2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1989, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * Poul-Henning Kamp of the FreeBSD Project.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
41 #include "opt_ktrace.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/capsicum.h>
46 #include <sys/counter.h>
47 #include <sys/filedesc.h>
48 #include <sys/fnv_hash.h>
49 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/fcntl.h>
55 #include <sys/mount.h>
56 #include <sys/namei.h>
62 #include <sys/syscallsubr.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/vnode.h>
68 #include <sys/ktrace.h>
71 #include <sys/capsicum.h>
73 #include <security/audit/audit.h>
74 #include <security/mac/mac_framework.h>
82 SDT_PROVIDER_DECLARE(vfs);
83 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
85 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
87 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
89 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
90 "struct namecache *", "int", "int");
91 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
92 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
93 "char *", "struct vnode *");
94 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
95 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
96 "struct vnode *", "char *");
97 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
99 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
100 "struct vnode *", "char *");
101 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
103 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
104 "struct componentname *");
105 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
106 "struct componentname *");
107 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
108 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
109 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
110 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
112 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
114 SDT_PROBE_DEFINE2(vfs, namecache, shrink_negative, done, "struct vnode *",
117 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
118 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
119 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
122 * This structure describes the elements in the cache of recent
123 * names looked up by namei.
128 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
129 "the state must fit in a union with a pointer without growing it");
132 LIST_ENTRY(namecache) nc_src; /* source vnode list */
133 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */
134 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
135 struct vnode *nc_dvp; /* vnode of parent of name */
137 struct vnode *nu_vp; /* vnode the name refers to */
138 struct negstate nu_neg;/* negative entry state */
140 u_char nc_flag; /* flag bits */
141 u_char nc_nlen; /* length of name */
142 char nc_name[0]; /* segment name + nul */
146 * struct namecache_ts repeats struct namecache layout up to the
148 * struct namecache_ts is used in place of struct namecache when time(s) need
149 * to be stored. The nc_dotdottime field is used when a cache entry is mapping
150 * both a non-dotdot directory name plus dotdot for the directory's
153 * See below for alignment requirement.
155 struct namecache_ts {
156 struct timespec nc_time; /* timespec provided by fs */
157 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */
158 int nc_ticks; /* ticks value when entry was added */
159 struct namecache nc_nc;
163 * At least mips n32 performs 64-bit accesses to timespec as found
164 * in namecache_ts and requires them to be aligned. Since others
165 * may be in the same spot suffer a little bit and enforce the
166 * alignment for everyone. Note this is a nop for 64-bit platforms.
168 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t)
169 #define CACHE_PATH_CUTOFF 39
171 #define CACHE_ZONE_SMALL_SIZE (sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1)
172 #define CACHE_ZONE_SMALL_TS_SIZE (sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1)
173 #define CACHE_ZONE_LARGE_SIZE (sizeof(struct namecache) + NAME_MAX + 1)
174 #define CACHE_ZONE_LARGE_TS_SIZE (sizeof(struct namecache_ts) + NAME_MAX + 1)
176 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
177 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
178 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
179 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
181 #define nc_vp n_un.nu_vp
182 #define nc_neg n_un.nu_neg
185 * Flags in namecache.nc_flag
187 #define NCF_WHITE 0x01
188 #define NCF_ISDOTDOT 0x02
191 #define NCF_DVDROP 0x10
192 #define NCF_NEGATIVE 0x20
193 #define NCF_INVALID 0x40
197 * Flags in negstate.neg_flag
202 * Mark an entry as invalid.
204 * This is called before it starts getting deconstructed.
207 cache_ncp_invalidate(struct namecache *ncp)
210 KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
211 ("%s: entry %p already invalid", __func__, ncp));
212 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
213 atomic_thread_fence_rel();
217 * Check whether the entry can be safely used.
219 * All places which elide locks are supposed to call this after they are
220 * done with reading from an entry.
223 cache_ncp_canuse(struct namecache *ncp)
226 atomic_thread_fence_acq();
227 return ((atomic_load_char(&ncp->nc_flag) & (NCF_INVALID | NCF_WIP)) == 0);
231 * Name caching works as follows:
233 * Names found by directory scans are retained in a cache
234 * for future reference. It is managed LRU, so frequently
235 * used names will hang around. Cache is indexed by hash value
236 * obtained from (dvp, name) where dvp refers to the directory
239 * If it is a "negative" entry, (i.e. for a name that is known NOT to
240 * exist) the vnode pointer will be NULL.
242 * Upon reaching the last segment of a path, if the reference
243 * is for DELETE, or NOCACHE is set (rewrite), and the
244 * name is located in the cache, it will be dropped.
246 * These locks are used (in the order in which they can be taken):
248 * vnodelock mtx vnode lists and v_cache_dd field protection
249 * bucketlock mtx for access to given set of hash buckets
250 * neglist mtx negative entry LRU management
252 * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
253 * shrinking the LRU list.
255 * It is legal to take multiple vnodelock and bucketlock locks. The locking
256 * order is lower address first. Both are recursive.
258 * "." lookups are lockless.
260 * ".." and vnode -> name lookups require vnodelock.
262 * name -> vnode lookup requires the relevant bucketlock to be held for reading.
264 * Insertions and removals of entries require involved vnodes and bucketlocks
265 * to be locked to provide safe operation against other threads modifying the
268 * Some lookups result in removal of the found entry (e.g. getting rid of a
269 * negative entry with the intent to create a positive one), which poses a
270 * problem when multiple threads reach the state. Similarly, two different
271 * threads can purge two different vnodes and try to remove the same name.
273 * If the already held vnode lock is lower than the second required lock, we
274 * can just take the other lock. However, in the opposite case, this could
275 * deadlock. As such, this is resolved by trylocking and if that fails unlocking
276 * the first node, locking everything in order and revalidating the state.
282 * Structures associated with name caching.
284 #define NCHHASH(hash) \
285 (&nchashtbl[(hash) & nchash])
286 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
287 static u_long __read_mostly nchash; /* size of hash table */
288 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
289 "Size of namecache hash table");
290 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */
291 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
292 "Ratio of negative namecache entries");
293 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */
294 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */
295 u_int ncsizefactor = 2;
296 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
297 "Size factor for namecache");
298 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */
300 struct nchstats nchstats; /* cache effectiveness statistics */
302 static bool __read_frequently cache_fast_revlookup = true;
303 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
304 &cache_fast_revlookup, 0, "");
306 static struct mtx __exclusive_cache_line ncneg_shrink_lock;
310 TAILQ_HEAD(, namecache) nl_list;
311 } __aligned(CACHE_LINE_SIZE);
313 static struct neglist __read_mostly *neglists;
314 static struct neglist ncneg_hot;
315 static u_long numhotneg;
318 #define numneglists (ncneghash + 1)
319 static inline struct neglist *
320 NCP2NEGLIST(struct namecache *ncp)
323 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
326 static inline struct negstate *
327 NCP2NEGSTATE(struct namecache *ncp)
330 MPASS(ncp->nc_flag & NCF_NEGATIVE);
331 return (&ncp->nc_neg);
334 #define numbucketlocks (ncbuckethash + 1)
335 static u_int __read_mostly ncbuckethash;
336 static struct mtx_padalign __read_mostly *bucketlocks;
337 #define HASH2BUCKETLOCK(hash) \
338 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
340 #define numvnodelocks (ncvnodehash + 1)
341 static u_int __read_mostly ncvnodehash;
342 static struct mtx __read_mostly *vnodelocks;
343 static inline struct mtx *
344 VP2VNODELOCK(struct vnode *vp)
347 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
351 * UMA zones for the VFS cache.
353 * The small cache is used for entries with short names, which are the
354 * most common. The large cache is used for entries which are too big to
355 * fit in the small cache.
357 static uma_zone_t __read_mostly cache_zone_small;
358 static uma_zone_t __read_mostly cache_zone_small_ts;
359 static uma_zone_t __read_mostly cache_zone_large;
360 static uma_zone_t __read_mostly cache_zone_large_ts;
362 static struct namecache *
363 cache_alloc(int len, int ts)
365 struct namecache_ts *ncp_ts;
366 struct namecache *ncp;
368 if (__predict_false(ts)) {
369 if (len <= CACHE_PATH_CUTOFF)
370 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
372 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
373 ncp = &ncp_ts->nc_nc;
375 if (len <= CACHE_PATH_CUTOFF)
376 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
378 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
384 cache_free(struct namecache *ncp)
386 struct namecache_ts *ncp_ts;
389 if ((ncp->nc_flag & NCF_DVDROP) != 0)
391 if (__predict_false(ncp->nc_flag & NCF_TS)) {
392 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
393 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
394 uma_zfree_smr(cache_zone_small_ts, ncp_ts);
396 uma_zfree_smr(cache_zone_large_ts, ncp_ts);
398 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
399 uma_zfree_smr(cache_zone_small, ncp);
401 uma_zfree_smr(cache_zone_large, ncp);
406 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
408 struct namecache_ts *ncp_ts;
410 KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
411 (tsp == NULL && ticksp == NULL),
417 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
418 *tsp = ncp_ts->nc_time;
419 *ticksp = ncp_ts->nc_ticks;
423 static int __read_mostly doingcache = 1; /* 1 => enable the cache */
424 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
425 "VFS namecache enabled");
428 /* Export size information to userland */
429 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
430 sizeof(struct namecache), "sizeof(struct namecache)");
433 * The new name cache statistics
435 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
436 "Name cache statistics");
437 #define STATNODE_ULONG(name, descr) \
438 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
439 #define STATNODE_COUNTER(name, descr) \
440 static COUNTER_U64_DEFINE_EARLY(name); \
441 SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, \
443 STATNODE_ULONG(numneg, "Number of negative cache entries");
444 STATNODE_ULONG(numcache, "Number of cache entries");
445 STATNODE_COUNTER(numcachehv, "Number of namecache entries with vnodes held");
446 STATNODE_COUNTER(numdrops, "Number of dropped entries due to reaching the limit");
447 STATNODE_COUNTER(dothits, "Number of '.' hits");
448 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
449 STATNODE_COUNTER(nummiss, "Number of cache misses");
450 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
451 STATNODE_COUNTER(numposzaps,
452 "Number of cache hits (positive) we do not want to cache");
453 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
454 STATNODE_COUNTER(numnegzaps,
455 "Number of cache hits (negative) we do not want to cache");
456 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
457 /* These count for vn_getcwd(), too. */
458 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
459 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
460 STATNODE_COUNTER(numfullpathfail2,
461 "Number of fullpath search errors (VOP_VPTOCNP failures)");
462 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
463 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
464 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
465 "Number of successful removals after relocking");
466 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
467 "Number of times zap_and_exit failed to lock");
468 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
469 "Number of times zap_and_exit failed to lock");
470 static long cache_lock_vnodes_cel_3_failures;
471 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
472 "Number of times 3-way vnode locking failed");
473 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
474 STATNODE_COUNTER(numneg_evicted,
475 "Number of negative entries evicted when adding a new entry");
476 STATNODE_COUNTER(shrinking_skipped,
477 "Number of times shrinking was already in progress");
479 static void cache_zap_locked(struct namecache *ncp);
480 static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf,
481 char **freebuf, size_t *buflen);
482 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
483 char **retbuf, size_t *buflen, bool slash_prefixed, size_t addend);
484 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
485 char **retbuf, size_t *buflen);
486 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
487 char **retbuf, size_t *len, bool slash_prefixed, size_t addend);
489 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
492 cache_assert_vlp_locked(struct mtx *vlp)
496 mtx_assert(vlp, MA_OWNED);
500 cache_assert_vnode_locked(struct vnode *vp)
504 vlp = VP2VNODELOCK(vp);
505 cache_assert_vlp_locked(vlp);
509 * TODO: With the value stored we can do better than computing the hash based
510 * on the address. The choice of FNV should also be revisited.
513 cache_prehash(struct vnode *vp)
516 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
520 cache_get_hash(char *name, u_char len, struct vnode *dvp)
523 return (fnv_32_buf(name, len, dvp->v_nchash));
526 static inline struct nchashhead *
527 NCP2BUCKET(struct namecache *ncp)
531 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
532 return (NCHHASH(hash));
535 static inline struct mtx *
536 NCP2BUCKETLOCK(struct namecache *ncp)
540 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
541 return (HASH2BUCKETLOCK(hash));
546 cache_assert_bucket_locked(struct namecache *ncp)
550 blp = NCP2BUCKETLOCK(ncp);
551 mtx_assert(blp, MA_OWNED);
555 cache_assert_bucket_unlocked(struct namecache *ncp)
559 blp = NCP2BUCKETLOCK(ncp);
560 mtx_assert(blp, MA_NOTOWNED);
563 #define cache_assert_bucket_locked(x) do { } while (0)
564 #define cache_assert_bucket_unlocked(x) do { } while (0)
567 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y))
569 _cache_sort_vnodes(void **p1, void **p2)
573 MPASS(*p1 != NULL || *p2 != NULL);
583 cache_lock_all_buckets(void)
587 for (i = 0; i < numbucketlocks; i++)
588 mtx_lock(&bucketlocks[i]);
592 cache_unlock_all_buckets(void)
596 for (i = 0; i < numbucketlocks; i++)
597 mtx_unlock(&bucketlocks[i]);
601 cache_lock_all_vnodes(void)
605 for (i = 0; i < numvnodelocks; i++)
606 mtx_lock(&vnodelocks[i]);
610 cache_unlock_all_vnodes(void)
614 for (i = 0; i < numvnodelocks; i++)
615 mtx_unlock(&vnodelocks[i]);
619 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
622 cache_sort_vnodes(&vlp1, &vlp2);
625 if (!mtx_trylock(vlp1))
628 if (!mtx_trylock(vlp2)) {
638 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
641 MPASS(vlp1 != NULL || vlp2 != NULL);
651 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
654 MPASS(vlp1 != NULL || vlp2 != NULL);
663 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
665 struct nchstats snap;
667 if (req->oldptr == NULL)
668 return (SYSCTL_OUT(req, 0, sizeof(snap)));
671 snap.ncs_goodhits = counter_u64_fetch(numposhits);
672 snap.ncs_neghits = counter_u64_fetch(numneghits);
673 snap.ncs_badhits = counter_u64_fetch(numposzaps) +
674 counter_u64_fetch(numnegzaps);
675 snap.ncs_miss = counter_u64_fetch(nummisszap) +
676 counter_u64_fetch(nummiss);
678 return (SYSCTL_OUT(req, &snap, sizeof(snap)));
680 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
681 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
682 "VFS cache effectiveness statistics");
686 * Grab an atomic snapshot of the name cache hash chain lengths
688 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
689 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
693 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
695 struct nchashhead *ncpp;
696 struct namecache *ncp;
697 int i, error, n_nchash, *cntbuf;
700 n_nchash = nchash + 1; /* nchash is max index, not count */
701 if (req->oldptr == NULL)
702 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
703 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
704 cache_lock_all_buckets();
705 if (n_nchash != nchash + 1) {
706 cache_unlock_all_buckets();
707 free(cntbuf, M_TEMP);
710 /* Scan hash tables counting entries */
711 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
712 CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
714 cache_unlock_all_buckets();
715 for (error = 0, i = 0; i < n_nchash; i++)
716 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
718 free(cntbuf, M_TEMP);
721 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
722 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
723 "nchash chain lengths");
726 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
729 struct nchashhead *ncpp;
730 struct namecache *ncp;
732 int count, maxlength, used, pct;
735 return SYSCTL_OUT(req, 0, 4 * sizeof(int));
737 cache_lock_all_buckets();
738 n_nchash = nchash + 1; /* nchash is max index, not count */
742 /* Scan hash tables for applicable entries */
743 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
745 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
750 if (maxlength < count)
753 n_nchash = nchash + 1;
754 cache_unlock_all_buckets();
755 pct = (used * 100) / (n_nchash / 100);
756 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
759 error = SYSCTL_OUT(req, &used, sizeof(used));
762 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
765 error = SYSCTL_OUT(req, &pct, sizeof(pct));
770 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
771 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
772 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
776 * Negative entries management
778 * A variation of LRU scheme is used. New entries are hashed into one of
779 * numneglists cold lists. Entries get promoted to the hot list on first hit.
781 * The shrinker will demote hot list head and evict from the cold list in a
782 * round-robin manner.
785 cache_negative_init(struct namecache *ncp)
787 struct negstate *negstate;
789 ncp->nc_flag |= NCF_NEGATIVE;
790 negstate = NCP2NEGSTATE(ncp);
791 negstate->neg_flag = 0;
795 cache_negative_hit(struct namecache *ncp)
797 struct neglist *neglist;
798 struct negstate *negstate;
800 negstate = NCP2NEGSTATE(ncp);
801 if ((negstate->neg_flag & NEG_HOT) != 0)
803 neglist = NCP2NEGLIST(ncp);
804 mtx_lock(&ncneg_hot.nl_lock);
805 mtx_lock(&neglist->nl_lock);
806 if ((negstate->neg_flag & NEG_HOT) == 0) {
808 TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
809 TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
810 negstate->neg_flag |= NEG_HOT;
812 mtx_unlock(&neglist->nl_lock);
813 mtx_unlock(&ncneg_hot.nl_lock);
817 cache_negative_insert(struct namecache *ncp)
819 struct neglist *neglist;
821 MPASS(ncp->nc_flag & NCF_NEGATIVE);
822 cache_assert_bucket_locked(ncp);
823 neglist = NCP2NEGLIST(ncp);
824 mtx_lock(&neglist->nl_lock);
825 TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
826 mtx_unlock(&neglist->nl_lock);
827 atomic_add_long(&numneg, 1);
831 cache_negative_remove(struct namecache *ncp)
833 struct neglist *neglist;
834 struct negstate *negstate;
835 bool hot_locked = false;
836 bool list_locked = false;
838 cache_assert_bucket_locked(ncp);
839 neglist = NCP2NEGLIST(ncp);
840 negstate = NCP2NEGSTATE(ncp);
841 if ((negstate->neg_flag & NEG_HOT) != 0) {
843 mtx_lock(&ncneg_hot.nl_lock);
844 if ((negstate->neg_flag & NEG_HOT) == 0) {
846 mtx_lock(&neglist->nl_lock);
850 mtx_lock(&neglist->nl_lock);
852 * We may be racing against promotion in lockless lookup.
854 if ((negstate->neg_flag & NEG_HOT) != 0) {
855 mtx_unlock(&neglist->nl_lock);
857 mtx_lock(&ncneg_hot.nl_lock);
858 mtx_lock(&neglist->nl_lock);
861 if ((negstate->neg_flag & NEG_HOT) != 0) {
862 mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
863 TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
866 mtx_assert(&neglist->nl_lock, MA_OWNED);
867 TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
870 mtx_unlock(&neglist->nl_lock);
872 mtx_unlock(&ncneg_hot.nl_lock);
873 atomic_subtract_long(&numneg, 1);
877 cache_negative_shrink_select(struct namecache **ncpp,
878 struct neglist **neglistpp)
880 struct neglist *neglist;
881 struct namecache *ncp;
887 for (i = 0; i < numneglists; i++) {
888 neglist = &neglists[(cycle + i) % numneglists];
889 if (TAILQ_FIRST(&neglist->nl_list) == NULL)
891 mtx_lock(&neglist->nl_lock);
892 ncp = TAILQ_FIRST(&neglist->nl_list);
895 mtx_unlock(&neglist->nl_lock);
898 *neglistpp = neglist;
904 cache_negative_zap_one(void)
906 struct namecache *ncp, *ncp2;
907 struct neglist *neglist;
908 struct negstate *negstate;
912 if (mtx_owner(&ncneg_shrink_lock) != NULL ||
913 !mtx_trylock(&ncneg_shrink_lock)) {
914 counter_u64_add(shrinking_skipped, 1);
918 mtx_lock(&ncneg_hot.nl_lock);
919 ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
921 neglist = NCP2NEGLIST(ncp);
922 negstate = NCP2NEGSTATE(ncp);
923 mtx_lock(&neglist->nl_lock);
924 MPASS((negstate->neg_flag & NEG_HOT) != 0);
925 TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
926 TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
927 negstate->neg_flag &= ~NEG_HOT;
929 mtx_unlock(&neglist->nl_lock);
931 mtx_unlock(&ncneg_hot.nl_lock);
933 cache_negative_shrink_select(&ncp, &neglist);
935 mtx_unlock(&ncneg_shrink_lock);
939 MPASS(ncp->nc_flag & NCF_NEGATIVE);
940 dvlp = VP2VNODELOCK(ncp->nc_dvp);
941 blp = NCP2BUCKETLOCK(ncp);
942 mtx_unlock(&neglist->nl_lock);
946 * Enter SMR to safely check the negative list.
947 * Even if the found pointer matches, the entry may now be reallocated
948 * and used by a different vnode.
951 ncp2 = TAILQ_FIRST(&neglist->nl_list);
952 if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
953 blp != NCP2BUCKETLOCK(ncp2)) {
958 SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
960 cache_zap_locked(ncp);
961 counter_u64_add(numneg_evicted, 1);
970 * cache_zap_locked():
972 * Removes a namecache entry from cache, whether it contains an actual
973 * pointer to a vnode or if it is just a negative cache entry.
976 cache_zap_locked(struct namecache *ncp)
978 struct nchashhead *ncpp;
980 if (!(ncp->nc_flag & NCF_NEGATIVE))
981 cache_assert_vnode_locked(ncp->nc_vp);
982 cache_assert_vnode_locked(ncp->nc_dvp);
983 cache_assert_bucket_locked(ncp);
985 cache_ncp_invalidate(ncp);
987 ncpp = NCP2BUCKET(ncp);
988 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
989 if (!(ncp->nc_flag & NCF_NEGATIVE)) {
990 SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
991 ncp->nc_name, ncp->nc_vp);
992 TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
993 if (ncp == ncp->nc_vp->v_cache_dd) {
994 vn_seqc_write_begin_unheld(ncp->nc_vp);
995 ncp->nc_vp->v_cache_dd = NULL;
996 vn_seqc_write_end(ncp->nc_vp);
999 SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
1001 cache_negative_remove(ncp);
1003 if (ncp->nc_flag & NCF_ISDOTDOT) {
1004 if (ncp == ncp->nc_dvp->v_cache_dd) {
1005 vn_seqc_write_begin_unheld(ncp->nc_dvp);
1006 ncp->nc_dvp->v_cache_dd = NULL;
1007 vn_seqc_write_end(ncp->nc_dvp);
1010 LIST_REMOVE(ncp, nc_src);
1011 if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
1012 ncp->nc_flag |= NCF_DVDROP;
1013 counter_u64_add(numcachehv, -1);
1016 atomic_subtract_long(&numcache, 1);
1020 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1024 MPASS(ncp->nc_dvp == vp);
1025 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1026 cache_assert_vnode_locked(vp);
1028 blp = NCP2BUCKETLOCK(ncp);
1030 cache_zap_locked(ncp);
1035 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1038 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1041 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1042 cache_assert_vnode_locked(vp);
1044 if (ncp->nc_flag & NCF_NEGATIVE) {
1045 if (*vlpp != NULL) {
1049 cache_zap_negative_locked_vnode_kl(ncp, vp);
1053 pvlp = VP2VNODELOCK(vp);
1054 blp = NCP2BUCKETLOCK(ncp);
1055 vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1056 vlp2 = VP2VNODELOCK(ncp->nc_vp);
1058 if (*vlpp == vlp1 || *vlpp == vlp2) {
1062 if (*vlpp != NULL) {
1066 cache_sort_vnodes(&vlp1, &vlp2);
1071 if (!mtx_trylock(vlp1))
1077 cache_zap_locked(ncp);
1079 if (to_unlock != NULL)
1080 mtx_unlock(to_unlock);
1087 MPASS(*vlpp == NULL);
1092 static int __noinline
1093 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
1095 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1099 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1100 cache_assert_vnode_locked(vp);
1102 pvlp = VP2VNODELOCK(vp);
1103 if (ncp->nc_flag & NCF_NEGATIVE) {
1104 cache_zap_negative_locked_vnode_kl(ncp, vp);
1108 blp = NCP2BUCKETLOCK(ncp);
1109 vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1110 vlp2 = VP2VNODELOCK(ncp->nc_vp);
1111 cache_sort_vnodes(&vlp1, &vlp2);
1116 if (!mtx_trylock(vlp1)) {
1118 * TODO: Very wasteful but rare.
1130 cache_zap_locked(ncp);
1132 mtx_unlock(to_unlock);
1139 * If trylocking failed we can get here. We know enough to take all needed locks
1140 * in the right order and re-lookup the entry.
1143 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1144 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1147 struct namecache *rncp;
1149 cache_assert_bucket_unlocked(ncp);
1151 cache_sort_vnodes(&dvlp, &vlp);
1152 cache_lock_vnodes(dvlp, vlp);
1154 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1155 if (rncp == ncp && rncp->nc_dvp == dvp &&
1156 rncp->nc_nlen == cnp->cn_namelen &&
1157 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1161 cache_zap_locked(rncp);
1163 cache_unlock_vnodes(dvlp, vlp);
1164 counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1169 cache_unlock_vnodes(dvlp, vlp);
1173 static int __noinline
1174 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1175 uint32_t hash, struct mtx *blp)
1177 struct mtx *dvlp, *vlp;
1180 cache_assert_bucket_locked(ncp);
1182 dvlp = VP2VNODELOCK(ncp->nc_dvp);
1184 if (!(ncp->nc_flag & NCF_NEGATIVE))
1185 vlp = VP2VNODELOCK(ncp->nc_vp);
1186 if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1187 cache_zap_locked(ncp);
1189 cache_unlock_vnodes(dvlp, vlp);
1195 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1198 static __noinline int
1199 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1201 struct namecache *ncp;
1203 struct mtx *dvlp, *dvlp2;
1207 if (cnp->cn_namelen == 2 &&
1208 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1209 dvlp = VP2VNODELOCK(dvp);
1213 ncp = dvp->v_cache_dd;
1218 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1221 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1222 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1224 MPASS(dvp->v_cache_dd == NULL);
1230 vn_seqc_write_begin(dvp);
1231 dvp->v_cache_dd = NULL;
1232 vn_seqc_write_end(dvp);
1237 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1241 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1242 blp = HASH2BUCKETLOCK(hash);
1244 if (CK_SLIST_EMPTY(NCHHASH(hash)))
1249 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1250 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1251 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1260 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1261 if (__predict_false(error != 0)) {
1262 zap_and_exit_bucket_fail++;
1265 counter_u64_add(numposzaps, 1);
1266 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1270 counter_u64_add(nummisszap, 1);
1271 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1275 static int __noinline
1276 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1277 struct timespec *tsp, int *ticksp)
1282 counter_u64_add(dothits, 1);
1283 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1290 * When we lookup "." we still can be asked to lock it
1293 ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1294 if (ltype != VOP_ISLOCKED(*vpp)) {
1295 if (ltype == LK_EXCLUSIVE) {
1296 vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1297 if (VN_IS_DOOMED((*vpp))) {
1298 /* forced unmount */
1304 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1309 static int __noinline
1310 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1311 struct timespec *tsp, int *ticksp)
1313 struct namecache_ts *ncp_ts;
1314 struct namecache *ncp;
1320 MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1322 if ((cnp->cn_flags & MAKEENTRY) == 0) {
1323 cache_remove_cnp(dvp, cnp);
1327 counter_u64_add(dotdothits, 1);
1329 dvlp = VP2VNODELOCK(dvp);
1331 ncp = dvp->v_cache_dd;
1333 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL);
1337 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1338 if (ncp->nc_flag & NCF_NEGATIVE)
1345 goto negative_success;
1346 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1347 cache_out_ts(ncp, tsp, ticksp);
1348 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1349 NCF_DTS && tsp != NULL) {
1350 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1351 *tsp = ncp_ts->nc_dotdottime;
1355 ltype = VOP_ISLOCKED(dvp);
1357 vs = vget_prep(*vpp);
1359 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1360 vn_lock(dvp, ltype | LK_RETRY);
1361 if (VN_IS_DOOMED(dvp)) {
1373 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1374 if (cnp->cn_flags & ISLASTCN) {
1375 counter_u64_add(numnegzaps, 1);
1376 error = cache_zap_locked_vnode(ncp, dvp);
1377 if (__predict_false(error != 0)) {
1378 zap_and_exit_bucket_fail2++;
1386 SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, ncp->nc_name);
1387 cache_out_ts(ncp, tsp, ticksp);
1388 counter_u64_add(numneghits, 1);
1389 whiteout = (ncp->nc_flag & NCF_WHITE);
1390 cache_negative_hit(ncp);
1393 cnp->cn_flags |= ISWHITEOUT;
1398 * Lookup a name in the name cache
1402 * - dvp: Parent directory in which to search.
1403 * - vpp: Return argument. Will contain desired vnode on cache hit.
1404 * - cnp: Parameters of the name search. The most interesting bits of
1405 * the cn_flags field have the following meanings:
1406 * - MAKEENTRY: If clear, free an entry from the cache rather than look
1408 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".."
1409 * - tsp: Return storage for cache timestamp. On a successful (positive
1410 * or negative) lookup, tsp will be filled with any timespec that
1411 * was stored when this cache entry was created. However, it will
1412 * be clear for "." entries.
1413 * - ticks: Return storage for alternate cache timestamp. On a successful
1414 * (positive or negative) lookup, it will contain the ticks value
1415 * that was current when the cache entry was created, unless cnp
1418 * Either both tsp and ticks have to be provided or neither of them.
1422 * - -1: A positive cache hit. vpp will contain the desired vnode.
1423 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due
1424 * to a forced unmount. vpp will not be modified. If the entry
1425 * is a whiteout, then the ISWHITEOUT flag will be set in
1427 * - 0: A cache miss. vpp will not be modified.
1431 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up
1432 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the
1433 * lock is not recursively acquired.
1435 static int __noinline
1436 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1437 struct timespec *tsp, int *ticksp)
1439 struct namecache *ncp;
1446 MPASS((cnp->cn_flags & (MAKEENTRY | ISDOTDOT)) == MAKEENTRY);
1449 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1450 blp = HASH2BUCKETLOCK(hash);
1453 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1454 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1455 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1459 if (__predict_false(ncp == NULL)) {
1461 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1463 counter_u64_add(nummiss, 1);
1467 if (ncp->nc_flag & NCF_NEGATIVE)
1468 goto negative_success;
1470 counter_u64_add(numposhits, 1);
1472 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1473 cache_out_ts(ncp, tsp, ticksp);
1475 vs = vget_prep(*vpp);
1477 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1484 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1485 if (cnp->cn_flags & ISLASTCN) {
1486 counter_u64_add(numnegzaps, 1);
1487 error = cache_zap_locked_vnode(ncp, dvp);
1488 if (__predict_false(error != 0)) {
1489 zap_and_exit_bucket_fail2++;
1497 SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, ncp->nc_name);
1498 cache_out_ts(ncp, tsp, ticksp);
1499 counter_u64_add(numneghits, 1);
1500 whiteout = (ncp->nc_flag & NCF_WHITE);
1501 cache_negative_hit(ncp);
1504 cnp->cn_flags |= ISWHITEOUT;
1509 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1510 struct timespec *tsp, int *ticksp)
1512 struct namecache *ncp;
1513 struct negstate *negstate;
1520 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
1523 if (__predict_false(!doingcache)) {
1524 cnp->cn_flags &= ~MAKEENTRY;
1529 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1530 if (cnp->cn_namelen == 1)
1531 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1532 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
1533 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
1536 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1538 if ((cnp->cn_flags & MAKEENTRY) == 0) {
1539 cache_remove_cnp(dvp, cnp);
1543 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1546 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1547 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1548 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1552 if (__predict_false(ncp == NULL)) {
1554 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1556 counter_u64_add(nummiss, 1);
1560 nc_flag = atomic_load_char(&ncp->nc_flag);
1561 if (nc_flag & NCF_NEGATIVE)
1562 goto negative_success;
1564 counter_u64_add(numposhits, 1);
1566 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1567 cache_out_ts(ncp, tsp, ticksp);
1569 if (!cache_ncp_canuse(ncp)) {
1574 vs = vget_prep_smr(*vpp);
1576 if (__predict_false(vs == VGET_NONE)) {
1580 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1587 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1588 if (cnp->cn_flags & ISLASTCN) {
1594 SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, ncp->nc_name);
1595 cache_out_ts(ncp, tsp, ticksp);
1596 counter_u64_add(numneghits, 1);
1597 whiteout = (ncp->nc_flag & NCF_WHITE);
1599 * TODO: We need to take locks to promote an entry. Code doing it
1600 * in SMR lookup can be modified to be shared.
1602 negstate = NCP2NEGSTATE(ncp);
1603 if ((negstate->neg_flag & NEG_HOT) == 0 ||
1604 !cache_ncp_canuse(ncp)) {
1610 cnp->cn_flags |= ISWHITEOUT;
1613 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
1616 struct celockstate {
1620 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1621 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1624 cache_celockstate_init(struct celockstate *cel)
1627 bzero(cel, sizeof(*cel));
1631 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1634 struct mtx *vlp1, *vlp2;
1636 MPASS(cel->vlp[0] == NULL);
1637 MPASS(cel->vlp[1] == NULL);
1638 MPASS(cel->vlp[2] == NULL);
1640 MPASS(vp != NULL || dvp != NULL);
1642 vlp1 = VP2VNODELOCK(vp);
1643 vlp2 = VP2VNODELOCK(dvp);
1644 cache_sort_vnodes(&vlp1, &vlp2);
1655 cache_unlock_vnodes_cel(struct celockstate *cel)
1658 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1660 if (cel->vlp[0] != NULL)
1661 mtx_unlock(cel->vlp[0]);
1662 if (cel->vlp[1] != NULL)
1663 mtx_unlock(cel->vlp[1]);
1664 if (cel->vlp[2] != NULL)
1665 mtx_unlock(cel->vlp[2]);
1669 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1674 cache_assert_vlp_locked(cel->vlp[0]);
1675 cache_assert_vlp_locked(cel->vlp[1]);
1676 MPASS(cel->vlp[2] == NULL);
1679 vlp = VP2VNODELOCK(vp);
1682 if (vlp >= cel->vlp[1]) {
1685 if (mtx_trylock(vlp))
1687 cache_lock_vnodes_cel_3_failures++;
1688 cache_unlock_vnodes_cel(cel);
1689 if (vlp < cel->vlp[0]) {
1691 mtx_lock(cel->vlp[0]);
1692 mtx_lock(cel->vlp[1]);
1694 if (cel->vlp[0] != NULL)
1695 mtx_lock(cel->vlp[0]);
1697 mtx_lock(cel->vlp[1]);
1707 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
1711 MPASS(cel->blp[0] == NULL);
1712 MPASS(cel->blp[1] == NULL);
1714 cache_sort_vnodes(&blp1, &blp2);
1725 cache_unlock_buckets_cel(struct celockstate *cel)
1728 if (cel->blp[0] != NULL)
1729 mtx_unlock(cel->blp[0]);
1730 mtx_unlock(cel->blp[1]);
1734 * Lock part of the cache affected by the insertion.
1736 * This means vnodelocks for dvp, vp and the relevant bucketlock.
1737 * However, insertion can result in removal of an old entry. In this
1738 * case we have an additional vnode and bucketlock pair to lock.
1740 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1741 * preserving the locking order (smaller address first).
1744 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1747 struct namecache *ncp;
1748 struct mtx *blps[2];
1750 blps[0] = HASH2BUCKETLOCK(hash);
1753 cache_lock_vnodes_cel(cel, dvp, vp);
1754 if (vp == NULL || vp->v_type != VDIR)
1756 ncp = vp->v_cache_dd;
1759 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1761 MPASS(ncp->nc_dvp == vp);
1762 blps[1] = NCP2BUCKETLOCK(ncp);
1763 if (ncp->nc_flag & NCF_NEGATIVE)
1765 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1768 * All vnodes got re-locked. Re-validate the state and if
1769 * nothing changed we are done. Otherwise restart.
1771 if (ncp == vp->v_cache_dd &&
1772 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1773 blps[1] == NCP2BUCKETLOCK(ncp) &&
1774 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1776 cache_unlock_vnodes_cel(cel);
1781 cache_lock_buckets_cel(cel, blps[0], blps[1]);
1785 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1788 struct namecache *ncp;
1789 struct mtx *blps[2];
1791 blps[0] = HASH2BUCKETLOCK(hash);
1794 cache_lock_vnodes_cel(cel, dvp, vp);
1795 ncp = dvp->v_cache_dd;
1798 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1800 MPASS(ncp->nc_dvp == dvp);
1801 blps[1] = NCP2BUCKETLOCK(ncp);
1802 if (ncp->nc_flag & NCF_NEGATIVE)
1804 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1806 if (ncp == dvp->v_cache_dd &&
1807 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1808 blps[1] == NCP2BUCKETLOCK(ncp) &&
1809 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1811 cache_unlock_vnodes_cel(cel);
1816 cache_lock_buckets_cel(cel, blps[0], blps[1]);
1820 cache_enter_unlock(struct celockstate *cel)
1823 cache_unlock_buckets_cel(cel);
1824 cache_unlock_vnodes_cel(cel);
1827 static void __noinline
1828 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
1829 struct componentname *cnp)
1831 struct celockstate cel;
1832 struct namecache *ncp;
1836 if (dvp->v_cache_dd == NULL)
1838 len = cnp->cn_namelen;
1839 cache_celockstate_init(&cel);
1840 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1841 cache_enter_lock_dd(&cel, dvp, vp, hash);
1842 vn_seqc_write_begin(dvp);
1843 ncp = dvp->v_cache_dd;
1844 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
1845 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
1846 cache_zap_locked(ncp);
1850 dvp->v_cache_dd = NULL;
1851 vn_seqc_write_end(dvp);
1852 cache_enter_unlock(&cel);
1858 * Add an entry to the cache.
1861 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1862 struct timespec *tsp, struct timespec *dtsp)
1864 struct celockstate cel;
1865 struct namecache *ncp, *n2, *ndd;
1866 struct namecache_ts *ncp_ts;
1867 struct nchashhead *ncpp;
1873 VNPASS(!VN_IS_DOOMED(dvp), dvp);
1874 VNPASS(dvp->v_type != VNON, dvp);
1876 VNPASS(!VN_IS_DOOMED(vp), vp);
1877 VNPASS(vp->v_type != VNON, vp);
1881 if (__predict_false(!doingcache))
1886 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1887 if (cnp->cn_namelen == 1)
1889 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1890 cache_enter_dotdot_prep(dvp, vp, cnp);
1891 flag = NCF_ISDOTDOT;
1896 * Avoid blowout in namecache entries.
1898 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1899 if (__predict_false(lnumcache >= ncsize)) {
1900 atomic_subtract_long(&numcache, 1);
1901 counter_u64_add(numdrops, 1);
1905 cache_celockstate_init(&cel);
1910 * Calculate the hash key and setup as much of the new
1911 * namecache entry as possible before acquiring the lock.
1913 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1914 ncp->nc_flag = flag | NCF_WIP;
1917 cache_negative_init(ncp);
1920 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1921 ncp_ts->nc_time = *tsp;
1922 ncp_ts->nc_ticks = ticks;
1923 ncp_ts->nc_nc.nc_flag |= NCF_TS;
1925 ncp_ts->nc_dotdottime = *dtsp;
1926 ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1929 len = ncp->nc_nlen = cnp->cn_namelen;
1930 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1931 memcpy(ncp->nc_name, cnp->cn_nameptr, len);
1932 ncp->nc_name[len] = '\0';
1933 cache_enter_lock(&cel, dvp, vp, hash);
1936 * See if this vnode or negative entry is already in the cache
1937 * with this name. This can happen with concurrent lookups of
1938 * the same path name.
1940 ncpp = NCHHASH(hash);
1941 CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
1942 if (n2->nc_dvp == dvp &&
1943 n2->nc_nlen == cnp->cn_namelen &&
1944 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1945 MPASS(cache_ncp_canuse(n2));
1946 if ((n2->nc_flag & NCF_NEGATIVE) != 0)
1948 ("%s: found entry pointing to a different vnode (%p != %p)",
1949 __func__, NULL, vp));
1951 KASSERT(n2->nc_vp == vp,
1952 ("%s: found entry pointing to a different vnode (%p != %p)",
1953 __func__, n2->nc_vp, vp));
1955 * Entries are supposed to be immutable unless in the
1956 * process of getting destroyed. Accommodating for
1957 * changing timestamps is possible but not worth it.
1958 * This should be harmless in terms of correctness, in
1959 * the worst case resulting in an earlier expiration.
1960 * Alternatively, the found entry can be replaced
1963 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
1966 KASSERT((n2->nc_flag & NCF_TS) != 0,
1968 n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1969 n2_ts->nc_time = ncp_ts->nc_time;
1970 n2_ts->nc_ticks = ncp_ts->nc_ticks;
1972 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1973 n2_ts->nc_nc.nc_flag |= NCF_DTS;
1977 goto out_unlock_free;
1981 if (flag == NCF_ISDOTDOT) {
1983 * See if we are trying to add .. entry, but some other lookup
1984 * has populated v_cache_dd pointer already.
1986 if (dvp->v_cache_dd != NULL)
1987 goto out_unlock_free;
1988 KASSERT(vp == NULL || vp->v_type == VDIR,
1989 ("wrong vnode type %p", vp));
1990 vn_seqc_write_begin(dvp);
1991 dvp->v_cache_dd = ncp;
1992 vn_seqc_write_end(dvp);
1996 if (flag != NCF_ISDOTDOT) {
1998 * For this case, the cache entry maps both the
1999 * directory name in it and the name ".." for the
2000 * directory's parent.
2002 vn_seqc_write_begin(vp);
2003 if ((ndd = vp->v_cache_dd) != NULL) {
2004 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2005 cache_zap_locked(ndd);
2009 vp->v_cache_dd = ncp;
2010 vn_seqc_write_end(vp);
2011 } else if (vp->v_type != VDIR) {
2012 if (vp->v_cache_dd != NULL) {
2013 vn_seqc_write_begin(vp);
2014 vp->v_cache_dd = NULL;
2015 vn_seqc_write_end(vp);
2020 if (flag != NCF_ISDOTDOT) {
2021 if (LIST_EMPTY(&dvp->v_cache_src)) {
2023 counter_u64_add(numcachehv, 1);
2025 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2029 * If the entry is "negative", we place it into the
2030 * "negative" cache queue, otherwise, we place it into the
2031 * destination vnode's cache entries queue.
2034 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2035 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2038 if (cnp->cn_flags & ISWHITEOUT)
2039 ncp->nc_flag |= NCF_WHITE;
2040 cache_negative_insert(ncp);
2041 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2046 * Insert the new namecache entry into the appropriate chain
2047 * within the cache entries table.
2049 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2051 atomic_thread_fence_rel();
2053 * Mark the entry as fully constructed.
2054 * It is immutable past this point until its removal.
2056 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2058 cache_enter_unlock(&cel);
2059 if (numneg * ncnegfactor > lnumcache)
2060 cache_negative_zap_one();
2065 cache_enter_unlock(&cel);
2066 atomic_subtract_long(&numcache, 1);
2072 cache_roundup_2(u_int val)
2076 for (res = 1; res <= val; res <<= 1)
2082 static struct nchashhead *
2083 nchinittbl(u_long elements, u_long *hashmask)
2085 struct nchashhead *hashtbl;
2088 hashsize = cache_roundup_2(elements) / 2;
2090 hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2091 for (i = 0; i < hashsize; i++)
2092 CK_SLIST_INIT(&hashtbl[i]);
2093 *hashmask = hashsize - 1;
2098 ncfreetbl(struct nchashhead *hashtbl)
2101 free(hashtbl, M_VFSCACHE);
2105 * Name cache initialization, from vfs_init() when we are booting
2108 nchinit(void *dummy __unused)
2112 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2113 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2114 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2115 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2116 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2117 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2118 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2119 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2121 VFS_SMR_ZONE_SET(cache_zone_small);
2122 VFS_SMR_ZONE_SET(cache_zone_small_ts);
2123 VFS_SMR_ZONE_SET(cache_zone_large);
2124 VFS_SMR_ZONE_SET(cache_zone_large_ts);
2126 ncsize = desiredvnodes * ncsizefactor;
2127 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2128 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2129 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2131 if (ncbuckethash > nchash)
2132 ncbuckethash = nchash;
2133 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2135 for (i = 0; i < numbucketlocks; i++)
2136 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2137 ncvnodehash = ncbuckethash;
2138 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2140 for (i = 0; i < numvnodelocks; i++)
2141 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2143 neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
2145 for (i = 0; i < numneglists; i++) {
2146 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2147 TAILQ_INIT(&neglists[i].nl_list);
2149 mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
2150 TAILQ_INIT(&ncneg_hot.nl_list);
2152 mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
2154 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2157 cache_vnode_init(struct vnode *vp)
2160 LIST_INIT(&vp->v_cache_src);
2161 TAILQ_INIT(&vp->v_cache_dst);
2162 vp->v_cache_dd = NULL;
2167 cache_changesize(u_long newmaxvnodes)
2169 struct nchashhead *new_nchashtbl, *old_nchashtbl;
2170 u_long new_nchash, old_nchash;
2171 struct namecache *ncp;
2176 newncsize = newmaxvnodes * ncsizefactor;
2177 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2178 if (newmaxvnodes < numbucketlocks)
2179 newmaxvnodes = numbucketlocks;
2181 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2182 /* If same hash table size, nothing to do */
2183 if (nchash == new_nchash) {
2184 ncfreetbl(new_nchashtbl);
2188 * Move everything from the old hash table to the new table.
2189 * None of the namecache entries in the table can be removed
2190 * because to do so, they have to be removed from the hash table.
2192 cache_lock_all_vnodes();
2193 cache_lock_all_buckets();
2194 old_nchashtbl = nchashtbl;
2195 old_nchash = nchash;
2196 nchashtbl = new_nchashtbl;
2197 nchash = new_nchash;
2198 for (i = 0; i <= old_nchash; i++) {
2199 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2200 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2202 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2203 CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2207 cache_unlock_all_buckets();
2208 cache_unlock_all_vnodes();
2209 ncfreetbl(old_nchashtbl);
2213 * Invalidate all entries from and to a particular vnode.
2216 cache_purge_impl(struct vnode *vp)
2218 TAILQ_HEAD(, namecache) ncps;
2219 struct namecache *ncp, *nnp;
2220 struct mtx *vlp, *vlp2;
2223 vlp = VP2VNODELOCK(vp);
2225 mtx_assert(vlp, MA_OWNED);
2227 while (!LIST_EMPTY(&vp->v_cache_src)) {
2228 ncp = LIST_FIRST(&vp->v_cache_src);
2229 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2231 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2233 while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2234 ncp = TAILQ_FIRST(&vp->v_cache_dst);
2235 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2237 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2239 ncp = vp->v_cache_dd;
2241 KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2242 ("lost dotdot link"));
2243 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2245 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2247 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2251 TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2257 * Opportunistic check to see if there is anything to do.
2260 cache_has_entries(struct vnode *vp)
2263 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2264 vp->v_cache_dd == NULL)
2270 cache_purge(struct vnode *vp)
2274 SDT_PROBE1(vfs, namecache, purge, done, vp);
2275 if (!cache_has_entries(vp))
2277 vlp = VP2VNODELOCK(vp);
2279 cache_purge_impl(vp);
2283 * Only to be used by vgone.
2286 cache_purge_vgone(struct vnode *vp)
2290 VNPASS(VN_IS_DOOMED(vp), vp);
2291 vlp = VP2VNODELOCK(vp);
2292 if (!(LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2293 vp->v_cache_dd == NULL)) {
2295 cache_purge_impl(vp);
2296 mtx_assert(vlp, MA_NOTOWNED);
2301 * All the NULL pointer state we found above may be transient.
2302 * Serialize against a possible thread doing cache_purge.
2304 mtx_wait_unlocked(vlp);
2305 if (!(LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2306 vp->v_cache_dd == NULL)) {
2308 cache_purge_impl(vp);
2309 mtx_assert(vlp, MA_NOTOWNED);
2316 * Invalidate all negative entries for a particular directory vnode.
2319 cache_purge_negative(struct vnode *vp)
2321 TAILQ_HEAD(, namecache) ncps;
2322 struct namecache *ncp, *nnp;
2325 SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2326 if (LIST_EMPTY(&vp->v_cache_src))
2329 vlp = VP2VNODELOCK(vp);
2331 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2332 if (!(ncp->nc_flag & NCF_NEGATIVE))
2334 cache_zap_negative_locked_vnode_kl(ncp, vp);
2335 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2338 TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2344 cache_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2345 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2348 ASSERT_VOP_IN_SEQC(fdvp);
2349 ASSERT_VOP_IN_SEQC(fvp);
2350 ASSERT_VOP_IN_SEQC(tdvp);
2352 ASSERT_VOP_IN_SEQC(tvp);
2357 KASSERT(!cache_remove_cnp(tdvp, tcnp),
2358 ("%s: lingering negative entry", __func__));
2360 cache_remove_cnp(tdvp, tcnp);
2365 * Flush all entries referencing a particular filesystem.
2368 cache_purgevfs(struct mount *mp)
2370 struct vnode *vp, *mvp;
2372 SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2374 * Somewhat wasteful iteration over all vnodes. Would be better to
2375 * support filtering and avoid the interlock to begin with.
2377 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2378 if (!cache_has_entries(vp)) {
2390 * Perform canonical checks and cache lookup and pass on to filesystem
2391 * through the vop_cachedlookup only if needed.
2395 vfs_cache_lookup(struct vop_lookup_args *ap)
2399 struct vnode **vpp = ap->a_vpp;
2400 struct componentname *cnp = ap->a_cnp;
2401 int flags = cnp->cn_flags;
2406 if (dvp->v_type != VDIR)
2409 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2410 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2413 error = vn_dir_check_exec(dvp, cnp);
2417 error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2419 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2425 /* Implementation of the getcwd syscall. */
2427 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2433 buflen = uap->buflen;
2434 if (__predict_false(buflen < 2))
2436 if (buflen > MAXPATHLEN)
2437 buflen = MAXPATHLEN;
2439 buf = uma_zalloc(namei_zone, M_WAITOK);
2440 error = vn_getcwd(buf, &retbuf, &buflen);
2442 error = copyout(retbuf, uap->buf, buflen);
2443 uma_zfree(namei_zone, buf);
2448 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
2454 pwd = pwd_get_smr();
2455 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
2457 VFS_SMR_ASSERT_NOT_ENTERED();
2459 pwd = pwd_hold(curthread);
2460 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
2466 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2473 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2474 size_t size, int flags, enum uio_seg pathseg)
2476 struct nameidata nd;
2477 char *retbuf, *freebuf;
2482 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2483 pathseg, path, fd, &cap_fstat_rights, td);
2484 if ((error = namei(&nd)) != 0)
2486 error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size);
2488 error = copyout(retbuf, buf, size);
2489 free(freebuf, M_TEMP);
2496 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2499 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2500 uap->flags, UIO_USERSPACE));
2504 * Retrieve the full filesystem path that correspond to a vnode from the name
2505 * cache (if available)
2508 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
2515 if (__predict_false(vp == NULL))
2518 buflen = MAXPATHLEN;
2519 buf = malloc(buflen, M_TEMP, M_WAITOK);
2521 pwd = pwd_get_smr();
2522 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, false, 0);
2523 VFS_SMR_ASSERT_NOT_ENTERED();
2525 pwd = pwd_hold(curthread);
2526 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
2537 * This function is similar to vn_fullpath, but it attempts to lookup the
2538 * pathname relative to the global root mount point. This is required for the
2539 * auditing sub-system, as audited pathnames must be absolute, relative to the
2540 * global root mount point.
2543 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
2549 if (__predict_false(vp == NULL))
2551 buflen = MAXPATHLEN;
2552 buf = malloc(buflen, M_TEMP, M_WAITOK);
2554 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, false, 0);
2555 VFS_SMR_ASSERT_NOT_ENTERED();
2557 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
2566 static struct namecache *
2567 vn_dd_from_dst(struct vnode *vp)
2569 struct namecache *ncp;
2571 cache_assert_vnode_locked(vp);
2572 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
2573 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2580 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, size_t *buflen)
2583 struct namecache *ncp;
2587 vlp = VP2VNODELOCK(*vp);
2589 ncp = (*vp)->v_cache_dd;
2590 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
2591 KASSERT(ncp == vn_dd_from_dst(*vp),
2592 ("%s: mismatch for dd entry (%p != %p)", __func__,
2593 ncp, vn_dd_from_dst(*vp)));
2595 ncp = vn_dd_from_dst(*vp);
2598 if (*buflen < ncp->nc_nlen) {
2601 counter_u64_add(numfullpathfail4, 1);
2603 SDT_PROBE3(vfs, namecache, fullpath, return, error,
2607 *buflen -= ncp->nc_nlen;
2608 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2609 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2618 SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2621 vn_lock(*vp, LK_SHARED | LK_RETRY);
2622 error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2625 counter_u64_add(numfullpathfail2, 1);
2626 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2631 if (VN_IS_DOOMED(dvp)) {
2632 /* forced unmount */
2635 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2639 * *vp has its use count incremented still.
2646 * Resolve a directory to a pathname.
2648 * The name of the directory can always be found in the namecache or fetched
2649 * from the filesystem. There is also guaranteed to be only one parent, meaning
2650 * we can just follow vnodes up until we find the root.
2652 * The vnode must be referenced.
2655 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
2656 size_t *len, bool slash_prefixed, size_t addend)
2658 #ifdef KDTRACE_HOOKS
2659 struct vnode *startvp = vp;
2665 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
2666 VNPASS(vp->v_usecount > 0, vp);
2670 if (!slash_prefixed) {
2678 SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2679 counter_u64_add(numfullpathcalls, 1);
2680 while (vp != rdir && vp != rootvnode) {
2682 * The vp vnode must be already fully constructed,
2683 * since it is either found in namecache or obtained
2684 * from VOP_VPTOCNP(). We may test for VV_ROOT safely
2685 * without obtaining the vnode lock.
2687 if ((vp->v_vflag & VV_ROOT) != 0) {
2688 vn_lock(vp, LK_RETRY | LK_SHARED);
2691 * With the vnode locked, check for races with
2692 * unmount, forced or not. Note that we
2693 * already verified that vp is not equal to
2694 * the root vnode, which means that
2695 * mnt_vnodecovered can be NULL only for the
2698 if (VN_IS_DOOMED(vp) ||
2699 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2700 vp1->v_mountedhere != vp->v_mount) {
2703 SDT_PROBE3(vfs, namecache, fullpath, return,
2713 if (vp->v_type != VDIR) {
2715 counter_u64_add(numfullpathfail1, 1);
2717 SDT_PROBE3(vfs, namecache, fullpath, return,
2721 error = vn_vptocnp(&vp, curthread->td_ucred, buf, &buflen);
2727 SDT_PROBE3(vfs, namecache, fullpath, return, error,
2731 buf[--buflen] = '/';
2732 slash_prefixed = true;
2736 if (!slash_prefixed) {
2739 counter_u64_add(numfullpathfail4, 1);
2740 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2744 buf[--buflen] = '/';
2746 counter_u64_add(numfullpathfound, 1);
2749 *retbuf = buf + buflen;
2750 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
2757 * Resolve an arbitrary vnode to a pathname.
2760 * - hardlinks are not tracked, thus if the vnode is not a directory this can
2761 * resolve to a different path than the one used to find it
2762 * - namecache is not mandatory, meaning names are not guaranteed to be added
2763 * (in which case resolving fails)
2765 static void __inline
2766 cache_rev_failed_impl(int *reason, int line)
2771 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__)
2774 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
2775 char **retbuf, size_t *buflen, bool slash_prefixed, size_t addend)
2777 #ifdef KDTRACE_HOOKS
2778 struct vnode *startvp = vp;
2782 struct namecache *ncp;
2786 #ifdef KDTRACE_HOOKS
2789 seqc_t vp_seqc, tvp_seqc;
2792 VFS_SMR_ASSERT_ENTERED();
2794 if (!cache_fast_revlookup) {
2799 orig_buflen = *buflen;
2801 if (!slash_prefixed) {
2802 MPASS(*buflen >= 2);
2804 buf[*buflen] = '\0';
2807 if (vp == rdir || vp == rootvnode) {
2808 if (!slash_prefixed) {
2815 #ifdef KDTRACE_HOOKS
2819 ncp = NULL; /* for sdt probe down below */
2820 vp_seqc = vn_seqc_read_any(vp);
2821 if (seqc_in_modify(vp_seqc)) {
2822 cache_rev_failed(&reason);
2827 #ifdef KDTRACE_HOOKS
2830 if ((vp->v_vflag & VV_ROOT) != 0) {
2831 mp = atomic_load_ptr(&vp->v_mount);
2833 cache_rev_failed(&reason);
2836 tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
2837 tvp_seqc = vn_seqc_read_any(tvp);
2838 if (seqc_in_modify(tvp_seqc)) {
2839 cache_rev_failed(&reason);
2842 if (!vn_seqc_consistent(vp, vp_seqc)) {
2843 cache_rev_failed(&reason);
2850 ncp = atomic_load_ptr(&vp->v_cache_dd);
2852 cache_rev_failed(&reason);
2855 nc_flag = atomic_load_char(&ncp->nc_flag);
2856 if ((nc_flag & NCF_ISDOTDOT) != 0) {
2857 cache_rev_failed(&reason);
2860 if (!cache_ncp_canuse(ncp)) {
2861 cache_rev_failed(&reason);
2864 if (ncp->nc_nlen >= *buflen) {
2865 cache_rev_failed(&reason);
2869 *buflen -= ncp->nc_nlen;
2870 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2874 tvp_seqc = vn_seqc_read_any(tvp);
2875 if (seqc_in_modify(tvp_seqc)) {
2876 cache_rev_failed(&reason);
2879 if (!vn_seqc_consistent(vp, vp_seqc)) {
2880 cache_rev_failed(&reason);
2885 if (vp == rdir || vp == rootvnode)
2890 *retbuf = buf + *buflen;
2891 *buflen = orig_buflen - *buflen + addend;
2892 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
2896 *buflen = orig_buflen;
2897 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
2903 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
2907 bool slash_prefixed;
2913 orig_buflen = *buflen;
2916 slash_prefixed = false;
2917 if (vp->v_type != VDIR) {
2919 buf[*buflen] = '\0';
2920 error = vn_vptocnp(&vp, curthread->td_ucred, buf, buflen);
2929 slash_prefixed = true;
2932 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, slash_prefixed,
2933 orig_buflen - *buflen));
2937 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
2939 * Since the namecache does not track handlings, the caller is expected to first
2940 * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
2942 * Then we have 2 cases:
2943 * - if the found vnode is a directory, the path can be constructed just by
2944 * fullowing names up the chain
2945 * - otherwise we populate the buffer with the saved name and start resolving
2949 vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf,
2954 struct componentname *cnp;
2958 bool slash_prefixed;
2963 if (*buflen > MAXPATHLEN)
2964 *buflen = MAXPATHLEN;
2966 slash_prefixed = false;
2968 buf = malloc(*buflen, M_TEMP, M_WAITOK);
2973 * Check for VBAD to work around the vp_crossmp bug in lookup().
2975 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
2976 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
2977 * If the type is VDIR (like in this very case) we can skip looking
2978 * at ni_dvp in the first place. However, since vnodes get passed here
2979 * unlocked the target may transition to doomed state (type == VBAD)
2980 * before we get to evaluate the condition. If this happens, we will
2981 * populate part of the buffer and descend to vn_fullpath_dir with
2982 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
2984 * This should be atomic_load(&vp->v_type) but it is ilegal to take
2985 * an address of a bit field, even if said field is sized to char.
2986 * Work around the problem by reading the value into a full-sized enum
2987 * and then re-reading it with atomic_load which will still prevent
2988 * the compiler from re-reading down the road.
2991 type = atomic_load_int(&type);
2998 addend = cnp->cn_namelen + 2;
2999 if (*buflen < addend) {
3004 tmpbuf = buf + *buflen;
3006 memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
3007 tmpbuf[addend - 1] = '\0';
3008 slash_prefixed = true;
3013 pwd = pwd_get_smr();
3014 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3015 slash_prefixed, addend);
3016 VFS_SMR_ASSERT_NOT_ENTERED();
3018 pwd = pwd_hold(curthread);
3020 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3021 slash_prefixed, addend);
3036 vn_dir_dd_ino(struct vnode *vp)
3038 struct namecache *ncp;
3043 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3044 vlp = VP2VNODELOCK(vp);
3046 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3047 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3050 vs = vget_prep(ddvp);
3052 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3061 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3063 struct namecache *ncp;
3067 vlp = VP2VNODELOCK(vp);
3069 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3070 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3076 l = min(ncp->nc_nlen, buflen - 1);
3077 memcpy(buf, ncp->nc_name, l);
3084 * This function updates path string to vnode's full global path
3085 * and checks the size of the new path string against the pathlen argument.
3087 * Requires a locked, referenced vnode.
3088 * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3090 * If vp is a directory, the call to vn_fullpath_global() always succeeds
3091 * because it falls back to the ".." lookup if the namecache lookup fails.
3094 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3097 struct nameidata nd;
3102 ASSERT_VOP_ELOCKED(vp, __func__);
3104 /* Construct global filesystem path from vp. */
3106 error = vn_fullpath_global(vp, &rpath, &fbuf);
3113 if (strlen(rpath) >= pathlen) {
3115 error = ENAMETOOLONG;
3120 * Re-lookup the vnode by path to detect a possible rename.
3121 * As a side effect, the vnode is relocked.
3122 * If vnode was renamed, return ENOENT.
3124 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3125 UIO_SYSSPACE, path, td);
3131 NDFREE(&nd, NDF_ONLY_PNBUF);
3135 strcpy(path, rpath);
3148 db_print_vpath(struct vnode *vp)
3151 while (vp != NULL) {
3152 db_printf("%p: ", vp);
3153 if (vp == rootvnode) {
3157 if (vp->v_vflag & VV_ROOT) {
3158 db_printf("<mount point>");
3159 vp = vp->v_mount->mnt_vnodecovered;
3161 struct namecache *ncp;
3165 ncp = TAILQ_FIRST(&vp->v_cache_dst);
3168 for (i = 0; i < ncp->nc_nlen; i++)
3169 db_printf("%c", *ncn++);
3182 DB_SHOW_COMMAND(vpath, db_show_vpath)
3187 db_printf("usage: show vpath <struct vnode *>\n");
3191 vp = (struct vnode *)addr;
3197 static bool __read_frequently cache_fast_lookup = true;
3198 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_lookup, CTLFLAG_RW,
3199 &cache_fast_lookup, 0, "");
3201 #define CACHE_FPL_FAILED -2020
3204 cache_fpl_cleanup_cnp(struct componentname *cnp)
3207 uma_zfree(namei_zone, cnp->cn_pnbuf);
3209 cnp->cn_pnbuf = NULL;
3210 cnp->cn_nameptr = NULL;
3215 cache_fpl_handle_root(struct nameidata *ndp, struct vnode **dpp)
3217 struct componentname *cnp;
3220 while (*(cnp->cn_nameptr) == '/') {
3225 *dpp = ndp->ni_rootdir;
3229 * Components of nameidata (or objects it can point to) which may
3230 * need restoring in case fast path lookup fails.
3232 struct nameidata_saved {
3240 struct nameidata *ndp;
3241 struct componentname *cnp;
3247 struct nameidata_saved snd;
3249 enum cache_fpl_status status:8;
3254 cache_fpl_checkpoint(struct cache_fpl *fpl, struct nameidata_saved *snd)
3257 snd->cn_flags = fpl->ndp->ni_cnd.cn_flags;
3258 snd->cn_namelen = fpl->ndp->ni_cnd.cn_namelen;
3259 snd->cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3260 snd->ni_pathlen = fpl->ndp->ni_pathlen;
3264 cache_fpl_restore(struct cache_fpl *fpl, struct nameidata_saved *snd)
3267 fpl->ndp->ni_cnd.cn_flags = snd->cn_flags;
3268 fpl->ndp->ni_cnd.cn_namelen = snd->cn_namelen;
3269 fpl->ndp->ni_cnd.cn_nameptr = snd->cn_nameptr;
3270 fpl->ndp->ni_pathlen = snd->ni_pathlen;
3274 #define cache_fpl_smr_assert_entered(fpl) ({ \
3275 struct cache_fpl *_fpl = (fpl); \
3276 MPASS(_fpl->in_smr == true); \
3277 VFS_SMR_ASSERT_ENTERED(); \
3279 #define cache_fpl_smr_assert_not_entered(fpl) ({ \
3280 struct cache_fpl *_fpl = (fpl); \
3281 MPASS(_fpl->in_smr == false); \
3282 VFS_SMR_ASSERT_NOT_ENTERED(); \
3285 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
3286 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
3289 #define cache_fpl_smr_enter_initial(fpl) ({ \
3290 struct cache_fpl *_fpl = (fpl); \
3292 _fpl->in_smr = true; \
3295 #define cache_fpl_smr_enter(fpl) ({ \
3296 struct cache_fpl *_fpl = (fpl); \
3297 MPASS(_fpl->in_smr == false); \
3299 _fpl->in_smr = true; \
3302 #define cache_fpl_smr_exit(fpl) ({ \
3303 struct cache_fpl *_fpl = (fpl); \
3304 MPASS(_fpl->in_smr == true); \
3306 _fpl->in_smr = false; \
3310 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
3313 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3314 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3315 ("%s: converting to abort from %d at %d, set at %d\n",
3316 __func__, fpl->status, line, fpl->line));
3318 fpl->status = CACHE_FPL_STATUS_ABORTED;
3320 return (CACHE_FPL_FAILED);
3323 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__)
3326 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
3329 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3330 ("%s: setting to partial at %d, but already set to %d at %d\n",
3331 __func__, line, fpl->status, fpl->line));
3332 cache_fpl_smr_assert_entered(fpl);
3333 fpl->status = CACHE_FPL_STATUS_PARTIAL;
3335 return (CACHE_FPL_FAILED);
3338 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__)
3341 cache_fpl_handled_impl(struct cache_fpl *fpl, int error, int line)
3344 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3345 ("%s: setting to handled at %d, but already set to %d at %d\n",
3346 __func__, line, fpl->status, fpl->line));
3347 cache_fpl_smr_assert_not_entered(fpl);
3348 MPASS(error != CACHE_FPL_FAILED);
3349 fpl->status = CACHE_FPL_STATUS_HANDLED;
3354 #define cache_fpl_handled(x, e) cache_fpl_handled_impl((x), (e), __LINE__)
3356 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
3357 (LOCKLEAF | LOCKPARENT | WANTPARENT | NOCACHE | FOLLOW | LOCKSHARED | SAVENAME | \
3358 SAVESTART | WILLBEDIR | ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
3360 #define CACHE_FPL_INTERNAL_CN_FLAGS \
3361 (ISDOTDOT | MAKEENTRY | ISLASTCN)
3363 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
3364 "supported and internal flags overlap");
3367 cache_fpl_islastcn(struct nameidata *ndp)
3370 return (*ndp->ni_next == 0);
3374 cache_fpl_isdotdot(struct componentname *cnp)
3377 if (cnp->cn_namelen == 2 &&
3378 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
3384 cache_can_fplookup(struct cache_fpl *fpl)
3386 struct nameidata *ndp;
3387 struct componentname *cnp;
3392 td = cnp->cn_thread;
3394 if (!cache_fast_lookup) {
3395 cache_fpl_aborted(fpl);
3399 if (mac_vnode_check_lookup_enabled()) {
3400 cache_fpl_aborted(fpl);
3404 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
3405 cache_fpl_aborted(fpl);
3408 if (ndp->ni_dirfd != AT_FDCWD) {
3409 cache_fpl_aborted(fpl);
3412 if (IN_CAPABILITY_MODE(td)) {
3413 cache_fpl_aborted(fpl);
3416 if (AUDITING_TD(td)) {
3417 cache_fpl_aborted(fpl);
3420 if (ndp->ni_startdir != NULL) {
3421 cache_fpl_aborted(fpl);
3428 cache_fplookup_vnode_supported(struct vnode *vp)
3431 return (vp->v_type != VLNK);
3435 * Move a negative entry to the hot list.
3437 * We have to take locks, but they may be contended and in the worst
3438 * case we may need to go off CPU. We don't want to spin within the
3439 * smr section and we can't block with it. Instead we are going to
3440 * look up the entry again.
3442 static int __noinline
3443 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
3446 struct componentname *cnp;
3447 struct namecache *ncp;
3448 struct neglist *neglist;
3449 struct negstate *negstate;
3456 if (!vhold_smr(dvp))
3457 return (cache_fpl_aborted(fpl));
3459 neglist = NCP2NEGLIST(oncp);
3460 cache_fpl_smr_exit(fpl);
3462 mtx_lock(&ncneg_hot.nl_lock);
3463 mtx_lock(&neglist->nl_lock);
3465 * For hash iteration.
3467 cache_fpl_smr_enter(fpl);
3470 * Avoid all surprises by only succeeding if we got the same entry and
3471 * bailing completely otherwise.
3473 * In particular at this point there can be a new ncp which matches the
3474 * search but hashes to a different neglist.
3476 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
3482 * No match to begin with.
3484 if (__predict_false(ncp == NULL)) {
3489 * The newly found entry may be something different...
3491 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
3492 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
3497 * ... and not even negative.
3499 nc_flag = atomic_load_char(&ncp->nc_flag);
3500 if ((nc_flag & NCF_NEGATIVE) == 0) {
3504 if (__predict_false(!cache_ncp_canuse(ncp))) {
3508 negstate = NCP2NEGSTATE(ncp);
3509 if ((negstate->neg_flag & NEG_HOT) == 0) {
3511 TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
3512 TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
3513 negstate->neg_flag |= NEG_HOT;
3516 SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, ncp->nc_name);
3517 counter_u64_add(numneghits, 1);
3518 cache_fpl_smr_exit(fpl);
3519 mtx_unlock(&neglist->nl_lock);
3520 mtx_unlock(&ncneg_hot.nl_lock);
3522 return (cache_fpl_handled(fpl, ENOENT));
3524 cache_fpl_smr_exit(fpl);
3525 mtx_unlock(&neglist->nl_lock);
3526 mtx_unlock(&ncneg_hot.nl_lock);
3528 return (cache_fpl_aborted(fpl));
3532 * The target vnode is not supported, prepare for the slow path to take over.
3534 static int __noinline
3535 cache_fplookup_partial_setup(struct cache_fpl *fpl)
3537 struct nameidata *ndp;
3538 struct componentname *cnp;
3547 dvp_seqc = fpl->dvp_seqc;
3549 dvs = vget_prep_smr(dvp);
3550 if (__predict_false(dvs == VGET_NONE)) {
3551 cache_fpl_smr_exit(fpl);
3552 return (cache_fpl_aborted(fpl));
3555 cache_fpl_smr_exit(fpl);
3557 vget_finish_ref(dvp, dvs);
3558 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3560 return (cache_fpl_aborted(fpl));
3563 pwd = pwd_hold(curthread);
3564 if (fpl->pwd != pwd) {
3567 return (cache_fpl_aborted(fpl));
3570 cache_fpl_restore(fpl, &fpl->snd);
3572 ndp->ni_startdir = dvp;
3573 cnp->cn_flags |= MAKEENTRY;
3574 if (cache_fpl_islastcn(ndp))
3575 cnp->cn_flags |= ISLASTCN;
3576 if (cache_fpl_isdotdot(cnp))
3577 cnp->cn_flags |= ISDOTDOT;
3583 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
3585 struct componentname *cnp;
3592 tvp_seqc = fpl->tvp_seqc;
3594 if ((cnp->cn_flags & LOCKLEAF) != 0) {
3595 lkflags = LK_SHARED;
3596 if ((cnp->cn_flags & LOCKSHARED) == 0)
3597 lkflags = LK_EXCLUSIVE;
3598 error = vget_finish(tvp, lkflags, tvs);
3599 if (__predict_false(error != 0)) {
3600 return (cache_fpl_aborted(fpl));
3603 vget_finish_ref(tvp, tvs);
3606 if (!vn_seqc_consistent(tvp, tvp_seqc)) {
3607 if ((cnp->cn_flags & LOCKLEAF) != 0)
3611 return (cache_fpl_aborted(fpl));
3614 return (cache_fpl_handled(fpl, 0));
3618 * They want to possibly modify the state of the namecache.
3620 * Don't try to match the API contract, just leave.
3621 * TODO: this leaves scalability on the table
3624 cache_fplookup_final_modifying(struct cache_fpl *fpl)
3626 struct componentname *cnp;
3629 MPASS(cnp->cn_nameiop != LOOKUP);
3630 return (cache_fpl_partial(fpl));
3633 static int __noinline
3634 cache_fplookup_final_withparent(struct cache_fpl *fpl)
3636 struct componentname *cnp;
3637 enum vgetstate dvs, tvs;
3638 struct vnode *dvp, *tvp;
3644 dvp_seqc = fpl->dvp_seqc;
3647 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
3650 * This is less efficient than it can be for simplicity.
3652 dvs = vget_prep_smr(dvp);
3653 if (__predict_false(dvs == VGET_NONE)) {
3654 return (cache_fpl_aborted(fpl));
3656 tvs = vget_prep_smr(tvp);
3657 if (__predict_false(tvs == VGET_NONE)) {
3658 cache_fpl_smr_exit(fpl);
3659 vget_abort(dvp, dvs);
3660 return (cache_fpl_aborted(fpl));
3663 cache_fpl_smr_exit(fpl);
3665 if ((cnp->cn_flags & LOCKPARENT) != 0) {
3666 error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
3667 if (__predict_false(error != 0)) {
3668 vget_abort(tvp, tvs);
3669 return (cache_fpl_aborted(fpl));
3672 vget_finish_ref(dvp, dvs);
3675 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3676 vget_abort(tvp, tvs);
3677 if ((cnp->cn_flags & LOCKPARENT) != 0)
3681 return (cache_fpl_aborted(fpl));
3684 error = cache_fplookup_final_child(fpl, tvs);
3685 if (__predict_false(error != 0)) {
3686 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED);
3687 if ((cnp->cn_flags & LOCKPARENT) != 0)
3694 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
3699 cache_fplookup_final(struct cache_fpl *fpl)
3701 struct componentname *cnp;
3703 struct vnode *dvp, *tvp;
3708 dvp_seqc = fpl->dvp_seqc;
3711 VNPASS(cache_fplookup_vnode_supported(dvp), dvp);
3713 if (cnp->cn_nameiop != LOOKUP) {
3714 return (cache_fplookup_final_modifying(fpl));
3717 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
3718 return (cache_fplookup_final_withparent(fpl));
3720 tvs = vget_prep_smr(tvp);
3721 if (__predict_false(tvs == VGET_NONE)) {
3722 return (cache_fpl_partial(fpl));
3725 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3726 cache_fpl_smr_exit(fpl);
3727 vget_abort(tvp, tvs);
3728 return (cache_fpl_aborted(fpl));
3731 cache_fpl_smr_exit(fpl);
3732 return (cache_fplookup_final_child(fpl, tvs));
3735 static int __noinline
3736 cache_fplookup_dot(struct cache_fpl *fpl)
3743 fpl->tvp_seqc = vn_seqc_read_any(dvp);
3744 if (seqc_in_modify(fpl->tvp_seqc)) {
3745 return (cache_fpl_aborted(fpl));
3748 counter_u64_add(dothits, 1);
3749 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", dvp);
3754 static int __noinline
3755 cache_fplookup_dotdot(struct cache_fpl *fpl)
3757 struct nameidata *ndp;
3758 struct componentname *cnp;
3759 struct namecache *ncp;
3769 * XXX this is racy the same way regular lookup is
3771 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
3773 if (dvp == pr->pr_root)
3776 if (dvp == ndp->ni_rootdir ||
3777 dvp == ndp->ni_topdir ||
3781 fpl->tvp_seqc = vn_seqc_read_any(dvp);
3782 if (seqc_in_modify(fpl->tvp_seqc)) {
3783 return (cache_fpl_aborted(fpl));
3788 if ((dvp->v_vflag & VV_ROOT) != 0) {
3791 * The opposite of climb mount is needed here.
3793 return (cache_fpl_aborted(fpl));
3796 ncp = atomic_load_ptr(&dvp->v_cache_dd);
3798 return (cache_fpl_aborted(fpl));
3801 nc_flag = atomic_load_char(&ncp->nc_flag);
3802 if ((nc_flag & NCF_ISDOTDOT) != 0) {
3803 if ((nc_flag & NCF_NEGATIVE) != 0)
3804 return (cache_fpl_aborted(fpl));
3805 fpl->tvp = ncp->nc_vp;
3807 fpl->tvp = ncp->nc_dvp;
3810 if (__predict_false(!cache_ncp_canuse(ncp))) {
3811 return (cache_fpl_aborted(fpl));
3814 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
3815 if (seqc_in_modify(fpl->tvp_seqc)) {
3816 return (cache_fpl_partial(fpl));
3819 counter_u64_add(dotdothits, 1);
3824 cache_fplookup_next(struct cache_fpl *fpl)
3826 struct componentname *cnp;
3827 struct namecache *ncp;
3828 struct negstate *negstate;
3829 struct vnode *dvp, *tvp;
3837 if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.')) {
3838 return (cache_fplookup_dot(fpl));
3841 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
3843 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
3844 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
3845 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
3850 * If there is no entry we have to punt to the slow path to perform
3851 * actual lookup. Should there be nothing with this name a negative
3852 * entry will be created.
3854 if (__predict_false(ncp == NULL)) {
3855 return (cache_fpl_partial(fpl));
3858 tvp = atomic_load_ptr(&ncp->nc_vp);
3859 nc_flag = atomic_load_char(&ncp->nc_flag);
3860 if ((nc_flag & NCF_NEGATIVE) != 0) {
3862 * If they want to create an entry we need to replace this one.
3864 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
3865 return (cache_fpl_partial(fpl));
3867 negstate = NCP2NEGSTATE(ncp);
3868 neg_hot = ((negstate->neg_flag & NEG_HOT) != 0);
3869 if (__predict_false(!cache_ncp_canuse(ncp))) {
3870 return (cache_fpl_partial(fpl));
3872 if (__predict_false((nc_flag & NCF_WHITE) != 0)) {
3873 return (cache_fpl_partial(fpl));
3876 return (cache_fplookup_negative_promote(fpl, ncp, hash));
3878 SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
3880 counter_u64_add(numneghits, 1);
3881 cache_fpl_smr_exit(fpl);
3882 return (cache_fpl_handled(fpl, ENOENT));
3885 if (__predict_false(!cache_ncp_canuse(ncp))) {
3886 return (cache_fpl_partial(fpl));
3890 fpl->tvp_seqc = vn_seqc_read_any(tvp);
3891 if (seqc_in_modify(fpl->tvp_seqc)) {
3892 return (cache_fpl_partial(fpl));
3895 if (!cache_fplookup_vnode_supported(tvp)) {
3896 return (cache_fpl_partial(fpl));
3899 counter_u64_add(numposhits, 1);
3900 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
3905 cache_fplookup_mp_supported(struct mount *mp)
3910 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
3916 * Walk up the mount stack (if any).
3918 * Correctness is provided in the following ways:
3919 * - all vnodes are protected from freeing with SMR
3920 * - struct mount objects are type stable making them always safe to access
3921 * - stability of the particular mount is provided by busying it
3922 * - relationship between the vnode which is mounted on and the mount is
3923 * verified with the vnode sequence counter after busying
3924 * - association between root vnode of the mount and the mount is protected
3927 * From that point on we can read the sequence counter of the root vnode
3928 * and get the next mount on the stack (if any) using the same protection.
3930 * By the end of successful walk we are guaranteed the reached state was
3931 * indeed present at least at some point which matches the regular lookup.
3933 static int __noinline
3934 cache_fplookup_climb_mount(struct cache_fpl *fpl)
3936 struct mount *mp, *prev_mp;
3941 vp_seqc = fpl->tvp_seqc;
3943 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
3944 mp = atomic_load_ptr(&vp->v_mountedhere);
3950 if (!vfs_op_thread_enter_crit(mp)) {
3951 if (prev_mp != NULL)
3952 vfs_op_thread_exit_crit(prev_mp);
3953 return (cache_fpl_partial(fpl));
3955 if (prev_mp != NULL)
3956 vfs_op_thread_exit_crit(prev_mp);
3957 if (!vn_seqc_consistent(vp, vp_seqc)) {
3958 vfs_op_thread_exit_crit(mp);
3959 return (cache_fpl_partial(fpl));
3961 if (!cache_fplookup_mp_supported(mp)) {
3962 vfs_op_thread_exit_crit(mp);
3963 return (cache_fpl_partial(fpl));
3965 vp = atomic_load_ptr(&mp->mnt_rootvnode);
3966 if (vp == NULL || VN_IS_DOOMED(vp)) {
3967 vfs_op_thread_exit_crit(mp);
3968 return (cache_fpl_partial(fpl));
3970 vp_seqc = vn_seqc_read_any(vp);
3971 if (seqc_in_modify(vp_seqc)) {
3972 vfs_op_thread_exit_crit(mp);
3973 return (cache_fpl_partial(fpl));
3976 mp = atomic_load_ptr(&vp->v_mountedhere);
3981 vfs_op_thread_exit_crit(prev_mp);
3983 fpl->tvp_seqc = vp_seqc;
3988 cache_fplookup_need_climb_mount(struct cache_fpl *fpl)
3996 * Hack: while this is a union, the pointer tends to be NULL so save on
3999 mp = atomic_load_ptr(&vp->v_mountedhere);
4002 if (vp->v_type == VDIR)
4010 * The code is mostly copy-pasted from regular lookup, see lookup().
4011 * The structure is maintained along with comments for easier maintenance.
4012 * Deduplicating the code will become feasible after fast path lookup
4013 * becomes more feature-complete.
4016 cache_fplookup_parse(struct cache_fpl *fpl)
4018 struct nameidata *ndp;
4019 struct componentname *cnp;
4026 * Search a new directory.
4028 * The last component of the filename is left accessible via
4029 * cnp->cn_nameptr for callers that need the name. Callers needing
4030 * the name set the SAVENAME flag. When done, they assume
4031 * responsibility for freeing the pathname buffer.
4033 for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
4035 cnp->cn_namelen = cp - cnp->cn_nameptr;
4036 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4037 cache_fpl_smr_exit(fpl);
4038 return (cache_fpl_handled(fpl, ENAMETOOLONG));
4040 ndp->ni_pathlen -= cnp->cn_namelen;
4041 KASSERT(ndp->ni_pathlen <= PATH_MAX,
4042 ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
4046 * Replace multiple slashes by a single slash and trailing slashes
4047 * by a null. This must be done before VOP_LOOKUP() because some
4048 * fs's don't know about trailing slashes. Remember if there were
4049 * trailing slashes to handle symlinks, existing non-directories
4050 * and non-existing files that won't be directories specially later.
4052 while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
4058 * Regular lookup performs the following:
4059 * *ndp->ni_next = '\0';
4060 * cnp->cn_flags |= TRAILINGSLASH;
4062 * Which is problematic since it modifies data read
4063 * from userspace. Then if fast path lookup was to
4064 * abort we would have to either restore it or convey
4065 * the flag. Since this is a corner case just ignore
4066 * it for simplicity.
4068 return (cache_fpl_partial(fpl));
4074 * Check for degenerate name (e.g. / or "")
4075 * which is a way of talking about a directory,
4076 * e.g. like "/." or ".".
4079 * Another corner case handled by the regular lookup
4081 if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
4082 return (cache_fpl_partial(fpl));
4088 cache_fplookup_parse_advance(struct cache_fpl *fpl)
4090 struct nameidata *ndp;
4091 struct componentname *cnp;
4096 cnp->cn_nameptr = ndp->ni_next;
4097 while (*cnp->cn_nameptr == '/') {
4103 static int __noinline
4104 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
4110 * Can happen when racing against vgone.
4113 cache_fpl_partial(fpl);
4117 * See the API contract for VOP_FPLOOKUP_VEXEC.
4119 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
4120 error = cache_fpl_aborted(fpl);
4122 cache_fpl_smr_exit(fpl);
4123 cache_fpl_handled(fpl, error);
4131 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
4133 struct nameidata *ndp;
4134 struct componentname *cnp;
4138 error = CACHE_FPL_FAILED;
4142 cache_fpl_checkpoint(fpl, &fpl->snd);
4145 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
4146 if (seqc_in_modify(fpl->dvp_seqc)) {
4147 cache_fpl_aborted(fpl);
4150 mp = atomic_load_ptr(&fpl->dvp->v_mount);
4151 if (!cache_fplookup_mp_supported(mp)) {
4152 cache_fpl_aborted(fpl);
4156 VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
4159 error = cache_fplookup_parse(fpl);
4160 if (__predict_false(error != 0)) {
4164 VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
4166 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
4167 if (__predict_false(error != 0)) {
4168 error = cache_fplookup_failed_vexec(fpl, error);
4172 if (__predict_false(cache_fpl_isdotdot(cnp))) {
4173 error = cache_fplookup_dotdot(fpl);
4174 if (__predict_false(error != 0)) {
4178 error = cache_fplookup_next(fpl);
4179 if (__predict_false(error != 0)) {
4183 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
4185 if (cache_fplookup_need_climb_mount(fpl)) {
4186 error = cache_fplookup_climb_mount(fpl);
4187 if (__predict_false(error != 0)) {
4193 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
4195 if (cache_fpl_islastcn(ndp)) {
4196 error = cache_fplookup_final(fpl);
4200 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
4201 error = cache_fpl_aborted(fpl);
4205 fpl->dvp = fpl->tvp;
4206 fpl->dvp_seqc = fpl->tvp_seqc;
4208 cache_fplookup_parse_advance(fpl);
4209 cache_fpl_checkpoint(fpl, &fpl->snd);
4212 switch (fpl->status) {
4213 case CACHE_FPL_STATUS_UNSET:
4214 __assert_unreachable();
4216 case CACHE_FPL_STATUS_PARTIAL:
4217 cache_fpl_smr_assert_entered(fpl);
4218 return (cache_fplookup_partial_setup(fpl));
4219 case CACHE_FPL_STATUS_ABORTED:
4221 cache_fpl_smr_exit(fpl);
4222 return (CACHE_FPL_FAILED);
4223 case CACHE_FPL_STATUS_HANDLED:
4224 MPASS(error != CACHE_FPL_FAILED);
4225 cache_fpl_smr_assert_not_entered(fpl);
4226 if (__predict_false(error != 0)) {
4229 cache_fpl_cleanup_cnp(cnp);
4232 ndp->ni_dvp = fpl->dvp;
4233 ndp->ni_vp = fpl->tvp;
4234 if (cnp->cn_flags & SAVENAME)
4235 cnp->cn_flags |= HASBUF;
4237 cache_fpl_cleanup_cnp(cnp);
4243 * Fast path lookup protected with SMR and sequence counters.
4245 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
4247 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
4250 * Traditional vnode lookup conceptually looks like this:
4256 * vn_unlock(current);
4263 * Each jump to the next vnode is safe memory-wise and atomic with respect to
4264 * any modifications thanks to holding respective locks.
4266 * The same guarantee can be provided with a combination of safe memory
4267 * reclamation and sequence counters instead. If all operations which affect
4268 * the relationship between the current vnode and the one we are looking for
4269 * also modify the counter, we can verify whether all the conditions held as
4270 * we made the jump. This includes things like permissions, mount points etc.
4271 * Counter modification is provided by enclosing relevant places in
4272 * vn_seqc_write_begin()/end() calls.
4274 * Thus this translates to:
4277 * dvp_seqc = seqc_read_any(dvp);
4278 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
4282 * tvp_seqc = seqc_read_any(tvp);
4283 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
4285 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
4287 * dvp = tvp; // we know nothing of importance has changed
4288 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
4292 * vget(); // secure the vnode
4293 * if (!seqc_consistent(tvp, tvp_seqc) // final check
4295 * // at this point we know nothing has changed for any parent<->child pair
4296 * // as they were crossed during the lookup, meaning we matched the guarantee
4297 * // of the locked variant
4300 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
4301 * - they are called while within vfs_smr protection which they must never exit
4302 * - EAGAIN can be returned to denote checking could not be performed, it is
4303 * always valid to return it
4304 * - if the sequence counter has not changed the result must be valid
4305 * - if the sequence counter has changed both false positives and false negatives
4306 * are permitted (since the result will be rejected later)
4307 * - for simple cases of unix permission checks vaccess_vexec_smr can be used
4309 * Caveats to watch out for:
4310 * - vnodes are passed unlocked and unreferenced with nothing stopping
4311 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
4312 * to use atomic_load_ptr to fetch it.
4313 * - the aforementioned object can also get freed, meaning absent other means it
4314 * should be protected with vfs_smr
4315 * - either safely checking permissions as they are modified or guaranteeing
4316 * their stability is left to the routine
4319 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
4322 struct cache_fpl fpl;
4325 struct componentname *cnp;
4326 struct nameidata_saved orig;
4329 MPASS(ndp->ni_lcf == 0);
4331 fpl.status = CACHE_FPL_STATUS_UNSET;
4333 fpl.cnp = &ndp->ni_cnd;
4334 MPASS(curthread == fpl.cnp->cn_thread);
4336 if ((fpl.cnp->cn_flags & SAVESTART) != 0)
4337 MPASS(fpl.cnp->cn_nameiop != LOOKUP);
4339 if (!cache_can_fplookup(&fpl)) {
4340 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
4341 *status = fpl.status;
4342 return (EOPNOTSUPP);
4345 cache_fpl_checkpoint(&fpl, &orig);
4347 cache_fpl_smr_enter_initial(&fpl);
4348 pwd = pwd_get_smr();
4350 ndp->ni_rootdir = pwd->pwd_rdir;
4351 ndp->ni_topdir = pwd->pwd_jdir;
4354 cnp->cn_nameptr = cnp->cn_pnbuf;
4355 if (cnp->cn_pnbuf[0] == '/') {
4356 cache_fpl_handle_root(ndp, &dvp);
4358 MPASS(ndp->ni_dirfd == AT_FDCWD);
4359 dvp = pwd->pwd_cdir;
4362 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
4364 error = cache_fplookup_impl(dvp, &fpl);
4365 cache_fpl_smr_assert_not_entered(&fpl);
4366 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
4368 *status = fpl.status;
4369 switch (fpl.status) {
4370 case CACHE_FPL_STATUS_UNSET:
4371 __assert_unreachable();
4373 case CACHE_FPL_STATUS_HANDLED:
4374 SDT_PROBE3(vfs, namei, lookup, return, error,
4375 (error == 0 ? ndp->ni_vp : NULL), true);
4377 case CACHE_FPL_STATUS_PARTIAL:
4380 * Status restored by cache_fplookup_partial_setup.
4383 case CACHE_FPL_STATUS_ABORTED:
4384 cache_fpl_restore(&fpl, &orig);