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 <machine/_inttypes.h>
74 #include <sys/capsicum.h>
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
85 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
88 SDT_PROVIDER_DECLARE(vfs);
89 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
91 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *",
93 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
95 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
97 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
98 "struct namecache *", "int", "int");
99 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
100 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
101 "char *", "struct vnode *");
102 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
103 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
104 "struct vnode *", "char *");
105 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
107 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
108 "struct vnode *", "char *");
109 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
111 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
112 "struct componentname *");
113 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
114 "struct componentname *");
115 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
116 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int");
117 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
118 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
119 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
121 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
123 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *",
126 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
127 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
128 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
131 * This structure describes the elements in the cache of recent
132 * names looked up by namei.
138 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
139 "the state must fit in a union with a pointer without growing it");
142 LIST_ENTRY(namecache) nc_src; /* source vnode list */
143 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */
144 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
145 struct vnode *nc_dvp; /* vnode of parent of name */
147 struct vnode *nu_vp; /* vnode the name refers to */
148 struct negstate nu_neg;/* negative entry state */
150 u_char nc_flag; /* flag bits */
151 u_char nc_nlen; /* length of name */
152 char nc_name[0]; /* segment name + nul */
156 * struct namecache_ts repeats struct namecache layout up to the
158 * struct namecache_ts is used in place of struct namecache when time(s) need
159 * to be stored. The nc_dotdottime field is used when a cache entry is mapping
160 * both a non-dotdot directory name plus dotdot for the directory's
163 * See below for alignment requirement.
165 struct namecache_ts {
166 struct timespec nc_time; /* timespec provided by fs */
167 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */
168 int nc_ticks; /* ticks value when entry was added */
170 struct namecache nc_nc;
173 TAILQ_HEAD(cache_freebatch, namecache);
176 * At least mips n32 performs 64-bit accesses to timespec as found
177 * in namecache_ts and requires them to be aligned. Since others
178 * may be in the same spot suffer a little bit and enforce the
179 * alignment for everyone. Note this is a nop for 64-bit platforms.
181 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t)
184 * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
185 * 4.4 BSD codebase. Later on struct namecache was tweaked to become
186 * smaller and the value was bumped to retain the total size, but it
187 * was never re-evaluated for suitability. A simple test counting
188 * lengths during package building shows that the value of 45 covers
189 * about 86% of all added entries, reaching 99% at 65.
191 * Regardless of the above, use of dedicated zones instead of malloc may be
192 * inducing additional waste. This may be hard to address as said zones are
193 * tied to VFS SMR. Even if retaining them, the current split should be
197 #define CACHE_PATH_CUTOFF 45
198 #define CACHE_LARGE_PAD 6
200 #define CACHE_PATH_CUTOFF 41
201 #define CACHE_LARGE_PAD 2
204 #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
205 #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
206 #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
207 #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
209 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
210 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
211 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
212 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
214 #define nc_vp n_un.nu_vp
215 #define nc_neg n_un.nu_neg
218 * Flags in namecache.nc_flag
220 #define NCF_WHITE 0x01
221 #define NCF_ISDOTDOT 0x02
224 #define NCF_DVDROP 0x10
225 #define NCF_NEGATIVE 0x20
226 #define NCF_INVALID 0x40
230 * Flags in negstate.neg_flag
234 static bool cache_neg_evict_cond(u_long lnumcache);
237 * Mark an entry as invalid.
239 * This is called before it starts getting deconstructed.
242 cache_ncp_invalidate(struct namecache *ncp)
245 KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
246 ("%s: entry %p already invalid", __func__, ncp));
247 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
248 atomic_thread_fence_rel();
252 * Check whether the entry can be safely used.
254 * All places which elide locks are supposed to call this after they are
255 * done with reading from an entry.
257 #define cache_ncp_canuse(ncp) ({ \
258 struct namecache *_ncp = (ncp); \
261 atomic_thread_fence_acq(); \
262 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
263 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \
267 * Like the above but also checks NCF_WHITE.
269 #define cache_fpl_neg_ncp_canuse(ncp) ({ \
270 struct namecache *_ncp = (ncp); \
273 atomic_thread_fence_acq(); \
274 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
275 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0); \
279 * Name caching works as follows:
281 * Names found by directory scans are retained in a cache
282 * for future reference. It is managed LRU, so frequently
283 * used names will hang around. Cache is indexed by hash value
284 * obtained from (dvp, name) where dvp refers to the directory
287 * If it is a "negative" entry, (i.e. for a name that is known NOT to
288 * exist) the vnode pointer will be NULL.
290 * Upon reaching the last segment of a path, if the reference
291 * is for DELETE, or NOCACHE is set (rewrite), and the
292 * name is located in the cache, it will be dropped.
294 * These locks are used (in the order in which they can be taken):
296 * vnodelock mtx vnode lists and v_cache_dd field protection
297 * bucketlock mtx for access to given set of hash buckets
298 * neglist mtx negative entry LRU management
300 * It is legal to take multiple vnodelock and bucketlock locks. The locking
301 * order is lower address first. Both are recursive.
303 * "." lookups are lockless.
305 * ".." and vnode -> name lookups require vnodelock.
307 * name -> vnode lookup requires the relevant bucketlock to be held for reading.
309 * Insertions and removals of entries require involved vnodes and bucketlocks
310 * to be locked to provide safe operation against other threads modifying the
313 * Some lookups result in removal of the found entry (e.g. getting rid of a
314 * negative entry with the intent to create a positive one), which poses a
315 * problem when multiple threads reach the state. Similarly, two different
316 * threads can purge two different vnodes and try to remove the same name.
318 * If the already held vnode lock is lower than the second required lock, we
319 * can just take the other lock. However, in the opposite case, this could
320 * deadlock. As such, this is resolved by trylocking and if that fails unlocking
321 * the first node, locking everything in order and revalidating the state.
326 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
327 "Name cache parameters");
329 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */
330 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0,
331 "Total namecache capacity");
333 u_int ncsizefactor = 2;
334 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
335 "Size factor for namecache");
337 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */
338 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
339 "Ratio of negative namecache entries");
342 * Negative entry % of namecache capacity above which automatic eviction is allowed.
344 * Check cache_neg_evict_cond for details.
346 static u_int ncnegminpct = 3;
348 static u_int __read_mostly neg_min; /* the above recomputed against ncsize */
349 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
350 "Negative entry count above which automatic eviction is allowed");
353 * Structures associated with name caching.
355 #define NCHHASH(hash) \
356 (&nchashtbl[(hash) & nchash])
357 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
358 static u_long __read_mostly nchash; /* size of hash table */
359 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
360 "Size of namecache hash table");
361 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */
362 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */
364 struct nchstats nchstats; /* cache effectiveness statistics */
366 static bool __read_frequently cache_fast_revlookup = true;
367 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
368 &cache_fast_revlookup, 0, "");
370 static u_int __exclusive_cache_line neg_cycle;
373 #define numneglists (ncneghash + 1)
376 struct mtx nl_evict_lock;
377 struct mtx nl_lock __aligned(CACHE_LINE_SIZE);
378 TAILQ_HEAD(, namecache) nl_list;
379 TAILQ_HEAD(, namecache) nl_hotlist;
381 } __aligned(CACHE_LINE_SIZE);
383 static struct neglist neglists[numneglists];
385 static inline struct neglist *
386 NCP2NEGLIST(struct namecache *ncp)
389 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
392 static inline struct negstate *
393 NCP2NEGSTATE(struct namecache *ncp)
396 MPASS(ncp->nc_flag & NCF_NEGATIVE);
397 return (&ncp->nc_neg);
400 #define numbucketlocks (ncbuckethash + 1)
401 static u_int __read_mostly ncbuckethash;
402 static struct mtx_padalign __read_mostly *bucketlocks;
403 #define HASH2BUCKETLOCK(hash) \
404 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
406 #define numvnodelocks (ncvnodehash + 1)
407 static u_int __read_mostly ncvnodehash;
408 static struct mtx __read_mostly *vnodelocks;
409 static inline struct mtx *
410 VP2VNODELOCK(struct vnode *vp)
413 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
417 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
419 struct namecache_ts *ncp_ts;
421 KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
422 (tsp == NULL && ticksp == NULL),
428 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
429 *tsp = ncp_ts->nc_time;
430 *ticksp = ncp_ts->nc_ticks;
434 static int __read_mostly doingcache = 1; /* 1 => enable the cache */
435 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
436 "VFS namecache enabled");
439 /* Export size information to userland */
440 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
441 sizeof(struct namecache), "sizeof(struct namecache)");
444 * The new name cache statistics
446 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
447 "Name cache statistics");
449 #define STATNODE_ULONG(name, varname, descr) \
450 SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
451 #define STATNODE_COUNTER(name, varname, descr) \
452 static COUNTER_U64_DEFINE_EARLY(varname); \
453 SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
455 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
456 STATNODE_ULONG(count, numcache, "Number of cache entries");
457 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
458 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
459 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits");
460 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits");
461 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
462 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
463 STATNODE_COUNTER(posszaps, numposzaps,
464 "Number of cache hits (positive) we do not want to cache");
465 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
466 STATNODE_COUNTER(negzaps, numnegzaps,
467 "Number of cache hits (negative) we do not want to cache");
468 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
469 /* These count for vn_getcwd(), too. */
470 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
471 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
472 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
473 "Number of fullpath search errors (VOP_VPTOCNP failures)");
474 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
475 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
478 * Debug or developer statistics.
480 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
481 "Name cache debugging");
482 #define DEBUGNODE_ULONG(name, varname, descr) \
483 SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
484 #define DEBUGNODE_COUNTER(name, varname, descr) \
485 static COUNTER_U64_DEFINE_EARLY(varname); \
486 SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \
488 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success,
489 "Number of successful removals after relocking");
490 static long zap_bucket_fail;
491 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
492 static long zap_bucket_fail2;
493 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
494 static long cache_lock_vnodes_cel_3_failures;
495 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
496 "Number of times 3-way vnode locking failed");
498 static void cache_zap_locked(struct namecache *ncp);
499 static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf,
500 char **freebuf, size_t *buflen);
501 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
502 char **retbuf, size_t *buflen, size_t addend);
503 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
504 char **retbuf, size_t *buflen);
505 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
506 char **retbuf, size_t *len, size_t addend);
508 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
511 cache_assert_vlp_locked(struct mtx *vlp)
515 mtx_assert(vlp, MA_OWNED);
519 cache_assert_vnode_locked(struct vnode *vp)
523 vlp = VP2VNODELOCK(vp);
524 cache_assert_vlp_locked(vlp);
528 * Directory vnodes with entries are held for two reasons:
529 * 1. make them less of a target for reclamation in vnlru
530 * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
532 * It will be feasible to stop doing it altogether if all filesystems start
533 * supporting lockless lookup.
536 cache_hold_vnode(struct vnode *vp)
539 cache_assert_vnode_locked(vp);
540 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
542 counter_u64_add(numcachehv, 1);
546 cache_drop_vnode(struct vnode *vp)
550 * Called after all locks are dropped, meaning we can't assert
551 * on the state of v_cache_src.
554 counter_u64_add(numcachehv, -1);
560 static uma_zone_t __read_mostly cache_zone_small;
561 static uma_zone_t __read_mostly cache_zone_small_ts;
562 static uma_zone_t __read_mostly cache_zone_large;
563 static uma_zone_t __read_mostly cache_zone_large_ts;
565 static struct namecache *
566 cache_alloc_uma(int len, bool ts)
568 struct namecache_ts *ncp_ts;
569 struct namecache *ncp;
571 if (__predict_false(ts)) {
572 if (len <= CACHE_PATH_CUTOFF)
573 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
575 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
576 ncp = &ncp_ts->nc_nc;
578 if (len <= CACHE_PATH_CUTOFF)
579 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
581 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
587 cache_free_uma(struct namecache *ncp)
589 struct namecache_ts *ncp_ts;
591 if (__predict_false(ncp->nc_flag & NCF_TS)) {
592 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
593 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
594 uma_zfree_smr(cache_zone_small_ts, ncp_ts);
596 uma_zfree_smr(cache_zone_large_ts, ncp_ts);
598 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
599 uma_zfree_smr(cache_zone_small, ncp);
601 uma_zfree_smr(cache_zone_large, ncp);
605 static struct namecache *
606 cache_alloc(int len, bool ts)
611 * Avoid blowout in namecache entries.
614 * 1. filesystems may end up trying to add an already existing entry
615 * (for example this can happen after a cache miss during concurrent
616 * lookup), in which case we will call cache_neg_evict despite not
618 * 2. the routine may fail to free anything and no provisions are made
619 * to make it try harder (see the inside for failure modes)
620 * 3. it only ever looks at negative entries.
622 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
623 if (cache_neg_evict_cond(lnumcache)) {
624 lnumcache = atomic_load_long(&numcache);
626 if (__predict_false(lnumcache >= ncsize)) {
627 atomic_subtract_long(&numcache, 1);
628 counter_u64_add(numdrops, 1);
631 return (cache_alloc_uma(len, ts));
635 cache_free(struct namecache *ncp)
639 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
640 cache_drop_vnode(ncp->nc_dvp);
643 atomic_subtract_long(&numcache, 1);
647 cache_free_batch(struct cache_freebatch *batch)
649 struct namecache *ncp, *nnp;
653 if (TAILQ_EMPTY(batch))
655 TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
656 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
657 cache_drop_vnode(ncp->nc_dvp);
662 atomic_subtract_long(&numcache, i);
664 SDT_PROBE1(vfs, namecache, purge, batch, i);
668 * TODO: With the value stored we can do better than computing the hash based
669 * on the address. The choice of FNV should also be revisited.
672 cache_prehash(struct vnode *vp)
675 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
679 cache_get_hash(char *name, u_char len, struct vnode *dvp)
682 return (fnv_32_buf(name, len, dvp->v_nchash));
685 static inline struct nchashhead *
686 NCP2BUCKET(struct namecache *ncp)
690 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
691 return (NCHHASH(hash));
694 static inline struct mtx *
695 NCP2BUCKETLOCK(struct namecache *ncp)
699 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
700 return (HASH2BUCKETLOCK(hash));
705 cache_assert_bucket_locked(struct namecache *ncp)
709 blp = NCP2BUCKETLOCK(ncp);
710 mtx_assert(blp, MA_OWNED);
714 cache_assert_bucket_unlocked(struct namecache *ncp)
718 blp = NCP2BUCKETLOCK(ncp);
719 mtx_assert(blp, MA_NOTOWNED);
722 #define cache_assert_bucket_locked(x) do { } while (0)
723 #define cache_assert_bucket_unlocked(x) do { } while (0)
726 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y))
728 _cache_sort_vnodes(void **p1, void **p2)
732 MPASS(*p1 != NULL || *p2 != NULL);
742 cache_lock_all_buckets(void)
746 for (i = 0; i < numbucketlocks; i++)
747 mtx_lock(&bucketlocks[i]);
751 cache_unlock_all_buckets(void)
755 for (i = 0; i < numbucketlocks; i++)
756 mtx_unlock(&bucketlocks[i]);
760 cache_lock_all_vnodes(void)
764 for (i = 0; i < numvnodelocks; i++)
765 mtx_lock(&vnodelocks[i]);
769 cache_unlock_all_vnodes(void)
773 for (i = 0; i < numvnodelocks; i++)
774 mtx_unlock(&vnodelocks[i]);
778 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
781 cache_sort_vnodes(&vlp1, &vlp2);
784 if (!mtx_trylock(vlp1))
787 if (!mtx_trylock(vlp2)) {
797 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
800 MPASS(vlp1 != NULL || vlp2 != NULL);
810 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
813 MPASS(vlp1 != NULL || vlp2 != NULL);
822 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
824 struct nchstats snap;
826 if (req->oldptr == NULL)
827 return (SYSCTL_OUT(req, 0, sizeof(snap)));
830 snap.ncs_goodhits = counter_u64_fetch(numposhits);
831 snap.ncs_neghits = counter_u64_fetch(numneghits);
832 snap.ncs_badhits = counter_u64_fetch(numposzaps) +
833 counter_u64_fetch(numnegzaps);
834 snap.ncs_miss = counter_u64_fetch(nummisszap) +
835 counter_u64_fetch(nummiss);
837 return (SYSCTL_OUT(req, &snap, sizeof(snap)));
839 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
840 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
841 "VFS cache effectiveness statistics");
844 cache_recalc_neg_min(u_int val)
847 neg_min = (ncsize * val) / 100;
851 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
857 error = sysctl_handle_int(oidp, &val, 0, req);
858 if (error != 0 || req->newptr == NULL)
861 if (val == ncnegminpct)
863 if (val < 0 || val > 99)
866 cache_recalc_neg_min(val);
870 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
871 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
872 "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
876 * Grab an atomic snapshot of the name cache hash chain lengths
878 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
879 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
883 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
885 struct nchashhead *ncpp;
886 struct namecache *ncp;
887 int i, error, n_nchash, *cntbuf;
890 n_nchash = nchash + 1; /* nchash is max index, not count */
891 if (req->oldptr == NULL)
892 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
893 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
894 cache_lock_all_buckets();
895 if (n_nchash != nchash + 1) {
896 cache_unlock_all_buckets();
897 free(cntbuf, M_TEMP);
900 /* Scan hash tables counting entries */
901 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
902 CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
904 cache_unlock_all_buckets();
905 for (error = 0, i = 0; i < n_nchash; i++)
906 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
908 free(cntbuf, M_TEMP);
911 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
912 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
913 "nchash chain lengths");
916 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
919 struct nchashhead *ncpp;
920 struct namecache *ncp;
922 int count, maxlength, used, pct;
925 return SYSCTL_OUT(req, 0, 4 * sizeof(int));
927 cache_lock_all_buckets();
928 n_nchash = nchash + 1; /* nchash is max index, not count */
932 /* Scan hash tables for applicable entries */
933 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
935 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
940 if (maxlength < count)
943 n_nchash = nchash + 1;
944 cache_unlock_all_buckets();
945 pct = (used * 100) / (n_nchash / 100);
946 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
949 error = SYSCTL_OUT(req, &used, sizeof(used));
952 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
955 error = SYSCTL_OUT(req, &pct, sizeof(pct));
960 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
961 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
962 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
966 * Negative entries management
968 * Various workloads create plenty of negative entries and barely use them
969 * afterwards. Moreover malicious users can keep performing bogus lookups
970 * adding even more entries. For example "make tinderbox" as of writing this
971 * comment ends up with 2.6M namecache entries in total, 1.2M of which are
974 * As such, a rather aggressive eviction method is needed. The currently
975 * employed method is a placeholder.
977 * Entries are split over numneglists separate lists, each of which is further
978 * split into hot and cold entries. Entries get promoted after getting a hit.
979 * Eviction happens on addition of new entry.
981 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
982 "Name cache negative entry statistics");
984 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
985 "Number of negative cache entries");
987 static COUNTER_U64_DEFINE_EARLY(neg_created);
988 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
989 "Number of created negative entries");
991 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
992 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
993 "Number of evicted negative entries");
995 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
996 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
997 &neg_evict_skipped_empty,
998 "Number of times evicting failed due to lack of entries");
1000 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
1001 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
1002 &neg_evict_skipped_missed,
1003 "Number of times evicting failed due to target entry disappearing");
1005 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
1006 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
1007 &neg_evict_skipped_contended,
1008 "Number of times evicting failed due to contention");
1010 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
1011 "Number of cache hits (negative)");
1014 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1019 for (i = 0; i < numneglists; i++)
1020 out += neglists[i].nl_hotnum;
1022 return (SYSCTL_OUT(req, &out, sizeof(out)));
1024 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1025 CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1026 "Number of hot negative entries");
1029 cache_neg_init(struct namecache *ncp)
1031 struct negstate *ns;
1033 ncp->nc_flag |= NCF_NEGATIVE;
1034 ns = NCP2NEGSTATE(ncp);
1037 counter_u64_add(neg_created, 1);
1040 #define CACHE_NEG_PROMOTION_THRESH 2
1043 cache_neg_hit_prep(struct namecache *ncp)
1045 struct negstate *ns;
1048 ns = NCP2NEGSTATE(ncp);
1049 n = atomic_load_char(&ns->neg_hit);
1051 if (n >= CACHE_NEG_PROMOTION_THRESH)
1053 if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1056 return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1060 * Nothing to do here but it is provided for completeness as some
1061 * cache_neg_hit_prep callers may end up returning without even
1062 * trying to promote.
1064 #define cache_neg_hit_abort(ncp) do { } while (0)
1067 cache_neg_hit_finish(struct namecache *ncp)
1070 SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1071 counter_u64_add(numneghits, 1);
1075 * Move a negative entry to the hot list.
1078 cache_neg_promote_locked(struct namecache *ncp)
1081 struct negstate *ns;
1083 ns = NCP2NEGSTATE(ncp);
1084 nl = NCP2NEGLIST(ncp);
1085 mtx_assert(&nl->nl_lock, MA_OWNED);
1086 if ((ns->neg_flag & NEG_HOT) == 0) {
1087 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1088 TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1090 ns->neg_flag |= NEG_HOT;
1095 * Move a hot negative entry to the cold list.
1098 cache_neg_demote_locked(struct namecache *ncp)
1101 struct negstate *ns;
1103 ns = NCP2NEGSTATE(ncp);
1104 nl = NCP2NEGLIST(ncp);
1105 mtx_assert(&nl->nl_lock, MA_OWNED);
1106 MPASS(ns->neg_flag & NEG_HOT);
1107 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1108 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1110 ns->neg_flag &= ~NEG_HOT;
1111 atomic_store_char(&ns->neg_hit, 0);
1115 * Move a negative entry to the hot list if it matches the lookup.
1117 * We have to take locks, but they may be contended and in the worst
1118 * case we may need to go off CPU. We don't want to spin within the
1119 * smr section and we can't block with it. Exiting the section means
1120 * the found entry could have been evicted. We are going to look it
1124 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1125 struct namecache *oncp, uint32_t hash)
1127 struct namecache *ncp;
1131 nl = NCP2NEGLIST(oncp);
1133 mtx_lock(&nl->nl_lock);
1135 * For hash iteration.
1140 * Avoid all surprises by only succeeding if we got the same entry and
1141 * bailing completely otherwise.
1142 * XXX There are no provisions to keep the vnode around, meaning we may
1143 * end up promoting a negative entry for a *new* vnode and returning
1144 * ENOENT on its account. This is the error we want to return anyway
1145 * and promotion is harmless.
1147 * In particular at this point there can be a new ncp which matches the
1148 * search but hashes to a different neglist.
1150 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1156 * No match to begin with.
1158 if (__predict_false(ncp == NULL)) {
1163 * The newly found entry may be something different...
1165 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1166 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
1171 * ... and not even negative.
1173 nc_flag = atomic_load_char(&ncp->nc_flag);
1174 if ((nc_flag & NCF_NEGATIVE) == 0) {
1178 if (!cache_ncp_canuse(ncp)) {
1182 cache_neg_promote_locked(ncp);
1183 cache_neg_hit_finish(ncp);
1185 mtx_unlock(&nl->nl_lock);
1189 mtx_unlock(&nl->nl_lock);
1194 cache_neg_promote(struct namecache *ncp)
1198 nl = NCP2NEGLIST(ncp);
1199 mtx_lock(&nl->nl_lock);
1200 cache_neg_promote_locked(ncp);
1201 mtx_unlock(&nl->nl_lock);
1205 cache_neg_insert(struct namecache *ncp)
1209 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1210 cache_assert_bucket_locked(ncp);
1211 nl = NCP2NEGLIST(ncp);
1212 mtx_lock(&nl->nl_lock);
1213 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1214 mtx_unlock(&nl->nl_lock);
1215 atomic_add_long(&numneg, 1);
1219 cache_neg_remove(struct namecache *ncp)
1222 struct negstate *ns;
1224 cache_assert_bucket_locked(ncp);
1225 nl = NCP2NEGLIST(ncp);
1226 ns = NCP2NEGSTATE(ncp);
1227 mtx_lock(&nl->nl_lock);
1228 if ((ns->neg_flag & NEG_HOT) != 0) {
1229 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1232 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1234 mtx_unlock(&nl->nl_lock);
1235 atomic_subtract_long(&numneg, 1);
1238 static struct neglist *
1239 cache_neg_evict_select_list(void)
1244 c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1245 nl = &neglists[c % numneglists];
1246 if (!mtx_trylock(&nl->nl_evict_lock)) {
1247 counter_u64_add(neg_evict_skipped_contended, 1);
1253 static struct namecache *
1254 cache_neg_evict_select_entry(struct neglist *nl)
1256 struct namecache *ncp, *lncp;
1257 struct negstate *ns, *lns;
1260 mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1261 mtx_assert(&nl->nl_lock, MA_OWNED);
1262 ncp = TAILQ_FIRST(&nl->nl_list);
1266 lns = NCP2NEGSTATE(lncp);
1267 for (i = 1; i < 4; i++) {
1268 ncp = TAILQ_NEXT(ncp, nc_dst);
1271 ns = NCP2NEGSTATE(ncp);
1272 if (ns->neg_hit < lns->neg_hit) {
1281 cache_neg_evict(void)
1283 struct namecache *ncp, *ncp2;
1292 nl = cache_neg_evict_select_list();
1297 mtx_lock(&nl->nl_lock);
1298 ncp = TAILQ_FIRST(&nl->nl_hotlist);
1300 cache_neg_demote_locked(ncp);
1302 ncp = cache_neg_evict_select_entry(nl);
1304 counter_u64_add(neg_evict_skipped_empty, 1);
1305 mtx_unlock(&nl->nl_lock);
1306 mtx_unlock(&nl->nl_evict_lock);
1309 nlen = ncp->nc_nlen;
1311 hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1312 dvlp = VP2VNODELOCK(dvp);
1313 blp = HASH2BUCKETLOCK(hash);
1314 mtx_unlock(&nl->nl_lock);
1315 mtx_unlock(&nl->nl_evict_lock);
1319 * Note that since all locks were dropped above, the entry may be
1320 * gone or reallocated to be something else.
1322 CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1323 if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1324 ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1328 counter_u64_add(neg_evict_skipped_missed, 1);
1332 MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1333 MPASS(blp == NCP2BUCKETLOCK(ncp));
1334 SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1336 cache_zap_locked(ncp);
1337 counter_u64_add(neg_evicted, 1);
1348 * Maybe evict a negative entry to create more room.
1350 * The ncnegfactor parameter limits what fraction of the total count
1351 * can comprise of negative entries. However, if the cache is just
1352 * warming up this leads to excessive evictions. As such, ncnegminpct
1353 * (recomputed to neg_min) dictates whether the above should be
1356 * Try evicting if the cache is close to full capacity regardless of
1357 * other considerations.
1360 cache_neg_evict_cond(u_long lnumcache)
1364 if (ncsize - 1000 < lnumcache)
1366 lnumneg = atomic_load_long(&numneg);
1367 if (lnumneg < neg_min)
1369 if (lnumneg * ncnegfactor < lnumcache)
1372 return (cache_neg_evict());
1376 * cache_zap_locked():
1378 * Removes a namecache entry from cache, whether it contains an actual
1379 * pointer to a vnode or if it is just a negative cache entry.
1382 cache_zap_locked(struct namecache *ncp)
1384 struct nchashhead *ncpp;
1386 if (!(ncp->nc_flag & NCF_NEGATIVE))
1387 cache_assert_vnode_locked(ncp->nc_vp);
1388 cache_assert_vnode_locked(ncp->nc_dvp);
1389 cache_assert_bucket_locked(ncp);
1391 cache_ncp_invalidate(ncp);
1393 ncpp = NCP2BUCKET(ncp);
1394 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1395 if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1396 SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
1397 ncp->nc_name, ncp->nc_vp);
1398 TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
1399 if (ncp == ncp->nc_vp->v_cache_dd) {
1400 vn_seqc_write_begin_unheld(ncp->nc_vp);
1401 ncp->nc_vp->v_cache_dd = NULL;
1402 vn_seqc_write_end(ncp->nc_vp);
1405 SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
1407 cache_neg_remove(ncp);
1409 if (ncp->nc_flag & NCF_ISDOTDOT) {
1410 if (ncp == ncp->nc_dvp->v_cache_dd) {
1411 vn_seqc_write_begin_unheld(ncp->nc_dvp);
1412 ncp->nc_dvp->v_cache_dd = NULL;
1413 vn_seqc_write_end(ncp->nc_dvp);
1416 LIST_REMOVE(ncp, nc_src);
1417 if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
1418 ncp->nc_flag |= NCF_DVDROP;
1424 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1428 MPASS(ncp->nc_dvp == vp);
1429 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1430 cache_assert_vnode_locked(vp);
1432 blp = NCP2BUCKETLOCK(ncp);
1434 cache_zap_locked(ncp);
1439 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1442 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1445 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1446 cache_assert_vnode_locked(vp);
1448 if (ncp->nc_flag & NCF_NEGATIVE) {
1449 if (*vlpp != NULL) {
1453 cache_zap_negative_locked_vnode_kl(ncp, vp);
1457 pvlp = VP2VNODELOCK(vp);
1458 blp = NCP2BUCKETLOCK(ncp);
1459 vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1460 vlp2 = VP2VNODELOCK(ncp->nc_vp);
1462 if (*vlpp == vlp1 || *vlpp == vlp2) {
1466 if (*vlpp != NULL) {
1470 cache_sort_vnodes(&vlp1, &vlp2);
1475 if (!mtx_trylock(vlp1))
1481 cache_zap_locked(ncp);
1483 if (to_unlock != NULL)
1484 mtx_unlock(to_unlock);
1491 MPASS(*vlpp == NULL);
1497 * If trylocking failed we can get here. We know enough to take all needed locks
1498 * in the right order and re-lookup the entry.
1501 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1502 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1505 struct namecache *rncp;
1507 cache_assert_bucket_unlocked(ncp);
1509 cache_sort_vnodes(&dvlp, &vlp);
1510 cache_lock_vnodes(dvlp, vlp);
1512 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1513 if (rncp == ncp && rncp->nc_dvp == dvp &&
1514 rncp->nc_nlen == cnp->cn_namelen &&
1515 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1519 cache_zap_locked(rncp);
1521 cache_unlock_vnodes(dvlp, vlp);
1522 counter_u64_add(zap_bucket_relock_success, 1);
1527 cache_unlock_vnodes(dvlp, vlp);
1531 static int __noinline
1532 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1533 uint32_t hash, struct mtx *blp)
1535 struct mtx *dvlp, *vlp;
1538 cache_assert_bucket_locked(ncp);
1540 dvlp = VP2VNODELOCK(ncp->nc_dvp);
1542 if (!(ncp->nc_flag & NCF_NEGATIVE))
1543 vlp = VP2VNODELOCK(ncp->nc_vp);
1544 if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1545 cache_zap_locked(ncp);
1547 cache_unlock_vnodes(dvlp, vlp);
1553 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1556 static __noinline int
1557 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1559 struct namecache *ncp;
1561 struct mtx *dvlp, *dvlp2;
1565 if (cnp->cn_namelen == 2 &&
1566 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1567 dvlp = VP2VNODELOCK(dvp);
1571 ncp = dvp->v_cache_dd;
1576 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1579 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1580 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1582 MPASS(dvp->v_cache_dd == NULL);
1588 vn_seqc_write_begin(dvp);
1589 dvp->v_cache_dd = NULL;
1590 vn_seqc_write_end(dvp);
1595 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1599 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1600 blp = HASH2BUCKETLOCK(hash);
1602 if (CK_SLIST_EMPTY(NCHHASH(hash)))
1607 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1608 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1609 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1618 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1619 if (__predict_false(error != 0)) {
1623 counter_u64_add(numposzaps, 1);
1624 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1628 counter_u64_add(nummisszap, 1);
1629 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1633 static int __noinline
1634 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1635 struct timespec *tsp, int *ticksp)
1640 counter_u64_add(dothits, 1);
1641 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1648 * When we lookup "." we still can be asked to lock it
1651 ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1652 if (ltype != VOP_ISLOCKED(*vpp)) {
1653 if (ltype == LK_EXCLUSIVE) {
1654 vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1655 if (VN_IS_DOOMED((*vpp))) {
1656 /* forced unmount */
1662 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1667 static int __noinline
1668 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1669 struct timespec *tsp, int *ticksp)
1671 struct namecache_ts *ncp_ts;
1672 struct namecache *ncp;
1678 MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1680 if ((cnp->cn_flags & MAKEENTRY) == 0) {
1681 cache_remove_cnp(dvp, cnp);
1685 counter_u64_add(dotdothits, 1);
1687 dvlp = VP2VNODELOCK(dvp);
1689 ncp = dvp->v_cache_dd;
1691 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL);
1695 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1696 if (ncp->nc_flag & NCF_NEGATIVE)
1703 goto negative_success;
1704 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1705 cache_out_ts(ncp, tsp, ticksp);
1706 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1707 NCF_DTS && tsp != NULL) {
1708 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1709 *tsp = ncp_ts->nc_dotdottime;
1713 ltype = VOP_ISLOCKED(dvp);
1715 vs = vget_prep(*vpp);
1717 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1718 vn_lock(dvp, ltype | LK_RETRY);
1719 if (VN_IS_DOOMED(dvp)) {
1731 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1732 if (cnp->cn_flags & ISLASTCN) {
1733 counter_u64_add(numnegzaps, 1);
1734 cache_zap_negative_locked_vnode_kl(ncp, dvp);
1741 whiteout = (ncp->nc_flag & NCF_WHITE);
1742 cache_out_ts(ncp, tsp, ticksp);
1743 if (cache_neg_hit_prep(ncp))
1744 cache_neg_promote(ncp);
1746 cache_neg_hit_finish(ncp);
1749 cnp->cn_flags |= ISWHITEOUT;
1754 * Lookup a name in the name cache
1758 * - dvp: Parent directory in which to search.
1759 * - vpp: Return argument. Will contain desired vnode on cache hit.
1760 * - cnp: Parameters of the name search. The most interesting bits of
1761 * the cn_flags field have the following meanings:
1762 * - MAKEENTRY: If clear, free an entry from the cache rather than look
1764 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".."
1765 * - tsp: Return storage for cache timestamp. On a successful (positive
1766 * or negative) lookup, tsp will be filled with any timespec that
1767 * was stored when this cache entry was created. However, it will
1768 * be clear for "." entries.
1769 * - ticks: Return storage for alternate cache timestamp. On a successful
1770 * (positive or negative) lookup, it will contain the ticks value
1771 * that was current when the cache entry was created, unless cnp
1774 * Either both tsp and ticks have to be provided or neither of them.
1778 * - -1: A positive cache hit. vpp will contain the desired vnode.
1779 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due
1780 * to a forced unmount. vpp will not be modified. If the entry
1781 * is a whiteout, then the ISWHITEOUT flag will be set in
1783 * - 0: A cache miss. vpp will not be modified.
1787 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up
1788 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the
1789 * lock is not recursively acquired.
1791 static int __noinline
1792 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1793 struct timespec *tsp, int *ticksp)
1795 struct namecache *ncp;
1802 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1803 MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
1806 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1807 blp = HASH2BUCKETLOCK(hash);
1810 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1811 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1812 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1816 if (__predict_false(ncp == NULL)) {
1818 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1820 counter_u64_add(nummiss, 1);
1824 if (ncp->nc_flag & NCF_NEGATIVE)
1825 goto negative_success;
1827 counter_u64_add(numposhits, 1);
1829 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1830 cache_out_ts(ncp, tsp, ticksp);
1832 vs = vget_prep(*vpp);
1834 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1842 * We don't get here with regular lookup apart from corner cases.
1844 if (__predict_true(cnp->cn_nameiop == CREATE)) {
1845 if (cnp->cn_flags & ISLASTCN) {
1846 counter_u64_add(numnegzaps, 1);
1847 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1848 if (__predict_false(error != 0)) {
1857 whiteout = (ncp->nc_flag & NCF_WHITE);
1858 cache_out_ts(ncp, tsp, ticksp);
1859 if (cache_neg_hit_prep(ncp))
1860 cache_neg_promote(ncp);
1862 cache_neg_hit_finish(ncp);
1865 cnp->cn_flags |= ISWHITEOUT;
1870 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1871 struct timespec *tsp, int *ticksp)
1873 struct namecache *ncp;
1877 bool whiteout, neg_promote;
1880 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
1883 if (__predict_false(!doingcache)) {
1884 cnp->cn_flags &= ~MAKEENTRY;
1889 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1890 if (cnp->cn_namelen == 1)
1891 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1892 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
1893 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
1896 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1898 if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
1899 cache_remove_cnp(dvp, cnp);
1903 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1906 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1907 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1908 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1912 if (__predict_false(ncp == NULL)) {
1914 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1916 counter_u64_add(nummiss, 1);
1920 nc_flag = atomic_load_char(&ncp->nc_flag);
1921 if (nc_flag & NCF_NEGATIVE)
1922 goto negative_success;
1924 counter_u64_add(numposhits, 1);
1926 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1927 cache_out_ts(ncp, tsp, ticksp);
1929 if (!cache_ncp_canuse(ncp)) {
1934 vs = vget_prep_smr(*vpp);
1936 if (__predict_false(vs == VGET_NONE)) {
1940 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1947 if (cnp->cn_nameiop == CREATE) {
1948 if (cnp->cn_flags & ISLASTCN) {
1954 cache_out_ts(ncp, tsp, ticksp);
1955 whiteout = (ncp->nc_flag & NCF_WHITE);
1956 neg_promote = cache_neg_hit_prep(ncp);
1957 if (!cache_ncp_canuse(ncp)) {
1958 cache_neg_hit_abort(ncp);
1964 if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
1967 cache_neg_hit_finish(ncp);
1971 cnp->cn_flags |= ISWHITEOUT;
1974 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
1977 struct celockstate {
1981 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1982 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1985 cache_celockstate_init(struct celockstate *cel)
1988 bzero(cel, sizeof(*cel));
1992 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1995 struct mtx *vlp1, *vlp2;
1997 MPASS(cel->vlp[0] == NULL);
1998 MPASS(cel->vlp[1] == NULL);
1999 MPASS(cel->vlp[2] == NULL);
2001 MPASS(vp != NULL || dvp != NULL);
2003 vlp1 = VP2VNODELOCK(vp);
2004 vlp2 = VP2VNODELOCK(dvp);
2005 cache_sort_vnodes(&vlp1, &vlp2);
2016 cache_unlock_vnodes_cel(struct celockstate *cel)
2019 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2021 if (cel->vlp[0] != NULL)
2022 mtx_unlock(cel->vlp[0]);
2023 if (cel->vlp[1] != NULL)
2024 mtx_unlock(cel->vlp[1]);
2025 if (cel->vlp[2] != NULL)
2026 mtx_unlock(cel->vlp[2]);
2030 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2035 cache_assert_vlp_locked(cel->vlp[0]);
2036 cache_assert_vlp_locked(cel->vlp[1]);
2037 MPASS(cel->vlp[2] == NULL);
2040 vlp = VP2VNODELOCK(vp);
2043 if (vlp >= cel->vlp[1]) {
2046 if (mtx_trylock(vlp))
2048 cache_lock_vnodes_cel_3_failures++;
2049 cache_unlock_vnodes_cel(cel);
2050 if (vlp < cel->vlp[0]) {
2052 mtx_lock(cel->vlp[0]);
2053 mtx_lock(cel->vlp[1]);
2055 if (cel->vlp[0] != NULL)
2056 mtx_lock(cel->vlp[0]);
2058 mtx_lock(cel->vlp[1]);
2068 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2072 MPASS(cel->blp[0] == NULL);
2073 MPASS(cel->blp[1] == NULL);
2075 cache_sort_vnodes(&blp1, &blp2);
2086 cache_unlock_buckets_cel(struct celockstate *cel)
2089 if (cel->blp[0] != NULL)
2090 mtx_unlock(cel->blp[0]);
2091 mtx_unlock(cel->blp[1]);
2095 * Lock part of the cache affected by the insertion.
2097 * This means vnodelocks for dvp, vp and the relevant bucketlock.
2098 * However, insertion can result in removal of an old entry. In this
2099 * case we have an additional vnode and bucketlock pair to lock.
2101 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2102 * preserving the locking order (smaller address first).
2105 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2108 struct namecache *ncp;
2109 struct mtx *blps[2];
2111 blps[0] = HASH2BUCKETLOCK(hash);
2114 cache_lock_vnodes_cel(cel, dvp, vp);
2115 if (vp == NULL || vp->v_type != VDIR)
2117 ncp = vp->v_cache_dd;
2120 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2122 MPASS(ncp->nc_dvp == vp);
2123 blps[1] = NCP2BUCKETLOCK(ncp);
2124 if (ncp->nc_flag & NCF_NEGATIVE)
2126 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2129 * All vnodes got re-locked. Re-validate the state and if
2130 * nothing changed we are done. Otherwise restart.
2132 if (ncp == vp->v_cache_dd &&
2133 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2134 blps[1] == NCP2BUCKETLOCK(ncp) &&
2135 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2137 cache_unlock_vnodes_cel(cel);
2142 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2146 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2149 struct namecache *ncp;
2150 struct mtx *blps[2];
2152 blps[0] = HASH2BUCKETLOCK(hash);
2155 cache_lock_vnodes_cel(cel, dvp, vp);
2156 ncp = dvp->v_cache_dd;
2159 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2161 MPASS(ncp->nc_dvp == dvp);
2162 blps[1] = NCP2BUCKETLOCK(ncp);
2163 if (ncp->nc_flag & NCF_NEGATIVE)
2165 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2167 if (ncp == dvp->v_cache_dd &&
2168 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2169 blps[1] == NCP2BUCKETLOCK(ncp) &&
2170 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2172 cache_unlock_vnodes_cel(cel);
2177 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2181 cache_enter_unlock(struct celockstate *cel)
2184 cache_unlock_buckets_cel(cel);
2185 cache_unlock_vnodes_cel(cel);
2188 static void __noinline
2189 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2190 struct componentname *cnp)
2192 struct celockstate cel;
2193 struct namecache *ncp;
2197 if (dvp->v_cache_dd == NULL)
2199 len = cnp->cn_namelen;
2200 cache_celockstate_init(&cel);
2201 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2202 cache_enter_lock_dd(&cel, dvp, vp, hash);
2203 vn_seqc_write_begin(dvp);
2204 ncp = dvp->v_cache_dd;
2205 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2206 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2207 cache_zap_locked(ncp);
2211 dvp->v_cache_dd = NULL;
2212 vn_seqc_write_end(dvp);
2213 cache_enter_unlock(&cel);
2219 * Add an entry to the cache.
2222 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2223 struct timespec *tsp, struct timespec *dtsp)
2225 struct celockstate cel;
2226 struct namecache *ncp, *n2, *ndd;
2227 struct namecache_ts *ncp_ts;
2228 struct nchashhead *ncpp;
2233 VNPASS(dvp != vp, dvp);
2234 VNPASS(!VN_IS_DOOMED(dvp), dvp);
2235 VNPASS(dvp->v_type != VNON, dvp);
2237 VNPASS(!VN_IS_DOOMED(vp), vp);
2238 VNPASS(vp->v_type != VNON, vp);
2242 if (__predict_false(!doingcache))
2247 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2248 if (cnp->cn_namelen == 1)
2250 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2251 cache_enter_dotdot_prep(dvp, vp, cnp);
2252 flag = NCF_ISDOTDOT;
2256 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2260 cache_celockstate_init(&cel);
2265 * Calculate the hash key and setup as much of the new
2266 * namecache entry as possible before acquiring the lock.
2268 ncp->nc_flag = flag | NCF_WIP;
2271 cache_neg_init(ncp);
2274 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2275 ncp_ts->nc_time = *tsp;
2276 ncp_ts->nc_ticks = ticks;
2277 ncp_ts->nc_nc.nc_flag |= NCF_TS;
2279 ncp_ts->nc_dotdottime = *dtsp;
2280 ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2283 len = ncp->nc_nlen = cnp->cn_namelen;
2284 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2285 memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2286 ncp->nc_name[len] = '\0';
2287 cache_enter_lock(&cel, dvp, vp, hash);
2290 * See if this vnode or negative entry is already in the cache
2291 * with this name. This can happen with concurrent lookups of
2292 * the same path name.
2294 ncpp = NCHHASH(hash);
2295 CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
2296 if (n2->nc_dvp == dvp &&
2297 n2->nc_nlen == cnp->cn_namelen &&
2298 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
2299 MPASS(cache_ncp_canuse(n2));
2300 if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2302 ("%s: found entry pointing to a different vnode (%p != %p)",
2303 __func__, NULL, vp));
2305 KASSERT(n2->nc_vp == vp,
2306 ("%s: found entry pointing to a different vnode (%p != %p)",
2307 __func__, n2->nc_vp, vp));
2309 * Entries are supposed to be immutable unless in the
2310 * process of getting destroyed. Accommodating for
2311 * changing timestamps is possible but not worth it.
2312 * This should be harmless in terms of correctness, in
2313 * the worst case resulting in an earlier expiration.
2314 * Alternatively, the found entry can be replaced
2317 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2320 KASSERT((n2->nc_flag & NCF_TS) != 0,
2322 n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2323 n2_ts->nc_time = ncp_ts->nc_time;
2324 n2_ts->nc_ticks = ncp_ts->nc_ticks;
2326 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2327 n2_ts->nc_nc.nc_flag |= NCF_DTS;
2331 SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2333 goto out_unlock_free;
2337 if (flag == NCF_ISDOTDOT) {
2339 * See if we are trying to add .. entry, but some other lookup
2340 * has populated v_cache_dd pointer already.
2342 if (dvp->v_cache_dd != NULL)
2343 goto out_unlock_free;
2344 KASSERT(vp == NULL || vp->v_type == VDIR,
2345 ("wrong vnode type %p", vp));
2346 vn_seqc_write_begin(dvp);
2347 dvp->v_cache_dd = ncp;
2348 vn_seqc_write_end(dvp);
2352 if (flag != NCF_ISDOTDOT) {
2354 * For this case, the cache entry maps both the
2355 * directory name in it and the name ".." for the
2356 * directory's parent.
2358 vn_seqc_write_begin(vp);
2359 if ((ndd = vp->v_cache_dd) != NULL) {
2360 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2361 cache_zap_locked(ndd);
2365 vp->v_cache_dd = ncp;
2366 vn_seqc_write_end(vp);
2367 } else if (vp->v_type != VDIR) {
2368 if (vp->v_cache_dd != NULL) {
2369 vn_seqc_write_begin(vp);
2370 vp->v_cache_dd = NULL;
2371 vn_seqc_write_end(vp);
2376 if (flag != NCF_ISDOTDOT) {
2377 if (LIST_EMPTY(&dvp->v_cache_src)) {
2378 cache_hold_vnode(dvp);
2380 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2384 * If the entry is "negative", we place it into the
2385 * "negative" cache queue, otherwise, we place it into the
2386 * destination vnode's cache entries queue.
2389 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2390 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2393 if (cnp->cn_flags & ISWHITEOUT)
2394 ncp->nc_flag |= NCF_WHITE;
2395 cache_neg_insert(ncp);
2396 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2401 * Insert the new namecache entry into the appropriate chain
2402 * within the cache entries table.
2404 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2406 atomic_thread_fence_rel();
2408 * Mark the entry as fully constructed.
2409 * It is immutable past this point until its removal.
2411 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2413 cache_enter_unlock(&cel);
2418 cache_enter_unlock(&cel);
2424 cache_roundup_2(u_int val)
2428 for (res = 1; res <= val; res <<= 1)
2434 static struct nchashhead *
2435 nchinittbl(u_long elements, u_long *hashmask)
2437 struct nchashhead *hashtbl;
2440 hashsize = cache_roundup_2(elements) / 2;
2442 hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2443 for (i = 0; i < hashsize; i++)
2444 CK_SLIST_INIT(&hashtbl[i]);
2445 *hashmask = hashsize - 1;
2450 ncfreetbl(struct nchashhead *hashtbl)
2453 free(hashtbl, M_VFSCACHE);
2457 * Name cache initialization, from vfs_init() when we are booting
2460 nchinit(void *dummy __unused)
2464 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2465 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2466 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2467 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2468 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2469 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2470 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2471 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2473 VFS_SMR_ZONE_SET(cache_zone_small);
2474 VFS_SMR_ZONE_SET(cache_zone_small_ts);
2475 VFS_SMR_ZONE_SET(cache_zone_large);
2476 VFS_SMR_ZONE_SET(cache_zone_large_ts);
2478 ncsize = desiredvnodes * ncsizefactor;
2479 cache_recalc_neg_min(ncnegminpct);
2480 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2481 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2482 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2484 if (ncbuckethash > nchash)
2485 ncbuckethash = nchash;
2486 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2488 for (i = 0; i < numbucketlocks; i++)
2489 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2490 ncvnodehash = ncbuckethash;
2491 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2493 for (i = 0; i < numvnodelocks; i++)
2494 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2496 for (i = 0; i < numneglists; i++) {
2497 mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2498 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2499 TAILQ_INIT(&neglists[i].nl_list);
2500 TAILQ_INIT(&neglists[i].nl_hotlist);
2503 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2506 cache_vnode_init(struct vnode *vp)
2509 LIST_INIT(&vp->v_cache_src);
2510 TAILQ_INIT(&vp->v_cache_dst);
2511 vp->v_cache_dd = NULL;
2516 cache_changesize(u_long newmaxvnodes)
2518 struct nchashhead *new_nchashtbl, *old_nchashtbl;
2519 u_long new_nchash, old_nchash;
2520 struct namecache *ncp;
2525 newncsize = newmaxvnodes * ncsizefactor;
2526 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2527 if (newmaxvnodes < numbucketlocks)
2528 newmaxvnodes = numbucketlocks;
2530 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2531 /* If same hash table size, nothing to do */
2532 if (nchash == new_nchash) {
2533 ncfreetbl(new_nchashtbl);
2537 * Move everything from the old hash table to the new table.
2538 * None of the namecache entries in the table can be removed
2539 * because to do so, they have to be removed from the hash table.
2541 cache_lock_all_vnodes();
2542 cache_lock_all_buckets();
2543 old_nchashtbl = nchashtbl;
2544 old_nchash = nchash;
2545 nchashtbl = new_nchashtbl;
2546 nchash = new_nchash;
2547 for (i = 0; i <= old_nchash; i++) {
2548 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2549 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2551 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2552 CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2556 cache_recalc_neg_min(ncnegminpct);
2557 cache_unlock_all_buckets();
2558 cache_unlock_all_vnodes();
2559 ncfreetbl(old_nchashtbl);
2563 * Invalidate all entries from and to a particular vnode.
2566 cache_purge_impl(struct vnode *vp)
2568 struct cache_freebatch batch;
2569 struct namecache *ncp;
2570 struct mtx *vlp, *vlp2;
2573 vlp = VP2VNODELOCK(vp);
2577 while (!LIST_EMPTY(&vp->v_cache_src)) {
2578 ncp = LIST_FIRST(&vp->v_cache_src);
2579 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2581 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2583 while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2584 ncp = TAILQ_FIRST(&vp->v_cache_dst);
2585 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2587 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2589 ncp = vp->v_cache_dd;
2591 KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2592 ("lost dotdot link"));
2593 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2595 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2597 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2601 cache_free_batch(&batch);
2605 * Opportunistic check to see if there is anything to do.
2608 cache_has_entries(struct vnode *vp)
2611 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2612 vp->v_cache_dd == NULL)
2618 cache_purge(struct vnode *vp)
2621 SDT_PROBE1(vfs, namecache, purge, done, vp);
2622 if (!cache_has_entries(vp))
2624 cache_purge_impl(vp);
2628 * Only to be used by vgone.
2631 cache_purge_vgone(struct vnode *vp)
2635 VNPASS(VN_IS_DOOMED(vp), vp);
2636 if (cache_has_entries(vp)) {
2637 cache_purge_impl(vp);
2642 * Serialize against a potential thread doing cache_purge.
2644 vlp = VP2VNODELOCK(vp);
2645 mtx_wait_unlocked(vlp);
2646 if (cache_has_entries(vp)) {
2647 cache_purge_impl(vp);
2654 * Invalidate all negative entries for a particular directory vnode.
2657 cache_purge_negative(struct vnode *vp)
2659 struct cache_freebatch batch;
2660 struct namecache *ncp, *nnp;
2663 SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2664 if (LIST_EMPTY(&vp->v_cache_src))
2667 vlp = VP2VNODELOCK(vp);
2669 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2670 if (!(ncp->nc_flag & NCF_NEGATIVE))
2672 cache_zap_negative_locked_vnode_kl(ncp, vp);
2673 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2676 cache_free_batch(&batch);
2680 * Entry points for modifying VOP operations.
2683 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2684 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2687 ASSERT_VOP_IN_SEQC(fdvp);
2688 ASSERT_VOP_IN_SEQC(fvp);
2689 ASSERT_VOP_IN_SEQC(tdvp);
2691 ASSERT_VOP_IN_SEQC(tvp);
2696 KASSERT(!cache_remove_cnp(tdvp, tcnp),
2697 ("%s: lingering negative entry", __func__));
2699 cache_remove_cnp(tdvp, tcnp);
2704 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
2707 ASSERT_VOP_IN_SEQC(dvp);
2708 ASSERT_VOP_IN_SEQC(vp);
2714 * Validate that if an entry exists it matches.
2717 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2719 struct namecache *ncp;
2723 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2724 if (CK_SLIST_EMPTY(NCHHASH(hash)))
2726 blp = HASH2BUCKETLOCK(hash);
2728 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2729 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2730 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) {
2731 if (ncp->nc_vp != vp)
2732 panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p vp %p\n",
2733 __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp,
2742 * Flush all entries referencing a particular filesystem.
2745 cache_purgevfs(struct mount *mp)
2747 struct vnode *vp, *mvp;
2749 SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2751 * Somewhat wasteful iteration over all vnodes. Would be better to
2752 * support filtering and avoid the interlock to begin with.
2754 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2755 if (!cache_has_entries(vp)) {
2767 * Perform canonical checks and cache lookup and pass on to filesystem
2768 * through the vop_cachedlookup only if needed.
2772 vfs_cache_lookup(struct vop_lookup_args *ap)
2776 struct vnode **vpp = ap->a_vpp;
2777 struct componentname *cnp = ap->a_cnp;
2778 int flags = cnp->cn_flags;
2783 if (dvp->v_type != VDIR)
2786 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2787 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2790 error = vn_dir_check_exec(dvp, cnp);
2794 error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2796 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2802 /* Implementation of the getcwd syscall. */
2804 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2810 buflen = uap->buflen;
2811 if (__predict_false(buflen < 2))
2813 if (buflen > MAXPATHLEN)
2814 buflen = MAXPATHLEN;
2816 buf = uma_zalloc(namei_zone, M_WAITOK);
2817 error = vn_getcwd(buf, &retbuf, &buflen);
2819 error = copyout(retbuf, uap->buf, buflen);
2820 uma_zfree(namei_zone, buf);
2825 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
2831 pwd = pwd_get_smr();
2832 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
2834 VFS_SMR_ASSERT_NOT_ENTERED();
2836 pwd = pwd_hold(curthread);
2837 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
2843 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2850 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2851 size_t size, int flags, enum uio_seg pathseg)
2853 struct nameidata nd;
2854 char *retbuf, *freebuf;
2859 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2860 pathseg, path, fd, &cap_fstat_rights, td);
2861 if ((error = namei(&nd)) != 0)
2863 error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size);
2865 error = copyout(retbuf, buf, size);
2866 free(freebuf, M_TEMP);
2873 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2876 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2877 uap->flags, UIO_USERSPACE));
2881 * Retrieve the full filesystem path that correspond to a vnode from the name
2882 * cache (if available)
2885 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
2892 if (__predict_false(vp == NULL))
2895 buflen = MAXPATHLEN;
2896 buf = malloc(buflen, M_TEMP, M_WAITOK);
2898 pwd = pwd_get_smr();
2899 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
2900 VFS_SMR_ASSERT_NOT_ENTERED();
2902 pwd = pwd_hold(curthread);
2903 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
2914 * This function is similar to vn_fullpath, but it attempts to lookup the
2915 * pathname relative to the global root mount point. This is required for the
2916 * auditing sub-system, as audited pathnames must be absolute, relative to the
2917 * global root mount point.
2920 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
2926 if (__predict_false(vp == NULL))
2928 buflen = MAXPATHLEN;
2929 buf = malloc(buflen, M_TEMP, M_WAITOK);
2931 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
2932 VFS_SMR_ASSERT_NOT_ENTERED();
2934 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
2943 static struct namecache *
2944 vn_dd_from_dst(struct vnode *vp)
2946 struct namecache *ncp;
2948 cache_assert_vnode_locked(vp);
2949 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
2950 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2957 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
2960 struct namecache *ncp;
2964 vlp = VP2VNODELOCK(*vp);
2966 ncp = (*vp)->v_cache_dd;
2967 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
2968 KASSERT(ncp == vn_dd_from_dst(*vp),
2969 ("%s: mismatch for dd entry (%p != %p)", __func__,
2970 ncp, vn_dd_from_dst(*vp)));
2972 ncp = vn_dd_from_dst(*vp);
2975 if (*buflen < ncp->nc_nlen) {
2978 counter_u64_add(numfullpathfail4, 1);
2980 SDT_PROBE3(vfs, namecache, fullpath, return, error,
2984 *buflen -= ncp->nc_nlen;
2985 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2986 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2995 SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2998 vn_lock(*vp, LK_SHARED | LK_RETRY);
2999 error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
3002 counter_u64_add(numfullpathfail2, 1);
3003 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3008 if (VN_IS_DOOMED(dvp)) {
3009 /* forced unmount */
3012 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3016 * *vp has its use count incremented still.
3023 * Resolve a directory to a pathname.
3025 * The name of the directory can always be found in the namecache or fetched
3026 * from the filesystem. There is also guaranteed to be only one parent, meaning
3027 * we can just follow vnodes up until we find the root.
3029 * The vnode must be referenced.
3032 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3033 size_t *len, size_t addend)
3035 #ifdef KDTRACE_HOOKS
3036 struct vnode *startvp = vp;
3041 bool slash_prefixed;
3043 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3044 VNPASS(vp->v_usecount > 0, vp);
3048 slash_prefixed = true;
3053 slash_prefixed = false;
3058 SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3059 counter_u64_add(numfullpathcalls, 1);
3060 while (vp != rdir && vp != rootvnode) {
3062 * The vp vnode must be already fully constructed,
3063 * since it is either found in namecache or obtained
3064 * from VOP_VPTOCNP(). We may test for VV_ROOT safely
3065 * without obtaining the vnode lock.
3067 if ((vp->v_vflag & VV_ROOT) != 0) {
3068 vn_lock(vp, LK_RETRY | LK_SHARED);
3071 * With the vnode locked, check for races with
3072 * unmount, forced or not. Note that we
3073 * already verified that vp is not equal to
3074 * the root vnode, which means that
3075 * mnt_vnodecovered can be NULL only for the
3078 if (VN_IS_DOOMED(vp) ||
3079 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3080 vp1->v_mountedhere != vp->v_mount) {
3083 SDT_PROBE3(vfs, namecache, fullpath, return,
3093 if (vp->v_type != VDIR) {
3095 counter_u64_add(numfullpathfail1, 1);
3097 SDT_PROBE3(vfs, namecache, fullpath, return,
3101 error = vn_vptocnp(&vp, buf, &buflen);
3107 SDT_PROBE3(vfs, namecache, fullpath, return, error,
3111 buf[--buflen] = '/';
3112 slash_prefixed = true;
3116 if (!slash_prefixed) {
3119 counter_u64_add(numfullpathfail4, 1);
3120 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3124 buf[--buflen] = '/';
3126 counter_u64_add(numfullpathfound, 1);
3129 *retbuf = buf + buflen;
3130 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3137 * Resolve an arbitrary vnode to a pathname.
3140 * - hardlinks are not tracked, thus if the vnode is not a directory this can
3141 * resolve to a different path than the one used to find it
3142 * - namecache is not mandatory, meaning names are not guaranteed to be added
3143 * (in which case resolving fails)
3145 static void __inline
3146 cache_rev_failed_impl(int *reason, int line)
3151 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__)
3154 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3155 char **retbuf, size_t *buflen, size_t addend)
3157 #ifdef KDTRACE_HOOKS
3158 struct vnode *startvp = vp;
3162 struct namecache *ncp;
3166 #ifdef KDTRACE_HOOKS
3169 seqc_t vp_seqc, tvp_seqc;
3172 VFS_SMR_ASSERT_ENTERED();
3174 if (!cache_fast_revlookup) {
3179 orig_buflen = *buflen;
3182 MPASS(*buflen >= 2);
3184 buf[*buflen] = '\0';
3187 if (vp == rdir || vp == rootvnode) {
3195 #ifdef KDTRACE_HOOKS
3199 ncp = NULL; /* for sdt probe down below */
3200 vp_seqc = vn_seqc_read_any(vp);
3201 if (seqc_in_modify(vp_seqc)) {
3202 cache_rev_failed(&reason);
3207 #ifdef KDTRACE_HOOKS
3210 if ((vp->v_vflag & VV_ROOT) != 0) {
3211 mp = atomic_load_ptr(&vp->v_mount);
3213 cache_rev_failed(&reason);
3216 tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3217 tvp_seqc = vn_seqc_read_any(tvp);
3218 if (seqc_in_modify(tvp_seqc)) {
3219 cache_rev_failed(&reason);
3222 if (!vn_seqc_consistent(vp, vp_seqc)) {
3223 cache_rev_failed(&reason);
3230 ncp = atomic_load_ptr(&vp->v_cache_dd);
3232 cache_rev_failed(&reason);
3235 nc_flag = atomic_load_char(&ncp->nc_flag);
3236 if ((nc_flag & NCF_ISDOTDOT) != 0) {
3237 cache_rev_failed(&reason);
3240 if (!cache_ncp_canuse(ncp)) {
3241 cache_rev_failed(&reason);
3244 if (ncp->nc_nlen >= *buflen) {
3245 cache_rev_failed(&reason);
3249 *buflen -= ncp->nc_nlen;
3250 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3254 tvp_seqc = vn_seqc_read_any(tvp);
3255 if (seqc_in_modify(tvp_seqc)) {
3256 cache_rev_failed(&reason);
3259 if (!vn_seqc_consistent(vp, vp_seqc)) {
3260 cache_rev_failed(&reason);
3265 if (vp == rdir || vp == rootvnode)
3270 *retbuf = buf + *buflen;
3271 *buflen = orig_buflen - *buflen + addend;
3272 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3276 *buflen = orig_buflen;
3277 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3283 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3286 size_t orig_buflen, addend;
3292 orig_buflen = *buflen;
3296 if (vp->v_type != VDIR) {
3298 buf[*buflen] = '\0';
3299 error = vn_vptocnp(&vp, buf, buflen);
3308 addend = orig_buflen - *buflen;
3311 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3315 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3317 * Since the namecache does not track hardlinks, the caller is expected to first
3318 * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
3320 * Then we have 2 cases:
3321 * - if the found vnode is a directory, the path can be constructed just by
3322 * following names up the chain
3323 * - otherwise we populate the buffer with the saved name and start resolving
3327 vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf,
3332 struct componentname *cnp;
3340 if (*buflen > MAXPATHLEN)
3341 *buflen = MAXPATHLEN;
3343 buf = malloc(*buflen, M_TEMP, M_WAITOK);
3348 * Check for VBAD to work around the vp_crossmp bug in lookup().
3350 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3351 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3352 * If the type is VDIR (like in this very case) we can skip looking
3353 * at ni_dvp in the first place. However, since vnodes get passed here
3354 * unlocked the target may transition to doomed state (type == VBAD)
3355 * before we get to evaluate the condition. If this happens, we will
3356 * populate part of the buffer and descend to vn_fullpath_dir with
3357 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3359 * This should be atomic_load(&vp->v_type) but it is illegal to take
3360 * an address of a bit field, even if said field is sized to char.
3361 * Work around the problem by reading the value into a full-sized enum
3362 * and then re-reading it with atomic_load which will still prevent
3363 * the compiler from re-reading down the road.
3366 type = atomic_load_int(&type);
3373 addend = cnp->cn_namelen + 2;
3374 if (*buflen < addend) {
3379 tmpbuf = buf + *buflen;
3381 memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
3382 tmpbuf[addend - 1] = '\0';
3387 pwd = pwd_get_smr();
3388 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3390 VFS_SMR_ASSERT_NOT_ENTERED();
3392 pwd = pwd_hold(curthread);
3394 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3410 vn_dir_dd_ino(struct vnode *vp)
3412 struct namecache *ncp;
3417 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3418 vlp = VP2VNODELOCK(vp);
3420 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3421 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3424 vs = vget_prep(ddvp);
3426 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3435 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3437 struct namecache *ncp;
3441 vlp = VP2VNODELOCK(vp);
3443 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3444 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3450 l = min(ncp->nc_nlen, buflen - 1);
3451 memcpy(buf, ncp->nc_name, l);
3458 * This function updates path string to vnode's full global path
3459 * and checks the size of the new path string against the pathlen argument.
3461 * Requires a locked, referenced vnode.
3462 * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3464 * If vp is a directory, the call to vn_fullpath_global() always succeeds
3465 * because it falls back to the ".." lookup if the namecache lookup fails.
3468 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3471 struct nameidata nd;
3476 ASSERT_VOP_ELOCKED(vp, __func__);
3478 /* Construct global filesystem path from vp. */
3480 error = vn_fullpath_global(vp, &rpath, &fbuf);
3487 if (strlen(rpath) >= pathlen) {
3489 error = ENAMETOOLONG;
3494 * Re-lookup the vnode by path to detect a possible rename.
3495 * As a side effect, the vnode is relocked.
3496 * If vnode was renamed, return ENOENT.
3498 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3499 UIO_SYSSPACE, path, td);
3505 NDFREE(&nd, NDF_ONLY_PNBUF);
3509 strcpy(path, rpath);
3522 db_print_vpath(struct vnode *vp)
3525 while (vp != NULL) {
3526 db_printf("%p: ", vp);
3527 if (vp == rootvnode) {
3531 if (vp->v_vflag & VV_ROOT) {
3532 db_printf("<mount point>");
3533 vp = vp->v_mount->mnt_vnodecovered;
3535 struct namecache *ncp;
3539 ncp = TAILQ_FIRST(&vp->v_cache_dst);
3542 for (i = 0; i < ncp->nc_nlen; i++)
3543 db_printf("%c", *ncn++);
3556 DB_SHOW_COMMAND(vpath, db_show_vpath)
3561 db_printf("usage: show vpath <struct vnode *>\n");
3565 vp = (struct vnode *)addr;
3571 static bool __read_frequently cache_fast_lookup = true;
3572 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_lookup, CTLFLAG_RW,
3573 &cache_fast_lookup, 0, "");
3575 #define CACHE_FPL_FAILED -2020
3578 * Components of nameidata (or objects it can point to) which may
3579 * need restoring in case fast path lookup fails.
3581 struct nameidata_saved {
3589 struct nameidata *ndp;
3590 struct componentname *cnp;
3596 struct nameidata_saved snd;
3598 enum cache_fpl_status status:8;
3604 cache_fpl_cleanup_cnp(struct componentname *cnp)
3607 uma_zfree(namei_zone, cnp->cn_pnbuf);
3609 cnp->cn_pnbuf = NULL;
3610 cnp->cn_nameptr = NULL;
3614 static struct vnode *
3615 cache_fpl_handle_root(struct cache_fpl *fpl)
3617 struct nameidata *ndp;
3618 struct componentname *cnp;
3623 while (*(cnp->cn_nameptr) == '/') {
3628 return (ndp->ni_rootdir);
3632 cache_fpl_checkpoint(struct cache_fpl *fpl, struct nameidata_saved *snd)
3635 snd->cn_flags = fpl->ndp->ni_cnd.cn_flags;
3636 snd->cn_namelen = fpl->ndp->ni_cnd.cn_namelen;
3637 snd->cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3638 snd->ni_pathlen = fpl->ndp->ni_pathlen;
3642 cache_fpl_restore_partial(struct cache_fpl *fpl, struct nameidata_saved *snd)
3645 fpl->ndp->ni_cnd.cn_flags = snd->cn_flags;
3646 fpl->ndp->ni_cnd.cn_namelen = snd->cn_namelen;
3647 fpl->ndp->ni_cnd.cn_nameptr = snd->cn_nameptr;
3648 fpl->ndp->ni_pathlen = snd->ni_pathlen;
3652 cache_fpl_restore_abort(struct cache_fpl *fpl, struct nameidata_saved *snd)
3655 cache_fpl_restore_partial(fpl, snd);
3657 * It is 0 on entry by API contract.
3659 fpl->ndp->ni_resflags = 0;
3663 #define cache_fpl_smr_assert_entered(fpl) ({ \
3664 struct cache_fpl *_fpl = (fpl); \
3665 MPASS(_fpl->in_smr == true); \
3666 VFS_SMR_ASSERT_ENTERED(); \
3668 #define cache_fpl_smr_assert_not_entered(fpl) ({ \
3669 struct cache_fpl *_fpl = (fpl); \
3670 MPASS(_fpl->in_smr == false); \
3671 VFS_SMR_ASSERT_NOT_ENTERED(); \
3674 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
3675 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
3678 #define cache_fpl_smr_enter_initial(fpl) ({ \
3679 struct cache_fpl *_fpl = (fpl); \
3681 _fpl->in_smr = true; \
3684 #define cache_fpl_smr_enter(fpl) ({ \
3685 struct cache_fpl *_fpl = (fpl); \
3686 MPASS(_fpl->in_smr == false); \
3688 _fpl->in_smr = true; \
3691 #define cache_fpl_smr_exit(fpl) ({ \
3692 struct cache_fpl *_fpl = (fpl); \
3693 MPASS(_fpl->in_smr == true); \
3695 _fpl->in_smr = false; \
3699 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
3702 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3703 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3704 ("%s: converting to abort from %d at %d, set at %d\n",
3705 __func__, fpl->status, line, fpl->line));
3707 fpl->status = CACHE_FPL_STATUS_ABORTED;
3709 return (CACHE_FPL_FAILED);
3712 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__)
3715 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
3718 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3719 ("%s: setting to partial at %d, but already set to %d at %d\n",
3720 __func__, line, fpl->status, fpl->line));
3721 cache_fpl_smr_assert_entered(fpl);
3722 fpl->status = CACHE_FPL_STATUS_PARTIAL;
3724 return (CACHE_FPL_FAILED);
3727 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__)
3730 cache_fpl_handled_impl(struct cache_fpl *fpl, int error, int line)
3733 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3734 ("%s: setting to handled at %d, but already set to %d at %d\n",
3735 __func__, line, fpl->status, fpl->line));
3736 cache_fpl_smr_assert_not_entered(fpl);
3737 MPASS(error != CACHE_FPL_FAILED);
3738 fpl->status = CACHE_FPL_STATUS_HANDLED;
3743 #define cache_fpl_handled(x, e) cache_fpl_handled_impl((x), (e), __LINE__)
3746 cache_fpl_terminated(struct cache_fpl *fpl)
3749 return (fpl->status != CACHE_FPL_STATUS_UNSET);
3752 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
3753 (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
3754 FAILIFEXISTS | FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | \
3755 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
3757 #define CACHE_FPL_INTERNAL_CN_FLAGS \
3758 (ISDOTDOT | MAKEENTRY | ISLASTCN)
3760 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
3761 "supported and internal flags overlap");
3763 static bool cache_fplookup_is_mp(struct cache_fpl *fpl);
3764 static int cache_fplookup_cross_mount(struct cache_fpl *fpl);
3767 cache_fpl_islastcn(struct nameidata *ndp)
3770 return (*ndp->ni_next == 0);
3774 cache_fpl_isdotdot(struct componentname *cnp)
3777 if (cnp->cn_namelen == 2 &&
3778 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
3784 cache_can_fplookup(struct cache_fpl *fpl)
3786 struct nameidata *ndp;
3787 struct componentname *cnp;
3792 td = cnp->cn_thread;
3794 if (!cache_fast_lookup) {
3795 cache_fpl_aborted(fpl);
3799 if (mac_vnode_check_lookup_enabled()) {
3800 cache_fpl_aborted(fpl);
3804 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
3805 cache_fpl_aborted(fpl);
3808 if (IN_CAPABILITY_MODE(td)) {
3809 cache_fpl_aborted(fpl);
3812 if (AUDITING_TD(td)) {
3813 cache_fpl_aborted(fpl);
3816 if (ndp->ni_startdir != NULL) {
3817 cache_fpl_aborted(fpl);
3824 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
3826 struct nameidata *ndp;
3831 error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch);
3832 if (__predict_false(error != 0)) {
3833 cache_fpl_smr_exit(fpl);
3834 return (cache_fpl_aborted(fpl));
3836 fpl->fsearch = fsearch;
3841 cache_fplookup_vnode_supported(struct vnode *vp)
3844 return (vp->v_type != VLNK);
3847 static int __noinline
3848 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
3851 struct componentname *cnp;
3857 cache_fpl_smr_exit(fpl);
3858 if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
3859 return (cache_fpl_handled(fpl, ENOENT));
3861 return (cache_fpl_aborted(fpl));
3865 * The target vnode is not supported, prepare for the slow path to take over.
3867 static int __noinline
3868 cache_fplookup_partial_setup(struct cache_fpl *fpl)
3870 struct nameidata *ndp;
3871 struct componentname *cnp;
3881 dvp_seqc = fpl->dvp_seqc;
3883 if (!pwd_hold_smr(pwd)) {
3884 cache_fpl_smr_exit(fpl);
3885 return (cache_fpl_aborted(fpl));
3889 * Note that seqc is checked before the vnode is locked, so by
3890 * the time regular lookup gets to it it may have moved.
3892 * Ultimately this does not affect correctness, any lookup errors
3893 * are userspace racing with itself. It is guaranteed that any
3894 * path which ultimatley gets found could also have been found
3895 * by regular lookup going all the way in absence of concurrent
3898 dvs = vget_prep_smr(dvp);
3899 cache_fpl_smr_exit(fpl);
3900 if (__predict_false(dvs == VGET_NONE)) {
3902 return (cache_fpl_aborted(fpl));
3905 vget_finish_ref(dvp, dvs);
3906 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3909 return (cache_fpl_aborted(fpl));
3912 cache_fpl_restore_partial(fpl, &fpl->snd);
3914 ndp->ni_startdir = dvp;
3915 cnp->cn_flags |= MAKEENTRY;
3916 if (cache_fpl_islastcn(ndp))
3917 cnp->cn_flags |= ISLASTCN;
3918 if (cache_fpl_isdotdot(cnp))
3919 cnp->cn_flags |= ISDOTDOT;
3925 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
3927 struct componentname *cnp;
3934 tvp_seqc = fpl->tvp_seqc;
3936 if ((cnp->cn_flags & LOCKLEAF) != 0) {
3937 lkflags = LK_SHARED;
3938 if ((cnp->cn_flags & LOCKSHARED) == 0)
3939 lkflags = LK_EXCLUSIVE;
3940 error = vget_finish(tvp, lkflags, tvs);
3941 if (__predict_false(error != 0)) {
3942 return (cache_fpl_aborted(fpl));
3945 vget_finish_ref(tvp, tvs);
3948 if (!vn_seqc_consistent(tvp, tvp_seqc)) {
3949 if ((cnp->cn_flags & LOCKLEAF) != 0)
3953 return (cache_fpl_aborted(fpl));
3956 return (cache_fpl_handled(fpl, 0));
3960 * They want to possibly modify the state of the namecache.
3962 static int __noinline
3963 cache_fplookup_final_modifying(struct cache_fpl *fpl)
3965 struct nameidata *ndp;
3966 struct componentname *cnp;
3968 struct vnode *dvp, *tvp;
3977 dvp_seqc = fpl->dvp_seqc;
3979 MPASS(cache_fpl_islastcn(ndp));
3980 if ((cnp->cn_flags & LOCKPARENT) == 0)
3981 MPASS((cnp->cn_flags & WANTPARENT) != 0);
3982 MPASS((cnp->cn_flags & TRAILINGSLASH) == 0);
3983 MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE ||
3984 cnp->cn_nameiop == RENAME);
3985 MPASS((cnp->cn_flags & MAKEENTRY) == 0);
3986 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
3988 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
3989 if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)
3992 mp = atomic_load_ptr(&dvp->v_mount);
3993 if (__predict_false(mp == NULL)) {
3994 return (cache_fpl_aborted(fpl));
3997 if (__predict_false(mp->mnt_flag & MNT_RDONLY)) {
3998 cache_fpl_smr_exit(fpl);
4000 * Original code keeps not checking for CREATE which
4001 * might be a bug. For now let the old lookup decide.
4003 if (cnp->cn_nameiop == CREATE) {
4004 return (cache_fpl_aborted(fpl));
4006 return (cache_fpl_handled(fpl, EROFS));
4009 if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) {
4010 cache_fpl_smr_exit(fpl);
4011 return (cache_fpl_handled(fpl, EEXIST));
4015 * Secure access to dvp; check cache_fplookup_partial_setup for
4018 * XXX At least UFS requires its lookup routine to be called for
4019 * the last path component, which leads to some level of complicaton
4021 * - the target routine always locks the target vnode, but our caller
4022 * may not need it locked
4023 * - some of the VOP machinery asserts that the parent is locked, which
4024 * once more may be not required
4026 * TODO: add a flag for filesystems which don't need this.
4028 dvs = vget_prep_smr(dvp);
4029 cache_fpl_smr_exit(fpl);
4030 if (__predict_false(dvs == VGET_NONE)) {
4031 return (cache_fpl_aborted(fpl));
4034 vget_finish_ref(dvp, dvs);
4035 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4037 return (cache_fpl_aborted(fpl));
4040 error = vn_lock(dvp, LK_EXCLUSIVE);
4041 if (__predict_false(error != 0)) {
4043 return (cache_fpl_aborted(fpl));
4047 cnp->cn_flags |= ISLASTCN;
4049 cnp->cn_flags |= MAKEENTRY;
4050 if (cache_fpl_isdotdot(cnp))
4051 cnp->cn_flags |= ISDOTDOT;
4052 cnp->cn_lkflags = LK_EXCLUSIVE;
4053 error = VOP_LOOKUP(dvp, &tvp, cnp);
4061 return (cache_fpl_handled(fpl, error));
4064 return (cache_fpl_aborted(fpl));
4070 if ((cnp->cn_flags & SAVESTART) != 0) {
4071 ndp->ni_startdir = dvp;
4072 vrefact(ndp->ni_startdir);
4073 cnp->cn_flags |= SAVENAME;
4075 MPASS(error == EJUSTRETURN);
4076 if ((cnp->cn_flags & LOCKPARENT) == 0) {
4079 return (cache_fpl_handled(fpl, 0));
4083 * There are very hairy corner cases concerning various flag combinations
4084 * and locking state. In particular here we only hold one lock instead of
4087 * Skip the complexity as it is of no significance for normal workloads.
4089 if (__predict_false(tvp == dvp)) {
4092 return (cache_fpl_aborted(fpl));
4096 * Check if the target is either a symlink or a mount point.
4097 * Since we expect this to be the terminal vnode it should
4098 * almost never be true.
4100 if (__predict_false(!cache_fplookup_vnode_supported(tvp) ||
4101 cache_fplookup_is_mp(fpl))) {
4104 return (cache_fpl_aborted(fpl));
4107 if ((cnp->cn_flags & FAILIFEXISTS) != 0) {
4110 return (cache_fpl_handled(fpl, EEXIST));
4113 if ((cnp->cn_flags & LOCKLEAF) == 0) {
4117 if ((cnp->cn_flags & LOCKPARENT) == 0) {
4121 if ((cnp->cn_flags & SAVESTART) != 0) {
4122 ndp->ni_startdir = dvp;
4123 vrefact(ndp->ni_startdir);
4124 cnp->cn_flags |= SAVENAME;
4127 return (cache_fpl_handled(fpl, 0));
4130 static int __noinline
4131 cache_fplookup_modifying(struct cache_fpl *fpl)
4133 struct nameidata *ndp;
4137 if (!cache_fpl_islastcn(ndp)) {
4138 return (cache_fpl_partial(fpl));
4140 return (cache_fplookup_final_modifying(fpl));
4143 static int __noinline
4144 cache_fplookup_final_withparent(struct cache_fpl *fpl)
4146 struct componentname *cnp;
4147 enum vgetstate dvs, tvs;
4148 struct vnode *dvp, *tvp;
4154 dvp_seqc = fpl->dvp_seqc;
4157 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
4160 * This is less efficient than it can be for simplicity.
4162 dvs = vget_prep_smr(dvp);
4163 if (__predict_false(dvs == VGET_NONE)) {
4164 return (cache_fpl_aborted(fpl));
4166 tvs = vget_prep_smr(tvp);
4167 if (__predict_false(tvs == VGET_NONE)) {
4168 cache_fpl_smr_exit(fpl);
4169 vget_abort(dvp, dvs);
4170 return (cache_fpl_aborted(fpl));
4173 cache_fpl_smr_exit(fpl);
4175 if ((cnp->cn_flags & LOCKPARENT) != 0) {
4176 error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
4177 if (__predict_false(error != 0)) {
4178 vget_abort(tvp, tvs);
4179 return (cache_fpl_aborted(fpl));
4182 vget_finish_ref(dvp, dvs);
4185 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4186 vget_abort(tvp, tvs);
4187 if ((cnp->cn_flags & LOCKPARENT) != 0)
4191 return (cache_fpl_aborted(fpl));
4194 error = cache_fplookup_final_child(fpl, tvs);
4195 if (__predict_false(error != 0)) {
4196 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED);
4197 if ((cnp->cn_flags & LOCKPARENT) != 0)
4204 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
4209 cache_fplookup_final(struct cache_fpl *fpl)
4211 struct componentname *cnp;
4213 struct vnode *dvp, *tvp;
4218 dvp_seqc = fpl->dvp_seqc;
4221 if (cnp->cn_nameiop != LOOKUP) {
4222 return (cache_fplookup_final_modifying(fpl));
4225 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
4226 return (cache_fplookup_final_withparent(fpl));
4228 tvs = vget_prep_smr(tvp);
4229 if (__predict_false(tvs == VGET_NONE)) {
4230 return (cache_fpl_partial(fpl));
4233 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4234 cache_fpl_smr_exit(fpl);
4235 vget_abort(tvp, tvs);
4236 return (cache_fpl_aborted(fpl));
4239 cache_fpl_smr_exit(fpl);
4240 return (cache_fplookup_final_child(fpl, tvs));
4244 * Comment from locked lookup:
4245 * Check for degenerate name (e.g. / or "") which is a way of talking about a
4246 * directory, e.g. like "/." or ".".
4248 static int __noinline
4249 cache_fplookup_degenerate(struct cache_fpl *fpl)
4251 struct componentname *cnp;
4256 fpl->tvp = fpl->dvp;
4257 fpl->tvp_seqc = fpl->dvp_seqc;
4262 if (__predict_false(cnp->cn_nameiop != LOOKUP)) {
4263 cache_fpl_smr_exit(fpl);
4264 return (cache_fpl_handled(fpl, EISDIR));
4267 MPASS((cnp->cn_flags & SAVESTART) == 0);
4269 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) {
4270 return (cache_fplookup_final_withparent(fpl));
4273 dvs = vget_prep_smr(dvp);
4274 cache_fpl_smr_exit(fpl);
4275 if (__predict_false(dvs == VGET_NONE)) {
4276 return (cache_fpl_aborted(fpl));
4279 if ((cnp->cn_flags & LOCKLEAF) != 0) {
4280 lkflags = LK_SHARED;
4281 if ((cnp->cn_flags & LOCKSHARED) == 0)
4282 lkflags = LK_EXCLUSIVE;
4283 error = vget_finish(dvp, lkflags, dvs);
4284 if (__predict_false(error != 0)) {
4285 return (cache_fpl_aborted(fpl));
4288 vget_finish_ref(dvp, dvs);
4290 return (cache_fpl_handled(fpl, 0));
4293 static int __noinline
4294 cache_fplookup_noentry(struct cache_fpl *fpl)
4296 struct nameidata *ndp;
4297 struct componentname *cnp;
4299 struct vnode *dvp, *tvp;
4307 dvp_seqc = fpl->dvp_seqc;
4309 MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4310 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4311 MPASS(!cache_fpl_isdotdot(cnp));
4313 if (cnp->cn_nameiop != LOOKUP) {
4315 return (cache_fplookup_modifying(fpl));
4318 MPASS((cnp->cn_flags & SAVESTART) == 0);
4321 * Only try to fill in the component if it is the last one,
4322 * otherwise not only there may be several to handle but the
4323 * walk may be complicated.
4325 if (!cache_fpl_islastcn(ndp)) {
4326 return (cache_fpl_partial(fpl));
4330 * Secure access to dvp; check cache_fplookup_partial_setup for
4333 dvs = vget_prep_smr(dvp);
4334 cache_fpl_smr_exit(fpl);
4335 if (__predict_false(dvs == VGET_NONE)) {
4336 return (cache_fpl_aborted(fpl));
4339 vget_finish_ref(dvp, dvs);
4340 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4342 return (cache_fpl_aborted(fpl));
4345 error = vn_lock(dvp, LK_SHARED);
4346 if (__predict_false(error != 0)) {
4348 return (cache_fpl_aborted(fpl));
4353 * TODO: provide variants which don't require locking either vnode.
4355 cnp->cn_flags |= ISLASTCN;
4356 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4358 cnp->cn_flags |= MAKEENTRY;
4359 cnp->cn_lkflags = LK_SHARED;
4360 if ((cnp->cn_flags & LOCKSHARED) == 0) {
4361 cnp->cn_lkflags = LK_EXCLUSIVE;
4363 error = VOP_LOOKUP(dvp, &tvp, cnp);
4371 return (cache_fpl_handled(fpl, error));
4374 return (cache_fpl_aborted(fpl));
4380 MPASS(error == EJUSTRETURN);
4381 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4383 } else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4386 return (cache_fpl_handled(fpl, 0));
4389 if (__predict_false(!cache_fplookup_vnode_supported(tvp) ||
4390 cache_fplookup_is_mp(fpl))) {
4393 return (cache_fpl_aborted(fpl));
4396 if ((cnp->cn_flags & LOCKLEAF) == 0) {
4400 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4402 } else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4405 return (cache_fpl_handled(fpl, 0));
4408 static int __noinline
4409 cache_fplookup_dot(struct cache_fpl *fpl)
4413 MPASS(!seqc_in_modify(fpl->dvp_seqc));
4415 * Just re-assign the value. seqc will be checked later for the first
4416 * non-dot path component in line and/or before deciding to return the
4419 fpl->tvp = fpl->dvp;
4420 fpl->tvp_seqc = fpl->dvp_seqc;
4421 if (cache_fplookup_is_mp(fpl)) {
4422 error = cache_fplookup_cross_mount(fpl);
4423 if (__predict_false(error != 0)) {
4428 counter_u64_add(dothits, 1);
4429 SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp);
4433 static int __noinline
4434 cache_fplookup_dotdot(struct cache_fpl *fpl)
4436 struct nameidata *ndp;
4437 struct componentname *cnp;
4438 struct namecache *ncp;
4447 MPASS(cache_fpl_isdotdot(cnp));
4450 * XXX this is racy the same way regular lookup is
4452 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
4454 if (dvp == pr->pr_root)
4457 if (dvp == ndp->ni_rootdir ||
4458 dvp == ndp->ni_topdir ||
4462 fpl->tvp_seqc = vn_seqc_read_any(dvp);
4463 if (seqc_in_modify(fpl->tvp_seqc)) {
4464 return (cache_fpl_aborted(fpl));
4469 if ((dvp->v_vflag & VV_ROOT) != 0) {
4472 * The opposite of climb mount is needed here.
4474 return (cache_fpl_aborted(fpl));
4477 ncp = atomic_load_ptr(&dvp->v_cache_dd);
4479 return (cache_fpl_aborted(fpl));
4482 nc_flag = atomic_load_char(&ncp->nc_flag);
4483 if ((nc_flag & NCF_ISDOTDOT) != 0) {
4484 if ((nc_flag & NCF_NEGATIVE) != 0)
4485 return (cache_fpl_aborted(fpl));
4486 fpl->tvp = ncp->nc_vp;
4488 fpl->tvp = ncp->nc_dvp;
4491 if (!cache_ncp_canuse(ncp)) {
4492 return (cache_fpl_aborted(fpl));
4495 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
4496 if (seqc_in_modify(fpl->tvp_seqc)) {
4497 return (cache_fpl_partial(fpl));
4500 counter_u64_add(dotdothits, 1);
4504 static int __noinline
4505 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
4510 nc_flag = atomic_load_char(&ncp->nc_flag);
4511 MPASS((nc_flag & NCF_NEGATIVE) != 0);
4513 * If they want to create an entry we need to replace this one.
4515 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
4517 return (cache_fplookup_modifying(fpl));
4519 neg_promote = cache_neg_hit_prep(ncp);
4520 if (!cache_fpl_neg_ncp_canuse(ncp)) {
4521 cache_neg_hit_abort(ncp);
4522 return (cache_fpl_partial(fpl));
4525 return (cache_fplookup_negative_promote(fpl, ncp, hash));
4527 cache_neg_hit_finish(ncp);
4528 cache_fpl_smr_exit(fpl);
4529 return (cache_fpl_handled(fpl, ENOENT));
4533 cache_fplookup_next(struct cache_fpl *fpl)
4535 struct componentname *cnp;
4536 struct namecache *ncp;
4537 struct vnode *dvp, *tvp;
4545 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
4546 if (cnp->cn_namelen == 1) {
4547 return (cache_fplookup_dot(fpl));
4549 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
4550 return (cache_fplookup_dotdot(fpl));
4554 MPASS(!cache_fpl_isdotdot(cnp));
4556 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
4558 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
4559 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
4560 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
4564 if (__predict_false(ncp == NULL)) {
4565 return (cache_fplookup_noentry(fpl));
4568 tvp = atomic_load_ptr(&ncp->nc_vp);
4569 nc_flag = atomic_load_char(&ncp->nc_flag);
4570 if ((nc_flag & NCF_NEGATIVE) != 0) {
4571 return (cache_fplookup_neg(fpl, ncp, hash));
4574 if (!cache_ncp_canuse(ncp)) {
4575 return (cache_fpl_partial(fpl));
4579 fpl->tvp_seqc = vn_seqc_read_any(tvp);
4580 if (seqc_in_modify(fpl->tvp_seqc)) {
4581 return (cache_fpl_partial(fpl));
4584 if (!cache_fplookup_vnode_supported(tvp)) {
4585 return (cache_fpl_partial(fpl));
4588 if (cache_fplookup_is_mp(fpl)) {
4589 error = cache_fplookup_cross_mount(fpl);
4590 if (__predict_false(error != 0)) {
4595 counter_u64_add(numposhits, 1);
4596 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
4601 cache_fplookup_mp_supported(struct mount *mp)
4605 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
4611 * Walk up the mount stack (if any).
4613 * Correctness is provided in the following ways:
4614 * - all vnodes are protected from freeing with SMR
4615 * - struct mount objects are type stable making them always safe to access
4616 * - stability of the particular mount is provided by busying it
4617 * - relationship between the vnode which is mounted on and the mount is
4618 * verified with the vnode sequence counter after busying
4619 * - association between root vnode of the mount and the mount is protected
4622 * From that point on we can read the sequence counter of the root vnode
4623 * and get the next mount on the stack (if any) using the same protection.
4625 * By the end of successful walk we are guaranteed the reached state was
4626 * indeed present at least at some point which matches the regular lookup.
4628 static int __noinline
4629 cache_fplookup_climb_mount(struct cache_fpl *fpl)
4631 struct mount *mp, *prev_mp;
4632 struct mount_pcpu *mpcpu, *prev_mpcpu;
4637 vp_seqc = fpl->tvp_seqc;
4639 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
4640 mp = atomic_load_ptr(&vp->v_mountedhere);
4641 if (__predict_false(mp == NULL)) {
4647 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
4648 if (prev_mp != NULL)
4649 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4650 return (cache_fpl_partial(fpl));
4652 if (prev_mp != NULL)
4653 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4654 if (!vn_seqc_consistent(vp, vp_seqc)) {
4655 vfs_op_thread_exit_crit(mp, mpcpu);
4656 return (cache_fpl_partial(fpl));
4658 if (!cache_fplookup_mp_supported(mp)) {
4659 vfs_op_thread_exit_crit(mp, mpcpu);
4660 return (cache_fpl_partial(fpl));
4662 vp = atomic_load_ptr(&mp->mnt_rootvnode);
4663 if (vp == NULL || VN_IS_DOOMED(vp)) {
4664 vfs_op_thread_exit_crit(mp, mpcpu);
4665 return (cache_fpl_partial(fpl));
4667 vp_seqc = vn_seqc_read_any(vp);
4668 if (seqc_in_modify(vp_seqc)) {
4669 vfs_op_thread_exit_crit(mp, mpcpu);
4670 return (cache_fpl_partial(fpl));
4674 mp = atomic_load_ptr(&vp->v_mountedhere);
4679 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4681 fpl->tvp_seqc = vp_seqc;
4685 static int __noinline
4686 cache_fplookup_cross_mount(struct cache_fpl *fpl)
4689 struct mount_pcpu *mpcpu;
4694 vp_seqc = fpl->tvp_seqc;
4696 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
4697 mp = atomic_load_ptr(&vp->v_mountedhere);
4698 if (__predict_false(mp == NULL)) {
4702 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
4703 return (cache_fpl_partial(fpl));
4705 if (!vn_seqc_consistent(vp, vp_seqc)) {
4706 vfs_op_thread_exit_crit(mp, mpcpu);
4707 return (cache_fpl_partial(fpl));
4709 if (!cache_fplookup_mp_supported(mp)) {
4710 vfs_op_thread_exit_crit(mp, mpcpu);
4711 return (cache_fpl_partial(fpl));
4713 vp = atomic_load_ptr(&mp->mnt_rootvnode);
4714 if (__predict_false(vp == NULL || VN_IS_DOOMED(vp))) {
4715 vfs_op_thread_exit_crit(mp, mpcpu);
4716 return (cache_fpl_partial(fpl));
4718 vp_seqc = vn_seqc_read_any(vp);
4719 vfs_op_thread_exit_crit(mp, mpcpu);
4720 if (seqc_in_modify(vp_seqc)) {
4721 return (cache_fpl_partial(fpl));
4723 mp = atomic_load_ptr(&vp->v_mountedhere);
4724 if (__predict_false(mp != NULL)) {
4726 * There are possibly more mount points on top.
4727 * Normally this does not happen so for simplicity just start
4730 return (cache_fplookup_climb_mount(fpl));
4734 fpl->tvp_seqc = vp_seqc;
4739 * Check if a vnode is mounted on.
4742 cache_fplookup_is_mp(struct cache_fpl *fpl)
4750 * Hack: while this is a union, the pointer tends to be NULL so save on
4753 mp = atomic_load_ptr(&vp->v_mountedhere);
4756 if (vp->v_type == VDIR)
4764 * The code was originally copy-pasted from regular lookup and despite
4765 * clean ups leaves performance on the table. Any modifications here
4766 * must take into account that in case off fallback the resulting
4767 * nameidata state has to be compatible with the original.
4770 cache_fplookup_preparse(struct cache_fpl *fpl)
4772 struct nameidata *ndp;
4773 struct componentname *cnp;
4778 if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
4779 return (cache_fplookup_degenerate(fpl));
4783 * By this point the shortest possible pathname is one character + nul
4784 * terminator, hence 2.
4786 KASSERT(ndp->ni_pathlen >= 2, ("%s: ni_pathlen %zu\n", __func__,
4789 if (__predict_false(cnp->cn_nameptr[ndp->ni_pathlen - 2] == '/')) {
4792 * Regular lookup performs the following:
4793 * *ndp->ni_next = '\0';
4794 * cnp->cn_flags |= TRAILINGSLASH;
4796 * Which is problematic since it modifies data read
4797 * from userspace. Then if fast path lookup was to
4798 * abort we would have to either restore it or convey
4799 * the flag. Since this is a corner case just ignore
4800 * it for simplicity.
4802 return (cache_fpl_aborted(fpl));
4808 cache_fplookup_parse(struct cache_fpl *fpl)
4810 struct nameidata *ndp;
4811 struct componentname *cnp;
4818 * Find the end of this path component, it is either / or nul.
4820 * Store / as a temporary sentinel so that we only have one character
4821 * to test for. Pathnames tend to be short so this should not be
4822 * resulting in cache misses.
4824 KASSERT(cnp->cn_nameptr[ndp->ni_pathlen - 1] == '\0',
4825 ("%s: expected nul at %p + %zu; string [%s]\n", __func__,
4826 cnp->cn_nameptr, ndp->ni_pathlen - 1, cnp->cn_nameptr));
4827 cnp->cn_nameptr[ndp->ni_pathlen - 1] = '/';
4828 for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
4829 KASSERT(*cp != '\0',
4830 ("%s: encountered unexpected nul; string [%s]\n", __func__,
4834 cnp->cn_nameptr[ndp->ni_pathlen - 1] = '\0';
4836 cnp->cn_namelen = cp - cnp->cn_nameptr;
4837 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4838 cache_fpl_smr_exit(fpl);
4839 return (cache_fpl_handled(fpl, ENAMETOOLONG));
4841 ndp->ni_pathlen -= cnp->cn_namelen;
4842 KASSERT(ndp->ni_pathlen <= PATH_MAX,
4843 ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
4848 * Code below is only here to assure compatibility with regular lookup.
4849 * It covers handling of trailing slashles and names like "/", both of
4850 * which of can be taken care of upfront which lockless lookup does
4851 * in cache_fplookup_preparse. Regular lookup performs these for each
4854 while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
4857 panic("%s: ran into TRAILINGSLASH handling from [%s]\n",
4858 __func__, cnp->cn_pnbuf);
4862 if (cnp->cn_nameptr[0] == '\0') {
4863 panic("%s: ran into degenerate name from [%s]\n", __func__, cnp->cn_pnbuf);
4870 cache_fplookup_parse_advance(struct cache_fpl *fpl)
4872 struct nameidata *ndp;
4873 struct componentname *cnp;
4878 cnp->cn_nameptr = ndp->ni_next;
4879 while (*cnp->cn_nameptr == '/') {
4886 * See the API contract for VOP_FPLOOKUP_VEXEC.
4888 static int __noinline
4889 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
4895 dvp_seqc = fpl->dvp_seqc;
4898 * Hack: they may be looking up foo/bar, where foo is a
4899 * regular file. In such a case we need to turn ENOTDIR,
4900 * but we may happen to get here with a different error.
4902 if (dvp->v_type != VDIR) {
4904 * The check here is predominantly to catch
4905 * EOPNOTSUPP from dead_vnodeops. If the vnode
4906 * gets doomed past this point it is going to
4907 * fail seqc verification.
4909 if (VN_IS_DOOMED(dvp)) {
4910 return (cache_fpl_aborted(fpl));
4916 * Hack: handle O_SEARCH.
4918 * Open Group Base Specifications Issue 7, 2018 edition states:
4919 * If the access mode of the open file description associated with the
4920 * file descriptor is not O_SEARCH, the function shall check whether
4921 * directory searches are permitted using the current permissions of
4922 * the directory underlying the file descriptor. If the access mode is
4923 * O_SEARCH, the function shall not perform the check.
4925 * Regular lookup tests for the NOEXECCHECK flag for every path
4926 * component to decide whether to do the permission check. However,
4927 * since most lookups never have the flag (and when they do it is only
4928 * present for the first path component), lockless lookup only acts on
4929 * it if there is a permission problem. Here the flag is represented
4930 * with a boolean so that we don't have to clear it on the way out.
4932 * For simplicity this always aborts.
4933 * TODO: check if this is the first lookup and ignore the permission
4934 * problem. Note the flag has to survive fallback (if it happens to be
4938 return (cache_fpl_aborted(fpl));
4943 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4944 error = cache_fpl_aborted(fpl);
4946 cache_fpl_partial(fpl);
4950 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4951 error = cache_fpl_aborted(fpl);
4953 cache_fpl_smr_exit(fpl);
4954 cache_fpl_handled(fpl, error);
4962 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
4964 struct nameidata *ndp;
4965 struct componentname *cnp;
4969 error = CACHE_FPL_FAILED;
4973 cache_fpl_checkpoint(fpl, &fpl->snd);
4976 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
4977 if (seqc_in_modify(fpl->dvp_seqc)) {
4978 cache_fpl_aborted(fpl);
4981 mp = atomic_load_ptr(&dvp->v_mount);
4982 if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) {
4983 cache_fpl_aborted(fpl);
4987 VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
4989 error = cache_fplookup_preparse(fpl);
4990 if (__predict_false(cache_fpl_terminated(fpl))) {
4995 error = cache_fplookup_parse(fpl);
4996 if (__predict_false(error != 0)) {
5000 VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
5002 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
5003 if (__predict_false(error != 0)) {
5004 error = cache_fplookup_failed_vexec(fpl, error);
5008 error = cache_fplookup_next(fpl);
5009 if (__predict_false(cache_fpl_terminated(fpl))) {
5013 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
5015 if (cache_fpl_islastcn(ndp)) {
5016 error = cache_fplookup_final(fpl);
5020 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
5021 error = cache_fpl_aborted(fpl);
5025 fpl->dvp = fpl->tvp;
5026 fpl->dvp_seqc = fpl->tvp_seqc;
5028 cache_fplookup_parse_advance(fpl);
5029 cache_fpl_checkpoint(fpl, &fpl->snd);
5032 switch (fpl->status) {
5033 case CACHE_FPL_STATUS_UNSET:
5034 __assert_unreachable();
5036 case CACHE_FPL_STATUS_PARTIAL:
5037 cache_fpl_smr_assert_entered(fpl);
5038 return (cache_fplookup_partial_setup(fpl));
5039 case CACHE_FPL_STATUS_ABORTED:
5041 cache_fpl_smr_exit(fpl);
5042 return (CACHE_FPL_FAILED);
5043 case CACHE_FPL_STATUS_HANDLED:
5044 MPASS(error != CACHE_FPL_FAILED);
5045 cache_fpl_smr_assert_not_entered(fpl);
5047 * A common error is ENOENT.
5052 cache_fpl_cleanup_cnp(cnp);
5055 ndp->ni_dvp = fpl->dvp;
5056 ndp->ni_vp = fpl->tvp;
5057 if (cnp->cn_flags & SAVENAME)
5058 cnp->cn_flags |= HASBUF;
5060 cache_fpl_cleanup_cnp(cnp);
5063 __assert_unreachable();
5067 * Fast path lookup protected with SMR and sequence counters.
5069 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
5071 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
5074 * Traditional vnode lookup conceptually looks like this:
5080 * vn_unlock(current);
5087 * Each jump to the next vnode is safe memory-wise and atomic with respect to
5088 * any modifications thanks to holding respective locks.
5090 * The same guarantee can be provided with a combination of safe memory
5091 * reclamation and sequence counters instead. If all operations which affect
5092 * the relationship between the current vnode and the one we are looking for
5093 * also modify the counter, we can verify whether all the conditions held as
5094 * we made the jump. This includes things like permissions, mount points etc.
5095 * Counter modification is provided by enclosing relevant places in
5096 * vn_seqc_write_begin()/end() calls.
5098 * Thus this translates to:
5101 * dvp_seqc = seqc_read_any(dvp);
5102 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
5106 * tvp_seqc = seqc_read_any(tvp);
5107 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
5109 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
5111 * dvp = tvp; // we know nothing of importance has changed
5112 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
5116 * vget(); // secure the vnode
5117 * if (!seqc_consistent(tvp, tvp_seqc) // final check
5119 * // at this point we know nothing has changed for any parent<->child pair
5120 * // as they were crossed during the lookup, meaning we matched the guarantee
5121 * // of the locked variant
5124 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
5125 * - they are called while within vfs_smr protection which they must never exit
5126 * - EAGAIN can be returned to denote checking could not be performed, it is
5127 * always valid to return it
5128 * - if the sequence counter has not changed the result must be valid
5129 * - if the sequence counter has changed both false positives and false negatives
5130 * are permitted (since the result will be rejected later)
5131 * - for simple cases of unix permission checks vaccess_vexec_smr can be used
5133 * Caveats to watch out for:
5134 * - vnodes are passed unlocked and unreferenced with nothing stopping
5135 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
5136 * to use atomic_load_ptr to fetch it.
5137 * - the aforementioned object can also get freed, meaning absent other means it
5138 * should be protected with vfs_smr
5139 * - either safely checking permissions as they are modified or guaranteeing
5140 * their stability is left to the routine
5143 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
5146 struct cache_fpl fpl;
5149 struct componentname *cnp;
5150 struct nameidata_saved orig;
5153 MPASS(ndp->ni_lcf == 0);
5155 fpl.status = CACHE_FPL_STATUS_UNSET;
5157 fpl.cnp = &ndp->ni_cnd;
5158 MPASS(curthread == fpl.cnp->cn_thread);
5159 KASSERT ((fpl.cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
5160 ("%s: internal flags found in cn_flags %" PRIx64, __func__,
5161 fpl.cnp->cn_flags));
5163 if ((fpl.cnp->cn_flags & SAVESTART) != 0)
5164 MPASS(fpl.cnp->cn_nameiop != LOOKUP);
5166 if (!cache_can_fplookup(&fpl)) {
5167 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5168 *status = fpl.status;
5169 return (EOPNOTSUPP);
5172 cache_fpl_checkpoint(&fpl, &orig);
5174 cache_fpl_smr_enter_initial(&fpl);
5175 fpl.fsearch = false;
5176 pwd = pwd_get_smr();
5178 ndp->ni_rootdir = pwd->pwd_rdir;
5179 ndp->ni_topdir = pwd->pwd_jdir;
5182 cnp->cn_nameptr = cnp->cn_pnbuf;
5183 if (cnp->cn_pnbuf[0] == '/') {
5184 dvp = cache_fpl_handle_root(&fpl);
5185 ndp->ni_resflags |= NIRES_ABS;
5187 if (ndp->ni_dirfd == AT_FDCWD) {
5188 dvp = pwd->pwd_cdir;
5190 error = cache_fplookup_dirfd(&fpl, &dvp);
5191 if (__predict_false(error != 0)) {
5197 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
5199 error = cache_fplookup_impl(dvp, &fpl);
5201 cache_fpl_smr_assert_not_entered(&fpl);
5202 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5204 *status = fpl.status;
5205 switch (fpl.status) {
5206 case CACHE_FPL_STATUS_UNSET:
5207 __assert_unreachable();
5209 case CACHE_FPL_STATUS_HANDLED:
5211 MPASS(ndp->ni_vp == NULL);
5212 SDT_PROBE3(vfs, namei, lookup, return, error, ndp->ni_vp, true);
5214 case CACHE_FPL_STATUS_PARTIAL:
5217 * Status restored by cache_fplookup_partial_setup.
5220 case CACHE_FPL_STATUS_ABORTED:
5221 cache_fpl_restore_abort(&fpl, &orig);