_nresolve() code calls this function to automatically
1338 * set a resolved cache element to unresolved if it has timed out
1339 * or if it is a negative cache hit and the mount point namecache_gen
1343 _cache_auto_unresolve_test(struct mount *mp, struct namecache *ncp)
1346 * Try to zap entries that have timed out. We have
1347 * to be careful here because locked leafs may depend
1348 * on the vnode remaining intact in a parent, so only
1349 * do this under very specific conditions.
1351 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 &&
1352 TAILQ_EMPTY(&ncp->nc_list)) {
1357 * If a resolved negative cache hit is invalid due to
1358 * the mount's namecache generation being bumped, zap it.
1360 if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) {
1365 * Otherwise we are good
1370 static __inline void
1371 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp)
1374 * Already in an unresolved state, nothing to do.
1376 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1377 if (_cache_auto_unresolve_test(mp, ncp))
1378 _cache_setunresolved(ncp);
1386 cache_setunresolved(struct nchandle *nch)
1388 _cache_setunresolved(nch->ncp);
1392 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
1393 * looking for matches. This flag tells the lookup code when it must
1394 * check for a mount linkage and also prevents the directories in question
1395 * from being deleted or renamed.
1399 cache_clrmountpt_callback(struct mount *mp, void *data)
1401 struct nchandle *nch = data;
1403 if (mp->mnt_ncmounton.ncp == nch->ncp)
1405 if (mp->mnt_ncmountpt.ncp == nch->ncp)
1414 cache_clrmountpt(struct nchandle *nch)
1418 count = mountlist_scan(cache_clrmountpt_callback, nch,
1419 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1421 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
1425 * Invalidate portions of the namecache topology given a starting entry.
1426 * The passed ncp is set to an unresolved state and:
1428 * The passed ncp must be referencxed and locked. The routine may unlock
1429 * and relock ncp several times, and will recheck the children and loop
1430 * to catch races. When done the passed ncp will be returned with the
1431 * reference and lock intact.
1433 * CINV_DESTROY - Set a flag in the passed ncp entry indicating
1434 * that the physical underlying nodes have been
1435 * destroyed... as in deleted. For example, when
1436 * a directory is removed. This will cause record
1437 * lookups on the name to no longer be able to find
1438 * the record and tells the resolver to return failure
1439 * rather then trying to resolve through the parent.
1441 * The topology itself, including ncp->nc_name,
1444 * This only applies to the passed ncp, if CINV_CHILDREN
1445 * is specified the children are not flagged.
1447 * CINV_CHILDREN - Set all children (recursively) to an unresolved
1450 * Note that this will also have the side effect of
1451 * cleaning out any unreferenced nodes in the topology
1452 * from the leaves up as the recursion backs out.
1454 * Note that the topology for any referenced nodes remains intact, but
1455 * the nodes will be marked as having been destroyed and will be set
1456 * to an unresolved state.
1458 * It is possible for cache_inval() to race a cache_resolve(), meaning that
1459 * the namecache entry may not actually be invalidated on return if it was
1460 * revalidated while recursing down into its children. This code guarentees
1461 * that the node(s) will go through an invalidation cycle, but does not
1462 * guarentee that they will remain in an invalidated state.
1464 * Returns non-zero if a revalidation was detected during the invalidation
1465 * recursion, zero otherwise. Note that since only the original ncp is
1466 * locked the revalidation ultimately can only indicate that the original ncp
1467 * *MIGHT* no have been reresolved.
1469 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
1470 * have to avoid blowing out the kernel stack. We do this by saving the
1471 * deep namecache node and aborting the recursion, then re-recursing at that
1472 * node using a depth-first algorithm in order to allow multiple deep
1473 * recursions to chain through each other, then we restart the invalidation
1478 struct namecache *resume_ncp;
1482 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
1486 _cache_inval(struct namecache *ncp, int flags)
1488 struct cinvtrack track;
1489 struct namecache *ncp2;
1493 track.resume_ncp = NULL;
1496 r = _cache_inval_internal(ncp, flags, &track);
1497 if (track.resume_ncp == NULL)
1499 kprintf("Warning: deep namecache recursion at %s\n",
1502 while ((ncp2 = track.resume_ncp) != NULL) {
1503 track.resume_ncp = NULL;
1505 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
1515 cache_inval(struct nchandle *nch, int flags)
1517 return(_cache_inval(nch->ncp, flags));
1521 * Helper for _cache_inval(). The passed ncp is refd and locked and
1522 * remains that way on return, but may be unlocked/relocked multiple
1523 * times by the routine.
1526 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
1528 struct namecache *kid;
1529 struct namecache *nextkid;
1532 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
1534 _cache_setunresolved(ncp);
1535 if (flags & CINV_DESTROY)
1536 ncp->nc_flag |= NCF_DESTROYED;
1537 if ((flags & CINV_CHILDREN) &&
1538 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
1541 if (++track->depth > MAX_RECURSION_DEPTH) {
1542 track->resume_ncp = ncp;
1548 if (track->resume_ncp) {
1552 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
1553 _cache_hold(nextkid);
1554 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
1555 TAILQ_FIRST(&kid->nc_list)
1558 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
1569 * Someone could have gotten in there while ncp was unlocked,
1572 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1578 * Invalidate a vnode's namecache associations. To avoid races against
1579 * the resolver we do not invalidate a node which we previously invalidated
1580 * but which was then re-resolved while we were in the invalidation loop.
1582 * Returns non-zero if any namecache entries remain after the invalidation
1585 * NOTE: Unlike the namecache topology which guarentees that ncp's will not
1586 * be ripped out of the topology while held, the vnode's v_namecache
1587 * list has no such restriction. NCP's can be ripped out of the list
1588 * at virtually any time if not locked, even if held.
1590 * In addition, the v_namecache list itself must be locked via
1591 * the vnode's spinlock.
1594 cache_inval_vp(struct vnode *vp, int flags)
1596 struct namecache *ncp;
1597 struct namecache *next;
1600 spin_lock(&vp->v_spin);
1601 ncp = TAILQ_FIRST(&vp->v_namecache);
1605 /* loop entered with ncp held and vp spin-locked */
1606 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1608 spin_unlock(&vp->v_spin);
1610 if (ncp->nc_vp != vp) {
1611 kprintf("Warning: cache_inval_vp: race-A detected on "
1612 "%s\n", ncp->nc_name);
1618 _cache_inval(ncp, flags);
1619 _cache_put(ncp); /* also releases reference */
1621 spin_lock(&vp->v_spin);
1622 if (ncp && ncp->nc_vp != vp) {
1623 spin_unlock(&vp->v_spin);
1624 kprintf("Warning: cache_inval_vp: race-B detected on "
1625 "%s\n", ncp->nc_name);
1630 spin_unlock(&vp->v_spin);
1631 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1635 * This routine is used instead of the normal cache_inval_vp() when we
1636 * are trying to recycle otherwise good vnodes.
1638 * Return 0 on success, non-zero if not all namecache records could be
1639 * disassociated from the vnode (for various reasons).
1642 cache_inval_vp_nonblock(struct vnode *vp)
1644 struct namecache *ncp;
1645 struct namecache *next;
1647 spin_lock(&vp->v_spin);
1648 ncp = TAILQ_FIRST(&vp->v_namecache);
1652 /* loop entered with ncp held */
1653 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1655 spin_unlock(&vp->v_spin);
1656 if (_cache_lock_nonblock(ncp)) {
1662 if (ncp->nc_vp != vp) {
1663 kprintf("Warning: cache_inval_vp: race-A detected on "
1664 "%s\n", ncp->nc_name);
1670 _cache_inval(ncp, 0);
1671 _cache_put(ncp); /* also releases reference */
1673 spin_lock(&vp->v_spin);
1674 if (ncp && ncp->nc_vp != vp) {
1675 spin_unlock(&vp->v_spin);
1676 kprintf("Warning: cache_inval_vp: race-B detected on "
1677 "%s\n", ncp->nc_name);
1682 spin_unlock(&vp->v_spin);
1684 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1688 * The source ncp has been renamed to the target ncp. Both fncp and tncp
1689 * must be locked. The target ncp is destroyed (as a normal rename-over
1690 * would destroy the target file or directory).
1692 * Because there may be references to the source ncp we cannot copy its
1693 * contents to the target. Instead the source ncp is relinked as the target
1694 * and the target ncp is removed from the namecache topology.
1697 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
1699 struct namecache *fncp = fnch->ncp;
1700 struct namecache *tncp = tnch->ncp;
1701 struct namecache *tncp_par;
1702 struct nchash_head *nchpp;
1707 if (tncp->nc_nlen) {
1708 nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK);
1709 bcopy(tncp->nc_name, nname, tncp->nc_nlen);
1710 nname[tncp->nc_nlen] = 0;
1716 * Rename fncp (unlink)
1718 _cache_unlink_parent(fncp);
1719 oname = fncp->nc_name;
1720 fncp->nc_name = nname;
1721 fncp->nc_nlen = tncp->nc_nlen;
1723 kfree(oname, M_VFSCACHE);
1725 tncp_par = tncp->nc_parent;
1726 _cache_hold(tncp_par);
1727 _cache_lock(tncp_par);
1730 * Rename fncp (relink)
1732 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT);
1733 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash);
1734 nchpp = NCHHASH(hash);
1736 spin_lock(&nchpp->spin);
1737 _cache_link_parent(fncp, tncp_par, nchpp);
1738 spin_unlock(&nchpp->spin);
1740 _cache_put(tncp_par);
1743 * Get rid of the overwritten tncp (unlink)
1745 _cache_unlink(tncp);
1749 * Perform actions consistent with unlinking a file. The passed-in ncp
1752 * The ncp is marked DESTROYED so it no longer shows up in searches,
1753 * and will be physically deleted when the vnode goes away.
1755 * If the related vnode has no refs then we cycle it through vget()/vput()
1756 * to (possibly if we don't have a ref race) trigger a deactivation,
1757 * allowing the VFS to trivially detect and recycle the deleted vnode
1758 * via VOP_INACTIVE().
1760 * NOTE: _cache_rename() will automatically call _cache_unlink() on the
1764 cache_unlink(struct nchandle *nch)
1766 _cache_unlink(nch->ncp);
1770 _cache_unlink(struct namecache *ncp)
1775 * Causes lookups to fail and allows another ncp with the same
1776 * name to be created under ncp->nc_parent.
1778 ncp->nc_flag |= NCF_DESTROYED;
1781 * Attempt to trigger a deactivation.
1783 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1784 (vp = ncp->nc_vp) != NULL &&
1785 !sysref_isactive(&vp->v_sysref)) {
1786 if (vget(vp, LK_SHARED) == 0)
1792 * vget the vnode associated with the namecache entry. Resolve the namecache
1793 * entry if necessary. The passed ncp must be referenced and locked.
1795 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked
1796 * (depending on the passed lk_type) will be returned in *vpp with an error
1797 * of 0, or NULL will be returned in *vpp with a non-0 error code. The
1798 * most typical error is ENOENT, meaning that the ncp represents a negative
1799 * cache hit and there is no vnode to retrieve, but other errors can occur
1802 * The vget() can race a reclaim. If this occurs we re-resolve the
1805 * There are numerous places in the kernel where vget() is called on a
1806 * vnode while one or more of its namecache entries is locked. Releasing
1807 * a vnode never deadlocks against locked namecache entries (the vnode
1808 * will not get recycled while referenced ncp's exist). This means we
1809 * can safely acquire the vnode. In fact, we MUST NOT release the ncp
1810 * lock when acquiring the vp lock or we might cause a deadlock.
1812 * NOTE: The passed-in ncp must be locked exclusively if it is initially
1813 * unresolved. If a reclaim race occurs the passed-in ncp will be
1814 * relocked exclusively before being re-resolved.
1817 cache_vget(struct nchandle *nch, struct ucred *cred,
1818 int lk_type, struct vnode **vpp)
1820 struct namecache *ncp;
1827 if (ncp->nc_flag & NCF_UNRESOLVED)
1828 error = cache_resolve(nch, cred);
1832 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1833 error = vget(vp, lk_type);
1838 if (error == ENOENT) {
1839 kprintf("Warning: vnode reclaim race detected "
1840 "in cache_vget on %p (%s)\n",
1844 _cache_setunresolved(ncp);
1849 * Not a reclaim race, some other error.
1851 KKASSERT(ncp->nc_vp == vp);
1854 KKASSERT(ncp->nc_vp == vp);
1855 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1858 if (error == 0 && vp == NULL)
1865 * Similar to cache_vget() but only acquires a ref on the vnode.
1867 * NOTE: The passed-in ncp must be locked exclusively if it is initially
1868 * unresolved. If a reclaim race occurs the passed-in ncp will be
1869 * relocked exclusively before being re-resolved.
1872 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
1874 struct namecache *ncp;
1881 if (ncp->nc_flag & NCF_UNRESOLVED)
1882 error = cache_resolve(nch, cred);
1886 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1887 error = vget(vp, LK_SHARED);
1892 if (error == ENOENT) {
1893 kprintf("Warning: vnode reclaim race detected "
1894 "in cache_vget on %p (%s)\n",
1898 _cache_setunresolved(ncp);
1903 * Not a reclaim race, some other error.
1905 KKASSERT(ncp->nc_vp == vp);
1908 KKASSERT(ncp->nc_vp == vp);
1909 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1910 /* caller does not want a lock */
1914 if (error == 0 && vp == NULL)
1921 * Return a referenced vnode representing the parent directory of
1924 * Because the caller has locked the ncp it should not be possible for
1925 * the parent ncp to go away. However, the parent can unresolve its
1926 * dvp at any time so we must be able to acquire a lock on the parent
1927 * to safely access nc_vp.
1929 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock,
1930 * so use vhold()/vdrop() while holding the lock to prevent dvp from
1931 * getting destroyed.
1933 * NOTE: vhold() is allowed when dvp has 0 refs if we hold a
1934 * lock on the ncp in question..
1936 static struct vnode *
1937 cache_dvpref(struct namecache *ncp)
1939 struct namecache *par;
1943 if ((par = ncp->nc_parent) != NULL) {
1946 if ((par->nc_flag & NCF_UNRESOLVED) == 0) {
1947 if ((dvp = par->nc_vp) != NULL)
1952 if (vget(dvp, LK_SHARED) == 0) {
1955 /* return refd, unlocked dvp */
1967 * Convert a directory vnode to a namecache record without any other
1968 * knowledge of the topology. This ONLY works with directory vnodes and
1969 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the
1970 * returned ncp (if not NULL) will be held and unlocked.
1972 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1973 * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1974 * for dvp. This will fail only if the directory has been deleted out from
1977 * Callers must always check for a NULL return no matter the value of 'makeit'.
1979 * To avoid underflowing the kernel stack each recursive call increments
1980 * the makeit variable.
1983 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1984 struct vnode *dvp, char *fakename);
1985 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1986 struct vnode **saved_dvp);
1989 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
1990 struct nchandle *nch)
1992 struct vnode *saved_dvp;
1998 nch->mount = dvp->v_mount;
2003 * Handle the makeit == 0 degenerate case
2006 spin_lock(&dvp->v_spin);
2007 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
2010 spin_unlock(&dvp->v_spin);
2014 * Loop until resolution, inside code will break out on error.
2018 * Break out if we successfully acquire a working ncp.
2020 spin_lock(&dvp->v_spin);
2021 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
2024 spin_unlock(&dvp->v_spin);
2027 spin_unlock(&dvp->v_spin);
2030 * If dvp is the root of its filesystem it should already
2031 * have a namecache pointer associated with it as a side
2032 * effect of the mount, but it may have been disassociated.
2034 if (dvp->v_flag & VROOT) {
2035 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
2036 error = cache_resolve_mp(nch->mount);
2037 _cache_put(nch->ncp);
2039 kprintf("cache_fromdvp: resolve root of mount %p error %d",
2040 dvp->v_mount, error);
2044 kprintf(" failed\n");
2049 kprintf(" succeeded\n");
2054 * If we are recursed too deeply resort to an O(n^2)
2055 * algorithm to resolve the namecache topology. The
2056 * resolved pvp is left referenced in saved_dvp to
2057 * prevent the tree from being destroyed while we loop.
2060 error = cache_fromdvp_try(dvp, cred, &saved_dvp);
2062 kprintf("lookupdotdot(longpath) failed %d "
2063 "dvp %p\n", error, dvp);
2071 * Get the parent directory and resolve its ncp.
2074 kfree(fakename, M_TEMP);
2077 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
2080 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
2086 * Reuse makeit as a recursion depth counter. On success
2087 * nch will be fully referenced.
2089 cache_fromdvp(pvp, cred, makeit + 1, nch);
2091 if (nch->ncp == NULL)
2095 * Do an inefficient scan of pvp (embodied by ncp) to look
2096 * for dvp. This will create a namecache record for dvp on
2097 * success. We loop up to recheck on success.
2099 * ncp and dvp are both held but not locked.
2101 error = cache_inefficient_scan(nch, cred, dvp, fakename);
2103 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
2104 pvp, nch->ncp->nc_name, dvp);
2106 /* nch was NULLed out, reload mount */
2107 nch->mount = dvp->v_mount;
2111 kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
2112 pvp, nch->ncp->nc_name);
2115 /* nch was NULLed out, reload mount */
2116 nch->mount = dvp->v_mount;
2120 * If nch->ncp is non-NULL it will have been held already.
2123 kfree(fakename, M_TEMP);
2132 * Go up the chain of parent directories until we find something
2133 * we can resolve into the namecache. This is very inefficient.
2137 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
2138 struct vnode **saved_dvp)
2140 struct nchandle nch;
2143 static time_t last_fromdvp_report;
2147 * Loop getting the parent directory vnode until we get something we
2148 * can resolve in the namecache.
2151 nch.mount = dvp->v_mount;
2157 kfree(fakename, M_TEMP);
2160 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
2167 spin_lock(&pvp->v_spin);
2168 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
2169 _cache_hold(nch.ncp);
2170 spin_unlock(&pvp->v_spin);
2174 spin_unlock(&pvp->v_spin);
2175 if (pvp->v_flag & VROOT) {
2176 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
2177 error = cache_resolve_mp(nch.mount);
2178 _cache_unlock(nch.ncp);
2181 _cache_drop(nch.ncp);
2191 if (last_fromdvp_report != time_uptime) {
2192 last_fromdvp_report = time_uptime;
2193 kprintf("Warning: extremely inefficient path "
2194 "resolution on %s\n",
2197 error = cache_inefficient_scan(&nch, cred, dvp, fakename);
2200 * Hopefully dvp now has a namecache record associated with
2201 * it. Leave it referenced to prevent the kernel from
2202 * recycling the vnode. Otherwise extremely long directory
2203 * paths could result in endless recycling.
2208 _cache_drop(nch.ncp);
2211 kfree(fakename, M_TEMP);
2216 * Do an inefficient scan of the directory represented by ncp looking for
2217 * the directory vnode dvp. ncp must be held but not locked on entry and
2218 * will be held on return. dvp must be refd but not locked on entry and
2219 * will remain refd on return.
2221 * Why do this at all? Well, due to its stateless nature the NFS server
2222 * converts file handles directly to vnodes without necessarily going through
2223 * the namecache ops that would otherwise create the namecache topology
2224 * leading to the vnode. We could either (1) Change the namecache algorithms
2225 * to allow disconnect namecache records that are re-merged opportunistically,
2226 * or (2) Make the NFS server backtrack and scan to recover a connected
2227 * namecache topology in order to then be able to issue new API lookups.
2229 * It turns out that (1) is a huge mess. It takes a nice clean set of
2230 * namecache algorithms and introduces a lot of complication in every subsystem
2231 * that calls into the namecache to deal with the re-merge case, especially
2232 * since we are using the namecache to placehold negative lookups and the
2233 * vnode might not be immediately assigned. (2) is certainly far less
2234 * efficient then (1), but since we are only talking about directories here
2235 * (which are likely to remain cached), the case does not actually run all
2236 * that often and has the supreme advantage of not polluting the namecache
2239 * If a fakename is supplied just construct a namecache entry using the
2243 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
2244 struct vnode *dvp, char *fakename)
2246 struct nlcomponent nlc;
2247 struct nchandle rncp;
2259 vat.va_blocksize = 0;
2260 if ((error = VOP_GETATTR(dvp, &vat)) != 0)
2263 error = cache_vref(nch, cred, &pvp);
2268 kprintf("inefficient_scan: directory iosize %ld "
2269 "vattr fileid = %lld\n",
2271 (long long)vat.va_fileid);
2275 * Use the supplied fakename if not NULL. Fake names are typically
2276 * not in the actual filesystem hierarchy. This is used by HAMMER
2277 * to glue @@timestamp recursions together.
2280 nlc.nlc_nameptr = fakename;
2281 nlc.nlc_namelen = strlen(fakename);
2282 rncp = cache_nlookup(nch, &nlc);
2286 if ((blksize = vat.va_blocksize) == 0)
2287 blksize = DEV_BSIZE;
2288 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
2294 iov.iov_base = rbuf;
2295 iov.iov_len = blksize;
2298 uio.uio_resid = blksize;
2299 uio.uio_segflg = UIO_SYSSPACE;
2300 uio.uio_rw = UIO_READ;
2301 uio.uio_td = curthread;
2303 if (ncvp_debug >= 2)
2304 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
2305 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
2307 den = (struct dirent *)rbuf;
2308 bytes = blksize - uio.uio_resid;
2311 if (ncvp_debug >= 2) {
2312 kprintf("cache_inefficient_scan: %*.*s\n",
2313 den->d_namlen, den->d_namlen,
2316 if (den->d_type != DT_WHT &&
2317 den->d_ino == vat.va_fileid) {
2319 kprintf("cache_inefficient_scan: "
2320 "MATCHED inode %lld path %s/%*.*s\n",
2321 (long long)vat.va_fileid,
2323 den->d_namlen, den->d_namlen,
2326 nlc.nlc_nameptr = den->d_name;
2327 nlc.nlc_namelen = den->d_namlen;
2328 rncp = cache_nlookup(nch, &nlc);
2329 KKASSERT(rncp.ncp != NULL);
2332 bytes -= _DIRENT_DIRSIZ(den);
2333 den = _DIRENT_NEXT(den);
2335 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
2338 kfree(rbuf, M_TEMP);
2342 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
2343 _cache_setvp(rncp.mount, rncp.ncp, dvp);
2344 if (ncvp_debug >= 2) {
2345 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
2346 nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
2349 if (ncvp_debug >= 2) {
2350 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
2351 nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
2355 if (rncp.ncp->nc_vp == NULL)
2356 error = rncp.ncp->nc_error;
2358 * Release rncp after a successful nlookup. rncp was fully
2363 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
2364 dvp, nch->ncp->nc_name);
2371 * Zap a namecache entry. The ncp is unconditionally set to an unresolved
2372 * state, which disassociates it from its vnode or ncneglist.
2374 * Then, if there are no additional references to the ncp and no children,
2375 * the ncp is removed from the topology and destroyed.
2377 * References and/or children may exist if the ncp is in the middle of the
2378 * topology, preventing the ncp from being destroyed.
2380 * This function must be called with the ncp held and locked and will unlock
2381 * and drop it during zapping.
2383 * If nonblock is non-zero and the parent ncp cannot be locked we give up.
2384 * This case can occur in the cache_drop() path.
2386 * This function may returned a held (but NOT locked) parent node which the
2387 * caller must drop. We do this so _cache_drop() can loop, to avoid
2388 * blowing out the kernel stack.
2390 * WARNING! For MPSAFE operation this routine must acquire up to three
2391 * spin locks to be able to safely test nc_refs. Lock order is
2394 * hash spinlock if on hash list
2395 * parent spinlock if child of parent
2396 * (the ncp is unresolved so there is no vnode association)
2398 static struct namecache *
2399 cache_zap(struct namecache *ncp, int nonblock)
2401 struct namecache *par;
2402 struct vnode *dropvp;
2406 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
2408 _cache_setunresolved(ncp);
2411 * Try to scrap the entry and possibly tail-recurse on its parent.
2412 * We only scrap unref'd (other then our ref) unresolved entries,
2413 * we do not scrap 'live' entries.
2415 * Note that once the spinlocks are acquired if nc_refs == 1 no
2416 * other references are possible. If it isn't, however, we have
2417 * to decrement but also be sure to avoid a 1->0 transition.
2419 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
2420 KKASSERT(ncp->nc_refs > 0);
2423 * Acquire locks. Note that the parent can't go away while we hold
2426 if ((par = ncp->nc_parent) != NULL) {
2429 if (_cache_lock_nonblock(par) == 0)
2431 refs = ncp->nc_refs;
2432 ncp->nc_flag |= NCF_DEFEREDZAP;
2433 ++numdefered; /* MP race ok */
2434 if (atomic_cmpset_int(&ncp->nc_refs,
2446 spin_lock(&ncp->nc_head->spin);
2450 * If someone other then us has a ref or we have children
2451 * we cannot zap the entry. The 1->0 transition and any
2452 * further list operation is protected by the spinlocks
2453 * we have acquired but other transitions are not.
2456 refs = ncp->nc_refs;
2457 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list))
2459 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) {
2461 spin_unlock(&ncp->nc_head->spin);
2471 * We are the only ref and with the spinlocks held no further
2472 * refs can be acquired by others.
2474 * Remove us from the hash list and parent list. We have to
2475 * drop a ref on the parent's vp if the parent's list becomes
2480 struct nchash_head *nchpp = ncp->nc_head;
2482 KKASSERT(nchpp != NULL);
2483 LIST_REMOVE(ncp, nc_hash);
2484 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
2485 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
2486 dropvp = par->nc_vp;
2487 ncp->nc_head = NULL;
2488 ncp->nc_parent = NULL;
2489 spin_unlock(&nchpp->spin);
2492 KKASSERT(ncp->nc_head == NULL);
2496 * ncp should not have picked up any refs. Physically
2499 KKASSERT(ncp->nc_refs == 1);
2500 /* _cache_unlock(ncp) not required */
2501 ncp->nc_refs = -1; /* safety */
2503 kfree(ncp->nc_name, M_VFSCACHE);
2504 kfree(ncp, M_VFSCACHE);
2507 * Delayed drop (we had to release our spinlocks)
2509 * The refed parent (if not NULL) must be dropped. The
2510 * caller is responsible for looping.
2518 * Clean up dangling negative cache and defered-drop entries in the
2521 * This routine is called in the critical path and also called from
2522 * vnlru(). When called from vnlru we use a lower limit to try to
2523 * deal with the negative cache before the critical path has to start
2526 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t;
2528 static cache_hs_t neg_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW };
2529 static cache_hs_t pos_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW };
2532 cache_hysteresis(int critpath)
2535 int neglimit = desiredvnodes / ncnegfactor;
2536 int xnumcache = numcache;
2539 neglimit = neglimit * 8 / 10;
2542 * Don't cache too many negative hits. We use hysteresis to reduce
2543 * the impact on the critical path.
2545 switch(neg_cache_hysteresis_state[critpath]) {
2547 if (numneg > MINNEG && numneg > neglimit) {
2549 _cache_cleanneg(ncnegflush);
2551 _cache_cleanneg(ncnegflush +
2553 neg_cache_hysteresis_state[critpath] = CHI_HIGH;
2557 if (numneg > MINNEG * 9 / 10 &&
2558 numneg * 9 / 10 > neglimit
2561 _cache_cleanneg(ncnegflush);
2563 _cache_cleanneg(ncnegflush +
2564 numneg * 9 / 10 - neglimit);
2566 neg_cache_hysteresis_state[critpath] = CHI_LOW;
2572 * Don't cache too many positive hits. We use hysteresis to reduce
2573 * the impact on the critical path.
2575 * Excessive positive hits can accumulate due to large numbers of
2576 * hardlinks (the vnode cache will not prevent hl ncps from growing
2579 if ((poslimit = ncposlimit) == 0)
2580 poslimit = desiredvnodes * 2;
2582 poslimit = poslimit * 8 / 10;
2584 switch(pos_cache_hysteresis_state[critpath]) {
2586 if (xnumcache > poslimit && xnumcache > MINPOS) {
2588 _cache_cleanpos(ncposflush);
2590 _cache_cleanpos(ncposflush +
2591 xnumcache - poslimit);
2592 pos_cache_hysteresis_state[critpath] = CHI_HIGH;
2596 if (xnumcache > poslimit * 5 / 6 && xnumcache > MINPOS) {
2598 _cache_cleanpos(ncposflush);
2600 _cache_cleanpos(ncposflush +
2601 xnumcache - poslimit * 5 / 6);
2603 pos_cache_hysteresis_state[critpath] = CHI_LOW;
2609 * Clean out dangling defered-zap ncps which could not
2610 * be cleanly dropped if too many build up. Note
2611 * that numdefered is not an exact number as such ncps
2612 * can be reused and the counter is not handled in a MP
2613 * safe manner by design.
2615 if (numdefered > neglimit) {
2616 _cache_cleandefered();
2621 * NEW NAMECACHE LOOKUP API
2623 * Lookup an entry in the namecache. The passed par_nch must be referenced
2624 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp
2625 * is ALWAYS returned, eve if the supplied component is illegal.
2627 * The resulting namecache entry should be returned to the system with
2628 * cache_put() or cache_unlock() + cache_drop().
2630 * namecache locks are recursive but care must be taken to avoid lock order
2631 * reversals (hence why the passed par_nch must be unlocked). Locking
2632 * rules are to order for parent traversals, not for child traversals.
2634 * Nobody else will be able to manipulate the associated namespace (e.g.
2635 * create, delete, rename, rename-target) until the caller unlocks the
2638 * The returned entry will be in one of three states: positive hit (non-null
2639 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
2640 * Unresolved entries must be resolved through the filesystem to associate the
2641 * vnode and/or determine whether a positive or negative hit has occured.
2643 * It is not necessary to lock a directory in order to lock namespace under
2644 * that directory. In fact, it is explicitly not allowed to do that. A
2645 * directory is typically only locked when being created, renamed, or
2648 * The directory (par) may be unresolved, in which case any returned child
2649 * will likely also be marked unresolved. Likely but not guarenteed. Since
2650 * the filesystem lookup requires a resolved directory vnode the caller is
2651 * responsible for resolving the namecache chain top-down. This API
2652 * specifically allows whole chains to be created in an unresolved state.
2655 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
2657 struct nchandle nch;
2658 struct namecache *ncp;
2659 struct namecache *new_ncp;
2660 struct nchash_head *nchpp;
2668 mp = par_nch->mount;
2672 * This is a good time to call it, no ncp's are locked by
2675 cache_hysteresis(1);
2678 * Try to locate an existing entry
2680 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2681 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2683 nchpp = NCHHASH(hash);
2685 spin_lock(&nchpp->spin);
2686 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2690 * Break out if we find a matching entry. Note that
2691 * UNRESOLVED entries may match, but DESTROYED entries
2694 if (ncp->nc_parent == par_nch->ncp &&
2695 ncp->nc_nlen == nlc->nlc_namelen &&
2696 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2697 (ncp->nc_flag & NCF_DESTROYED) == 0
2700 spin_unlock(&nchpp->spin);
2702 _cache_unlock(par_nch->ncp);
2705 if (_cache_lock_special(ncp) == 0) {
2706 _cache_auto_unresolve(mp, ncp);
2708 _cache_free(new_ncp);
2719 * We failed to locate an entry, create a new entry and add it to
2720 * the cache. The parent ncp must also be locked so we
2723 * We have to relookup after possibly blocking in kmalloc or
2724 * when locking par_nch.
2726 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2727 * mount case, in which case nc_name will be NULL.
2729 if (new_ncp == NULL) {
2730 spin_unlock(&nchpp->spin);
2731 new_ncp = cache_alloc(nlc->nlc_namelen);
2732 if (nlc->nlc_namelen) {
2733 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2735 new_ncp->nc_name[nlc->nlc_namelen] = 0;
2739 if (par_locked == 0) {
2740 spin_unlock(&nchpp->spin);
2741 _cache_lock(par_nch->ncp);
2747 * WARNING! We still hold the spinlock. We have to set the hash
2748 * table entry atomically.
2751 _cache_link_parent(ncp, par_nch->ncp, nchpp);
2752 spin_unlock(&nchpp->spin);
2753 _cache_unlock(par_nch->ncp);
2754 /* par_locked = 0 - not used */
2757 * stats and namecache size management
2759 if (ncp->nc_flag & NCF_UNRESOLVED)
2760 ++gd->gd_nchstats->ncs_miss;
2761 else if (ncp->nc_vp)
2762 ++gd->gd_nchstats->ncs_goodhits;
2764 ++gd->gd_nchstats->ncs_neghits;
2767 atomic_add_int(&nch.mount->mnt_refs, 1);
2772 * Attempt to lookup a namecache entry and return with a shared namecache
2776 cache_nlookup_maybe_shared(struct nchandle *par_nch, struct nlcomponent *nlc,
2777 int excl, struct nchandle *res_nch)
2779 struct namecache *ncp;
2780 struct nchash_head *nchpp;
2786 * If exclusive requested or shared namecache locks are disabled,
2789 if (ncp_shared_lock_disable || excl)
2790 return(EWOULDBLOCK);
2794 mp = par_nch->mount;
2797 * This is a good time to call it, no ncp's are locked by
2800 cache_hysteresis(1);
2803 * Try to locate an existing entry
2805 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2806 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2807 nchpp = NCHHASH(hash);
2809 spin_lock(&nchpp->spin);
2811 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2815 * Break out if we find a matching entry. Note that
2816 * UNRESOLVED entries may match, but DESTROYED entries
2819 if (ncp->nc_parent == par_nch->ncp &&
2820 ncp->nc_nlen == nlc->nlc_namelen &&
2821 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2822 (ncp->nc_flag & NCF_DESTROYED) == 0
2825 spin_unlock(&nchpp->spin);
2826 if (_cache_lock_shared_special(ncp) == 0) {
2827 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
2828 (ncp->nc_flag & NCF_DESTROYED) == 0 &&
2829 _cache_auto_unresolve_test(mp, ncp) == 0) {
2835 spin_lock(&nchpp->spin);
2843 spin_unlock(&nchpp->spin);
2844 return(EWOULDBLOCK);
2849 * Note that nc_error might be non-zero (e.g ENOENT).
2852 res_nch->mount = mp;
2854 ++gd->gd_nchstats->ncs_goodhits;
2855 atomic_add_int(&res_nch->mount->mnt_refs, 1);
2857 KKASSERT(ncp->nc_error != EWOULDBLOCK);
2858 return(ncp->nc_error);
2862 * This is a non-blocking verison of cache_nlookup() used by
2863 * nfs_readdirplusrpc_uio(). It can fail for any reason and
2864 * will return nch.ncp == NULL in that case.
2867 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc)
2869 struct nchandle nch;
2870 struct namecache *ncp;
2871 struct namecache *new_ncp;
2872 struct nchash_head *nchpp;
2880 mp = par_nch->mount;
2884 * Try to locate an existing entry
2886 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2887 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2889 nchpp = NCHHASH(hash);
2891 spin_lock(&nchpp->spin);
2892 LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2896 * Break out if we find a matching entry. Note that
2897 * UNRESOLVED entries may match, but DESTROYED entries
2900 if (ncp->nc_parent == par_nch->ncp &&
2901 ncp->nc_nlen == nlc->nlc_namelen &&
2902 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2903 (ncp->nc_flag & NCF_DESTROYED) == 0
2906 spin_unlock(&nchpp->spin);
2908 _cache_unlock(par_nch->ncp);
2911 if (_cache_lock_special(ncp) == 0) {
2912 _cache_auto_unresolve(mp, ncp);
2914 _cache_free(new_ncp);
2925 * We failed to locate an entry, create a new entry and add it to
2926 * the cache. The parent ncp must also be locked so we
2929 * We have to relookup after possibly blocking in kmalloc or
2930 * when locking par_nch.
2932 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2933 * mount case, in which case nc_name will be NULL.
2935 if (new_ncp == NULL) {
2936 spin_unlock(&nchpp->spin);
2937 new_ncp = cache_alloc(nlc->nlc_namelen);
2938 if (nlc->nlc_namelen) {
2939 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2941 new_ncp->nc_name[nlc->nlc_namelen] = 0;
2945 if (par_locked == 0) {
2946 spin_unlock(&nchpp->spin);
2947 if (_cache_lock_nonblock(par_nch->ncp) == 0) {
2955 * WARNING! We still hold the spinlock. We have to set the hash
2956 * table entry atomically.
2959 _cache_link_parent(ncp, par_nch->ncp, nchpp);
2960 spin_unlock(&nchpp->spin);
2961 _cache_unlock(par_nch->ncp);
2962 /* par_locked = 0 - not used */
2965 * stats and namecache size management
2967 if (ncp->nc_flag & NCF_UNRESOLVED)
2968 ++gd->gd_nchstats->ncs_miss;
2969 else if (ncp->nc_vp)
2970 ++gd->gd_nchstats->ncs_goodhits;
2972 ++gd->gd_nchstats->ncs_neghits;
2975 atomic_add_int(&nch.mount->mnt_refs, 1);
2979 _cache_free(new_ncp);
2988 * The namecache entry is marked as being used as a mount point.
2989 * Locate the mount if it is visible to the caller. The DragonFly
2990 * mount system allows arbitrary loops in the topology and disentangles
2991 * those loops by matching against (mp, ncp) rather than just (ncp).
2992 * This means any given ncp can dive any number of mounts, depending
2993 * on the relative mount (e.g. nullfs) the caller is at in the topology.
2995 * We use a very simple frontend cache to reduce SMP conflicts,
2996 * which we have to do because the mountlist scan needs an exclusive
2997 * lock around its ripout info list. Not to mention that there might
2998 * be a lot of mounts.
3000 struct findmount_info {
3001 struct mount *result;
3002 struct mount *nch_mount;
3003 struct namecache *nch_ncp;
3007 struct ncmount_cache *
3008 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp)
3012 hash = ((int)(intptr_t)mp / sizeof(*mp)) ^
3013 ((int)(intptr_t)ncp / sizeof(*ncp));
3014 hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE;
3015 return (&ncmount_cache[hash]);
3020 cache_findmount_callback(struct mount *mp, void *data)
3022 struct findmount_info *info = data;
3025 * Check the mount's mounted-on point against the passed nch.
3027 if (mp->mnt_ncmounton.mount == info->nch_mount &&
3028 mp->mnt_ncmounton.ncp == info->nch_ncp
3031 atomic_add_int(&mp->mnt_refs, 1);
3038 cache_findmount(struct nchandle *nch)
3040 struct findmount_info info;
3041 struct ncmount_cache *ncc;
3047 if (ncmount_cache_enable == 0) {
3051 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3052 if (ncc->ncp == nch->ncp) {
3053 spin_lock_shared(&ncc->spin);
3054 if (ncc->isneg == 0 &&
3055 ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) {
3056 if (mp->mnt_ncmounton.mount == nch->mount &&
3057 mp->mnt_ncmounton.ncp == nch->ncp) {
3059 * Cache hit (positive)
3061 atomic_add_int(&mp->mnt_refs, 1);
3062 spin_unlock_shared(&ncc->spin);
3063 ++ncmount_cache_hit;
3066 /* else cache miss */
3069 ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3071 * Cache hit (negative)
3073 spin_unlock_shared(&ncc->spin);
3074 ++ncmount_cache_hit;
3077 spin_unlock_shared(&ncc->spin);
3085 info.nch_mount = nch->mount;
3086 info.nch_ncp = nch->ncp;
3087 mountlist_scan(cache_findmount_callback, &info,
3088 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
3093 * Negative lookups: We cache the originating {ncp,mp}. (mp) is
3094 * only used for pointer comparisons and is not
3095 * referenced (otherwise there would be dangling
3098 * Positive lookups: We cache the originating {ncp} and the target
3099 * (mp). (mp) is referenced.
3101 * Indeterminant: If the match is undergoing an unmount we do
3102 * not cache it to avoid racing cache_unmounting(),
3103 * but still return the match.
3106 spin_lock(&ncc->spin);
3107 if (info.result == NULL) {
3108 if (ncc->isneg == 0 && ncc->mp)
3109 atomic_add_int(&ncc->mp->mnt_refs, -1);
3110 ncc->ncp = nch->ncp;
3111 ncc->mp = nch->mount;
3113 spin_unlock(&ncc->spin);
3114 ++ncmount_cache_overwrite;
3115 } else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) {
3116 if (ncc->isneg == 0 && ncc->mp)
3117 atomic_add_int(&ncc->mp->mnt_refs, -1);
3118 atomic_add_int(&info.result->mnt_refs, 1);
3119 ncc->ncp = nch->ncp;
3120 ncc->mp = info.result;
3122 spin_unlock(&ncc->spin);
3123 ++ncmount_cache_overwrite;
3125 spin_unlock(&ncc->spin);
3127 ++ncmount_cache_miss;
3129 return(info.result);
3133 cache_dropmount(struct mount *mp)
3135 atomic_add_int(&mp->mnt_refs, -1);
3139 cache_ismounting(struct mount *mp)
3141 struct nchandle *nch = &mp->mnt_ncmounton;
3142 struct ncmount_cache *ncc;
3144 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3146 ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3147 spin_lock(&ncc->spin);
3149 ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3153 spin_unlock(&ncc->spin);
3158 cache_unmounting(struct mount *mp)
3160 struct nchandle *nch = &mp->mnt_ncmounton;
3161 struct ncmount_cache *ncc;
3163 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3164 if (ncc->isneg == 0 &&
3165 ncc->ncp == nch->ncp && ncc->mp == mp) {
3166 spin_lock(&ncc->spin);
3167 if (ncc->isneg == 0 &&
3168 ncc->ncp == nch->ncp && ncc->mp == mp) {
3169 atomic_add_int(&mp->mnt_refs, -1);
3173 spin_unlock(&ncc->spin);
3178 * Resolve an unresolved namecache entry, generally by looking it up.
3179 * The passed ncp must be locked and refd.
3181 * Theoretically since a vnode cannot be recycled while held, and since
3182 * the nc_parent chain holds its vnode as long as children exist, the
3183 * direct parent of the cache entry we are trying to resolve should
3184 * have a valid vnode. If not then generate an error that we can
3185 * determine is related to a resolver bug.
3187 * However, if a vnode was in the middle of a recyclement when the NCP
3188 * got locked, ncp->nc_vp might point to a vnode that is about to become
3189 * invalid. cache_resolve() handles this case by unresolving the entry
3190 * and then re-resolving it.
3192 * Note that successful resolution does not necessarily return an error
3193 * code of 0. If the ncp resolves to a negative cache hit then ENOENT
3197 cache_resolve(struct nchandle *nch, struct ucred *cred)
3199 struct namecache *par_tmp;
3200 struct namecache *par;
3201 struct namecache *ncp;
3202 struct nchandle nctmp;
3209 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
3212 * If the ncp is already resolved we have nothing to do. However,
3213 * we do want to guarentee that a usable vnode is returned when
3214 * a vnode is present, so make sure it hasn't been reclaimed.
3216 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3217 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
3218 _cache_setunresolved(ncp);
3219 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
3220 return (ncp->nc_error);
3224 * If the ncp was destroyed it will never resolve again. This
3225 * can basically only happen when someone is chdir'd into an
3226 * empty directory which is then rmdir'd. We want to catch this
3227 * here and not dive the VFS because the VFS might actually
3228 * have a way to re-resolve the disconnected ncp, which will
3229 * result in inconsistencies in the cdir/nch for proc->p_fd.
3231 if (ncp->nc_flag & NCF_DESTROYED) {
3232 kprintf("Warning: cache_resolve: ncp '%s' was unlinked\n",
3238 * Mount points need special handling because the parent does not
3239 * belong to the same filesystem as the ncp.
3241 if (ncp == mp->mnt_ncmountpt.ncp)
3242 return (cache_resolve_mp(mp));
3245 * We expect an unbroken chain of ncps to at least the mount point,
3246 * and even all the way to root (but this code doesn't have to go
3247 * past the mount point).
3249 if (ncp->nc_parent == NULL) {
3250 kprintf("EXDEV case 1 %p %*.*s\n", ncp,
3251 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
3252 ncp->nc_error = EXDEV;
3253 return(ncp->nc_error);
3257 * The vp's of the parent directories in the chain are held via vhold()
3258 * due to the existance of the child, and should not disappear.
3259 * However, there are cases where they can disappear:
3261 * - due to filesystem I/O errors.
3262 * - due to NFS being stupid about tracking the namespace and
3263 * destroys the namespace for entire directories quite often.
3264 * - due to forced unmounts.
3265 * - due to an rmdir (parent will be marked DESTROYED)
3267 * When this occurs we have to track the chain backwards and resolve
3268 * it, looping until the resolver catches up to the current node. We
3269 * could recurse here but we might run ourselves out of kernel stack
3270 * so we do it in a more painful manner. This situation really should
3271 * not occur all that often, or if it does not have to go back too
3272 * many nodes to resolve the ncp.
3274 while ((dvp = cache_dvpref(ncp)) == NULL) {
3276 * This case can occur if a process is CD'd into a
3277 * directory which is then rmdir'd. If the parent is marked
3278 * destroyed there is no point trying to resolve it.
3280 if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
3282 par = ncp->nc_parent;
3285 while ((par_tmp = par->nc_parent) != NULL &&
3286 par_tmp->nc_vp == NULL) {
3287 _cache_hold(par_tmp);
3288 _cache_lock(par_tmp);
3292 if (par->nc_parent == NULL) {
3293 kprintf("EXDEV case 2 %*.*s\n",
3294 par->nc_nlen, par->nc_nlen, par->nc_name);
3298 kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
3299 par->nc_nlen, par->nc_nlen, par->nc_name);
3301 * The parent is not set in stone, ref and lock it to prevent
3302 * it from disappearing. Also note that due to renames it
3303 * is possible for our ncp to move and for par to no longer
3304 * be one of its parents. We resolve it anyway, the loop
3305 * will handle any moves.
3307 _cache_get(par); /* additional hold/lock */
3308 _cache_put(par); /* from earlier hold/lock */
3309 if (par == nch->mount->mnt_ncmountpt.ncp) {
3310 cache_resolve_mp(nch->mount);
3311 } else if ((dvp = cache_dvpref(par)) == NULL) {
3312 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
3316 if (par->nc_flag & NCF_UNRESOLVED) {
3319 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
3323 if ((error = par->nc_error) != 0) {
3324 if (par->nc_error != EAGAIN) {
3325 kprintf("EXDEV case 3 %*.*s error %d\n",
3326 par->nc_nlen, par->nc_nlen, par->nc_name,
3331 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
3332 par, par->nc_nlen, par->nc_nlen, par->nc_name);
3339 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
3340 * ncp's and reattach them. If this occurs the original ncp is marked
3341 * EAGAIN to force a relookup.
3343 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
3344 * ncp must already be resolved.
3349 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
3352 ncp->nc_error = EPERM;
3354 if (ncp->nc_error == EAGAIN) {
3355 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
3356 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
3359 return(ncp->nc_error);
3363 * Resolve the ncp associated with a mount point. Such ncp's almost always
3364 * remain resolved and this routine is rarely called. NFS MPs tends to force
3365 * re-resolution more often due to its mac-truck-smash-the-namecache
3366 * method of tracking namespace changes.
3368 * The semantics for this call is that the passed ncp must be locked on
3369 * entry and will be locked on return. However, if we actually have to
3370 * resolve the mount point we temporarily unlock the entry in order to
3371 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of
3372 * the unlock we have to recheck the flags after we relock.
3375 cache_resolve_mp(struct mount *mp)
3377 struct namecache *ncp = mp->mnt_ncmountpt.ncp;
3381 KKASSERT(mp != NULL);
3384 * If the ncp is already resolved we have nothing to do. However,
3385 * we do want to guarentee that a usable vnode is returned when
3386 * a vnode is present, so make sure it hasn't been reclaimed.
3388 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3389 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
3390 _cache_setunresolved(ncp);
3393 if (ncp->nc_flag & NCF_UNRESOLVED) {
3395 while (vfs_busy(mp, 0))
3397 error = VFS_ROOT(mp, &vp);
3401 * recheck the ncp state after relocking.
3403 if (ncp->nc_flag & NCF_UNRESOLVED) {
3404 ncp->nc_error = error;
3406 _cache_setvp(mp, ncp, vp);
3409 kprintf("[diagnostic] cache_resolve_mp: failed"
3410 " to resolve mount %p err=%d ncp=%p\n",
3412 _cache_setvp(mp, ncp, NULL);
3414 } else if (error == 0) {
3419 return(ncp->nc_error);
3423 * Clean out negative cache entries when too many have accumulated.
3426 _cache_cleanneg(int count)
3428 struct namecache *ncp;
3431 * Attempt to clean out the specified number of negative cache
3436 ncp = TAILQ_FIRST(&ncneglist);
3438 spin_unlock(&ncspin);
3441 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
3442 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
3444 spin_unlock(&ncspin);
3447 * This can race, so we must re-check that the ncp
3448 * is on the ncneglist after successfully locking it.
3450 if (_cache_lock_special(ncp) == 0) {
3451 if (ncp->nc_vp == NULL &&
3452 (ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3453 ncp = cache_zap(ncp, 1);
3457 kprintf("cache_cleanneg: race avoided\n");
3468 * Clean out positive cache entries when too many have accumulated.
3471 _cache_cleanpos(int count)
3473 static volatile int rover;
3474 struct nchash_head *nchpp;
3475 struct namecache *ncp;
3479 * Attempt to clean out the specified number of negative cache
3483 rover_copy = ++rover; /* MPSAFEENOUGH */
3485 nchpp = NCHHASH(rover_copy);
3487 spin_lock(&nchpp->spin);
3488 ncp = LIST_FIRST(&nchpp->list);
3489 while (ncp && (ncp->nc_flag & NCF_DESTROYED))
3490 ncp = LIST_NEXT(ncp, nc_hash);
3493 spin_unlock(&nchpp->spin);
3496 if (_cache_lock_special(ncp) == 0) {
3497 ncp = cache_zap(ncp, 1);
3509 * This is a kitchen sink function to clean out ncps which we
3510 * tried to zap from cache_drop() but failed because we were
3511 * unable to acquire the parent lock.
3513 * Such entries can also be removed via cache_inval_vp(), such
3514 * as when unmounting.
3517 _cache_cleandefered(void)
3519 struct nchash_head *nchpp;
3520 struct namecache *ncp;
3521 struct namecache dummy;
3525 bzero(&dummy, sizeof(dummy));
3526 dummy.nc_flag = NCF_DESTROYED;
3529 for (i = 0; i <= nchash; ++i) {
3530 nchpp = &nchashtbl[i];
3532 spin_lock(&nchpp->spin);
3533 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash);
3535 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) {
3536 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0)
3538 LIST_REMOVE(&dummy, nc_hash);
3539 LIST_INSERT_AFTER(ncp, &dummy, nc_hash);
3541 spin_unlock(&nchpp->spin);
3542 if (_cache_lock_nonblock(ncp) == 0) {
3543 ncp->nc_flag &= ~NCF_DEFEREDZAP;
3547 spin_lock(&nchpp->spin);
3550 LIST_REMOVE(&dummy, nc_hash);
3551 spin_unlock(&nchpp->spin);
3556 * Name cache initialization, from vfsinit() when we are booting
3564 /* initialise per-cpu namecache effectiveness statistics. */
3565 for (i = 0; i < ncpus; ++i) {
3566 gd = globaldata_find(i);
3567 gd->gd_nchstats = &nchstats[i];
3569 TAILQ_INIT(&ncneglist);
3571 nchashtbl = hashinit_ext(desiredvnodes / 2,
3572 sizeof(struct nchash_head),
3573 M_VFSCACHE, &nchash);
3574 for (i = 0; i <= (int)nchash; ++i) {
3575 LIST_INIT(&nchashtbl[i].list);
3576 spin_init(&nchashtbl[i].spin);
3578 for (i = 0; i < NCMOUNT_NUMCACHE; ++i)
3579 spin_init(&ncmount_cache[i].spin);
3580 nclockwarn = 5 * hz;
3584 * Called from start_init() to bootstrap the root filesystem. Returns
3585 * a referenced, unlocked namecache record.
3588 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
3590 nch->ncp = cache_alloc(0);
3592 atomic_add_int(&mp->mnt_refs, 1);
3594 _cache_setvp(nch->mount, nch->ncp, vp);
3598 * vfs_cache_setroot()
3600 * Create an association between the root of our namecache and
3601 * the root vnode. This routine may be called several times during
3604 * If the caller intends to save the returned namecache pointer somewhere
3605 * it must cache_hold() it.
3608 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
3611 struct nchandle onch;
3619 cache_zero(&rootnch);
3627 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache
3628 * topology and is being removed as quickly as possible. The new VOP_N*()
3629 * API calls are required to make specific adjustments using the supplied
3630 * ncp pointers rather then just bogusly purging random vnodes.
3632 * Invalidate all namecache entries to a particular vnode as well as
3633 * any direct children of that vnode in the namecache. This is a
3634 * 'catch all' purge used by filesystems that do not know any better.
3636 * Note that the linkage between the vnode and its namecache entries will
3637 * be removed, but the namecache entries themselves might stay put due to
3638 * active references from elsewhere in the system or due to the existance of
3639 * the children. The namecache topology is left intact even if we do not
3640 * know what the vnode association is. Such entries will be marked
3644 cache_purge(struct vnode *vp)
3646 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
3650 * Flush all entries referencing a particular filesystem.
3652 * Since we need to check it anyway, we will flush all the invalid
3653 * entries at the same time.
3658 cache_purgevfs(struct mount *mp)
3660 struct nchash_head *nchpp;
3661 struct namecache *ncp, *nnp;
3664 * Scan hash tables for applicable entries.
3666 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
3667 spin_lock_wr(&nchpp->spin); XXX
3668 ncp = LIST_FIRST(&nchpp->list);
3672 nnp = LIST_NEXT(ncp, nc_hash);
3675 if (ncp->nc_mount == mp) {
3677 ncp = cache_zap(ncp, 0);
3685 spin_unlock_wr(&nchpp->spin); XXX
3691 static int disablecwd;
3692 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
3695 static u_long numcwdcalls;
3696 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0,
3697 "Number of current directory resolution calls");
3698 static u_long numcwdfailnf;
3699 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0,
3700 "Number of current directory failures due to lack of file");
3701 static u_long numcwdfailsz;
3702 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0,
3703 "Number of current directory failures due to large result");
3704 static u_long numcwdfound;
3705 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0,
3706 "Number of current directory resolution successes");
3712 sys___getcwd(struct __getcwd_args *uap)
3722 buflen = uap->buflen;
3725 if (buflen > MAXPATHLEN)
3726 buflen = MAXPATHLEN;
3728 buf = kmalloc(buflen, M_TEMP, M_WAITOK);
3729 bp = kern_getcwd(buf, buflen, &error);
3731 error = copyout(bp, uap->buf, strlen(bp) + 1);
3737 kern_getcwd(char *buf, size_t buflen, int *error)
3739 struct proc *p = curproc;
3741 int i, slash_prefixed;
3742 struct filedesc *fdp;
3743 struct nchandle nch;
3744 struct namecache *ncp;
3753 nch = fdp->fd_ncdir;
3758 while (ncp && (ncp != fdp->fd_nrdir.ncp ||
3759 nch.mount != fdp->fd_nrdir.mount)
3762 * While traversing upwards if we encounter the root
3763 * of the current mount we have to skip to the mount point
3764 * in the underlying filesystem.
3766 if (ncp == nch.mount->mnt_ncmountpt.ncp) {
3767 nch = nch.mount->mnt_ncmounton;
3776 * Prepend the path segment
3778 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3785 *--bp = ncp->nc_name[i];
3797 * Go up a directory. This isn't a mount point so we don't
3798 * have to check again.
3800 while ((nch.ncp = ncp->nc_parent) != NULL) {
3801 if (ncp_shared_lock_disable)
3804 _cache_lock_shared(ncp);
3805 if (nch.ncp != ncp->nc_parent) {
3809 _cache_hold(nch.ncp);
3822 if (!slash_prefixed) {
3840 * Thus begins the fullpath magic.
3842 * The passed nchp is referenced but not locked.
3844 static int disablefullpath;
3845 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
3846 &disablefullpath, 0,
3847 "Disable fullpath lookups");
3849 static u_int numfullpathcalls;
3850 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD,
3851 &numfullpathcalls, 0,
3852 "Number of full path resolutions in progress");
3853 static u_int numfullpathfailnf;
3854 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD,
3855 &numfullpathfailnf, 0,
3856 "Number of full path resolution failures due to lack of file");
3857 static u_int numfullpathfailsz;
3858 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD,
3859 &numfullpathfailsz, 0,
3860 "Number of full path resolution failures due to insufficient memory");
3861 static u_int numfullpathfound;
3862 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD,
3863 &numfullpathfound, 0,
3864 "Number of full path resolution successes");
3867 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase,
3868 char **retbuf, char **freebuf, int guess)
3870 struct nchandle fd_nrdir;
3871 struct nchandle nch;
3872 struct namecache *ncp;
3873 struct mount *mp, *new_mp;
3879 atomic_add_int(&numfullpathcalls, -1);
3884 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
3885 bp = buf + MAXPATHLEN - 1;
3888 fd_nrdir = *nchbase;
3890 fd_nrdir = p->p_fd->fd_nrdir;
3900 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) {
3904 * If we are asked to guess the upwards path, we do so whenever
3905 * we encounter an ncp marked as a mountpoint. We try to find
3906 * the actual mountpoint by finding the mountpoint with this
3909 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) {
3910 new_mp = mount_get_by_nc(ncp);
3913 * While traversing upwards if we encounter the root
3914 * of the current mount we have to skip to the mount point.
3916 if (ncp == mp->mnt_ncmountpt.ncp) {
3920 nch = new_mp->mnt_ncmounton;
3930 * Prepend the path segment
3932 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3934 numfullpathfailsz++;
3939 *--bp = ncp->nc_name[i];
3942 numfullpathfailsz++;
3951 * Go up a directory. This isn't a mount point so we don't
3952 * have to check again.
3954 * We can only safely access nc_parent with ncp held locked.
3956 while ((nch.ncp = ncp->nc_parent) != NULL) {
3958 if (nch.ncp != ncp->nc_parent) {
3962 _cache_hold(nch.ncp);
3970 numfullpathfailnf++;
3976 if (!slash_prefixed) {
3978 numfullpathfailsz++;
3996 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf,
3999 struct namecache *ncp;
4000 struct nchandle nch;
4004 atomic_add_int(&numfullpathcalls, 1);
4005 if (disablefullpath)
4011 /* vn is NULL, client wants us to use p->p_textvp */
4013 if ((vn = p->p_textvp) == NULL)
4016 spin_lock(&vn->v_spin);
4017 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
4022 spin_unlock(&vn->v_spin);
4026 spin_unlock(&vn->v_spin);
4028 atomic_add_int(&numfullpathcalls, -1);
4030 nch.mount = vn->v_mount;
4031 error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess);