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/spinlock2.h>
86 #define MAX_RECURSION_DEPTH 64
89 * Random lookups in the cache are accomplished with a hash table using
90 * a hash key of (nc_src_vp, name). Each hash chain has its own spin lock.
92 * Negative entries may exist and correspond to resolved namecache
93 * structures where nc_vp is NULL. In a negative entry, NCF_WHITEOUT
94 * will be set if the entry corresponds to a whited-out directory entry
95 * (verses simply not finding the entry at all). pcpu_ncache[n].neg_list
96 * is locked via pcpu_ncache[n].neg_spin;
100 * (1) A ncp must be referenced before it can be locked.
102 * (2) A ncp must be locked in order to modify it.
104 * (3) ncp locks are always ordered child -> parent. That may seem
105 * backwards but forward scans use the hash table and thus can hold
106 * the parent unlocked when traversing downward.
108 * This allows insert/rename/delete/dot-dot and other operations
109 * to use ncp->nc_parent links.
111 * This also prevents a locked up e.g. NFS node from creating a
112 * chain reaction all the way back to the root vnode / namecache.
114 * (4) parent linkages require both the parent and child to be locked.
118 * Structures associated with name cacheing.
120 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash])
123 #define NCMOUNT_NUMCACHE 16301 /* prime number */
125 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
127 TAILQ_HEAD(nchash_list, namecache);
130 * Don't cachealign, but at least pad to 32 bytes so entries
131 * don't cross a cache line.
134 struct nchash_list list; /* 16 bytes */
135 struct spinlock spin; /* 8 bytes */
136 long pad01; /* 8 bytes */
139 struct ncmount_cache {
140 struct spinlock spin;
141 struct namecache *ncp;
143 int isneg; /* if != 0 mp is originator and not target */
147 struct spinlock neg_spin; /* for neg_list and neg_count */
148 struct namecache_list neg_list;
155 static struct nchash_head *nchashtbl;
156 static struct pcpu_ncache *pcpu_ncache;
157 static struct ncmount_cache ncmount_cache[NCMOUNT_NUMCACHE];
160 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server
161 * to create the namecache infrastructure leading to a dangling vnode.
163 * 0 Only errors are reported
164 * 1 Successes are reported
165 * 2 Successes + the whole directory scan is reported
166 * 3 Force the directory scan code run as if the parent vnode did not
167 * have a namecache record, even if it does have one.
169 static int ncvp_debug;
170 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0,
171 "Namecache debug level (0-3)");
173 static u_long nchash; /* size of hash table */
174 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
175 "Size of namecache hash table");
177 static int ncnegflush = 10; /* burst for negative flush */
178 SYSCTL_INT(_debug, OID_AUTO, ncnegflush, CTLFLAG_RW, &ncnegflush, 0,
179 "Batch flush negative entries");
181 static int ncposflush = 10; /* burst for positive flush */
182 SYSCTL_INT(_debug, OID_AUTO, ncposflush, CTLFLAG_RW, &ncposflush, 0,
183 "Batch flush positive entries");
185 static int ncnegfactor = 16; /* ratio of negative entries */
186 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
187 "Ratio of namecache negative entries");
189 static int nclockwarn; /* warn on locked entries in ticks */
190 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0,
191 "Warn on locked namecache entries in ticks");
193 static int numdefered; /* number of cache entries allocated */
194 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0,
195 "Number of cache entries allocated");
197 static int ncposlimit; /* number of cache entries allocated */
198 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0,
199 "Number of cache entries allocated");
201 static int ncp_shared_lock_disable = 0;
202 SYSCTL_INT(_debug, OID_AUTO, ncp_shared_lock_disable, CTLFLAG_RW,
203 &ncp_shared_lock_disable, 0, "Disable shared namecache locks");
205 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode),
206 "sizeof(struct vnode)");
207 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache),
208 "sizeof(struct namecache)");
210 static int ncmount_cache_enable = 1;
211 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW,
212 &ncmount_cache_enable, 0, "mount point cache");
214 static __inline void _cache_drop(struct namecache *ncp);
215 static int cache_resolve_mp(struct mount *mp);
216 static struct vnode *cache_dvpref(struct namecache *ncp);
217 static void _cache_lock(struct namecache *ncp);
218 static void _cache_setunresolved(struct namecache *ncp);
219 static void _cache_cleanneg(long count);
220 static void _cache_cleanpos(long count);
221 static void _cache_cleandefered(void);
222 static void _cache_unlink(struct namecache *ncp);
224 static void vfscache_rollup_all(void);
228 * The new name cache statistics
230 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
231 static long vfscache_negs;
232 SYSCTL_LONG(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &vfscache_negs, 0,
233 "Number of negative namecache entries");
234 static long vfscache_count;
235 SYSCTL_LONG(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &vfscache_count, 0,
236 "Number of namecaches entries");
237 static long vfscache_leafs;
238 SYSCTL_LONG(_vfs_cache, OID_AUTO, numleafs, CTLFLAG_RD, &vfscache_leafs, 0,
239 "Number of namecaches entries");
241 struct nchstats nchstats[SMP_MAXCPU];
243 * Export VFS cache effectiveness statistics to user-land.
245 * The statistics are left for aggregation to user-land so
246 * neat things can be achieved, like observing per-CPU cache
250 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
252 struct globaldata *gd;
256 for (i = 0; i < ncpus; ++i) {
257 gd = globaldata_find(i);
258 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
259 sizeof(struct nchstats))))
265 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
266 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
268 static struct namecache *cache_zap(struct namecache *ncp, int nonblock);
271 * Cache mount points and namecache records in order to avoid unnecessary
272 * atomic ops on mnt_refs and ncp->refs. This improves concurrent SMP
273 * performance and is particularly important on multi-socket systems to
274 * reduce cache-line ping-ponging.
276 * Try to keep the pcpu structure within one cache line (~64 bytes).
278 #define MNTCACHE_COUNT 5
281 struct mount *mntary[MNTCACHE_COUNT];
282 struct namecache *ncp1;
283 struct namecache *ncp2;
284 struct nchandle ncdir;
289 static struct mntcache pcpu_mntcache[MAXCPU];
293 _cache_mntref(struct mount *mp)
295 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid];
298 for (i = 0; i < MNTCACHE_COUNT; ++i) {
299 if (cache->mntary[i] != mp)
301 if (atomic_cmpset_ptr((void *)&cache->mntary[i], mp, NULL))
304 atomic_add_int(&mp->mnt_refs, 1);
309 _cache_mntrel(struct mount *mp)
311 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid];
314 for (i = 0; i < MNTCACHE_COUNT; ++i) {
315 if (cache->mntary[i] == NULL) {
316 mp = atomic_swap_ptr((void *)&cache->mntary[i], mp);
321 i = (int)((uint32_t)++cache->iter % (uint32_t)MNTCACHE_COUNT);
322 mp = atomic_swap_ptr((void *)&cache->mntary[i], mp);
324 atomic_add_int(&mp->mnt_refs, -1);
328 * Clears all cached mount points on all cpus. This routine should only
329 * be called when we are waiting for a mount to clear, e.g. so we can
333 cache_clearmntcache(void)
337 for (n = 0; n < ncpus; ++n) {
338 struct mntcache *cache = &pcpu_mntcache[n];
339 struct namecache *ncp;
343 for (i = 0; i < MNTCACHE_COUNT; ++i) {
344 if (cache->mntary[i]) {
345 mp = atomic_swap_ptr(
346 (void *)&cache->mntary[i], NULL);
348 atomic_add_int(&mp->mnt_refs, -1);
352 ncp = atomic_swap_ptr((void *)&cache->ncp1, NULL);
357 ncp = atomic_swap_ptr((void *)&cache->ncp2, NULL);
361 if (cache->ncdir.ncp) {
362 ncp = atomic_swap_ptr((void *)&cache->ncdir.ncp, NULL);
366 if (cache->ncdir.mount) {
367 mp = atomic_swap_ptr((void *)&cache->ncdir.mount, NULL);
369 atomic_add_int(&mp->mnt_refs, -1);
376 * Namespace locking. The caller must already hold a reference to the
377 * namecache structure in order to lock/unlock it. This function prevents
378 * the namespace from being created or destroyed by accessors other then
381 * Note that holding a locked namecache structure prevents other threads
382 * from making namespace changes (e.g. deleting or creating), prevents
383 * vnode association state changes by other threads, and prevents the
384 * namecache entry from being resolved or unresolved by other threads.
386 * An exclusive lock owner has full authority to associate/disassociate
387 * vnodes and resolve/unresolve the locked ncp.
389 * A shared lock owner only has authority to acquire the underlying vnode,
392 * The primary lock field is nc_lockstatus. nc_locktd is set after the
393 * fact (when locking) or cleared prior to unlocking.
395 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed
396 * or recycled, but it does NOT help you if the vnode had already
397 * initiated a recyclement. If this is important, use cache_get()
398 * rather then cache_lock() (and deal with the differences in the
399 * way the refs counter is handled). Or, alternatively, make an
400 * unconditional call to cache_validate() or cache_resolve()
401 * after cache_lock() returns.
405 _cache_lock(struct namecache *ncp)
413 KKASSERT(ncp->nc_refs != 0);
419 count = ncp->nc_lockstatus;
422 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) {
423 if (atomic_cmpset_int(&ncp->nc_lockstatus,
426 * The vp associated with a locked ncp must
427 * be held to prevent it from being recycled.
429 * WARNING! If VRECLAIMED is set the vnode
430 * could already be in the middle of a recycle.
431 * Callers must use cache_vref() or
432 * cache_vget() on the locked ncp to
433 * validate the vp or set the cache entry
436 * NOTE! vhold() is allowed if we hold a
437 * lock on the ncp (which we do).
447 if (ncp->nc_locktd == td) {
448 KKASSERT((count & NC_SHLOCK_FLAG) == 0);
449 if (atomic_cmpset_int(&ncp->nc_lockstatus,
456 tsleep_interlock(&ncp->nc_locktd, 0);
457 if (atomic_cmpset_int(&ncp->nc_lockstatus, count,
458 count | NC_EXLOCK_REQ) == 0) {
464 error = tsleep(&ncp->nc_locktd, PINTERLOCKED,
465 "clock", nclockwarn);
466 if (error == EWOULDBLOCK) {
469 kprintf("[diagnostic] cache_lock: "
470 "%s blocked on %p %08x",
471 td->td_comm, ncp, count);
472 kprintf(" \"%*.*s\"\n",
473 ncp->nc_nlen, ncp->nc_nlen,
480 kprintf("[diagnostic] cache_lock: %s unblocked %*.*s after "
483 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
484 (int)(ticks + (hz / 2) - begticks) / hz);
489 * The shared lock works similarly to the exclusive lock except
490 * nc_locktd is left NULL and we need an interlock (VHOLD) to
491 * prevent vhold() races, since the moment our cmpset_int succeeds
492 * another cpu can come in and get its own shared lock.
494 * A critical section is needed to prevent interruption during the
499 _cache_lock_shared(struct namecache *ncp)
504 u_int optreq = NC_EXLOCK_REQ;
506 KKASSERT(ncp->nc_refs != 0);
510 count = ncp->nc_lockstatus;
513 if ((count & ~NC_SHLOCK_REQ) == 0) {
515 if (atomic_cmpset_int(&ncp->nc_lockstatus,
517 (count + 1) | NC_SHLOCK_FLAG |
520 * The vp associated with a locked ncp must
521 * be held to prevent it from being recycled.
523 * WARNING! If VRECLAIMED is set the vnode
524 * could already be in the middle of a recycle.
525 * Callers must use cache_vref() or
526 * cache_vget() on the locked ncp to
527 * validate the vp or set the cache entry
530 * NOTE! vhold() is allowed if we hold a
531 * lock on the ncp (which we do).
535 atomic_clear_int(&ncp->nc_lockstatus,
546 * If already held shared we can just bump the count, but
547 * only allow this if nobody is trying to get the lock
548 * exclusively. If we are blocking too long ignore excl
549 * requests (which can race/deadlock us).
551 * VHOLD is a bit of a hack. Even though we successfully
552 * added another shared ref, the cpu that got the first
553 * shared ref might not yet have held the vnode.
555 if ((count & (optreq|NC_SHLOCK_FLAG)) == NC_SHLOCK_FLAG) {
556 KKASSERT((count & ~(NC_EXLOCK_REQ |
558 NC_SHLOCK_FLAG)) > 0);
559 if (atomic_cmpset_int(&ncp->nc_lockstatus,
561 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD)
567 tsleep_interlock(ncp, 0);
568 if (atomic_cmpset_int(&ncp->nc_lockstatus, count,
569 count | NC_SHLOCK_REQ) == 0) {
573 error = tsleep(ncp, PINTERLOCKED, "clocksh", nclockwarn);
574 if (error == EWOULDBLOCK) {
577 didwarn = ticks - nclockwarn;
578 kprintf("[diagnostic] cache_lock_shared: "
579 "%s blocked on %p %08x "
581 curthread->td_comm, ncp, count,
582 ncp->nc_nlen, ncp->nc_nlen,
589 kprintf("[diagnostic] cache_lock_shared: "
590 "%s unblocked %*.*s after %d secs\n",
592 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
593 (int)(ticks - didwarn) / hz);
598 * Lock ncp exclusively, return 0 on success.
600 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance,
601 * such as the case where one of its children is locked.
605 _cache_lock_nonblock(struct namecache *ncp)
613 count = ncp->nc_lockstatus;
615 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) {
616 if (atomic_cmpset_int(&ncp->nc_lockstatus,
619 * The vp associated with a locked ncp must
620 * be held to prevent it from being recycled.
622 * WARNING! If VRECLAIMED is set the vnode
623 * could already be in the middle of a recycle.
624 * Callers must use cache_vref() or
625 * cache_vget() on the locked ncp to
626 * validate the vp or set the cache entry
629 * NOTE! vhold() is allowed if we hold a
630 * lock on the ncp (which we do).
640 if (ncp->nc_locktd == td) {
641 if (atomic_cmpset_int(&ncp->nc_lockstatus,
654 * The shared lock works similarly to the exclusive lock except
655 * nc_locktd is left NULL and we need an interlock (VHOLD) to
656 * prevent vhold() races, since the moment our cmpset_int succeeds
657 * another cpu can come in and get its own shared lock.
659 * A critical section is needed to prevent interruption during the
664 _cache_lock_shared_nonblock(struct namecache *ncp)
669 count = ncp->nc_lockstatus;
671 if ((count & ~NC_SHLOCK_REQ) == 0) {
673 if (atomic_cmpset_int(&ncp->nc_lockstatus,
675 (count + 1) | NC_SHLOCK_FLAG |
678 * The vp associated with a locked ncp must
679 * be held to prevent it from being recycled.
681 * WARNING! If VRECLAIMED is set the vnode
682 * could already be in the middle of a recycle.
683 * Callers must use cache_vref() or
684 * cache_vget() on the locked ncp to
685 * validate the vp or set the cache entry
688 * NOTE! vhold() is allowed if we hold a
689 * lock on the ncp (which we do).
693 atomic_clear_int(&ncp->nc_lockstatus,
704 * If already held shared we can just bump the count, but
705 * only allow this if nobody is trying to get the lock
708 * VHOLD is a bit of a hack. Even though we successfully
709 * added another shared ref, the cpu that got the first
710 * shared ref might not yet have held the vnode.
712 if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) ==
714 KKASSERT((count & ~(NC_EXLOCK_REQ |
716 NC_SHLOCK_FLAG)) > 0);
717 if (atomic_cmpset_int(&ncp->nc_lockstatus,
719 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD)
733 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop).
735 * nc_locktd must be NULLed out prior to nc_lockstatus getting cleared.
739 _cache_unlock(struct namecache *ncp)
741 thread_t td __debugvar = curthread;
744 struct vnode *dropvp;
746 KKASSERT(ncp->nc_refs >= 0);
747 KKASSERT((ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) > 0);
748 KKASSERT((ncp->nc_lockstatus & NC_SHLOCK_FLAG) || ncp->nc_locktd == td);
750 count = ncp->nc_lockstatus;
754 * Clear nc_locktd prior to the atomic op (excl lock only)
756 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1)
757 ncp->nc_locktd = NULL;
762 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ|NC_SHLOCK_FLAG)) == 1) {
764 if (count & NC_EXLOCK_REQ)
765 ncount = count & NC_SHLOCK_REQ; /* cnt->0 */
769 if (atomic_cmpset_int(&ncp->nc_lockstatus,
771 if (count & NC_EXLOCK_REQ)
772 wakeup(&ncp->nc_locktd);
773 else if (count & NC_SHLOCK_REQ)
779 KKASSERT((count & NC_SHLOCK_VHOLD) == 0);
780 KKASSERT((count & ~(NC_EXLOCK_REQ |
782 NC_SHLOCK_FLAG)) > 1);
783 if (atomic_cmpset_int(&ncp->nc_lockstatus,
788 count = ncp->nc_lockstatus;
793 * Don't actually drop the vp until we successfully clean out
794 * the lock, otherwise we may race another shared lock.
802 _cache_lockstatus(struct namecache *ncp)
804 if (ncp->nc_locktd == curthread)
805 return(LK_EXCLUSIVE);
806 if (ncp->nc_lockstatus & NC_SHLOCK_FLAG)
812 * cache_hold() and cache_drop() prevent the premature deletion of a
813 * namecache entry but do not prevent operations (such as zapping) on
814 * that namecache entry.
816 * This routine may only be called from outside this source module if
817 * nc_refs is already at least 1.
819 * This is a rare case where callers are allowed to hold a spinlock,
820 * so we can't ourselves.
824 _cache_hold(struct namecache *ncp)
826 atomic_add_int(&ncp->nc_refs, 1);
831 * Drop a cache entry, taking care to deal with races.
833 * For potential 1->0 transitions we must hold the ncp lock to safely
834 * test its flags. An unresolved entry with no children must be zapped
837 * The call to cache_zap() itself will handle all remaining races and
838 * will decrement the ncp's refs regardless. If we are resolved or
839 * have children nc_refs can safely be dropped to 0 without having to
842 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion.
844 * NOTE: cache_zap() may return a non-NULL referenced parent which must
845 * be dropped in a loop.
849 _cache_drop(struct namecache *ncp)
854 KKASSERT(ncp->nc_refs > 0);
858 if (_cache_lock_nonblock(ncp) == 0) {
859 ncp->nc_flag &= ~NCF_DEFEREDZAP;
860 if ((ncp->nc_flag & NCF_UNRESOLVED) &&
861 TAILQ_EMPTY(&ncp->nc_list)) {
862 ncp = cache_zap(ncp, 1);
865 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) {
872 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1))
880 * Link a new namecache entry to its parent and to the hash table. Be
881 * careful to avoid races if vhold() blocks in the future.
883 * Both ncp and par must be referenced and locked.
885 * NOTE: The hash table spinlock is held during this call, we can't do
889 _cache_link_parent(struct namecache *ncp, struct namecache *par,
890 struct nchash_head *nchpp)
892 struct pcpu_ncache *pn = &pcpu_ncache[mycpu->gd_cpuid];
894 KKASSERT(ncp->nc_parent == NULL);
895 ncp->nc_parent = par;
896 ncp->nc_head = nchpp;
899 * Set inheritance flags. Note that the parent flags may be
900 * stale due to getattr potentially not having been run yet
901 * (it gets run during nlookup()'s).
903 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE);
904 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE))
905 ncp->nc_flag |= NCF_SF_PNOCACHE;
906 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE))
907 ncp->nc_flag |= NCF_UF_PCACHE;
910 * Add to hash table and parent, adjust accounting
912 TAILQ_INSERT_HEAD(&nchpp->list, ncp, nc_hash);
913 atomic_add_long(&pn->vfscache_count, 1);
914 if (TAILQ_EMPTY(&ncp->nc_list))
915 atomic_add_long(&pn->vfscache_leafs, 1);
917 if (TAILQ_EMPTY(&par->nc_list)) {
918 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
919 atomic_add_long(&pn->vfscache_leafs, -1);
921 * Any vp associated with an ncp which has children must
922 * be held to prevent it from being recycled.
927 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
932 * Remove the parent and hash associations from a namecache structure.
933 * If this is the last child of the parent the cache_drop(par) will
934 * attempt to recursively zap the parent.
936 * ncp must be locked. This routine will acquire a temporary lock on
937 * the parent as wlel as the appropriate hash chain.
940 _cache_unlink_parent(struct namecache *ncp)
942 struct pcpu_ncache *pn = &pcpu_ncache[mycpu->gd_cpuid];
943 struct namecache *par;
944 struct vnode *dropvp;
946 if ((par = ncp->nc_parent) != NULL) {
947 KKASSERT(ncp->nc_parent == par);
950 spin_lock(&ncp->nc_head->spin);
953 * Remove from hash table and parent, adjust accounting
955 TAILQ_REMOVE(&ncp->nc_head->list, ncp, nc_hash);
956 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
957 atomic_add_long(&pn->vfscache_count, -1);
958 if (TAILQ_EMPTY(&ncp->nc_list))
959 atomic_add_long(&pn->vfscache_leafs, -1);
962 if (TAILQ_EMPTY(&par->nc_list)) {
963 atomic_add_long(&pn->vfscache_leafs, 1);
967 spin_unlock(&ncp->nc_head->spin);
968 ncp->nc_parent = NULL;
974 * We can only safely vdrop with no spinlocks held.
982 * Allocate a new namecache structure. Most of the code does not require
983 * zero-termination of the string but it makes vop_compat_ncreate() easier.
985 static struct namecache *
986 cache_alloc(int nlen)
988 struct namecache *ncp;
990 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
992 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
994 ncp->nc_flag = NCF_UNRESOLVED;
995 ncp->nc_error = ENOTCONN; /* needs to be resolved */
998 TAILQ_INIT(&ncp->nc_list);
1004 * Can only be called for the case where the ncp has never been
1005 * associated with anything (so no spinlocks are needed).
1008 _cache_free(struct namecache *ncp)
1010 KKASSERT(ncp->nc_refs == 1 && ncp->nc_lockstatus == 1);
1012 kfree(ncp->nc_name, M_VFSCACHE);
1013 kfree(ncp, M_VFSCACHE);
1017 * [re]initialize a nchandle.
1020 cache_zero(struct nchandle *nch)
1027 * Ref and deref a namecache structure.
1029 * The caller must specify a stable ncp pointer, typically meaning the
1030 * ncp is already referenced but this can also occur indirectly through
1031 * e.g. holding a lock on a direct child.
1033 * WARNING: Caller may hold an unrelated read spinlock, which means we can't
1034 * use read spinlocks here.
1037 cache_hold(struct nchandle *nch)
1039 _cache_hold(nch->ncp);
1040 _cache_mntref(nch->mount);
1045 * Create a copy of a namecache handle for an already-referenced
1049 cache_copy(struct nchandle *nch, struct nchandle *target)
1051 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid];
1052 struct namecache *ncp;
1055 _cache_mntref(target->mount);
1058 if (ncp == cache->ncp1) {
1059 if (atomic_cmpset_ptr((void *)&cache->ncp1, ncp, NULL))
1062 if (ncp == cache->ncp2) {
1063 if (atomic_cmpset_ptr((void *)&cache->ncp2, ncp, NULL))
1071 * Caller wants to copy the current directory, copy it out from our
1072 * pcpu cache if possible (the entire critical path is just two localized
1073 * cmpset ops). If the pcpu cache has a snapshot at all it will be a
1074 * valid one, so we don't have to lock p->p_fd even though we are loading
1077 * This has a limited effect since nlookup must still ref and shlock the
1078 * vnode to check perms. We do avoid the per-proc spin-lock though, which
1079 * can aid threaded programs.
1082 cache_copy_ncdir(struct proc *p, struct nchandle *target)
1084 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid];
1086 *target = p->p_fd->fd_ncdir;
1087 if (target->ncp == cache->ncdir.ncp &&
1088 target->mount == cache->ncdir.mount) {
1089 if (atomic_cmpset_ptr((void *)&cache->ncdir.ncp,
1090 target->ncp, NULL)) {
1091 if (atomic_cmpset_ptr((void *)&cache->ncdir.mount,
1092 target->mount, NULL)) {
1096 _cache_drop(target->ncp);
1099 spin_lock_shared(&p->p_fd->fd_spin);
1100 cache_copy(&p->p_fd->fd_ncdir, target);
1101 spin_unlock_shared(&p->p_fd->fd_spin);
1105 cache_changemount(struct nchandle *nch, struct mount *mp)
1108 _cache_mntrel(nch->mount);
1113 cache_drop(struct nchandle *nch)
1115 _cache_mntrel(nch->mount);
1116 _cache_drop(nch->ncp);
1122 * Drop the nchandle, but try to cache the ref to avoid global atomic
1123 * ops. This is typically done on the system root and jail root nchandles.
1126 cache_drop_and_cache(struct nchandle *nch)
1128 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid];
1129 struct namecache *ncp;
1131 _cache_mntrel(nch->mount);
1133 if (cache->ncp1 == NULL) {
1134 ncp = atomic_swap_ptr((void *)&cache->ncp1, ncp);
1138 if (cache->ncp2 == NULL) {
1139 ncp = atomic_swap_ptr((void *)&cache->ncp2, ncp);
1143 if (++cache->iter & 1)
1144 ncp = atomic_swap_ptr((void *)&cache->ncp2, ncp);
1146 ncp = atomic_swap_ptr((void *)&cache->ncp1, ncp);
1155 * We are dropping what the caller believes is the current directory,
1156 * unconditionally store it in our pcpu cache. Anything already in
1157 * the cache will be discarded.
1160 cache_drop_ncdir(struct nchandle *nch)
1162 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid];
1164 nch->ncp = atomic_swap_ptr((void *)&cache->ncdir.ncp, nch->ncp);
1165 nch->mount = atomic_swap_ptr((void *)&cache->ncdir.mount, nch->mount);
1167 _cache_drop(nch->ncp);
1169 _cache_mntrel(nch->mount);
1175 cache_lockstatus(struct nchandle *nch)
1177 return(_cache_lockstatus(nch->ncp));
1181 cache_lock(struct nchandle *nch)
1183 _cache_lock(nch->ncp);
1187 cache_lock_maybe_shared(struct nchandle *nch, int excl)
1189 struct namecache *ncp = nch->ncp;
1191 if (ncp_shared_lock_disable || excl ||
1192 (ncp->nc_flag & NCF_UNRESOLVED)) {
1195 _cache_lock_shared(ncp);
1196 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1197 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) {
1209 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller
1210 * is responsible for checking both for validity on return as they
1211 * may have become invalid.
1213 * We have to deal with potential deadlocks here, just ping pong
1214 * the lock until we get it (we will always block somewhere when
1215 * looping so this is not cpu-intensive).
1217 * which = 0 nch1 not locked, nch2 is locked
1218 * which = 1 nch1 is locked, nch2 is not locked
1221 cache_relock(struct nchandle *nch1, struct ucred *cred1,
1222 struct nchandle *nch2, struct ucred *cred2)
1230 if (cache_lock_nonblock(nch1) == 0) {
1231 cache_resolve(nch1, cred1);
1236 cache_resolve(nch1, cred1);
1239 if (cache_lock_nonblock(nch2) == 0) {
1240 cache_resolve(nch2, cred2);
1245 cache_resolve(nch2, cred2);
1252 cache_lock_nonblock(struct nchandle *nch)
1254 return(_cache_lock_nonblock(nch->ncp));
1258 cache_unlock(struct nchandle *nch)
1260 _cache_unlock(nch->ncp);
1264 * ref-and-lock, unlock-and-deref functions.
1266 * This function is primarily used by nlookup. Even though cache_lock
1267 * holds the vnode, it is possible that the vnode may have already
1268 * initiated a recyclement.
1270 * We want cache_get() to return a definitively usable vnode or a
1271 * definitively unresolved ncp.
1275 _cache_get(struct namecache *ncp)
1279 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1280 _cache_setunresolved(ncp);
1285 * Attempt to obtain a shared lock on the ncp. A shared lock will only
1286 * be obtained if the ncp is resolved and the vnode (if not ENOENT) is
1287 * valid. Otherwise an exclusive lock will be acquired instead.
1291 _cache_get_maybe_shared(struct namecache *ncp, int excl)
1293 if (ncp_shared_lock_disable || excl ||
1294 (ncp->nc_flag & NCF_UNRESOLVED)) {
1295 return(_cache_get(ncp));
1298 _cache_lock_shared(ncp);
1299 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1300 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) {
1302 ncp = _cache_get(ncp);
1307 ncp = _cache_get(ncp);
1314 * This is a special form of _cache_lock() which only succeeds if
1315 * it can get a pristine, non-recursive lock. The caller must have
1316 * already ref'd the ncp.
1318 * On success the ncp will be locked, on failure it will not. The
1319 * ref count does not change either way.
1321 * We want _cache_lock_special() (on success) to return a definitively
1322 * usable vnode or a definitively unresolved ncp.
1325 _cache_lock_special(struct namecache *ncp)
1327 if (_cache_lock_nonblock(ncp) == 0) {
1328 if ((ncp->nc_lockstatus &
1329 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) {
1330 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1331 _cache_setunresolved(ncp);
1336 return(EWOULDBLOCK);
1340 * This function tries to get a shared lock but will back-off to an exclusive
1343 * (1) Some other thread is trying to obtain an exclusive lock
1344 * (to prevent the exclusive requester from getting livelocked out
1345 * by many shared locks).
1347 * (2) The current thread already owns an exclusive lock (to avoid
1350 * WARNING! On machines with lots of cores we really want to try hard to
1351 * get a shared lock or concurrent path lookups can chain-react
1352 * into a very high-latency exclusive lock.
1355 _cache_lock_shared_special(struct namecache *ncp)
1358 * Only honor a successful shared lock (returning 0) if there is
1359 * no exclusive request pending and the vnode, if present, is not
1360 * in a reclaimed state.
1362 if (_cache_lock_shared_nonblock(ncp) == 0) {
1363 if ((ncp->nc_lockstatus & NC_EXLOCK_REQ) == 0) {
1364 if (ncp->nc_vp == NULL ||
1365 (ncp->nc_vp->v_flag & VRECLAIMED) == 0) {
1370 return(EWOULDBLOCK);
1374 * Non-blocking shared lock failed. If we already own the exclusive
1375 * lock just acquire another exclusive lock (instead of deadlocking).
1376 * Otherwise acquire a shared lock.
1378 if (ncp->nc_locktd == curthread) {
1382 _cache_lock_shared(ncp);
1388 * NOTE: The same nchandle can be passed for both arguments.
1391 cache_get(struct nchandle *nch, struct nchandle *target)
1393 KKASSERT(nch->ncp->nc_refs > 0);
1394 target->mount = nch->mount;
1395 target->ncp = _cache_get(nch->ncp);
1396 _cache_mntref(target->mount);
1400 cache_get_maybe_shared(struct nchandle *nch, struct nchandle *target, int excl)
1402 KKASSERT(nch->ncp->nc_refs > 0);
1403 target->mount = nch->mount;
1404 target->ncp = _cache_get_maybe_shared(nch->ncp, excl);
1405 _cache_mntref(target->mount);
1413 _cache_put(struct namecache *ncp)
1423 cache_put(struct nchandle *nch)
1425 _cache_mntrel(nch->mount);
1426 _cache_put(nch->ncp);
1432 * Resolve an unresolved ncp by associating a vnode with it. If the
1433 * vnode is NULL, a negative cache entry is created.
1435 * The ncp should be locked on entry and will remain locked on return.
1439 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp)
1441 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
1442 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
1446 * Any vp associated with an ncp which has children must
1447 * be held. Any vp associated with a locked ncp must be held.
1449 if (!TAILQ_EMPTY(&ncp->nc_list))
1451 spin_lock(&vp->v_spin);
1453 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
1454 spin_unlock(&vp->v_spin);
1455 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ))
1459 * Set auxiliary flags
1461 switch(vp->v_type) {
1463 ncp->nc_flag |= NCF_ISDIR;
1466 ncp->nc_flag |= NCF_ISSYMLINK;
1467 /* XXX cache the contents of the symlink */
1473 /* XXX: this is a hack to work-around the lack of a real pfs vfs
1476 if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0)
1480 * When creating a negative cache hit we set the
1481 * namecache_gen. A later resolve will clean out the
1482 * negative cache hit if the mount point's namecache_gen
1483 * has changed. Used by devfs, could also be used by
1486 struct pcpu_ncache *pn = &pcpu_ncache[mycpu->gd_cpuid];
1489 ncp->nc_negcpu = mycpu->gd_cpuid;
1490 spin_lock(&pn->neg_spin);
1491 TAILQ_INSERT_TAIL(&pn->neg_list, ncp, nc_vnode);
1493 spin_unlock(&pn->neg_spin);
1494 atomic_add_long(&pn->vfscache_negs, 1);
1496 ncp->nc_error = ENOENT;
1498 VFS_NCPGEN_SET(mp, ncp);
1500 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP);
1507 cache_setvp(struct nchandle *nch, struct vnode *vp)
1509 _cache_setvp(nch->mount, nch->ncp, vp);
1516 cache_settimeout(struct nchandle *nch, int nticks)
1518 struct namecache *ncp = nch->ncp;
1520 if ((ncp->nc_timeout = ticks + nticks) == 0)
1521 ncp->nc_timeout = 1;
1525 * Disassociate the vnode or negative-cache association and mark a
1526 * namecache entry as unresolved again. Note that the ncp is still
1527 * left in the hash table and still linked to its parent.
1529 * The ncp should be locked and refd on entry and will remain locked and refd
1532 * This routine is normally never called on a directory containing children.
1533 * However, NFS often does just that in its rename() code as a cop-out to
1534 * avoid complex namespace operations. This disconnects a directory vnode
1535 * from its namecache and can cause the OLDAPI and NEWAPI to get out of
1541 _cache_setunresolved(struct namecache *ncp)
1545 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1546 ncp->nc_flag |= NCF_UNRESOLVED;
1547 ncp->nc_timeout = 0;
1548 ncp->nc_error = ENOTCONN;
1549 if ((vp = ncp->nc_vp) != NULL) {
1550 spin_lock(&vp->v_spin);
1552 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
1553 spin_unlock(&vp->v_spin);
1556 * Any vp associated with an ncp with children is
1557 * held by that ncp. Any vp associated with a locked
1558 * ncp is held by that ncp. These conditions must be
1559 * undone when the vp is cleared out from the ncp.
1561 if (!TAILQ_EMPTY(&ncp->nc_list))
1563 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ))
1566 struct pcpu_ncache *pn;
1568 pn = &pcpu_ncache[ncp->nc_negcpu];
1570 atomic_add_long(&pn->vfscache_negs, -1);
1571 spin_lock(&pn->neg_spin);
1572 TAILQ_REMOVE(&pn->neg_list, ncp, nc_vnode);
1574 spin_unlock(&pn->neg_spin);
1576 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK);
1581 * The cache_nresolve() code calls this function to automatically
1582 * set a resolved cache element to unresolved if it has timed out
1583 * or if it is a negative cache hit and the mount point namecache_gen
1587 _cache_auto_unresolve_test(struct mount *mp, struct namecache *ncp)
1590 * Try to zap entries that have timed out. We have
1591 * to be careful here because locked leafs may depend
1592 * on the vnode remaining intact in a parent, so only
1593 * do this under very specific conditions.
1595 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 &&
1596 TAILQ_EMPTY(&ncp->nc_list)) {
1601 * If a resolved negative cache hit is invalid due to
1602 * the mount's namecache generation being bumped, zap it.
1604 if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) {
1609 * Otherwise we are good
1614 static __inline void
1615 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp)
1618 * Already in an unresolved state, nothing to do.
1620 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1621 if (_cache_auto_unresolve_test(mp, ncp))
1622 _cache_setunresolved(ncp);
1630 cache_setunresolved(struct nchandle *nch)
1632 _cache_setunresolved(nch->ncp);
1636 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
1637 * looking for matches. This flag tells the lookup code when it must
1638 * check for a mount linkage and also prevents the directories in question
1639 * from being deleted or renamed.
1643 cache_clrmountpt_callback(struct mount *mp, void *data)
1645 struct nchandle *nch = data;
1647 if (mp->mnt_ncmounton.ncp == nch->ncp)
1649 if (mp->mnt_ncmountpt.ncp == nch->ncp)
1655 * Clear NCF_ISMOUNTPT on nch->ncp if it is no longer associated
1656 * with a mount point.
1659 cache_clrmountpt(struct nchandle *nch)
1663 count = mountlist_scan(cache_clrmountpt_callback, nch,
1664 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1666 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
1670 * Invalidate portions of the namecache topology given a starting entry.
1671 * The passed ncp is set to an unresolved state and:
1673 * The passed ncp must be referencxed and locked. The routine may unlock
1674 * and relock ncp several times, and will recheck the children and loop
1675 * to catch races. When done the passed ncp will be returned with the
1676 * reference and lock intact.
1678 * CINV_DESTROY - Set a flag in the passed ncp entry indicating
1679 * that the physical underlying nodes have been
1680 * destroyed... as in deleted. For example, when
1681 * a directory is removed. This will cause record
1682 * lookups on the name to no longer be able to find
1683 * the record and tells the resolver to return failure
1684 * rather then trying to resolve through the parent.
1686 * The topology itself, including ncp->nc_name,
1689 * This only applies to the passed ncp, if CINV_CHILDREN
1690 * is specified the children are not flagged.
1692 * CINV_CHILDREN - Set all children (recursively) to an unresolved
1695 * Note that this will also have the side effect of
1696 * cleaning out any unreferenced nodes in the topology
1697 * from the leaves up as the recursion backs out.
1699 * Note that the topology for any referenced nodes remains intact, but
1700 * the nodes will be marked as having been destroyed and will be set
1701 * to an unresolved state.
1703 * It is possible for cache_inval() to race a cache_resolve(), meaning that
1704 * the namecache entry may not actually be invalidated on return if it was
1705 * revalidated while recursing down into its children. This code guarentees
1706 * that the node(s) will go through an invalidation cycle, but does not
1707 * guarentee that they will remain in an invalidated state.
1709 * Returns non-zero if a revalidation was detected during the invalidation
1710 * recursion, zero otherwise. Note that since only the original ncp is
1711 * locked the revalidation ultimately can only indicate that the original ncp
1712 * *MIGHT* no have been reresolved.
1714 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
1715 * have to avoid blowing out the kernel stack. We do this by saving the
1716 * deep namecache node and aborting the recursion, then re-recursing at that
1717 * node using a depth-first algorithm in order to allow multiple deep
1718 * recursions to chain through each other, then we restart the invalidation
1723 struct namecache *resume_ncp;
1727 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
1731 _cache_inval(struct namecache *ncp, int flags)
1733 struct cinvtrack track;
1734 struct namecache *ncp2;
1738 track.resume_ncp = NULL;
1741 r = _cache_inval_internal(ncp, flags, &track);
1742 if (track.resume_ncp == NULL)
1745 while ((ncp2 = track.resume_ncp) != NULL) {
1746 track.resume_ncp = NULL;
1748 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
1758 cache_inval(struct nchandle *nch, int flags)
1760 return(_cache_inval(nch->ncp, flags));
1764 * Helper for _cache_inval(). The passed ncp is refd and locked and
1765 * remains that way on return, but may be unlocked/relocked multiple
1766 * times by the routine.
1769 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
1771 struct namecache *nextkid;
1774 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
1776 _cache_setunresolved(ncp);
1777 if (flags & CINV_DESTROY) {
1778 ncp->nc_flag |= NCF_DESTROYED;
1779 ++ncp->nc_generation;
1781 while ((flags & CINV_CHILDREN) &&
1782 (nextkid = TAILQ_FIRST(&ncp->nc_list)) != NULL
1784 struct namecache *kid;
1788 _cache_hold(nextkid);
1789 if (++track->depth > MAX_RECURSION_DEPTH) {
1790 track->resume_ncp = ncp;
1794 while ((kid = nextkid) != NULL) {
1796 * Parent (ncp) must be locked for the iteration.
1799 if (kid->nc_parent != ncp) {
1801 kprintf("cache_inval_internal restartA %s\n",
1806 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
1807 _cache_hold(nextkid);
1810 * Parent unlocked for this section to avoid
1814 if (track->resume_ncp) {
1819 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
1820 TAILQ_FIRST(&kid->nc_list)
1823 if (kid->nc_parent != ncp) {
1824 kprintf("cache_inval_internal "
1834 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
1841 _cache_drop(nextkid);
1848 * Someone could have gotten in there while ncp was unlocked,
1851 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1857 * Invalidate a vnode's namecache associations. To avoid races against
1858 * the resolver we do not invalidate a node which we previously invalidated
1859 * but which was then re-resolved while we were in the invalidation loop.
1861 * Returns non-zero if any namecache entries remain after the invalidation
1864 * NOTE: Unlike the namecache topology which guarentees that ncp's will not
1865 * be ripped out of the topology while held, the vnode's v_namecache
1866 * list has no such restriction. NCP's can be ripped out of the list
1867 * at virtually any time if not locked, even if held.
1869 * In addition, the v_namecache list itself must be locked via
1870 * the vnode's spinlock.
1873 cache_inval_vp(struct vnode *vp, int flags)
1875 struct namecache *ncp;
1876 struct namecache *next;
1879 spin_lock(&vp->v_spin);
1880 ncp = TAILQ_FIRST(&vp->v_namecache);
1884 /* loop entered with ncp held and vp spin-locked */
1885 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1887 spin_unlock(&vp->v_spin);
1889 if (ncp->nc_vp != vp) {
1890 kprintf("Warning: cache_inval_vp: race-A detected on "
1891 "%s\n", ncp->nc_name);
1897 _cache_inval(ncp, flags);
1898 _cache_put(ncp); /* also releases reference */
1900 spin_lock(&vp->v_spin);
1901 if (ncp && ncp->nc_vp != vp) {
1902 spin_unlock(&vp->v_spin);
1903 kprintf("Warning: cache_inval_vp: race-B detected on "
1904 "%s\n", ncp->nc_name);
1909 spin_unlock(&vp->v_spin);
1910 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1914 * This routine is used instead of the normal cache_inval_vp() when we
1915 * are trying to recycle otherwise good vnodes.
1917 * Return 0 on success, non-zero if not all namecache records could be
1918 * disassociated from the vnode (for various reasons).
1921 cache_inval_vp_nonblock(struct vnode *vp)
1923 struct namecache *ncp;
1924 struct namecache *next;
1926 spin_lock(&vp->v_spin);
1927 ncp = TAILQ_FIRST(&vp->v_namecache);
1931 /* loop entered with ncp held */
1932 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1934 spin_unlock(&vp->v_spin);
1935 if (_cache_lock_nonblock(ncp)) {
1941 if (ncp->nc_vp != vp) {
1942 kprintf("Warning: cache_inval_vp: race-A detected on "
1943 "%s\n", ncp->nc_name);
1949 _cache_inval(ncp, 0);
1950 _cache_put(ncp); /* also releases reference */
1952 spin_lock(&vp->v_spin);
1953 if (ncp && ncp->nc_vp != vp) {
1954 spin_unlock(&vp->v_spin);
1955 kprintf("Warning: cache_inval_vp: race-B detected on "
1956 "%s\n", ncp->nc_name);
1961 spin_unlock(&vp->v_spin);
1963 return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1967 * Clears the universal directory search 'ok' flag. This flag allows
1968 * nlookup() to bypass normal vnode checks. This flag is a cached flag
1969 * so clearing it simply forces revalidation.
1972 cache_inval_wxok(struct vnode *vp)
1974 struct namecache *ncp;
1976 spin_lock(&vp->v_spin);
1977 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1978 if (ncp->nc_flag & NCF_WXOK)
1979 atomic_clear_short(&ncp->nc_flag, NCF_WXOK);
1981 spin_unlock(&vp->v_spin);
1985 * The source ncp has been renamed to the target ncp. Both fncp and tncp
1986 * must be locked. The target ncp is destroyed (as a normal rename-over
1987 * would destroy the target file or directory).
1989 * Because there may be references to the source ncp we cannot copy its
1990 * contents to the target. Instead the source ncp is relinked as the target
1991 * and the target ncp is removed from the namecache topology.
1994 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
1996 struct namecache *fncp = fnch->ncp;
1997 struct namecache *tncp = tnch->ncp;
1998 struct namecache *tncp_par;
1999 struct nchash_head *nchpp;
2004 ++fncp->nc_generation;
2005 ++tncp->nc_generation;
2006 if (tncp->nc_nlen) {
2007 nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK);
2008 bcopy(tncp->nc_name, nname, tncp->nc_nlen);
2009 nname[tncp->nc_nlen] = 0;
2015 * Rename fncp (unlink)
2017 _cache_unlink_parent(fncp);
2018 oname = fncp->nc_name;
2019 fncp->nc_name = nname;
2020 fncp->nc_nlen = tncp->nc_nlen;
2022 kfree(oname, M_VFSCACHE);
2024 tncp_par = tncp->nc_parent;
2025 _cache_hold(tncp_par);
2026 _cache_lock(tncp_par);
2029 * Rename fncp (relink)
2031 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT);
2032 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash);
2033 nchpp = NCHHASH(hash);
2035 spin_lock(&nchpp->spin);
2036 _cache_link_parent(fncp, tncp_par, nchpp);
2037 spin_unlock(&nchpp->spin);
2039 _cache_put(tncp_par);
2042 * Get rid of the overwritten tncp (unlink)
2044 _cache_unlink(tncp);
2048 * Perform actions consistent with unlinking a file. The passed-in ncp
2051 * The ncp is marked DESTROYED so it no longer shows up in searches,
2052 * and will be physically deleted when the vnode goes away.
2054 * If the related vnode has no refs then we cycle it through vget()/vput()
2055 * to (possibly if we don't have a ref race) trigger a deactivation,
2056 * allowing the VFS to trivially detect and recycle the deleted vnode
2057 * via VOP_INACTIVE().
2059 * NOTE: _cache_rename() will automatically call _cache_unlink() on the
2063 cache_unlink(struct nchandle *nch)
2065 _cache_unlink(nch->ncp);
2069 _cache_unlink(struct namecache *ncp)
2074 * Causes lookups to fail and allows another ncp with the same
2075 * name to be created under ncp->nc_parent.
2077 ncp->nc_flag |= NCF_DESTROYED;
2078 ++ncp->nc_generation;
2081 * Attempt to trigger a deactivation. Set VREF_FINALIZE to
2082 * force action on the 1->0 transition.
2084 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
2085 (vp = ncp->nc_vp) != NULL) {
2086 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE);
2087 if (VREFCNT(vp) <= 0) {
2088 if (vget(vp, LK_SHARED) == 0)
2095 * Return non-zero if the nch might be associated with an open and/or mmap()'d
2096 * file. The easy solution is to just return non-zero if the vnode has refs.
2097 * Used to interlock hammer2 reclaims (VREF_FINALIZE should already be set to
2098 * force the reclaim).
2101 cache_isopen(struct nchandle *nch)
2104 struct namecache *ncp = nch->ncp;
2106 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
2107 (vp = ncp->nc_vp) != NULL &&
2116 * vget the vnode associated with the namecache entry. Resolve the namecache
2117 * entry if necessary. The passed ncp must be referenced and locked. If
2118 * the ncp is resolved it might be locked shared.
2120 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked
2121 * (depending on the passed lk_type) will be returned in *vpp with an error
2122 * of 0, or NULL will be returned in *vpp with a non-0 error code. The
2123 * most typical error is ENOENT, meaning that the ncp represents a negative
2124 * cache hit and there is no vnode to retrieve, but other errors can occur
2127 * The vget() can race a reclaim. If this occurs we re-resolve the
2130 * There are numerous places in the kernel where vget() is called on a
2131 * vnode while one or more of its namecache entries is locked. Releasing
2132 * a vnode never deadlocks against locked namecache entries (the vnode
2133 * will not get recycled while referenced ncp's exist). This means we
2134 * can safely acquire the vnode. In fact, we MUST NOT release the ncp
2135 * lock when acquiring the vp lock or we might cause a deadlock.
2137 * NOTE: The passed-in ncp must be locked exclusively if it is initially
2138 * unresolved. If a reclaim race occurs the passed-in ncp will be
2139 * relocked exclusively before being re-resolved.
2142 cache_vget(struct nchandle *nch, struct ucred *cred,
2143 int lk_type, struct vnode **vpp)
2145 struct namecache *ncp;
2152 if (ncp->nc_flag & NCF_UNRESOLVED)
2153 error = cache_resolve(nch, cred);
2157 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
2158 error = vget(vp, lk_type);
2163 * The ncp may have been locked shared, we must relock
2164 * it exclusively before we can set it to unresolved.
2166 if (error == ENOENT) {
2167 kprintf("Warning: vnode reclaim race detected "
2168 "in cache_vget on %p (%s)\n",
2172 _cache_setunresolved(ncp);
2177 * Not a reclaim race, some other error.
2179 KKASSERT(ncp->nc_vp == vp);
2182 KKASSERT(ncp->nc_vp == vp);
2183 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
2186 if (error == 0 && vp == NULL)
2193 * Similar to cache_vget() but only acquires a ref on the vnode.
2195 * NOTE: The passed-in ncp must be locked exclusively if it is initially
2196 * unresolved. If a reclaim race occurs the passed-in ncp will be
2197 * relocked exclusively before being re-resolved.
2200 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
2202 struct namecache *ncp;
2209 if (ncp->nc_flag & NCF_UNRESOLVED)
2210 error = cache_resolve(nch, cred);
2214 if (error == 0 && (vp = ncp->nc_vp) != NULL) {
2215 error = vget(vp, LK_SHARED);
2220 if (error == ENOENT) {
2221 kprintf("Warning: vnode reclaim race detected "
2222 "in cache_vget on %p (%s)\n",
2226 _cache_setunresolved(ncp);
2231 * Not a reclaim race, some other error.
2233 KKASSERT(ncp->nc_vp == vp);
2236 KKASSERT(ncp->nc_vp == vp);
2237 KKASSERT((vp->v_flag & VRECLAIMED) == 0);
2238 /* caller does not want a lock */
2242 if (error == 0 && vp == NULL)
2249 * Return a referenced vnode representing the parent directory of
2252 * Because the caller has locked the ncp it should not be possible for
2253 * the parent ncp to go away. However, the parent can unresolve its
2254 * dvp at any time so we must be able to acquire a lock on the parent
2255 * to safely access nc_vp.
2257 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock,
2258 * so use vhold()/vdrop() while holding the lock to prevent dvp from
2259 * getting destroyed.
2261 * NOTE: vhold() is allowed when dvp has 0 refs if we hold a
2262 * lock on the ncp in question..
2264 static struct vnode *
2265 cache_dvpref(struct namecache *ncp)
2267 struct namecache *par;
2271 if ((par = ncp->nc_parent) != NULL) {
2274 if ((par->nc_flag & NCF_UNRESOLVED) == 0) {
2275 if ((dvp = par->nc_vp) != NULL)
2280 if (vget(dvp, LK_SHARED) == 0) {
2283 /* return refd, unlocked dvp */
2295 * Convert a directory vnode to a namecache record without any other
2296 * knowledge of the topology. This ONLY works with directory vnodes and
2297 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the
2298 * returned ncp (if not NULL) will be held and unlocked.
2300 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
2301 * If 'makeit' is 1 we attempt to track-down and create the namecache topology
2302 * for dvp. This will fail only if the directory has been deleted out from
2305 * Callers must always check for a NULL return no matter the value of 'makeit'.
2307 * To avoid underflowing the kernel stack each recursive call increments
2308 * the makeit variable.
2311 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
2312 struct vnode *dvp, char *fakename);
2313 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
2314 struct vnode **saved_dvp);
2317 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
2318 struct nchandle *nch)
2320 struct vnode *saved_dvp;
2326 nch->mount = dvp->v_mount;
2331 * Handle the makeit == 0 degenerate case
2334 spin_lock_shared(&dvp->v_spin);
2335 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
2338 spin_unlock_shared(&dvp->v_spin);
2342 * Loop until resolution, inside code will break out on error.
2346 * Break out if we successfully acquire a working ncp.
2348 spin_lock_shared(&dvp->v_spin);
2349 nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
2352 spin_unlock_shared(&dvp->v_spin);
2355 spin_unlock_shared(&dvp->v_spin);
2358 * If dvp is the root of its filesystem it should already
2359 * have a namecache pointer associated with it as a side
2360 * effect of the mount, but it may have been disassociated.
2362 if (dvp->v_flag & VROOT) {
2363 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
2364 error = cache_resolve_mp(nch->mount);
2365 _cache_put(nch->ncp);
2367 kprintf("cache_fromdvp: resolve root of mount %p error %d",
2368 dvp->v_mount, error);
2372 kprintf(" failed\n");
2377 kprintf(" succeeded\n");
2382 * If we are recursed too deeply resort to an O(n^2)
2383 * algorithm to resolve the namecache topology. The
2384 * resolved pvp is left referenced in saved_dvp to
2385 * prevent the tree from being destroyed while we loop.
2388 error = cache_fromdvp_try(dvp, cred, &saved_dvp);
2390 kprintf("lookupdotdot(longpath) failed %d "
2391 "dvp %p\n", error, dvp);
2399 * Get the parent directory and resolve its ncp.
2402 kfree(fakename, M_TEMP);
2405 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
2408 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
2414 * Reuse makeit as a recursion depth counter. On success
2415 * nch will be fully referenced.
2417 cache_fromdvp(pvp, cred, makeit + 1, nch);
2419 if (nch->ncp == NULL)
2423 * Do an inefficient scan of pvp (embodied by ncp) to look
2424 * for dvp. This will create a namecache record for dvp on
2425 * success. We loop up to recheck on success.
2427 * ncp and dvp are both held but not locked.
2429 error = cache_inefficient_scan(nch, cred, dvp, fakename);
2431 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
2432 pvp, nch->ncp->nc_name, dvp);
2434 /* nch was NULLed out, reload mount */
2435 nch->mount = dvp->v_mount;
2439 kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
2440 pvp, nch->ncp->nc_name);
2443 /* nch was NULLed out, reload mount */
2444 nch->mount = dvp->v_mount;
2448 * If nch->ncp is non-NULL it will have been held already.
2451 kfree(fakename, M_TEMP);
2460 * Go up the chain of parent directories until we find something
2461 * we can resolve into the namecache. This is very inefficient.
2465 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
2466 struct vnode **saved_dvp)
2468 struct nchandle nch;
2471 static time_t last_fromdvp_report;
2475 * Loop getting the parent directory vnode until we get something we
2476 * can resolve in the namecache.
2479 nch.mount = dvp->v_mount;
2485 kfree(fakename, M_TEMP);
2488 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
2495 spin_lock_shared(&pvp->v_spin);
2496 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
2497 _cache_hold(nch.ncp);
2498 spin_unlock_shared(&pvp->v_spin);
2502 spin_unlock_shared(&pvp->v_spin);
2503 if (pvp->v_flag & VROOT) {
2504 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
2505 error = cache_resolve_mp(nch.mount);
2506 _cache_unlock(nch.ncp);
2509 _cache_drop(nch.ncp);
2519 if (last_fromdvp_report != time_uptime) {
2520 last_fromdvp_report = time_uptime;
2521 kprintf("Warning: extremely inefficient path "
2522 "resolution on %s\n",
2525 error = cache_inefficient_scan(&nch, cred, dvp, fakename);
2528 * Hopefully dvp now has a namecache record associated with
2529 * it. Leave it referenced to prevent the kernel from
2530 * recycling the vnode. Otherwise extremely long directory
2531 * paths could result in endless recycling.
2536 _cache_drop(nch.ncp);
2539 kfree(fakename, M_TEMP);
2544 * Do an inefficient scan of the directory represented by ncp looking for
2545 * the directory vnode dvp. ncp must be held but not locked on entry and
2546 * will be held on return. dvp must be refd but not locked on entry and
2547 * will remain refd on return.
2549 * Why do this at all? Well, due to its stateless nature the NFS server
2550 * converts file handles directly to vnodes without necessarily going through
2551 * the namecache ops that would otherwise create the namecache topology
2552 * leading to the vnode. We could either (1) Change the namecache algorithms
2553 * to allow disconnect namecache records that are re-merged opportunistically,
2554 * or (2) Make the NFS server backtrack and scan to recover a connected
2555 * namecache topology in order to then be able to issue new API lookups.
2557 * It turns out that (1) is a huge mess. It takes a nice clean set of
2558 * namecache algorithms and introduces a lot of complication in every subsystem
2559 * that calls into the namecache to deal with the re-merge case, especially
2560 * since we are using the namecache to placehold negative lookups and the
2561 * vnode might not be immediately assigned. (2) is certainly far less
2562 * efficient then (1), but since we are only talking about directories here
2563 * (which are likely to remain cached), the case does not actually run all
2564 * that often and has the supreme advantage of not polluting the namecache
2567 * If a fakename is supplied just construct a namecache entry using the
2571 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
2572 struct vnode *dvp, char *fakename)
2574 struct nlcomponent nlc;
2575 struct nchandle rncp;
2587 vat.va_blocksize = 0;
2588 if ((error = VOP_GETATTR(dvp, &vat)) != 0)
2591 error = cache_vref(nch, cred, &pvp);
2596 kprintf("inefficient_scan of (%p,%s): directory iosize %ld "
2597 "vattr fileid = %lld\n",
2598 nch->ncp, nch->ncp->nc_name,
2600 (long long)vat.va_fileid);
2604 * Use the supplied fakename if not NULL. Fake names are typically
2605 * not in the actual filesystem hierarchy. This is used by HAMMER
2606 * to glue @@timestamp recursions together.
2609 nlc.nlc_nameptr = fakename;
2610 nlc.nlc_namelen = strlen(fakename);
2611 rncp = cache_nlookup(nch, &nlc);
2615 if ((blksize = vat.va_blocksize) == 0)
2616 blksize = DEV_BSIZE;
2617 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
2623 iov.iov_base = rbuf;
2624 iov.iov_len = blksize;
2627 uio.uio_resid = blksize;
2628 uio.uio_segflg = UIO_SYSSPACE;
2629 uio.uio_rw = UIO_READ;
2630 uio.uio_td = curthread;
2632 if (ncvp_debug >= 2)
2633 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
2634 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
2636 den = (struct dirent *)rbuf;
2637 bytes = blksize - uio.uio_resid;
2640 if (ncvp_debug >= 2) {
2641 kprintf("cache_inefficient_scan: %*.*s\n",
2642 den->d_namlen, den->d_namlen,
2645 if (den->d_type != DT_WHT &&
2646 den->d_ino == vat.va_fileid) {
2648 kprintf("cache_inefficient_scan: "
2649 "MATCHED inode %lld path %s/%*.*s\n",
2650 (long long)vat.va_fileid,
2652 den->d_namlen, den->d_namlen,
2655 nlc.nlc_nameptr = den->d_name;
2656 nlc.nlc_namelen = den->d_namlen;
2657 rncp = cache_nlookup(nch, &nlc);
2658 KKASSERT(rncp.ncp != NULL);
2661 bytes -= _DIRENT_DIRSIZ(den);
2662 den = _DIRENT_NEXT(den);
2664 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
2667 kfree(rbuf, M_TEMP);
2671 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
2672 _cache_setvp(rncp.mount, rncp.ncp, dvp);
2673 if (ncvp_debug >= 2) {
2674 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
2675 nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
2678 if (ncvp_debug >= 2) {
2679 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
2680 nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
2684 if (rncp.ncp->nc_vp == NULL)
2685 error = rncp.ncp->nc_error;
2687 * Release rncp after a successful nlookup. rncp was fully
2692 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
2693 dvp, nch->ncp->nc_name);
2700 * Zap a namecache entry. The ncp is unconditionally set to an unresolved
2701 * state, which disassociates it from its vnode or pcpu_ncache[n].neg_list.
2703 * Then, if there are no additional references to the ncp and no children,
2704 * the ncp is removed from the topology and destroyed.
2706 * References and/or children may exist if the ncp is in the middle of the
2707 * topology, preventing the ncp from being destroyed.
2709 * This function must be called with the ncp held and locked and will unlock
2710 * and drop it during zapping.
2712 * If nonblock is non-zero and the parent ncp cannot be locked we give up.
2713 * This case can occur in the cache_drop() path.
2715 * This function may returned a held (but NOT locked) parent node which the
2716 * caller must drop. We do this so _cache_drop() can loop, to avoid
2717 * blowing out the kernel stack.
2719 * WARNING! For MPSAFE operation this routine must acquire up to three
2720 * spin locks to be able to safely test nc_refs. Lock order is
2723 * hash spinlock if on hash list
2724 * parent spinlock if child of parent
2725 * (the ncp is unresolved so there is no vnode association)
2727 static struct namecache *
2728 cache_zap(struct namecache *ncp, int nonblock)
2730 struct namecache *par;
2731 struct vnode *dropvp;
2732 struct nchash_head *nchpp;
2736 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
2738 _cache_setunresolved(ncp);
2741 * Try to scrap the entry and possibly tail-recurse on its parent.
2742 * We only scrap unref'd (other then our ref) unresolved entries,
2743 * we do not scrap 'live' entries.
2745 * Note that once the spinlocks are acquired if nc_refs == 1 no
2746 * other references are possible. If it isn't, however, we have
2747 * to decrement but also be sure to avoid a 1->0 transition.
2749 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
2750 KKASSERT(ncp->nc_refs > 0);
2753 * Acquire locks. Note that the parent can't go away while we hold
2757 if ((par = ncp->nc_parent) != NULL) {
2760 if (_cache_lock_nonblock(par) == 0)
2762 refs = ncp->nc_refs;
2763 ncp->nc_flag |= NCF_DEFEREDZAP;
2764 ++numdefered; /* MP race ok */
2765 if (atomic_cmpset_int(&ncp->nc_refs,
2777 nchpp = ncp->nc_head;
2778 spin_lock(&nchpp->spin);
2782 * At this point if we find refs == 1 it should not be possible for
2783 * anyone else to have access to the ncp. We are holding the only
2784 * possible access point left (nchpp) spin-locked.
2786 * If someone other then us has a ref or we have children
2787 * we cannot zap the entry. The 1->0 transition and any
2788 * further list operation is protected by the spinlocks
2789 * we have acquired but other transitions are not.
2792 refs = ncp->nc_refs;
2794 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list))
2796 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) {
2798 spin_unlock(&nchpp->spin);
2808 * We are the only ref and with the spinlocks held no further
2809 * refs can be acquired by others.
2811 * Remove us from the hash list and parent list. We have to
2812 * drop a ref on the parent's vp if the parent's list becomes
2817 struct pcpu_ncache *pn = &pcpu_ncache[mycpu->gd_cpuid];
2819 KKASSERT(nchpp == ncp->nc_head);
2820 TAILQ_REMOVE(&ncp->nc_head->list, ncp, nc_hash);
2821 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
2822 atomic_add_long(&pn->vfscache_count, -1);
2823 if (TAILQ_EMPTY(&ncp->nc_list))
2824 atomic_add_long(&pn->vfscache_leafs, -1);
2826 if (TAILQ_EMPTY(&par->nc_list)) {
2827 atomic_add_long(&pn->vfscache_leafs, 1);
2829 dropvp = par->nc_vp;
2831 ncp->nc_head = NULL;
2832 ncp->nc_parent = NULL;
2833 spin_unlock(&nchpp->spin);
2836 KKASSERT(ncp->nc_head == NULL);
2840 * ncp should not have picked up any refs. Physically
2843 if (ncp->nc_refs != 1) {
2844 int save_refs = ncp->nc_refs;
2846 panic("cache_zap: %p bad refs %d (%d)\n",
2847 ncp, save_refs, atomic_fetchadd_int(&ncp->nc_refs, 0));
2849 KKASSERT(ncp->nc_refs == 1);
2850 /* _cache_unlock(ncp) not required */
2851 ncp->nc_refs = -1; /* safety */
2853 kfree(ncp->nc_name, M_VFSCACHE);
2854 kfree(ncp, M_VFSCACHE);
2857 * Delayed drop (we had to release our spinlocks)
2859 * The refed parent (if not NULL) must be dropped. The
2860 * caller is responsible for looping.
2868 * Clean up dangling negative cache and defered-drop entries in the
2871 * This routine is called in the critical path and also called from
2872 * vnlru(). When called from vnlru we use a lower limit to try to
2873 * deal with the negative cache before the critical path has to start
2876 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t;
2878 static cache_hs_t neg_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW };
2879 static cache_hs_t pos_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW };
2882 cache_hysteresis(int critpath)
2885 long neglimit = maxvnodes / ncnegfactor;
2886 long xnumcache = vfscache_leafs;
2889 neglimit = neglimit * 8 / 10;
2892 * Don't cache too many negative hits. We use hysteresis to reduce
2893 * the impact on the critical path.
2895 switch(neg_cache_hysteresis_state[critpath]) {
2897 if (vfscache_negs > MINNEG && vfscache_negs > neglimit) {
2899 _cache_cleanneg(ncnegflush);
2901 _cache_cleanneg(ncnegflush +
2902 vfscache_negs - neglimit);
2903 neg_cache_hysteresis_state[critpath] = CHI_HIGH;
2907 if (vfscache_negs > MINNEG * 9 / 10 &&
2908 vfscache_negs * 9 / 10 > neglimit
2911 _cache_cleanneg(ncnegflush);
2913 _cache_cleanneg(ncnegflush +
2914 vfscache_negs * 9 / 10 -
2917 neg_cache_hysteresis_state[critpath] = CHI_LOW;
2923 * Don't cache too many positive hits. We use hysteresis to reduce
2924 * the impact on the critical path.
2926 * Excessive positive hits can accumulate due to large numbers of
2927 * hardlinks (the vnode cache will not prevent hl ncps from growing
2930 if ((poslimit = ncposlimit) == 0)
2931 poslimit = maxvnodes * 2;
2933 poslimit = poslimit * 8 / 10;
2935 switch(pos_cache_hysteresis_state[critpath]) {
2937 if (xnumcache > poslimit && xnumcache > MINPOS) {
2939 _cache_cleanpos(ncposflush);
2941 _cache_cleanpos(ncposflush +
2942 xnumcache - poslimit);
2943 pos_cache_hysteresis_state[critpath] = CHI_HIGH;
2947 if (xnumcache > poslimit * 5 / 6 && xnumcache > MINPOS) {
2949 _cache_cleanpos(ncposflush);
2951 _cache_cleanpos(ncposflush +
2952 xnumcache - poslimit * 5 / 6);
2954 pos_cache_hysteresis_state[critpath] = CHI_LOW;
2960 * Clean out dangling defered-zap ncps which could not
2961 * be cleanly dropped if too many build up. Note
2962 * that numdefered is not an exact number as such ncps
2963 * can be reused and the counter is not handled in a MP
2964 * safe manner by design.
2966 if (numdefered > neglimit) {
2967 _cache_cleandefered();
2972 * NEW NAMECACHE LOOKUP API
2974 * Lookup an entry in the namecache. The passed par_nch must be referenced
2975 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp
2976 * is ALWAYS returned, eve if the supplied component is illegal.
2978 * The resulting namecache entry should be returned to the system with
2979 * cache_put() or cache_unlock() + cache_drop().
2981 * namecache locks are recursive but care must be taken to avoid lock order
2982 * reversals (hence why the passed par_nch must be unlocked). Locking
2983 * rules are to order for parent traversals, not for child traversals.
2985 * Nobody else will be able to manipulate the associated namespace (e.g.
2986 * create, delete, rename, rename-target) until the caller unlocks the
2989 * The returned entry will be in one of three states: positive hit (non-null
2990 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
2991 * Unresolved entries must be resolved through the filesystem to associate the
2992 * vnode and/or determine whether a positive or negative hit has occured.
2994 * It is not necessary to lock a directory in order to lock namespace under
2995 * that directory. In fact, it is explicitly not allowed to do that. A
2996 * directory is typically only locked when being created, renamed, or
2999 * The directory (par) may be unresolved, in which case any returned child
3000 * will likely also be marked unresolved. Likely but not guarenteed. Since
3001 * the filesystem lookup requires a resolved directory vnode the caller is
3002 * responsible for resolving the namecache chain top-down. This API
3003 * specifically allows whole chains to be created in an unresolved state.
3006 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
3008 struct nchandle nch;
3009 struct namecache *ncp;
3010 struct namecache *new_ncp;
3011 struct namecache *rep_ncp; /* reuse a destroyed ncp */
3012 struct nchash_head *nchpp;
3019 mp = par_nch->mount;
3023 * This is a good time to call it, no ncp's are locked by
3026 cache_hysteresis(1);
3029 * Try to locate an existing entry
3031 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
3032 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
3034 nchpp = NCHHASH(hash);
3038 spin_lock(&nchpp->spin);
3040 spin_lock_shared(&nchpp->spin);
3042 TAILQ_FOREACH(ncp, &nchpp->list, nc_hash) {
3044 * Break out if we find a matching entry. Note that
3045 * UNRESOLVED entries may match, but DESTROYED entries
3048 * We may be able to reuse DESTROYED entries that we come
3049 * across, even if the name does not match, as long as
3050 * nc_nlen is correct.
3052 if (ncp->nc_parent == par_nch->ncp &&
3053 ncp->nc_nlen == nlc->nlc_namelen) {
3054 if (ncp->nc_flag & NCF_DESTROYED) {
3055 if (ncp->nc_refs == 0 && rep_ncp == NULL)
3059 if (bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen))
3063 spin_unlock(&nchpp->spin);
3065 spin_unlock_shared(&nchpp->spin);
3067 _cache_unlock(par_nch->ncp);
3070 if (_cache_lock_special(ncp) == 0) {
3072 * Successfully locked but we must re-test
3073 * conditions that might have changed since
3074 * we did not have the lock before.
3076 if (ncp->nc_parent != par_nch->ncp ||
3077 ncp->nc_nlen != nlc->nlc_namelen ||
3078 bcmp(ncp->nc_name, nlc->nlc_nameptr,
3080 (ncp->nc_flag & NCF_DESTROYED)) {
3084 _cache_auto_unresolve(mp, ncp);
3086 _cache_free(new_ncp);
3089 _cache_get(ncp); /* cycle the lock to block */
3097 * We failed to locate the entry, try to resurrect a destroyed
3098 * entry that we did find that is already correctly linked into
3099 * nchpp and the parent. We must re-test conditions after
3100 * successfully locking rep_ncp.
3102 * This case can occur under heavy loads due to not being able
3103 * to safely lock the parent in cache_zap(). Nominally a repeated
3104 * create/unlink load, but only the namelen needs to match.
3106 if (rep_ncp && new_ncp == NULL) {
3107 if (_cache_lock_nonblock(rep_ncp) == 0) {
3108 _cache_hold(rep_ncp);
3109 if (rep_ncp->nc_parent == par_nch->ncp &&
3110 rep_ncp->nc_nlen == nlc->nlc_namelen &&
3111 (rep_ncp->nc_flag & NCF_DESTROYED)) {
3113 * Update nc_name as reuse as new.
3116 bcopy(nlc->nlc_nameptr, ncp->nc_name,
3118 spin_unlock_shared(&nchpp->spin);
3119 _cache_setunresolved(ncp);
3120 ncp->nc_flag = NCF_UNRESOLVED;
3121 ncp->nc_error = ENOTCONN;
3124 _cache_put(rep_ncp);
3129 * Otherwise create a new entry and add it to the cache. The parent
3130 * ncp must also be locked so we can link into it.
3132 * We have to relookup after possibly blocking in kmalloc or
3133 * when locking par_nch.
3135 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
3136 * mount case, in which case nc_name will be NULL.
3138 if (new_ncp == NULL) {
3139 spin_unlock_shared(&nchpp->spin);
3140 new_ncp = cache_alloc(nlc->nlc_namelen);
3141 if (nlc->nlc_namelen) {
3142 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
3144 new_ncp->nc_name[nlc->nlc_namelen] = 0;
3150 * NOTE! The spinlock is held exclusively here because new_ncp
3153 if (par_locked == 0) {
3154 spin_unlock(&nchpp->spin);
3155 _cache_lock(par_nch->ncp);
3161 * WARNING! We still hold the spinlock. We have to set the hash
3162 * table entry atomically.
3165 _cache_link_parent(ncp, par_nch->ncp, nchpp);
3166 spin_unlock(&nchpp->spin);
3167 _cache_unlock(par_nch->ncp);
3168 /* par_locked = 0 - not used */
3171 * stats and namecache size management
3173 if (ncp->nc_flag & NCF_UNRESOLVED)
3174 ++gd->gd_nchstats->ncs_miss;
3175 else if (ncp->nc_vp)
3176 ++gd->gd_nchstats->ncs_goodhits;
3178 ++gd->gd_nchstats->ncs_neghits;
3181 _cache_mntref(nch.mount);
3187 * Attempt to lookup a namecache entry and return with a shared namecache
3191 cache_nlookup_maybe_shared(struct nchandle *par_nch, struct nlcomponent *nlc,
3192 int excl, struct nchandle *res_nch)
3194 struct namecache *ncp;
3195 struct nchash_head *nchpp;
3201 * If exclusive requested or shared namecache locks are disabled,
3204 if (ncp_shared_lock_disable || excl)
3205 return(EWOULDBLOCK);
3208 mp = par_nch->mount;
3211 * This is a good time to call it, no ncp's are locked by
3214 cache_hysteresis(1);
3217 * Try to locate an existing entry
3219 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
3220 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
3221 nchpp = NCHHASH(hash);
3223 spin_lock_shared(&nchpp->spin);
3225 TAILQ_FOREACH(ncp, &nchpp->list, nc_hash) {
3227 * Break out if we find a matching entry. Note that
3228 * UNRESOLVED entries may match, but DESTROYED entries
3231 if (ncp->nc_parent == par_nch->ncp &&
3232 ncp->nc_nlen == nlc->nlc_namelen &&
3233 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
3234 (ncp->nc_flag & NCF_DESTROYED) == 0
3237 spin_unlock_shared(&nchpp->spin);
3238 if (_cache_lock_shared_special(ncp) == 0) {
3239 if (ncp->nc_parent == par_nch->ncp &&
3240 ncp->nc_nlen == nlc->nlc_namelen &&
3241 bcmp(ncp->nc_name, nlc->nlc_nameptr,
3242 ncp->nc_nlen) == 0 &&
3243 (ncp->nc_flag & NCF_DESTROYED) == 0 &&
3244 (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
3245 _cache_auto_unresolve_test(mp, ncp) == 0) {
3251 spin_lock_shared(&nchpp->spin);
3259 spin_unlock_shared(&nchpp->spin);
3260 return(EWOULDBLOCK);
3265 * Note that nc_error might be non-zero (e.g ENOENT).
3268 res_nch->mount = mp;
3270 ++gd->gd_nchstats->ncs_goodhits;
3271 _cache_mntref(res_nch->mount);
3273 KKASSERT(ncp->nc_error != EWOULDBLOCK);
3274 return(ncp->nc_error);
3278 * This is a non-blocking verison of cache_nlookup() used by
3279 * nfs_readdirplusrpc_uio(). It can fail for any reason and
3280 * will return nch.ncp == NULL in that case.
3283 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc)
3285 struct nchandle nch;
3286 struct namecache *ncp;
3287 struct namecache *new_ncp;
3288 struct nchash_head *nchpp;
3295 mp = par_nch->mount;
3299 * Try to locate an existing entry
3301 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
3302 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
3304 nchpp = NCHHASH(hash);
3306 spin_lock(&nchpp->spin);
3307 TAILQ_FOREACH(ncp, &nchpp->list, nc_hash) {
3309 * Break out if we find a matching entry. Note that
3310 * UNRESOLVED entries may match, but DESTROYED entries
3313 if (ncp->nc_parent == par_nch->ncp &&
3314 ncp->nc_nlen == nlc->nlc_namelen &&
3315 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
3316 (ncp->nc_flag & NCF_DESTROYED) == 0
3319 spin_unlock(&nchpp->spin);
3321 _cache_unlock(par_nch->ncp);
3324 if (_cache_lock_special(ncp) == 0) {
3325 if (ncp->nc_parent != par_nch->ncp ||
3326 ncp->nc_nlen != nlc->nlc_namelen ||
3327 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) ||
3328 (ncp->nc_flag & NCF_DESTROYED)) {
3329 kprintf("cache_lookup_nonblock: "
3330 "ncp-race %p %*.*s\n",
3339 _cache_auto_unresolve(mp, ncp);
3341 _cache_free(new_ncp);
3352 * We failed to locate an entry, create a new entry and add it to
3353 * the cache. The parent ncp must also be locked so we
3356 * We have to relookup after possibly blocking in kmalloc or
3357 * when locking par_nch.
3359 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
3360 * mount case, in which case nc_name will be NULL.
3362 if (new_ncp == NULL) {
3363 spin_unlock(&nchpp->spin);
3364 new_ncp = cache_alloc(nlc->nlc_namelen);
3365 if (nlc->nlc_namelen) {
3366 bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
3368 new_ncp->nc_name[nlc->nlc_namelen] = 0;
3372 if (par_locked == 0) {
3373 spin_unlock(&nchpp->spin);
3374 if (_cache_lock_nonblock(par_nch->ncp) == 0) {
3382 * WARNING! We still hold the spinlock. We have to set the hash
3383 * table entry atomically.
3386 _cache_link_parent(ncp, par_nch->ncp, nchpp);
3387 spin_unlock(&nchpp->spin);
3388 _cache_unlock(par_nch->ncp);
3389 /* par_locked = 0 - not used */
3392 * stats and namecache size management
3394 if (ncp->nc_flag & NCF_UNRESOLVED)
3395 ++gd->gd_nchstats->ncs_miss;
3396 else if (ncp->nc_vp)
3397 ++gd->gd_nchstats->ncs_goodhits;
3399 ++gd->gd_nchstats->ncs_neghits;
3402 _cache_mntref(nch.mount);
3407 _cache_free(new_ncp);
3416 * The namecache entry is marked as being used as a mount point.
3417 * Locate the mount if it is visible to the caller. The DragonFly
3418 * mount system allows arbitrary loops in the topology and disentangles
3419 * those loops by matching against (mp, ncp) rather than just (ncp).
3420 * This means any given ncp can dive any number of mounts, depending
3421 * on the relative mount (e.g. nullfs) the caller is at in the topology.
3423 * We use a very simple frontend cache to reduce SMP conflicts,
3424 * which we have to do because the mountlist scan needs an exclusive
3425 * lock around its ripout info list. Not to mention that there might
3426 * be a lot of mounts.
3428 struct findmount_info {
3429 struct mount *result;
3430 struct mount *nch_mount;
3431 struct namecache *nch_ncp;
3435 struct ncmount_cache *
3436 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp)
3440 hash = (uintptr_t)mp + ((uintptr_t)mp >> 18);
3441 hash += (uintptr_t)ncp + ((uintptr_t)ncp >> 16);
3442 hash = (hash >> 1) % NCMOUNT_NUMCACHE;
3444 return (&ncmount_cache[hash]);
3449 cache_findmount_callback(struct mount *mp, void *data)
3451 struct findmount_info *info = data;
3454 * Check the mount's mounted-on point against the passed nch.
3456 if (mp->mnt_ncmounton.mount == info->nch_mount &&
3457 mp->mnt_ncmounton.ncp == info->nch_ncp
3467 cache_findmount(struct nchandle *nch)
3469 struct findmount_info info;
3470 struct ncmount_cache *ncc;
3476 if (ncmount_cache_enable == 0) {
3480 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3481 if (ncc->ncp == nch->ncp) {
3482 spin_lock_shared(&ncc->spin);
3483 if (ncc->isneg == 0 &&
3484 ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) {
3485 if (mp->mnt_ncmounton.mount == nch->mount &&
3486 mp->mnt_ncmounton.ncp == nch->ncp) {
3488 * Cache hit (positive)
3491 spin_unlock_shared(&ncc->spin);
3494 /* else cache miss */
3497 ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3499 * Cache hit (negative)
3501 spin_unlock_shared(&ncc->spin);
3504 spin_unlock_shared(&ncc->spin);
3512 info.nch_mount = nch->mount;
3513 info.nch_ncp = nch->ncp;
3514 mountlist_scan(cache_findmount_callback, &info,
3515 MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
3520 * Negative lookups: We cache the originating {ncp,mp}. (mp) is
3521 * only used for pointer comparisons and is not
3522 * referenced (otherwise there would be dangling
3525 * Positive lookups: We cache the originating {ncp} and the target
3526 * (mp). (mp) is referenced.
3528 * Indeterminant: If the match is undergoing an unmount we do
3529 * not cache it to avoid racing cache_unmounting(),
3530 * but still return the match.
3533 spin_lock(&ncc->spin);
3534 if (info.result == NULL) {
3535 if (ncc->isneg == 0 && ncc->mp)
3536 _cache_mntrel(ncc->mp);
3537 ncc->ncp = nch->ncp;
3538 ncc->mp = nch->mount;
3540 spin_unlock(&ncc->spin);
3541 } else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) {
3542 if (ncc->isneg == 0 && ncc->mp)
3543 _cache_mntrel(ncc->mp);
3544 _cache_mntref(info.result);
3545 ncc->ncp = nch->ncp;
3546 ncc->mp = info.result;
3548 spin_unlock(&ncc->spin);
3550 spin_unlock(&ncc->spin);
3553 return(info.result);
3557 cache_dropmount(struct mount *mp)
3563 cache_ismounting(struct mount *mp)
3565 struct nchandle *nch = &mp->mnt_ncmounton;
3566 struct ncmount_cache *ncc;
3568 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3570 ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3571 spin_lock(&ncc->spin);
3573 ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3577 spin_unlock(&ncc->spin);
3582 cache_unmounting(struct mount *mp)
3584 struct nchandle *nch = &mp->mnt_ncmounton;
3585 struct ncmount_cache *ncc;
3587 ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3588 if (ncc->isneg == 0 &&
3589 ncc->ncp == nch->ncp && ncc->mp == mp) {
3590 spin_lock(&ncc->spin);
3591 if (ncc->isneg == 0 &&
3592 ncc->ncp == nch->ncp && ncc->mp == mp) {
3597 spin_unlock(&ncc->spin);
3602 * Resolve an unresolved namecache entry, generally by looking it up.
3603 * The passed ncp must be locked and refd.
3605 * Theoretically since a vnode cannot be recycled while held, and since
3606 * the nc_parent chain holds its vnode as long as children exist, the
3607 * direct parent of the cache entry we are trying to resolve should
3608 * have a valid vnode. If not then generate an error that we can
3609 * determine is related to a resolver bug.
3611 * However, if a vnode was in the middle of a recyclement when the NCP
3612 * got locked, ncp->nc_vp might point to a vnode that is about to become
3613 * invalid. cache_resolve() handles this case by unresolving the entry
3614 * and then re-resolving it.
3616 * Note that successful resolution does not necessarily return an error
3617 * code of 0. If the ncp resolves to a negative cache hit then ENOENT
3621 cache_resolve(struct nchandle *nch, struct ucred *cred)
3623 struct namecache *par_tmp;
3624 struct namecache *par;
3625 struct namecache *ncp;
3626 struct nchandle nctmp;
3633 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
3636 * If the ncp is already resolved we have nothing to do. However,
3637 * we do want to guarentee that a usable vnode is returned when
3638 * a vnode is present, so make sure it hasn't been reclaimed.
3640 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3641 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
3642 _cache_setunresolved(ncp);
3643 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
3644 return (ncp->nc_error);
3648 * If the ncp was destroyed it will never resolve again. This
3649 * can basically only happen when someone is chdir'd into an
3650 * empty directory which is then rmdir'd. We want to catch this
3651 * here and not dive the VFS because the VFS might actually
3652 * have a way to re-resolve the disconnected ncp, which will
3653 * result in inconsistencies in the cdir/nch for proc->p_fd.
3655 if (ncp->nc_flag & NCF_DESTROYED)
3659 * Mount points need special handling because the parent does not
3660 * belong to the same filesystem as the ncp.
3662 if (ncp == mp->mnt_ncmountpt.ncp)
3663 return (cache_resolve_mp(mp));
3666 * We expect an unbroken chain of ncps to at least the mount point,
3667 * and even all the way to root (but this code doesn't have to go
3668 * past the mount point).
3670 if (ncp->nc_parent == NULL) {
3671 kprintf("EXDEV case 1 %p %*.*s\n", ncp,
3672 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
3673 ncp->nc_error = EXDEV;
3674 return(ncp->nc_error);
3678 * The vp's of the parent directories in the chain are held via vhold()
3679 * due to the existance of the child, and should not disappear.
3680 * However, there are cases where they can disappear:
3682 * - due to filesystem I/O errors.
3683 * - due to NFS being stupid about tracking the namespace and
3684 * destroys the namespace for entire directories quite often.
3685 * - due to forced unmounts.
3686 * - due to an rmdir (parent will be marked DESTROYED)
3688 * When this occurs we have to track the chain backwards and resolve
3689 * it, looping until the resolver catches up to the current node. We
3690 * could recurse here but we might run ourselves out of kernel stack
3691 * so we do it in a more painful manner. This situation really should
3692 * not occur all that often, or if it does not have to go back too
3693 * many nodes to resolve the ncp.
3695 while ((dvp = cache_dvpref(ncp)) == NULL) {
3697 * This case can occur if a process is CD'd into a
3698 * directory which is then rmdir'd. If the parent is marked
3699 * destroyed there is no point trying to resolve it.
3701 if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
3703 par = ncp->nc_parent;
3706 while ((par_tmp = par->nc_parent) != NULL &&
3707 par_tmp->nc_vp == NULL) {
3708 _cache_hold(par_tmp);
3709 _cache_lock(par_tmp);
3713 if (par->nc_parent == NULL) {
3714 kprintf("EXDEV case 2 %*.*s\n",
3715 par->nc_nlen, par->nc_nlen, par->nc_name);
3720 * The parent is not set in stone, ref and lock it to prevent
3721 * it from disappearing. Also note that due to renames it
3722 * is possible for our ncp to move and for par to no longer
3723 * be one of its parents. We resolve it anyway, the loop
3724 * will handle any moves.
3726 _cache_get(par); /* additional hold/lock */
3727 _cache_put(par); /* from earlier hold/lock */
3728 if (par == nch->mount->mnt_ncmountpt.ncp) {
3729 cache_resolve_mp(nch->mount);
3730 } else if ((dvp = cache_dvpref(par)) == NULL) {
3731 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
3735 if (par->nc_flag & NCF_UNRESOLVED) {
3738 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
3742 if ((error = par->nc_error) != 0) {
3743 if (par->nc_error != EAGAIN) {
3744 kprintf("EXDEV case 3 %*.*s error %d\n",
3745 par->nc_nlen, par->nc_nlen, par->nc_name,
3750 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
3751 par, par->nc_nlen, par->nc_nlen, par->nc_name);
3758 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
3759 * ncp's and reattach them. If this occurs the original ncp is marked
3760 * EAGAIN to force a relookup.
3762 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
3763 * ncp must already be resolved.
3768 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
3771 ncp->nc_error = EPERM;
3773 if (ncp->nc_error == EAGAIN) {
3774 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
3775 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
3778 return(ncp->nc_error);
3782 * Resolve the ncp associated with a mount point. Such ncp's almost always
3783 * remain resolved and this routine is rarely called. NFS MPs tends to force
3784 * re-resolution more often due to its mac-truck-smash-the-namecache
3785 * method of tracking namespace changes.
3787 * The semantics for this call is that the passed ncp must be locked on
3788 * entry and will be locked on return. However, if we actually have to
3789 * resolve the mount point we temporarily unlock the entry in order to
3790 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of
3791 * the unlock we have to recheck the flags after we relock.
3794 cache_resolve_mp(struct mount *mp)
3796 struct namecache *ncp = mp->mnt_ncmountpt.ncp;
3800 KKASSERT(mp != NULL);
3803 * If the ncp is already resolved we have nothing to do. However,
3804 * we do want to guarentee that a usable vnode is returned when
3805 * a vnode is present, so make sure it hasn't been reclaimed.
3807 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3808 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
3809 _cache_setunresolved(ncp);
3812 if (ncp->nc_flag & NCF_UNRESOLVED) {
3814 while (vfs_busy(mp, 0))
3816 error = VFS_ROOT(mp, &vp);
3820 * recheck the ncp state after relocking.
3822 if (ncp->nc_flag & NCF_UNRESOLVED) {
3823 ncp->nc_error = error;
3825 _cache_setvp(mp, ncp, vp);
3828 kprintf("[diagnostic] cache_resolve_mp: failed"
3829 " to resolve mount %p err=%d ncp=%p\n",
3831 _cache_setvp(mp, ncp, NULL);
3833 } else if (error == 0) {
3838 return(ncp->nc_error);
3842 * Clean out negative cache entries when too many have accumulated.
3845 _cache_cleanneg(long count)
3847 struct pcpu_ncache *pn;
3848 struct namecache *ncp;
3849 static uint32_t neg_rover;
3853 n = neg_rover++; /* SMP heuristical, race ok */
3855 n = n % (uint32_t)ncpus;
3858 * Normalize vfscache_negs and count. count is sometimes based
3859 * on vfscache_negs. vfscache_negs is heuristical and can sometimes
3860 * have crazy values.
3862 vnegs = vfscache_negs;
3864 if (vnegs <= MINNEG)
3869 pn = &pcpu_ncache[n];
3870 spin_lock(&pn->neg_spin);
3871 count = pn->neg_count * count / vnegs + 1;
3872 spin_unlock(&pn->neg_spin);
3875 * Attempt to clean out the specified number of negative cache
3879 spin_lock(&pn->neg_spin);
3880 ncp = TAILQ_FIRST(&pn->neg_list);
3882 spin_unlock(&pn->neg_spin);
3885 TAILQ_REMOVE(&pn->neg_list, ncp, nc_vnode);
3886 TAILQ_INSERT_TAIL(&pn->neg_list, ncp, nc_vnode);
3888 spin_unlock(&pn->neg_spin);
3891 * This can race, so we must re-check that the ncp
3892 * is on the ncneg.list after successfully locking it.
3894 if (_cache_lock_special(ncp) == 0) {
3895 if (ncp->nc_vp == NULL &&
3896 (ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3897 ncp = cache_zap(ncp, 1);
3912 * Clean out positive cache entries when too many have accumulated.
3915 _cache_cleanpos(long count)
3917 static volatile int rover;
3918 struct nchash_head *nchpp;
3919 struct namecache *ncp;
3923 * Attempt to clean out the specified number of negative cache
3927 rover_copy = ++rover; /* MPSAFEENOUGH */
3929 nchpp = NCHHASH(rover_copy);
3931 if (TAILQ_FIRST(&nchpp->list) == NULL) {
3937 * Cycle ncp on list, ignore and do not move DUMMY
3938 * ncps. These are temporary list iterators.
3940 * We must cycle the ncp to the end of the list to
3941 * ensure that all ncp's have an equal chance of
3944 spin_lock(&nchpp->spin);
3945 ncp = TAILQ_FIRST(&nchpp->list);
3946 while (ncp && (ncp->nc_flag & NCF_DUMMY))
3947 ncp = TAILQ_NEXT(ncp, nc_hash);
3949 TAILQ_REMOVE(&nchpp->list, ncp, nc_hash);
3950 TAILQ_INSERT_TAIL(&nchpp->list, ncp, nc_hash);
3953 spin_unlock(&nchpp->spin);
3956 if (_cache_lock_special(ncp) == 0) {
3957 ncp = cache_zap(ncp, 1);
3969 * This is a kitchen sink function to clean out ncps which we
3970 * tried to zap from cache_drop() but failed because we were
3971 * unable to acquire the parent lock.
3973 * Such entries can also be removed via cache_inval_vp(), such
3974 * as when unmounting.
3977 _cache_cleandefered(void)
3979 struct nchash_head *nchpp;
3980 struct namecache *ncp;
3981 struct namecache dummy;
3985 * Create a list iterator. DUMMY indicates that this is a list
3986 * iterator, DESTROYED prevents matches by lookup functions.
3989 bzero(&dummy, sizeof(dummy));
3990 dummy.nc_flag = NCF_DESTROYED | NCF_DUMMY;
3993 for (i = 0; i <= nchash; ++i) {
3994 nchpp = &nchashtbl[i];
3996 spin_lock(&nchpp->spin);
3997 TAILQ_INSERT_HEAD(&nchpp->list, &dummy, nc_hash);
3999 while ((ncp = TAILQ_NEXT(ncp, nc_hash)) != NULL) {
4000 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0)
4002 TAILQ_REMOVE(&nchpp->list, &dummy, nc_hash);
4003 TAILQ_INSERT_AFTER(&nchpp->list, ncp, &dummy, nc_hash);
4005 spin_unlock(&nchpp->spin);
4006 if (_cache_lock_nonblock(ncp) == 0) {
4007 ncp->nc_flag &= ~NCF_DEFEREDZAP;
4011 spin_lock(&nchpp->spin);
4014 TAILQ_REMOVE(&nchpp->list, &dummy, nc_hash);
4015 spin_unlock(&nchpp->spin);
4020 * Name cache initialization, from vfsinit() when we are booting
4025 struct pcpu_ncache *pn;
4030 * Per-cpu accounting and negative hit list
4032 pcpu_ncache = kmalloc(sizeof(*pcpu_ncache) * ncpus,
4033 M_VFSCACHE, M_WAITOK|M_ZERO);
4034 for (i = 0; i < ncpus; ++i) {
4035 pn = &pcpu_ncache[i];
4036 TAILQ_INIT(&pn->neg_list);
4037 spin_init(&pn->neg_spin, "ncneg");
4041 * Initialise per-cpu namecache effectiveness statistics.
4043 for (i = 0; i < ncpus; ++i) {
4044 gd = globaldata_find(i);
4045 gd->gd_nchstats = &nchstats[i];
4049 * Create a generous namecache hash table
4051 nchashtbl = hashinit_ext(vfs_inodehashsize(),
4052 sizeof(struct nchash_head),
4053 M_VFSCACHE, &nchash);
4054 for (i = 0; i <= (int)nchash; ++i) {
4055 TAILQ_INIT(&nchashtbl[i].list);
4056 spin_init(&nchashtbl[i].spin, "nchinit_hash");
4058 for (i = 0; i < NCMOUNT_NUMCACHE; ++i)
4059 spin_init(&ncmount_cache[i].spin, "nchinit_cache");
4060 nclockwarn = 5 * hz;
4064 * Called from start_init() to bootstrap the root filesystem. Returns
4065 * a referenced, unlocked namecache record.
4068 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
4070 nch->ncp = cache_alloc(0);
4074 _cache_setvp(nch->mount, nch->ncp, vp);
4078 * vfs_cache_setroot()
4080 * Create an association between the root of our namecache and
4081 * the root vnode. This routine may be called several times during
4084 * If the caller intends to save the returned namecache pointer somewhere
4085 * it must cache_hold() it.
4088 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
4091 struct nchandle onch;
4099 cache_zero(&rootnch);
4107 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache
4108 * topology and is being removed as quickly as possible. The new VOP_N*()
4109 * API calls are required to make specific adjustments using the supplied
4110 * ncp pointers rather then just bogusly purging random vnodes.
4112 * Invalidate all namecache entries to a particular vnode as well as
4113 * any direct children of that vnode in the namecache. This is a
4114 * 'catch all' purge used by filesystems that do not know any better.
4116 * Note that the linkage between the vnode and its namecache entries will
4117 * be removed, but the namecache entries themselves might stay put due to
4118 * active references from elsewhere in the system or due to the existance of
4119 * the children. The namecache topology is left intact even if we do not
4120 * know what the vnode association is. Such entries will be marked
4124 cache_purge(struct vnode *vp)
4126 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
4129 static int disablecwd;
4130 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
4133 static u_long numcwdcalls;
4134 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0,
4135 "Number of current directory resolution calls");
4136 static u_long numcwdfailnf;
4137 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0,
4138 "Number of current directory failures due to lack of file");
4139 static u_long numcwdfailsz;
4140 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0,
4141 "Number of current directory failures due to large result");
4142 static u_long numcwdfound;
4143 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0,
4144 "Number of current directory resolution successes");
4150 sys___getcwd(struct __getcwd_args *uap)
4160 buflen = uap->buflen;
4163 if (buflen > MAXPATHLEN)
4164 buflen = MAXPATHLEN;
4166 buf = kmalloc(buflen, M_TEMP, M_WAITOK);
4167 bp = kern_getcwd(buf, buflen, &error);
4169 error = copyout(bp, uap->buf, strlen(bp) + 1);
4175 kern_getcwd(char *buf, size_t buflen, int *error)
4177 struct proc *p = curproc;
4179 int i, slash_prefixed;
4180 struct filedesc *fdp;
4181 struct nchandle nch;
4182 struct namecache *ncp;
4191 nch = fdp->fd_ncdir;
4196 while (ncp && (ncp != fdp->fd_nrdir.ncp ||
4197 nch.mount != fdp->fd_nrdir.mount)
4200 * While traversing upwards if we encounter the root
4201 * of the current mount we have to skip to the mount point
4202 * in the underlying filesystem.
4204 if (ncp == nch.mount->mnt_ncmountpt.ncp) {
4205 nch = nch.mount->mnt_ncmounton;
4214 * Prepend the path segment
4216 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
4223 *--bp = ncp->nc_name[i];
4235 * Go up a directory. This isn't a mount point so we don't
4236 * have to check again.
4238 while ((nch.ncp = ncp->nc_parent) != NULL) {
4239 if (ncp_shared_lock_disable)
4242 _cache_lock_shared(ncp);
4243 if (nch.ncp != ncp->nc_parent) {
4247 _cache_hold(nch.ncp);
4260 if (!slash_prefixed) {
4278 * Thus begins the fullpath magic.
4280 * The passed nchp is referenced but not locked.
4282 static int disablefullpath;
4283 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
4284 &disablefullpath, 0,
4285 "Disable fullpath lookups");
4288 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase,
4289 char **retbuf, char **freebuf, int guess)
4291 struct nchandle fd_nrdir;
4292 struct nchandle nch;
4293 struct namecache *ncp;
4294 struct mount *mp, *new_mp;
4303 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
4304 bp = buf + MAXPATHLEN - 1;
4307 fd_nrdir = *nchbase;
4309 fd_nrdir = p->p_fd->fd_nrdir;
4319 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) {
4323 * If we are asked to guess the upwards path, we do so whenever
4324 * we encounter an ncp marked as a mountpoint. We try to find
4325 * the actual mountpoint by finding the mountpoint with this
4328 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) {
4329 new_mp = mount_get_by_nc(ncp);
4332 * While traversing upwards if we encounter the root
4333 * of the current mount we have to skip to the mount point.
4335 if (ncp == mp->mnt_ncmountpt.ncp) {
4339 nch = new_mp->mnt_ncmounton;
4349 * Prepend the path segment
4351 for (i = ncp->nc_nlen - 1; i >= 0; i--) {
4357 *--bp = ncp->nc_name[i];
4368 * Go up a directory. This isn't a mount point so we don't
4369 * have to check again.
4371 * We can only safely access nc_parent with ncp held locked.
4373 while ((nch.ncp = ncp->nc_parent) != NULL) {
4375 if (nch.ncp != ncp->nc_parent) {
4379 _cache_hold(nch.ncp);
4392 if (!slash_prefixed) {
4410 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf,
4411 char **freebuf, int guess)
4413 struct namecache *ncp;
4414 struct nchandle nch;
4418 if (disablefullpath)
4424 /* vn is NULL, client wants us to use p->p_textvp */
4426 if ((vn = p->p_textvp) == NULL)
4429 spin_lock_shared(&vn->v_spin);
4430 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
4435 spin_unlock_shared(&vn->v_spin);
4439 spin_unlock_shared(&vn->v_spin);
4442 nch.mount = vn->v_mount;
4443 error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess);
4449 vfscache_rollup_cpu(struct globaldata *gd)
4451 struct pcpu_ncache *pn;
4454 if (pcpu_ncache == NULL)
4456 pn = &pcpu_ncache[gd->gd_cpuid];
4458 if (pn->vfscache_count) {
4459 count = atomic_swap_long(&pn->vfscache_count, 0);
4460 atomic_add_long(&vfscache_count, count);
4462 if (pn->vfscache_leafs) {
4463 count = atomic_swap_long(&pn->vfscache_leafs, 0);
4464 atomic_add_long(&vfscache_leafs, count);
4466 if (pn->vfscache_negs) {
4467 count = atomic_swap_long(&pn->vfscache_negs, 0);
4468 atomic_add_long(&vfscache_negs, count);
4474 vfscache_rollup_all(void)
4478 for (n = 0; n < ncpus; ++n)
4479 vfscache_rollup_cpu(globaldata_find(n));