2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1989, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * Poul-Henning Kamp of the FreeBSD Project.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
41 #include "opt_ktrace.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/capsicum.h>
46 #include <sys/counter.h>
47 #include <sys/filedesc.h>
48 #include <sys/fnv_hash.h>
49 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/fcntl.h>
55 #include <sys/mount.h>
56 #include <sys/namei.h>
62 #include <sys/syscallsubr.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/vnode.h>
68 #include <sys/ktrace.h>
71 #include <sys/capsicum.h>
73 #include <security/audit/audit.h>
74 #include <security/mac/mac_framework.h>
82 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
85 SDT_PROVIDER_DECLARE(vfs);
86 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
88 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *",
90 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
92 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
94 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
95 "struct namecache *", "int", "int");
96 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
97 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
98 "char *", "struct vnode *");
99 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
100 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
101 "struct vnode *", "char *");
102 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
104 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
105 "struct vnode *", "char *");
106 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
108 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
109 "struct componentname *");
110 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
111 "struct componentname *");
112 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
113 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int");
114 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
115 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
116 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
118 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
120 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *",
123 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
124 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
125 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
128 * This structure describes the elements in the cache of recent
129 * names looked up by namei.
135 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
136 "the state must fit in a union with a pointer without growing it");
139 LIST_ENTRY(namecache) nc_src; /* source vnode list */
140 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */
141 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
142 struct vnode *nc_dvp; /* vnode of parent of name */
144 struct vnode *nu_vp; /* vnode the name refers to */
145 struct negstate nu_neg;/* negative entry state */
147 u_char nc_flag; /* flag bits */
148 u_char nc_nlen; /* length of name */
149 char nc_name[0]; /* segment name + nul */
153 * struct namecache_ts repeats struct namecache layout up to the
155 * struct namecache_ts is used in place of struct namecache when time(s) need
156 * to be stored. The nc_dotdottime field is used when a cache entry is mapping
157 * both a non-dotdot directory name plus dotdot for the directory's
160 * See below for alignment requirement.
162 struct namecache_ts {
163 struct timespec nc_time; /* timespec provided by fs */
164 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */
165 int nc_ticks; /* ticks value when entry was added */
167 struct namecache nc_nc;
170 TAILQ_HEAD(cache_freebatch, namecache);
173 * At least mips n32 performs 64-bit accesses to timespec as found
174 * in namecache_ts and requires them to be aligned. Since others
175 * may be in the same spot suffer a little bit and enforce the
176 * alignment for everyone. Note this is a nop for 64-bit platforms.
178 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t)
181 * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
182 * 4.4 BSD codebase. Later on struct namecache was tweaked to become
183 * smaller and the value was bumped to retain the total size, but it
184 * was never re-evaluated for suitability. A simple test counting
185 * lengths during package building shows that the value of 45 covers
186 * about 86% of all added entries, reaching 99% at 65.
188 * Regardless of the above, use of dedicated zones instead of malloc may be
189 * inducing additional waste. This may be hard to address as said zones are
190 * tied to VFS SMR. Even if retaining them, the current split should be
194 #define CACHE_PATH_CUTOFF 45
195 #define CACHE_LARGE_PAD 6
197 #define CACHE_PATH_CUTOFF 41
198 #define CACHE_LARGE_PAD 2
201 #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
202 #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
203 #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
204 #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
206 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
207 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
208 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
209 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
211 #define nc_vp n_un.nu_vp
212 #define nc_neg n_un.nu_neg
215 * Flags in namecache.nc_flag
217 #define NCF_WHITE 0x01
218 #define NCF_ISDOTDOT 0x02
221 #define NCF_DVDROP 0x10
222 #define NCF_NEGATIVE 0x20
223 #define NCF_INVALID 0x40
227 * Flags in negstate.neg_flag
231 static bool cache_neg_evict_cond(u_long lnumcache);
234 * Mark an entry as invalid.
236 * This is called before it starts getting deconstructed.
239 cache_ncp_invalidate(struct namecache *ncp)
242 KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
243 ("%s: entry %p already invalid", __func__, ncp));
244 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
245 atomic_thread_fence_rel();
249 * Check whether the entry can be safely used.
251 * All places which elide locks are supposed to call this after they are
252 * done with reading from an entry.
254 #define cache_ncp_canuse(ncp) ({ \
255 struct namecache *_ncp = (ncp); \
258 atomic_thread_fence_acq(); \
259 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
260 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \
264 * Name caching works as follows:
266 * Names found by directory scans are retained in a cache
267 * for future reference. It is managed LRU, so frequently
268 * used names will hang around. Cache is indexed by hash value
269 * obtained from (dvp, name) where dvp refers to the directory
272 * If it is a "negative" entry, (i.e. for a name that is known NOT to
273 * exist) the vnode pointer will be NULL.
275 * Upon reaching the last segment of a path, if the reference
276 * is for DELETE, or NOCACHE is set (rewrite), and the
277 * name is located in the cache, it will be dropped.
279 * These locks are used (in the order in which they can be taken):
281 * vnodelock mtx vnode lists and v_cache_dd field protection
282 * bucketlock mtx for access to given set of hash buckets
283 * neglist mtx negative entry LRU management
285 * It is legal to take multiple vnodelock and bucketlock locks. The locking
286 * order is lower address first. Both are recursive.
288 * "." lookups are lockless.
290 * ".." and vnode -> name lookups require vnodelock.
292 * name -> vnode lookup requires the relevant bucketlock to be held for reading.
294 * Insertions and removals of entries require involved vnodes and bucketlocks
295 * to be locked to provide safe operation against other threads modifying the
298 * Some lookups result in removal of the found entry (e.g. getting rid of a
299 * negative entry with the intent to create a positive one), which poses a
300 * problem when multiple threads reach the state. Similarly, two different
301 * threads can purge two different vnodes and try to remove the same name.
303 * If the already held vnode lock is lower than the second required lock, we
304 * can just take the other lock. However, in the opposite case, this could
305 * deadlock. As such, this is resolved by trylocking and if that fails unlocking
306 * the first node, locking everything in order and revalidating the state.
311 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
312 "Name cache parameters");
314 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */
315 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0,
316 "Total namecache capacity");
318 u_int ncsizefactor = 2;
319 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
320 "Size factor for namecache");
322 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */
323 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
324 "Ratio of negative namecache entries");
327 * Negative entry % of namecache capacity above which automatic eviction is allowed.
329 * Check cache_neg_evict_cond for details.
331 static u_int ncnegminpct = 3;
333 static u_int __read_mostly neg_min; /* the above recomputed against ncsize */
334 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
335 "Negative entry count above which automatic eviction is allowed");
338 * Structures associated with name caching.
340 #define NCHHASH(hash) \
341 (&nchashtbl[(hash) & nchash])
342 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
343 static u_long __read_mostly nchash; /* size of hash table */
344 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
345 "Size of namecache hash table");
346 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */
347 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */
349 struct nchstats nchstats; /* cache effectiveness statistics */
351 static bool __read_frequently cache_fast_revlookup = true;
352 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
353 &cache_fast_revlookup, 0, "");
355 static u_int __exclusive_cache_line neg_cycle;
358 #define numneglists (ncneghash + 1)
361 struct mtx nl_evict_lock;
362 struct mtx nl_lock __aligned(CACHE_LINE_SIZE);
363 TAILQ_HEAD(, namecache) nl_list;
364 TAILQ_HEAD(, namecache) nl_hotlist;
366 } __aligned(CACHE_LINE_SIZE);
368 static struct neglist neglists[numneglists];
370 static inline struct neglist *
371 NCP2NEGLIST(struct namecache *ncp)
374 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
377 static inline struct negstate *
378 NCP2NEGSTATE(struct namecache *ncp)
381 MPASS(ncp->nc_flag & NCF_NEGATIVE);
382 return (&ncp->nc_neg);
385 #define numbucketlocks (ncbuckethash + 1)
386 static u_int __read_mostly ncbuckethash;
387 static struct mtx_padalign __read_mostly *bucketlocks;
388 #define HASH2BUCKETLOCK(hash) \
389 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
391 #define numvnodelocks (ncvnodehash + 1)
392 static u_int __read_mostly ncvnodehash;
393 static struct mtx __read_mostly *vnodelocks;
394 static inline struct mtx *
395 VP2VNODELOCK(struct vnode *vp)
398 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
402 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
404 struct namecache_ts *ncp_ts;
406 KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
407 (tsp == NULL && ticksp == NULL),
413 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
414 *tsp = ncp_ts->nc_time;
415 *ticksp = ncp_ts->nc_ticks;
419 static int __read_mostly doingcache = 1; /* 1 => enable the cache */
420 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
421 "VFS namecache enabled");
424 /* Export size information to userland */
425 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
426 sizeof(struct namecache), "sizeof(struct namecache)");
429 * The new name cache statistics
431 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
432 "Name cache statistics");
434 #define STATNODE_ULONG(name, varname, descr) \
435 SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
436 #define STATNODE_COUNTER(name, varname, descr) \
437 static COUNTER_U64_DEFINE_EARLY(varname); \
438 SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
440 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
441 STATNODE_ULONG(count, numcache, "Number of cache entries");
442 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
443 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
444 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits");
445 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits");
446 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
447 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
448 STATNODE_COUNTER(posszaps, numposzaps,
449 "Number of cache hits (positive) we do not want to cache");
450 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
451 STATNODE_COUNTER(negzaps, numnegzaps,
452 "Number of cache hits (negative) we do not want to cache");
453 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
454 /* These count for vn_getcwd(), too. */
455 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
456 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
457 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
458 "Number of fullpath search errors (VOP_VPTOCNP failures)");
459 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
460 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
463 * Debug or developer statistics.
465 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
466 "Name cache debugging");
467 #define DEBUGNODE_ULONG(name, varname, descr) \
468 SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
469 #define DEBUGNODE_COUNTER(name, varname, descr) \
470 static COUNTER_U64_DEFINE_EARLY(varname); \
471 SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \
473 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success,
474 "Number of successful removals after relocking");
475 static long zap_bucket_fail;
476 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
477 static long zap_bucket_fail2;
478 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
479 static long cache_lock_vnodes_cel_3_failures;
480 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
481 "Number of times 3-way vnode locking failed");
483 static void cache_zap_locked(struct namecache *ncp);
484 static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf,
485 char **freebuf, size_t *buflen);
486 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
487 char **retbuf, size_t *buflen, size_t addend);
488 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
489 char **retbuf, size_t *buflen);
490 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
491 char **retbuf, size_t *len, size_t addend);
493 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
496 cache_assert_vlp_locked(struct mtx *vlp)
500 mtx_assert(vlp, MA_OWNED);
504 cache_assert_vnode_locked(struct vnode *vp)
508 vlp = VP2VNODELOCK(vp);
509 cache_assert_vlp_locked(vlp);
513 * Directory vnodes with entries are held for two reasons:
514 * 1. make them less of a target for reclamation in vnlru
515 * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
517 * Note this preferably would not be done and it's a hold over from. It will be
518 * feasible to eliminate altogether if all filesystems start supporting
522 cache_hold_vnode(struct vnode *vp)
525 cache_assert_vnode_locked(vp);
526 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
528 counter_u64_add(numcachehv, 1);
532 cache_drop_vnode(struct vnode *vp)
536 * Called after all locks are dropped, meaning we can't assert
537 * on the state of v_cache_src.
540 counter_u64_add(numcachehv, -1);
546 static uma_zone_t __read_mostly cache_zone_small;
547 static uma_zone_t __read_mostly cache_zone_small_ts;
548 static uma_zone_t __read_mostly cache_zone_large;
549 static uma_zone_t __read_mostly cache_zone_large_ts;
551 static struct namecache *
552 cache_alloc_uma(int len, bool ts)
554 struct namecache_ts *ncp_ts;
555 struct namecache *ncp;
557 if (__predict_false(ts)) {
558 if (len <= CACHE_PATH_CUTOFF)
559 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
561 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
562 ncp = &ncp_ts->nc_nc;
564 if (len <= CACHE_PATH_CUTOFF)
565 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
567 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
573 cache_free_uma(struct namecache *ncp)
575 struct namecache_ts *ncp_ts;
577 if (__predict_false(ncp->nc_flag & NCF_TS)) {
578 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
579 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
580 uma_zfree_smr(cache_zone_small_ts, ncp_ts);
582 uma_zfree_smr(cache_zone_large_ts, ncp_ts);
584 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
585 uma_zfree_smr(cache_zone_small, ncp);
587 uma_zfree_smr(cache_zone_large, ncp);
591 static struct namecache *
592 cache_alloc(int len, bool ts)
597 * Avoid blowout in namecache entries.
600 * 1. filesystems may end up trying to add an already existing entry
601 * (for example this can happen after a cache miss during concurrent
602 * lookup), in which case we will call cache_neg_evict despite not
604 * 2. the routine may fail to free anything and no provisions are made
605 * to make it try harder (see the inside for failure modes)
606 * 3. it only ever looks at negative entries.
608 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
609 if (cache_neg_evict_cond(lnumcache)) {
610 lnumcache = atomic_load_long(&numcache);
612 if (__predict_false(lnumcache >= ncsize)) {
613 atomic_subtract_long(&numcache, 1);
614 counter_u64_add(numdrops, 1);
617 return (cache_alloc_uma(len, ts));
621 cache_free(struct namecache *ncp)
625 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
626 cache_drop_vnode(ncp->nc_dvp);
629 atomic_subtract_long(&numcache, 1);
633 cache_free_batch(struct cache_freebatch *batch)
635 struct namecache *ncp, *nnp;
639 if (TAILQ_EMPTY(batch))
641 TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
642 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
643 cache_drop_vnode(ncp->nc_dvp);
648 atomic_subtract_long(&numcache, i);
650 SDT_PROBE1(vfs, namecache, purge, batch, i);
654 * TODO: With the value stored we can do better than computing the hash based
655 * on the address. The choice of FNV should also be revisited.
658 cache_prehash(struct vnode *vp)
661 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
665 cache_get_hash(char *name, u_char len, struct vnode *dvp)
668 return (fnv_32_buf(name, len, dvp->v_nchash));
671 static inline struct nchashhead *
672 NCP2BUCKET(struct namecache *ncp)
676 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
677 return (NCHHASH(hash));
680 static inline struct mtx *
681 NCP2BUCKETLOCK(struct namecache *ncp)
685 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
686 return (HASH2BUCKETLOCK(hash));
691 cache_assert_bucket_locked(struct namecache *ncp)
695 blp = NCP2BUCKETLOCK(ncp);
696 mtx_assert(blp, MA_OWNED);
700 cache_assert_bucket_unlocked(struct namecache *ncp)
704 blp = NCP2BUCKETLOCK(ncp);
705 mtx_assert(blp, MA_NOTOWNED);
708 #define cache_assert_bucket_locked(x) do { } while (0)
709 #define cache_assert_bucket_unlocked(x) do { } while (0)
712 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y))
714 _cache_sort_vnodes(void **p1, void **p2)
718 MPASS(*p1 != NULL || *p2 != NULL);
728 cache_lock_all_buckets(void)
732 for (i = 0; i < numbucketlocks; i++)
733 mtx_lock(&bucketlocks[i]);
737 cache_unlock_all_buckets(void)
741 for (i = 0; i < numbucketlocks; i++)
742 mtx_unlock(&bucketlocks[i]);
746 cache_lock_all_vnodes(void)
750 for (i = 0; i < numvnodelocks; i++)
751 mtx_lock(&vnodelocks[i]);
755 cache_unlock_all_vnodes(void)
759 for (i = 0; i < numvnodelocks; i++)
760 mtx_unlock(&vnodelocks[i]);
764 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
767 cache_sort_vnodes(&vlp1, &vlp2);
770 if (!mtx_trylock(vlp1))
773 if (!mtx_trylock(vlp2)) {
783 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
786 MPASS(vlp1 != NULL || vlp2 != NULL);
796 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
799 MPASS(vlp1 != NULL || vlp2 != NULL);
808 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
810 struct nchstats snap;
812 if (req->oldptr == NULL)
813 return (SYSCTL_OUT(req, 0, sizeof(snap)));
816 snap.ncs_goodhits = counter_u64_fetch(numposhits);
817 snap.ncs_neghits = counter_u64_fetch(numneghits);
818 snap.ncs_badhits = counter_u64_fetch(numposzaps) +
819 counter_u64_fetch(numnegzaps);
820 snap.ncs_miss = counter_u64_fetch(nummisszap) +
821 counter_u64_fetch(nummiss);
823 return (SYSCTL_OUT(req, &snap, sizeof(snap)));
825 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
826 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
827 "VFS cache effectiveness statistics");
830 cache_recalc_neg_min(u_int val)
833 neg_min = (ncsize * val) / 100;
837 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
843 error = sysctl_handle_int(oidp, &val, 0, req);
844 if (error != 0 || req->newptr == NULL)
847 if (val == ncnegminpct)
849 if (val < 0 || val > 99)
852 cache_recalc_neg_min(val);
856 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
857 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
858 "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
862 * Grab an atomic snapshot of the name cache hash chain lengths
864 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
865 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
869 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
871 struct nchashhead *ncpp;
872 struct namecache *ncp;
873 int i, error, n_nchash, *cntbuf;
876 n_nchash = nchash + 1; /* nchash is max index, not count */
877 if (req->oldptr == NULL)
878 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
879 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
880 cache_lock_all_buckets();
881 if (n_nchash != nchash + 1) {
882 cache_unlock_all_buckets();
883 free(cntbuf, M_TEMP);
886 /* Scan hash tables counting entries */
887 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
888 CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
890 cache_unlock_all_buckets();
891 for (error = 0, i = 0; i < n_nchash; i++)
892 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
894 free(cntbuf, M_TEMP);
897 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
898 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
899 "nchash chain lengths");
902 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
905 struct nchashhead *ncpp;
906 struct namecache *ncp;
908 int count, maxlength, used, pct;
911 return SYSCTL_OUT(req, 0, 4 * sizeof(int));
913 cache_lock_all_buckets();
914 n_nchash = nchash + 1; /* nchash is max index, not count */
918 /* Scan hash tables for applicable entries */
919 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
921 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
926 if (maxlength < count)
929 n_nchash = nchash + 1;
930 cache_unlock_all_buckets();
931 pct = (used * 100) / (n_nchash / 100);
932 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
935 error = SYSCTL_OUT(req, &used, sizeof(used));
938 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
941 error = SYSCTL_OUT(req, &pct, sizeof(pct));
946 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
947 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
948 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
952 * Negative entries management
954 * Various workloads create plenty of negative entries and barely use them
955 * afterwards. Moreover malicious users can keep performing bogus lookups
956 * adding even more entries. For example "make tinderbox" as of writing this
957 * comment ends up with 2.6M namecache entries in total, 1.2M of which are
960 * As such, a rather aggressive eviction method is needed. The currently
961 * employed method is a placeholder.
963 * Entries are split over numneglists separate lists, each of which is further
964 * split into hot and cold entries. Entries get promoted after getting a hit.
965 * Eviction happens on addition of new entry.
967 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
968 "Name cache negative entry statistics");
970 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
971 "Number of negative cache entries");
973 static COUNTER_U64_DEFINE_EARLY(neg_created);
974 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
975 "Number of created negative entries");
977 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
978 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
979 "Number of evicted negative entries");
981 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
982 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
983 &neg_evict_skipped_empty,
984 "Number of times evicting failed due to lack of entries");
986 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
987 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
988 &neg_evict_skipped_missed,
989 "Number of times evicting failed due to target entry disappearing");
991 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
992 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
993 &neg_evict_skipped_contended,
994 "Number of times evicting failed due to contention");
996 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
997 "Number of cache hits (negative)");
1000 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1005 for (i = 0; i < numneglists; i++)
1006 out += neglists[i].nl_hotnum;
1008 return (SYSCTL_OUT(req, &out, sizeof(out)));
1010 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1011 CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1012 "Number of hot negative entries");
1015 cache_neg_init(struct namecache *ncp)
1017 struct negstate *ns;
1019 ncp->nc_flag |= NCF_NEGATIVE;
1020 ns = NCP2NEGSTATE(ncp);
1023 counter_u64_add(neg_created, 1);
1026 #define CACHE_NEG_PROMOTION_THRESH 2
1029 cache_neg_hit_prep(struct namecache *ncp)
1031 struct negstate *ns;
1034 ns = NCP2NEGSTATE(ncp);
1035 n = atomic_load_char(&ns->neg_hit);
1037 if (n >= CACHE_NEG_PROMOTION_THRESH)
1039 if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1042 return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1046 * Nothing to do here but it is provided for completeness as some
1047 * cache_neg_hit_prep callers may end up returning without even
1048 * trying to promote.
1050 #define cache_neg_hit_abort(ncp) do { } while (0)
1053 cache_neg_hit_finish(struct namecache *ncp)
1056 SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1057 counter_u64_add(numneghits, 1);
1061 * Move a negative entry to the hot list.
1064 cache_neg_promote_locked(struct namecache *ncp)
1067 struct negstate *ns;
1069 ns = NCP2NEGSTATE(ncp);
1070 nl = NCP2NEGLIST(ncp);
1071 mtx_assert(&nl->nl_lock, MA_OWNED);
1072 if ((ns->neg_flag & NEG_HOT) == 0) {
1073 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1074 TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1076 ns->neg_flag |= NEG_HOT;
1081 * Move a hot negative entry to the cold list.
1084 cache_neg_demote_locked(struct namecache *ncp)
1087 struct negstate *ns;
1089 ns = NCP2NEGSTATE(ncp);
1090 nl = NCP2NEGLIST(ncp);
1091 mtx_assert(&nl->nl_lock, MA_OWNED);
1092 MPASS(ns->neg_flag & NEG_HOT);
1093 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1094 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1096 ns->neg_flag &= ~NEG_HOT;
1097 atomic_store_char(&ns->neg_hit, 0);
1101 * Move a negative entry to the hot list if it matches the lookup.
1103 * We have to take locks, but they may be contended and in the worst
1104 * case we may need to go off CPU. We don't want to spin within the
1105 * smr section and we can't block with it. Exiting the section means
1106 * the found entry could have been evicted. We are going to look it
1110 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1111 struct namecache *oncp, uint32_t hash)
1113 struct namecache *ncp;
1117 nl = NCP2NEGLIST(oncp);
1119 mtx_lock(&nl->nl_lock);
1121 * For hash iteration.
1126 * Avoid all surprises by only succeeding if we got the same entry and
1127 * bailing completely otherwise.
1128 * XXX There are no provisions to keep the vnode around, meaning we may
1129 * end up promoting a negative entry for a *new* vnode and returning
1130 * ENOENT on its account. This is the error we want to return anyway
1131 * and promotion is harmless.
1133 * In particular at this point there can be a new ncp which matches the
1134 * search but hashes to a different neglist.
1136 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1142 * No match to begin with.
1144 if (__predict_false(ncp == NULL)) {
1149 * The newly found entry may be something different...
1151 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1152 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
1157 * ... and not even negative.
1159 nc_flag = atomic_load_char(&ncp->nc_flag);
1160 if ((nc_flag & NCF_NEGATIVE) == 0) {
1164 if (!cache_ncp_canuse(ncp)) {
1168 cache_neg_promote_locked(ncp);
1169 cache_neg_hit_finish(ncp);
1171 mtx_unlock(&nl->nl_lock);
1175 mtx_unlock(&nl->nl_lock);
1180 cache_neg_promote(struct namecache *ncp)
1184 nl = NCP2NEGLIST(ncp);
1185 mtx_lock(&nl->nl_lock);
1186 cache_neg_promote_locked(ncp);
1187 mtx_unlock(&nl->nl_lock);
1191 cache_neg_insert(struct namecache *ncp)
1195 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1196 cache_assert_bucket_locked(ncp);
1197 nl = NCP2NEGLIST(ncp);
1198 mtx_lock(&nl->nl_lock);
1199 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1200 mtx_unlock(&nl->nl_lock);
1201 atomic_add_long(&numneg, 1);
1205 cache_neg_remove(struct namecache *ncp)
1208 struct negstate *ns;
1210 cache_assert_bucket_locked(ncp);
1211 nl = NCP2NEGLIST(ncp);
1212 ns = NCP2NEGSTATE(ncp);
1213 mtx_lock(&nl->nl_lock);
1214 if ((ns->neg_flag & NEG_HOT) != 0) {
1215 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1218 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1220 mtx_unlock(&nl->nl_lock);
1221 atomic_subtract_long(&numneg, 1);
1224 static struct neglist *
1225 cache_neg_evict_select_list(void)
1230 c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1231 nl = &neglists[c % numneglists];
1232 if (!mtx_trylock(&nl->nl_evict_lock)) {
1233 counter_u64_add(neg_evict_skipped_contended, 1);
1239 static struct namecache *
1240 cache_neg_evict_select_entry(struct neglist *nl)
1242 struct namecache *ncp, *lncp;
1243 struct negstate *ns, *lns;
1246 mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1247 mtx_assert(&nl->nl_lock, MA_OWNED);
1248 ncp = TAILQ_FIRST(&nl->nl_list);
1252 lns = NCP2NEGSTATE(lncp);
1253 for (i = 1; i < 4; i++) {
1254 ncp = TAILQ_NEXT(ncp, nc_dst);
1257 ns = NCP2NEGSTATE(ncp);
1258 if (ns->neg_hit < lns->neg_hit) {
1267 cache_neg_evict(void)
1269 struct namecache *ncp, *ncp2;
1278 nl = cache_neg_evict_select_list();
1283 mtx_lock(&nl->nl_lock);
1284 ncp = TAILQ_FIRST(&nl->nl_hotlist);
1286 cache_neg_demote_locked(ncp);
1288 ncp = cache_neg_evict_select_entry(nl);
1290 counter_u64_add(neg_evict_skipped_empty, 1);
1291 mtx_unlock(&nl->nl_lock);
1292 mtx_unlock(&nl->nl_evict_lock);
1295 nlen = ncp->nc_nlen;
1297 hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1298 dvlp = VP2VNODELOCK(dvp);
1299 blp = HASH2BUCKETLOCK(hash);
1300 mtx_unlock(&nl->nl_lock);
1301 mtx_unlock(&nl->nl_evict_lock);
1305 * Note that since all locks were dropped above, the entry may be
1306 * gone or reallocated to be something else.
1308 CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1309 if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1310 ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1314 counter_u64_add(neg_evict_skipped_missed, 1);
1318 MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1319 MPASS(blp == NCP2BUCKETLOCK(ncp));
1320 SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1322 cache_zap_locked(ncp);
1323 counter_u64_add(neg_evicted, 1);
1334 * Maybe evict a negative entry to create more room.
1336 * The ncnegfactor parameter limits what fraction of the total count
1337 * can comprise of negative entries. However, if the cache is just
1338 * warming up this leads to excessive evictions. As such, ncnegminpct
1339 * (recomputed to neg_min) dictates whether the above should be
1342 * Try evicting if the cache is close to full capacity regardless of
1343 * other considerations.
1346 cache_neg_evict_cond(u_long lnumcache)
1350 if (ncsize - 1000 < lnumcache)
1352 lnumneg = atomic_load_long(&numneg);
1353 if (lnumneg < neg_min)
1355 if (lnumneg * ncnegfactor < lnumcache)
1358 return (cache_neg_evict());
1362 * cache_zap_locked():
1364 * Removes a namecache entry from cache, whether it contains an actual
1365 * pointer to a vnode or if it is just a negative cache entry.
1368 cache_zap_locked(struct namecache *ncp)
1370 struct nchashhead *ncpp;
1372 if (!(ncp->nc_flag & NCF_NEGATIVE))
1373 cache_assert_vnode_locked(ncp->nc_vp);
1374 cache_assert_vnode_locked(ncp->nc_dvp);
1375 cache_assert_bucket_locked(ncp);
1377 cache_ncp_invalidate(ncp);
1379 ncpp = NCP2BUCKET(ncp);
1380 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1381 if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1382 SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
1383 ncp->nc_name, ncp->nc_vp);
1384 TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
1385 if (ncp == ncp->nc_vp->v_cache_dd) {
1386 vn_seqc_write_begin_unheld(ncp->nc_vp);
1387 ncp->nc_vp->v_cache_dd = NULL;
1388 vn_seqc_write_end(ncp->nc_vp);
1391 SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
1393 cache_neg_remove(ncp);
1395 if (ncp->nc_flag & NCF_ISDOTDOT) {
1396 if (ncp == ncp->nc_dvp->v_cache_dd) {
1397 vn_seqc_write_begin_unheld(ncp->nc_dvp);
1398 ncp->nc_dvp->v_cache_dd = NULL;
1399 vn_seqc_write_end(ncp->nc_dvp);
1402 LIST_REMOVE(ncp, nc_src);
1403 if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
1404 ncp->nc_flag |= NCF_DVDROP;
1410 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1414 MPASS(ncp->nc_dvp == vp);
1415 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1416 cache_assert_vnode_locked(vp);
1418 blp = NCP2BUCKETLOCK(ncp);
1420 cache_zap_locked(ncp);
1425 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1428 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1431 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1432 cache_assert_vnode_locked(vp);
1434 if (ncp->nc_flag & NCF_NEGATIVE) {
1435 if (*vlpp != NULL) {
1439 cache_zap_negative_locked_vnode_kl(ncp, vp);
1443 pvlp = VP2VNODELOCK(vp);
1444 blp = NCP2BUCKETLOCK(ncp);
1445 vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1446 vlp2 = VP2VNODELOCK(ncp->nc_vp);
1448 if (*vlpp == vlp1 || *vlpp == vlp2) {
1452 if (*vlpp != NULL) {
1456 cache_sort_vnodes(&vlp1, &vlp2);
1461 if (!mtx_trylock(vlp1))
1467 cache_zap_locked(ncp);
1469 if (to_unlock != NULL)
1470 mtx_unlock(to_unlock);
1477 MPASS(*vlpp == NULL);
1483 * If trylocking failed we can get here. We know enough to take all needed locks
1484 * in the right order and re-lookup the entry.
1487 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1488 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1491 struct namecache *rncp;
1493 cache_assert_bucket_unlocked(ncp);
1495 cache_sort_vnodes(&dvlp, &vlp);
1496 cache_lock_vnodes(dvlp, vlp);
1498 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1499 if (rncp == ncp && rncp->nc_dvp == dvp &&
1500 rncp->nc_nlen == cnp->cn_namelen &&
1501 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1505 cache_zap_locked(rncp);
1507 cache_unlock_vnodes(dvlp, vlp);
1508 counter_u64_add(zap_bucket_relock_success, 1);
1513 cache_unlock_vnodes(dvlp, vlp);
1517 static int __noinline
1518 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1519 uint32_t hash, struct mtx *blp)
1521 struct mtx *dvlp, *vlp;
1524 cache_assert_bucket_locked(ncp);
1526 dvlp = VP2VNODELOCK(ncp->nc_dvp);
1528 if (!(ncp->nc_flag & NCF_NEGATIVE))
1529 vlp = VP2VNODELOCK(ncp->nc_vp);
1530 if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1531 cache_zap_locked(ncp);
1533 cache_unlock_vnodes(dvlp, vlp);
1539 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1542 static __noinline int
1543 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1545 struct namecache *ncp;
1547 struct mtx *dvlp, *dvlp2;
1551 if (cnp->cn_namelen == 2 &&
1552 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1553 dvlp = VP2VNODELOCK(dvp);
1557 ncp = dvp->v_cache_dd;
1562 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1565 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1566 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1568 MPASS(dvp->v_cache_dd == NULL);
1574 vn_seqc_write_begin(dvp);
1575 dvp->v_cache_dd = NULL;
1576 vn_seqc_write_end(dvp);
1581 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1585 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1586 blp = HASH2BUCKETLOCK(hash);
1588 if (CK_SLIST_EMPTY(NCHHASH(hash)))
1593 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1594 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1595 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1604 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1605 if (__predict_false(error != 0)) {
1609 counter_u64_add(numposzaps, 1);
1610 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1614 counter_u64_add(nummisszap, 1);
1615 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1619 static int __noinline
1620 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1621 struct timespec *tsp, int *ticksp)
1626 counter_u64_add(dothits, 1);
1627 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1634 * When we lookup "." we still can be asked to lock it
1637 ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1638 if (ltype != VOP_ISLOCKED(*vpp)) {
1639 if (ltype == LK_EXCLUSIVE) {
1640 vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1641 if (VN_IS_DOOMED((*vpp))) {
1642 /* forced unmount */
1648 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1653 static int __noinline
1654 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1655 struct timespec *tsp, int *ticksp)
1657 struct namecache_ts *ncp_ts;
1658 struct namecache *ncp;
1664 MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1666 if ((cnp->cn_flags & MAKEENTRY) == 0) {
1667 cache_remove_cnp(dvp, cnp);
1671 counter_u64_add(dotdothits, 1);
1673 dvlp = VP2VNODELOCK(dvp);
1675 ncp = dvp->v_cache_dd;
1677 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL);
1681 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1682 if (ncp->nc_flag & NCF_NEGATIVE)
1689 goto negative_success;
1690 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1691 cache_out_ts(ncp, tsp, ticksp);
1692 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1693 NCF_DTS && tsp != NULL) {
1694 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1695 *tsp = ncp_ts->nc_dotdottime;
1699 ltype = VOP_ISLOCKED(dvp);
1701 vs = vget_prep(*vpp);
1703 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1704 vn_lock(dvp, ltype | LK_RETRY);
1705 if (VN_IS_DOOMED(dvp)) {
1717 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1718 if (cnp->cn_flags & ISLASTCN) {
1719 counter_u64_add(numnegzaps, 1);
1720 cache_zap_negative_locked_vnode_kl(ncp, dvp);
1727 whiteout = (ncp->nc_flag & NCF_WHITE);
1728 cache_out_ts(ncp, tsp, ticksp);
1729 if (cache_neg_hit_prep(ncp))
1730 cache_neg_promote(ncp);
1732 cache_neg_hit_finish(ncp);
1735 cnp->cn_flags |= ISWHITEOUT;
1740 * Lookup a name in the name cache
1744 * - dvp: Parent directory in which to search.
1745 * - vpp: Return argument. Will contain desired vnode on cache hit.
1746 * - cnp: Parameters of the name search. The most interesting bits of
1747 * the cn_flags field have the following meanings:
1748 * - MAKEENTRY: If clear, free an entry from the cache rather than look
1750 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".."
1751 * - tsp: Return storage for cache timestamp. On a successful (positive
1752 * or negative) lookup, tsp will be filled with any timespec that
1753 * was stored when this cache entry was created. However, it will
1754 * be clear for "." entries.
1755 * - ticks: Return storage for alternate cache timestamp. On a successful
1756 * (positive or negative) lookup, it will contain the ticks value
1757 * that was current when the cache entry was created, unless cnp
1760 * Either both tsp and ticks have to be provided or neither of them.
1764 * - -1: A positive cache hit. vpp will contain the desired vnode.
1765 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due
1766 * to a forced unmount. vpp will not be modified. If the entry
1767 * is a whiteout, then the ISWHITEOUT flag will be set in
1769 * - 0: A cache miss. vpp will not be modified.
1773 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up
1774 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the
1775 * lock is not recursively acquired.
1777 static int __noinline
1778 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1779 struct timespec *tsp, int *ticksp)
1781 struct namecache *ncp;
1788 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1789 MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
1792 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1793 blp = HASH2BUCKETLOCK(hash);
1796 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1797 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1798 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1802 if (__predict_false(ncp == NULL)) {
1804 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1806 counter_u64_add(nummiss, 1);
1810 if (ncp->nc_flag & NCF_NEGATIVE)
1811 goto negative_success;
1813 counter_u64_add(numposhits, 1);
1815 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1816 cache_out_ts(ncp, tsp, ticksp);
1818 vs = vget_prep(*vpp);
1820 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1827 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1828 if (cnp->cn_flags & ISLASTCN) {
1829 counter_u64_add(numnegzaps, 1);
1830 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1831 if (__predict_false(error != 0)) {
1840 whiteout = (ncp->nc_flag & NCF_WHITE);
1841 cache_out_ts(ncp, tsp, ticksp);
1842 if (cache_neg_hit_prep(ncp))
1843 cache_neg_promote(ncp);
1845 cache_neg_hit_finish(ncp);
1848 cnp->cn_flags |= ISWHITEOUT;
1853 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1854 struct timespec *tsp, int *ticksp)
1856 struct namecache *ncp;
1860 bool whiteout, neg_promote;
1863 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
1866 if (__predict_false(!doingcache)) {
1867 cnp->cn_flags &= ~MAKEENTRY;
1872 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1873 if (cnp->cn_namelen == 1)
1874 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1875 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
1876 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
1879 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1881 if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
1882 cache_remove_cnp(dvp, cnp);
1886 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1889 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1890 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1891 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1895 if (__predict_false(ncp == NULL)) {
1897 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1899 counter_u64_add(nummiss, 1);
1903 nc_flag = atomic_load_char(&ncp->nc_flag);
1904 if (nc_flag & NCF_NEGATIVE)
1905 goto negative_success;
1907 counter_u64_add(numposhits, 1);
1909 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1910 cache_out_ts(ncp, tsp, ticksp);
1912 if (!cache_ncp_canuse(ncp)) {
1917 vs = vget_prep_smr(*vpp);
1919 if (__predict_false(vs == VGET_NONE)) {
1923 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1930 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1931 if (cnp->cn_flags & ISLASTCN) {
1937 cache_out_ts(ncp, tsp, ticksp);
1938 whiteout = (ncp->nc_flag & NCF_WHITE);
1939 neg_promote = cache_neg_hit_prep(ncp);
1940 if (!cache_ncp_canuse(ncp)) {
1941 cache_neg_hit_abort(ncp);
1947 if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
1950 cache_neg_hit_finish(ncp);
1954 cnp->cn_flags |= ISWHITEOUT;
1957 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
1960 struct celockstate {
1964 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1965 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1968 cache_celockstate_init(struct celockstate *cel)
1971 bzero(cel, sizeof(*cel));
1975 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1978 struct mtx *vlp1, *vlp2;
1980 MPASS(cel->vlp[0] == NULL);
1981 MPASS(cel->vlp[1] == NULL);
1982 MPASS(cel->vlp[2] == NULL);
1984 MPASS(vp != NULL || dvp != NULL);
1986 vlp1 = VP2VNODELOCK(vp);
1987 vlp2 = VP2VNODELOCK(dvp);
1988 cache_sort_vnodes(&vlp1, &vlp2);
1999 cache_unlock_vnodes_cel(struct celockstate *cel)
2002 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2004 if (cel->vlp[0] != NULL)
2005 mtx_unlock(cel->vlp[0]);
2006 if (cel->vlp[1] != NULL)
2007 mtx_unlock(cel->vlp[1]);
2008 if (cel->vlp[2] != NULL)
2009 mtx_unlock(cel->vlp[2]);
2013 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2018 cache_assert_vlp_locked(cel->vlp[0]);
2019 cache_assert_vlp_locked(cel->vlp[1]);
2020 MPASS(cel->vlp[2] == NULL);
2023 vlp = VP2VNODELOCK(vp);
2026 if (vlp >= cel->vlp[1]) {
2029 if (mtx_trylock(vlp))
2031 cache_lock_vnodes_cel_3_failures++;
2032 cache_unlock_vnodes_cel(cel);
2033 if (vlp < cel->vlp[0]) {
2035 mtx_lock(cel->vlp[0]);
2036 mtx_lock(cel->vlp[1]);
2038 if (cel->vlp[0] != NULL)
2039 mtx_lock(cel->vlp[0]);
2041 mtx_lock(cel->vlp[1]);
2051 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2055 MPASS(cel->blp[0] == NULL);
2056 MPASS(cel->blp[1] == NULL);
2058 cache_sort_vnodes(&blp1, &blp2);
2069 cache_unlock_buckets_cel(struct celockstate *cel)
2072 if (cel->blp[0] != NULL)
2073 mtx_unlock(cel->blp[0]);
2074 mtx_unlock(cel->blp[1]);
2078 * Lock part of the cache affected by the insertion.
2080 * This means vnodelocks for dvp, vp and the relevant bucketlock.
2081 * However, insertion can result in removal of an old entry. In this
2082 * case we have an additional vnode and bucketlock pair to lock.
2084 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2085 * preserving the locking order (smaller address first).
2088 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2091 struct namecache *ncp;
2092 struct mtx *blps[2];
2094 blps[0] = HASH2BUCKETLOCK(hash);
2097 cache_lock_vnodes_cel(cel, dvp, vp);
2098 if (vp == NULL || vp->v_type != VDIR)
2100 ncp = vp->v_cache_dd;
2103 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2105 MPASS(ncp->nc_dvp == vp);
2106 blps[1] = NCP2BUCKETLOCK(ncp);
2107 if (ncp->nc_flag & NCF_NEGATIVE)
2109 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2112 * All vnodes got re-locked. Re-validate the state and if
2113 * nothing changed we are done. Otherwise restart.
2115 if (ncp == vp->v_cache_dd &&
2116 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2117 blps[1] == NCP2BUCKETLOCK(ncp) &&
2118 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2120 cache_unlock_vnodes_cel(cel);
2125 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2129 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2132 struct namecache *ncp;
2133 struct mtx *blps[2];
2135 blps[0] = HASH2BUCKETLOCK(hash);
2138 cache_lock_vnodes_cel(cel, dvp, vp);
2139 ncp = dvp->v_cache_dd;
2142 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2144 MPASS(ncp->nc_dvp == dvp);
2145 blps[1] = NCP2BUCKETLOCK(ncp);
2146 if (ncp->nc_flag & NCF_NEGATIVE)
2148 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2150 if (ncp == dvp->v_cache_dd &&
2151 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2152 blps[1] == NCP2BUCKETLOCK(ncp) &&
2153 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2155 cache_unlock_vnodes_cel(cel);
2160 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2164 cache_enter_unlock(struct celockstate *cel)
2167 cache_unlock_buckets_cel(cel);
2168 cache_unlock_vnodes_cel(cel);
2171 static void __noinline
2172 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2173 struct componentname *cnp)
2175 struct celockstate cel;
2176 struct namecache *ncp;
2180 if (dvp->v_cache_dd == NULL)
2182 len = cnp->cn_namelen;
2183 cache_celockstate_init(&cel);
2184 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2185 cache_enter_lock_dd(&cel, dvp, vp, hash);
2186 vn_seqc_write_begin(dvp);
2187 ncp = dvp->v_cache_dd;
2188 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2189 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2190 cache_zap_locked(ncp);
2194 dvp->v_cache_dd = NULL;
2195 vn_seqc_write_end(dvp);
2196 cache_enter_unlock(&cel);
2202 * Add an entry to the cache.
2205 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2206 struct timespec *tsp, struct timespec *dtsp)
2208 struct celockstate cel;
2209 struct namecache *ncp, *n2, *ndd;
2210 struct namecache_ts *ncp_ts;
2211 struct nchashhead *ncpp;
2216 VNPASS(dvp != vp, dvp);
2217 VNPASS(!VN_IS_DOOMED(dvp), dvp);
2218 VNPASS(dvp->v_type != VNON, dvp);
2220 VNPASS(!VN_IS_DOOMED(vp), vp);
2221 VNPASS(vp->v_type != VNON, vp);
2225 if (__predict_false(!doingcache))
2230 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2231 if (cnp->cn_namelen == 1)
2233 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2234 cache_enter_dotdot_prep(dvp, vp, cnp);
2235 flag = NCF_ISDOTDOT;
2239 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2243 cache_celockstate_init(&cel);
2248 * Calculate the hash key and setup as much of the new
2249 * namecache entry as possible before acquiring the lock.
2251 ncp->nc_flag = flag | NCF_WIP;
2254 cache_neg_init(ncp);
2257 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2258 ncp_ts->nc_time = *tsp;
2259 ncp_ts->nc_ticks = ticks;
2260 ncp_ts->nc_nc.nc_flag |= NCF_TS;
2262 ncp_ts->nc_dotdottime = *dtsp;
2263 ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2266 len = ncp->nc_nlen = cnp->cn_namelen;
2267 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2268 memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2269 ncp->nc_name[len] = '\0';
2270 cache_enter_lock(&cel, dvp, vp, hash);
2273 * See if this vnode or negative entry is already in the cache
2274 * with this name. This can happen with concurrent lookups of
2275 * the same path name.
2277 ncpp = NCHHASH(hash);
2278 CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
2279 if (n2->nc_dvp == dvp &&
2280 n2->nc_nlen == cnp->cn_namelen &&
2281 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
2282 MPASS(cache_ncp_canuse(n2));
2283 if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2285 ("%s: found entry pointing to a different vnode (%p != %p)",
2286 __func__, NULL, vp));
2288 KASSERT(n2->nc_vp == vp,
2289 ("%s: found entry pointing to a different vnode (%p != %p)",
2290 __func__, n2->nc_vp, vp));
2292 * Entries are supposed to be immutable unless in the
2293 * process of getting destroyed. Accommodating for
2294 * changing timestamps is possible but not worth it.
2295 * This should be harmless in terms of correctness, in
2296 * the worst case resulting in an earlier expiration.
2297 * Alternatively, the found entry can be replaced
2300 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2303 KASSERT((n2->nc_flag & NCF_TS) != 0,
2305 n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2306 n2_ts->nc_time = ncp_ts->nc_time;
2307 n2_ts->nc_ticks = ncp_ts->nc_ticks;
2309 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2310 n2_ts->nc_nc.nc_flag |= NCF_DTS;
2314 SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2316 goto out_unlock_free;
2320 if (flag == NCF_ISDOTDOT) {
2322 * See if we are trying to add .. entry, but some other lookup
2323 * has populated v_cache_dd pointer already.
2325 if (dvp->v_cache_dd != NULL)
2326 goto out_unlock_free;
2327 KASSERT(vp == NULL || vp->v_type == VDIR,
2328 ("wrong vnode type %p", vp));
2329 vn_seqc_write_begin(dvp);
2330 dvp->v_cache_dd = ncp;
2331 vn_seqc_write_end(dvp);
2335 if (flag != NCF_ISDOTDOT) {
2337 * For this case, the cache entry maps both the
2338 * directory name in it and the name ".." for the
2339 * directory's parent.
2341 vn_seqc_write_begin(vp);
2342 if ((ndd = vp->v_cache_dd) != NULL) {
2343 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2344 cache_zap_locked(ndd);
2348 vp->v_cache_dd = ncp;
2349 vn_seqc_write_end(vp);
2350 } else if (vp->v_type != VDIR) {
2351 if (vp->v_cache_dd != NULL) {
2352 vn_seqc_write_begin(vp);
2353 vp->v_cache_dd = NULL;
2354 vn_seqc_write_end(vp);
2359 if (flag != NCF_ISDOTDOT) {
2360 if (LIST_EMPTY(&dvp->v_cache_src)) {
2361 cache_hold_vnode(dvp);
2363 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2367 * If the entry is "negative", we place it into the
2368 * "negative" cache queue, otherwise, we place it into the
2369 * destination vnode's cache entries queue.
2372 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2373 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2376 if (cnp->cn_flags & ISWHITEOUT)
2377 ncp->nc_flag |= NCF_WHITE;
2378 cache_neg_insert(ncp);
2379 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2384 * Insert the new namecache entry into the appropriate chain
2385 * within the cache entries table.
2387 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2389 atomic_thread_fence_rel();
2391 * Mark the entry as fully constructed.
2392 * It is immutable past this point until its removal.
2394 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2396 cache_enter_unlock(&cel);
2401 cache_enter_unlock(&cel);
2407 cache_roundup_2(u_int val)
2411 for (res = 1; res <= val; res <<= 1)
2417 static struct nchashhead *
2418 nchinittbl(u_long elements, u_long *hashmask)
2420 struct nchashhead *hashtbl;
2423 hashsize = cache_roundup_2(elements) / 2;
2425 hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2426 for (i = 0; i < hashsize; i++)
2427 CK_SLIST_INIT(&hashtbl[i]);
2428 *hashmask = hashsize - 1;
2433 ncfreetbl(struct nchashhead *hashtbl)
2436 free(hashtbl, M_VFSCACHE);
2440 * Name cache initialization, from vfs_init() when we are booting
2443 nchinit(void *dummy __unused)
2447 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2448 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2449 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2450 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2451 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2452 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2453 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2454 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2456 VFS_SMR_ZONE_SET(cache_zone_small);
2457 VFS_SMR_ZONE_SET(cache_zone_small_ts);
2458 VFS_SMR_ZONE_SET(cache_zone_large);
2459 VFS_SMR_ZONE_SET(cache_zone_large_ts);
2461 ncsize = desiredvnodes * ncsizefactor;
2462 cache_recalc_neg_min(ncnegminpct);
2463 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2464 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2465 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2467 if (ncbuckethash > nchash)
2468 ncbuckethash = nchash;
2469 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2471 for (i = 0; i < numbucketlocks; i++)
2472 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2473 ncvnodehash = ncbuckethash;
2474 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2476 for (i = 0; i < numvnodelocks; i++)
2477 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2479 for (i = 0; i < numneglists; i++) {
2480 mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2481 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2482 TAILQ_INIT(&neglists[i].nl_list);
2483 TAILQ_INIT(&neglists[i].nl_hotlist);
2486 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2489 cache_vnode_init(struct vnode *vp)
2492 LIST_INIT(&vp->v_cache_src);
2493 TAILQ_INIT(&vp->v_cache_dst);
2494 vp->v_cache_dd = NULL;
2499 cache_changesize(u_long newmaxvnodes)
2501 struct nchashhead *new_nchashtbl, *old_nchashtbl;
2502 u_long new_nchash, old_nchash;
2503 struct namecache *ncp;
2508 newncsize = newmaxvnodes * ncsizefactor;
2509 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2510 if (newmaxvnodes < numbucketlocks)
2511 newmaxvnodes = numbucketlocks;
2513 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2514 /* If same hash table size, nothing to do */
2515 if (nchash == new_nchash) {
2516 ncfreetbl(new_nchashtbl);
2520 * Move everything from the old hash table to the new table.
2521 * None of the namecache entries in the table can be removed
2522 * because to do so, they have to be removed from the hash table.
2524 cache_lock_all_vnodes();
2525 cache_lock_all_buckets();
2526 old_nchashtbl = nchashtbl;
2527 old_nchash = nchash;
2528 nchashtbl = new_nchashtbl;
2529 nchash = new_nchash;
2530 for (i = 0; i <= old_nchash; i++) {
2531 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2532 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2534 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2535 CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2539 cache_recalc_neg_min(ncnegminpct);
2540 cache_unlock_all_buckets();
2541 cache_unlock_all_vnodes();
2542 ncfreetbl(old_nchashtbl);
2546 * Invalidate all entries from and to a particular vnode.
2549 cache_purge_impl(struct vnode *vp)
2551 struct cache_freebatch batch;
2552 struct namecache *ncp;
2553 struct mtx *vlp, *vlp2;
2556 vlp = VP2VNODELOCK(vp);
2560 while (!LIST_EMPTY(&vp->v_cache_src)) {
2561 ncp = LIST_FIRST(&vp->v_cache_src);
2562 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2564 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2566 while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2567 ncp = TAILQ_FIRST(&vp->v_cache_dst);
2568 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2570 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2572 ncp = vp->v_cache_dd;
2574 KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2575 ("lost dotdot link"));
2576 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2578 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2580 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2584 cache_free_batch(&batch);
2588 * Opportunistic check to see if there is anything to do.
2591 cache_has_entries(struct vnode *vp)
2594 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2595 vp->v_cache_dd == NULL)
2601 cache_purge(struct vnode *vp)
2604 SDT_PROBE1(vfs, namecache, purge, done, vp);
2605 if (!cache_has_entries(vp))
2607 cache_purge_impl(vp);
2611 * Only to be used by vgone.
2614 cache_purge_vgone(struct vnode *vp)
2618 VNPASS(VN_IS_DOOMED(vp), vp);
2619 if (cache_has_entries(vp)) {
2620 cache_purge_impl(vp);
2625 * Serialize against a potential thread doing cache_purge.
2627 vlp = VP2VNODELOCK(vp);
2628 mtx_wait_unlocked(vlp);
2629 if (cache_has_entries(vp)) {
2630 cache_purge_impl(vp);
2637 * Invalidate all negative entries for a particular directory vnode.
2640 cache_purge_negative(struct vnode *vp)
2642 struct cache_freebatch batch;
2643 struct namecache *ncp, *nnp;
2646 SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2647 if (LIST_EMPTY(&vp->v_cache_src))
2650 vlp = VP2VNODELOCK(vp);
2652 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2653 if (!(ncp->nc_flag & NCF_NEGATIVE))
2655 cache_zap_negative_locked_vnode_kl(ncp, vp);
2656 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2659 cache_free_batch(&batch);
2663 * Entry points for modifying VOP operations.
2666 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2667 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2670 ASSERT_VOP_IN_SEQC(fdvp);
2671 ASSERT_VOP_IN_SEQC(fvp);
2672 ASSERT_VOP_IN_SEQC(tdvp);
2674 ASSERT_VOP_IN_SEQC(tvp);
2679 KASSERT(!cache_remove_cnp(tdvp, tcnp),
2680 ("%s: lingering negative entry", __func__));
2682 cache_remove_cnp(tdvp, tcnp);
2687 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
2690 ASSERT_VOP_IN_SEQC(dvp);
2691 ASSERT_VOP_IN_SEQC(vp);
2697 * Validate that if an entry exists it matches.
2700 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2702 struct namecache *ncp;
2706 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2707 if (CK_SLIST_EMPTY(NCHHASH(hash)))
2709 blp = HASH2BUCKETLOCK(hash);
2711 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2712 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2713 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) {
2714 if (ncp->nc_vp != vp)
2715 panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p vp %p\n",
2716 __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp,
2725 * Flush all entries referencing a particular filesystem.
2728 cache_purgevfs(struct mount *mp)
2730 struct vnode *vp, *mvp;
2732 SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2734 * Somewhat wasteful iteration over all vnodes. Would be better to
2735 * support filtering and avoid the interlock to begin with.
2737 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2738 if (!cache_has_entries(vp)) {
2750 * Perform canonical checks and cache lookup and pass on to filesystem
2751 * through the vop_cachedlookup only if needed.
2755 vfs_cache_lookup(struct vop_lookup_args *ap)
2759 struct vnode **vpp = ap->a_vpp;
2760 struct componentname *cnp = ap->a_cnp;
2761 int flags = cnp->cn_flags;
2766 if (dvp->v_type != VDIR)
2769 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2770 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2773 error = vn_dir_check_exec(dvp, cnp);
2777 error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2779 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2785 /* Implementation of the getcwd syscall. */
2787 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2793 buflen = uap->buflen;
2794 if (__predict_false(buflen < 2))
2796 if (buflen > MAXPATHLEN)
2797 buflen = MAXPATHLEN;
2799 buf = uma_zalloc(namei_zone, M_WAITOK);
2800 error = vn_getcwd(buf, &retbuf, &buflen);
2802 error = copyout(retbuf, uap->buf, buflen);
2803 uma_zfree(namei_zone, buf);
2808 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
2814 pwd = pwd_get_smr();
2815 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
2817 VFS_SMR_ASSERT_NOT_ENTERED();
2819 pwd = pwd_hold(curthread);
2820 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
2826 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2833 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2834 size_t size, int flags, enum uio_seg pathseg)
2836 struct nameidata nd;
2837 char *retbuf, *freebuf;
2842 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2843 pathseg, path, fd, &cap_fstat_rights, td);
2844 if ((error = namei(&nd)) != 0)
2846 error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size);
2848 error = copyout(retbuf, buf, size);
2849 free(freebuf, M_TEMP);
2856 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2859 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2860 uap->flags, UIO_USERSPACE));
2864 * Retrieve the full filesystem path that correspond to a vnode from the name
2865 * cache (if available)
2868 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
2875 if (__predict_false(vp == NULL))
2878 buflen = MAXPATHLEN;
2879 buf = malloc(buflen, M_TEMP, M_WAITOK);
2881 pwd = pwd_get_smr();
2882 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
2883 VFS_SMR_ASSERT_NOT_ENTERED();
2885 pwd = pwd_hold(curthread);
2886 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
2897 * This function is similar to vn_fullpath, but it attempts to lookup the
2898 * pathname relative to the global root mount point. This is required for the
2899 * auditing sub-system, as audited pathnames must be absolute, relative to the
2900 * global root mount point.
2903 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
2909 if (__predict_false(vp == NULL))
2911 buflen = MAXPATHLEN;
2912 buf = malloc(buflen, M_TEMP, M_WAITOK);
2914 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
2915 VFS_SMR_ASSERT_NOT_ENTERED();
2917 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
2926 static struct namecache *
2927 vn_dd_from_dst(struct vnode *vp)
2929 struct namecache *ncp;
2931 cache_assert_vnode_locked(vp);
2932 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
2933 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2940 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
2943 struct namecache *ncp;
2947 vlp = VP2VNODELOCK(*vp);
2949 ncp = (*vp)->v_cache_dd;
2950 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
2951 KASSERT(ncp == vn_dd_from_dst(*vp),
2952 ("%s: mismatch for dd entry (%p != %p)", __func__,
2953 ncp, vn_dd_from_dst(*vp)));
2955 ncp = vn_dd_from_dst(*vp);
2958 if (*buflen < ncp->nc_nlen) {
2961 counter_u64_add(numfullpathfail4, 1);
2963 SDT_PROBE3(vfs, namecache, fullpath, return, error,
2967 *buflen -= ncp->nc_nlen;
2968 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2969 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2978 SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2981 vn_lock(*vp, LK_SHARED | LK_RETRY);
2982 error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
2985 counter_u64_add(numfullpathfail2, 1);
2986 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2991 if (VN_IS_DOOMED(dvp)) {
2992 /* forced unmount */
2995 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2999 * *vp has its use count incremented still.
3006 * Resolve a directory to a pathname.
3008 * The name of the directory can always be found in the namecache or fetched
3009 * from the filesystem. There is also guaranteed to be only one parent, meaning
3010 * we can just follow vnodes up until we find the root.
3012 * The vnode must be referenced.
3015 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3016 size_t *len, size_t addend)
3018 #ifdef KDTRACE_HOOKS
3019 struct vnode *startvp = vp;
3024 bool slash_prefixed;
3026 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3027 VNPASS(vp->v_usecount > 0, vp);
3031 slash_prefixed = true;
3036 slash_prefixed = false;
3041 SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3042 counter_u64_add(numfullpathcalls, 1);
3043 while (vp != rdir && vp != rootvnode) {
3045 * The vp vnode must be already fully constructed,
3046 * since it is either found in namecache or obtained
3047 * from VOP_VPTOCNP(). We may test for VV_ROOT safely
3048 * without obtaining the vnode lock.
3050 if ((vp->v_vflag & VV_ROOT) != 0) {
3051 vn_lock(vp, LK_RETRY | LK_SHARED);
3054 * With the vnode locked, check for races with
3055 * unmount, forced or not. Note that we
3056 * already verified that vp is not equal to
3057 * the root vnode, which means that
3058 * mnt_vnodecovered can be NULL only for the
3061 if (VN_IS_DOOMED(vp) ||
3062 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3063 vp1->v_mountedhere != vp->v_mount) {
3066 SDT_PROBE3(vfs, namecache, fullpath, return,
3076 if (vp->v_type != VDIR) {
3078 counter_u64_add(numfullpathfail1, 1);
3080 SDT_PROBE3(vfs, namecache, fullpath, return,
3084 error = vn_vptocnp(&vp, buf, &buflen);
3090 SDT_PROBE3(vfs, namecache, fullpath, return, error,
3094 buf[--buflen] = '/';
3095 slash_prefixed = true;
3099 if (!slash_prefixed) {
3102 counter_u64_add(numfullpathfail4, 1);
3103 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3107 buf[--buflen] = '/';
3109 counter_u64_add(numfullpathfound, 1);
3112 *retbuf = buf + buflen;
3113 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3120 * Resolve an arbitrary vnode to a pathname.
3123 * - hardlinks are not tracked, thus if the vnode is not a directory this can
3124 * resolve to a different path than the one used to find it
3125 * - namecache is not mandatory, meaning names are not guaranteed to be added
3126 * (in which case resolving fails)
3128 static void __inline
3129 cache_rev_failed_impl(int *reason, int line)
3134 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__)
3137 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3138 char **retbuf, size_t *buflen, size_t addend)
3140 #ifdef KDTRACE_HOOKS
3141 struct vnode *startvp = vp;
3145 struct namecache *ncp;
3149 #ifdef KDTRACE_HOOKS
3152 seqc_t vp_seqc, tvp_seqc;
3155 VFS_SMR_ASSERT_ENTERED();
3157 if (!cache_fast_revlookup) {
3162 orig_buflen = *buflen;
3165 MPASS(*buflen >= 2);
3167 buf[*buflen] = '\0';
3170 if (vp == rdir || vp == rootvnode) {
3178 #ifdef KDTRACE_HOOKS
3182 ncp = NULL; /* for sdt probe down below */
3183 vp_seqc = vn_seqc_read_any(vp);
3184 if (seqc_in_modify(vp_seqc)) {
3185 cache_rev_failed(&reason);
3190 #ifdef KDTRACE_HOOKS
3193 if ((vp->v_vflag & VV_ROOT) != 0) {
3194 mp = atomic_load_ptr(&vp->v_mount);
3196 cache_rev_failed(&reason);
3199 tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3200 tvp_seqc = vn_seqc_read_any(tvp);
3201 if (seqc_in_modify(tvp_seqc)) {
3202 cache_rev_failed(&reason);
3205 if (!vn_seqc_consistent(vp, vp_seqc)) {
3206 cache_rev_failed(&reason);
3213 ncp = atomic_load_ptr(&vp->v_cache_dd);
3215 cache_rev_failed(&reason);
3218 nc_flag = atomic_load_char(&ncp->nc_flag);
3219 if ((nc_flag & NCF_ISDOTDOT) != 0) {
3220 cache_rev_failed(&reason);
3223 if (!cache_ncp_canuse(ncp)) {
3224 cache_rev_failed(&reason);
3227 if (ncp->nc_nlen >= *buflen) {
3228 cache_rev_failed(&reason);
3232 *buflen -= ncp->nc_nlen;
3233 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3237 tvp_seqc = vn_seqc_read_any(tvp);
3238 if (seqc_in_modify(tvp_seqc)) {
3239 cache_rev_failed(&reason);
3242 if (!vn_seqc_consistent(vp, vp_seqc)) {
3243 cache_rev_failed(&reason);
3248 if (vp == rdir || vp == rootvnode)
3253 *retbuf = buf + *buflen;
3254 *buflen = orig_buflen - *buflen + addend;
3255 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3259 *buflen = orig_buflen;
3260 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3266 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3269 size_t orig_buflen, addend;
3275 orig_buflen = *buflen;
3279 if (vp->v_type != VDIR) {
3281 buf[*buflen] = '\0';
3282 error = vn_vptocnp(&vp, buf, buflen);
3291 addend = orig_buflen - *buflen;
3294 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3298 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3300 * Since the namecache does not track hardlinks, the caller is expected to first
3301 * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
3303 * Then we have 2 cases:
3304 * - if the found vnode is a directory, the path can be constructed just by
3305 * following names up the chain
3306 * - otherwise we populate the buffer with the saved name and start resolving
3310 vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf,
3315 struct componentname *cnp;
3323 if (*buflen > MAXPATHLEN)
3324 *buflen = MAXPATHLEN;
3326 buf = malloc(*buflen, M_TEMP, M_WAITOK);
3331 * Check for VBAD to work around the vp_crossmp bug in lookup().
3333 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3334 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3335 * If the type is VDIR (like in this very case) we can skip looking
3336 * at ni_dvp in the first place. However, since vnodes get passed here
3337 * unlocked the target may transition to doomed state (type == VBAD)
3338 * before we get to evaluate the condition. If this happens, we will
3339 * populate part of the buffer and descend to vn_fullpath_dir with
3340 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3342 * This should be atomic_load(&vp->v_type) but it is illegal to take
3343 * an address of a bit field, even if said field is sized to char.
3344 * Work around the problem by reading the value into a full-sized enum
3345 * and then re-reading it with atomic_load which will still prevent
3346 * the compiler from re-reading down the road.
3349 type = atomic_load_int(&type);
3356 addend = cnp->cn_namelen + 2;
3357 if (*buflen < addend) {
3362 tmpbuf = buf + *buflen;
3364 memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
3365 tmpbuf[addend - 1] = '\0';
3370 pwd = pwd_get_smr();
3371 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3373 VFS_SMR_ASSERT_NOT_ENTERED();
3375 pwd = pwd_hold(curthread);
3377 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3393 vn_dir_dd_ino(struct vnode *vp)
3395 struct namecache *ncp;
3400 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3401 vlp = VP2VNODELOCK(vp);
3403 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3404 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3407 vs = vget_prep(ddvp);
3409 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3418 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3420 struct namecache *ncp;
3424 vlp = VP2VNODELOCK(vp);
3426 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3427 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3433 l = min(ncp->nc_nlen, buflen - 1);
3434 memcpy(buf, ncp->nc_name, l);
3441 * This function updates path string to vnode's full global path
3442 * and checks the size of the new path string against the pathlen argument.
3444 * Requires a locked, referenced vnode.
3445 * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3447 * If vp is a directory, the call to vn_fullpath_global() always succeeds
3448 * because it falls back to the ".." lookup if the namecache lookup fails.
3451 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3454 struct nameidata nd;
3459 ASSERT_VOP_ELOCKED(vp, __func__);
3461 /* Construct global filesystem path from vp. */
3463 error = vn_fullpath_global(vp, &rpath, &fbuf);
3470 if (strlen(rpath) >= pathlen) {
3472 error = ENAMETOOLONG;
3477 * Re-lookup the vnode by path to detect a possible rename.
3478 * As a side effect, the vnode is relocked.
3479 * If vnode was renamed, return ENOENT.
3481 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3482 UIO_SYSSPACE, path, td);
3488 NDFREE(&nd, NDF_ONLY_PNBUF);
3492 strcpy(path, rpath);
3505 db_print_vpath(struct vnode *vp)
3508 while (vp != NULL) {
3509 db_printf("%p: ", vp);
3510 if (vp == rootvnode) {
3514 if (vp->v_vflag & VV_ROOT) {
3515 db_printf("<mount point>");
3516 vp = vp->v_mount->mnt_vnodecovered;
3518 struct namecache *ncp;
3522 ncp = TAILQ_FIRST(&vp->v_cache_dst);
3525 for (i = 0; i < ncp->nc_nlen; i++)
3526 db_printf("%c", *ncn++);
3539 DB_SHOW_COMMAND(vpath, db_show_vpath)
3544 db_printf("usage: show vpath <struct vnode *>\n");
3548 vp = (struct vnode *)addr;
3554 static bool __read_frequently cache_fast_lookup = true;
3555 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_lookup, CTLFLAG_RW,
3556 &cache_fast_lookup, 0, "");
3558 #define CACHE_FPL_FAILED -2020
3561 cache_fpl_cleanup_cnp(struct componentname *cnp)
3564 uma_zfree(namei_zone, cnp->cn_pnbuf);
3566 cnp->cn_pnbuf = NULL;
3567 cnp->cn_nameptr = NULL;
3572 cache_fpl_handle_root(struct nameidata *ndp, struct vnode **dpp)
3574 struct componentname *cnp;
3577 while (*(cnp->cn_nameptr) == '/') {
3582 *dpp = ndp->ni_rootdir;
3586 * Components of nameidata (or objects it can point to) which may
3587 * need restoring in case fast path lookup fails.
3589 struct nameidata_saved {
3597 struct nameidata *ndp;
3598 struct componentname *cnp;
3604 struct nameidata_saved snd;
3606 enum cache_fpl_status status:8;
3612 cache_fpl_checkpoint(struct cache_fpl *fpl, struct nameidata_saved *snd)
3615 snd->cn_flags = fpl->ndp->ni_cnd.cn_flags;
3616 snd->cn_namelen = fpl->ndp->ni_cnd.cn_namelen;
3617 snd->cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3618 snd->ni_pathlen = fpl->ndp->ni_pathlen;
3622 cache_fpl_restore_partial(struct cache_fpl *fpl, struct nameidata_saved *snd)
3625 fpl->ndp->ni_cnd.cn_flags = snd->cn_flags;
3626 fpl->ndp->ni_cnd.cn_namelen = snd->cn_namelen;
3627 fpl->ndp->ni_cnd.cn_nameptr = snd->cn_nameptr;
3628 fpl->ndp->ni_pathlen = snd->ni_pathlen;
3632 cache_fpl_restore_abort(struct cache_fpl *fpl, struct nameidata_saved *snd)
3635 cache_fpl_restore_partial(fpl, snd);
3637 * It is 0 on entry by API contract.
3639 fpl->ndp->ni_resflags = 0;
3643 #define cache_fpl_smr_assert_entered(fpl) ({ \
3644 struct cache_fpl *_fpl = (fpl); \
3645 MPASS(_fpl->in_smr == true); \
3646 VFS_SMR_ASSERT_ENTERED(); \
3648 #define cache_fpl_smr_assert_not_entered(fpl) ({ \
3649 struct cache_fpl *_fpl = (fpl); \
3650 MPASS(_fpl->in_smr == false); \
3651 VFS_SMR_ASSERT_NOT_ENTERED(); \
3654 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
3655 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
3658 #define cache_fpl_smr_enter_initial(fpl) ({ \
3659 struct cache_fpl *_fpl = (fpl); \
3661 _fpl->in_smr = true; \
3664 #define cache_fpl_smr_enter(fpl) ({ \
3665 struct cache_fpl *_fpl = (fpl); \
3666 MPASS(_fpl->in_smr == false); \
3668 _fpl->in_smr = true; \
3671 #define cache_fpl_smr_exit(fpl) ({ \
3672 struct cache_fpl *_fpl = (fpl); \
3673 MPASS(_fpl->in_smr == true); \
3675 _fpl->in_smr = false; \
3679 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
3682 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3683 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3684 ("%s: converting to abort from %d at %d, set at %d\n",
3685 __func__, fpl->status, line, fpl->line));
3687 fpl->status = CACHE_FPL_STATUS_ABORTED;
3689 return (CACHE_FPL_FAILED);
3692 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__)
3695 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
3698 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3699 ("%s: setting to partial at %d, but already set to %d at %d\n",
3700 __func__, line, fpl->status, fpl->line));
3701 cache_fpl_smr_assert_entered(fpl);
3702 fpl->status = CACHE_FPL_STATUS_PARTIAL;
3704 return (CACHE_FPL_FAILED);
3707 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__)
3710 cache_fpl_handled_impl(struct cache_fpl *fpl, int error, int line)
3713 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3714 ("%s: setting to handled at %d, but already set to %d at %d\n",
3715 __func__, line, fpl->status, fpl->line));
3716 cache_fpl_smr_assert_not_entered(fpl);
3717 MPASS(error != CACHE_FPL_FAILED);
3718 fpl->status = CACHE_FPL_STATUS_HANDLED;
3723 #define cache_fpl_handled(x, e) cache_fpl_handled_impl((x), (e), __LINE__)
3725 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
3726 (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
3727 FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | ISOPEN | \
3728 NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
3730 #define CACHE_FPL_INTERNAL_CN_FLAGS \
3731 (ISDOTDOT | MAKEENTRY | ISLASTCN)
3733 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
3734 "supported and internal flags overlap");
3737 cache_fpl_islastcn(struct nameidata *ndp)
3740 return (*ndp->ni_next == 0);
3744 cache_fpl_isdotdot(struct componentname *cnp)
3747 if (cnp->cn_namelen == 2 &&
3748 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
3754 cache_can_fplookup(struct cache_fpl *fpl)
3756 struct nameidata *ndp;
3757 struct componentname *cnp;
3762 td = cnp->cn_thread;
3764 if (!cache_fast_lookup) {
3765 cache_fpl_aborted(fpl);
3769 if (mac_vnode_check_lookup_enabled()) {
3770 cache_fpl_aborted(fpl);
3774 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
3775 cache_fpl_aborted(fpl);
3778 if (IN_CAPABILITY_MODE(td)) {
3779 cache_fpl_aborted(fpl);
3782 if (AUDITING_TD(td)) {
3783 cache_fpl_aborted(fpl);
3786 if (ndp->ni_startdir != NULL) {
3787 cache_fpl_aborted(fpl);
3794 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
3796 struct nameidata *ndp;
3801 error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch);
3802 if (__predict_false(error != 0)) {
3803 cache_fpl_smr_exit(fpl);
3804 return (cache_fpl_aborted(fpl));
3806 fpl->fsearch = fsearch;
3811 cache_fplookup_vnode_supported(struct vnode *vp)
3814 return (vp->v_type != VLNK);
3817 static int __noinline
3818 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
3821 struct componentname *cnp;
3827 cache_fpl_smr_exit(fpl);
3828 if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
3829 return (cache_fpl_handled(fpl, ENOENT));
3831 return (cache_fpl_aborted(fpl));
3835 * The target vnode is not supported, prepare for the slow path to take over.
3837 static int __noinline
3838 cache_fplookup_partial_setup(struct cache_fpl *fpl)
3840 struct nameidata *ndp;
3841 struct componentname *cnp;
3851 dvp_seqc = fpl->dvp_seqc;
3853 if (!pwd_hold_smr(pwd)) {
3854 cache_fpl_smr_exit(fpl);
3855 return (cache_fpl_aborted(fpl));
3858 dvs = vget_prep_smr(dvp);
3859 cache_fpl_smr_exit(fpl);
3860 if (__predict_false(dvs == VGET_NONE)) {
3862 return (cache_fpl_aborted(fpl));
3865 vget_finish_ref(dvp, dvs);
3866 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3869 return (cache_fpl_aborted(fpl));
3872 cache_fpl_restore_partial(fpl, &fpl->snd);
3874 ndp->ni_startdir = dvp;
3875 cnp->cn_flags |= MAKEENTRY;
3876 if (cache_fpl_islastcn(ndp))
3877 cnp->cn_flags |= ISLASTCN;
3878 if (cache_fpl_isdotdot(cnp))
3879 cnp->cn_flags |= ISDOTDOT;
3885 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
3887 struct componentname *cnp;
3894 tvp_seqc = fpl->tvp_seqc;
3896 if ((cnp->cn_flags & LOCKLEAF) != 0) {
3897 lkflags = LK_SHARED;
3898 if ((cnp->cn_flags & LOCKSHARED) == 0)
3899 lkflags = LK_EXCLUSIVE;
3900 error = vget_finish(tvp, lkflags, tvs);
3901 if (__predict_false(error != 0)) {
3902 return (cache_fpl_aborted(fpl));
3905 vget_finish_ref(tvp, tvs);
3908 if (!vn_seqc_consistent(tvp, tvp_seqc)) {
3909 if ((cnp->cn_flags & LOCKLEAF) != 0)
3913 return (cache_fpl_aborted(fpl));
3916 return (cache_fpl_handled(fpl, 0));
3920 * They want to possibly modify the state of the namecache.
3922 * Don't try to match the API contract, just leave.
3923 * TODO: this leaves scalability on the table
3926 cache_fplookup_final_modifying(struct cache_fpl *fpl)
3928 struct componentname *cnp;
3931 MPASS(cnp->cn_nameiop != LOOKUP);
3932 return (cache_fpl_partial(fpl));
3935 static int __noinline
3936 cache_fplookup_final_withparent(struct cache_fpl *fpl)
3938 struct componentname *cnp;
3939 enum vgetstate dvs, tvs;
3940 struct vnode *dvp, *tvp;
3946 dvp_seqc = fpl->dvp_seqc;
3949 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
3952 * This is less efficient than it can be for simplicity.
3954 dvs = vget_prep_smr(dvp);
3955 if (__predict_false(dvs == VGET_NONE)) {
3956 return (cache_fpl_aborted(fpl));
3958 tvs = vget_prep_smr(tvp);
3959 if (__predict_false(tvs == VGET_NONE)) {
3960 cache_fpl_smr_exit(fpl);
3961 vget_abort(dvp, dvs);
3962 return (cache_fpl_aborted(fpl));
3965 cache_fpl_smr_exit(fpl);
3967 if ((cnp->cn_flags & LOCKPARENT) != 0) {
3968 error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
3969 if (__predict_false(error != 0)) {
3970 vget_abort(tvp, tvs);
3971 return (cache_fpl_aborted(fpl));
3974 vget_finish_ref(dvp, dvs);
3977 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3978 vget_abort(tvp, tvs);
3979 if ((cnp->cn_flags & LOCKPARENT) != 0)
3983 return (cache_fpl_aborted(fpl));
3986 error = cache_fplookup_final_child(fpl, tvs);
3987 if (__predict_false(error != 0)) {
3988 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED);
3989 if ((cnp->cn_flags & LOCKPARENT) != 0)
3996 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
4001 cache_fplookup_final(struct cache_fpl *fpl)
4003 struct componentname *cnp;
4005 struct vnode *dvp, *tvp;
4010 dvp_seqc = fpl->dvp_seqc;
4013 VNPASS(cache_fplookup_vnode_supported(dvp), dvp);
4015 if (cnp->cn_nameiop != LOOKUP) {
4016 return (cache_fplookup_final_modifying(fpl));
4019 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
4020 return (cache_fplookup_final_withparent(fpl));
4022 tvs = vget_prep_smr(tvp);
4023 if (__predict_false(tvs == VGET_NONE)) {
4024 return (cache_fpl_partial(fpl));
4027 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4028 cache_fpl_smr_exit(fpl);
4029 vget_abort(tvp, tvs);
4030 return (cache_fpl_aborted(fpl));
4033 cache_fpl_smr_exit(fpl);
4034 return (cache_fplookup_final_child(fpl, tvs));
4037 static int __noinline
4038 cache_fplookup_dot(struct cache_fpl *fpl)
4045 fpl->tvp_seqc = vn_seqc_read_any(dvp);
4046 if (seqc_in_modify(fpl->tvp_seqc)) {
4047 return (cache_fpl_aborted(fpl));
4050 counter_u64_add(dothits, 1);
4051 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", dvp);
4056 static int __noinline
4057 cache_fplookup_dotdot(struct cache_fpl *fpl)
4059 struct nameidata *ndp;
4060 struct componentname *cnp;
4061 struct namecache *ncp;
4071 * XXX this is racy the same way regular lookup is
4073 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
4075 if (dvp == pr->pr_root)
4078 if (dvp == ndp->ni_rootdir ||
4079 dvp == ndp->ni_topdir ||
4083 fpl->tvp_seqc = vn_seqc_read_any(dvp);
4084 if (seqc_in_modify(fpl->tvp_seqc)) {
4085 return (cache_fpl_aborted(fpl));
4090 if ((dvp->v_vflag & VV_ROOT) != 0) {
4093 * The opposite of climb mount is needed here.
4095 return (cache_fpl_aborted(fpl));
4098 ncp = atomic_load_ptr(&dvp->v_cache_dd);
4100 return (cache_fpl_aborted(fpl));
4103 nc_flag = atomic_load_char(&ncp->nc_flag);
4104 if ((nc_flag & NCF_ISDOTDOT) != 0) {
4105 if ((nc_flag & NCF_NEGATIVE) != 0)
4106 return (cache_fpl_aborted(fpl));
4107 fpl->tvp = ncp->nc_vp;
4109 fpl->tvp = ncp->nc_dvp;
4112 if (!cache_ncp_canuse(ncp)) {
4113 return (cache_fpl_aborted(fpl));
4116 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
4117 if (seqc_in_modify(fpl->tvp_seqc)) {
4118 return (cache_fpl_partial(fpl));
4121 counter_u64_add(dotdothits, 1);
4125 static int __noinline
4126 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
4131 nc_flag = atomic_load_char(&ncp->nc_flag);
4132 MPASS((nc_flag & NCF_NEGATIVE) != 0);
4134 * If they want to create an entry we need to replace this one.
4136 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
4139 * This should call something similar to
4140 * cache_fplookup_final_modifying.
4142 return (cache_fpl_partial(fpl));
4144 neg_promote = cache_neg_hit_prep(ncp);
4145 if (!cache_ncp_canuse(ncp)) {
4146 cache_neg_hit_abort(ncp);
4147 return (cache_fpl_partial(fpl));
4149 if (__predict_false((nc_flag & NCF_WHITE) != 0)) {
4150 cache_neg_hit_abort(ncp);
4151 return (cache_fpl_partial(fpl));
4154 return (cache_fplookup_negative_promote(fpl, ncp, hash));
4156 cache_neg_hit_finish(ncp);
4157 cache_fpl_smr_exit(fpl);
4158 return (cache_fpl_handled(fpl, ENOENT));
4162 cache_fplookup_next(struct cache_fpl *fpl)
4164 struct componentname *cnp;
4165 struct namecache *ncp;
4166 struct vnode *dvp, *tvp;
4173 if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.')) {
4174 return (cache_fplookup_dot(fpl));
4177 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
4179 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
4180 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
4181 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
4186 * If there is no entry we have to punt to the slow path to perform
4187 * actual lookup. Should there be nothing with this name a negative
4188 * entry will be created.
4190 if (__predict_false(ncp == NULL)) {
4191 return (cache_fpl_partial(fpl));
4194 tvp = atomic_load_ptr(&ncp->nc_vp);
4195 nc_flag = atomic_load_char(&ncp->nc_flag);
4196 if ((nc_flag & NCF_NEGATIVE) != 0) {
4197 return (cache_fplookup_neg(fpl, ncp, hash));
4200 if (!cache_ncp_canuse(ncp)) {
4201 return (cache_fpl_partial(fpl));
4205 fpl->tvp_seqc = vn_seqc_read_any(tvp);
4206 if (seqc_in_modify(fpl->tvp_seqc)) {
4207 return (cache_fpl_partial(fpl));
4210 if (!cache_fplookup_vnode_supported(tvp)) {
4211 return (cache_fpl_partial(fpl));
4214 counter_u64_add(numposhits, 1);
4215 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
4220 cache_fplookup_mp_supported(struct mount *mp)
4225 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
4231 * Walk up the mount stack (if any).
4233 * Correctness is provided in the following ways:
4234 * - all vnodes are protected from freeing with SMR
4235 * - struct mount objects are type stable making them always safe to access
4236 * - stability of the particular mount is provided by busying it
4237 * - relationship between the vnode which is mounted on and the mount is
4238 * verified with the vnode sequence counter after busying
4239 * - association between root vnode of the mount and the mount is protected
4242 * From that point on we can read the sequence counter of the root vnode
4243 * and get the next mount on the stack (if any) using the same protection.
4245 * By the end of successful walk we are guaranteed the reached state was
4246 * indeed present at least at some point which matches the regular lookup.
4248 static int __noinline
4249 cache_fplookup_climb_mount(struct cache_fpl *fpl)
4251 struct mount *mp, *prev_mp;
4252 struct mount_pcpu *mpcpu, *prev_mpcpu;
4257 vp_seqc = fpl->tvp_seqc;
4259 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
4260 mp = atomic_load_ptr(&vp->v_mountedhere);
4266 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
4267 if (prev_mp != NULL)
4268 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4269 return (cache_fpl_partial(fpl));
4271 if (prev_mp != NULL)
4272 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4273 if (!vn_seqc_consistent(vp, vp_seqc)) {
4274 vfs_op_thread_exit_crit(mp, mpcpu);
4275 return (cache_fpl_partial(fpl));
4277 if (!cache_fplookup_mp_supported(mp)) {
4278 vfs_op_thread_exit_crit(mp, mpcpu);
4279 return (cache_fpl_partial(fpl));
4281 vp = atomic_load_ptr(&mp->mnt_rootvnode);
4282 if (vp == NULL || VN_IS_DOOMED(vp)) {
4283 vfs_op_thread_exit_crit(mp, mpcpu);
4284 return (cache_fpl_partial(fpl));
4286 vp_seqc = vn_seqc_read_any(vp);
4287 if (seqc_in_modify(vp_seqc)) {
4288 vfs_op_thread_exit_crit(mp, mpcpu);
4289 return (cache_fpl_partial(fpl));
4293 mp = atomic_load_ptr(&vp->v_mountedhere);
4298 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4300 fpl->tvp_seqc = vp_seqc;
4305 cache_fplookup_need_climb_mount(struct cache_fpl *fpl)
4313 * Hack: while this is a union, the pointer tends to be NULL so save on
4316 mp = atomic_load_ptr(&vp->v_mountedhere);
4319 if (vp->v_type == VDIR)
4327 * The code was originally copy-pasted from regular lookup and despite
4328 * clean ups leaves performance on the table. Any modifications here
4329 * must take into account that in case off fallback the resulting
4330 * nameidata state has to be compatible with the original.
4333 cache_fplookup_parse(struct cache_fpl *fpl)
4335 struct nameidata *ndp;
4336 struct componentname *cnp;
4343 * Search a new directory.
4345 * The last component of the filename is left accessible via
4346 * cnp->cn_nameptr for callers that need the name. Callers needing
4347 * the name set the SAVENAME flag. When done, they assume
4348 * responsibility for freeing the pathname buffer.
4350 for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++)
4352 cnp->cn_namelen = cp - cnp->cn_nameptr;
4353 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4354 cache_fpl_smr_exit(fpl);
4355 return (cache_fpl_handled(fpl, ENAMETOOLONG));
4357 ndp->ni_pathlen -= cnp->cn_namelen;
4358 KASSERT(ndp->ni_pathlen <= PATH_MAX,
4359 ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
4363 * Replace multiple slashes by a single slash and trailing slashes
4364 * by a null. This must be done before VOP_LOOKUP() because some
4365 * fs's don't know about trailing slashes. Remember if there were
4366 * trailing slashes to handle symlinks, existing non-directories
4367 * and non-existing files that won't be directories specially later.
4369 while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
4375 * Regular lookup performs the following:
4376 * *ndp->ni_next = '\0';
4377 * cnp->cn_flags |= TRAILINGSLASH;
4379 * Which is problematic since it modifies data read
4380 * from userspace. Then if fast path lookup was to
4381 * abort we would have to either restore it or convey
4382 * the flag. Since this is a corner case just ignore
4383 * it for simplicity.
4385 return (cache_fpl_partial(fpl));
4391 * Check for degenerate name (e.g. / or "")
4392 * which is a way of talking about a directory,
4393 * e.g. like "/." or ".".
4396 * Another corner case handled by the regular lookup
4398 if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
4399 return (cache_fpl_partial(fpl));
4405 cache_fplookup_parse_advance(struct cache_fpl *fpl)
4407 struct nameidata *ndp;
4408 struct componentname *cnp;
4413 cnp->cn_nameptr = ndp->ni_next;
4414 while (*cnp->cn_nameptr == '/') {
4421 * See the API contract for VOP_FPLOOKUP_VEXEC.
4423 static int __noinline
4424 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
4430 dvp_seqc = fpl->dvp_seqc;
4433 * Hack: they may be looking up foo/bar, where foo is a
4434 * regular file. In such a case we need to turn ENOTDIR,
4435 * but we may happen to get here with a different error.
4437 if (dvp->v_type != VDIR) {
4439 * The check here is predominantly to catch
4440 * EOPNOTSUPP from dead_vnodeops. If the vnode
4441 * gets doomed past this point it is going to
4442 * fail seqc verification.
4444 if (VN_IS_DOOMED(dvp)) {
4445 return (cache_fpl_aborted(fpl));
4451 * Hack: handle O_SEARCH.
4453 * Open Group Base Specifications Issue 7, 2018 edition states:
4454 * If the access mode of the open file description associated with the
4455 * file descriptor is not O_SEARCH, the function shall check whether
4456 * directory searches are permitted using the current permissions of
4457 * the directory underlying the file descriptor. If the access mode is
4458 * O_SEARCH, the function shall not perform the check.
4460 * Regular lookup tests for the NOEXECCHECK flag for every path
4461 * component to decide whether to do the permission check. However,
4462 * since most lookups never have the flag (and when they do it is only
4463 * present for the first path component), lockless lookup only acts on
4464 * it if there is a permission problem. Here the flag is represented
4465 * with a boolean so that we don't have to clear it on the way out.
4467 * For simplicity this always aborts.
4468 * TODO: check if this is the first lookup and ignore the permission
4469 * problem. Note the flag has to survive fallback (if it happens to be
4473 return (cache_fpl_aborted(fpl));
4478 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4479 error = cache_fpl_aborted(fpl);
4481 cache_fpl_partial(fpl);
4485 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4486 error = cache_fpl_aborted(fpl);
4488 cache_fpl_smr_exit(fpl);
4489 cache_fpl_handled(fpl, error);
4497 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
4499 struct nameidata *ndp;
4500 struct componentname *cnp;
4504 error = CACHE_FPL_FAILED;
4508 cache_fpl_checkpoint(fpl, &fpl->snd);
4511 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
4512 if (seqc_in_modify(fpl->dvp_seqc)) {
4513 cache_fpl_aborted(fpl);
4516 mp = atomic_load_ptr(&fpl->dvp->v_mount);
4517 if (!cache_fplookup_mp_supported(mp)) {
4518 cache_fpl_aborted(fpl);
4522 VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
4525 error = cache_fplookup_parse(fpl);
4526 if (__predict_false(error != 0)) {
4530 VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
4532 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
4533 if (__predict_false(error != 0)) {
4534 error = cache_fplookup_failed_vexec(fpl, error);
4538 if (__predict_false(cache_fpl_isdotdot(cnp))) {
4539 error = cache_fplookup_dotdot(fpl);
4540 if (__predict_false(error != 0)) {
4544 error = cache_fplookup_next(fpl);
4545 if (__predict_false(error != 0)) {
4549 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
4551 if (cache_fplookup_need_climb_mount(fpl)) {
4552 error = cache_fplookup_climb_mount(fpl);
4553 if (__predict_false(error != 0)) {
4559 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
4561 if (cache_fpl_islastcn(ndp)) {
4562 error = cache_fplookup_final(fpl);
4566 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
4567 error = cache_fpl_aborted(fpl);
4571 fpl->dvp = fpl->tvp;
4572 fpl->dvp_seqc = fpl->tvp_seqc;
4574 cache_fplookup_parse_advance(fpl);
4575 cache_fpl_checkpoint(fpl, &fpl->snd);
4578 switch (fpl->status) {
4579 case CACHE_FPL_STATUS_UNSET:
4580 __assert_unreachable();
4582 case CACHE_FPL_STATUS_PARTIAL:
4583 cache_fpl_smr_assert_entered(fpl);
4584 return (cache_fplookup_partial_setup(fpl));
4585 case CACHE_FPL_STATUS_ABORTED:
4587 cache_fpl_smr_exit(fpl);
4588 return (CACHE_FPL_FAILED);
4589 case CACHE_FPL_STATUS_HANDLED:
4590 MPASS(error != CACHE_FPL_FAILED);
4591 cache_fpl_smr_assert_not_entered(fpl);
4592 if (__predict_false(error != 0)) {
4595 cache_fpl_cleanup_cnp(cnp);
4598 ndp->ni_dvp = fpl->dvp;
4599 ndp->ni_vp = fpl->tvp;
4600 if (cnp->cn_flags & SAVENAME)
4601 cnp->cn_flags |= HASBUF;
4603 cache_fpl_cleanup_cnp(cnp);
4609 * Fast path lookup protected with SMR and sequence counters.
4611 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
4613 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
4616 * Traditional vnode lookup conceptually looks like this:
4622 * vn_unlock(current);
4629 * Each jump to the next vnode is safe memory-wise and atomic with respect to
4630 * any modifications thanks to holding respective locks.
4632 * The same guarantee can be provided with a combination of safe memory
4633 * reclamation and sequence counters instead. If all operations which affect
4634 * the relationship between the current vnode and the one we are looking for
4635 * also modify the counter, we can verify whether all the conditions held as
4636 * we made the jump. This includes things like permissions, mount points etc.
4637 * Counter modification is provided by enclosing relevant places in
4638 * vn_seqc_write_begin()/end() calls.
4640 * Thus this translates to:
4643 * dvp_seqc = seqc_read_any(dvp);
4644 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
4648 * tvp_seqc = seqc_read_any(tvp);
4649 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
4651 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
4653 * dvp = tvp; // we know nothing of importance has changed
4654 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
4658 * vget(); // secure the vnode
4659 * if (!seqc_consistent(tvp, tvp_seqc) // final check
4661 * // at this point we know nothing has changed for any parent<->child pair
4662 * // as they were crossed during the lookup, meaning we matched the guarantee
4663 * // of the locked variant
4666 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
4667 * - they are called while within vfs_smr protection which they must never exit
4668 * - EAGAIN can be returned to denote checking could not be performed, it is
4669 * always valid to return it
4670 * - if the sequence counter has not changed the result must be valid
4671 * - if the sequence counter has changed both false positives and false negatives
4672 * are permitted (since the result will be rejected later)
4673 * - for simple cases of unix permission checks vaccess_vexec_smr can be used
4675 * Caveats to watch out for:
4676 * - vnodes are passed unlocked and unreferenced with nothing stopping
4677 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
4678 * to use atomic_load_ptr to fetch it.
4679 * - the aforementioned object can also get freed, meaning absent other means it
4680 * should be protected with vfs_smr
4681 * - either safely checking permissions as they are modified or guaranteeing
4682 * their stability is left to the routine
4685 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
4688 struct cache_fpl fpl;
4691 struct componentname *cnp;
4692 struct nameidata_saved orig;
4695 MPASS(ndp->ni_lcf == 0);
4697 fpl.status = CACHE_FPL_STATUS_UNSET;
4699 fpl.cnp = &ndp->ni_cnd;
4700 MPASS(curthread == fpl.cnp->cn_thread);
4702 if ((fpl.cnp->cn_flags & SAVESTART) != 0)
4703 MPASS(fpl.cnp->cn_nameiop != LOOKUP);
4705 if (!cache_can_fplookup(&fpl)) {
4706 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
4707 *status = fpl.status;
4708 return (EOPNOTSUPP);
4711 cache_fpl_checkpoint(&fpl, &orig);
4713 cache_fpl_smr_enter_initial(&fpl);
4714 fpl.fsearch = false;
4715 pwd = pwd_get_smr();
4717 ndp->ni_rootdir = pwd->pwd_rdir;
4718 ndp->ni_topdir = pwd->pwd_jdir;
4721 cnp->cn_nameptr = cnp->cn_pnbuf;
4722 if (cnp->cn_pnbuf[0] == '/') {
4723 cache_fpl_handle_root(ndp, &dvp);
4724 ndp->ni_resflags |= NIRES_ABS;
4726 if (ndp->ni_dirfd == AT_FDCWD) {
4727 dvp = pwd->pwd_cdir;
4729 error = cache_fplookup_dirfd(&fpl, &dvp);
4730 if (__predict_false(error != 0)) {
4736 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
4738 error = cache_fplookup_impl(dvp, &fpl);
4740 cache_fpl_smr_assert_not_entered(&fpl);
4741 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
4743 *status = fpl.status;
4744 switch (fpl.status) {
4745 case CACHE_FPL_STATUS_UNSET:
4746 __assert_unreachable();
4748 case CACHE_FPL_STATUS_HANDLED:
4749 SDT_PROBE3(vfs, namei, lookup, return, error,
4750 (error == 0 ? ndp->ni_vp : NULL), true);
4752 case CACHE_FPL_STATUS_PARTIAL:
4755 * Status restored by cache_fplookup_partial_setup.
4758 case CACHE_FPL_STATUS_ABORTED:
4759 cache_fpl_restore_abort(&fpl, &orig);